ci: add a manual flake hunt for #203 (Linux x Node version)

#203 has been seen four times, always on Linux CI, never once on macOS across 1000+
runs in two independent experiments. The documented next step is a Linux + Node 24
reproduction, and there is currently no way to run one without reddening an unrelated
PR and waiting for chance.

workflow_dispatch only — it never runs on push or pull_request, so it costs nothing
until someone asks for it.

Inputs are the two suspected variables:

- `node` is a choice (24 / 22 / 26) rather than pinned, because 22-vs-24 is a
  comparison worth running deliberately. A previous Linux VM attempt was invalidated
  by Node 22's `node:sqlite` ExperimentalWarning landing on stderr, which the boot gate
  read as "server did not start" — ~23 of 50 runs failed spuriously. The workflow says
  so, so the next person interprets a 22 result against that confound instead of
  rediscovering it.
- `concurrency` is the knob that actually reproduces, not round count: test() is
  fire-and-forget for async bodies, so ONE suite run already spawns 15 concurrent
  server.mjs children across 11 ltBoot tests. Several suites at once is what multiplies
  cross-process contention.

Classification is anchored on the failure marker AND the test name, which is not
cosmetic. ltDiag samples the first 900B of child stdout (#204) — the boot banner — so a
log where only the GATE test failed still contains "Local tools: ON". Validated against
real logs rather than reasoned about: 3 clean runs + 1 gate-mutation run gave

    unanchored 'Local tools'  -> 1 hit, attributed to the WRONG category
    anchored                  -> gate=1, announce=0        (correct)
    totals                    -> total=4 clean=3           (correct)

My first draft had the unanchored form with a comment claiming the failure mode was
"matches every log". That was wrong — the real defect is cross-category attribution,
caused by the diagnostic quoting the child's own output. Found by running the logic on
real logs; the comment now states the measured reason.

The job does NOT fail on a reproduction: catching the flake is the goal, and a red X
would read as "the hunt is broken" rather than "the flake was caught". Results go to
the step summary; logs upload as an artifact.

server.mjs: unchanged (CI-only; ALIGNMENT.md requires no cli.js citation).
Verified: YAML parses, all three run blocks pass `bash -n`, classify logic exercised
against real suite output.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017gbqUZ8HfBZpjjbzQ85oH8

.github/workflows/flake-hunt.yml

@@ -0,0 +1,145 @@
+name: flake hunt (#203)
+
+# Manual only. This never runs on push or PR — it exists to reproduce a flake that has
+# been seen FOUR times, always on Linux CI, and never once on macOS (0/1000+ across two
+# independent experiments). The suspected variables are the platform and the Node version,
+# so both are inputs.
+#
+# Background: an attempt to run this on a Linux VM was invalidated because Node 22 emits a
+# `node:sqlite` ExperimentalWarning on stderr, which the boot gate read as "server did not
+# start" — ~23 of 50 runs failed spuriously. That is exactly why `node` is an input here
+# rather than pinned: 22-vs-24 is a comparison worth running deliberately, not an accident
+# to stumble into. Interpret any 22 result against that known confound.
+#
+# Concurrency is the load-bearing knob, not round count. test-features.mjs makes test()
+# fire-and-forget for async bodies, so one suite run already spawns 15 concurrent
+# `server.mjs` children across 11 ltBoot tests; running several suites at once is what
+# multiplies cross-process contention. See AGENTS.md § "Testing: reaching faults inside
+# server.mjs".
+
+on:
+  workflow_dispatch:
+    inputs:
+      node:
+        description: 'Node major version (24 = what CI and released OCP run; 22 has the SQLite stderr confound)'
+        default: '24'
+        type: choice
+        options: ['24', '22', '26']
+      rounds:
+        description: 'Rounds; each round runs <concurrency> suites at once and waits'
+        default: '25'
+      concurrency:
+        description: 'Concurrent suite processes per round (this is the knob that reproduces)'
+        default: '4'
+
+jobs:
+  hunt:
+    name: 'hunt: node ${{ inputs.node }} x ${{ inputs.rounds }} rounds x ${{ inputs.concurrency }}'
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: ${{ inputs.node }}
+
+      - name: Record the environment
+        run: |
+          set -euo pipefail
+          {
+            echo "node       $(node --version)"
+            echo "kernel     $(uname -srm)"
+            echo "cpus       $(nproc)"
+            echo "ephemeral  $(cat /proc/sys/net/ipv4/ip_local_port_range)"
+          } | tee env.txt
+          # The ephemeral range matters: the fixed ports this suite used before #204
+          # (39321-39364) sit INSIDE Linux's default 32768-60999 but OUTSIDE macOS's
+          # 49152-65535, which is independent support for "CI is more exposed than a Mac".
+
+      - name: Hunt
+        run: |
+          set -uo pipefail          # NOT -e: a failing suite run is the DATA, not an error
+          mkdir -p logs
+          R=${{ inputs.rounds }}
+          C=${{ inputs.concurrency }}
+          for r in $(seq 1 "$R"); do
+            for c in $(seq 1 "$C"); do
+              npm test > "logs/r${r}c${c}.log" 2>&1 &
+            done
+            wait
+            printf '.'
+          done
+          echo
+
+      - name: Classify
+        if: always()
+        run: |
+          set -uo pipefail
+
+          # EVERY pattern is anchored on the failure marker AND on the test's name, because
+          # the categories cross-contaminate otherwise. ltDiag samples the first 900B of the
+          # child's stdout (PR #204), which is the boot banner — so a log where only the GATE
+          # test failed still contains the string "Local tools: ON", and an unanchored
+          # `grep -l 'Local tools'` scores it as an announcement failure too. Verified against
+          # real logs: 3 clean runs + 1 gate-mutation run gave
+          #   unanchored 'Local tools' -> 1 hit, attributed to the WRONG category
+          #   anchored                 -> gate=1, announce=0     (correct)
+          # A diagnostic that quotes the program's own output makes substring classification
+          # unsound; anchor on what the runner prints, not on what the child printed.
+          n() { grep -l "✗.*$1" logs/*.log 2>/dev/null | wc -l | tr -d ' '; }
+
+          total=$(ls logs/*.log | wc -l | tr -d ' ')
+          clean=$(grep -l ', 0 failed ===' logs/*.log 2>/dev/null | wc -l | tr -d ' ')
+
+          # A category scoring 0 has NOT been proven absent — it may simply not fire in this
+          # configuration. Only a non-zero count is evidence.
+          gate=$(n 'boot gate REFUSES')
+          tui=$(n 'announced INERT')
+          announce=$(n 'BOOTS past the gate')
+          # These two are runtime errors rather than named tests, so they are matched on the
+          # error string anywhere in the log, not on a ✗ line.
+          eaddr=$(grep -l 'EADDRINUSE' logs/*.log 2>/dev/null | wc -l | tr -d ' ')
+          enotempty=$(grep -l 'ENOTEMPTY' logs/*.log 2>/dev/null | wc -l | tr -d ' ')
+
+          {
+            echo "## flake hunt — node ${{ inputs.node }}"
+            echo
+            echo '```'
+            cat env.txt
+            echo '```'
+            echo
+            echo "| category | runs affected |"
+            echo "|---|---|"
+            echo "| **clean (0 failed)** | **$clean / $total** |"
+            echo "| #203 boot gate | $gate |"
+            echo "| #199 TUI inert warning | $tui |"
+            echo "| \`Local tools\` announce | $announce |"
+            echo "| EADDRINUSE | $eaddr |"
+            echo "| ENOTEMPTY teardown | $enotempty |"
+            echo
+            echo "A zero is not proof of absence — only a non-zero count is evidence."
+          } >> "$GITHUB_STEP_SUMMARY"
+
+          # Surface the diagnostics inline so a repro is readable without downloading.
+          # ltDiag prints exit/signal/closed/closeMs/node plus both streams (PR #204).
+          if [ "$clean" -lt "$total" ]; then
+            echo "### first failing run" >> "$GITHUB_STEP_SUMMARY"
+            echo '```' >> "$GITHUB_STEP_SUMMARY"
+            grep -h '✗' $(grep -L ', 0 failed ===' logs/*.log | head -1) | head -20 >> "$GITHUB_STEP_SUMMARY"
+            echo '```' >> "$GITHUB_STEP_SUMMARY"
+          fi
+
+      - name: Upload logs
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: flake-hunt-node${{ inputs.node }}
+          path: |
+            logs/
+            env.txt
+          retention-days: 14
+
+      # This job does NOT fail on a reproduction. Reproducing is the goal, and a red X
+      # would read as "the hunt is broken" rather than "the flake was caught".

.github/workflows/release.yml

@@ -21,24 +21,31 @@ jobs:
       - name: Extract CHANGELOG section
         id: notes
         run: |
+          set -euo pipefail
           VERSION="${{ steps.ver.outputs.version }}"
+          NOTES=/tmp/release-notes.md
           # Extract section for this version from CHANGELOG.md
           # Pattern: "## v${VERSION}" through the next "## " or EOF
           if [ ! -f CHANGELOG.md ]; then
             echo "No CHANGELOG.md found; using minimal release notes"
-            echo "notes=Release v${VERSION}" >> $GITHUB_OUTPUT
+            # MUST write the file, not just an output: the create step consumes a FILE, so an
+            # early exit here used to leave --notes-file pointing at a path that never existed,
+            # turning "degrade to minimal notes" into a failed release job (#202).
+            echo "Release v${VERSION}" > "$NOTES"
+            echo "notes_file=$NOTES" >> $GITHUB_OUTPUT
             exit 0
           fi
           awk -v ver="v${VERSION}" '
             $0 ~ "^## " ver { found=1; print; next }
             found && /^## v/ { exit }
             found { print }
-          ' CHANGELOG.md > /tmp/release-notes.md
-          if [ ! -s /tmp/release-notes.md ]; then
+          ' CHANGELOG.md > "$NOTES"
+          if [ ! -s "$NOTES" ]; then
             echo "No matching section in CHANGELOG for v${VERSION}; using minimal notes"
-            echo "Release v${VERSION}" > /tmp/release-notes.md
+            echo "Release v${VERSION}" > "$NOTES"
           fi
-          echo "notes_file=/tmp/release-notes.md" >> $GITHUB_OUTPUT
+          echo "--- release notes (${VERSION}) ---"; cat "$NOTES"
+          echo "notes_file=$NOTES" >> $GITHUB_OUTPUT
       - name: Create GitHub Release
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -49,5 +56,5 @@ jobs:
           fi
           gh release create "v${{ steps.ver.outputs.version }}" \
             --title "v${{ steps.ver.outputs.version }}" \
-            --notes-file /tmp/release-notes.md \
+            --notes-file "${{ steps.notes.outputs.notes_file }}" \
             --latest

AGENTS.md

@@ -30,6 +30,7 @@ Runtime: Node.js (ESM, `.mjs` throughout). No build step. No bundler. `server.mj
 
 - `server.mjs` — the proxy itself; every request path lives here. Governed by `ALIGNMENT.md`.
 - `models.json` — single source of truth for model IDs, aliases, and context windows. See ADR 0003.
+- `models.schema.json` — the schema `models.json` declares in its `$schema`. CI validates the SPOT against it (`test-features.mjs`) using the repo's own `validateJsonSchema`, so a malformed entry fails the build instead of surfacing downstream in OpenClaw.
 - `setup.mjs` — first-time installer; reads `models.json` to derive bootstrap config.
 - `scripts/sync-openclaw.mjs` — idempotent OpenClaw registry sync invoked by `ocp update`. See ADR 0004.
 - `ocp` — user-facing CLI (install, update, start, stop, status, logs, etc.).
@@ -52,6 +53,22 @@ Runtime: Node.js (ESM, `.mjs` throughout). No build step. No bundler. `server.mj
 
 ---
 
+## Testing: reaching faults inside `server.mjs`
+
+`test-features.mjs` cannot `import` `server.mjs` (it calls `server.listen()` at top level), and that has twice led to the wrong conclusion that a class of bug is untestable. It isn't. Read this before writing "no regression test is possible here".
+
+**There is a real live-server fixture.** `ltBoot(env, dir, nodeArgs)` (around `test-features.mjs:990`) spawns the actual `server.mjs` as a child with a **fake `claude` binary**, so integration tests cost no quota. `ltPost` / `ltPostStatus` / `ltWait` / `ltFreePort` round it out. It already covers boot gates, cache-epoch invalidation across two boots sharing one SQLite store, and system-prompt capture.
+
+**`--stack-size` is a fault lever.** `ltBoot`'s third argument passes V8 flags to the child, which puts recursion- and argument-count-limited failures in reach at a much smaller input. `#193` needed a *synchronous throw* deep inside `spawnClaudeProcess`; `buildCliArgs` does `args.push("--allowedTools", ...ALLOWED_TOOLS)`, and under `--stack-size=200` that spread throws at ~24k elements instead of ~124k — which is what brings the trigger under Linux's `MAX_ARG_STRLEN` (131072 bytes for a single env string) so the test runs in CI rather than skipping. **No production fault hook was needed.**
+
+Three rules that made it hold up, all learned the hard way:
+
+- **Discover the threshold in a child under the same flags**, never in the test process — the parent's stack is not the one that matters.
+- **Assert that the fault actually fired**, not just that the outcome looks right. `#193` asserts HTTP 500 *and* that the body carries `call stack size exceeded`; a control mutation (trigger neutered, bug still present) proves the test fails rather than passing vacuously.
+- **Wait for the thing you are about to assert**, not a proxy for it. Waiting on `listening on` and then asserting a different line is a race (`#199`); waiting for the process to *exit* and then reading its `stderr` is another, because a terminated child's pipes may still hold unread data (`#203` — wait for the stdio to close, not for the exit).
+
+Allocate ports with `ltFreePort()`. Fixed ports have caused at least one flake here.
+
 ## Release protocol
 
 OCP follows the machine-readable `release_kit:` overlay in `CLAUDE.md` (Iron Rule 5.5). Before any version bump or tag push, re-read that YAML block and walk every item in `new_feature_doc_expectations` and `bootstrap_quirk_policy`. Tag push triggers `.github/workflows/release.yml`, which creates the GitHub Release automatically — do not create the release manually.

CHANGELOG.md

@@ -1,5 +1,112 @@
 # Changelog
 
+## Unreleased
+
+### Changed
+
+- **`maxTokens` now matches the CLI registry instead of a uniform 16384 (#195).** Every Opus entry and `claude-sonnet-5` go to **64000**, `claude-sonnet-4-6` and `claude-haiku-4-5` to **32000** — the `max_output_tokens.default` each model declares in the compiled CLI 2.1.220 registry. This corrects **advertised metadata only**: `models.json` is the SPOT (ADR 0003) and every value in it should be the truth about the model. **It changes nothing about how OCP behaves.** OCP never enforces `maxTokens` — `buildCliArgs` passes no output-token flag to the CLI at all — and OpenClaw addresses a local OCP over `openai-completions`, whose request field (`max_completion_tokens`) appears nowhere in this repo. The value is consumed only by clients that choose to honour it, via `setup.mjs` / `scripts/sync-openclaw.mjs` / `ocp-connect`. **Expect no change in answer length or quota burn.** `ocp-connect`'s independent family table moves to the floor over each family's current `models.json` members (opus 64000, sonnet 32000, haiku 32000), since prefixes cannot distinguish versions. Its unknown-id fallback stays at **8192** — the registry's global minimum, held by `claude-3-5-haiku` and `claude-3-5-sonnet`, which are the only real ids that reach it (`claude-3-5-*` matches no family prefix).
+
+## v3.25.0 — 2026-07-27
+
+Minor release. Headline: **Claude Opus 5** joins the model list and the `opus` alias now resolves to it. Alongside it, two `server.mjs` correctness fixes found by review rather than by users — a monotonic in-flight-counter leak, and cache keys that were hashing the alias string instead of the model it resolves to. The three TUI/prompt fixes that landed after v3.24.0 are included here too.
+
+Every code PR carried an independent fresh-context reviewer (Iron Rule 10); #192 additionally went through the external codex gate, which is what surfaced the cache-key defect. No new endpoint, no new env var, no new `cli.js` wire behavior.
+
+### Added
+
+- **Claude Opus 5 (`claude-opus-5`) (#192).** New `models.json` entry — `/v1/models` goes 6 → 7 and OpenClaw picks it up on the next `ocp update` (via `scripts/sync-openclaw.mjs`). Verified against the installed CLI rather than assumed: the compiled `claude` 2.1.220 bundle carries `latest_per_family:{fable:"claude-fable-5",opus:"claude-opus-5",sonnet:"claude-sonnet-5",haiku:"claude-haiku-4-5"}`. Availability confirmed with a live `claude -p --model claude-opus-5` completion on the subscription pool. `claude-opus-4-8` is retained for pinning.
+- `contextWindow` is deliberately **200000**, not Opus 5's native 1M. Two reasons, both verified: (1) `MAX_PROMPT_CHARS` is a **single global** budget — `derivePromptCharBudget` takes `max(contextWindow) × 3` across *all* entries (`lib/prompt.mjs`), so a 1M entry would raise the truncation ceiling to 3,000,000 chars for `claude-haiku-4-5` too, which is genuinely 200k native, converting clean OCP-side truncation into an upstream API rejection; (2) OpenClaw scales its history budget linearly off this value (`contextWindow × maxHistoryShare × SAFETY_MARGIN` = `× 0.6`, plus an oversized-message threshold at `× 0.5`, per `compaction-planning` in OpenClaw 2026.7.1), and its own bundled registry hardcodes 200000 for Claude — the upstream request to raise it to 1M ([openclaw#22979](https://github.com/openclaw/openclaw/issues/22979)) was closed *not planned*. A new regression test pins the invariant so a future 1M entry has to be a deliberate, reviewed change. Raising it for real needs per-model budgets — tracked separately, ADR-level.
+
+### Changed
+
+- **The `opus` alias now resolves to `claude-opus-5` instead of `claude-opus-4-8` (#192).** Every request that names `opus` — and OpenClaw's opus entry — moves to Opus 5 on upgrade. This mirrors what the CLI itself defaults to (`latest_per_family.opus` above), the same reasoning as #168's `sonnet` → `claude-sonnet-5` repoint in v3.23.0. **Pricing is unchanged** ($5/$25 per MTok; CLI registry `pricing:"tier_5_25"` for both Opus 4.8 and Opus 5), so this carries no cost change. Pin `claude-opus-4-8` explicitly to stay on the previous model.
+
+### Fixed
+
+- **Cache keys hashed the alias, not the model it resolves to (#194).** `model` enters `cacheHash` exactly as the client sent it, so a request for `"opus"` was cached under the literal `"opus"` — and since `models.json` is read once at boot while the SQLite `response_cache` outlives a restart, repointing an alias kept serving the **old model's** answers under it until TTL expiry. That would have silently defeated this release's own `opus` repoint for anyone running with the cache on. All three call sites now hash `Object.hasOwn(MODEL_MAP, model) ? MODEL_MAP[model] : model`, which fixes the normal, structured **and** single-flight keys at once and covers `legacyAliases` for free, with precise invalidation (only the repointed alias's entries change key) rather than a whole-cache flush. (`hasOwn` rather than a bare lookup: `MODEL_MAP` is a plain object, so `MODEL_MAP["__proto__"]` would return a truthy *object* and not even fall through to the `|| model` guard.) **Also closes a latent gap from #177:** the structured and dedup keys never passed `configEpoch` at all, so a `CLAUDE_SYSTEM_PROMPT` change — the original #176 scenario — kept serving structured answers composed under the old config, live since #153. Found by the external codex review of #192.
+- **In-flight request counter leaked permanently on a pre-spawn throw (#193, reported by @konceptnet in #180).** `stats.activeRequests` was incremented ~40 lines before the child spawn, while its only decrement is reached through that process's events — so any synchronous throw in between (`buildCliArgs`, env assembly, the spawn decision, or `spawn()` itself) leaked `+1` forever, and `/health` and `/status` over-reported in-flight work monotonically. The increment now sits immediately after `activeProcesses.add(proc)`, structurally pairing it with the decrement; no reconciliation pass and no try/catch. Observability-only field, so no admission decision changes.
+- **TUI: the host `CLAUDE.md` could leak into proxied turns (#187, contributed by @sumlin).** The TUI pane now spawns with `--safe-mode`.
+- **TUI: `shift+tab to cycle` is accepted as an input-ready marker (#188, contributed by @sumlin).** Claude renders one of two ready-state footers depending on the build — the classic `? for shortcuts` hint, or `shift+tab to cycle` (as part of `⏵⏵ bypass permissions on (shift+tab to cycle)`) on newer 2.1.x. The matcher only knew the classic string, so on those builds it silently reported "never ready": every boot timed out with `tui_pane_not_ready`, and with the warm pool on, every pre-boot failed.
+- **`OCP_LOCAL_TOOLS` no longer hard-codes a tool list in its wrapper (#191, closes #185).** The positive wrapper claimed a fixed set of tools regardless of `CLAUDE_ALLOWED_TOOLS`, so a narrowed tool surface was described inaccurately to the model.
+
+### Testing
+
+- The response-cache and counter fixes ship with **mutation-proven integration tests** built on the existing `ltBoot` child-process fixture (real `server.mjs`, fake `claude` binary, so no quota cost). The counter test reaches a synchronous fault *inside* `spawnClaudeProcess` from environment alone — no production fault hook — by running the child under `--stack-size=200` to lower the spread-throw threshold enough to fit Linux's `MAX_ARG_STRLEN`. Suite: **447 → 457** across this release (per PR: #188 +1, #187 +1, #191 +0, #194 +4, #192 +3, #193 +1).
+
+## v3.24.0 — 2026-07-21
+
+Minor release. Headline: two long-requested **OpenAI-compat features** land — **multimodal vision** (`image_url` parts) and **structured outputs** (`response_format` / JSON schema). Also: the prompt-char budget now derives from the model SPOT instead of a hand-set constant, an agentic-turn bug that dropped the model's final answer is fixed, and `OCP_LOCAL_TOOLS` supports the OpenClaw-backend use case. Four of the six landed from external contributors (@vvlasy-openclaw). Every code PR carried a fresh-context reviewer (Iron Rule 10); no new endpoint, no new `cli.js` wire behavior.
+
+### Added
+
+- **Multimodal vision — OpenAI `image_url` parts (#154, contributed by @vvlasy-openclaw).** `/v1/chat/completions` forwards OpenAI `image_url` content parts to `claude` as native Anthropic image blocks via `--input-format stream-json` (the CLI's own contract — no invented wire shape, verified live). Base64 `data:` URIs by default; remote `http(s)` URLs are off unless `CLAUDE_IMAGE_ALLOW_URL=1` (and even then OCP never fetches them — no SSRF surface). Byte/count caps (`CLAUDE_MAX_IMAGE_BYTES`, `CLAUDE_MAX_IMAGES`, `CLAUDE_MAX_IMAGE_TOTAL_BYTES`), all fail-closed on a misconfigured value. TUI mode returns `400 images_unsupported_in_tui_mode` (it can't carry image blocks) and an image present only in a `system` message returns `400` rather than being silently dropped. README § "Images / Multimodal".
+- **Structured outputs — OpenAI `response_format` (#153, contributed by @vvlasy-openclaw).** `/v1/chat/completions` honors `response_format: { type: "json_schema" | "json_object" }` so OpenAI-SDK clients (Home Assistant AI Tasks, Honcho, scripts) get machine-parseable JSON in `content`. Validates against the schema (incl. `$ref`/`$defs` + `allOf`/`anyOf`/`oneOf` — the shapes the OpenAI SDK emits), retries with a stronger instruction up to `OCP_STRUCTURED_MAX_ATTEMPTS` (default 3, fail-closed), and on exhaustion returns OpenAI's own `refusal` field (200/`content:null`) rather than an invented error. Cyclic-`$ref` schemas fail closed (no stack overflow); a pathologically deep model reply returns a refusal, not a 500 (#181). Single-flight dedup + structured-keyed cache bound the cost. Class B.1 (ADR 0006). README § "Structured Outputs".
+- **SPOT-derived prompt-char budget (#179, ADR 0009).** `MAX_PROMPT_CHARS` default now derives from `max(models.json contextWindow) × 3 chars/token` (600,000 today) instead of the hand-set 150,000 (~37.5k tokens) that silently under-delivered the advertised window ~5×. `CLAUDE_MAX_PROMPT_CHARS` and the settings API remain absolute overrides; a garbage value fails closed to the derived default.
+- **`OCP_LOCAL_TOOLS` — positive local-tools system-prompt wrapper (single-user, loopback only; default off) (#182, contributed by @vvlasy-openclaw).** The `-p` path prepends a wrapper telling the model it has no local filesystem/shell access — correct for a shared gateway, but it makes a personal instance's model (e.g. an OpenClaw agent on its own local OCP) refuse to use the server-side `claude` tools it legitimately has. `=1` swaps in a positive wrapper. Changes **only the prompt**, never the tool surface (`--allowedTools`/`--disallowedTools` untouched; multi-tenant still disallows the FS surface); it does **not** enable client-side `tool_calls` (still unsupported by design). Fail-closed boot gate mirroring `OCP_TUI_FULL_TOOLS` (ADR 0007): refuses to start under `CLAUDE_AUTH_MODE=multi`, a non-loopback bind, or `PROXY_ANONYMOUS_KEY`. Inert (and logged as such) in TUI mode. The active wrapper is folded into the config epoch so toggling it invalidates the standard response cache. No new `cli.js` wire behavior (reuses the already-cited `--system-prompt` flag).
+
+### Fixed
+
+- **Agentic turns dropped the model's final answer (#183, contributed by @vvlasy-openclaw).** On a tool-using turn, `/v1/chat/completions` returned only the opening preamble ("I'll find the repo…") and silently discarded the post-tool-use final answer: aggregate-`assistant` extraction was gated on `isFirstDelta` (which flips false after the first text), and OCP runs pure-aggregate mode (no `--include-partial-messages`), so each of an agentic turn's several assistant messages after the first was lost. Now guards on `sawTextDelta` and accumulates every assistant message (streaming and buffered paths assemble byte-identically).
+- **Deep structured reply returned a 500 instead of a refusal (#181 / #184).** `validateJsonSchema` recurses on the model reply's nesting depth; a ~2000-level-deep reply overflowed the stack → caught `RangeError` → generic 500. A crash-safe façade converts that (only) into a validation miss → refusal; any other throw still surfaces.
+
+## v3.23.0 — 2026-07-17
+
+Minor release. Headline: **the default `sonnet` alias now resolves to Claude Sonnet 5** — a behavior change for every request that omits `model` (pin `claude-sonnet-4-6` by full ID to keep the previous default). Also: Windows-safe upgrade snapshots, two upgrade-system reliability fixes from a live fleet update, the `CLAUDE_SYSTEM_PROMPT` env var made functional, cache-key honesty for config changes, a billing-policy status correction (the 2026-06-15 `-p` split is PAUSED by Anthropic), and a major README restructure. No new endpoint; no new `cli.js` wire behavior. Every code PR carried a fresh-context reviewer (Iron Rule 10).
+
+### Changed
+
+- **Default `sonnet` alias → `claude-sonnet-5` (#168, contributed by @vvlasy-openclaw).** The `sonnet` alias (the model used for every `/v1/chat/completions` request that omits `model`, and OpenClaw's OCP primary via `ocp-connect`) now resolves to `claude-sonnet-5` instead of `claude-sonnet-4-6`. `claude-sonnet-4-6` remains available by full ID for pinning. Mirrors the shipped Claude CLI's own `latest_per_family` mapping (`sonnet → claude-sonnet-5`, verified from binary 2.1.211). Split out from the additive model entry (#152) per Iron Rule 11.
+- **`CLAUDE_SYSTEM_PROMPT` is now functional (#175).** The var was read, documented, and echoed on `/health.systemPrompt` but never reached a request (dead since the `APPEND_SYSTEM_PROMPT` retirement). It is now appended (last, trimmed) to the composed system prompt on the default `-p` path via the new pure `lib/prompt.mjs`; TUI-mode panes are unaffected. Unset ⇒ byte-identical composition to before. README § Environment Variables documents it, including the cache caveat below.
+
+### Fixed
+
+- **Windows-safe upgrade snapshot paths (#167, contributed by @nyxst4ck).** Snapshot directory timestamps now use `-` instead of `:` (Windows forbids `:` in names); legacy colon-named snapshots keep parsing, and `listSnapshots` now orders by **parsed timestamp** (with a deterministic name tie-breaker) so mixed legacy/new names sort chronologically — the initial revision's raw-string sort could delete the newest recovery snapshot at the format boundary and was caught in review; regression tests pin the same-hour mixed-format case.
+- **`ocp update` reliability — two live-incident fixes (#174, closes #173).** (1) The doctor now runs `git fetch --tags` (offline-tolerant) before computing `latest_version` — previously it compared against the locally cached `origin/main`, so machines that hadn't pulled since a release reported "Already at latest" forever. (2) Post-flight now asserts `/health.version` equals the upgrade target (new `postFlightOk` predicate) instead of accepting any `auth.ok` — a stale orphan process holding the port used to pass post-flight while still serving the old version; the failure message now reports the last-seen version and points at `ss -ltnp`/`lsof -i`.
+- **Response-cache key now carries a boot-config epoch (#177, closes #176).** The persistent cache keyed on model+key+params+messages but not on server config that shapes answers (`CLAUDE_SYSTEM_PROMPT`, wrapper text, `CLAUDE_ALLOWED_TOOLS`, `CLAUDE_NO_CONTEXT`) — changing any of these and restarting could serve stale-config answers until TTL expiry. A sha256 config-epoch is folded into every key; any config change is an instant whole-cache invalidation. One-time side effect: existing cache entries miss once after this upgrade.
+
+### Docs
+
+- **Billing-policy status corrected (#171).** Anthropic **paused** the announced 2026-06-15 `claude -p` billing split on its effective date (official help-article citation in README § How It Works): the default `-p` path currently bills the subscription, and TUI-mode is reframed as the ready-made **hedge** for if/when a reworked change lands. All in-force assertions of the split are now date-stamped and conditioned.
+- **LAN mode scoped to chat-class workloads (#171).** New "workload fit" paragraph: multi-device OCP is for text-in/text-out workloads; client-machine coding agents are architecturally out of scope (tools execute on the OCP host).
+- **README restructured, 1205 → ~500 lines (#172).** Operations-manual content moved to `docs/lan-mode.md`, `docs/tui-mode.md`, `docs/troubleshooting.md`, `docs/upgrading.md` (verbatim moves + two canonical dedups; zero content loss verified section-by-section). README keeps the quickstart, the release-kit-pinned reference tables, and summary stubs with links. Plus a staleness sweep (#170): 6-model examples, removal of the never-existed `ocp stop` command, `ocp-connect` claims corrected, current version examples.
+
+## v3.22.1 — 2026-07-17
+
+Minor release: TUI-mode latency and streaming features — **all opt-in and off by default**, so the default request path (`-p` / `--output-format stream-json`) is byte-for-byte unchanged — plus hardening from an independent (Codex) re-review of the streaming work, Windows `claude.exe` startup resolution, and the Claude Sonnet 5 model entry. No new `cli.js` wire behavior and no new endpoint; the new surface is entirely OCP-owned TUI-mode configuration (env vars), startup binary discovery, model metadata, and `/health` observation. Every code PR carried a fresh-context reviewer (Iron Rule 10). (Version note: v3.22.0 was prepared but never tagged; its contents ship here as v3.22.1 together with the additions below.)
+
+### Added
+
+- **Claude Sonnet 5 in the model SPOT (#152, contributed by @vvlasy-openclaw)** — `claude-sonnet-5` added to `models.json` (`contextWindow` 200000 / `maxTokens` 16384 / `reasoning` true, consistent with existing entries), exposed via `/v1/models` and the OpenClaw sync. Purely additive: the `sonnet` alias still resolves to `claude-sonnet-4-6` (the repoint is tracked separately in #168). `ocp-connect`'s model classifier now matches on the model *family* prefix (`claude-sonnet`/`claude-opus`/`claude-haiku`) instead of version-pinned prefixes, so current and future versioned IDs register with correct `reasoning`/`maxTokens` metadata. New referential-integrity tests guard that every alias target exists in `models[]`.
+- **Windows `claude.exe` startup resolution (#161, contributed by @nyxst4ck, diagnosis credit #147 @Justinsato)** — on Windows, `resolveClaude()` now discovers a native `claude.exe` (`%USERPROFILE%\.local\bin`, WinGet Links, WindowsApps, then `where.exe`) and rejects npm `.cmd`/`.bat`/`.ps1` shims, which cannot be spawned without a shell — previously startup resolved a shim and failed. A non-`.exe` `CLAUDE_BIN` on Windows is a fatal error with an actionable hint. The macOS/Linux path is byte-for-byte unchanged. Note: this is startup binary resolution only — full Windows support is not yet claimed (snapshot-path portability is tracked in #167).
+
+### Added — TUI mode (all opt-in, default off)
+
+- **Spawn effort control — `OCP_TUI_EFFORT` (default `low`) (#156)** — the interactive `claude` is now spawned with an explicit `--effort` flag. `low` cuts measured TTFT p50 by ~40% and collapses run-to-run variance ~15× versus an inherited `xhigh`; proxied requests rarely benefit from extended thinking. Set `inherit` to omit the flag and restore the pre-flag HOME-dependent behaviour. Banner-verified to stay on the subscription pool (`· Claude Max`); an invalid value warns and falls back to `low`. README § "Environment Variables".
+- **Warm pane pool — `OCP_TUI_POOL_SIZE` (default `0` / off) (#158)** — pre-boots up to 4 single-use `claude` panes so a request skips the cold boot: measured end-to-end p50 `10.17s` → `6.00s` (−41%) on a Mac mini (Sonnet 4.6, `--effort low`). Opt-in because each warm pane is a live idle process held whether or not a request ever arrives. Panes are single-use (one turn, then killed and replaced in the background), port-scoped (`ocp-tui-<port>-p<hex>`), and coexist with the zombie reaper by a synchronous drain→reap→resume sweep. README §§ "Environment Variables" + "How It Works".
+- **Real SSE streaming — `OCP_TUI_STREAM` (default `0` / off) (#159, #160)** — `stream:true` turns emit real `delta.content` chunks as `claude` generates them, sourced from `claude`'s own `MessageDisplay` hook (registered via `--settings` on the ordinary interactive spawn — banner-verified on the subscription pool). Granularity is block-level, and it moves the *first* byte, not the last. The transcript stays authoritative: streamed text is asserted equal to it at end-of-turn, the auth-banner and truncation gates still run before anything is committed, and a turn whose stream cannot be reconciled is **refused** (SSE error frame, not cached) and counted on `/health` (`tui.streamDivergences`; a silent total-hook-failure is counted separately as `tui.streamZeroDeltaTurns`). Tunables: `OCP_TUI_STREAM_HOLDBACK` (default `100`), `OCP_TUI_STREAM_DIR`, `OCP_TUI_STREAM_POLL_MS`. See ADR 0007 (2026-07-13 amendment). README §§ "Environment Variables" + "How It Works".
+
+### Fixed
+
+- **Streaming auth-banner guard: a null `message_id` on the first hook fire (#160)** — a first `MessageDisplay` fire with a null `message_id` could disarm the auth-banner guard; re-landed after a #159 squash dropped it (`lib/tui/stream.mjs`).
+- **Test suite wrote live, unrevoked API keys into the operator's real key store (#163)** — `npm test` had been opening `~/.ocp/ocp.db` (the running server's DB) and writing two junk `api_keys` rows per run (737 accumulated on the maintainer's host), because the isolation the comments claimed was never wired (ESM import hoisting). `keys.mjs` now honors `OCP_DIR_OVERRIDE` under `NODE_ENV=test` and the suite points at a scratch dir; a child-process probe verifies a production process (no `NODE_ENV`) cannot be redirected.
+- **Streaming holdback floor + billing-pool observation on failed turns (#164)** — (A1) `OCP_TUI_STREAM_HOLDBACK` now clamps up to the safe floor (`100`) with a boot warning, closing a latent auth-banner leak when an operator set a sub-floor value. (A3) the `cc_entrypoint` (billing-pool) observation is now recorded before the honesty gates that throw, so `/health` no longer goes blind to exactly the failed turns most likely to signal a silent degrade to the metered Agent SDK pool.
+- **Test-only key-store redirection vars can no longer reach a server OCP launches (#165)** — (A4) `NODE_ENV`/`OCP_DIR_OVERRIDE` are stripped from every service unit `setup.mjs` writes (`plist-merge`'s `NEVER_PRESERVE`) and from the `ocp restart` manual nohup fallback (`env -u`); #163's overstated "a prod server can NEVER be redirected" comments were softened to name the one residual hand-launch path and the loud `getDb()` "NOT the default" backstop.
+
+### Docs
+
+- **README billing honesty (#162, closes #136)** — removed a feature bullet that promised what the § "honest limits" section forbids.
+- **TUI latency plans + streaming-achievability spike (#155, #157)** — measured latency decomposition, backlog, and the `MessageDisplay`-hook streaming prereq spike under `docs/plans/2026-07-13-tui-latency/`.
+
+## v3.21.1 — 2026-07-07
+
+Patch release: three bug fixes from an independent concurrency/session-lifecycle audit, each its own PR with a fresh-context reviewer (Iron Rule 10). No new `cli.js` wire behavior, no new endpoint, header, or env var; the `/health` field set is unchanged (only value truthfulness improved).
+
+### Fixed
+
+- **TUI session-scope / boot-reap (#148)** — `lib/tui/session.mjs`'s tmux session prefix is now scoped per-instance by listen port (`ocp-tui-<port>-`) instead of a bare host-wide `ocp-tui-` constant, so a second OCP instance on the same host (e.g. a temporary verification instance) can no longer have its live TUI sessions reaped or `kill-server`'d by another instance's boot/periodic sweep. The one-time boot reap also claims exact-shape legacy `ocp-tui-<8hex>` sessions (pre-fix naming) once, to clean up zombies left behind across an in-place upgrade.
+- **`-p` spawn-token mutex + keychain caching (#150)** — the real-HOME token fallback used when the keychain token is within its 5-minute expiry window is now serialized behind a mutex, so concurrent `-p` spawns no longer race the same single-use refresh token against each other (the credential-fork hazard). Added a 30s TTL cache + last-good-label memoization for the keychain read, cutting per-spawn event-loop blocking. The isolation decision (`/health` isolated/real-home reporting) is now re-evaluated per spawn instead of memoized forever, so `/health` no longer misreports a stale decision. New module `lib/spawn-auth.mjs` extracts the pure, unit-testable primitives (mutex, TTL cache, expiry gate, label ordering).
+- **Concurrency queue / disconnect handling (#149)** — the shared semaphore now honors a runtime-lowered `maxConcurrent` immediately (previously a decrease was silently ignored until in-flight tasks finished on their own) and wakes queued waiters right away when the limit is raised. Queued `-p`/TUI requests are now linked to the client's HTTP connection via `AbortSignal`; a client that disconnects while queued is spliced out of the queue instead of still spawning `claude` once a slot frees. A singleflight follower whose leader disconnected now retries instead of inheriting a spurious 500, and a queued-then-disconnected request is no longer recorded as a usage failure or logged as an error (quiet disconnect handling).
+
 ## v3.21.0 — 2026-06-25
 
 Cleanup + docs release: TUI dead-code removal, docs honesty, and release prep. No new `cli.js` wire behavior; the default path (`CLAUDE_TUI_MODE` unset) is byte-for-byte unchanged.

docs/adr/0007-tui-interactive-mode.md

@@ -56,7 +56,7 @@ Add `CLAUDE_TUI_MODE=true` as an opt-in flag in `server.mjs`.
 3. The serialized prompt (from `messagesToPrompt`) is pasted via `tmux send-keys … "$(cat file)"` + a separate `Enter` key event.
 4. The answer is read from claude's native JSONL transcript at `<HOME>/.claude/projects/<encoded-cwd>/<session-id>.jsonl`, polling until a `turn_duration` system event or the wall-clock cap (`CLAUDE_TUI_WALLCLOCK_MS`, default 120 s).
 5. The string answer is returned to OCP's existing downstream (singleflight → cache write-back → `completionResponse` / `streamStringAsSSE`) — **same contract as `callClaude`**.
-6. Streaming requests are buffered then replayed as chunked SSE (no real token streaming — deliberate; "don't build fragile features").
+6. Streaming requests are buffered then replayed as chunked SSE (no real token streaming — deliberate; "don't build fragile features"). **Superseded for `stream:true` when `OCP_TUI_STREAM=1` — see the 2026-07-13 amendment below. The buffered path remains the default and is unchanged.**
 
 ### Billing-classifier labeling (`OCP_TUI_ENTRYPOINT`, PR-4)
 
@@ -333,6 +333,61 @@ The original "Home strategy" section and PR-C's `prepareTuiHome` comment warned
 
 ---
 
+## Amendment (2026-07-13) — real SSE streaming via the `MessageDisplay` hook (`OCP_TUI_STREAM`)
+
+**Supersedes**: Request-flow step 6 above ("no real token streaming — deliberate"), for `stream:true`
+requests when `OCP_TUI_STREAM=1`. The buffered path stays the default and is byte-for-byte unchanged.
+
+**Context.** Step 6 was written when the interactive CLI appeared to expose no byte-faithful
+incremental source. A prereq spike (`docs/plans/2026-07-13-tui-latency/streaming-spike.md`) confirmed
+three obvious sources are dead ends — the transcript JSONL grows one *whole event* at a time (the
+answer lands as a single line ~0.3 s before the terminal marker); `tmux capture-pane` yields a
+*rendered* view whose markdown source is unrecoverable (an H2 and a bold span produce identical ANSI);
+`--debug-file` logs stream *timing*, never stream *content*. Every interface that does emit
+`text_delta` (`--output-format stream-json`) requires `-p`, which moves the request to the **metered**
+`sdk-cli` pool — precisely what TUI-mode exists to avoid.
+
+**Decision.** Consume `claude`'s own **`MessageDisplay`** hook, registered via `--settings` on the
+ordinary interactive spawn (no `-p`, no `--bare`). Each fire delivers the **raw markdown source** of an
+incremental `delta` on the hook's stdin. Verified live (claude 2.1.207, sonnet-4-6): banner stays
+`· Claude Max` and the transcript `entrypoint` stays `cli` (subscription pool); `concat(deltas) === T`
+byte-exactly; `T.startsWith(concat(deltas[0..n]))` at every *n*. This is **forwarding, not inventing**
+— ALIGNMENT.md **Class B**. No `cli.js` citation applies: the TUI spawn is OCP-owned surface (this
+ADR), the hook payload is claude's own published contract, and the SSE wire shapes are the OpenAI
+chat/completions streaming spec adopted by **ADR 0006** (the emitters are literally the `-p` path's).
+
+**The transcript remains authoritative.** It is still the terminal-turn signal, still the source of the
+returned/cached text `T`, and still the input to the honesty gates (auth-banner detection C-1,
+`truncated` C-2). The delta stream is a low-latency **mirror**, never a replacement. At end of turn OCP
+asserts the streamed bytes against `T`: equal → serve; a strict *prefix* of `T` → top up from the
+transcript (client still receives exactly `T`); **not** a prefix → **refuse the turn** (SSE error frame,
+no cache, `tui.streamDivergences++`). Serving text the transcript disagrees with is the failure class
+ALIGNMENT.md exists to prevent, so streaming fails loud rather than degrading quietly.
+
+**Consequences / constraints recorded for future authors:**
+
+- **Opt-in, default OFF.** The buffered path is stable production; streaming does not change it.
+- **Per-`session_id` sink is mandatory, not an optimization.** `OCP_TUI_MAX_CONCURRENT` defaults to
+  **2** — two `claude` panes already run concurrently. A single shared sink would interleave one
+  client's deltas into another's stream. The hook writes to `<dir>/<session_id>.jsonl`, the path
+  delivered through the *pane's own env* (`OCP_TUI_STREAM_FILE`); OCP reads only its own turn's file.
+  Verified with two concurrent streamed turns (ALPHA/BRAVO): zero cross-contamination.
+- **Warm-pool compatible (a separate in-flight PR depends on this).** The hook script and the settings
+  file are **static** — nothing request-specific is baked in at spawn time. The sink path derives from
+  the session-id, which for a pre-booted pane is fixed at boot.
+- **The hook is synchronous** (`forceSyncExecution: true` — `claude` *blocks* on it). The hook script
+  must write and exit; it does one `cat` append and nothing else. Measured: p50 **7.2 ms** per fire,
+  ~50 ms across a whole turn — noise against a 6–10 s turn. Do not add work to it.
+- **Thinking blocks do not fire the hook** — verified on a substantive Opus/`xhigh` reasoning turn (see
+  the PR evidence), not merely inferred from the `final:true` call site. This must be **re-verified** if
+  the hook is ever pointed at a new model/effort tier: a thinking delta reaching a client would be
+  unretractable, and the `concat === T` assertion can only *detect* that after the fact, never prevent
+  it. The first-bytes **holdback** (`OCP_TUI_STREAM_HOLDBACK`, default 100 chars) is the same
+  prevention-not-detection reasoning applied to the auth-banner gate.
+- **Block-level granularity**, scaling with answer length — not token-level. Do not promise otherwise.
+- **It moves the first byte, not the last.** Only a progressively-rendering consumer benefits; it does
+  not move TUI-mode's ~6 s TTFT floor.
+
 ## Provenance
 
 TUI-mode originated in a prototype contributed via PR #101 (see the PR for author attribution). The productionization design is in `docs/superpowers/specs/2026-05-30-tui-mode-production-design.md`. Spikes S1–S6 / T1–T6 were validated live on the test host against `claude v2.1.158`.

docs/adr/0008-tui-warm-pane-pool.md

@@ -0,0 +1,208 @@
+# ADR 0008 — TUI Warm Pane Pool
+
+**Date:** 2026-07-13
+**Status:** Proposed
+**Extends:** [ADR 0007](0007-tui-interactive-mode.md) (TUI interactive mode). This ADR does not
+change ADR 0007's billing-pool argument, security posture, or kill-switch — it adds a latency
+optimization *inside* the TUI spawn machinery ADR 0007 owns.
+
+---
+
+## Context
+
+TUI mode (ADR 0007) serves every request by cold-booting a fresh `tmux` session running an
+interactive `claude`, submitting one prompt, reading the native transcript, and killing the
+session. That cold boot is paid on **every** request.
+
+[`docs/plans/2026-07-13-tui-latency/`](../plans/2026-07-13-tui-latency/README.md) measured the
+TUI path and listed a warm pane pool as backlog item #3, costed at "**~1.0 s**" (the observed
+boot-to-input-bar time). Instrumenting the real request path showed that estimate is **~4×
+too low**. Phase decomposition of the cold path (n=6 medians, Sonnet 4.6, `--effort low`,
+through a real OCP instance):
+
+| Phase | Median |
+|---|---|
+| prep (trust cwd, write prompt file) | 2 ms |
+| `tmux new-session` | 27 ms |
+| **boot → input bar ready** | **1232 ms** |
+| paste (`load-buffer` + `paste-buffer`) | 8 ms |
+| paste-verify poll | 426 ms |
+| **submit → transcript terminal** | **8458 ms** |
+| teardown | 8 ms |
+| **total** | **10162 ms** |
+| *claude's own reported `turn_duration`* | *5539 ms* |
+| **OCP-side overhead** | **4490 ms** |
+
+The `submit → terminal` phase exceeds claude's own `turn_duration` by **~2.9 s**. That gap is
+**post-input-bar initialization inside `claude`** — work that a pane which has merely *sat idle
+for a few seconds* has already completed. A direct spike confirmed it: an identical pane, idle
+12 s before receiving the same prompt, completed its turn in a median 5537 ms versus 7980 ms
+cold.
+
+So a warm pane recovers **~1.26 s of boot *and* ~2.9 s of in-`claude` cold start** — not the
+~1.0 s the plan predicted.
+
+The reason this was worth a pool rather than a "keep one session and reuse it" cache is a
+hazard already flagged in the code. `lib/tui/transcript.mjs` returns the **last text-bearing
+assistant entry in the whole transcript file**, which is correct *only* under OCP's
+one-session-per-request model, and it says so:
+
+> *"If a future warm-pool ever reuses a session WITHOUT a fresh session-id / clear, earlier-turn
+> text could leak — that author must add user-line scoping here."*
+
+Reusing a pane for a second turn puts two exchanges in one transcript and would leak the earlier
+turn's text into the later turn's answer — a **cross-request data leak**, not merely a bug.
+
+---
+
+## Decision
+
+Add an **opt-in pool of pre-booted, single-use `claude` panes**, `OCP_TUI_POOL_SIZE` (default
+`0` = off, max `4`). Implementation: `lib/tui/pool.mjs`.
+
+### 1. Panes are SINGLE-USE. This is the load-bearing rule.
+
+A pooled pane serves **exactly one turn**, then is killed and replaced in the background. Each
+pane is booted with its **own fresh `--session-id`**, fixed at spawn, and the turn locates its
+transcript by that id.
+
+This preserves one-session-per-request exactly, so the `transcript.mjs` hazard above **does not
+arise** and no user-line scoping was needed. The warning in `transcript.mjs` is deliberately
+left standing, now annotated: it still binds anyone who later wants a pane to serve a second
+turn, or to reset a session with `/clear` and reuse it. **Neither is permitted without first
+adding user-line scoping to the transcript reader.**
+
+Rejected alternative — *reuse a pane for N turns, `/clear` between* — is strictly cheaper
+(no re-boot per request) and was rejected on exactly this basis. The latency win is not worth a
+cross-request text-leak surface guarded only by a `/clear` that we cannot verify landed.
+
+### 2. The pool is keyed by model, and a MISS is always safe.
+
+`--model` is fixed at spawn, so a pane can only serve the model it booted with. A pool miss
+falls back to the existing cold-boot path with **zero behavioural difference**. There is no
+boot-time pre-warm and no configured model: OCP cannot know which model the next caller wants,
+so the pool warms the **most recently requested** model. Consequence, stated plainly: **the
+first request after start, and the first after any model switch, is always a cold miss.**
+
+### 3. The pool and the session reaper coexist by an explicit invariant.
+
+This is the subtle part. `reapStaleTuiSessions()` kills every session matching this instance's
+`ocp-tui-<port>-` prefix, and issues `tmux kill-server` when no foreign session remains (the
+only mechanism that can reap `<defunct>` `claude` zombies — the pane's `claude` is a child of
+the tmux *server*, not of node). A warm pooled pane **is** one of our own sessions, alive and
+idle **by design** — and the periodic sweep runs precisely **when the instance is idle**, i.e.
+exactly when the pool is full.
+
+The invariant, stated in a comment above `reapStaleTuiSessions` and pinned by tests:
+
+1. **A live pooled pane is never reaped — including one that is still BOOTING.** The reaper
+   takes a `spare` set of **exact session names** supplied by the pool's live registry.
+2. **An orphaned pooled pane IS still reaped.** Membership is by **exact name from a live
+   in-memory registry, never by name shape**. A pane the pool no longer owns — handed out,
+   dropped, cancelled, or left behind by a previous process generation (whose registry died with
+   it) — is absent from `spare` and is killed like any other stale session. **Fail-safe:
+   omitting `spare` reaps *more*, never less.** Pool panes are named `ocp-tui-<port>-p<hex>`
+   purely for operator legibility; that shape is *not* the exemption mechanism.
+3. **`kill-server` is suppressed while any pane is spared** (it would kill a live child of the
+   tmux server). Therefore **the pool is DRAINED immediately before every sweep**, so `spare` is
+   empty on the normal tick and `kill-server` still fires. Without the drain, a permanently-full
+   pool would **permanently disable zombie reaping** — the pool would silently break the thing
+   the sweep exists to do. The drain costs one pane re-boot per tick (15 min).
+
+The `spare` mechanism is belt-and-braces given the drain: it makes it impossible for a reap call
+site that *forgets* to drain to kill a live pane.
+
+### 4. The pool tracks its in-flight boot BY NAME, not as a count.
+
+`bootTuiPane` creates the tmux session **synchronously** and only *then* waits (up to
+`POOL_BOOT_MS`, 20 s) for the input bar. So **a pooled tmux session can be live for ~20 s before
+its boot resolves.** A pool that tracked in-flight boots as a *count* could not name that
+session, and this produced two real bugs (both caught in review, both now regression-tested):
+
+- the periodic sweep **killed the booting pane** (it could not be spared), then left the pool
+  empty with nothing scheduled, and logged the exact `tui_pool_boot_failed` warning operators are
+  told to alert on — for a completely healthy drain;
+- graceful shutdown **orphaned a live, authenticated, idle `claude`**: `gracefulShutdown` calls
+  `process.exit(0)` in the same tick as the drain (TUI panes are tmux children, so node's
+  `activeProcesses` set is empty and the "wait for children" path exits immediately), so any
+  cleanup deferred to a `.then()` never ran.
+
+The pool therefore **mints each pane's identity up front** (`{sessionId, name}`, name derived
+from the session-id so `tmux ls` correlates to the transcript file) and holds it in
+`_bootingPane`. `liveNames()` includes it; `drain()` kills it **synchronously**. A generation
+counter distinguishes *"cancelled by us"* from *"genuinely failed"*, so a drain never inflates
+`bootFailures` and `resume()` reliably starts a fresh boot.
+
+### 5. Refills take no concurrency slot, and are serialized.
+
+A refill boot deliberately does **not** take a `TuiSemaphore` slot: those slots bound concurrent
+*turns* and belong to real requests, and charging a background pre-boot against them would let
+the pool starve the traffic it exists to speed up. It cannot leak a slot either, since it never
+holds one. Boots are **serialized** (one at a time): two cold boots racing an in-flight turn were
+observed to overrun even the generous pool readiness cap. A genuinely failed boot does **not**
+re-kick the chain (backoff — a broken `claude` must not respawn forever).
+
+Background boots get a more generous readiness cap (`POOL_BOOT_MS` = 5 × `BOOT_MS`): `BOOT_MS` is
+tight because a *client* is blocked on it, which is not true of a pre-boot. Slow ≠ broken.
+
+---
+
+## Consequences
+
+### Cost — standing processes, paid whether or not a request arrives
+
+**A warm pane is a live idle `claude` process.** Peak process count is
+`OCP_TUI_POOL_SIZE` + `OCP_TUI_MAX_CONCURRENT` + 1 (booting replacement). This is the whole
+reason the pool is **default-off**: an operator must opt into holding processes for traffic that
+may never come. Size is clamped to `POOL_MAX_SIZE` = 4; an unparseable value **disables** the
+pool rather than guessing.
+
+Panes carry a 10-minute TTL and are health-checked at hand-out; a dead or degraded pane becomes
+a **miss** (cold path), never a hung turn.
+
+### Benefit
+
+Measured end-to-end through a real OCP instance (Sonnet 4.6, `--effort low`):
+**p50 10.17 s (n=6, pool off) → 6.00 s (n=12 warm hits) — −4.2 s / −41%.**
+
+### The floor is unchanged
+
+The pool does not touch the **~6 s TTFT floor** documented in the latency plan (claude always
+prefills the full Claude Code system prompt). TUI mode remains unsuitable for interactive /
+real-time consumers; it is for batch and background work. This ADR does not change that
+conclusion.
+
+### Observability
+
+`/health`'s `tui` block gains a `pool` sub-object (`null` when off): `size`, `warm`, `booting`,
+`model`, `hits`, `misses`, `boots`, `bootFailures`, `cancelled`, `dropped`. A climbing
+`bootFailures` means panes are not reaching their input bar — the pool then degrades safely to
+the cold path, but latency reverts to the un-pooled numbers. A steadily climbing `dropped` is
+**normal** (the 15-min sweep drains and re-boots the pool on every tick, by design — see
+Decision 3).
+
+### ALIGNMENT authorization
+
+- **Class B / OCP-owned.** The warm pool is process management around the `claude` CLI — the
+  same category as the existing tmux session lifecycle and the defunct-session reaper it extends.
+  **`cli.js` does not perform this operation, and no `cli.js` citation applies**; the authority
+  is ADR 0007 (which owns the TUI spawn machinery) plus this ADR. This is `ALIGNMENT.md` Rule 2's
+  Class B citation requirement, discharged explicitly rather than by silence.
+- **The `/health` extension** adds sub-fields to the `tui` block. That block is **owned by ADR
+  0007** and post-dates ADR 0006's v3.16.4 grandfather snapshot, so it is not part of the frozen
+  B.2 inventory. The change is additive — every pre-existing `/health` field keeps a
+  byte-identical value, and `pool` is `null` unless the operator opts in — which is the
+  behaviour-preserving bar ADR 0006 sets. This ADR records that authorization.
+- **No spawn argument changed.** `buildTuiCmd` is byte-identical; the pool calls it with the same
+  arguments. Banner-verified on live pooled panes: `· Claude Max`, never `API Usage Billing`
+  (the `--bare` trap documented in the latency plan).
+
+### What a future contributor must not undo
+
+- **Do not let a pane serve a second turn** (or `/clear`-and-reuse one) without first adding
+  user-line scoping to `lib/tui/transcript.mjs`. That is a cross-request text leak, not a perf
+  tweak. See Decision 1.
+- **Do not remove the drain-before-sweep.** It is what keeps `kill-server` zombie reaping alive.
+  See Decision 3.
+- **Do not go back to counting in-flight boots.** The pool must be able to *name* a session that
+  exists but has not finished booting. See Decision 4.

docs/adr/0009-spot-derived-prompt-budget.md

@@ -0,0 +1,54 @@
+# ADR 0009 — Prompt-char budget derives from the models.json SPOT
+
+Date: 2026-07-18
+Status: Accepted (maintainer directive, 2026-07-18: "37.5k 截断未免太短了吧 … 这个在现在还适用吗")
+
+## Context
+
+`MAX_PROMPT_CHARS` (the tail-first truncation guard in `messagesToPrompt`) defaulted to a
+hand-set constant of 150,000 chars ≈ 37.5k English tokens — set in the 200k-window era as a
+runaway-context guard. Meanwhile `models.json` advertises `contextWindow: 200000` for every
+model (and the underlying CLI registry carries 1M native windows for Opus 4.8 / Sonnet 5), and
+`scripts/sync-openclaw.mjs` feeds that 200k into OpenClaw's compaction budget. The result was
+a standing dishonesty identified in the PR #152 review: **no advertised contextWindow value was
+true**, because the proxy silently guillotined every request at ~37.5k tokens — roughly 5×
+below the advertised window — logging only a server-side warning the client never sees.
+
+Raising the constant to another hand-set number would rot the same way. Following the model's
+native 1M directly is also wrong: chars ≠ tokens (CJK runs ~1–1.5 chars/token vs ~4 for
+English, so a 1M-token char cap would let CJK text sail past the model's real window into an
+upstream rejection), single near-window requests can consume a large fraction of a 5-hour
+subscription quota window, and the TUI paste path is untested at megabyte scale.
+
+## Decision
+
+The default budget **derives from the SPOT** instead of being a constant:
+
+```
+MAX_PROMPT_CHARS (default) = max(models.json models[].contextWindow) × 3 chars/token
+                            = 200000 × 3 = 600,000 chars today
+```
+
+Implemented as the pure `derivePromptCharBudget(models, {charsPerToken = 3, floor = 150000})`
+in `lib/prompt.mjs` (unit-tested; floor guards degenerate SPOT states). The multiplier ×3 is
+deliberately conservative: full window for English, and CJK text reaches the model's real
+window at roughly the same point the cap fires — so OCP truncates gracefully (tail-first)
+instead of the upstream rejecting outright.
+
+`CLAUDE_MAX_PROMPT_CHARS` (env) and the runtime settings API remain **absolute overrides**;
+the derivation applies only when neither is set.
+
+## Consequences
+
+- The advertised `contextWindow: 200000` becomes honest: the proxy now actually accepts
+  prompts of that order (English ≈150–200k tokens) before truncating.
+- If `models.json` ever advertises a larger window (e.g. 1M for the 1M-native models), the
+  budget scales automatically — no code change. Whether to advertise 1M is a **separate,
+  deliberate decision** (quota burn per request, OpenClaw compaction memory, TUI paste
+  limits) and is explicitly NOT made by this ADR; the current recommendation is to keep
+  200000 advertised until a real >200k use case appears.
+- One-time behavior change: requests between 150k and 600k chars that were previously
+  truncated now pass through whole — longer TTFT and higher quota consumption for those
+  requests, by design.
+- The truncation mechanism, logging, and the multimodal-path budget threading (PR #154's F2,
+  pending) are unchanged — only the default value's provenance changed.

docs/adr/README.md

@@ -23,6 +23,9 @@ New ADRs increment from the highest existing number. Filenames are
 | [0004](0004-openclaw-auto-sync.md) | OpenClaw Auto-Sync | Why `scripts/sync-openclaw.mjs` runs on `ocp update`, what its scope boundary is (writes only `models.providers["claude-local"].models` and `agents.defaults.models["claude-local/*"]`), and the idempotency contract. |
 | [0005](0005-no-multi-provider.md) | No Multi-Provider | Why OCP stays single-provider (Anthropic-via-cli.js) and does not extend to OpenAI / Gemini / OpenRouter. Cost estimate: ~7 weeks for a v1 that buys neither moat nor commercial readiness. Separate commercial work starts in a separate repo. |
 | [0006](0006-openai-shim-scope.md) | OpenAI Shim Scope | The Class A / Class B taxonomy. Class A endpoints (`cli.js`-mirror) keep Rules 1–5 verbatim; Class B endpoints (OCP-owned compatibility surface — `/v1/chat/completions`, `/v1/models`, admin endpoints) are anchored to OpenAI's spec (B.1) or to an authorizing ADR (B.2). Triggered by PR #99 (external `response_format` honoring). Grandfathers the existing B.2 inventory at v3.16.4. |
+| [0007](0007-tui-interactive-mode.md) | TUI Interactive Mode | Why TUI-mode spawns an interactive `claude` in a tmux pane (no `-p`) to reach the **subscription** billing pool (`cc_entrypoint=cli`) rather than the metered Agent SDK pool. Owns the TUI spawn machinery: entrypoint labeling, credential-isolated home, MCP hard-disable, session namespace + defunct-session reaping, the independent concurrency bound, and the `/health` `tui` block. **Single-user only** — hard FATAL on multi-user configs. |
+| [0008](0008-tui-warm-pane-pool.md) | TUI Warm Pane Pool | Why `OCP_TUI_POOL_SIZE` pre-boots **single-use** `claude` panes (one turn each, own `--session-id`) — and why reuse is forbidden (`transcript.mjs` returns the last assistant entry in the file, so a reused session leaks the earlier turn's text). Measured −41% end-to-end. Defines the pool↔reaper invariant (exemption by exact name from a live registry; drain before every sweep so `kill-server` zombie reaping survives) and the standing idle-process cost. Extends ADR 0007. |
+| [0009](0009-spot-derived-prompt-budget.md) | SPOT-Derived Prompt Budget | Why `MAX_PROMPT_CHARS`'s default is `max(models.json contextWindow) × 3 chars/token` (600k chars today) instead of a hand-set constant — the old 150k silently under-delivered the advertised window ~5×. ×3 is the CJK-safe multiplier; env/settings stay absolute overrides; whether to advertise 1M windows is explicitly a separate decision. |
 
 ## When to write a new ADR
 

docs/lan-mode.md

@@ -0,0 +1,397 @@
+Part of [OCP](../README.md) — LAN & multi-user: server setup, client connect, API-key management, per-key quotas, anonymous access, and the deployment/security model (including the honest limits of sharing).
+
+# LAN & multi-user
+
+OCP has two roles: **Server** (runs the proxy, needs Claude CLI) and **Client** (connects to a server, zero dependencies).
+
+```
+┌─ Server (always-on device) ─────────────────────────────┐
+│  Mac mini / NAS / Raspberry Pi / Desktop                │
+│  Claude CLI + OCP server → bound to 0.0.0.0:3456       │
+└───────────────────────┬─────────────────────────────────┘
+                        │ LAN
+    ┌───────────────────┼───────────────────┐
+    ▼                   ▼                   ▼
+ Laptop             Phone/Tablet        Pi / Server
+ (client)           (browser)           (client)
+```
+
+## Server Setup
+
+> **Recommended:** Install OCP on a device that stays powered on — Mac mini, NAS, Raspberry Pi, or a desktop that doesn't sleep. This ensures all clients always have access.
+
+**Prerequisites:**
+- macOS or Linux (Windows is not supported — `setup.mjs` installs launchd / systemd auto-start)
+- Node.js 22.5+ (Node 23+ recommended — `node:sqlite` is fully stable without flags from 23.0; on 22.5–22.x it works behind `--experimental-sqlite`)
+- `git`
+- [Claude CLI](https://docs.anthropic.com/en/docs/claude-cli) — install and authenticate:
+  ```bash
+  npm install -g @anthropic-ai/claude-code
+  claude auth login   # prints a URL + code — open URL on any browser, sign in, paste code back
+  ```
+  Headless servers (Pi / NAS / VPS without a desktop browser): see [Headless install notes](#headless-install-notes) below.
+
+```bash
+# 1. Clone and run setup
+git clone https://github.com/dtzp555-max/ocp.git
+cd ocp
+node setup.mjs
+```
+
+The setup script will:
+1. Verify Claude CLI is installed and authenticated
+2. Start the proxy on port 3456
+3. Install auto-start (launchd on macOS, systemd on Linux)
+
+After install the `ocp` CLI lives at `~/ocp/ocp`. To put it on your PATH, either symlink it manually (`ln -sf ~/ocp/ocp ~/.local/bin/ocp` if `~/.local/bin` is on your PATH, or `sudo ln -sf ~/ocp/ocp /usr/local/bin/ocp` for a system-wide symlink) or add an alias (`alias ocp=~/ocp/ocp`). Otherwise invoke it as `~/ocp/ocp <subcommand>`. The rest of this document assumes `ocp` is on your PATH.
+
+> **Cloud/Linux servers:** If `ocp: command not found` after a cloud install, the binary isn't in PATH. Full path in that layout: `~/.openclaw/projects/ocp/ocp`
+
+**Single-machine use** — just set your IDE to use the proxy:
+```bash
+export OPENAI_BASE_URL=http://127.0.0.1:3456/v1
+```
+
+**LAN mode** — reach OCP from your own devices on the network (Claude Pro/Max are per-user accounts — see [Sharing with family / a team — honest limits](#deployment-model--security-read-this) before extending access to other people):
+```bash
+# Enable LAN access with per-user auth (recommended)
+node setup.mjs --bind 0.0.0.0 --auth-mode multi
+```
+
+Then create API keys for each person/device:
+```bash
+# Generate a strong admin key (one-time — save it for later key management):
+export OCP_ADMIN_KEY=$(openssl rand -base64 32)
+# Add the same export line to ~/.zshrc or ~/.bashrc so it persists.
+
+ocp keys add wife-laptop
+#  ✓ Key created for "wife-laptop"
+#    API Key: ocp_example12345abcde...
+#    Copy this key now — you won't see it again.
+
+ocp keys add son-ipad
+ocp keys add pi-server
+```
+
+Run `ocp lan` to see your IP and ready-to-share instructions.
+
+**Verify:**
+```bash
+curl http://127.0.0.1:3456/v1/models
+# Returns: claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-sonnet-5, claude-sonnet-4-6, claude-haiku-4-5-20251001
+```
+
+### Headless install notes
+
+OCP is designed for always-on devices that often don't have a desktop browser — Mac mini, NAS, Raspberry Pi, cloud VPS. The Claude CLI auth flow still works headless:
+
+**Option 1 — interactive OAuth over SSH (one-shot).** `claude auth login` prints a URL + 8-digit code. Open the URL on **any** device with a browser (your laptop, phone), sign in to your Anthropic account, and paste the code back into the SSH session. No browser needed on the server itself.
+
+**Option 2 — long-lived token (auth once, no re-prompts).**
+
+```bash
+claude setup-token   # subscription-backed long-lived token
+```
+
+Same Claude subscription as Option 1; the token is stored in Claude CLI's normal config location. Useful when you'd rather not redo the OAuth flow when sessions expire.
+
+If `claude auth login` errors out with something like `cannot open browser`, you've hit the same case — fall back to either option above.
+
+## AI-assisted install prompts
+
+If you've got Claude Code, Cursor, or any other AI coding assistant on this machine, you can copy-paste one of these prompts and let the AI walk through the install for you. Each prompt pins the AI to the right README section, names the verification step, and forbids silent retries — so you stay in the loop.
+
+**Single-machine use** — install OCP for IDEs on this same machine only:
+
+```text
+I want to install OCP on this machine to use my Claude Pro/Max subscription
+as an OpenAI-compatible API for local IDEs.
+
+Please follow https://github.com/dtzp555-max/ocp/blob/main/README.md
+§Quickstart (single-machine install):
+
+1. Verify prerequisites: macOS or Linux, Node.js 22.5+, git, Claude CLI
+   installed and logged in (`claude auth status`). Install missing pieces
+   using my system's package manager.
+2. git clone the repo, cd in, and run `node setup.mjs`.
+3. Verify with `curl http://127.0.0.1:3456/v1/models` (should list 7 models).
+4. Add `export OPENAI_BASE_URL=http://127.0.0.1:3456/v1` to my shell rc.
+5. Tell me to reload my shell and try a tool like Cline / Continue / Cursor.
+
+Before each step, tell me what you'll run and wait for confirmation.
+On any error, diagnose first — don't auto-retry.
+```
+
+**LAN mode (server)** — install OCP as a server so your own devices on the LAN can reach it (Claude Pro/Max are per-user accounts — review Anthropic's Usage Policy before extending access to other people):
+
+```text
+I want to install OCP on this device as a LAN server so my own devices on the
+network can reach my Claude Pro/Max subscription through a local
+OpenAI-compatible endpoint.
+
+Please follow https://github.com/dtzp555-max/ocp/blob/main/docs/lan-mode.md
+"Server Setup" → "LAN mode" path:
+
+1. Verify prerequisites: macOS or Linux (Windows not supported), Node.js
+   22.5+, git, Claude CLI installed and authenticated.
+2. Generate a strong admin key with `openssl rand -base64 32`. Save it —
+   I'll need it to manage per-user keys later.
+3. git clone https://github.com/dtzp555-max/ocp.git && cd ocp
+4. Run `node setup.mjs --bind 0.0.0.0 --auth-mode multi`.
+5. Add OCP_ADMIN_KEY to my shell rc (~/.zshrc or ~/.bashrc).
+6. Run `ocp lan` to show me the LAN IP and connect command.
+7. Optionally create example keys: `ocp keys add laptop`, `ocp keys add tablet`.
+8. Verify: `curl http://127.0.0.1:3456/v1/models` returns 7 models.
+
+Tell me each step before running it. On error, diagnose before retrying.
+```
+
+**Client connect** — configure this device to use an existing OCP server on your LAN:
+
+```text
+There's an OCP server at <SERVER_IP> on my LAN. Configure this machine to
+use it for any local IDEs (Cursor, Cline, Continue.dev, OpenCode, OpenClaw).
+
+Server IP: <SERVER_IP>
+API key (leave blank if the server has anonymous mode enabled): <OPTIONAL_KEY>
+
+Please follow https://github.com/dtzp555-max/ocp/blob/main/docs/lan-mode.md
+"Client Setup" path:
+
+1. Download ocp-connect:
+     curl -fsSL https://raw.githubusercontent.com/dtzp555-max/ocp/main/ocp-connect -o ocp-connect
+     chmod +x ocp-connect
+2. Run `./ocp-connect <SERVER_IP>` (add `--key <KEY>` if you have one).
+3. Follow any IDE-specific manual hints it prints.
+4. Verify: `curl http://<SERVER_IP>:3456/v1/models` returns 7 models.
+5. Tell me to reload my shell + restart any IDE that was already running.
+
+Don't auto-retry on error. Tell me the failure mode first.
+```
+
+## Client Setup
+
+> Clients do **not** need to install Node.js, Claude CLI, or the OCP repo. Only `curl` and `python3` are required (pre-installed on most Linux/Mac systems).
+>
+> **Find the server's LAN IP** by running `ocp lan` on the server machine — it prints both the IP and a ready-to-share connect command.
+
+**One-command setup** — download the lightweight `ocp-connect` script:
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/dtzp555-max/ocp/main/ocp-connect -o ocp-connect
+chmod +x ocp-connect
+./ocp-connect <server-ip>
+```
+
+**Zero-config** — when the server admin has set `PROXY_ANONYMOUS_KEY` *and* opted in with `PROXY_ADVERTISE_ANON_KEY=1` (see [Anonymous Access](#anonymous-access-optional) below), just pass the server IP and nothing else. `ocp-connect` reads the anonymous key from `/health` and uses it automatically. Without the opt-in, `/health` does not expose the key (issue #109); pass `--key` or rely on anonymous access instead:
+
+```bash
+./ocp-connect <server-ip>
+```
+
+If the server requires a key, pass it with `--key`:
+```bash
+./ocp-connect <server-ip> --key <your-api-key>
+```
+
+Or as a one-liner (no file saved):
+```bash
+curl -fsSL https://raw.githubusercontent.com/dtzp555-max/ocp/main/ocp-connect | bash -s -- <server-ip>
+```
+
+Example:
+```
+$ ./ocp-connect 192.168.1.100
+
+OCP Connect v1.3.0
+─────────────────────────────────────
+  Remote: http://192.168.1.100:3456
+
+  Checking connectivity...
+  ✓ Connected
+
+  Remote OCP v3.11.0  (auth: multi)
+
+  ⓘ Using server-advertised anonymous key: ocp_publ...n_v1
+    (set by admin via PROXY_ANONYMOUS_KEY; see issue #12 §14 Path A)
+
+  Testing API access...
+  ✓ API accessible (7 models available)
+
+  Shell config:
+    ✓ .bashrc
+    ✓ .zshrc
+    OPENAI_BASE_URL=http://192.168.1.100:3456/v1
+
+  System-level (launchctl):
+    ✓ OPENAI_BASE_URL set for GUI apps and daemons
+
+  IDE Configuration
+  ─────────────────────────────────────
+  Detected: OpenClaw (~/.openclaw/openclaw.json)
+
+  Configure OpenClaw to use this OCP? [Y/n] y
+  Provider name (models show as <name>/model-id) [ocp]: ocp
+
+  How should OCP models be configured?
+    1) Primary — use OCP by default, keep existing models as backup
+    2) Backup  — keep current primary, add OCP as additional option
+
+  Choice [1]: 1
+
+  Writing OpenClaw config...
+    ✓ Per-agent auth profile seeded (2):
+      • ~/.openclaw/agents/main/agent/auth-profiles.json
+      • ~/.openclaw/agents/macbook_bot/agent/auth-profiles.json
+  ✓ OpenClaw configured
+    Provider: ocp
+    Models:
+      • ocp/claude-opus-5
+      • ocp/claude-opus-4-8
+      • ocp/claude-opus-4-7
+      • ocp/claude-opus-4-6
+      • ocp/claude-sonnet-5
+      • ocp/claude-sonnet-4-6
+      • ocp/claude-haiku-4-5-20251001
+    Priority: PRIMARY (default model)
+
+    Restart OpenClaw to apply: openclaw gateway restart
+
+  Running smoke test...
+  ✓ Smoke test passed: OK
+    Note: smoke test only verifies OCP is reachable and the key is valid.
+    It does not verify your IDE/agent end-to-end. To verify OpenClaw works,
+    restart it (`openclaw gateway restart`) and send a test message to your bot.
+
+  Done. Reload your shell to apply:
+    source ~/.zshrc
+```
+
+The script automatically:
+- Writes env vars to all relevant shell rc files (`.bashrc`, `.zshrc`)
+- Sets system-level env vars (`launchctl setenv` on macOS, `environment.d` on Linux)
+- **Auto-discovers anonymous key** from `/health.anonymousKey` when no `--key` given (v1.3.0+, requires server v3.10.0+; server must also set `PROXY_ADVERTISE_ANON_KEY=1` — see [Anonymous Access](#anonymous-access-optional))
+- Configures OpenClaw automatically (including per-agent `auth-profiles.json` for multi-agent setups)
+- Detects Cline, Continue.dev, Cursor, and opencode, and prints setup hints (manual configuration required for these IDEs)
+
+On macOS, `launchctl setenv` vars reset on reboot — re-run `ocp-connect` after restart.
+
+**Manual setup** — if you prefer not to use the script:
+```bash
+export OPENAI_BASE_URL=http://<server-ip>:3456/v1
+export OPENAI_API_KEY=ocp_<your-key>
+```
+Add these lines to `~/.bashrc` or `~/.zshrc` to persist across sessions.
+
+## Monitoring (Server-side)
+
+```bash
+# Per-key usage stats
+ocp usage --by-key
+#  Key                  Reqs   OK  Err  Avg Time
+#  wife-laptop             5    5    0      8.0s
+#  son-ipad                3    3    0      6.2s
+
+# Manage keys
+ocp keys              # List all keys
+ocp keys revoke son-ipad   # Revoke a key
+```
+
+**Web Dashboard:** Open `http://<server-ip>:3456/dashboard` in any browser for real-time monitoring — per-key usage, request history, plan utilization, and system health.
+
+![OCP Dashboard](images/dashboard.png)
+
+## Auth Modes
+
+| Mode | Env | Use Case |
+|------|-----|----------|
+| `none` | `CLAUDE_AUTH_MODE=none` | Trusted home network, no auth needed |
+| `shared` | `CLAUDE_AUTH_MODE=shared` + `PROXY_API_KEY=xxx` | Everyone shares one key |
+| `multi` | `CLAUDE_AUTH_MODE=multi` + `OCP_ADMIN_KEY=xxx` | Per-person keys for usage tracking + quotas (trusted users only — see Deployment model below) |
+
+> **Usage scope (v3.14.0+):** `/api/usage` returns the caller's own rows by default. Admin callers must pass `?all=true` to retrieve data for all keys; doing so emits an audit log line.
+
+## Deployment model & security (read this)
+
+**What OCP is built for today: single-user, multi-IDE.** Run OCP as a server on one machine and point all of *your own* IDEs/devices at it — one Claude Pro/Max subscription, used everywhere. This is the primary, solid use case.
+
+**Sharing with family / a team — honest limits.** You *can* share OCP on a LAN, but be clear about what the auth modes do and don't give you:
+
+- The per-key modes (`shared` / `multi`) give per-key **usage tracking, quotas, and cache separation** — useful for seeing who used what and capping budgets.
+- They do **not** give a **security isolation boundary**. The spawned `claude` runs with the **operator's filesystem access** and is *not* sandboxed per key. **Only share with people you fully trust, on a trusted network.**
+- For simple trusted family sharing, the easiest setup is a single shared **anonymous key** (see [Anonymous Access](#anonymous-access-optional)) — no per-person separation, same trust assumption.
+- **Account terms and ToS — read before sharing with others.** Claude Pro/Max are *per-user* accounts. Pooling a single subscription across **multiple distinct people** may violate Anthropic's Consumer Terms of Service and risk account suspension by the abuse classifier. The defensible framing is **"one person, your own devices"** — sharing with friends or a team is not. OCP does not change your account terms, and whether any particular sharing setup complies with the ToS is the account holder's responsibility. Review Anthropic's Usage Policy before extending access to other people.
+
+**Real per-user isolation (sandboxed, multi-tenant-safe) is planned for after 2026-06-15** — per-key ephemeral home + tool lockdown + an OS sandbox. Until then, treat a multi-user OCP as a *trusted-group convenience*, not a security boundary. (This is also why `CLAUDE_TUI_MODE` is single-user-only — see [Subscription-pool (TUI) mode](tui-mode.md#subscription-pool-tui-mode).)
+
+## Anonymous Access (optional)
+
+In `multi` mode, the admin can designate a single well-known "anonymous" key that bypasses `validateKey()` and grants public read/write access. This is useful for letting LAN users (or clients like OpenClaw multi-agent setups) connect without individual per-user keys.
+
+**Enable**:
+
+The anonymous key is wired into the service unit (launchd plist on macOS, systemd unit on Linux) at install time. Export `PROXY_ANONYMOUS_KEY` in your shell before running `setup.mjs`, and `setup.mjs` will write it into the service unit env so the auto-started proxy picks it up:
+
+```bash
+export PROXY_ANONYMOUS_KEY=ocp_public_anon   # or any string of your choice
+node setup.mjs --bind 0.0.0.0 --auth-mode multi
+```
+
+If OCP is already installed without it, re-export the env var and re-run `node setup.mjs` (the installer is idempotent — it refreshes the service unit). Then `ocp restart` so the running proxy picks up the new env. Setting `PROXY_ANONYMOUS_KEY` only in your interactive shell **does not** affect the auto-started proxy — the service unit is the source of truth for its environment.
+
+**Client side**: the anonymous key value is exposed via `GET /health` as the field `anonymousKey` (null when not set) **only to localhost callers** or when the admin has also set `PROXY_ADVERTISE_ANON_KEY=1` (default off — see issue #109). With that opt-in, clients like `ocp-connect` can auto-discover and use it, so the end user doesn't need to get a personal key from the admin.
+
+**Security note**: setting this env var is an **opt-in** to public access — anyone who can reach your OCP endpoint can use it, up to any rate limits you configure. Don't enable this on internet-exposed OCP instances without additional protection.
+
+**Not a secret**: because `/health` is an unauthenticated endpoint, the anonymous key is **publicly readable** by anyone who can reach the server. That is intentional — the key exists so clients can self-configure without out-of-band coordination. Treat it as a convenience handle, not as an access credential.
+
+## Per-Key Quota (Budget Control)
+
+Prevent any single user from exhausting your subscription. Set daily, weekly, or monthly request limits per API key:
+
+```bash
+# Set a daily limit of 50 requests for a key
+curl -X PATCH http://127.0.0.1:3456/api/keys/wife-laptop/quota \
+  -H "Authorization: Bearer $OCP_ADMIN_KEY" \
+  -d '{"daily": 50}'
+
+# Set multiple limits at once
+curl -X PATCH http://127.0.0.1:3456/api/keys/son-ipad/quota \
+  -H "Authorization: Bearer $OCP_ADMIN_KEY" \
+  -d '{"daily": 20, "weekly": 100}'
+
+# Check current quota + usage
+curl http://127.0.0.1:3456/api/keys/wife-laptop/quota
+# → { "daily": { "limit": 50, "used": 12 }, "weekly": { "limit": null, "used": 34 }, ... }
+
+# Remove a limit (set to null)
+curl -X PATCH http://127.0.0.1:3456/api/keys/wife-laptop/quota \
+  -d '{"daily": null}'
+```
+
+When a key exceeds its quota, OCP returns HTTP 429 with a structured error:
+```json
+{
+  "error": {
+    "message": "Quota exceeded: 50/50 requests (daily). Resets 6h 12m.",
+    "type": "quota_exceeded",
+    "quota": { "period": "daily", "limit": 50, "used": 50, "resetsIn": "6h 12m" }
+  }
+}
+```
+
+- `null` = unlimited (default for all keys)
+- Only successful requests count toward quota
+- Admin and anonymous users are never subject to quotas
+- PATCH is a partial update — omitted fields are left unchanged
+
+> **Note:** quotas are best-effort. Under concurrent bursts a key can exceed its cap by up to the server's max-concurrency (default 8), and cache hits are not counted toward quota. They cap budgets for cooperative family use, not adversarial abuse.
+
+## Important Notes
+
+- All users share your Claude Pro/Max **rate limits** (5h session + 7d weekly)
+- `ocp usage` shows how much quota remains
+- Keys are stored in `~/.ocp/ocp.db` (SQLite, zero external dependencies)
+- Admin key is required for key management API endpoints
+- The dashboard (`/dashboard`) and health check (`/health`) are always public
+- File modes for `~/.ocp` (0700), `admin-key` + `ocp.db` (0600) are auto-tightened at server startup as of v3.14.0

docs/plans/2026-07-13-tui-latency/README.md

@@ -0,0 +1,236 @@
+# TUI-mode latency: measured floor, and the four things worth fixing
+
+**Date**: 2026-07-13
+**Status**: findings + backlog. **Superseded in part** — see the dated update boxes below.
+Item #1 shipped ([#156](https://github.com/dtzp555-max/ocp/pull/156)); item #2 is **dead**
+([`streaming-spike.md`](streaming-spike.md)); item #4 measured, **no effect**; item #3 stands.
+**Measured on**: Mac mini / macOS 26.5.2 / Claude Code **v2.1.207** / Sonnet 5 / Claude Max subscription / **real-home mode** (no `CLAUDE_CODE_OAUTH_TOKEN`, no `OCP_TUI_HOME` in the service env)
+**Evidence**: [`measurements.jsonl`](measurements.jsonl) — **n=15** (3 configs × 5) · banner captures [`billing-banner.txt`](billing-banner.txt) · harness [`floor.sh`](floor.sh)
+
+## Why this exists
+
+An external consumer (the 知音 AI project) benchmarked OCP's prompt path and measured
+**TTFT p50 ≈ 30–32 s**, and excluded OCP as a backend on that basis. That number is real,
+but it is *not* the model being slow — this document decomposes where the 30 seconds
+actually go, and what OCP can do about it.
+
+**The harness deliberately does not go through OCP.** It spawns `tmux` + `claude` directly
+(session prefix `zhiyin-floor-`, never `ocp-tui-*`) and polls `tmux capture-pane` for
+incremental render, so it measures the **true first-token time** of the underlying
+subscription path — the floor OCP could reach if it were perfect.
+
+---
+
+## Measurements
+
+All rows in [`measurements.jsonl`](measurements.jsonl); every number below is recomputable from it.
+
+| Config | n | boot→input-ready (median) | **TTFT (median)** | TTFT range | full answer (median) |
+|---|---|---|---|---|---|
+| baseline (inherits global `effortLevel: xhigh`) | 5 | 1.07 s | **10.35 s** | 8.32 – 17.19 s | 11.32 s |
+| **`--effort low`** | 5 | 1.03 s | **6.17 s** | **5.87 – 6.44 s** | 9.98 s |
+| `--bare` | 5 | 0.44 s | **no answer at all** (5/5 `ttft_ms: -1`) | — | — |
+
+> **Not from this harness**: the direct Anthropic API reference figure (TTFT 0.84–1.64 s, n=2)
+> comes from the 知音 AI project's own smoke test, not from `measurements.jsonl`. It is quoted
+> only to size the gap; do not look for it in the evidence file.
+
+### Where the 30 seconds go
+
+```
+ ~1.0 s   spawn → claude's input bar is ready          ← NOT the bottleneck
+ ~6-10 s  true TTFT (first token rendered in the pane)
+ ~20 s    ████ waiting for the whole turn to finish ████  ← this is the 30s
+```
+
+`runTuiTurn` blocks on the native transcript until a terminal event (`lib/tui/session.mjs`
+"Block on the native transcript … until terminal"; `readTuiTranscript` in
+`lib/tui/transcript.mjs`; ADR 0007 step 4) — i.e. it waits for the **entire turn** to complete
+before returning anything. There is no streaming path. The ~20 s delta between this harness's
+real TTFT and OCP's reported 30–32 s is exactly that.
+
+> **⚠️ 2026-07-13 correction — this decomposition attributes the ~20 s to the wrong thing.** It was
+> inferred from the external 30–32 s report, never measured *through* OCP. It has since been measured
+> through a real OCP instance (TUI mode, `claude-sonnet-4-6`, the same ~1850-token prompt, n=5):
+> **median 11.30 s** before [#156](https://github.com/dtzp555-max/ocp/pull/156), **9.55 s** after.
+> Same-turn decomposition (baseline row `i=5`): **11.563 s** wall through OCP vs `turn_duration:
+> 7.319 s` of CLI-internal time on that same turn → **OCP's own overhead ≈ 4.2 s** (n=1), **not
+> ~20 s**. The rest of any larger number is the model *generating a long answer*,
+> which the blocking wait does not cause and streaming would not shorten — it would only move the
+> first byte earlier. The 30–32 s figure therefore reflects a much longer output (and/or the
+> then-inherited `xhigh` effort), not 20 s of OCP dead time. See
+> [`streaming-spike.md`](streaming-spike.md) § "What streaming would have bought".
+
+---
+
+## ⚠️ Blocking constraint: `--bare` silently drops you off the subscription pool
+
+Captured live ([`billing-banner.txt`](billing-banner.txt)) — the startup banner is the **only**
+reliable indicator:
+
+```
+[]                     | Sonnet 5 with xhigh effort · Claude Max
+[--effort low]         | Sonnet 5 with low effort  · Claude Max
+[--bare]               | Sonnet 5 with xhigh effort · API Usage Billing     ← ❌
+```
+
+`--bare` ("skip hooks, LSP, plugin…") **also skips the subscription-credential resolution
+path**. It really does cut boot to 0.43–0.45 s — but you are no longer on the subscription,
+which defeats the entire purpose of TUI mode (ADR 0007 exists solely to reach the
+subscription pool).
+
+**The failure is silent.** All 5 `--bare` samples reached input-ready (boot 0.43–0.45 s), were
+sent the prompt, and then produced **no answer at all** — 60 s timeout, no error, no crash, the
+pane simply never rendered a token (the API-billing account had no credit balance). Nothing in
+the transcript or the exit status reveals this.
+
+**Anyone changing spawn flags must diff the banner line before and after.**
+
+---
+
+## Backlog — four items, ranked by value ÷ effort
+
+### 1. Pass `--effort` explicitly on spawn — **do this first**
+
+`buildTuiCmd` (`lib/tui/session.mjs`) does not pass `--effort` — `grep -rn -- "--effort\|effortLevel" lib/ server.mjs`
+returns zero hits. What the pane's `claude` ends up using therefore depends on **which HOME mode
+`resolveTuiHome()` picked**:
+
+| mode | HOME | effort the pane gets |
+|---|---|---|
+| **real-home** (legacy default — *current* service config: no `CLAUDE_CODE_OAUTH_TOKEN`, no `OCP_TUI_HOME`) | `~` | **inherits the operator's `~/.claude/settings.json` → `effortLevel: xhigh` on this host** |
+| env-token scratch (`CLAUDE_CODE_OAUTH_TOKEN` set — the direction #146/#150 pushed) | `~/.ocp-tui/home` | that settings.json contains only `permissions.additionalDirectories`; `prepareTuiHome()` never writes `effortLevel` → **claude's built-in default** |
+
+**Scope note**: TUI mode is currently *off* on this host (`CLAUDE_TUI_MODE=false`; `/health` →
+`"tui": {"enabled": false}`), so live traffic takes the `-p` path today. The statement below is
+about what happens **when TUI mode is enabled**.
+
+On the current HOME config, **every TUI request would run extended thinking** — pure waste
+for the typical "generate this JSON" request, and it makes latency depend on an unrelated global
+setting the operator may have changed for their own interactive use. And the mode split means
+the effort level silently changes if the operator ever switches to env-token mode.
+**Passing `--effort` explicitly fixes both problems at once.**
+
+- **Effect (real-home, measured)**: TTFT p50 **10.35 s → 6.17 s (−40 %)**, and the spread
+  collapses from 8.32–17.19 s to **5.87–6.44 s**. For a proxy, the variance reduction matters
+  more than the median.
+- **Cost**: one flag. Suggested: a new `OCP_TUI_EFFORT` env var (default `low`), documented in
+  README § "Environment Variables" per `release_kit.new_feature_doc_expectations`.
+- **Risk**: none — banner confirms it stays on `Claude Max` (see `billing-banner.txt`).
+- ⚠️ Do **not** reach for `--bare` to shave boot: see above.
+
+### 2. Real streaming instead of blocking on turn-terminal — **ACHIEVABLE → [`streaming-spike.md`](streaming-spike.md)**
+
+> **2026-07-13 update — the prereq spike was run. The answer is YES, but not from either source this
+> item guessed at.** (a) The transcript grows at *event* granularity (the whole answer lands in one
+> line, ~0.3 s before terminal) — dead. (b) The pane is a **rendered** view whose `capture-pane` text
+> no longer contains the answer's source bytes (`## `, `**`, code fences are gone) — dead, and worse
+> than "lossy": it is *not the model's text*. **But there is a third source neither this backlog nor
+> the first spike considered: `claude` fires a `MessageDisplay` hook carrying incremental,
+> byte-faithful `delta`s of the raw reply.** Verified live on a plain interactive TUI spawn (no `-p`),
+> banner `· Claude Max`: 7 fires spread across generation, `concat(deltas) === T` **byte-exactly**
+> (579 == 579), `T.startsWith(S)` true at every step, `## ` / `**` / ```` ```javascript ```` all
+> present in the deltas. Granularity is block-level (~5–7 chunks/answer), not token-level — plenty for
+> SSE. **Build it.**
+>
+> ⚠️ Two corrections to this item as written: the **"~20 s" is wrong** (inferred from an external
+> report, never measured through OCP — the same-turn decomposition puts OCP's own overhead at **~4 s**,
+> n=1), and **streaming moves the first byte, not the last** — so a consumer needing the *complete*
+> answer (the JSON-card case that motivated this) gains **nothing** from it. Build it for
+> progressively-rendering consumers, not as a throughput win.
+>
+> Full evidence + implementer caveats (the hook is `forceSyncExecution` — claude BLOCKS on it):
+> **[`streaming-spike.md`](streaming-spike.md)**. Original framing preserved below.
+
+Today `runTuiTurn` blocks on the transcript until the turn is *finished*. The pane is already
+rendering tokens incrementally the whole time — this harness proves you can observe first token
+at ~6 s by polling `tmux capture-pane`.
+
+- **Effect**: turns a 30 s wall into a ~6 s TTFT with progressive output; enables SSE streaming
+  on the OCP endpoint instead of a single blob at the end.
+- **Cost**: real work. Pane capture is ANSI/redraw-based and lossy for exact text (wrapping,
+  scrollback, spinner lines). Two candidate sources: (a) incremental reads of the transcript
+  JSONL, (b) `capture-pane` diffing with a stable start marker. (a) is much cleaner **if it
+  holds**.
+- **Prereq spike (do this before designing anything)**: does the transcript JSONL grow *during*
+  a turn, or only at the end? If only at the end, (a) is dead and you are stuck with (b).
+
+### 3. Warm pane pool — ~1 s
+
+Every request spawns a fresh tmux session + `claude` (`randomUUID()` + `new-session`, then
+`kill-session` in `finally`; `grep -rn "pool\|warm\|reuse" lib/tui/*.mjs` → zero hits). Boot to
+input-ready is ~1.0 s, paid on every request. A pool of pre-booted panes (single-use, replaced in
+the background) amortizes it to zero for any workload below the pool refill rate.
+
+- **Effect**: −1.0 s.
+- **Cost**: moderate; interacts with the session reaper and the per-port prefix scoping added in
+  #148 — pooled panes must not look like zombies to the sweep.
+- Lower priority than #1 and #2: it is the smallest slice.
+
+### 4. Trim the prefill — ~~probably not worth it~~ **MEASURED: no detectable benefit. Do not adopt.**
+
+> **2026-07-13 update.** `--exclude-dynamic-system-prompt-sections` was measured with the same
+> harness (`floor.sh`, n=5, Sonnet 5, on top of `--effort low`): **TTFT median 6.39 s**
+> (5.87–10.54 s) vs **6.17 s** (5.87–6.44 s) for `--effort low` alone — i.e. **0.22 s worse, inside
+> the noise band**, with one worse outlier; dropping that outlier does not change the verdict. n=5
+> cannot prove "zero", only "no benefit detectable above noise" — but there is also a **mechanistic**
+> reason not to expect one: `--help` says the flag *"Improves cross-user prompt-cache **reuse**"*, and
+> **OCP is single-user** — there is no cross-user cache to share, so the flag has nothing to buy here.
+> The banner stayed on `· Claude Max` (no billing-pool drop), but there is no win to bank. The ~6 s
+> floor stands as stated below. Raw rows: [`prefill-spike-measurements.jsonl`](prefill-spike-measurements.jsonl).
+
+
+After #1–#3, the floor is **~6 s**, and it does not go lower. `claude` always injects the full
+Claude Code system prompt + tool definitions (thousands to tens of thousands of prefill tokens)
+regardless of what you ask it. `--exclude-dynamic-system-prompt-sections` exists and may shave
+some of it — **unmeasured**; worth one spike, but do not expect to reach the direct API's
+~1 s.
+
+**Consequence to accept, and to state in the README**: even fully optimized, TUI mode has a
+**~6 s TTFT floor**, so it cannot serve real-time / interactive-latency consumers. It remains
+appropriate for batch, background, and cost-insensitive-latency use. The 知音 AI project
+excluded it on this basis (their prompt-latency budget is 2–4 s) *independently* of the ToS
+question already documented in the README.
+
+---
+
+## Reproduction
+
+```bash
+# harness never touches OCP's :3456 service or ocp-tui-* sessions, and never kill-server
+bash docs/plans/2026-07-13-tui-latency/floor.sh 5                                  # baseline
+TAG=effort-low EXTRA_ARGS="--effort low" bash .../floor.sh 5                       # −40 %
+TAG=bare       EXTRA_ARGS="--bare"       bash .../floor.sh 5                       # the trap
+
+# billing-pool check for ANY spawn-flag change — the banner is the only source of truth
+tmux new-session -d -s probe -x 200 -y 50 -c "$HOME" \
+  "claude --model claude-sonnet-5 --session-id $(uuidgen) <your-flags-here>"
+sleep 6; tmux capture-pane -p -t probe | grep -E "Claude Max|API Usage Billing"
+tmux kill-session -t probe
+```
+
+## Interaction with OCP while the harness runs
+
+- **Kill direction is safe both ways**: `reapStaleTuiSessions()` only `kill-session`s names
+  matching `ocp-tui-<port>-`, which `zhiyin-floor-*` never matches; and the harness only
+  `kill-session`s its own single session — it contains **no `kill-server`**.
+- **One benign interaction** (only when TUI mode is enabled — the reap tick is itself gated on
+  `TUI_MODE`): OCP's periodic `kill-server` (zombie reaping) is gated on
+  `othersRemain` — *any* foreign-prefixed tmux session suppresses it. So while the harness is
+  running, that sweep is skipped. This is the coexistence guard working as designed; it resumes
+  on the next tick.
+
+## Harness caveats (stated so the numbers are not over-trusted)
+
+- **n=5 per config**, single host, single model (Sonnet 5), single prompt size (~1850 tokens).
+  Enough to separate 6 s from 10 s from 30 s; **not** enough for a p95.
+- TTFT is "marker visible in `capture-pane`", which includes tmux render latency (small, but
+  nonzero) — it is an upper bound on the true first-token time.
+- **The harness's readiness marker is not OCP's.** `floor.sh` waits for `│ >|❯|Try "`; OCP's
+  `tuiInputReady()` matches `/\? for shortcuts/`. These are different events, so the ~1.0 s
+  boot figure is **not** directly comparable to OCP's `BOOT_MS` gate (default cap 4000 ms). It
+  does not affect the conclusions (1 s ≪ 6 s TTFT), but it is not apples-to-apples.
+- The first version of this harness reported TTFT **0.08 s** — a false positive: the prompt
+  literally contained the marker string it was grepping for, so the match fired the instant the
+  prompt was pasted. Fixed by describing the marker instead of spelling it. **The script exited 0
+  and "successfully" produced 5 samples both times** — exit status proves nothing here.

docs/plans/2026-07-13-tui-latency/billing-banner.txt

@@ -0,0 +1,3 @@
+[]                     | ▝▜█████▛▘  Sonnet 5 with xhigh effort · Claude Max
+[--effort low]         | ▝▜█████▛▘  Sonnet 5 with low effort · Claude Max
+[--bare]               | ▝▜█████▛▘  Sonnet 5 with xhigh effort · API Usage Billing

docs/plans/2026-07-13-tui-latency/floor.sh

@@ -0,0 +1,128 @@
+#!/usr/bin/env bash
+# OCP TUI-mode latency floor harness — see README.md in this directory.
+#
+# 目的：回答一个问题——如果把 OCP 现有的两个已知开销砍掉
+#   (a) 每请求 spawn + boot（可用预热进程池消除）
+#   (b) 假流式（等 turn_duration 才返回，可用增量读 pane 消除）
+# 之后，订阅池路径的**真实 TTFT 地板**是多少？
+#
+# 判据：地板 ≤ 4s → OCP 作为"省钱选项"可行；> 8s → 死透，不再讨论。
+#
+# 红线：
+# - 不经过生产 OCP 服务（:3456）—— 直接起 tmux+claude，OCP 进程零干扰
+# - tmux session 前缀用 zhiyin-floor-（**不是** ocp-tui-），避免被 OCP 的
+#   reaper 当成自己的会话杀掉，也避免我们杀到它的
+# - 用 real HOME（凭据）—— scratch HOME + symlink 凭据会 fork OAuth 导致 401
+#   （见跨机记忆 tui_scratch_home_credential_fork）
+set -uo pipefail
+
+N=${1:-5}
+MODEL=${MODEL:-claude-sonnet-5}
+EXTRA_ARGS=${EXTRA_ARGS:-}          # 额外 CLI 参数（如 --effort low --bare）
+TAG=${TAG:-baseline}
+OUT=${OUT:-$(dirname "$0")/measurements.jsonl}
+PROMPT_FILE=$(mktemp)
+PREFIX="zhiyin-floor"
+
+mkdir -p "$(dirname "$OUT")"
+
+# ── 构造提示：~2000 token 的假会议转写 + 明确的起始标记 ────────────────
+# 单行（多行会在 tmux send-keys 时提前触发 Enter）
+build_prompt() {
+  local seg="Speaker A said the quarterly pipeline is tracking behind plan and the enterprise segment needs a different motion. Speaker B replied that the current onboarding flow loses roughly a third of trial accounts before the first integration is complete. They debated whether the fix belongs in product or in customer success. "
+  local body=""
+  for _ in $(seq 1 22); do body+="$seg"; done
+  printf '%s' "You are a real-time meeting copilot. Meeting transcript so far: $body --- Task: produce ONE prompt card as compact JSON with keys: points (array of 3 short Chinese bullet points), keyline (one English sentence the user can read aloud). IMPORTANT: your reply MUST begin with three hash characters immediately followed by the uppercase word CARD (no space between them), then the JSON. No preamble, no markdown fences." > "$PROMPT_FILE"
+}
+build_prompt
+PROMPT_CHARS=$(wc -c < "$PROMPT_FILE" | tr -d ' ')
+
+now_ms() { python3 -c 'import time;print(int(time.time()*1000))'; }
+
+echo "配置: $TAG   参数: [$EXTRA_ARGS]"
+echo "模型: $MODEL   样本: $N   提示长度: ${PROMPT_CHARS} chars (≈$((PROMPT_CHARS/4)) token)"
+echo "输出: $OUT"
+echo
+
+for i in $(seq 1 "$N"); do
+  SESS="${PREFIX}-$$-$i"
+  SID=$(uuidgen)
+
+  # ── 冷启动：spawn + 等输入框就绪 ─────────────────────────────────
+  T_SPAWN=$(now_ms)
+  tmux new-session -d -s "$SESS" -x 200 -y 50 \
+    -e CLAUDE_CODE_DISABLE_CLAUDE_MDS=1 \
+    -e CLAUDE_CODE_DISABLE_AUTO_MEMORY=1 \
+    -c "$HOME" \
+    "claude --model $MODEL --session-id $SID --strict-mcp-config --disallowedTools 'mcp__*' $EXTRA_ARGS" 2>/dev/null
+  if [ $? -ne 0 ]; then echo "[$i] tmux spawn 失败，跳过"; continue; fi
+
+  # 轮询输入框就绪（claude TUI 的输入提示符）
+  READY=0
+  for _ in $(seq 1 150); do   # 上限 15s
+    PANE=$(tmux capture-pane -p -t "$SESS" 2>/dev/null || true)
+    if grep -qE '│ >|❯|Try "' <<<"$PANE"; then READY=1; break; fi
+    sleep 0.1
+  done
+  T_READY=$(now_ms)
+  BOOT_MS=$((T_READY - T_SPAWN))
+  if [ "$READY" -ne 1 ]; then
+    echo "[$i] 启动超时（${BOOT_MS}ms），pane 末 3 行:"
+    tmux capture-pane -p -t "$SESS" 2>/dev/null | tail -3 | sed 's/^/      /'
+    tmux kill-session -t "$SESS" 2>/dev/null
+    continue
+  fi
+
+  # ── 热态：粘提示 → 回车 → 量首 token ─────────────────────────────
+  tmux send-keys -t "$SESS" -l "$(cat "$PROMPT_FILE")" 2>/dev/null
+  sleep 0.4                       # 让粘贴落地（OCP 用 400ms 轮询粒度）
+  T0=$(now_ms)
+  tmux send-keys -t "$SESS" Enter 2>/dev/null
+
+  TTFT_MS=-1
+  for _ in $(seq 1 600); do       # 上限 60s
+    if tmux capture-pane -p -t "$SESS" 2>/dev/null | grep -q '###CARD'; then
+      TTFT_MS=$(( $(now_ms) - T0 )); break
+    fi
+    sleep 0.1
+  done
+
+  # ── 完整回答：pane 连续 2s 不再变化 ──────────────────────────────
+  COMPLETE_MS=-1
+  if [ "$TTFT_MS" -ge 0 ]; then
+    LAST=""; STABLE=0
+    for _ in $(seq 1 900); do     # 上限 90s
+      CUR=$(tmux capture-pane -p -t "$SESS" 2>/dev/null | cksum)
+      if [ "$CUR" = "$LAST" ]; then
+        STABLE=$((STABLE+1))
+        [ "$STABLE" -ge 20 ] && { COMPLETE_MS=$(( $(now_ms) - T0 - 2000 )); break; }
+      else
+        STABLE=0; LAST="$CUR"
+      fi
+      sleep 0.1
+    done
+  fi
+
+  printf '{"i":%d,"tag":"%s","model":"%s","extra_args":"%s","prompt_chars":%s,"boot_ms":%d,"ttft_ms":%d,"complete_ms":%d}\n' \
+    "$i" "$TAG" "$MODEL" "$EXTRA_ARGS" "$PROMPT_CHARS" "$BOOT_MS" "$TTFT_MS" "$COMPLETE_MS" | tee -a "$OUT"
+
+  tmux kill-session -t "$SESS" 2>/dev/null
+  sleep 1
+done
+
+rm -f "$PROMPT_FILE"
+echo
+echo "=== 汇总 ==="
+python3 - "$OUT" <<'EOF'
+import json,sys,statistics
+rows=[json.loads(l) for l in open(sys.argv[1]) if l.strip()]
+ok=[r for r in rows if r['ttft_ms']>=0]
+if not ok: print("无有效样本"); sys.exit()
+def s(k):
+    v=[r[k] for r in ok if r[k]>=0]
+    return f"n={len(v)} 中位={statistics.median(v)/1000:.2f}s 最小={min(v)/1000:.2f}s 最大={max(v)/1000:.2f}s" if v else "无"
+print(f"  冷启动 boot      : {s('boot_ms')}   ← 预热进程池可完全消除")
+print(f"  TTFT（首 token） : {s('ttft_ms')}   ★ 这就是地板")
+print(f"  完整回答         : {s('complete_ms')}")
+print(f"\n  失败样本: {len(rows)-len(ok)}/{len(rows)}")
+EOF

docs/plans/2026-07-13-tui-latency/measurements.jsonl

@@ -0,0 +1,15 @@
+{"i": 1, "tag": "effort-low", "model": "claude-sonnet-5", "extra_args": "--effort low", "prompt_chars": 7451, "boot_ms": 1077, "ttft_ms": 6172, "complete_ms": 9929}
+{"i": 2, "tag": "effort-low", "model": "claude-sonnet-5", "extra_args": "--effort low", "prompt_chars": 7451, "boot_ms": 1026, "ttft_ms": 6160, "complete_ms": 9996}
+{"i": 3, "tag": "effort-low", "model": "claude-sonnet-5", "extra_args": "--effort low", "prompt_chars": 7451, "boot_ms": 1010, "ttft_ms": 6437, "complete_ms": 9977}
+{"i": 4, "tag": "effort-low", "model": "claude-sonnet-5", "extra_args": "--effort low", "prompt_chars": 7451, "boot_ms": 1033, "ttft_ms": 5872, "complete_ms": 9944}
+{"i": 5, "tag": "effort-low", "model": "claude-sonnet-5", "extra_args": "--effort low", "prompt_chars": 7451, "boot_ms": 1154, "ttft_ms": 6387, "complete_ms": 9993}
+{"i":1,"tag":"baseline","model":"claude-sonnet-5","extra_args":"","prompt_chars":7451,"boot_ms":1300,"ttft_ms":8321,"complete_ms":9939}
+{"i":2,"tag":"baseline","model":"claude-sonnet-5","extra_args":"","prompt_chars":7451,"boot_ms":1070,"ttft_ms":10347,"complete_ms":11320}
+{"i":3,"tag":"baseline","model":"claude-sonnet-5","extra_args":"","prompt_chars":7451,"boot_ms":911,"ttft_ms":13061,"complete_ms":15163}
+{"i":4,"tag":"baseline","model":"claude-sonnet-5","extra_args":"","prompt_chars":7451,"boot_ms":1441,"ttft_ms":9981,"complete_ms":11066}
+{"i":5,"tag":"baseline","model":"claude-sonnet-5","extra_args":"","prompt_chars":7451,"boot_ms":1036,"ttft_ms":17189,"complete_ms":17985}
+{"i":1,"tag":"bare","model":"claude-sonnet-5","extra_args":"--bare","prompt_chars":7451,"boot_ms":429,"ttft_ms":-1,"complete_ms":-1}
+{"i":2,"tag":"bare","model":"claude-sonnet-5","extra_args":"--bare","prompt_chars":7451,"boot_ms":437,"ttft_ms":-1,"complete_ms":-1}
+{"i":3,"tag":"bare","model":"claude-sonnet-5","extra_args":"--bare","prompt_chars":7451,"boot_ms":444,"ttft_ms":-1,"complete_ms":-1}
+{"i":4,"tag":"bare","model":"claude-sonnet-5","extra_args":"--bare","prompt_chars":7451,"boot_ms":446,"ttft_ms":-1,"complete_ms":-1}
+{"i":5,"tag":"bare","model":"claude-sonnet-5","extra_args":"--bare","prompt_chars":7451,"boot_ms":441,"ttft_ms":-1,"complete_ms":-1}

docs/plans/2026-07-13-tui-latency/messagedisplay-deltas.jsonl

@@ -0,0 +1,7 @@
+{"hook_event_name": "MessageDisplay", "index": 0, "final": false, "delta": "## Mutex\n\n"}
+{"hook_event_name": "MessageDisplay", "index": 1, "final": false, "delta": "A **mutual exclusion lock** prevents concurrent access to a shared resource, ensuring only one thread runs the critical section at a time.\n\n"}
+{"hook_event_name": "MessageDisplay", "index": 2, "final": false, "delta": "- Acquiring a locked mutex blocks the caller until the current holder releases it.\n"}
+{"hook_event_name": "MessageDisplay", "index": 3, "final": false, "delta": "- Failing to release a mutex causes a deadlock, freezing all waiting threads.\n\n```javascript\nconst { Mutex } = require('async-mutex');\n\nconst mutex = new Mutex();\n"}
+{"hook_event_name": "MessageDisplay", "index": 4, "final": false, "delta": "let counter = 0;\n\nasync function increment() {\n  const release = await mutex.acquire();\n  try {\n"}
+{"hook_event_name": "MessageDisplay", "index": 5, "final": false, "delta": "    counter++; // only one caller here at a time\n  } finally {\n    release();\n  }\n}\n"}
+{"hook_event_name": "MessageDisplay", "index": 6, "final": true, "delta": "```"}

docs/plans/2026-07-13-tui-latency/prefill-spike-measurements.jsonl

@@ -0,0 +1,5 @@
+{"i":1,"tag":"effort-low-exclude-dynamic","model":"claude-sonnet-5","extra_args":"--effort low --exclude-dynamic-system-prompt-sections","prompt_chars":7451,"boot_ms":934,"ttft_ms":5867,"complete_ms":9953}
+{"i":2,"tag":"effort-low-exclude-dynamic","model":"claude-sonnet-5","extra_args":"--effort low --exclude-dynamic-system-prompt-sections","prompt_chars":7451,"boot_ms":1275,"ttft_ms":6388,"complete_ms":9874}
+{"i":3,"tag":"effort-low-exclude-dynamic","model":"claude-sonnet-5","extra_args":"--effort low --exclude-dynamic-system-prompt-sections","prompt_chars":7451,"boot_ms":874,"ttft_ms":10537,"complete_ms":11782}
+{"i":4,"tag":"effort-low-exclude-dynamic","model":"claude-sonnet-5","extra_args":"--effort low --exclude-dynamic-system-prompt-sections","prompt_chars":7451,"boot_ms":1170,"ttft_ms":6379,"complete_ms":9947}
+{"i":5,"tag":"effort-low-exclude-dynamic","model":"claude-sonnet-5","extra_args":"--effort low --exclude-dynamic-system-prompt-sections","prompt_chars":7451,"boot_ms":1329,"ttft_ms":6443,"complete_ms":9884}

docs/plans/2026-07-13-tui-latency/streaming-spike.md

@@ -0,0 +1,256 @@
+# Backlog #2 (real streaming): **achievable** — via the `MessageDisplay` hook
+
+**Date**: 2026-07-13
+**Status**: prereq-spike result. **Streaming IS achievable on the TUI path**, byte-faithfully, on the
+subscription pool. Three obvious sources are dead ends; a fourth one works.
+**Scope**: answers the prereq spike that [`README.md`](README.md) § "Backlog #2" demanded *before* any
+streaming design:
+
+> **Prereq spike (do this before designing anything)**: does the transcript JSONL grow *during* a
+> turn, or only at the end? If only at the end, (a) is dead and you are stuck with (b).
+
+The answer: **(a) is dead, (b) is dead — and you are not stuck with either.** The CLI exposes its own
+streaming interface as a **hook**, which the backlog did not consider.
+
+**Measured on**: Mac mini / Claude Code **v2.1.207** / Sonnet 4.6 + Sonnet 5 / Claude Max /
+real-home mode. Every claim below is reproducible from the commands given.
+
+> **Honesty note on how this document was produced.** Its first version concluded the exact opposite —
+> "streaming is not achievable; the CLI exposes no byte-faithful incremental source" — and was **wrong**.
+> An adversarial reviewer, commissioned specifically to *refute* it, found `MessageDisplay` on a second
+> pass; its own first pass had enumerated the hook registry with a truncated grep (it reported 21
+> events — there are **30**). Both the wrong conclusion and its refutation are preserved here, because
+> "we checked, it's impossible" is the most expensive kind of claim to get wrong: it closes a door and
+> nobody re-opens it.
+
+---
+
+## ✅ The source that works: the `MessageDisplay` hook
+
+`claude` fires a **`MessageDisplay`** hook as it renders each block of the assistant's reply. The
+payload carries the **raw markdown source** of an incremental `delta`, plus a monotonic `index` and a
+`final` flag:
+
+```json
+{ "hook_event_name": "MessageDisplay",
+  "turn_id": "6cb31d21-…", "message_id": "84ab9832-…",
+  "index": 0, "final": false, "delta": "## Mutex\n\n" }
+```
+*(payload also carries `session_id`, `transcript_path`, `prompt_id`, `cwd`)*
+
+Registered as an ordinary command hook via `--settings` on a **plain interactive TUI spawn** (no `-p`,
+no `--bare`), `claude-sonnet-4-6`, `--effort low`. Banner verified:
+`▝▜█████▛▘ Sonnet 4.6 with low effort · Claude Max` — **subscription pool, not metered billing**.
+
+One live turn — 7 fires, spread across generation:
+
+```
+index=0 final=false  len=  10  '## Mutex\n\n'
+index=1 final=false  len= 140  'A **mutual exclusion lock** prevents concurrent access to a shar…'
+index=2 final=false  len=  83  '- Acquiring a locked mutex blocks the caller until the current h…'
+index=3 final=false  len= 163  '- Failing to release a mutex causes a deadlock, freezing all wai…'
+index=4 final=false  len=  96  'let counter = 0;\n\nasync function increment() {\n  const release =…'
+index=5 final=false  len=  84  '    counter++; // only one caller here at a time\n  } finally {\n …'
+index=6 final=true   len=   3  '```'
+```
+
+**Every invariant a proxy needs — all hold:**
+
+| requirement | result |
+|---|---|
+| **byte-faithful** — deltas are the model's *source*, not the rendered pane | ✅ `## `, `**`, ```` ```javascript ```` all present in the deltas |
+| **exactness** — `concat(deltas) === T` (the transcript-authoritative text) | ✅ **true**, 579 == 579 bytes |
+| **prefix-stable** — `T.startsWith(concat(deltas[0..n]))` at every n | ✅ **true at all 7 steps** |
+| **incremental** — arrives during generation, not at the end | ✅ 7 fires spread across the turn |
+| **no `-p`** — stays out of the metered `sdk-cli` pool | ✅ plain interactive TUI |
+| **subscription pool** | ✅ banner `· Claude Max` |
+
+This is exactly the contract a streaming design needs: deltas forward straight into SSE
+`delta.content` chunks, and the transcript's final text `T` stays a cheap end-of-turn assertion
+(`concat === T`) instead of a reconciliation problem.
+
+### Caveats for the implementer
+
+- **Block-level granularity, not token-level** — the hook fires **once per rendered block** (roughly one
+  per paragraph / list item / code block), so the chunk count **scales with answer length**: 7 fires for a
+  ~600-byte answer, **18 for a ~2 KB one**. Plenty for SSE (`delta.content` has no minimum size), but do
+  not promise token-by-token output, and do not hard-code any assumption about chunk count.
+- **🔴 The sink MUST be keyed by `session_id` — this is live TODAY, not a future concern.**
+  `OCP_TUI_MAX_CONCURRENT` defaults to **2**, so **two `claude` processes already run concurrently**. One
+  hook command writing to one shared sink would **interleave deltas from two different turns into one
+  stream** — request A's client receiving request B's text, the worst failure a proxy can have, and one a
+  single-request test will never surface. The payload carries `session_id` (and `turn_id` / `message_id`),
+  so demux is easy: derive the sink path from `session_id` (`<dir>/<session_id>.jsonl`) and read only your
+  own turn's file. This *also* keeps the design **warm-pool compatible**, because a pre-booted pane's
+  session-id is fixed at boot — one static hook script serves every pane. **Test it with ≥2 concurrent
+  streaming requests carrying distinguishable prompts and assert zero cross-contamination.**
+- **⚠️ `forceSyncExecution: true` in the hook's source — `claude` BLOCKS on the hook.** A slow hook
+  adds latency to *every* delta. The hook must write and exit immediately (e.g. write to a FIFO / unix
+  socket that OCP reads; never work inline). **Measure the added per-delta latency.**
+- **Thinking blocks appear to be excluded — but this is NOT yet stress-tested. Verify before shipping.**
+  The exclusion is inferred from `content.map(c => c.type === "text" ? c.text : "")` — but that snippet is
+  from the **`final:true`** call site, not the incremental one. Four live turns (incl. two at `--effort
+  high`) showed no thinking text in any delta and `concat === T` held — **but each transcript's thinking
+  block was empty (`thinking:""`, 0 chars)**, so the exclusion was never actually stressed. **The failure
+  mode is severe**: if thinking deltas *do* fire on some config (Opus, `xhigh`), `concat(deltas) !== T`
+  **and OCP streams the model's private reasoning to the caller**. The end-of-turn `concat === T` assertion
+  would *detect* that but **cannot prevent** it — SSE deltas cannot be un-sent. **Before shipping, run a
+  turn on a model+effort that produces substantive thinking** (a hard reasoning prompt on Opus / `xhigh`)
+  and confirm both (a) no thinking text in any delta and (b) `concat === T` still holds.
+- OCP already owns the spawn (isolated HOME, its own flags), so injecting `--settings` with a
+  `MessageDisplay` hook sits inside the existing architecture.
+- **`ALIGNMENT.md`**: this consumes `claude`'s **own** hook surface as emitted — forwarding, not
+  inventing. Not a new endpoint, not a fabricated protocol. (Class B / ADR 0007 — the TUI spawn is
+  OCP-owned; no `cli.js` citation applies.)
+
+### Reproduce in 60 seconds
+
+```bash
+# hook script: append the payload (arrives on stdin) and exit immediately
+printf '#!/bin/bash\ncat >> "$MD_LOG"; printf "\\n" >> "$MD_LOG"; exit 0\n' > /tmp/h.sh && chmod +x /tmp/h.sh
+echo '{"hooks":{"MessageDisplay":[{"hooks":[{"type":"command","command":"MD_LOG=/tmp/deltas.jsonl /tmp/h.sh"}]}]}}' > /tmp/s.json
+
+# plain interactive claude in tmux (prefix NOT ocp-tui-*, and never kill-server)
+tmux new-session -d -s md-probe -x 220 -y 50 \
+  "claude --model claude-sonnet-4-6 --effort low --session-id $(uuidgen) --settings /tmp/s.json"
+# …wait for '? for shortcuts', paste a markdown-producing prompt, press Enter…
+
+jq -r '"\(.index) \(.final) \(.delta|@json)"' /tmp/deltas.jsonl   # incremental raw-markdown deltas
+# then assert: concat(deltas) == extractLatestAssistantText(<transcript>.jsonl)
+```
+
+---
+
+## The three dead ends (still worth knowing — they say what NOT to build)
+
+### (a) Incremental transcript reads — **dead: event granularity, not token granularity**
+
+The transcript JSONL *does* grow during a turn, but one **whole event at a time**; the assistant's text
+event is written as **one complete line**, appearing only ~0.3 s before the terminal `turn_duration`.
+
+Observed (session `efd5b161`, `turn_duration: 7319 ms`):
+
+```
+ #6  t+0.0s   type=user       (the prompt)
+ #15 t+4.7s   type=assistant  blocks=thinking
+ #16 t+7.0s   type=assistant  blocks=text      ← the ENTIRE answer, in one line
+ #21 t+7.3s   type=system     subtype=turn_duration   ← terminal
+```
+
+Cross-checked at **20 ms polling + `fs.watch`** (25× finer): a partial line **never touches disk** —
+one write, `+1` line, carrying the complete answer. Also forced with the undocumented
+`CLAUDE_CODE_INCLUDE_PARTIAL_MESSAGES=1`: still 1 assistant event, 0 partials (interactive mode has no
+stream-json *sink* for it to write to).
+
+**The transcript is still needed** — as the terminal-turn signal, as the authoritative `concat === T`
+check, and as the input to the existing honesty gates (auth-banner detection, `truncated`). It is just
+not the *streaming* source.
+
+### (b) `tmux capture-pane` diffing — **dead: the pane is a RENDERED view, not the text**
+
+The backlog expected to fall back to this, calling it "lossy … (wrapping, scrollback, spinner lines)".
+The loss is far worse than formatting noise: **the pane does not contain the answer's source bytes at
+all.** The TUI *renders* markdown, and `capture-pane -p` strips the ANSI that rendering produced.
+
+Same turn, same lines:
+
+```
+TRANSCRIPT (authoritative T):          PANE (capture-pane -p -J -S -500):
+  '## Semaphore'                         '⏺ Semaphore'                 ← heading marker gone
+  ''                                     ''
+  'A **semaphore** is a synchro…'        '  A semaphore is a synchro…' ← bold markers gone, indented
+```
+
+| token in the answer | in `T` | in the pane's answer region |
+|---|---|---|
+| `## ` (ATX heading) | yes | **no** — rendered as `⏺` |
+| `**` (bold markers) | yes | **no** — rendered to ANSI bold, then stripped by `-p` |
+| ` ```javascript ` (fence + language) | yes | **no** — fence and language tag both gone |
+| `- ` (list item) | yes | yes |
+
+*(A literal `**` does appear elsewhere in the pane — in the **prompt echo**, because the prompt asked
+for bold. Not in the answer.)*
+
+**`capture-pane -e` (keeping the ANSI) does not rescue it — the inverse is provably non-unique.**
+With `T` = ``"## Alpha\n\n**bravo**\n\n```javascript\nlet x=1;\n```"``:
+
+```
+⏺\e[39m \e[1mAlpha\n\n\e[0m  \e[1mbravo\n\n\e[0m  \e[34mlet\e[39m x=\e[32m1\e[39m;
+```
+
+`## Alpha` → **SGR 1 (bold)**. `**bravo**` → **SGR 1 (bold)**. *Identical ANSI* — an H2 and a bold span
+are indistinguishable, never mind `**` vs `__`. The fence and its `javascript` tag are consumed by the
+syntax highlighter into colours; recovering the tag would mean inverting a highlighter, and
+`let x=1;` is valid in several languages.
+
+So `T.startsWith(paneText)` is **false** — raw and indent-stripped, on essentially every markdown
+answer. A proxy streaming pane text would be streaming **something the model did not say**. With
+`MessageDisplay` available there is no reason to go near it.
+
+### (c) `--debug-file` — **dead: it logs stream *timing*, never stream *content***
+
+Worth stating precisely, because a casual check misleads in **both** directions here.
+
+The default log level is `debug`, which **suppresses every `verbose` site**. Raise it and per-chunk
+lines *do* appear, spread across generation:
+
+```bash
+CLAUDE_CODE_DEBUG_LOG_LEVEL=verbose claude --debug-file /tmp/d.log …
+```
+```
+05:51:11.088 [VERBOSE] [shoji-engine] yield stream_event/-     ← 16 of these, mid-turn,
+05:51:11.537 [VERBOSE] [shoji-engine] yield stream_event/-        over ~3.9 s of generation
+05:51:15.192 [DEBUG]   [shoji-engine] turn 1 end (usage in=575 out=255 api=6736ms stop=end_turn resultLen=857)
+```
+
+**But they carry no payload** — the format is `yield <type>/<subtype>`, a bare presence marker. Run with
+no category filter (i.e. all categories) at verbose level: `content_block_delta` = **0**, `text_delta` =
+**0**, `content_block_start` / `message_start` = **0**. The only byte-exact text in the log is the
+end-of-turn `Stop` hook payload (`"last_assistant_message":"## Title\n\n**alpha bravo charlie**"`) —
+transcript granularity. The log tells you **when** tokens arrive, never **what** they are. It is also
+~2.7 MB per turn.
+
+### Also checked, also not the answer
+
+| candidate | outcome |
+|---|---|
+| `--output-format stream-json` (the one interface that emits `text_delta`) | **requires `--print`/`-p`** → `cc_entrypoint=sdk-cli` → the **metered** credit pool, which is exactly what TUI mode exists to avoid. Reproduced live. |
+| `--input-format stream-json` | `Error: --input-format=stream-json requires output-format=stream-json` → same gate. |
+| `CLAUDE_CODE_INCLUDE_PARTIAL_MESSAGES=1` (undocumented) | No stream-json sink in interactive mode → no partials. Banner stayed `· Claude Max`. |
+| `sessionMirror` (undocumented) | Gated on `outputFormat === "stream-json"` → the `-p` family. |
+| `--sdk-url` (hidden) | Forces stream-json + non-interactive → `sdk-cli`. *(inferred from the minified bundle; not banner-tested)* |
+| `~/.claude/sessions/<pid>.json` | Registry metadata only (`{pid, sessionId, cwd, status, version, entrypoint:"cli", kind:"interactive"}`). No assistant text. *(Its `entrypoint:"cli"` incidentally confirms the TUI path stays on the subscription pool.)* |
+| `~/.claude/history.jsonl` | User prompts only; the answer text is absent. |
+| Asking the model to emit plain text (so the pane renders faithfully) | Would mean **mutating the caller's prompt** — a correctness violation for a proxy, and still not byte-faithful (wrapping + indent remain). Rejected. |
+
+---
+
+## Value: what streaming actually buys (read before building)
+
+Streaming is *possible*. Whether it is *worth it* depends on the consumer, and the honest answer is
+uncomfortable:
+
+- **Streaming never makes the answer arrive sooner. It moves the *first* byte, not the *last*.** The
+  final token lands at the same wall-clock moment either way.
+- So a consumer that must have the **complete** answer before it can act — e.g. one parsing a structured
+  JSON reply, **which is exactly the 知音 AI use case that motivated this entire investigation** — gains
+  **nothing at all**. Only a **progressively-rendering** consumer (a chat UI) gains.
+
+And the number the backlog attached to this item was wrong:
+
+- The backlog's "~20 s" was inferred from an external 30–32 s report, **never measured through OCP**.
+  Measured through a real OCP instance (TUI mode, `claude-sonnet-4-6`, ~1850-token prompt, n=5):
+  **median 11.30 s** before [#156](https://github.com/dtzp555-max/ocp/pull/156), **9.55 s** after.
+- **Same-turn decomposition** (baseline row `i=5`): **11.563 s** wall through OCP vs `turn_duration:
+  7.319 s` of CLI-internal time on that same turn → **OCP's own overhead ≈ 4.2 s** (n=1, baseline
+  `effort=high` config). *Caveats*: n=1; and `turn_duration` is the CLI's internal duration of an
+  **OCP-driven** turn, not a separate "native" baseline. Do **not** subtract this `effort=high` 7.3 s
+  from the `effort=low` 9.55 s median — a low-effort turn generates faster, so mixing them
+  *understates* the overhead.
+- So OCP's own overhead is **single-digit seconds**, not ~20 s. The rest of any large number is the
+  model generating a long answer — which streaming hides but does not shorten.
+
+**Recommendation**: build it — the contract is clean and the cost is small — but size the expectation
+honestly. It is a *perceived-latency* feature for progressively-rendering consumers, not a throughput
+win, and it does not move the **~6 s TTFT floor** ([`README.md`](README.md)) that rules TUI mode out for
+interactive-latency consumers regardless.

docs/runbooks/615-canary.md

@@ -8,7 +8,7 @@ The billing classifier reading `cli` is **necessary but NOT sufficient** proof.
 
 ## Prerequisites
 
-- `CLAUDE_TUI_MODE=true` already set and OCP restarted (see [TUI-mode setup in README](../../README.md#enabling-tui-mode-opt-in))
+- `CLAUDE_TUI_MODE=true` already set and OCP restarted (see [TUI-mode setup](../tui-mode.md#enabling-tui-mode-opt-in))
 - `tmux` installed on the host
 - No other OCP traffic during the canary (quiesce — see below)
 - Access to your Anthropic account billing page (manual step — see below)
@@ -148,4 +148,4 @@ Run this after any major `claude` CLI upgrade. The `auto` mode lets the CLI's ow
 
 - [Flip/rollback runbook](./tui-flip-rollback.md) — how to set and unset `CLAUDE_TUI_MODE` on systemd and launchd hosts
 - [ADR 0007](../adr/0007-tui-interactive-mode.md) — TUI-mode architecture and governing rules
-- README § [Subscription-pool (TUI) mode](../../README.md#subscription-pool-tui-mode)
+- [Subscription-pool (TUI) mode](../tui-mode.md#subscription-pool-tui-mode)

docs/runbooks/tui-flip-rollback.md

@@ -176,5 +176,5 @@ If you want to continue using OCP without TUI-mode after 2026-06-15, budget for
 
 - [615-canary runbook](./615-canary.md) — how to verify billing pool routing after a flip
 - [ADR 0007](../adr/0007-tui-interactive-mode.md) — TUI-mode architecture; Kill-switch section
-- README § [Subscription-pool (TUI) mode](../../README.md#subscription-pool-tui-mode)
+- [Subscription-pool (TUI) mode](../tui-mode.md#subscription-pool-tui-mode)
 - README § [Environment Variables](../../README.md#environment-variables) — `CLAUDE_TUI_MODE`, `OCP_TUI_ALLOW_LAN=1`

docs/troubleshooting.md

@@ -0,0 +1,152 @@
+Part of [OCP](../README.md) — full troubleshooting manual. The README keeps a slim version with the most common issues and the one-time bootstrap quirks; everything else lives here.
+
+# Troubleshooting
+
+The simplest path: ask your AI.
+
+  Paste this prompt:
+
+  ```
+  Run `ocp doctor` and follow its `next_action`. Tell me if you hit
+  anything that needs human input.
+  ```
+
+The doctor produces a JSON `next_action` with `ai_executable[]` (commands
+the agent runs verbatim) and `human_required[]` (steps that need you,
+typically just OAuth).
+
+## Manual debugging
+
+### Setup fails with "claude: command not found"
+
+`setup.mjs` requires the Claude CLI to be on `PATH`. Install it via the [official guide](https://docs.anthropic.com/en/docs/claude-cli), confirm with `which claude`, then run `claude auth login` before re-running `node setup.mjs`.
+
+### Setup fails with "EADDRINUSE: port 3456 already in use"
+
+Something else is already bound to port 3456 — usually an old OCP instance. Check what:
+
+```bash
+lsof -nP -iTCP:3456 -sTCP:LISTEN
+```
+
+If it's an old OCP process, stop it before re-running setup:
+
+```bash
+launchctl bootout gui/$(id -u)/dev.ocp.proxy            # macOS launchd
+systemctl --user stop ocp-proxy                         # Linux systemd (installed as a --user unit)
+```
+
+(There is no `ocp stop` subcommand — the proxy runs as a service, so stopping it goes through the service manager above. `ocp restart` exists for the bounce case.)
+
+### Setup fails with "node: command not found" or version error
+
+OCP requires Node.js 22.5+. Install:
+
+```bash
+brew install node          # macOS
+# Linux: see https://nodejs.org/en/download for current install commands
+```
+
+Confirm with `node --version` (should be ≥ v22.5).
+
+### Requests fail or agents stuck
+
+```bash
+# Clear sessions and restart
+ocp clear
+ocp restart
+
+# If using OpenClaw gateway
+openclaw gateway restart
+```
+
+### Env var change (e.g. `CLAUDE_BIND`, `CLAUDE_CODE_OAUTH_TOKEN`) doesn't take effect after restart
+
+On **macOS**, `ocp restart` does a full `launchctl bootout` + `bootstrap` of the agent, which **re-reads the plist `EnvironmentVariables`** — so an env change you made (in `~/Library/LaunchAgents/dev.ocp.proxy.plist`) actually takes effect:
+
+```bash
+ocp restart
+```
+
+This is deliberate: the older `launchctl kickstart -k` only re-execs the process and **reuses launchd's cached environment**, so plist env edits would be silently ignored. If you ever restart the agent by hand, use bootout+bootstrap, not `kickstart -k`:
+
+```bash
+launchctl bootout   gui/$(id -u)/dev.ocp.proxy 2>/dev/null
+launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/dev.ocp.proxy.plist
+```
+
+Verify the new value reached the running process:
+
+```bash
+ps -E -p "$(launchctl print gui/$(id -u)/dev.ocp.proxy 2>/dev/null | awk '/pid =/{print $3}')" | tr ' ' '\n' | grep CLAUDE_
+```
+
+On **Linux**, `systemctl --user restart` already re-reads the unit's `EnvironmentFile`, so no special handling is needed.
+
+### Usage shows "unknown"
+
+Usually caused by an expired Claude CLI session. Fix:
+```bash
+claude auth login
+ocp restart
+```
+
+### Startup log warns "OpenClaw registry out of sync"
+
+On boot, OCP compares OpenClaw's registered models against [`models.json`](../models.json) and warns if they drift. Cause: someone (or an OpenClaw upgrade) modified `~/.openclaw/openclaw.json` and removed entries OCP expects. Fix:
+
+```bash
+node ~/ocp/scripts/sync-openclaw.mjs
+```
+
+This is read-only at startup; the warning never blocks the gateway from running.
+
+### A TUI session vanished right after upgrading OCP
+
+If you ran a pre-3.21.1 OCP instance and a post-3.21.1 instance on the same host at the same time during an upgrade, the new instance's one-time boot reap can, once, kill an old-format (`ocp-tui-<8hex>`) live TUI session belonging to the still-running old instance — restart the affected session (`ocp restart` or re-run your TUI turn) and it will come back under the new instance's port-scoped naming.
+
+### OpenClaw shows old models after `ocp update` (v3.10→v3.11 only)
+
+One-time bootstrap quirk for the v3.10.0 → v3.11.0 jump only — the running shell had the old `cmd_update` cached. Run once manually:
+
+```bash
+node ~/ocp/scripts/sync-openclaw.mjs
+openclaw gateway restart   # so OpenClaw re-reads the config
+```
+
+Future `ocp update` invocations sync automatically.
+
+<a id="cache-rekey-v3250"></a>
+### Response-cache hit rate drops once after upgrading to v3.25.0
+
+Only affects instances running with the response cache **on** (`CLAUDE_CACHE_TTL > 0`); it is off by default, so most installs see nothing.
+
+v3.25.0 keys the cache on the **resolved** model rather than on the string the client sent, so rows written for an alias (`opus`, `sonnet`, `haiku`, or a legacy alias like `claude-haiku-4-5`) no longer match. Those rows orphan and are reaped by the TTL cleanup interval within one window — **no migration script, no action required**; expect one window of extra misses and then normal hit rates.
+
+Two different scopes, worth being precise about:
+
+- **Normal cache** — only *alias-addressed* rows rekey. Rows written under a literal model id (`claude-sonnet-5`) keep matching — **unless the instance runs `OCP_LOCAL_TOOLS=1`**, in which case the entire normal cache rekeys once. That is a separate mechanism: v3.25.0 also reworded the local-tools wrapper, and the wrapper text is one of the four inputs to `CONFIG_EPOCH`, which every normal cache key folds in (established behavior since v3.23.0, not new here).
+- **Structured-output cache** — **every** row rekeys, alias or literal. The same change also folds the config epoch into the structured key, which it had never included; that gap meant a `CLAUDE_SYSTEM_PROMPT` change did not invalidate structured answers either. Structured caching only exists from v3.24.0, so there is at most one release worth of rows to orphan.
+
+This is deliberate, and it is what makes an alias repoint take effect. Before v3.25.0, changing where an alias pointed (v3.25.0 itself repoints `opus` → `claude-opus-5`) left the cache serving the **old** model's answers under that alias until TTL expiry, because `models.json` is read once at boot while the SQLite cache survives the restart. If you were running with the cache on and repointed an alias in an earlier version, that is why it appeared not to take.
+
+A side effect worth knowing: an alias and its canonical id now **share** a cache slot, since both produce an identical spawn. That is a small hit-rate improvement in steady state.
+
+<a id="tui-401"></a>
+### TUI-mode returns a permanent `Please run /login` 401 (re-login doesn't stick)
+
+A long-running TUI-mode host can get stuck returning a permanent 401 (`Please run /login · API Error: 401`) that re-login cannot fix.
+
+**Root cause (two layers):** interactive `claude` **prefers `~/.claude/.credentials.json` over the `CLAUDE_CODE_OAUTH_TOKEN` env var** (this is *unlike* the `-p` path, where the env token wins). So (a) a stale/corrupt `credentials.json` **shadows** the env token — passing the token is not enough on its own; and (b) when claude does use `credentials.json`, its single-use OAuth refresh token can be corrupted (ending up an empty string) by the per-request spawn + `kill-session` teardown racing claude's token rotation. Re-login writes a fresh token, but the next spawn re-corrupts it. Proven live on PI231: *env token passed + broken `credentials.json` present → 401; env token passed + `credentials.json` moved aside → works.*
+
+**Fix:** set `CLAUDE_CODE_OAUTH_TOKEN` on the OCP host and leave `OCP_TUI_HOME` **unset**. OCP then runs the TUI `claude` in a **credential-isolated home** (`$HOME/.ocp-tui/home`) that has **no `credentials.json`** at all, so the env token is the only credential (authoritative — nothing shadows it) and claude never runs the refresh path (so the single-use token can't be corrupted). Then restart — on systemd `daemon-reload`, on launchd `bootout`+`bootstrap`; `kickstart -k` does **not** reload env. Verify the env reached the process and the boot log shows the isolated home:
+
+```bash
+# Linux (systemd): confirm the token is in the service env
+tr '\0' '\n' < /proc/$(pgrep -f server.mjs | head -1)/environ | grep CLAUDE_CODE_OAUTH_TOKEN
+# Boot log should read: TUI-mode: ON home=$HOME/.ocp-tui/home ... auth=env-token (credential-isolated home — no credentials.json)
+```
+
+> If you previously set `OCP_TUI_HOME` to the real home (or any home that contains a `credentials.json`), **unset it** so the credential-isolated default takes effect — otherwise the shadowing `credentials.json` remains in play.
+
+See [Subscription-pool (TUI) mode](tui-mode.md#subscription-pool-tui-mode) and ADR 0007 PR-C / PR-D amendments.

docs/tui-mode.md

@@ -0,0 +1,196 @@
+Part of [OCP](../README.md) — subscription-pool (TUI) mode: serve requests through interactive `claude` so they bill the Pro/Max subscription pool instead of the metered Agent SDK path.
+
+# Subscription-pool (TUI) mode
+
+> **SECURITY — read before enabling.**  
+> TUI-mode is **single-user / single-operator only**. `claude` runs with the OCP process owner's filesystem access regardless of `HOME` setting. If OCP serves multiple users or guest API keys, a guest prompt could exfiltrate files or exhaust the subscription. **Never enable `CLAUDE_TUI_MODE=true` on a multi-user OCP.**
+
+## What it is and why
+
+> **⚠️ Status (as of 2026-07): the billing split below is PAUSED.** Anthropic announced it for 2026-06-15, then paused it on the effective date — *"For now, nothing has changed: Claude Agent SDK, `claude -p`, and third-party app usage still draw from your subscription's usage limits"* ([official help article](https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan)). While the pause holds, OCP's default `-p` path bills the subscription and **TUI-mode is a hedge, not a necessity**. The table describes the *announced* regime, kept here because Anthropic says a reworked change will return (with advance notice) — everything in this section is ready to flip on that day.
+
+The announced routing keys `claude` invocations by `cc_entrypoint`:
+
+| Launch method | `cc_entrypoint` | Billing pool (announced regime, currently paused) |
+|---------------|-----------------|-------------|
+| `claude -p` / `--output-format` (OCP default) | `sdk-cli` | Agent SDK credit pool (~$20/mo on Pro) |
+| Interactive `claude` (no flags) | `cli` | Pro/Max subscription pool |
+
+TUI-mode lets OCP serve requests via the interactive path so they bill against the subscription pool under that regime. The response is read from claude's native JSONL session transcript once the turn is complete, then replayed to the caller as a normal OpenAI completion or chunked SSE response.
+
+<a id="tui-entrypoint"></a>
+## Billing-classifier labeling (`OCP_TUI_ENTRYPOINT`)
+
+`OCP_TUI_ENTRYPOINT` (default `cli`) controls how `CLAUDE_CODE_ENTRYPOINT` is set on the spawn
+environment. The default (`cli`) pins the value deterministically — immune to a stray inherited
+env var or a future stdout-redirect bug silently flipping it to `sdk-cli`. This label is honest
+**only** when the spawn is a genuine interactive PTY (tmux pane, no `-p`, stdout not redirected,
+and `tmux new-session` verified to succeed). If you need to observe the raw TTY-derived value, set
+`OCP_TUI_ENTRYPOINT=auto`. See ADR 0007 for the full rationale and governing rule.
+
+## Enabling TUI-mode (opt-in)
+
+```bash
+# Prerequisites
+mkdir -p ~/.ocp-tui/work    # one-time scratch cwd setup
+# tmux must be installed: brew install tmux  /  apt install tmux
+
+# Enable
+export CLAUDE_TUI_MODE=true
+# STRONGLY RECOMMENDED on a TUI host — authenticate via the long-lived OAuth token.
+# With this set (and OCP_TUI_HOME left UNSET), OCP runs the interactive claude in a
+# credential-isolated home ($HOME/.ocp-tui/home, no credentials.json), so the env token
+# is the only credential and is authoritative. This both stops a stale credentials.json
+# from shadowing the token AND ends the refresh-token corruption that caused a permanent
+# "Please run /login" 401 (no credentials file → claude never runs the refresh path).
+# See the auth note below + ADR 0007 PR-D.
+export CLAUDE_CODE_OAUTH_TOKEN=sk-ant-oat01-...
+# Optionally tune:
+export CLAUDE_TUI_WALLCLOCK_MS=180000   # 3 min cap for long Opus turns
+export OCP_TUI_CWD=$HOME/.ocp-tui/work  # default; override if needed
+export OCP_TUI_ENTRYPOINT=cli           # default; use 'auto' to observe TTY-derived value
+# Do NOT set OCP_TUI_HOME for the recommended setup — leaving it unset is what enables
+# the credential-isolated home. Set it only to opt into the legacy symlinked-creds mode.
+```
+
+Then restart OCP. At boot you will see (with the env token set, isolated home auto-selected):
+
+```
+⚠️  TUI-mode ON — single-user only; do NOT enable on a multi-user OCP ...
+  TUI-mode: ON home=/home/user/.ocp-tui/home cwd=/home/user/.ocp-tui/work auth=env-token (credential-isolated home — no credentials.json) wallclock=120000ms maxConcurrent=2
+```
+
+## What changes / what doesn't
+
+- **Callers see no API change.** The response is a normal OpenAI completion object or chunked SSE — identical wire format.
+- **Real streaming is opt-in (`OCP_TUI_STREAM=1`), and off by default.** By default TUI-mode buffers the full response and replays it as chunked SSE — you see a delay, then the complete response. Set `OCP_TUI_STREAM=1` and `stream:true` turns emit real SSE `delta.content` chunks as `claude` renders them, sourced from `claude`'s own `MessageDisplay` hook (byte-faithful raw markdown, on the subscription pool, no `-p`). Two honest caveats: granularity is **block-level** — the hook fires once per rendered block, so a handful of chunks per answer, scaling with length, not token-by-token; and it moves the **first** byte, not the last, so a consumer that must parse a complete reply gains nothing. The transcript stays authoritative: every streamed turn is asserted against it at the end, and a turn whose stream disagrees is **failed rather than served** (watch `tui.streamDivergences` on `/health`). Evidence: [`plans/2026-07-13-tui-latency/streaming-spike.md`](plans/2026-07-13-tui-latency/streaming-spike.md).
+- **Cache and singleflight work normally.** TUI-mode writes the buffered response to the cache on success; cache-hits skip the interactive turn entirely.
+- **The host's `CLAUDE.md` / auto-memory is never injected.** OCP is a proxy — the proxied client (OpenClaw / your IDE) owns its own context and memory. TUI-mode runs `claude` with `--safe-mode`, which disables all host customizations (CLAUDE.md, skills, plugins, hooks, MCP servers) while leaving auth, model selection, built-in tools, and permissions untouched — so a `CLAUDE.md` on the OCP host can never leak into a proxied turn and steer the answer, and the session still bills the subscription pool (`cc_entrypoint=cli`, unlike `--bare`). The `CLAUDE_CODE_DISABLE_CLAUDE_MDS` / `CLAUDE_CODE_DISABLE_AUTO_MEMORY` env vars are kept as a fallback (see #4). Built-in tool schemas + the interactive system prompt remain (the inherent ~20–35K context floor of interactive mode); MCP is hard-disabled. **Exception:** the two opt-in modes that deliberately need a customization `--safe-mode` would strip — real streaming (`OCP_TUI_STREAM`, which registers a `MessageDisplay` **hook** via `--settings`) and `OCP_TUI_FULL_TOOLS` (which grants an **MCP** / skills surface) — run without `--safe-mode` and rely on the env-var suppression alone.
+- **Authenticate via `CLAUDE_CODE_OAUTH_TOKEN` in a credential-isolated home (recommended).** tmux does not forward the parent process's env to the pane, so OCP sets the token explicitly on the spawned `claude` when `CLAUDE_CODE_OAUTH_TOKEN` is present. With the env token set and `OCP_TUI_HOME` unset, OCP runs claude in a **credential-isolated home** (`$HOME/.ocp-tui/home`) that has **no `credentials.json`** — so the env token is the only credential and is authoritative, and claude never runs the token-refresh path. This both stops a stale `credentials.json` from shadowing the token and ends the refresh-token corruption behind the permanent `Please run /login · API Error: 401` (full two-layer root cause, live proof, and fix in [Troubleshooting § the permanent TUI-mode 401](troubleshooting.md#tui-401)). Transcripts land under the same isolated home, so the answer-reader is unaffected. Without the env token, claude falls back to the real home's `credentials.json` (byte-for-byte the previous behaviour). (The token is visible in `ps` on the pane command — acceptable for the single-user A-path; the multi-user B-path is refused at boot.) See ADR 0007 PR-C / PR-D amendments.
+- **Stale tmux sessions are reaped.** The pane's `claude` is a child of the tmux server (not OCP), so OCP cannot reap it directly; `claude` zombies can otherwise accumulate as `<defunct>` over a long-running host. OCP reaps them at boot and on a 15-min idle sweep by issuing `tmux kill-server` — but **only when no foreign tmux session remains** (it never disrupts a co-hosted `olp-tui-*` instance). See ADR 0007 PR-C amendment.
+- **Default path unchanged.** Unset `CLAUDE_TUI_MODE` and restart → `callClaude` / `callClaudeStreaming` are used again, byte-for-byte identical to today.
+- **Concurrency is bounded separately.** TUI turns are heavy (per-request cold-boot + long wallclock), so the TUI path has its own limiter — `OCP_TUI_MAX_CONCURRENT` (default `2`), independent of `CLAUDE_MAX_CONCURRENT`. Excess turns queue; a full queue returns a 503. Tune it up only on a host that can run more interactive `claude` sessions at once.
+- **Optional warm pane pool (`OCP_TUI_POOL_SIZE`, default off).** Pre-boots panes so a request skips the cold boot — measured p50 `10.17s` → `6.00s` (−41%). Pooled panes are **single-use** (one turn, then killed and replaced in the background), each carrying its own fresh `--session-id`, so one session still means one exchange and no earlier-turn text can leak into a later answer. They are named `ocp-tui-<port>-p<hex>` and coexist with the reaper by design: the sweep **drains the pool first**, then reaps (so `kill-server` still flushes `<defunct>` zombies), then the pool refills in the background. Drain→reap→resume is synchronous, so no request can land mid-sweep; a request arriving while the pool is still re-booting simply misses it and cold-boots. A live pooled pane is never reaped — **including one that is still booting**, whose tmux session already exists — while an *orphaned* one (left by a previous process generation) still is.
+
+## ⚠️ Latency: TUI mode has a ~6-second floor, and it is immovable
+
+**TUI mode cannot serve real-time or interactive-latency consumers.** This is a hard property of the
+path, stated plainly so you can rule it out before building on it:
+
+| | measured |
+|---|---|
+| **TTFT floor (first token)** | **≈ 6 s** — immovable |
+| cold boot → input bar ready | ~1 s (per request; not the bottleneck) |
+| OCP's own overhead above the CLI | ~4 s (n=1 same-turn decomposition) |
+| direct Anthropic API, same prompt (for scale) | 0.84–1.64 s |
+
+The ~6 s floor is the `claude` CLI itself: it always injects the full Claude Code system prompt plus
+its tool definitions before your prompt, on every turn, no matter what you ask. No flag removes it
+(`--exclude-dynamic-system-prompt-sections` was measured: **no effect** on the floor). Extended
+thinking is *not* the cause — `OCP_TUI_EFFORT` already defaults to `low`, which is what cuts a
+formerly-inherited `xhigh` down to this floor and collapses its variance.
+
+On top of the floor you pay the model's generation time (a function of output length). Progressive
+output is not wired up **yet** (see "No real token streaming" above — it is achievable and planned),
+so today a turn returns as one blob once generation completes. Note that streaming, when it lands,
+will move the *first* byte earlier — it does **not** shorten the turn, and a consumer that needs the
+complete answer gains nothing from it.
+
+**Use TUI mode for**: batch, background, and latency-insensitive work where the subscription pool is
+the point. **Do not use it for**: anything a person is waiting on interactively, or any consumer with
+a sub-5-second budget. Full measurements and methodology:
+[`plans/2026-07-13-tui-latency/`](plans/2026-07-13-tui-latency/).
+
+## Monitoring drift via `/health`
+
+`GET /health` includes a `tui` block so you can poll for a silent billing-pool drift (the top risk under the announced split, if it re-lands — a lost TTY flipping `cc_entrypoint` from `cli` to `sdk-cli` would still return answers but land in the metered pool). The block is **always present** (with `enabled:false` when TUI-mode is off):
+
+```jsonc
+"tui": {
+  "enabled": true,             // CLAUDE_TUI_MODE === "true"
+  "entrypointMode": "cli",     // OCP_TUI_ENTRYPOINT (cli | auto | off)
+  "lastEntrypoint": "cli",     // last cc_entrypoint observed in a transcript, or null
+  "entrypointMismatches": 0,   // count of cli-expected-but-got-other turns — ALERT if this climbs
+  "inflight": 1,               // TUI turns running right now
+  "queued": 0,                 // TUI turns waiting for a concurrency slot
+  "maxConcurrent": 2,          // OCP_TUI_MAX_CONCURRENT
+  "pool": {                    // warm pane pool — null when OCP_TUI_POOL_SIZE=0 (the default)
+    "size": 2,                 // target warm panes (OCP_TUI_POOL_SIZE)
+    "warm": 2,                 // panes ready right now — each is a LIVE idle claude process
+    "booting": 0,              // replacement panes currently pre-booting
+    "model": "claude-sonnet-4-6", // the model being warmed (the most recently requested one)
+    "hits": 12,                // requests served by a warm pane
+    "misses": 1,               // requests that fell back to the cold boot (the 1st is always one)
+    "boots": 14,               // panes successfully pre-booted
+    "bootFailures": 0,         // pre-boots that genuinely never reached the input bar — WATCH this
+    "cancelled": 4,            // in-flight boots OCP killed on purpose (drain / model switch) — not faults
+    "dropped": 8               // panes discarded unused (drain sweep / expired / unhealthy)
+  }
+}
+```
+
+Alert on `entrypointMismatches > 0` (or `lastEntrypoint !== "cli"`): it means a turn drew from the metered Agent SDK pool instead of the subscription. `inflight` / `queued` show how close the TUI path is to its concurrency cap.
+
+With the pool on, `hits` / `misses` is the hit rate (a steady single-model consumer should sit near 100% after the first request), and `warm` is your standing idle-process cost. A climbing `bootFailures` means panes are not reaching their input bar — the pool then degrades safely to the cold path, but latency reverts to the un-pooled numbers. `cancelled` counts boots OCP killed *on purpose* (a drain, a model switch) and is **not** a fault signal — do not alert on it. A steadily climbing `dropped` is likewise normal: the 15-min reap sweep drains and re-boots the pool on every tick so `kill-server` can still flush `<defunct>` zombies.
+
+## Kill-switch
+
+```bash
+unset CLAUDE_TUI_MODE
+# restart OCP
+```
+
+The stream-json path is restored immediately. No other change is needed.
+
+## Operator checklist for the (paused) billing split
+
+> **Status:** the 2026-06-15 split never took effect — Anthropic paused it on the effective date (see the status note at the top of this section). **Nothing needs flipping while the pause holds.** The checklist is retained verbatim as the runbook for if/when a reworked change lands (Anthropic has promised advance notice).
+
+Under the announced regime, every host serving traffic must be flipped to TUI-mode **and** canary-verified before the effective date, or it will bill the metered Agent SDK credit pool instead of the subscription.
+
+- **[Flip/rollback runbook](runbooks/tui-flip-rollback.md)** — how to set `CLAUDE_TUI_MODE=true` on systemd (Linux) and launchd (macOS) hosts. Covers the `daemon-reload` requirement (systemd) and the `bootout`+`bootstrap` cycle requirement (launchd — `launchctl kickstart -k` does not reload plist env).
+- **[615-canary runbook](runbooks/615-canary.md)** — after each flip, run one quiesced request and compare the Agent SDK credit balance before and after. `entrypoint:cli` in the transcript (the `cc_entrypoint` billing classifier) is necessary but not sufficient — only a stable credit balance confirms the subscription pool is being used. Balance check is a manual step (no known programmatic API for the Agent SDK credit pool balance).
+
+## Architecture and design decisions
+
+See [`adr/0007-tui-interactive-mode.md`](adr/0007-tui-interactive-mode.md) for the full rationale, home-strategy options, MCP-disable mechanism, coexistence rules, and the B-path (multi-tenant isolation) roadmap.
+
+## TUI-mode environment variables
+
+The README [Environment Variables](../README.md#environment-variables) table lists these as one-line pointers; the full behaviour of each lives here.
+
+<a id="ocp-tui-stream"></a>
+### `OCP_TUI_STREAM` — real SSE streaming (opt-in)
+
+`OCP_TUI_STREAM` default `0` (off). When `=1`, `stream:true` requests emit **real SSE `delta.content` chunks as `claude` generates them**, instead of buffering the turn and replaying it. Deltas come from `claude`'s own `MessageDisplay` hook (registered with `--settings` on the ordinary interactive spawn — banner-verified to stay on the subscription pool, `· Claude Max`). Granularity is **block-level**, not token-level. The transcript remains authoritative: the streamed text is asserted equal to it at end-of-turn, the auth-banner and truncation gates still run before anything is committed, and only the transcript text is cached. A turn whose stream cannot be reconciled with the transcript is **refused** (SSE error frame, not cached) and counted as `tui.streamDivergences` on `/health`. A total hook failure (e.g. `--settings` stops registering it after a `claude` version bump) is a *different, silent* failure mode — every streamed turn still succeeds, fully buffered, with no divergence and no error — so it is counted separately as `tui.streamZeroDeltaTurns` (streamed turns where the hook fired **zero** times) and logged as `tui_stream_zero_deltas`; watch it alongside `streamDivergences`. Default off — the buffered path is unchanged and remains the stable default. ⚠️ **Tool-using turns:** the transcript keeps only the model's **last** assistant message, so if the model narrates before calling a tool ("I'll check that file…") and that narration exceeds `OCP_TUI_STREAM_HOLDBACK`, it has already been streamed and cannot be retracted — the turn is then **refused** rather than served (measured live: Opus narrated 475 chars before a `Bash` call). If your deployment lets the model use tools (the TUI default, and anything with `OCP_TUI_FULL_TOOLS=1`), either raise `OCP_TUI_STREAM_HOLDBACK` above the typical narration length — the narration then stays held back and is correctly discarded, at the cost of a later first chunk — or leave streaming off. Streaming is best suited to tool-light chat proxying. See ADR 0007 (2026-07-13 amendment).
+
+Two related streaming knobs:
+
+- **`OCP_TUI_STREAM_DIR`** (default `$HOME/.ocp-tui/stream`) — directory holding the static `MessageDisplay` hook script + settings file, and the per-session delta sink (`<session-id>.jsonl`, removed at turn teardown). One sink **per session-id** — this is what keeps concurrent TUI turns (`OCP_TUI_MAX_CONCURRENT` ≥ 2) from interleaving one client's deltas into another's stream.
+- **`OCP_TUI_STREAM_POLL_MS`** (default `100`) — interval at which OCP drains the delta sink. The hook fires at block granularity (seconds apart), so a finer poll buys nothing.
+
+<a id="ocp-tui-stream-holdback"></a>
+### `OCP_TUI_STREAM_HOLDBACK`
+
+`OCP_TUI_STREAM_HOLDBACK` default `100`. (TUI-mode, streaming) Characters withheld before the first chunk reaches the client. Two jobs. (1) It keeps the **auth-banner gate** alive under streaming, via a guarantee with two required halves: (i) nothing is emitted for a message until its trimmed accumulation exceeds 100 chars — past the default banner detector's reach, since real banners are ≤100 chars — and (ii) once a message boundary follows an emit, nothing further is ever emitted for the rest of the turn, and the turn is refused outright. Half (i) alone only covers a turn's first message; half (ii) is what covers an error banner rendered as a *later* message (e.g. after tool-using prose). Raise the holdback if you replace the detector via `CLAUDE_TUI_ERROR_PATTERNS` with patterns that can match longer messages — that only affects half (i); OCP warns at boot if you do. (2) It is the knob for **tool-using turns** — see the `OCP_TUI_STREAM` caveat above. Answers shorter than the holdback are simply delivered whole at end-of-turn, exactly as the buffered path does.
+
+<a id="ocp-tui-pool-size"></a>
+### `OCP_TUI_POOL_SIZE` — warm pane pool
+
+`OCP_TUI_POOL_SIZE` default `0` (off). Number of **pre-booted warm `claude` panes** kept ready, so a request does not pay the cold boot. `0` disables the pool entirely — the request path is then exactly the cold-boot path. Max `4`; an unparseable value disables it rather than guessing. **Measured on a Mac mini (Sonnet 4.6, `--effort low`): end-to-end p50 `10.17s` (n=6, pool off) → `6.00s` (n=12 warm hits) — −4.2 s / −41%** — the pool recovers both the ~1.2 s boot *and* ~2.9 s of post-input-bar init that a pane which has been idle a moment has already finished. **Cost:** each warm pane is a *live idle `claude` process* held whether or not a request ever arrives (peak processes ≈ pool size + `OCP_TUI_MAX_CONCURRENT` + 1 booting replacement) — which is why it is opt-in. Panes are **single-use**: one turn, then killed and replaced in the background. The **first request after start (and after any model switch) is always a cold miss** — the pool warms the most recently requested model, since OCP cannot know which model the next caller wants. See [`plans/2026-07-13-tui-latency/`](plans/2026-07-13-tui-latency/).
+
+<a id="ocp-tui-full-tools"></a>
+### `OCP_TUI_FULL_TOOLS` — full tool surface (single-user only)
+
+`OCP_TUI_FULL_TOOLS` default *(unset)*. (TUI-mode, **single-user only**) When `=1`, grant the interactive session the **same tool surface as the `-p` path** — `--allowedTools` (+ optional `--mcp-config`, read from `CLAUDE_ALLOWED_TOOLS` / `CLAUDE_MCP_CONFIG`) — instead of the default MCP-walled, built-in-tools-only set. Lets a trusted single-operator TUI deployment run a **tool-using / MCP agent** (e.g. an OpenClaw assistant) on the subscription pool. Safe because TUI **refuses to boot under `AUTH_MODE=multi`** (hard exit) — no guest key can ever reach the TUI path, so this gate cannot expose tools to an untrusted caller. (Under `AUTH_MODE=shared` + `OCP_TUI_ALLOW_LAN=1`, anyone holding the single shared key reaches it — that is the existing TUI trust model, unchanged.) Note: `--dangerously-skip-permissions` / `CLAUDE_SKIP_PERMISSIONS` is **not** supported for TUI — claude v2.1.x shows an interactive bypass-acceptance screen in headless tmux that cannot be answered, bricking the pane. Use scratch-home `settings.json` `additionalDirectories` instead. See ADR 0007.
+
+<a id="tui-other-vars"></a>
+### Other TUI-mode variables
+
+- **`OCP_TUI_MAX_CONCURRENT`** (default `2`) — Max concurrent interactive TUI turns. **Independent** of `CLAUDE_MAX_CONCURRENT` (which bounds the `-p`/stream-json path; TUI never uses it). A TUI turn is heavy (per-request cold-boot of tmux+claude + up to `CLAUDE_TUI_WALLCLOCK_MS` wallclock), so the default is low to keep small hosts (e.g. a Pi 4) alive under a burst. Excess turns **queue** (bounded); a full queue yields a 503. See ADR 0007 PR-B amendment.
+- **`OCP_TUI_ENTRYPOINT`** (default `cli`) — Billing-classifier labeling: `cli` (default) pins `cc_entrypoint=cli` deterministically; `auto` lets claude self-classify via TTY detection; `off` leaves the inherited env untouched. Honest only when the spawn is a genuine interactive PTY — see the "Billing-classifier labeling" section above and ADR 0007.
+- **`OCP_TUI_EFFORT`** (default `low`) — Effort level passed to the interactive `claude` as an explicit `--effort` flag: `low` (default), `medium`, `high`, `xhigh`, `max`, or `inherit` to omit the flag (the pre-flag behaviour: the pane inherits a HOME-dependent effort — the operator's `~/.claude/settings.json` `effortLevel` in real-home mode, claude's built-in default in env-token scratch mode). Explicit `low` cuts measured TTFT p50 by ~40% and collapses run-to-run variance ~15× versus an inherited `xhigh` (see [`plans/2026-07-13-tui-latency/`](plans/2026-07-13-tui-latency/)); proxied requests rarely benefit from extended thinking. Banner-verified to stay on the subscription pool (`· Claude Max`). An invalid value logs a warning and falls back to `low`.
+- **`OCP_TUI_HOME`** (default *(auto)*) — `HOME` claude runs under. **When unset, OCP picks it for you:** if `CLAUDE_CODE_OAUTH_TOKEN` is set → a **credential-isolated** scratch home `$HOME/.ocp-tui/home` (no `credentials.json`, env-token auth — **recommended**); if no env token → the operator's real home (legacy shared `credentials.json`). Setting this to an **explicit** path overrides the auto-default. The credential handling at that path still follows the env token: **with** the env token it is credential-free (env-token auth, no `credentials.json` written); **without** the env token (and the path ≠ real home) it uses the legacy symlinked-credentials scratch mode, which carries the credential-fork caveat — see ADR 0007. If you previously set this to the real home (or any home containing a `credentials.json`) and hit a permanent 401, unset it — see [Troubleshooting § the permanent TUI-mode 401](troubleshooting.md#tui-401).
+- **`CLAUDE_TUI_WALLCLOCK_MS`** (default `120000`) — Maximum time in ms to wait for the native transcript to signal turn completion. Increase for long Opus thinking turns.
+- **`OCP_TUI_CWD`** (default `$HOME/.ocp-tui/work`) — Scratch working directory where interactive claude sessions run. Transcripts land under `<HOME>/.claude/projects/<encoded-cwd>/`. Created automatically.
+- **`CLAUDE_CODE_OAUTH_TOKEN`** — the recommended TUI credential; when set (and `OCP_TUI_HOME` unset) it selects the credential-isolated home. Full precedence and the 401 root cause it prevents are in [Troubleshooting § the permanent TUI-mode 401](troubleshooting.md#tui-401).

docs/upgrading.md

@@ -0,0 +1,79 @@
+Part of [OCP](../README.md) — the full upgrade manual (`ocp update` paths, manual flags, rollback, and OpenClaw auto-sync). The README keeps a short stub with the one-liner.
+
+# Upgrading
+
+The simplest path: ask your AI.
+
+  Paste this prompt:
+
+  ```
+  Upgrade my OCP. Run `ocp update` and follow whatever it says.
+  If it tells me to run `claude auth login`, I'll do that.
+  ```
+
+What `ocp update` does:
+
+- **Patch bump** (e.g. `v3.21.0 → v3.21.1`):
+  light path (git pull + npm install + restart).
+- **Cross-minor** (e.g. `v3.18 → v3.22`):
+  full path: pre-flight check, snapshot, `setup.mjs` (with plist env-merge),
+  service restart, post-flight `/health` and `/v1/models` verification.
+- **Old version** (< v3.4.0):
+  fresh-install. Pre-v3.4 lacked admin-key/usage-db, so there is nothing to
+  migrate. Your OAuth token (managed by the Claude Code CLI, not OCP) is
+  preserved; you do not need to re-OAuth unless your token expired
+  separately.
+
+Snapshots are saved to `~/.ocp/upgrade-snapshot-<ISO-ts>/` and never
+auto-deleted. Clean old ones with `rm -rf ~/.ocp/upgrade-snapshot-*` once
+you're confident the upgrade is stable.
+
+## Manual upgrade — same command, no AI
+
+```bash
+ocp update                  # smart-pick path
+ocp update --check          # show available updates, don't apply
+ocp update --dry-run        # preview plan
+ocp update --target v3.13.0 # pin a specific version
+ocp update --rollback --yes # restore most recent snapshot (--yes confirms)
+ocp update --rollback --list      # list snapshots, no mutation
+ocp update --rollback --dry-run   # preview rollback plan
+```
+
+## When upgrade fails
+
+`ocp update` prints a recovery line on failure. To restore from the snapshot:
+
+```bash
+ocp update --rollback --yes   # --yes confirms the destructive restore
+ocp doctor
+```
+
+If `ocp doctor` still reports problems after rollback, open a GitHub issue
+with the snapshot path and the doctor JSON output (`ocp doctor --json`).
+
+## OpenClaw Auto-Sync (v3.11.0+)
+
+Whenever the model list in [`models.json`](../models.json) changes, `ocp update` automatically reconciles your OpenClaw config so the model dropdown stays in sync — no more "I upgraded OCP but my Telegram bot still shows the old models" surprises.
+
+**What gets synced** (and only this — all other config keys are preserved):
+- `models.providers."claude-local".models` in `~/.openclaw/openclaw.json`
+- `agents.defaults.models["claude-local/*"]` aliases
+
+**Safety**:
+- Timestamped backup written before every change: `~/.openclaw/openclaw.json.bak.<ms>`
+- Idempotent — already-in-sync runs are a no-op (no backup, no rewrite)
+- Non-fatal — sync failure does NOT abort `ocp update`; `/v1/models` still works
+- Skips silently if OpenClaw is not installed (`~/.openclaw/openclaw.json` missing)
+
+**Manual trigger** (e.g. after fixing a hand-edited config, or for the one-time v3.10.0→v3.11.0 bootstrap quirk):
+```bash
+node ~/ocp/scripts/sync-openclaw.mjs
+node ~/ocp/scripts/sync-openclaw.mjs --quiet   # silent unless changes
+```
+
+**Opt-out**: `ocp update` only invokes the sync if `node` and `scripts/sync-openclaw.mjs` are both present. Removing the script disables auto-sync; the rest of `ocp update` still works.
+
+**One-time bootstrap caveat (v3.10.0 → v3.11.0 only)**: the first `ocp update` to v3.11.0 runs the *old* `cmd_update` already loaded into your shell, so the new sync hook does NOT fire on this single jump. Run `node ~/ocp/scripts/sync-openclaw.mjs` once manually. Every future update from v3.11.0+ syncs automatically. (Also captured in the README Troubleshooting section as a bootstrap quirk.)
+
+**Other IDEs** (Cline / Aider / Cursor / opencode) query `/v1/models` live, so they pick up new models on the next request — no sync needed. Continue.dev users edit their own `config.json` model id manually.

keys.mjs

@@ -6,26 +6,74 @@ import { join } from "node:path";
 import { mkdirSync, chmodSync } from "node:fs";
 import { homedir } from "node:os";
 
-const OCP_DIR = join(homedir(), ".ocp");
-mkdirSync(OCP_DIR, { recursive: true, mode: 0o700 });
-// Tighten the directory mode in case it already existed with broader permissions.
-try { chmodSync(OCP_DIR, 0o700); } catch { /* ignore EPERM on pre-existing dirs */ }
-const DB_PATH = join(OCP_DIR, "ocp.db");
+// Resolved LAZILY, on first getDb() — not at module top-level. Two reasons, and the second is
+// the bug this fixes:
+//
+//  1. Merely IMPORTING keys.mjs should not, as a side effect, create directories in the
+//     operator's home.
+//  2. OCP_DIR_OVERRIDE exists so the test suite can point the key store at a scratch dir — and
+//     because ESM hoists imports, a top-level `const OCP_DIR = ...` here would be evaluated
+//     BEFORE an importing module's body could set the env var. Eager resolution made the
+//     override unsettable in the one place that needs it. (test-features.mjs carried a comment
+//     claiming it could "set env before the first getDb() call" — it could not, because nothing
+//     here ever read an env var. So `npm test` wrote real, UNREVOKED api_keys rows into the
+//     operator's live ~/.ocp/ocp.db: two per run, unbounded — 737 junk keys against 12 real ones
+//     on the maintainer's host — and two concurrent runs raced one file, which is the ~1-in-6
+//     flake in `listKeys includes quota fields`.)
+//
+// The override is gated on NODE_ENV === "test", and that gate is the ACTUAL guard. An earlier
+// cut of this fix relied on the variable merely having an awkward name — i.e. a naming convention
+// plus a comment — which is precisely the failure mode this whole change exists to indict (a
+// comment describing an intention that nothing enforces). The two-key gate means NEITHER var
+// alone does anything: a stray OCP_DIR_OVERRIDE with no NODE_ENV is inert, and NODE_ENV=test with
+// no override just resolves the default dir.
+//
+// This gate does NOT, by itself, prove a production daemon can't be redirected — an earlier
+// version of this comment overclaimed that ("a production server runs without NODE_ENV, so it
+// CANNOT honor the override no matter how the variable got in"). That is only true while the
+// daemon's env actually lacks NODE_ENV=test, which is an assumption, not something this file can
+// enforce. What makes it hold in the shipped configuration is defense-in-depth in OCP's launchers:
+// the plist/systemd units strip both vars on every (re)install (scripts/lib/plist-merge.mjs
+// NEVER_PRESERVE), and `ocp` restart's manual nohup fallback strips them (`env -u`). So a server
+// OCP itself started cannot carry the test-only redirection. The one residual path is an operator
+// who hand-launches `node server.mjs` with BOTH vars explicitly exported, bypassing every
+// launcher — a case no library-level gate can catch. The loud getDb() log below ("NOT the default
+// ~/.ocp/ocp.db") is the backstop there: a wrong key store is at least never silent (in
+// AUTH_MODE=multi that would otherwise be a total auth outage with nothing on /health to show it).
+function resolveOcpDir() {
+  const override = process.env.NODE_ENV === "test" ? process.env.OCP_DIR_OVERRIDE : null;
+  const dir = override || join(homedir(), ".ocp");
+  mkdirSync(dir, { recursive: true, mode: 0o700 });
+  // Tighten the directory mode in case it already existed with broader permissions.
+  try { chmodSync(dir, 0o700); } catch { /* ignore EPERM on pre-existing dirs */ }
+  return dir;
+}
 
 let db;
+let dbPath;   // resolved on first open, alongside the db handle
 
 export function getDb() {
   if (!db) {
-    db = new DatabaseSync(DB_PATH);
+    dbPath = join(resolveOcpDir(), "ocp.db");
+    // Say which store we opened. Silence was the other half of the bug: a server on the wrong
+    // key store looks exactly like a server on the right one until every request 401s.
+    if (dbPath !== join(homedir(), ".ocp", "ocp.db")) {
+      console.error(`[keys] key store: ${dbPath} (NOT the default ~/.ocp/ocp.db)`);
+    }
+    db = new DatabaseSync(dbPath);
     db.exec("PRAGMA journal_mode = WAL");
     db.exec("PRAGMA foreign_keys = ON");
     initSchema();
     // Tighten mode on the DB file (0600) after creation / first open.
-    try { chmodSync(DB_PATH, 0o600); } catch { /* ignore — same-user access still works */ }
+    try { chmodSync(dbPath, 0o600); } catch { /* ignore — same-user access still works */ }
   }
   return db;
 }
 
+// Which file the key store actually opened. Exported so a test can ASSERT it is not the
+// operator's real db — the bug this replaced was invisible precisely because nothing checked.
+export function getDbPath() { return dbPath; }
+
 function initSchema() {
   db.exec(`
     CREATE TABLE IF NOT EXISTS api_keys (
@@ -307,6 +355,16 @@ export function cacheHash(model, messages, opts = {}) {
   if (opts.temperature != null) h.update(`t:${opts.temperature}`);
   if (opts.max_tokens != null) h.update(`mt:${opts.max_tokens}`);
   if (opts.top_p != null) h.update(`tp:${opts.top_p}`);
+  // #176: fold the server's boot-config epoch into the key, so a config change that shapes
+  // answers (operator system prompt, wrapper text, allowed tools, NO_CONTEXT) invalidates
+  // the persistent cache instead of serving answers composed under the old config. Callers
+  // that omit it (older paths, tests) hash byte-identically to before.
+  if (opts.configEpoch != null) h.update(`ce:${opts.configEpoch}|`);
+  // Structured-output (OpenAI response_format / json_mode) requests must never share a cache slot
+  // with the conversational answer to the same prompt, nor with a different schema — the steering
+  // instruction and validated JSON payload differ. Keying on the detected descriptor isolates them.
+  // Absent for normal requests → hashes are byte-identical to pre-change.
+  if (opts.structured != null) h.update(`s:${JSON.stringify(opts.structured)}`);
   for (const m of messages) {
     h.update(m.role || "");
     h.update(typeof m.content === "string" ? m.content : JSON.stringify(m.content));
@@ -382,11 +440,25 @@ export function getCacheStats() {
 // Per ADR 0005 / spec D4: in-process scope only (single Node process per host).
 const inflightMap = new Map();
 
-export function singleflight(hash, fn) {
+// `retryIf` (optional, audit finding M1): a predicate applied on the FOLLOWER path only.
+// When a follower joins an existing flight and the shared promise rejects with an error for
+// which retryIf(err) is true (in practice: the LEADER's client disconnected while queued —
+// an error that is personal to the leader, not a verdict about the upstream), the follower
+// does NOT inherit that rejection. Instead it re-enters singleflight with its OWN fn: it
+// either becomes the new leader (the map entry is already deleted — see the finally below,
+// which runs before any follower's catch because it is attached upstream of the promise the
+// followers await) or joins a flight another retrying follower just created. The leader's
+// own rejection is never retried here — its error belongs to it (leader path returns the
+// bare promise). Callers that pass no retryIf get the exact pre-M1 share-everything behavior.
+export function singleflight(hash, fn, retryIf) {
   const existing = inflightMap.get(hash);
   if (existing) {
     existing.requesters++;
-    return existing.promise;
+    if (!retryIf) return existing.promise;
+    return existing.promise.catch((err) => {
+      if (!retryIf(err)) throw err;
+      return singleflight(hash, fn, retryIf);
+    });
   }
   // Wrap fn() in Promise.resolve().then() so synchronous throws don't escape.
   const promise = Promise.resolve().then(fn).finally(() => {
@@ -412,5 +484,5 @@ export function findKey(idOrName) {
 }
 
 export function closeDb() {
-  if (db) { db.close(); db = null; }
+  if (db) { db.close(); db = null; dbPath = undefined; }  // clear both — a path to a closed db is a footgun
 }

lib/env.mjs

@@ -0,0 +1,29 @@
+// OCP env-var parsing helpers.
+//
+// Fail-closed positive-integer parsing for numeric caps (body size, image
+// byte/count limits). A misconfigured cap must NEVER silently disable a guard:
+// `parseInt("unlimited", 10)` is NaN and `x > NaN` is always false, so a naive
+// parse of CLAUDE_MAX_BODY_SIZE=unlimited would remove the body-size limit
+// entirely (unbounded body → OOM). Likewise CLAUDE_MAX_BODY_SIZE=5MB naively
+// parses to 5 (bytes) and bricks the proxy. So a present-but-invalid value is
+// REJECTED (default kept, caller warns), not accepted. (PR #154 review F3.)
+//
+// Pure (no env access, no IO) so it is unit-testable without a live server.
+
+// Parse `raw` as a strictly-positive base-10 integer of bytes/count (no unit
+// suffix). Returns { value, ok, reason }:
+//   - missing/empty        → { value: def, ok: true }               (use default)
+//   - valid positive int   → { value: n,  ok: true }
+//   - anything else        → { value: def, ok: false, reason }      (fail closed)
+// Rejects: NaN ("unlimited"), non-positive ("0", "-1"), unit-suffixed ("5MB"),
+// and fractional/ambiguous ("20.5", "0x10") values — String(n) !== trimmed catches
+// any input parseInt only partially consumed.
+export function parsePositiveInt(raw, def) {
+  if (raw === undefined || raw === null || raw === "") return { value: def, ok: true };
+  const trimmed = String(raw).trim();
+  const n = parseInt(trimmed, 10);
+  if (!Number.isFinite(n) || n <= 0 || String(n) !== trimmed) {
+    return { value: def, ok: false, reason: "not a strictly-positive integer (bytes/count, no unit suffix)" };
+  }
+  return { value: n, ok: true };
+}

lib/multimodal.mjs

@@ -0,0 +1,278 @@
+// OCP multimodal helpers — OpenAI `image_url` content parts → Anthropic image
+// blocks fed to `claude -p --input-format stream-json`. (issue #110)
+//
+// Class B.1 (OpenAI-compatibility surface). Protocol authority is OpenAI's
+// chat/completions spec — the multimodal `content` parts shape
+// (https://platform.openai.com/docs/guides/vision and
+// https://platform.openai.com/docs/api-reference/chat/create#chat-create-messages,
+// `image_url` part with `image_url.url` = data URI or http(s) URL). Authorized
+// by ADR 0006. This module introduces NO field beyond OpenAI's published shape:
+// the OpenAI-side vocabulary read here is `type:"image_url"` +
+// `image_url:{url, detail?}`; the Anthropic-side vocabulary written here
+// (`type:"image", source:{type:"base64"|"url", ...}`) is the CLI's native
+// stream-json input contract, not an OCP invention.
+//
+// Kept as a pure module (no I/O, no network, no process state) mirroring the
+// lib/*.mjs pattern so it is unit-testable without a live server. server.mjs is
+// the only consumer; it owns spawning, caps configuration, and HTTP status.
+
+// Anthropic vision-supported image media types. A data URI whose media type is
+// outside this set is rejected with a clear 4xx rather than forwarded (the API
+// would reject it anyway; failing early gives a better error).
+export const SUPPORTED_IMAGE_TYPES = new Set([
+  "image/jpeg",
+  "image/png",
+  "image/gif",
+  "image/webp",
+]);
+
+// Default caps. server.mjs overrides these from env; they live here so the
+// pure transform is self-contained and testable.
+export const DEFAULT_MULTIMODAL_OPTS = {
+  allowRemoteUrl: false,        // http(s) image URLs are OFF by default (v1: data URIs only)
+  maxImageBytes: 5 * 1024 * 1024,   // per-image decoded-byte cap
+  maxImages: 20,                // max image parts across the whole request
+  maxTotalImageBytes: 20 * 1024 * 1024, // aggregate decoded-byte cap
+  maxTextChars: Infinity,       // text-char budget (server passes MAX_PROMPT_CHARS); Infinity = no truncation
+};
+
+// Typed error so server.mjs can map to the right HTTP status + OpenAI-shaped
+// error body. `status` is the HTTP code; `type` is the OpenAI error `type`.
+export class MultimodalError extends Error {
+  constructor(code, status, message) {
+    super(message);
+    this.name = "MultimodalError";
+    this.code = code;
+    this.status = status;
+    this.type = "invalid_request_error"; // OpenAI error `type` for 4xx client errors
+  }
+}
+
+// True if any message carries an OpenAI `image_url` content part. Cheap guard so
+// the byte-for-byte text path is only left when an image is genuinely present.
+export function hasImageContent(messages) {
+  if (!Array.isArray(messages)) return false;
+  for (const m of messages) {
+    if (m && Array.isArray(m.content)) {
+      for (const part of m.content) {
+        if (part && part.type === "image_url") return true;
+      }
+    }
+  }
+  return false;
+}
+
+// Extract the URL string from an OpenAI image_url part. Spec form is
+// `{type:"image_url", image_url:{url, detail?}}`; many OpenAI-compatible clients
+// also send `image_url` as a bare string. Accept both (input leniency — no new
+// output field). `detail` (auto|low|high) is OpenAI-only and has no Anthropic
+// analogue, so it is read-and-ignored.
+function imageUrlOf(part) {
+  const iu = part.image_url;
+  if (typeof iu === "string") return iu;
+  if (iu && typeof iu.url === "string") return iu.url;
+  return null;
+}
+
+// Parse a base64 data URI: `data:[<media_type>][;base64],<data>`.
+// Returns { mediaType, data (base64), bytes (decoded size) } or throws MultimodalError.
+function parseDataUri(uri) {
+  const comma = uri.indexOf(",");
+  if (comma === -1) {
+    throw new MultimodalError("invalid_data_uri", 400, "Malformed image data URI (no comma).");
+  }
+  const meta = uri.slice(5, comma); // strip leading "data:"
+  const segs = meta.split(";");
+  const mediaType = (segs[0] || "").trim().toLowerCase();
+  const isBase64 = segs.slice(1).some((s) => s.trim().toLowerCase() === "base64");
+  if (!isBase64) {
+    throw new MultimodalError("invalid_data_uri", 400, "Only base64-encoded image data URIs are supported.");
+  }
+  if (!SUPPORTED_IMAGE_TYPES.has(mediaType)) {
+    throw new MultimodalError(
+      "unsupported_image_type",
+      400,
+      `Unsupported image media type '${mediaType || "(none)"}'. Supported: ${[...SUPPORTED_IMAGE_TYPES].join(", ")}.`
+    );
+  }
+  // Strip incidental whitespace/newlines some encoders insert into data URIs.
+  const data = uri.slice(comma + 1).replace(/\s/g, "");
+  if (data.length === 0 || !/^[A-Za-z0-9+/]+={0,2}$/.test(data)) {
+    throw new MultimodalError("invalid_data_uri", 400, "Image data URI payload is not valid base64.");
+  }
+  // Decoded size from base64 length (minus padding); avoids decoding the buffer
+  // just to measure it.
+  const padding = data.endsWith("==") ? 2 : data.endsWith("=") ? 1 : 0;
+  const bytes = Math.floor((data.length * 3) / 4) - padding;
+  return { mediaType, data, bytes };
+}
+
+// Convert a single OpenAI image_url part to an Anthropic image block, enforcing
+// caps via the mutable `acc` accumulator ({ images, bytes }). Throws MultimodalError.
+function imagePartToBlock(part, opts, acc) {
+  const url = imageUrlOf(part);
+  if (!url) {
+    throw new MultimodalError("invalid_image_url", 400, "image_url part is missing a URL.");
+  }
+
+  acc.images += 1;
+  if (acc.images > opts.maxImages) {
+    throw new MultimodalError("too_many_images", 413, `Too many images in request (max ${opts.maxImages}).`);
+  }
+
+  if (url.startsWith("data:")) {
+    const { mediaType, data, bytes } = parseDataUri(url);
+    if (bytes > opts.maxImageBytes) {
+      throw new MultimodalError("image_too_large", 413, `Image exceeds per-image size limit (${opts.maxImageBytes} bytes).`);
+    }
+    acc.bytes += bytes;
+    if (acc.bytes > opts.maxTotalImageBytes) {
+      throw new MultimodalError("images_too_large", 413, `Total image payload exceeds limit (${opts.maxTotalImageBytes} bytes).`);
+    }
+    return { type: "image", source: { type: "base64", media_type: mediaType, data } };
+  }
+
+  if (/^https?:\/\//i.test(url)) {
+    if (!opts.allowRemoteUrl) {
+      throw new MultimodalError(
+        "remote_url_disabled",
+        400,
+        "Remote image URLs are disabled. Enable CLAUDE_IMAGE_ALLOW_URL=1 to allow http(s) image URLs, or pass the image as a base64 data URI."
+      );
+    }
+    // Passthrough as an Anthropic url-source block. OCP does NOT fetch the URL
+    // itself (no OCP-side SSRF surface); the fetch is performed upstream by the
+    // Anthropic API. Best-effort: unreachable/blocked URLs surface as an API error.
+    return { type: "image", source: { type: "url", url } };
+  }
+
+  throw new MultimodalError("unsupported_url_scheme", 400, "image_url must be a base64 data URI or an http(s) URL.");
+}
+
+// Role prefix mirrors messagesToPrompt()'s text-path labeling so a multi-turn
+// conversation reads the same whether or not it carries images. System messages
+// are handled by the caller via --system-prompt and never reach here.
+function rolePrefix(role) {
+  if (role === "assistant") return "[Assistant] ";
+  return ""; // user / tool / anything else: verbatim, as in the text path
+}
+
+// Build the Anthropic content-block array for a single stream-json user
+// envelope. Mirrors the text path's "collapse the whole conversation into one
+// turn passed via stdin" model (OCP runs stateless, full context per spawn), but
+// preserves image position relative to text and keeps images out of the text
+// char budget entirely. Returns { blocks, stats } or throws MultimodalError.
+export function buildImageBlocks(messages, opts = {}) {
+  const o = { ...DEFAULT_MULTIMODAL_OPTS, ...opts };
+  const blocks = [];
+  const acc = { images: 0, bytes: 0 };
+  let textChars = 0;
+  let firstMessage = true;
+
+  const pushText = (text) => {
+    if (!text) return;
+    blocks.push({ type: "text", text });
+    textChars += text.length;
+  };
+
+  for (const m of messages) {
+    const prefix = rolePrefix(m.role);
+    // Separate messages with a blank line, matching messagesToPrompt's "\n\n" join.
+    const sep = firstMessage ? "" : "\n\n";
+    firstMessage = false;
+    let prefixEmitted = false;
+    const emitPrefixWith = (t) => {
+      if (prefixEmitted) return t;
+      prefixEmitted = true;
+      return sep + prefix + t;
+    };
+
+    if (typeof m.content === "string") {
+      pushText(emitPrefixWith(m.content));
+      continue;
+    }
+    if (!Array.isArray(m.content)) {
+      // null / object content: mirror contentToText's fallback.
+      const t = m.content == null ? "" : JSON.stringify(m.content);
+      pushText(emitPrefixWith(t));
+      continue;
+    }
+
+    for (const part of m.content) {
+      if (part && part.type === "text" && typeof part.text === "string") {
+        pushText(emitPrefixWith(part.text));
+      } else if (part && part.type === "image_url") {
+        // Ensure the role prefix isn't lost when a message leads with an image.
+        if (!prefixEmitted && prefix) pushText(emitPrefixWith(""));
+        blocks.push(imagePartToBlock(part, o, acc));
+      } else {
+        // audio / file / unknown parts: preserve the existing placeholder
+        // behavior (issue #110) — deferred to a future version.
+        pushText(emitPrefixWith("[non-text content omitted]"));
+      }
+    }
+  }
+
+  // Defensive: a stream-json user turn must have at least one content block.
+  if (blocks.length === 0) blocks.push({ type: "text", text: "" });
+
+  // Enforce the text-char budget (PR #154 review F2). Without this, attaching a
+  // single tiny image would let unbounded text bypass the gateway's runaway-context
+  // guard entirely — messagesToPrompt truncates the text path, so the multimodal
+  // path must too. Image blocks are preserved and are NOT counted (they are bounded
+  // by the byte/count caps above).
+  const budgeted = enforceTextBudget(blocks, o.maxTextChars);
+
+  return {
+    blocks: budgeted.blocks,
+    stats: {
+      imageCount: acc.images,
+      totalImageBytes: acc.bytes,
+      textChars: budgeted.textChars,
+      originalTextChars: budgeted.originalTextChars,
+      truncated: budgeted.truncated,
+    },
+  };
+}
+
+// Enforce a text-char budget over already-built content blocks, mirroring
+// messagesToPrompt's "keep the tail, drop the oldest" truncation. Image blocks are
+// preserved in place; only text blocks count against the budget. A truncation note
+// is prepended when anything is dropped. Returns
+// { blocks, truncated, originalTextChars, textChars }.
+function enforceTextBudget(blocks, maxTextChars) {
+  let originalTextChars = 0;
+  for (const b of blocks) if (b.type === "text") originalTextChars += b.text.length;
+  if (!(maxTextChars > 0) || originalTextChars <= maxTextChars) {
+    return { blocks, truncated: false, originalTextChars, textChars: originalTextChars };
+  }
+  // Keep the most recent text up to the budget; trim within the boundary block and
+  // drop older text blocks. Non-text (image) blocks always survive, in order.
+  let budget = maxTextChars;
+  const out = [];
+  for (let i = blocks.length - 1; i >= 0; i--) {
+    const b = blocks[i];
+    if (b.type !== "text") { out.unshift(b); continue; }
+    if (budget <= 0) continue; // older text fully dropped
+    if (b.text.length <= budget) {
+      out.unshift(b);
+      budget -= b.text.length;
+    } else {
+      out.unshift({ type: "text", text: b.text.slice(b.text.length - budget) });
+      budget = 0;
+    }
+  }
+  out.unshift({ type: "text", text: "[System] Note: older text content was truncated to fit the context limit." });
+  let textChars = 0;
+  for (const b of out) if (b.type === "text") textChars += b.text.length;
+  return { blocks: out, truncated: true, originalTextChars, textChars };
+}
+
+// Serialize the non-system conversation to a single newline-terminated
+// stream-json user message for `claude -p --input-format stream-json` stdin.
+// Returns { payload, stats } or throws MultimodalError.
+export function buildStreamJsonInput(messages, opts = {}) {
+  const { blocks, stats } = buildImageBlocks(messages, opts);
+  const envelope = { type: "user", message: { role: "user", content: blocks } };
+  return { payload: JSON.stringify(envelope) + "\n", stats };
+}

lib/prompt.mjs

@@ -0,0 +1,89 @@
+// lib/prompt.mjs — pure operator-append step for the system prompt.
+//
+// Extracted so the rule is unit-testable (the suite never imports server.mjs — it
+// boots a listener). server.mjs composes wrapper + client system messages exactly as
+// before, then passes the result through this. With CLAUDE_SYSTEM_PROMPT unset the
+// return is the INPUT STRING UNCHANGED — the default path stays byte-for-byte
+// identical, which is the repo's bar for touching a request-shaping function.
+//
+// The operator prompt goes LAST deliberately: a server-wide directive ("answer in
+// Chinese") should read as the final instruction, not something a client system
+// message overrides by coming later. Whitespace-only values are treated as unset —
+// a stray space in a service unit's Environment= line must not inject "\n\n " into
+// every request.
+export function appendOperatorPrompt(base, operatorAppend) {
+  const op = typeof operatorAppend === "string" ? operatorAppend.trim() : "";
+  return op ? `${base}\n\n${op}` : base;
+}
+
+// Derive the default prompt-char budget from the models.json SPOT (ADR 0009).
+//
+// The old default was a hand-set constant (150000 chars ≈ 37.5k English tokens) from the
+// 200k-window era — silently far below what the advertised contextWindow promises. Instead
+// of picking a new constant that will also rot, the default now FOLLOWS the SPOT:
+//
+//   budget = max(models[].contextWindow) × charsPerToken
+//
+// charsPerToken = 3 is deliberately conservative: English runs ~4 chars/token, CJK ~1–1.5.
+// At ×3, a 200k-token window yields 600,000 chars — full window for English, and CJK text
+// hits the model's real window at roughly the same point the cap fires, so we truncate
+// (graceful, tail-first) rather than let the upstream reject the request outright.
+//
+// The floor guards the degenerate cases (empty/missing models[], absent contextWindow):
+// fall back to the historical constant rather than 0 — a zero budget would truncate every
+// request to nothing, which is fail-OPEN in the "serve garbage" sense. CLAUDE_MAX_PROMPT_CHARS
+// remains an absolute operator override at the call site (server.mjs); this function is only
+// the unset-env default.
+export function derivePromptCharBudget(models, { charsPerToken = 3, floor = 150000 } = {}) {
+  const windows = (Array.isArray(models) ? models : [])
+    .map(m => m?.contextWindow)
+    .filter(w => Number.isFinite(w) && w > 0);
+  if (windows.length === 0) return floor;
+  return Math.max(floor, Math.max(...windows) * charsPerToken);
+}
+
+// Resolve the effective budget from the env var + SPOT. TRUTHINESS (not != null) on the env
+// value deliberately: an EMPTY value ("CLAUDE_MAX_PROMPT_CHARS=" in a systemd EnvironmentFile
+// or .env) must mean "use the default" — exactly the old `parseInt(env || "150000")` contract.
+// Treating "" as explicit gives parseInt("") = NaN, and a NaN cap silently DISABLES the
+// runaway-context guard while injecting a false "[System] Note: 0 older messages were
+// truncated" line into every prompt (caught in PR #179 review). Non-empty garbage still
+// parses to NaN — the pre-existing class, slated for parseIntEnv routing in PR #154.
+export function resolvePromptCharBudget(rawEnv, models, opts) {
+  return rawEnv ? parseInt(rawEnv, 10) : derivePromptCharBudget(models, opts);
+}
+
+// OCP_LOCAL_TOOLS system-prompt wrapper selection (pure).
+//
+// OCP's `-p` path prepends a fixed wrapper to every request's system prompt. The DEFAULT wrapper
+// tells the model it has NO local filesystem/shell/env access — the right posture for a shared or
+// multi-tenant gateway. But a single-user, loopback-bound instance (e.g. an OpenClaw agent talking
+// to its own local OCP) DOES legitimately have tools — the `-p` path already passes `--allowedTools`
+// and the CLI's built-in tools are available — so the default wrapper actively gaslights the model
+// into refusing to use tools it holds. `OCP_LOCAL_TOOLS=1` swaps in a positive wrapper for that case.
+//
+// This does NOT expand the tool surface: tools are governed solely by `--allowedTools` /
+// `--disallowedTools` (multi-tenant mode `--disallowedTools` the whole FS surface regardless of the
+// wrapper). It only changes the PROMPT the operator's own model reads. Pure so it is unit-testable.
+export function selectPromptWrapper(localToolsEnabled, negativeWrapper, positiveWrapper) {
+  return localToolsEnabled ? positiveWrapper : negativeWrapper;
+}
+
+// Boot-time safety gate for OCP_LOCAL_TOOLS, mirroring the OCP_TUI_FULL_TOOLS model (ADR 0007): a
+// positive "you may use local tools" wrapper must never reach an untrusted caller. Returns a fatal
+// message string when the flag is enabled in an unsafe deployment, or null when it is safe/disabled.
+// Fail-closed: any of multi-tenant auth, a non-loopback bind, or an anonymous key is refused. Pure —
+// the caller does the process.exit so this stays testable.
+export function localToolsSafetyError({ enabled, authMode, loopbackBind, anonymousKey }) {
+  if (!enabled) return null;
+  if (authMode === "multi") {
+    return "OCP_LOCAL_TOOLS=1 is incompatible with CLAUDE_AUTH_MODE=multi — a guest/anonymous prompt would be told it may drive the operator's filesystem/shell. Single-user only.";
+  }
+  if (!loopbackBind) {
+    return "OCP_LOCAL_TOOLS=1 requires a loopback bind (127.0.0.1/::1) — a network-exposed positive-tools wrapper could reach an untrusted peer. Bind to loopback, or leave OCP_LOCAL_TOOLS off.";
+  }
+  if (anonymousKey) {
+    return "OCP_LOCAL_TOOLS=1 is unsafe with PROXY_ANONYMOUS_KEY set — anonymous callers could reach the local-tools-enabled model without a named key. Remove PROXY_ANONYMOUS_KEY, or leave OCP_LOCAL_TOOLS off.";
+  }
+  return null;
+}

lib/spawn-auth.mjs

@@ -0,0 +1,67 @@
+// Pure, dependency-injected primitives for the `-p` spawn-token resolution + HOME-isolation
+// layer. Extracted from server.mjs (findings F3 / F5 / F6, 2026-07-07) so the concurrency,
+// caching and expiry logic is unit-testable WITHOUT booting the server or mocking execFileSync /
+// child_process.spawn / fs. server.mjs owns all I/O (macOS keychain exec, process spawn, fs);
+// this module owns only pure decision logic.
+//
+// ALIGNMENT NOTE: none of this touches the OAuth wire machinery (no endpoint / header / body).
+// OCP still NEVER performs a refresh_token grant itself — these helpers only READ + GATE a token
+// that some other process (the operator's real claude, or a spawned claude under the real HOME)
+// refreshes. That property is load-bearing (issue #112) and preserved.
+
+// Promise-chain mutex. `acquire()` resolves to a `release()` fn; the NEXT `acquire()` does not
+// resolve until the current holder calls its `release()`. Serializes async critical sections
+// without busy-waiting. release() is idempotent.
+export function createSerialMutex() {
+  let tail = Promise.resolve();
+  return {
+    acquire() {
+      let release;
+      const gate = new Promise((r) => { release = r; });
+      const prev = tail;
+      tail = tail.then(() => gate);
+      // Hand the caller its release fn only after the previous holder has released.
+      return prev.then(() => {
+        let released = false;
+        return function releaseMutex() { if (!released) { released = true; release(); } };
+      });
+    },
+  };
+}
+
+// Short-TTL memo. `get(produce, now)` returns the cached value while `now - storedAt < ttlMs`,
+// otherwise calls `produce()` and re-stores. A miss that produces null/undefined is STILL stored
+// (so a genuinely-absent source is not re-probed on every call within the TTL window). `now` is
+// injectable for testing.
+export function createTtlCache({ ttlMs }) {
+  let value;
+  let at = -Infinity;
+  let has = false;
+  return {
+    get(produce, now = Date.now()) {
+      if (has && now - at < ttlMs) return value;
+      value = produce();
+      at = now;
+      has = true;
+      return value;
+    },
+    clear() { has = false; value = undefined; at = -Infinity; },
+  };
+}
+
+// Pure expiry gate. Returns true when `creds` carries a known expiry that is at/within `bufferMs`
+// of `now`. Creds WITHOUT `expiresAt` (e.g. long-lived env tokens) are never treated as expiring.
+// This gate is applied to the CACHED creds on EVERY use — which is precisely why a short-TTL
+// keychain cache (createTtlCache) cannot reintroduce the #146 forever-stale-token regression: the
+// cache bounds how often we re-READ the keychain, but the expiry decision is recomputed per use.
+export function isTokenExpiring(creds, now = Date.now(), bufferMs = 300000) {
+  return !!(creds && creds.expiresAt && now + bufferMs >= creds.expiresAt);
+}
+
+// Order candidate keychain labels so the last-known-good label is tried first (avoids the
+// wrong-label miss that doubles the `security` exec count on the hot path). Pure: performs no
+// read. Returns a fresh array; input is not mutated.
+export function orderLabelsLastGoodFirst(labels, lastGood) {
+  if (!lastGood || !labels.includes(lastGood)) return labels.slice();
+  return [lastGood, ...labels.filter((l) => l !== lastGood)];
+}

lib/structured-output.mjs

@@ -0,0 +1,318 @@
+// ── OpenAI Structured Outputs (response_format) — pure helpers ───────────────
+//
+// OCP's `/v1/chat/completions` (Class B.1, ADR 0006) advertises OpenAI compatibility but forwards
+// to `claude -p`, which has no native `response_format`. Asked for JSON, the coding-assistant CLI
+// typically replies with prose, a Markdown table, or a ```json fenced block — none of which is
+// `JSON.parse`-able. These helpers implement the OpenAI `response_format` contract on top of that:
+// detect the request, build a strict JSON-only steering instruction, then extract and validate the
+// JSON the model returns. All functions here are pure (no I/O) so they are unit-tested directly;
+// the retry loop that calls the model lives in server.mjs (runStructuredCompletion).
+//
+// Spec authority (B.1, ADR 0006): OpenAI chat/completions `response_format`
+//   https://platform.openai.com/docs/api-reference/chat/create#chat-create-response_format
+// No field or behaviour beyond that published shape is introduced.
+
+export class StructuredOutputError extends Error {
+  constructor(reason, raw) {
+    super(`structured output could not be produced: ${reason}`);
+    this.name = "StructuredOutputError";
+    this.reason = reason;
+    this.raw = raw;
+  }
+}
+
+// Fail-closed parse of the OCP_STRUCTURED_MAX_ATTEMPTS retry cap. `Math.max(1, parseInt("abc",10))`
+// === `Math.max(1, NaN)` === NaN, and a retry loop bounded by `attempt < NaN` never runs → 0 spawns,
+// every structured request silently refuses. So any non-integer / non-finite / <1 value keeps the
+// documented default instead (and warns), rather than bricking the feature. (PR #153 review round 2.)
+export function resolveMaxAttempts(raw, { fallback = 3, warn } = {}) {
+  if (raw === undefined || raw === null || raw === "") return fallback;
+  const n = parseInt(raw, 10);
+  if (!Number.isFinite(n) || n < 1) {
+    if (typeof warn === "function") {
+      warn(`Ignoring invalid OCP_STRUCTURED_MAX_ATTEMPTS="${raw}" (want integer >= 1); using default ${fallback}.`);
+    }
+    return fallback;
+  }
+  return n;
+}
+
+// Returns { mode: "schema", schema, name?, strict } | { mode: "json_object" } | null.
+// Supports the OpenAI shapes: response_format:{type:"json_schema",json_schema:{schema,strict,name}}
+// and response_format:{type:"json_object"}, a lenient response_format:{schema} fallback, and the
+// widely-used (non-standard) top-level `json_mode: true` flag as a json_object alias.
+export function detectStructuredOutput(parsed) {
+  const rf = parsed?.response_format;
+  if (rf && typeof rf === "object") {
+    if (rf.type === "json_schema") {
+      const js = (rf.json_schema && typeof rf.json_schema === "object") ? rf.json_schema : {};
+      const schema = js.schema || rf.schema || null;
+      return { mode: "schema", schema, name: js.name, strict: js.strict === true };
+    }
+    if (rf.type === "json_object") return { mode: "json_object" };
+    if (rf.schema && typeof rf.schema === "object") {
+      return { mode: "schema", schema: rf.schema, strict: rf.strict === true };
+    }
+  }
+  // Non-standard convenience alias honored by several OpenAI-compatible clients.
+  if (parsed?.json_mode === true) return { mode: "json_object" };
+  return null;
+}
+
+export function jsonTypeOf(v) {
+  if (v === null) return "null";
+  if (Array.isArray(v)) return "array";
+  return typeof v; // object | string | number | boolean
+}
+
+export function jsonTypeMatches(t, value) {
+  switch (t) {
+    case "string": return typeof value === "string";
+    case "number": return typeof value === "number" && Number.isFinite(value);
+    case "integer": return typeof value === "number" && Number.isInteger(value);
+    case "boolean": return typeof value === "boolean";
+    case "object": return value !== null && typeof value === "object" && !Array.isArray(value);
+    case "array": return Array.isArray(value);
+    case "null": return value === null;
+    default: return true; // unknown type keyword → do not fail on it
+  }
+}
+
+const jsonDeepEqual = (a, b) => JSON.stringify(a) === JSON.stringify(b);
+
+// Resolve a local JSON-Pointer `$ref` (e.g. "#/$defs/Step" or "#/definitions/Step") against the
+// document root. Only same-document refs are supported (that is all the OpenAI SDK emits); a remote
+// or unresolvable ref returns null and the caller skips validation for it rather than failing.
+function resolveRef(ref, root) {
+  if (typeof ref !== "string" || !ref.startsWith("#/") || !root) return null;
+  const parts = ref.slice(2).split("/").map(p => p.replace(/~1/g, "/").replace(/~0/g, "~"));
+  let cur = root;
+  for (const p of parts) {
+    if (cur && typeof cur === "object" && Object.prototype.hasOwnProperty.call(cur, p)) cur = cur[p];
+    else return null;
+  }
+  return (cur && typeof cur === "object") ? cur : null;
+}
+
+// Minimal JSON-Schema validator: covers the subset OpenAI structured outputs use — type (incl.
+// arrays of types / integer), required, properties, additionalProperties (no invented keys), items
+// (list + tuple), enum, const, nullability (type:["x","null"] or nullable:true), min/maxItems, and
+// $ref/$defs + allOf/anyOf/oneOf composition (which the official OpenAI SDK emits heavily via
+// zodResponseFormat / client.beta.chat.completions.parse). `root` carries the top-level schema so
+// same-document $refs resolve. Returns error strings ([] = valid).
+// `refChain` tracks the $ref pointers resolved on the CURRENT path WITHOUT consuming data (a $ref
+// hop, or an allOf/anyOf/oneOf branch, all re-validate the same value). A pointer reappearing on
+// that chain is a pure ref→ref (or ref→composition→ref) cycle that recurses forever independent of
+// the data — we fail closed on it. Data-consuming recursion (properties/items/additionalProperties)
+// deliberately resets the chain (default []): it always terminates because a JSON value is a finite
+// tree, so a legitimately recursive schema (Node→child:Node) must NOT be flagged as a cycle. A depth
+// cap backstops any threading mistake. (PR #153 review round 2, cyclic-$ref blocker.)
+const REF_DEPTH_CAP = 512;
+export function validateJsonSchema(value, schema, path = "$", strict = false, root = schema, refChain = []) {
+  const errors = [];
+  if (!schema || typeof schema !== "object") return errors;
+  if (refChain.length > REF_DEPTH_CAP) { // defensive backstop; refChain cycle-check below is primary
+    errors.push(`${path}: $ref resolution too deep (possible cycle)`);
+    return errors;
+  }
+
+  // $ref: resolve against the document root ($defs / definitions) and validate the target. Without
+  // this a nested {$ref:"#/$defs/Step"} presents as {no type, no properties} — and under strict that
+  // used to wrongly reject every real key as "additional property not allowed" (the flagship OpenAI
+  // SDK shape). Sibling keywords alongside $ref (rare) are merged over the resolved target.
+  if (typeof schema.$ref === "string") {
+    if (refChain.includes(schema.$ref)) { // cyclic $ref (a→b→a, or self a→a) — fail closed.
+      errors.push(`${path}: cyclic $ref detected (${schema.$ref})`);
+      return errors;
+    }
+    const resolved = resolveRef(schema.$ref, root);
+    if (!resolved) return errors; // unresolvable ref → cannot validate; do not fail
+    const { $ref, ...siblings } = schema;
+    return validateJsonSchema(value, { ...resolved, ...siblings }, path, strict, root, [...refChain, schema.$ref]);
+  }
+
+  // Composition. allOf: every branch must pass. anyOf: at least one. oneOf: exactly one.
+  // These re-validate the SAME value → thread refChain so a ref cycle routed through a branch is caught.
+  if (Array.isArray(schema.allOf)) {
+    for (const sub of schema.allOf) errors.push(...validateJsonSchema(value, sub, path, strict, root, refChain));
+  }
+  if (Array.isArray(schema.anyOf)) {
+    if (!schema.anyOf.some(sub => validateJsonSchema(value, sub, path, strict, root, refChain).length === 0)) {
+      errors.push(`${path}: does not match any of the allowed schemas (anyOf)`);
+    }
+  }
+  if (Array.isArray(schema.oneOf)) {
+    const matches = schema.oneOf.filter(sub => validateJsonSchema(value, sub, path, strict, root, refChain).length === 0).length;
+    if (matches !== 1) errors.push(`${path}: must match exactly one allowed schema (oneOf), matched ${matches}`);
+  }
+
+  // Nullability takes precedence: a null value is valid whenever the schema permits null (its type
+  // union includes "null", or nullable:true), regardless of enum/const. This mirrors OpenAI
+  // structured-output behaviour — nullable fields accept null even when a bare enum (as generated by
+  // Home Assistant's extended_openai_conversation) omits null from its value list.
+  const allowsNull = schema.nullable === true
+    || (Array.isArray(schema.type) ? schema.type.includes("null") : schema.type === "null");
+  if (value === null && allowsNull) return errors;
+
+  if (Array.isArray(schema.enum) && !schema.enum.some(e => jsonDeepEqual(e, value))) {
+    errors.push(`${path}: not one of the allowed enum values`);
+  }
+  if ("const" in schema && !jsonDeepEqual(schema.const, value)) {
+    errors.push(`${path}: does not equal the required const value`);
+  }
+
+  if (schema.type !== undefined) {
+    const types = Array.isArray(schema.type) ? schema.type : [schema.type];
+    const nullable = schema.nullable === true || types.includes("null");
+    const ok = types.some(t => jsonTypeMatches(t, value)) || (value === null && nullable);
+    if (!ok) {
+      errors.push(`${path}: expected ${types.join("|")}${schema.nullable ? "|null" : ""}, got ${jsonTypeOf(value)}`);
+      return errors; // type mismatch — deeper checks are meaningless
+    }
+  }
+  if (value === null) return errors;
+
+  const vt = jsonTypeOf(value);
+  if (vt === "object") {
+    const props = schema.properties || {};
+    for (const r of (schema.required || [])) {
+      if (!Object.prototype.hasOwnProperty.call(value, r)) errors.push(`${path}.${r}: required property missing`);
+    }
+    const addl = schema.additionalProperties;
+    // Only treat strict as implying "no additional properties" when this object actually declares
+    // its own `properties` and is NOT a composite (allOf/anyOf/oneOf put the real keys in sub-schemas,
+    // which are validated separately above). Inferring closure from an EMPTY properties map — the
+    // shape an unresolved $ref or a pure-composition node presents — would reject every real key.
+    // An explicit additionalProperties:false is always honoured. (PR #153 review, finding 1.)
+    const isComposite = Array.isArray(schema.allOf) || Array.isArray(schema.anyOf) || Array.isArray(schema.oneOf);
+    const noExtra = addl === false || (strict && addl === undefined && Object.keys(props).length > 0 && !isComposite);
+    for (const k of Object.keys(value)) {
+      if (props[k]) {
+        errors.push(...validateJsonSchema(value[k], props[k], `${path}.${k}`, strict, root));
+      } else if (isComposite) {
+        // key may be defined in an allOf/anyOf/oneOf branch — already validated there; don't reject.
+      } else if (noExtra) {
+        errors.push(`${path}.${k}: additional property not allowed`);
+      } else if (addl && typeof addl === "object") {
+        errors.push(...validateJsonSchema(value[k], addl, `${path}.${k}`, strict, root));
+      }
+    }
+  } else if (vt === "array" && schema.items) {
+    if (Array.isArray(schema.items)) {
+      schema.items.forEach((s, i) => { if (i < value.length) errors.push(...validateJsonSchema(value[i], s, `${path}[${i}]`, strict, root)); });
+    } else {
+      value.forEach((item, i) => errors.push(...validateJsonSchema(item, schema.items, `${path}[${i}]`, strict, root)));
+    }
+    if (typeof schema.minItems === "number" && value.length < schema.minItems) errors.push(`${path}: fewer items than minItems ${schema.minItems}`);
+    if (typeof schema.maxItems === "number" && value.length > schema.maxItems) errors.push(`${path}: more items than maxItems ${schema.maxItems}`);
+  }
+  return errors;
+}
+
+// Crash-safe façade over validateJsonSchema (issue #181). The validator recurses on the DATA's
+// nesting depth (properties/items/additionalProperties), which the REF_DEPTH_CAP does NOT bound —
+// only the ref-chain is. A model reply nested ~2000 levels deep therefore overflowed the stack with
+// a RangeError, which the request handler caught as a generic HTTP 500 instead of the spec-correct
+// `refusal`. This wrapper converts ANY throw (the deep-data RangeError, or any future recursion
+// hazard) into a single validation error, so the structured-output retry loop treats a pathological
+// reply as "did not validate" → refusal — never a 500, never a crash. A well-formed reply is
+// unaffected: the inner validator returns and this just passes its errors through.
+export function validateJsonSchemaSafe(value, schema, path = "$", strict = false, root = schema) {
+  try {
+    return validateJsonSchema(value, schema, path, strict, root);
+  } catch (e) {
+    // Catch ONLY the deep-nesting stack overflow (the #181 vector) and turn it into a validation
+    // miss → retry → refusal, never a 500. Any OTHER throw is a genuine bug: re-throw it so it
+    // surfaces at error level instead of being silently masked as "did not validate" (reviewer
+    // finding — a catch-all would hide a future TypeError behind a warn-level structured_retry).
+    if (e instanceof RangeError) return [`${path}: schema validation aborted (value nesting too deep)`];
+    throw e;
+  }
+}
+
+function tryJsonParse(s) {
+  try { return { ok: true, value: JSON.parse(s) }; } catch { return { ok: false }; }
+}
+
+// Find the next brace-balanced JSON span at or after `from`. String-aware: brackets inside quoted
+// strings are ignored. Returns { text, start, end } for the first complete top-level span, or null.
+function balancedSlice(s, from) {
+  let start = -1;
+  for (let i = from; i < s.length; i++) { if (s[i] === "{" || s[i] === "[") { start = i; break; } }
+  if (start === -1) return null;
+  let depth = 0, inStr = false, esc = false;
+  for (let i = start; i < s.length; i++) {
+    const c = s[i];
+    if (inStr) {
+      if (esc) esc = false;
+      else if (c === "\\") esc = true;
+      else if (c === '"') inStr = false;
+      continue;
+    }
+    if (c === '"') { inStr = true; continue; }
+    if (c === "{" || c === "[") depth++;
+    else if (c === "}" || c === "]") { depth--; if (depth === 0) return { text: s.slice(start, i + 1), start, end: i }; }
+  }
+  return null;
+}
+
+// Extract a single JSON value from model text. Two modes (PR #153 review, finding 2):
+//
+//   opts.whole === true  (json_object mode): the ENTIRE reply, after trimming and stripping one code
+//     fence, must parse as a single JSON value. json_object has no schema to validate against, so
+//     this whole-reply parse is its ONLY guard — a reply like `I can't. The schema is {"type":"object"}`
+//     must NOT parse-and-serve the embedded object as if it were the answer.
+//
+//   default (schema mode): prose-wrapped JSON is tolerated (models often add a sentence), BUT a reply
+//     containing MORE THAN ONE top-level JSON value is rejected rather than silently serving the first
+//     — "Schema: {...}\n\nAnswer: {...}" or "Option A: {...} Option B: {...}" is ambiguous, not an
+//     answer. The extracted value is still schema-validated by the caller.
+//
+// Returns { ok:true, value } | { ok:false, reason? }.
+export function extractJsonPayload(text, opts = {}) {
+  if (typeof text !== "string") return { ok: false };
+  let s = text.trim();
+  const fence = s.match(/^```(?:json)?\s*([\s\S]*?)\s*```$/i);
+  if (fence) s = fence[1].trim();
+
+  if (opts.whole) {
+    const whole = tryJsonParse(s);
+    return whole.ok ? whole : { ok: false, reason: "reply was not a single JSON value" };
+  }
+
+  const direct = tryJsonParse(s);
+  if (direct.ok) return direct;
+
+  const first = balancedSlice(s, 0);
+  if (!first) return { ok: false };
+  const parsedFirst = tryJsonParse(first.text);
+  if (!parsedFirst.ok) return { ok: false };
+
+  // Reject ambiguity: a second parseable top-level JSON value means we cannot know which is the answer.
+  const second = balancedSlice(s, first.end + 1);
+  if (second && tryJsonParse(second.text).ok) {
+    return { ok: false, reason: "reply contained more than one JSON value" };
+  }
+  return parsedFirst;
+}
+
+// The strict JSON-only system instruction appended to the request (attempt 0), escalated with the
+// prior failure reason on retries.
+export function structuredSystemInstruction(structured, attempt, lastErr) {
+  const schemaBlock = (structured.mode === "schema" && structured.schema)
+    ? `Your JSON MUST validate EXACTLY against this JSON Schema:\n${JSON.stringify(structured.schema)}\n`
+      + `- Include every required property.\n`
+      + `- Do NOT add any property that is not defined in the schema.\n`
+      + `- Respect all declared types, enums, and nullability.`
+    : `Respond with a single valid JSON value.`;
+  let text =
+`You are a strict JSON generator. Output a SINGLE JSON value and NOTHING else.
+- The response MUST begin with { or [ and end with the matching } or ].
+- Do NOT wrap the JSON in Markdown or code fences (no \`\`\`).
+- Do NOT include any prose, explanation, heading, comment, reasoning, or XML — only the raw JSON.
+${schemaBlock}`;
+  if (attempt > 0) {
+    text = `YOUR PREVIOUS RESPONSE WAS REJECTED (${lastErr}). Output ONLY the corrected raw JSON now, with no other text.\n\n` + text;
+  }
+  return text;
+}

lib/tui/pool.mjs

@@ -0,0 +1,321 @@
+import { rmSync } from "node:fs";
+
+// TUI warm pane pool (docs/plans/2026-07-13-tui-latency backlog #3).
+//
+// WHAT IT IS: a small set of PRE-BOOTED `claude` panes, each already sitting at its
+// input bar, so a request does not pay the cold boot. Opt-in: OCP_TUI_POOL_SIZE=0
+// (default) disables it entirely and the request path is byte-for-byte today's.
+//
+// ── SINGLE-USE IS THE LOAD-BEARING RULE ─────────────────────────────────────
+// A pooled pane serves EXACTLY ONE turn and is then killed and replaced in the
+// background. Each pane carries its OWN fresh `--session-id`, fixed at boot, and the
+// turn locates its transcript by that id. So OCP's one-session-per-request model is
+// preserved: a session's transcript still holds exactly one logical exchange.
+// That is what keeps lib/tui/transcript.mjs's extractLatestAssistantText (which returns
+// the LAST text-bearing assistant entry in the whole file, not "text since the matching
+// user line") correct — see the scoping note there. A pane MUST NEVER serve a second
+// turn, and a session MUST NEVER be reset with /clear and reused: either would put two
+// exchanges in one transcript and leak the earlier turn's text into the later turn's
+// answer. Nothing here reuses a pane; keep it that way.
+//
+// ── WHY IT'S WORTH MORE THAN THE BOOT TIME ──────────────────────────────────
+// Measured on this host (n=6 through OCP, Sonnet 4.6, --effort low): the cold path
+// spends ~1.23 s reaching the input bar, but ALSO ~2.9 s inside the first turn beyond
+// what claude itself reports as the turn duration — post-input-bar init that a pane
+// which has been idle for a few seconds has already finished. A warm pane recovers both.
+//
+// ── COST (bounded, and paid whether or not a request arrives) ───────────────
+// Each warm pane is a LIVE `claude` process (plus its tmux pane) sitting idle. Peak
+// process count is (pool size) + (OCP_TUI_MAX_CONCURRENT in-flight turns) + (panes
+// currently booting as replacements). Pool size is clamped to POOL_MAX_SIZE.
+//
+// Pure + injectable (bootPane / killPane / paneHealthy / now) so test-features.mjs can
+// assert acquire / miss / refill / TTL / reaper-exemption with no tmux and no claude.
+
+// Hard cap on OCP_TUI_POOL_SIZE. Each pane is an idle claude process; 4 is already a
+// lot of resident memory on a small host (a Pi serving a family) for zero in-flight work.
+export const POOL_MAX_SIZE = 4;
+
+// A warm pane older than this is dropped on acquire rather than handed out. The periodic
+// reap tick (server.mjs) drains the pool every 15 min anyway, so this only bites when
+// that tick kept getting skipped because the TUI path was never idle. Guards against
+// handing out a pane whose `claude` has been sitting so long it may have drifted
+// (auto-compaction prompts, an idle-disconnect banner, an expired in-pane token).
+export const POOL_MAX_AGE_MS = 10 * 60 * 1000;
+
+// Clamp the operator-supplied size into [0, POOL_MAX_SIZE]. A garbage value disables the
+// pool rather than guessing — an unparseable size must never silently boot 4 processes.
+export function resolvePoolSize(raw) {
+  const n = parseInt(raw, 10);
+  if (!Number.isFinite(n) || n <= 0) return 0;
+  return Math.min(n, POOL_MAX_SIZE);
+}
+
+export class TuiPanePool {
+  // size:       target number of warm panes (0 = disabled).
+  // maxAgeMs:   per-pane TTL (see POOL_MAX_AGE_MS).
+  // mintPane:   () => ({ sessionId, name }) — mints the identity of the NEXT pane. The POOL,
+  //             not the boot function, owns this: the tmux session springs into existence the
+  //             instant bootPane starts, so the pool must already know its NAME (see
+  //             _bootingPane below). Deriving the name from the sessionId also makes `tmux ls`
+  //             correlate to the transcript file.
+  // bootPane:   async (model, {sessionId, name}) => { name, sessionId, model, bootedAt } —
+  //             boots ONE pane under exactly that identity and resolves only once it is
+  //             input-ready; throws if it never becomes ready.
+  // killPane:   (name) => void — tmux kill-session. MUST be synchronous (see drain).
+  // paneHealthy:(name) => bool — pane still exists AND is still at its input bar.
+  constructor({ size, maxAgeMs = POOL_MAX_AGE_MS, mintPane, bootPane, killPane, paneHealthy, now = Date.now, log = () => {} }) {
+    this.size = Math.max(0, Math.min(parseInt(size, 10) || 0, POOL_MAX_SIZE));
+    // Fail fast at CONSTRUCTION, not at request time. refill() is called synchronously from
+    // the request path (runTuiTurn), so a missing collaborator would otherwise surface as a
+    // 500 on a live request instead of a loud error at boot.
+    if (this.size > 0) {
+      for (const [k, fn] of [["mintPane", mintPane], ["bootPane", bootPane], ["killPane", killPane], ["paneHealthy", paneHealthy]]) {
+        if (typeof fn !== "function") throw new TypeError(`TuiPanePool: ${k} must be a function`);
+      }
+    }
+    this.maxAgeMs = maxAgeMs;
+    this._mintPane = mintPane;
+    this._bootPane = bootPane;
+    this._killPane = killPane;
+    this._paneHealthy = paneHealthy;
+    this._now = now;
+    this._log = log;
+
+    this._panes = [];        // warm, available panes: { name, sessionId, model, bootedAt }
+    // The pane currently BOOTING, BY NAME ({sessionId, name, model}) — or null.
+    //
+    // WHY A NAME AND NOT A COUNT (this is a fixed bug, don't regress it): bootTuiPane creates
+    // the tmux session SYNCHRONOUSLY and only THEN waits up to POOL_BOOT_MS (20 s) for the
+    // input bar. So for up to 20 s there is a LIVE pooled tmux session. When the pool tracked
+    // only a count, it could not NAME that session, so:
+    //   - liveNames() could not spare it and the periodic reap sweep KILLED it (and
+    //     kill-server'd on top), leaving the pool empty with nothing scheduled and firing the
+    //     very tui_pool_boot_failed WARN operators are told to alert on; and
+    //   - drain() could not kill it, so on shutdown it ORPHANED a live authenticated `claude`
+    //     (the boot's .then that was supposed to clean up never runs — gracefulShutdown calls
+    //     process.exit in the same tick).
+    // Both are fixed by holding the identity here, before the session exists.
+    this._bootingPane = null;
+    // Generation counter. Bumped whenever an in-flight boot is CANCELLED (drain / model
+    // switch). A boot compares the generation it started under against the current one:
+    // if they differ, its pane was already killed by us and its settle is inert — in
+    // particular a rejection is a CANCELLATION, not an operator-visible boot failure.
+    this._gen = 0;
+    this._paused = false;    // true while drained; refill() is a no-op until resume()
+    this.warmModel = null;   // the model the pool currently warms — learned from traffic (see acquire)
+
+    this.hits = 0;           // requests served by a warm pane
+    this.misses = 0;         // requests that fell back to the cold path
+    this.boots = 0;          // panes successfully pre-booted
+    this.bootFailures = 0;   // pre-boots that genuinely never reached the input bar
+    this.cancelled = 0;      // in-flight boots WE killed (drain / model switch) — not failures
+    this.dropped = 0;        // panes discarded unused (unhealthy / expired / wrong model / drained /
+                             // cancelled — a cancelled in-flight boot also lands here via _drop)
+  }
+
+  get enabled() { return this.size > 0; }
+  get warm() { return this._panes.length; }
+  get booting() { return this._bootingPane ? 1 : 0; }
+
+  // The reaper's spare set: the EXACT names of every pane the pool currently owns and has NOT
+  // handed out — the warm ones AND the one currently booting (whose tmux session is already
+  // live; see _bootingPane). See the POOL/REAPER INVARIANT in lib/tui/session.mjs.
+  // Fail-safe by construction: a pane leaves this set the instant it is acquired, dropped, or
+  // cancelled, and if the pool is empty (or the process restarted) the set is empty — so an
+  // orphaned pooled pane looks exactly like any other stale session and IS reaped.
+  liveNames() {
+    const names = new Set(this._panes.map((p) => p.name));
+    if (this._bootingPane) names.add(this._bootingPane.name);
+    return names;
+  }
+
+  // Take a warm pane for `model`, or null (caller must fall back to the cold path — a MISS
+  // is always safe, never an error). Synchronous: paneHealthy is a cheap tmux capture.
+  //
+  // The pool warms the MOST RECENTLY REQUESTED model (`warmModel`). There is no boot-time
+  // pre-warm and no configured model: OCP cannot know which model the next caller wants, and
+  // pre-booting a process for a model nobody asks for is pure waste. Consequence, stated
+  // plainly: the FIRST request after start (and the first after a model switch) is always a
+  // MISS. The pool pays off for the steady repeat traffic it exists to serve.
+  acquire(model) {
+    if (!this.enabled) return null;
+
+    // Retarget on a model switch: --model is fixed at spawn, so panes for another model are
+    // useless. Drop them now (they are replaced by the next refill) rather than holding
+    // processes for a model that is no longer being asked for. This includes any pane
+    // currently BOOTING for the old model — its tmux session already exists, so leaving it to
+    // die on resolve would both hold a useless process and block the next refill (one boot at
+    // a time) for up to POOL_BOOT_MS.
+    if (model !== this.warmModel) {
+      for (const p of this._panes) { this._drop(p, "model_switch"); }
+      this._panes = [];
+      this._cancelBooting("model_switch");
+      this.warmModel = model;
+    }
+
+    while (this._panes.length) {
+      const p = this._panes.shift();
+      if (this._now() - p.bootedAt > this.maxAgeMs) { this._drop(p, "expired"); continue; }
+      if (!this._paneHealthy(p.name)) { this._drop(p, "unhealthy"); continue; }
+      this.hits++;
+      return p; // caller OWNS it now: it is out of the registry (so out of the spare set),
+                // and the caller's finally MUST kill it. Single-use — never returned here.
+    }
+    this.misses++;
+    return null;
+  }
+
+  // Bring the pool back up to `size` warm panes for `warmModel`. Fire-and-forget: never
+  // awaited on the request path and never throws into it.
+  //
+  // SLOT ACCOUNTING: a refill boot deliberately does NOT take a TuiSemaphore slot. Those
+  // slots bound concurrent *turns* (each up to the 120 s wallclock) and belong to real
+  // requests; charging a background pre-boot against them would let the pool starve the
+  // traffic it exists to speed up. It cannot leak a slot either, because it never holds one.
+  //
+  // SERIALIZED, ONE BOOT AT A TIME (and re-kicked on success until the pool is at target).
+  // An earlier version launched all `want` boots at once; live at size=2 that put two cold
+  // `claude` boots plus an in-flight turn on the CPU together, and a refill overran even the
+  // generous pool readiness cap (tui_pool_boot_failed). Booting sequentially keeps each boot
+  // near its uncontended ~1.2 s, bounds the CPU burst the pool can cause, and still has the
+  // replacement pane warm long before the next request arrives.
+  //
+  // A genuinely FAILED boot deliberately does NOT re-kick the chain — that is the backoff. A
+  // persistently failing boot (bad claude binary, no auth) would otherwise spin, respawning
+  // forever. The next natural trigger (the following request's refill, or the reap tick's
+  // resume) retries it. A CANCELLED boot is different: we killed it on purpose, nothing is
+  // wrong, and resume() is expected to start a fresh one immediately.
+  refill() {
+    if (!this.enabled || this._paused || !this.warmModel) return;
+    if (this._bootingPane) return;                       // one boot in flight at a time
+    if (this._panes.length >= this.size) return;         // already at target
+
+    const model = this.warmModel;
+    const gen = this._gen;
+    // Mint the identity BEFORE booting: bootPane creates the tmux session synchronously, so
+    // the pool must be able to name (and therefore spare, and kill) it from this moment on.
+    const ident = this._mintPane();
+    this._bootingPane = { ...ident, model };
+    let enlisted = false;
+    Promise.resolve()
+      .then(() => this._bootPane(model, ident))
+      .then((pane) => {
+        // The world may have moved while we booted. If our generation was cancelled, kill the
+        // pane here rather than ASSUMING _cancelBooting already did.
+        //
+        // Why not just `return`: _cancelBooting kills by name, but the tmux session only EXISTS
+        // once _bootPane has actually run — and _bootPane is queued on a microtask (above). A
+        // caller that does refill() and then drain() in the SAME synchronous block would have
+        // _cancelBooting find nothing to kill (a no-op), bump the generation, and then this
+        // microtask would create the session, boot it fine, and — under a bare `return` — walk
+        // away from a LIVE authenticated `claude` that nothing owns. That is M1b in a new costume.
+        // No current call site does that, so this is defense-in-depth, not a live bug — but ADR
+        // 0008 and the reap-tick comment in server.mjs both explicitly contemplate a boot-time
+        // pre-warm, which is exactly the shape that would reach it.
+        //
+        // Killing an already-dead session is a harmless no-op (_drop swallows it), so this is
+        // idempotent whether or not _cancelBooting got there first.
+        if (gen !== this._gen) { this._drop(pane, "cancelled_late"); return; }
+        // Otherwise: still possible the pool filled or retargeted without a cancellation.
+        if (this._paused || model !== this.warmModel || this._panes.length >= this.size) {
+          this._drop(pane, "stale_boot");
+          return;
+        }
+        this._panes.push(pane);
+        this.boots++;
+        enlisted = true;
+      })
+      .catch((e) => {
+        // A rejection from a CANCELLED generation is not a fault: it is almost always
+        // "tui_pane_not_ready", thrown because WE killed the pane out from under the boot.
+        // Counting it as a bootFailure would fire the exact WARN operators are told to alert
+        // on, for a completely healthy drain. Stay silent — _cancelBooting already counted
+        // this as a cancellation, so do NOT count it again here.
+        if (gen !== this._gen) return;
+        this.bootFailures++;
+        this._log("warn", "tui_pool_boot_failed", { model, error: e && e.message });
+      })
+      .finally(() => {
+        // ONLY the current generation's boot owns the booting slot. A stale settle must not
+        // clear a slot that a newer boot (started by resume()) already holds.
+        if (gen === this._gen) this._bootingPane = null;
+        if (enlisted) this.refill(); // continue toward target, still one at a time
+      });
+  }
+
+  // Kill the in-flight boot's pane, SYNCHRONOUSLY, and invalidate its generation. Returns 1
+  // if there was one, else 0. The tmux session already exists (bootPane created it before it
+  // started waiting for readiness), so this is a real kill, not a cancellation flag.
+  _cancelBooting(reason) {
+    if (!this._bootingPane) return 0;
+    this._gen++;                              // the in-flight boot's settle is now inert
+    this._drop(this._bootingPane, reason);    // synchronous kill-session
+    this._bootingPane = null;
+    this.cancelled++;
+    return 1;
+  }
+
+  // Kill every pane the pool owns — warm AND currently booting — and stop refilling. Returns
+  // how many were killed.
+  //
+  // Called (a) before the periodic reap sweep — reapStaleTuiSessions can only reap defunct
+  // `claude` zombies via kill-server, and kill-server is suppressed while any live pooled pane
+  // exists (including a booting one), so without this drain the pool would permanently disable
+  // zombie reaping; and (b) on graceful shutdown, so no pane outlives the process as an orphan.
+  //
+  // EVERY KILL HERE IS SYNCHRONOUS, and that is load-bearing. It is NOT safe to leave the
+  // booting pane to clean itself up on resolve: gracefulShutdown calls process.exit() in the
+  // same tick as this drain (TUI panes are children of the tmux SERVER, not of node, so
+  // node's activeProcesses set is empty on a TUI host and the "wait for children" path exits
+  // immediately). A .then()/.catch() scheduled here would never run, and the pane would
+  // survive as an orphaned, authenticated, idle `claude`.
+  drain() {
+    this._paused = true;
+    let n = this._panes.length;
+    for (const p of this._panes) this._drop(p, "drain");
+    this._panes = [];
+    n += this._cancelBooting("drain_booting");
+    return n;
+  }
+
+  // Undo drain() and start refilling again. Because drain() CANCELLED the in-flight boot
+  // (rather than leaving it pending), the booting slot is free and this really does start a
+  // fresh boot — the pool is never left empty with nothing scheduled.
+  resume() {
+    this._paused = false;
+    this.refill();
+  }
+
+  // /health surface (additive).
+  stats() {
+    return {
+      size: this.size,
+      warm: this._panes.length,
+      booting: this.booting,
+      model: this.warmModel,
+      hits: this.hits,
+      misses: this.misses,
+      boots: this.boots,
+      bootFailures: this.bootFailures,
+      cancelled: this.cancelled,
+      dropped: this.dropped,
+    };
+  }
+
+  _drop(pane, reason) {
+    this.dropped++;
+    try { this._killPane(pane.name); } catch { /* already gone */ }
+    // F5: every drop path (expired / unhealthy / model_switch / drain / cancelled_late /
+    // stale_boot) ends up here, and the reap tick drains the WHOLE pool on every tick — so
+    // without this, every warm pane's sink orphans in streamDir with no GC path (killPane only
+    // reaches the tmux session, never the pane's OWN files). Best-effort: pane.streamFile is
+    // undefined for a still-booting identity (the sink path is only known once bootPane
+    // resolves) and rmSync(force:true) is already a no-op on a missing file, so this never
+    // throws into the reaper regardless of which drop path got here.
+    if (pane.streamFile) {
+      try { rmSync(pane.streamFile, { force: true }); } catch { /* best-effort GC */ }
+    }
+    this._log("info", "tui_pool_pane_dropped", { name: pane.name, reason });
+  }
+}

lib/tui/semaphore.mjs

@@ -20,6 +20,15 @@
 //
 // Pure + importable so test-features.mjs can assert the bound directly (no server boot).
 
+// Thrown by acquire() when the caller-supplied AbortSignal fires before a slot was granted
+// (audit finding F2 — a client that disconnects while queued must never receive a slot; the
+// queue entry is spliced out, not just flagged, so `queued` accounting stays exact). Distinct
+// `name` lets callers (server.mjs acquireClaudeSlot) tell "client went away" apart from
+// "queue is full" without string-matching the message.
+export class SemaphoreAbortError extends Error {
+  constructor(message) { super(message); this.name = "SemaphoreAbortError"; }
+}
+
 export class TuiSemaphore {
   // limit: max concurrent slots. maxQueue: max waiters before run() rejects with backpressure.
   constructor(limit, { maxQueue } = {}) {
@@ -34,9 +43,30 @@ export class TuiSemaphore {
   get inflight() { return this._inflight; }
   get queued() { return this._waiters.length; }
 
+  // Runtime-adjust the concurrency limit (audit finding F1 — a PATCH /settings maxConcurrent
+  // change must actually take effect, not just be ignored until every currently-inflight task
+  // happens to finish). Lowering the limit is handled lazily by release() (see below) — it
+  // simply stops re-granting until inflight drains under the new, lower limit. Raising the
+  // limit has immediate headroom, so we wake as many queued waiters as now fit.
+  setLimit(limit) {
+    this.limit = Math.max(1, parseInt(limit, 10) || 1);
+    while (this._inflight < this.limit && this._waiters.length > 0) {
+      const next = this._waiters.shift();
+      this._inflight++;
+      next();
+    }
+  }
+
   // Acquire a slot. Resolves once a slot is free (immediately if under the limit, otherwise
   // when an in-flight task releases). Rejects synchronously-ish if the wait queue is full.
-  acquire() {
+  // `signal` (optional AbortSignal, F2) lets the caller cancel a QUEUED wait — e.g. wired to
+  // a client's socket "close" event so a request that disconnects before a slot is granted
+  // is removed from the queue instead of eventually being handed a slot for a dead socket.
+  // If `signal` is already aborted, reject immediately without ever touching the queue.
+  acquire(signal) {
+    if (signal?.aborted) {
+      return Promise.reject(new SemaphoreAbortError("acquire aborted before requesting a slot"));
+    }
     if (this._inflight < this.limit) {
       this._inflight++;
       return Promise.resolve();
@@ -46,24 +76,44 @@ export class TuiSemaphore {
         `tui_queue_full: TUI concurrency limit (${this.limit}) reached and wait queue ` +
         `(${this.maxQueue}) is full`));
     }
-    return new Promise((resolve) => { this._waiters.push(resolve); });
+    return new Promise((resolve, reject) => {
+      let waiter; // the FIFO entry — captured so onAbort can find + splice exactly this one
+      const onAbort = () => {
+        const idx = this._waiters.indexOf(waiter);
+        if (idx === -1) return; // already granted a slot (shifted out by release()/setLimit) — too late to cancel
+        this._waiters.splice(idx, 1); // remove, not just flag — keeps `queued` accounting exact
+        reject(new SemaphoreAbortError("acquire aborted while queued"));
+      };
+      waiter = () => {
+        signal?.removeEventListener("abort", onAbort);
+        resolve();
+      };
+      signal?.addEventListener("abort", onAbort, { once: true });
+      this._waiters.push(waiter);
+    });
   }
 
-  // Release a slot. If a waiter is queued, hand the slot directly to it (inflight stays
-  // constant across the handoff); otherwise decrement.
+  // Release a slot. Always frees the caller's own slot first, then re-grants it to the next
+  // waiter ONLY if the (post-decrement) inflight count is still under the current limit (F1
+  // fix). This is what makes a runtime-lowered limit actually bite: if the limit was lowered
+  // while over-subscribed, releases stop re-granting and inflight drains toward the new limit
+  // instead of a freed slot being handed straight back out at the old, higher occupancy.
   release() {
-    const next = this._waiters.shift();
-    if (next) {
-      next(); // the woken waiter already "owns" the slot — inflight unchanged
-    } else if (this._inflight > 0) {
-      this._inflight--;
+    if (this._inflight > 0) this._inflight--;
+    if (this._inflight < this.limit) {
+      const next = this._waiters.shift();
+      if (next) {
+        this._inflight++;
+        next();
+      }
     }
   }
 
   // Run fn() under one slot. Releases in a finally so a throw (PR-A's honesty gates,
   // wallclock truncation, paste-not-landed, tmux spawn failure) NEVER leaks a slot.
-  async run(fn) {
-    await this.acquire();
+  // `signal` (optional, F2) is forwarded to acquire() so a queued run() can be cancelled.
+  async run(fn, signal) {
+    await this.acquire(signal);
     try {
       return await fn();
     } finally {
@@ -89,7 +139,40 @@ export function recordTuiEntrypoint(tuiStats, observed, expectedMode = "cli") {
 // Build the additive /health `tui` block (ADR 0007 PR-B amendment). Pure: given the
 // config + live counters, returns the exact object embedded in /health. New fields only —
 // behaviour-preserving for existing /health consumers (grandfathered B.2 under ADR 0006).
-export function buildTuiHealthBlock({ enabled, entrypointMode, maxConcurrent }, tuiStats, semaphore) {
+//
+// `pool` (optional, warm pane pool — lib/tui/pool.mjs): a TuiPanePool, or null/undefined
+// when the pool is off (the default). Reported as `pool: null` when off so the block's
+// shape stays stable, and as the pool's stats (size / warm / hits / misses / …) when on —
+// the operator's window onto both the hit rate and the standing idle-process cost.
+//
+// Streaming fields (backlog #2, OCP_TUI_STREAM) are ADDITIVE too:
+//   streamEnabled       — is real (MessageDisplay-hook) SSE streaming on for TUI turns?
+//   streamTurns         — streamed turns ATTEMPTED, counted before the truncation/auth-banner
+//                         gates run (F6) — so a turn REFUSED by those gates still shows up
+//                         here, which is exactly the turn an operator most wants visible.
+//                         Counting only turns that survived the gates would silently exclude
+//                         a turn's worst-case outcome from its own denominator.
+//   streamDeltas        — MessageDisplay hook fires OBSERVED, including held-back ones (F6) —
+//                         NOT only the ones forwarded to a client. This is what makes
+//                         streamZeroDeltaTurns meaningful: a turn can have streamDeltas
+//                         incrementing while still emitting nothing to the client (fully held
+//                         back, e.g. a short answer), which is healthy, vs. a hook that fired
+//                         zero times at all, which is not (see streamZeroDeltaTurns).
+//   streamTopUps        — turns where the delta stream was a safe PREFIX of the transcript but
+//                         not equal to it; OCP topped up from the transcript and served T.
+//                         Benign but worth watching — a persistent rate means the hook is
+//                         losing fires.
+//   streamDivergences   — turns REFUSED because emitted bytes were not a prefix of the
+//                         transcript. THE field to alert on for CORRECTNESS: it means the hook
+//                         and the transcript disagreed and OCP chose to fail rather than serve
+//                         unverifiable text.
+//   streamZeroDeltaTurns — streamed turns where the hook fired ZERO times (F7). THE field to
+//                         alert on for AVAILABILITY: streamTopUps climbing is one fire dropped
+//                         here and there (benign); this climbing means the hook is not firing
+//                         AT ALL — e.g. `--settings` silently stopped registering it (a claude
+//                         version bump), or F3's truncated-script failure mode — and every
+//                         streamed turn is quietly degrading to fully-buffered with no error.
+export function buildTuiHealthBlock({ enabled, entrypointMode, maxConcurrent, streamEnabled = false }, tuiStats, semaphore, pool = null) {
   return {
     enabled,
     entrypointMode,                                  // cli | auto | off
@@ -98,5 +181,12 @@ export function buildTuiHealthBlock({ enabled, entrypointMode, maxConcurrent },
     inflight: semaphore.inflight,                    // current concurrent TUI turns
     queued: semaphore.queued,                        // turns waiting for a slot
     maxConcurrent,
+    pool: pool ? pool.stats() : null,                // warm pane pool, or null when disabled
+    streamEnabled,
+    streamTurns: tuiStats.streamTurns ?? 0,
+    streamDeltas: tuiStats.streamDeltas ?? 0,
+    streamTopUps: tuiStats.streamTopUps ?? 0,
+    streamDivergences: tuiStats.streamDivergences ?? 0,
+    streamZeroDeltaTurns: tuiStats.streamZeroDeltaTurns ?? 0,
   };
 }

lib/tui/session.mjs

@@ -14,15 +14,50 @@ import { mkdtempSync, writeFileSync, readFileSync, mkdirSync, existsSync, rmSync
 import { tmpdir } from "node:os";
 import { randomUUID } from "node:crypto";
 import { readTuiTranscript } from "./transcript.mjs";
+import { prepareStreamHook, streamFilePath, parseDeltaChunk } from "./stream.mjs";
+
+// F7 fix (audit finding, LOW): the prefix used to be a bare, host-wide constant
+// ("ocp-tui-"), so a SECOND OCP instance on the same host (e.g. a temporary
+// verification instance stood up alongside production — a real pattern used during
+// PR #144/#146 verification) would boot-reap and potentially kill-server the OTHER
+// instance's LIVE sessions: the coexistence guard below only ever spared foreign
+// PRODUCT prefixes (olp-tui-*), never a second ocp-tui-* instance on a different port.
+//
+// Fix: scope the prefix to the instance's own listen port. The port is the natural
+// stable per-instance discriminator on one host (two OCP instances cannot share a
+// port), so `ocp-tui-<port>-` uniquely namespaces this instance's sessions and makes
+// a same-host sibling OCP instance look exactly like a foreign product (olp-tui-*) to
+// the coexistence guard — its `ocp-tui-<otherPort>-*` sessions never match our own
+// prefix and are therefore never reaped/kill-server'd by us.
+//
+// LEGACY_SESSION_PREFIX / LEGACY_SESSION_NAME_RE describe the OLD bare-prefix shape
+// (pre-this-fix), retained ONLY for the boot-time legacy-zombie migration handled in
+// reapStaleTuiSessions (see comment there). No code path in this version ever CREATES
+// a legacy-shaped session name again — sessionPrefixForPort() is the only session-name
+// prefix constructor used going forward.
+export const LEGACY_SESSION_PREFIX = "ocp-tui-";
+// Exact legacy shape: LEGACY_SESSION_PREFIX + sessionId.slice(0, 8), where sessionId is
+// a randomUUID() — so the suffix is always exactly 8 lowercase hex characters with NO
+// further separator. The new port-scoped shape always inserts a "-" between the port
+// digits and the 8-hex suffix (see sessionPrefixForPort), so this regex can never match
+// a new-shape name: a new-shape suffix is `<port digits>-<8 hex>` (contains a literal
+// "-"), which `[0-9a-f]{8}$` anchored immediately after the prefix cannot satisfy.
+export const LEGACY_SESSION_NAME_RE = /^ocp-tui-[0-9a-f]{8}$/;
+
+// Build this instance's own session-name prefix, scoped by its listen port so a
+// second OCP instance on the same host (different port) is never mistaken for "ours".
+export function sessionPrefixForPort(port) {
+  return `ocp-tui-${port}-`;
+}
 
-export const SESSION_PREFIX = "ocp-tui-"; // per-proxy namespace (coexistence rule)
 const TMUX = process.env.OCP_TUI_TMUX_BIN || "tmux";
 
 const defaultTmux = (args, opts = {}) =>
   spawnSync(TMUX, args, { encoding: "utf8", ...opts });
 
-// Kill ONLY our own stale sessions. Scoped to SESSION_PREFIX so a co-hosted
-// OLP test instance's `olp-tui-*` sessions are never touched.
+// Kill ONLY our own stale sessions. Scoped to sessionPrefixForPort(port) so a co-hosted
+// OLP test instance's `olp-tui-*` sessions — AND a co-hosted second OCP instance's
+// `ocp-tui-<otherPort>-*` sessions — are never touched (F7 fix).
 //
 // Defunct-reaping (PI231 incident): the pane's `claude` process is a child of the
 // long-lived tmux SERVER daemon, NOT of the OCP node process — `tmux new-session -d`
@@ -36,29 +71,86 @@ const defaultTmux = (args, opts = {}) =>
 // merely re-signalling — is to stop the tmux server: when the server exits, the kernel
 // reparents its surviving children to init (PID 1), which reaps them immediately.
 //
-// So after killing our own sessions, if the server has NO sessions left of ANY prefix
-// (i.e. nothing we could disrupt — no co-hosted `olp-tui-*` or other instance), we
-// `kill-server` to flush the defunct backlog. If ANY non-ocp session remains we leave the
-// server running (coexistence rule, ADR 0007) and let the next boot/periodic sweep retry
-// once the server is otherwise idle.
-export function reapStaleTuiSessions({ tmux = defaultTmux } = {}) {
+// `port` (required) is this instance's own listen port (server.mjs's PORT / lib/constants.mjs
+// DEFAULT_PORT resolution) — the SPOT for "which sessions are ours."
+//
+// ── POOL/REAPER INVARIANT (warm pane pool — lib/tui/pool.mjs) ───────────────────────────
+// A warm pooled pane is one of OUR OWN `ocp-tui-<port>-*` sessions that is ALIVE AND IDLE
+// BY DESIGN — and the periodic sweep runs precisely when the instance is idle, i.e. exactly
+// when the pool is full. Without an exemption the sweep would kill every warm pane on every
+// tick (and kill-server on top). The exemption is `spare`: a set of EXACT session names the
+// caller declares live. Three properties, all load-bearing:
+//
+//   1. A LIVE POOLED PANE IS NEVER REAPED — INCLUDING ONE THAT IS STILL BOOTING. It is in
+//      `spare` (the pool's live registry), so it is skipped by name. The booting case is not
+//      a footnote, it is the one that bit us: bootTuiPane creates the tmux session
+//      SYNCHRONOUSLY and only then waits up to POOL_BOOT_MS for the input bar, so a pooled
+//      session can be live for ~20 s before its boot resolves. The pool therefore mints the
+//      pane's NAME up front and holds it in `_bootingPane`, so liveNames() can name — and
+//      spare — a session whose boot has not finished. (An earlier version tracked only a
+//      COUNT of in-flight boots; the sweep could not name that session and killed it.)
+//   2. A LEAKED/ORPHANED POOLED PANE IS STILL REAPED. Membership is by EXACT NAME from a
+//      live in-memory registry — NOT by "looks pooled" (name shape). A pane the pool no
+//      longer owns (handed out, dropped, cancelled, or left behind by a previous process
+//      generation — whose registry died with it) is absent from `spare` and is killed like
+//      any other stale session. Fail-safe: forgetting to pass `spare` reaps MORE, never less.
+//   3. KILL-SERVER NEVER KILLS A LIVE POOL PANE. A spared session suppresses kill-server
+//      exactly as a foreign session does (it is a live child of the tmux server). The
+//      consequence — that a permanently-full pool would permanently disable the defunct-
+//      zombie reaping that ONLY kill-server can do — is resolved in server.mjs by DRAINING
+//      the pool immediately before the sweep, so `spare` is empty on the normal tick and
+//      kill-server still fires. `spare` is the belt-and-braces: a reap call site that
+//      forgets to drain still cannot kill a live pane.
+//
+// `spare` (default: none) — iterable of session names, or a Set. Ignored when the pool is off.
+//
+// `includeLegacy` (default false): when true, sessions matching the exact OLD bare-prefix
+// shape (LEGACY_SESSION_NAME_RE) are ALSO treated as ours for kill-session purposes. This is
+// the boot-time legacy migration: an operator upgrading past this fix could otherwise be left
+// with orphaned bare-prefix zombie sessions from the PREVIOUS (pre-fix) process generation of
+// this SAME instance, since no live instance of the new version ever creates that shape again
+// — a legacy-shaped session found at boot is therefore presumed to be this instance's own
+// leftover, not a stranger's. Passed true ONLY from the one-time boot-reap call site in
+// server.mjs; the periodic idle-reap sweep does NOT set it, so a lingering legacy session
+// during steady-state is conservatively treated as foreign (correctly blocking kill-server)
+// rather than assumed to be ours on every 15-minute tick. Residual (accepted, documented):
+// if a genuinely-still-running PRE-FIX OCP instance is coexisting on the same host at the
+// exact moment a new instance boots, its live legacy-shaped session could be reaped — the
+// same class of residual risk the audit finding itself accepts ("no live instance of the new
+// version creates them"); this PR does not regress that scenario, it only removes the far
+// more common same-version collision (the actual F7 finding).
+export function reapStaleTuiSessions({ tmux = defaultTmux, port, includeLegacy = false, spare = null } = {}) {
   const r = tmux(["list-sessions", "-F", "#{session_name}"]);
   if (!r || r.status !== 0) return 0; // no tmux server / no sessions
   const names = String(r.stdout || "").split("\n").map((s) => s.trim()).filter(Boolean);
+  const ownPrefix = sessionPrefixForPort(port);
+  const spared = spare instanceof Set ? spare : new Set(spare || []);
   let killed = 0;
   let othersRemain = false;
+  let sparedLive = 0;
   for (const name of names) {
-    if (name.startsWith(SESSION_PREFIX)) {
+    // Property 1+2: exemption is by EXACT NAME from the pool's live registry. A pooled-
+    // LOOKING name that is not in the registry is an orphan and falls through to the
+    // normal kill path below.
+    if (spared.has(name)) { sparedLive++; continue; }
+    const isOwn = name.startsWith(ownPrefix);
+    const isLegacyOwn = includeLegacy && LEGACY_SESSION_NAME_RE.test(name);
+    if (isOwn || isLegacyOwn) {
       tmux(["kill-session", "-t", name]);
       killed++;
     } else {
-      othersRemain = true; // a session we do NOT own (e.g. olp-tui-*) — never kill-server
+      othersRemain = true; // a session we do NOT own (olp-tui-*, a sibling ocp-tui-<otherPort>-*,
+                            // or — outside includeLegacy — a legacy-shaped name) — never kill-server
     }
   }
   // Reap defunct `claude` zombies: safe ONLY when the server is now ours-only/empty.
   // kill-server is what actually reaps (server exit reparents survivors to init); a
   // per-session kill cannot, since node is not the zombies' parent.
-  if (!othersRemain) {
+  //
+  // Property 3: a SPARED session is a live child of this tmux server, so kill-server would
+  // kill it — it therefore suppresses kill-server exactly as a foreign session does. On the
+  // normal sweep the pool is drained first, so sparedLive is 0 and kill-server still fires.
+  if (!othersRemain && sparedLive === 0) {
     tmux(["kill-server"]);
   }
   return killed;
@@ -68,8 +160,18 @@ export function reapStaleTuiSessions({ tmux = defaultTmux } = {}) {
 
 // Boot + paste-settle timing. Conservative defaults validated on PI231; env-tunable.
 const BOOT_MS         = parseInt(process.env.OCP_TUI_BOOT_MS  || "4000", 10);   // max wait for input-ready
+// Readiness cap for a POOL pre-boot. Deliberately far more generous than BOOT_MS: BOOT_MS is
+// tight because a client is blocked on it, whereas a warm-pane boot happens in the background
+// with nobody waiting. Observed live at size=2: a refill booting alongside an in-flight turn
+// exceeded 4000 ms and was discarded (tui_pool_boot_failed), quietly costing hit rate for a
+// pane that was merely slow, not broken. Scales with OCP_TUI_BOOT_MS if an operator raises it.
+export const POOL_BOOT_MS = BOOT_MS * 5;
 const READY_POLL_MS   = parseInt(process.env.OCP_TUI_READY_POLL_MS || "400", 10); // readiness / paste-verify poll interval
 const PASTE_VERIFY_MS = parseInt(process.env.OCP_TUI_PASTE_VERIFY_MS || "5000", 10); // max wait for pasted prompt to render
+// Hook-sink drain interval when streaming. 100ms: the hook fires at BLOCK granularity
+// (~5-7 fires per answer, seconds apart), so a finer poll buys nothing and a coarser one
+// would add visible lag to the first delta. Cheap — one readFileSync of a small file.
+const STREAM_POLL_MS  = parseInt(process.env.OCP_TUI_STREAM_POLL_MS || "100", 10);
 
 const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
 
@@ -80,8 +182,13 @@ function tuiCapturePane(tmux, tmuxName) {
 }
 
 // True once claude's input bar is rendered and ready for keystrokes.
+// Two known ready-state footers across claude versions: the classic `? for shortcuts`
+// hint, and the `shift+tab to cycle` hint (part of "⏵⏵ bypass permissions on (shift+tab
+// to cycle)") that newer claude 2.1.x renders in its place. Match EITHER — a matcher that
+// only knew the classic string silently reported "never ready" on those builds, timing out
+// every boot (tui_pane_not_ready) and, with the warm pool on, failing every pre-boot.
 function tuiInputReady(pane) {
-  return /\? for shortcuts/.test(pane);
+  return /\? for shortcuts|shift\+tab to cycle/.test(pane);
 }
 
 // True once the pasted prompt has POSITIVELY landed in the input box. We only trust
@@ -251,7 +358,25 @@ export function prepareTuiHome(realHome, tuiHome, cwd, { envTokenMode = false }
 // A-PATH ONLY: built-in tools are left enabled (acceptable single-user). Deployment B
 // (guest keys) MUST additionally pass --tools "" per spec §5.2(2) as the credential
 // wall before this argv is reachable for owner_tier=guest — guard that in PR-3 wiring.
-export function buildTuiCmd(claudeBin, model, sessionId, ehome, entrypointMode) {
+//
+// `stream` (optional, OCP_TUI_STREAM): { file, settings } — when present, the pane gets
+//   (a) OCP_TUI_STREAM_FILE in its env  — read by the static MessageDisplay hook script to
+//       decide WHERE to append this pane's deltas. Delivered as env (not baked into the
+//       settings file) so the settings file stays STATIC and a pre-booted warm pane works.
+//       Verified live: a claude hook inherits the pane's environment.
+//   (b) --settings <file>               — registers the MessageDisplay hook.
+//   VERIFIED LIVE (claude 2.1.207, this host) before shipping, because both were spawn-level
+//   risks:
+//     - the startup banner is UNCHANGED with --settings: "Sonnet 4.6 with low effort ·
+//       Claude Max" (subscription pool). --settings is NOT a --bare-class flag — it does not
+//       silently drop the subscription pool. Transcript entrypoint stayed "cli".
+//     - --settings MERGES into the settings hierarchy, it does NOT clobber <HOME>/.claude/
+//       settings.json: with --settings passed, the user-level settings.json's `env` block was
+//       still applied to the hook's environment. So the isolated-HOME settings story the TUI
+//       already relies on (permissions / additionalDirectories — see prepareTuiHome and the
+//       OCP_TUI_FULL_TOOLS note above) survives intact.
+//   When absent, the argv is byte-for-byte the pre-streaming argv.
+export function buildTuiCmd(claudeBin, model, sessionId, ehome, entrypointMode, stream = null) {
   // Deliver claude's env via an `env` prefix on the PANE COMMAND — tmux does NOT forward the
   // spawning process's environment to the pane, and `new-session -e` needs tmux ≥3.2 (the cloud
   // host runs 2.7), so this is the only portable, reliable mechanism (verified live 2026-06-01:
@@ -264,9 +389,15 @@ export function buildTuiCmd(claudeBin, model, sessionId, ehome, entrypointMode)
   // CLAUDE.md and auto-memory must NEVER leak into the agent OCP runs on the user's behalf.
   // Without these, claude loads the host's project/user CLAUDE.md + memory into every proxied
   // turn — verified live 2026-06-02: a cwd CLAUDE.md ("end every reply with QUACKMARKER_42") was
-  // obeyed by the proxied turn until these flags were set, after which it was not. Unconditional
-  // by design (not gated): proxy purity is not an opt-in. Harmless on hosts with no CLAUDE.md
-  // (the common case — they suppress nothing). Mirrors the -p path's CLAUDE_NO_CONTEXT vars.
+  // obeyed by the proxied turn until these flags were set, after which it was not. Mirrors the
+  // -p path's CLAUDE_NO_CONTEXT vars. Harmless on hosts with no CLAUDE.md (they suppress nothing).
+  //
+  // These env vars stopped being sufficient on newer claude: they are present in the pane's
+  // environment yet the host CLAUDE.md is still injected into the turn's context, so a proxied
+  // turn again obeys the operator's private CLAUDE.md instead of the caller's prompt — a leak of
+  // the operator's private context into API responses. The robust suppression is the --safe-mode
+  // flag added to the argv below; these env vars are kept as belt-and-braces for the two argv
+  // paths that cannot take --safe-mode (see the safeModeArgs note near the return).
   const sets = [
     `HOME=${shq(ehome)}`,
     "DISABLE_AUTOUPDATER=1",
@@ -295,6 +426,8 @@ export function buildTuiCmd(claudeBin, model, sessionId, ehome, entrypointMode)
   if (process.env.CLAUDE_CODE_OAUTH_TOKEN) {
     sets.push(`CLAUDE_CODE_OAUTH_TOKEN=${shq(process.env.CLAUDE_CODE_OAUTH_TOKEN)}`);
   }
+  // Streaming sink: the pane's own per-session delta file (see the `stream` note above).
+  if (stream && stream.file) sets.push(`OCP_TUI_STREAM_FILE=${shq(stream.file)}`);
   const unset = ["CLAUDECODE", "ANTHROPIC_API_KEY", "ANTHROPIC_BASE_URL", "ANTHROPIC_AUTH_TOKEN"];
   if (entrypointMode === "cli") sets.push("CLAUDE_CODE_ENTRYPOINT=cli");
   else if (entrypointMode === "auto") unset.push("CLAUDE_CODE_ENTRYPOINT"); // let claude self-classify via TTY
@@ -326,44 +459,130 @@ export function buildTuiCmd(claudeBin, model, sessionId, ehome, entrypointMode)
   } else {
     toolArgs = ["--strict-mcp-config", "--disallowedTools", shq("mcp__*")];
   }
+
+  // Effort: pass --effort EXPLICITLY. Without it, the pane's claude inherits a
+  // HOME-dependent effortLevel — real-home mode inherits the operator's
+  // ~/.claude/settings.json (whatever they set for their own interactive use),
+  // env-token scratch mode inherits claude's built-in default (prepareTuiHome never
+  // writes effortLevel) — so latency silently depends on which HOME mode
+  // resolveTuiHome() picked AND on an unrelated operator setting. Pinning it here
+  // removes both. Measured (docs/plans/2026-07-13-tui-latency): explicit low cuts
+  // direct-spawn TTFT p50 10.35s → 6.17s (−40%) and collapses the spread ~15×;
+  // banner-verified to stay on the subscription pool (`· Claude Max`).
+  // OCP_TUI_EFFORT=inherit restores the pre-flag argv byte-for-byte (no --effort).
+  // An unknown value falls back to the default rather than reaching claude's argv:
+  // a typo'd --effort value must not risk a spawn-time usage error in the pane.
+  const EFFORT_LEVELS = ["low", "medium", "high", "xhigh", "max"]; // claude 2.1.207 --help
+  const effortRaw = (process.env.OCP_TUI_EFFORT || "low").trim().toLowerCase();
+  let effortArgs;
+  if (effortRaw === "inherit") {
+    effortArgs = [];
+  } else if (EFFORT_LEVELS.includes(effortRaw)) {
+    effortArgs = ["--effort", effortRaw];
+  } else {
+    console.error(`[tui] invalid OCP_TUI_EFFORT=${JSON.stringify(process.env.OCP_TUI_EFFORT)}; using "low" (valid: ${EFFORT_LEVELS.join("|")}, or "inherit" to omit the flag)`);
+    effortArgs = ["--effort", "low"];
+  }
+
+  // --settings registers the MessageDisplay hook. Omitted entirely when streaming is off,
+  // so the OFF argv is byte-for-byte the pre-streaming argv.
+  const settingsArgs = stream && stream.settings ? ["--settings", shq(stream.settings)] : [];
+
+  // --safe-mode: disable ALL customizations (host CLAUDE.md, skills, plugins, hooks, MCP
+  // servers, custom commands/agents, output styles, ...) while leaving auth, model selection,
+  // built-in tools, and permissions untouched (claude 2.1.216 --help). This is the robust
+  // replacement for the CLAUDE_CODE_DISABLE_CLAUDE_MDS / _AUTO_MEMORY env vars above, which no
+  // longer suppress the host CLAUDE.md on newer claude — so the operator's private CLAUDE.md
+  // could otherwise leak into a proxied API response. Unlike --bare, --safe-mode does NOT drop
+  // the interactive session off the subscription pool (it keeps auth + model selection), so the
+  // billing classifier stays cc_entrypoint=cli.
+  //
+  // NOT applied when the pane deliberately relies on a customization --safe-mode would strip,
+  // because those would break SILENTLY:
+  //   - streaming (settingsArgs present): the MessageDisplay hook is a HOOK, which --safe-mode
+  //     disables — every streamed turn would then fire zero deltas (tui.streamZeroDeltaTurns)
+  //     and fall back to buffered, defeating OCP_TUI_STREAM.
+  //   - OCP_TUI_FULL_TOOLS=1: grants an MCP / skills / agent surface (--mcp-config, ...) that
+  //     --safe-mode disables wholesale.
+  // Those two opt-in paths keep the env-var suppression above as their best-effort fallback.
+  const safeModeArgs =
+    settingsArgs.length === 0 && process.env.OCP_TUI_FULL_TOOLS !== "1"
+      ? ["--safe-mode"]
+      : [];
+
   return [
     envPrefix,
     shq(claudeBin),
+    ...safeModeArgs,
     "--model",          shq(model),
     "--session-id",     sessionId,
     ...toolArgs,
+    ...effortArgs,
+    ...settingsArgs,
   ].join(" ");
 }
 
-// Full per-request TUI lifecycle:
-//   1. Pre-trust the scratch cwd (no trust dialog will appear).
-//   2. Write prompt to a 0600 temp file (no shell injection from prompt content).
-//   3. Boot an interactive `claude` in a fresh tmux session in the scratch cwd; poll
-//      capture-pane until the `? for shortcuts` input bar appears (readiness-poll
-//      replaces the old blind boot sleep). BOOT_MS is the max wait, not a fixed delay.
-//   4. Paste the prompt via tmux load-buffer + paste-buffer -p (bracketed paste) —
-//      reliable for large multi-line prompts where send-keys -l is not (issue #130).
-//      Poll-verify the prompt landed in the input (placeholder gone / [Pasted text]);
-//      fast-fail with tui_paste_not_landed if it never lands (prevents the 120s
-//      wallclock "stuck typing" hang). Then submit with a SEPARATE Enter key event.
-//   5. Block on the native JSONL transcript (located by session-id) until terminal
-//      marker or wall-clock cap.
-//   6. Always teardown: kill session + rm temp dir (even on throw).
-// Returns { text, entrypoint } from readTuiTranscript (entrypoint is the billing-pool
-// classifier, e.g. "cli", or null if the transcript did not include a turn_duration).
-export async function runTuiTurn({
-  prompt,
-  model,
-  claudeBin,
-  home,
-  realHome,
-  cwd,
-  wallclockMs = 120000,
-  entrypointMode = "cli",
-  tmux = defaultTmux,
+// Is a pane alive AND still sitting at its input bar? Used by the warm pool to decide,
+// at hand-out time, whether a pre-booted pane is still usable (a dead/degraded pane must
+// become a MISS → cold path, never a hung turn). capture-pane exits non-zero when the
+// session no longer exists, so this covers "pane gone" and "pane not ready" in one call.
+export function tuiPaneHealthy(tmux, tmuxName) {
+  const r = tmux(["capture-pane", "-p", "-t", tmuxName]);
+  if (!r || r.status !== 0 || typeof r.stdout !== "string") return false;
+  return tuiInputReady(r.stdout);
+}
+
+// Pool pane names carry a "p" marker after the port-scoped prefix:
+//   turn pane: ocp-tui-<port>-<8hex>      (unchanged)
+//   pool pane: ocp-tui-<port>-p<8hex>
+// Purely for operator legibility (`tmux ls` shows which panes are warm). It is NOT the
+// reaper's exemption mechanism — that is the exact-name spare set (see the POOL/REAPER
+// INVARIANT above), so a pooled-LOOKING orphan is still reaped. Both shapes start with
+// sessionPrefixForPort(port), so both remain reapable as "ours", and neither can match
+// LEGACY_SESSION_NAME_RE.
+export function poolPaneName(port, sessionId) {
+  return sessionPrefixForPort(port) + "p" + sessionId.slice(0, 8);
+}
+
+// Boot ONE interactive `claude` pane and wait for its input bar. Shared by the cold
+// request path (runTuiTurn) and the warm pool (lib/tui/pool.mjs) so a pooled pane is
+// spawned with byte-for-byte the same argv, HOME, cwd and trust preparation as a
+// cold-booted one — the pool must not become a second, drifting spawn path.
+//
+// Each pane gets its OWN fresh randomUUID() --session-id, fixed at boot. That is what
+// keeps a pooled pane single-use-safe: its transcript holds exactly one exchange.
+//
+// requireReady: the cold path tolerates a readiness timeout (it falls through and lets
+// the paste-verify decide — pre-existing behaviour, unchanged). The POOL sets it, because
+// a pane that never reached its input bar is worthless as a warm pane and must not be
+// enlisted: throw, let the pool count a bootFailure, and leave the request path to
+// cold-boot as usual.
+// bootMs: max wait for the input bar. Defaults to BOOT_MS (the REQUEST path's cap, which is
+// deliberately tight — a client is blocked on it). The POOL passes POOL_BOOT_MS instead: a
+// background pre-boot has nobody waiting on it, and capping it at the request-path's 4 s
+// made real refills fail (observed live: a refill booting alongside an in-flight turn took
+// >4 s and was discarded, silently lowering the hit rate). Slow != broken for a pre-boot.
+// `sessionId` / `name` (both optional): the caller may supply the pane's identity instead of
+// letting bootTuiPane mint it. The POOL does, because it must know the tmux session's NAME
+// before this function runs — the session is created synchronously below, well before the
+// readiness wait returns, so a pool that only learned the name on resolve could neither spare
+// the session from the reaper nor kill it on shutdown. Supplying BOTH also keeps the name's
+// hex suffix equal to the session-id's, so `tmux ls` correlates to the transcript file.
+// `streamDir` (optional, OCP_TUI_STREAM): install claude's MessageDisplay hook on this pane.
+// Done HERE, at boot — not at turn time — and that is the whole reason streaming survives the
+// WARM POOL: the hook script + settings file are STATIC (one pair per streamDir), and the only
+// per-turn thing, the sink path, is derived from the pane's own --session-id, which is fixed
+// right here. So a pre-booted pane already carries its hook and its own sink and streams exactly
+// like a cold-booted one; nothing request-specific is ever baked into the spawn.
+export async function bootTuiPane({
+  model, claudeBin, home, realHome, cwd, port, entrypointMode = "cli",
+  tmux = defaultTmux, sessionId = null, name = null, requireReady = false, bootMs = BOOT_MS,
+  streamDir = null,
 }) {
-  const sessionId = randomUUID();
-  const tmuxName  = SESSION_PREFIX + sessionId.slice(0, 8);
+  const sid = sessionId || randomUUID();
+  // Port-scoped session name (F7 fix) — see sessionPrefixForPort / reapStaleTuiSessions
+  // for why this instance's own listen port is the namespace discriminator.
+  const tmuxName  = name || (sessionPrefixForPort(port) + sid.slice(0, 8));
   const ehome     = home || process.env.HOME;       // HOME claude runs under (scratch or real)
   const rhome     = realHome || process.env.HOME;   // real home (OAuth + onboarded config source)
 
@@ -380,10 +599,14 @@ export async function runTuiTurn({
   if (!existsSync(cwd)) mkdirSync(cwd, { recursive: true });
   prepareTuiHome(rhome, ehome, cwd, { envTokenMode });
 
-  // Write prompt to a temp file (mode 0600) so the content never touches argv.
-  const tmpDir     = mkdtempSync(`${tmpdir()}/ocp-tui-`);
-  const promptFile = `${tmpDir}/prompt.txt`;
-  writeFileSync(promptFile, prompt, { mode: 0o600 });
+  // Streaming sink for THIS pane (see the streamDir note above). rmSync first so a
+  // re-used session-id can never replay a previous turn's deltas.
+  let streamFile = null, streamSettings = null;
+  if (streamDir) {
+    streamFile     = streamFilePath(streamDir, sid);
+    streamSettings = prepareStreamHook(streamDir);
+    try { rmSync(streamFile, { force: true }); } catch { /* start from a fresh sink */ }
+  }
 
   // Minimal env for spawnSync (tmux itself). The pane's claude env comes exclusively
   // from the `env` prefix string built inside buildTuiCmd — tmux does NOT forward the
@@ -391,31 +614,148 @@ export async function runTuiTurn({
   const env = { ...process.env };
   env.HOME = ehome; // tmux needs HOME; all claude-specific vars go via buildTuiCmd prefix
 
+  // Boot the interactive session inside tmux, rooted at the scratch cwd.
+  // Capture the result: if tmux new-session fails (status !== 0) there is no PTY, no
+  // interactive spawn — abort BEFORE the boot wait rather than paste into a non-existent
+  // session or issue a billing request without a verified interactive context.
+  const spawnResult = tmux(
+    ["new-session", "-d", "-s", tmuxName, "-x", "220", "-y", "50", "-c", cwd,
+     buildTuiCmd(claudeBin, model, sid, ehome, entrypointMode,
+                 streamFile ? { file: streamFile, settings: streamSettings } : null)],
+    { env },
+  );
+  if (!spawnResult || spawnResult.status !== 0) {
+    throw new Error("tui_spawn_failed: tmux session not created");
+  }
+
+  // Wait until claude's input bar is actually ready (not a blind sleep).
+  // bootMs is the MAX readiness wait, not a fixed delay.
+  const ready = await pollUntil(() => tuiInputReady(tuiCapturePane(tmux, tmuxName)),
+    { timeoutMs: bootMs, intervalMs: READY_POLL_MS });
+  if (!ready) {
+    if (requireReady) {
+      try { tmux(["kill-session", "-t", tmuxName]); } catch { /* already gone */ }
+      throw new Error("tui_pane_not_ready: input bar did not appear within " + bootMs + "ms");
+    }
+    // Cold path (pre-existing behaviour): readiness timed out; rely on paste-verify.
+    console.error("[tui] input_not_ready", tmuxName);
+  }
+  return { name: tmuxName, sessionId: sid, model, ehome, streamFile, bootedAt: Date.now() };
+}
+
+// Full per-request TUI lifecycle:
+//   1. Take a WARM pane from the pool if one is available for this model (opt-in;
+//      OCP_TUI_POOL_SIZE=0 => always null => steps 2-3 below are exactly today's path).
+//      A pooled pane is SINGLE-USE: it already carries its own fresh --session-id, it
+//      serves this one turn, and it is killed in the finally like any other pane.
+//   2. On a MISS: pre-trust the scratch cwd, boot an interactive `claude` in a fresh tmux
+//      session in the scratch cwd, poll capture-pane until the input bar appears — see
+//      tuiInputReady for the ready-state footers matched (bootTuiPane). BOOT_MS is the max
+//      wait, not a fixed delay.
+//   3. Write prompt to a 0600 temp file (no shell injection from prompt content).
+//   4. Paste the prompt via tmux load-buffer + paste-buffer -p (bracketed paste) —
+//      reliable for large multi-line prompts where send-keys -l is not (issue #130).
+//      Poll-verify the prompt landed in the input (placeholder gone / [Pasted text]);
+//      fast-fail with tui_paste_not_landed if it never lands (prevents the 120s
+//      wallclock "stuck typing" hang). Then submit with a SEPARATE Enter key event.
+//   5. Block on the native JSONL transcript (located by THIS pane's session-id) until
+//      terminal marker or wall-clock cap.
+//   6. Always teardown: kill session + rm temp dir (even on throw), and kick a background
+//      pool refill so the next request finds a warm pane.
+// Returns { text, entrypoint } from readTuiTranscript (entrypoint is the billing-pool
+// classifier, e.g. "cli", or null if the transcript did not include a turn_duration).
+//
+// STREAMING (OCP_TUI_STREAM, default off). Pass `onDelta` and `streamDir`, and the pane's
+// MessageDisplay hook (installed by bootTuiPane; see lib/tui/stream.mjs) appends each raw
+// delta payload to the pane's own sink. This driver polls that sink and invokes onDelta(payload)
+// per fire while the turn is still generating. A WARM pane already carries its sink from boot
+// (pane.streamFile), so the pooled and cold paths stream identically.
+//
+// `streamDir` IS PASSED TO THE COLD BOOT UNCONDITIONALLY (not gated on `onDelta`) — F4 fix. The
+// spawn argv is this project's billing-classification surface: a caller with OCP_TUI_STREAM on
+// but THIS particular request non-streaming (stream:false) must still get the SAME argv whether
+// it lands on a pool HIT or a cold-boot MISS, because a pre-booted pool pane cannot know in
+// advance whether the request it will eventually serve wants streaming — it installs the hook
+// unconditionally whenever the pool is warming at all (see server.mjs's bootPane closure). Gating
+// the cold boot's hook install on `onDelta` made a stream:false request's argv depend on whether
+// it happened to hit the pool or miss it — the exact drift this surface cannot tolerate. Whether
+// the hook is actually POLLED is a separate, correctly-scoped decision: see `streaming` below,
+// gated on onDelta && streamFile, so a non-streaming turn never reads its own sink even though
+// the hook is running.
+//
+// The transcript stays AUTHORITATIVE regardless: it is still the terminal-turn signal, still the
+// source of the returned `text`, and still the input to the caller's honesty gates. The delta
+// stream is a low-latency MIRROR of it, never a replacement, and the caller asserts the two
+// agree. With onDelta AND streamDir both omitted, nothing here changes: no poll, no hook.
+//
+// `abortSignal` (optional): aborts the transcript wait, so a client that disconnects mid-turn
+// tears the pane down NOW (the finally below) instead of holding the pane — and therefore the
+// caller's semaphore slot — until the turn or the wallclock cap ends.
+export async function runTuiTurn({
+  prompt,
+  model,
+  claudeBin,
+  home,
+  realHome,
+  cwd,
+  port,
+  wallclockMs = 120000,
+  entrypointMode = "cli",
+  tmux = defaultTmux,
+  pool = null,       // TuiPanePool | null — null (default) === today's cold-boot-only path
+  onPane = null,     // optional observer: ({ warm }) => void, for logging/metrics
+  onDelta = null,    // (payload) => void — invoked per MessageDisplay hook fire, mid-turn
+  streamDir = null,  // hook sink dir, passed to the COLD boot UNCONDITIONALLY (F4 — see above);
+                      // a warm pane brings its own, fixed at its own boot
+  abortSignal = null,
+}) {
+  // 1. Warm pane, or cold boot. A MISS is never an error — it is exactly today's path.
+  let pane = pool ? pool.acquire(model) : null;
+  const warm = !!pane;
+  // Kick the refill IMMEDIATELY (not after the turn): the replacement pane then boots
+  // CONCURRENTLY with this turn and is warm by the time the next request arrives. Also
+  // runs on a MISS — acquire() has just retargeted the pool to this model, so the miss
+  // that cold-boots today warms the pool for the next caller. Fire-and-forget; it takes
+  // no TuiSemaphore slot (see pool.refill's SLOT ACCOUNTING note).
+  if (pool) pool.refill();
+  if (onPane) { try { onPane({ warm }); } catch { /* observer must never break a turn */ } }
+  if (!pane) {
+    // streamDir passed AS-IS (not gated on onDelta) — F4: see the STREAMING comment above.
+    pane = await bootTuiPane({ model, claudeBin, home, realHome, cwd, port, entrypointMode, tmux,
+                              streamDir });
+  }
+  const tmuxName  = pane.name;
+  const sessionId = pane.sessionId;   // THIS pane's own session-id — one session, one turn
+  const ehome     = pane.ehome || home || process.env.HOME;
+
+  // Streaming state is read off the PANE, not recomputed here — a warm pane fixed its sink at
+  // boot, and a cold one just did the same above. If the pool was booted WITHOUT a streamDir
+  // while onDelta is set, streamFile is null and the turn degrades to buffered: correct, just
+  // not fast. (server.mjs wires the same streamDir into both paths so that cannot happen.)
+  const streamFile = pane.streamFile || null;
+  const streaming  = !!(onDelta && streamFile);
+  const streamCursor = { consumed: 0 };
+  let streamStopped = false;
+  let pollTimer = null;
+  // Drain every complete line appended since the last drain. Never throws into the turn: a
+  // malformed line is skipped by parseDeltaChunk, and an onDelta that throws is contained.
+  const drainDeltas = () => {
+    if (!streaming) return;
+    let text;
+    try { text = readFileSync(streamFile, "utf8"); } catch { return; } // absent until the first fire
+    const { deltas, consumed } = parseDeltaChunk(text, streamCursor.consumed);
+    streamCursor.consumed = consumed;
+    for (const d of deltas) {
+      try { onDelta(d); } catch { /* a sink error must never abort the turn */ }
+    }
+  };
+
+  // Write prompt to a temp file (mode 0600) so the content never touches argv.
+  const tmpDir     = mkdtempSync(`${tmpdir()}/ocp-tui-`);
+  const promptFile = `${tmpDir}/prompt.txt`;
+  writeFileSync(promptFile, prompt, { mode: 0o600 });
+
   try {
-    // 1. Boot the interactive session inside tmux, rooted at the scratch cwd.
-    //    Capture the result: if tmux new-session fails (status !== 0) there is no
-    //    PTY, no interactive spawn — abort BEFORE the boot sleep rather than paste
-    //    into a non-existent session or issue a billing request without a verified
-    //    interactive context. The finally teardown is still harmless (kill-session
-    //    is a no-op when the session never existed).
-    const spawnResult = tmux(
-      ["new-session", "-d", "-s", tmuxName, "-x", "220", "-y", "50", "-c", cwd,
-       buildTuiCmd(claudeBin, model, sessionId, ehome, entrypointMode)],
-      { env },
-    );
-    if (!spawnResult || spawnResult.status !== 0) {
-      throw new Error("tui_spawn_failed: tmux session not created");
-    }
-
-    // 2. Wait until claude's input bar is actually ready (was: blind sleep(BOOT_MS)).
-    //    BOOT_MS is now the MAX readiness wait, not a fixed delay.
-    const ready = await pollUntil(() => tuiInputReady(tuiCapturePane(tmux, tmuxName)),
-      { timeoutMs: BOOT_MS, intervalMs: READY_POLL_MS });
-    if (!ready) {
-      // (readiness timed out; relying on paste-verify)
-      console.error("[tui] input_not_ready", tmuxName);
-    }
-
     // 3. Paste the prompt via a tmux PASTE BUFFER with bracketed paste (-p), NOT
     //    `send-keys -l`. send-keys of a large multi-line prompt is unreliable: the
     //    embedded newlines arrive as separate key events (effectively repeated Enter),
@@ -441,12 +781,37 @@ export async function runTuiTurn({
     // Submit (separate Enter key event).
     tmux(["send-keys", "-t", tmuxName, "Enter"]);
 
-    // 4. Block on the native transcript (resolved by session-id) until terminal.
-    //    Returns { text, entrypoint } from readTuiTranscript.
-    return await readTuiTranscript({ home: ehome, sessionId, wallclockMs });
+    // 5a. Streaming only: start polling the hook sink. Runs CONCURRENTLY with the
+    //     transcript wait below — the deltas are what make the answer visible while the
+    //     turn is still generating; the transcript is what makes it authoritative.
+    if (streaming) {
+      const loop = () => {
+        if (streamStopped) return;
+        drainDeltas();
+        pollTimer = setTimeout(loop, STREAM_POLL_MS);
+      };
+      pollTimer = setTimeout(loop, STREAM_POLL_MS);
+    }
+
+    // 5b. Block on the native transcript (resolved by THIS pane's session-id) until terminal.
+    //     Returns { text, entrypoint, truncated } from readTuiTranscript.
+    const result = await readTuiTranscript({ home: ehome, sessionId, wallclockMs, abortSignal });
+
+    // 5c. FINAL drain. The terminal marker can land between two poll ticks, so the last
+    //     delta(s) may still be unread — without this the tail would be missing from the
+    //     stream and every turn would need a transcript top-up.
+    streamStopped = true;
+    if (pollTimer) clearTimeout(pollTimer);
+    drainDeltas();
+    return result;
   } finally {
-    // 5. Teardown — always, even on throw.
+    // 6. Teardown — always, even on throw (including an abortSignal disconnect, which is
+    //    exactly why the pane cannot outlive a client that walked away). A pooled pane is
+    //    torn down here exactly like a cold-booted one: SINGLE-USE, never returned (pool.mjs).
+    streamStopped = true;
+    if (pollTimer) clearTimeout(pollTimer);
     try { tmux(["kill-session", "-t", tmuxName]); } catch { /* already gone */ }
     try { rmSync(tmpDir, { recursive: true, force: true }); } catch { /* best effort */ }
+    if (streamFile) { try { rmSync(streamFile, { force: true }); } catch { /* best effort */ } }
   }
 }

lib/tui/stream.mjs

@@ -0,0 +1,288 @@
+// TUI-mode real SSE streaming — the `MessageDisplay` hook sink.
+//
+// WHAT THIS IS. `claude` fires a **MessageDisplay** hook per rendered block of the
+// assistant's reply, handing the hook the RAW MARKDOWN SOURCE of an incremental
+// `delta` on stdin. Registered via `--settings` on the ordinary interactive TUI spawn
+// (NO -p, NO --bare — the billing pool is untouched), it is the only byte-faithful
+// incremental source the interactive CLI exposes. Everything here consumes that hook
+// surface AS EMITTED — forwarding, not inventing.
+//
+// ALIGNMENT.md: **Class B**. We consume claude's own hook payload and re-emit it in the
+// OpenAI chat/completions streaming shapes OCP already speaks (ADR 0006). There is no
+// `cli.js` citation because no `cli.js` function is being mirrored: the TUI spawn is
+// OCP-owned surface (ADR 0007), and the hook payload is claude's own published contract.
+//
+// THE VERIFIED CONTRACT (docs/plans/2026-07-13-tui-latency/streaming-spike.md, and
+// independently reproduced on claude 2.1.207 / sonnet-4-6 / banner `· Claude Max`):
+//
+//   payload (stdin, one JSON object per fire):
+//     { hook_event_name:"MessageDisplay", session_id, transcript_path, prompt_id, cwd,
+//       turn_id, message_id, index, final, delta }
+//
+//   - deltas carry the raw markdown source (`## `, `**`, ```javascript all present)
+//   - concat(deltas of one message) === T, byte-exactly   (T = extractLatestAssistantText)
+//   - T.startsWith(concat(deltas[0..n])) at EVERY n       (prefix-stable)
+//   - block-level granularity (~5-7 fires per answer), NOT token-level
+//   - only `text` blocks fire it — thinking blocks are excluded (what OCP wants)
+//
+// ⚠️ THE HOOK IS SYNCHRONOUS. The hook's source sets `forceSyncExecution: true` —
+// `claude` BLOCKS on every fire. The hook script must therefore write and exit, doing
+// NO work inline. Measured cost of the script below: p50 7.2 ms / p90 14.7 ms per fire,
+// i.e. ~50 ms added blocking across a whole ~7-delta turn against a 6-10 s turn. That is
+// noise, so a plain append is the right sink — a FIFO would be faster on paper but a FIFO
+// blocks its writer until a reader attaches, which would hand `claude` a way to hang.
+//
+// WARM-POOL COMPATIBILITY (load-bearing — a warm pane pool is a separate in-flight PR).
+// The hook script and the settings file are BOTH STATIC: one copy per stream dir, written
+// once, never per-request. The per-turn destination is carried in the PANE'S OWN ENV as
+// `OCP_TUI_STREAM_FILE` (verified live: a hook inherits the pane's environment), and the
+// path is derived from the session-id — which for a pre-booted pane is fixed at BOOT.
+// Nothing about a request is baked into the settings file at spawn time, so a pane booted
+// before its request arrives streams exactly the same way.
+import { writeFileSync, mkdirSync, renameSync } from "node:fs";
+import { detectTuiUpstreamError } from "./transcript.mjs";
+
+// Default holdback before the first byte is released to the client. See TuiDeltaAssembler.
+export const DEFAULT_HOLDBACK_CHARS = 100;
+
+// Resolve OCP_TUI_STREAM_HOLDBACK to a SAFE value. The whole C-1 auth-banner guarantee rests
+// on the holdback being at least the default banner detector's max message length — which is
+// exactly DEFAULT_HOLDBACK_CHARS. So this is a FLOOR, not a hint: a smaller value (or a NaN
+// typo like "unlimited"/"5MB") would let a real banner fragment release before the terminal
+// detector could classify the whole message, silently reopening the leak the assembler exists
+// to prevent. The env var's own doc says "Only raise it"; this enforces that instead of trusting
+// it. Returns { value, clamped } so the caller can warn when it had to clamp — a silent floor is
+// less honest than a noticed one.
+export function resolveStreamHoldback(raw, floor = DEFAULT_HOLDBACK_CHARS) {
+  const parsed = parseInt(raw ?? "", 10);
+  if (!Number.isFinite(parsed)) return { value: floor, clamped: raw != null && String(raw).trim() !== "" };
+  if (parsed < floor) return { value: floor, clamped: true };
+  return { value: parsed, clamped: false };
+}
+
+// The hook script. POSIX sh, no interpreter startup beyond /bin/sh, one fork (`cat`).
+//
+//   - `printf` is a shell BUILTIN in sh/dash/bash, so the newline costs no fork.
+//   - the `{ cat; printf '\n'; } >>` group opens the file ONCE and appends both writes
+//     through the same O_APPEND fd, so a payload and its terminator can never be split
+//     by another writer. (They never race anyway: one file per pane, and MessageDisplay
+//     is synchronous within a pane.)
+//   - a payload JSON can never contain a literal newline — JSON.stringify escapes them —
+//     so "one line == one payload" holds, and a torn write is always a trailing partial
+//     line, which parseDeltaChunk() leaves unconsumed until it completes.
+//   - NO OCP_TUI_STREAM_FILE (e.g. a pane booted with streaming off, or any other claude
+//     session that happens to load this settings file) => swallow stdin and exit 0. The
+//     hook must NEVER fail or block: claude is waiting on it.
+export const HOOK_SCRIPT = `#!/bin/sh
+# OCP TUI streaming sink — claude fires this per MessageDisplay block and BLOCKS on it.
+# Write and exit. Never do work here.
+[ -n "\$OCP_TUI_STREAM_FILE" ] || exec cat >/dev/null
+{ cat; printf '\\n'; } >> "\$OCP_TUI_STREAM_FILE"
+`;
+
+// The --settings payload registering the hook. Static: no per-request data.
+export function buildStreamSettings(hookScriptPath) {
+  return { hooks: { MessageDisplay: [{ hooks: [{ type: "command", command: hookScriptPath }] }] } };
+}
+
+export const hookScriptPath   = (streamDir) => `${streamDir}/md-hook.sh`;
+export const streamSettingsPath = (streamDir) => `${streamDir}/settings.json`;
+// One file per session-id. For a pre-booted (warm) pane the session-id is fixed at boot,
+// so this path is knowable at boot — which is what keeps the pool compatible.
+export const streamFilePath   = (streamDir, sessionId) => `${streamDir}/${sessionId}.jsonl`;
+
+// Atomic write: temp file + rename (same-directory, same-filesystem, so rename is atomic on
+// POSIX). A process killed mid-`writeFileSync` leaves the TEMP file half-written, never the
+// real path — `path` always names either the old complete content or the new complete
+// content, never a torn one. That matters specifically for md-hook.sh: it is SYNCHRONOUS
+// (claude blocks on every fire), so a truncated script would still pass `existsSync`, still
+// get exec'd, and fail/hang on every single MessageDisplay fire with no operator-visible
+// symptom short of streaming going silently dead (F7's streamZeroDeltaTurns is the backstop
+// for exactly that). Mirrors ensureTuiCwdTrusted's tmp+renameSync pattern in session.mjs.
+function writeFileAtomic(path, content, mode) {
+  const tmp = `${path}.${process.pid}.tmp`;
+  writeFileSync(tmp, content, { mode });
+  renameSync(tmp, path);
+}
+
+// Write the static hook script + settings file into `streamDir`. UNCONDITIONAL, not
+// write-if-missing: these files persist across OCP restarts at `streamDir`, so a host that
+// booted once under an older version and never had its stream dir cleared would otherwise be
+// silently stuck on a stale HOOK_SCRIPT / buildStreamSettings() forever — no future OCP
+// upgrade could ever reach it. Safe to call every boot: the content is static (no per-request
+// data), so a same-content rewrite is the overwhelmingly common case and costs two tiny
+// atomic writes, not a per-turn expense. Returns the settings path to hand to `claude
+// --settings`.
+export function prepareStreamHook(streamDir) {
+  mkdirSync(streamDir, { recursive: true });
+  const script = hookScriptPath(streamDir);
+  const settings = streamSettingsPath(streamDir);
+  writeFileAtomic(script, HOOK_SCRIPT, 0o700);
+  writeFileAtomic(settings, JSON.stringify(buildStreamSettings(script), null, 2), 0o600);
+  return settings;
+}
+
+// Parse newly-appended sink lines. `consumed` is the number of COMPLETE lines already
+// taken; only lines terminated by "\n" are complete, so a payload caught mid-write stays
+// unconsumed until its terminator lands. Returns the fresh MessageDisplay payloads plus
+// the new consumed count. Pure — the caller owns the cursor.
+export function parseDeltaChunk(text, consumed = 0) {
+  const lines = String(text ?? "").split("\n");
+  const complete = lines.slice(0, -1); // the tail after the last "\n" is a partial line
+  const deltas = [];
+  for (const line of complete.slice(consumed)) {
+    const t = line.trim();
+    if (!t) continue;
+    try {
+      const o = JSON.parse(t);
+      if (o && o.hook_event_name === "MessageDisplay" && typeof o.delta === "string") deltas.push(o);
+    } catch { /* not ours / not parseable — skip, never throw into the request path */ }
+  }
+  return { deltas, consumed: complete.length };
+}
+
+// ── The assembler: hook deltas → client bytes, with the honesty gates intact ──
+//
+// Two jobs, both load-bearing.
+//
+// 1. THE AUTH-BANNER HOLDBACK (C-1 / issue #133 must survive streaming).
+//    The interactive CLI renders an auth failure as ordinary assistant TEXT — so an
+//    expired-credential turn fires MessageDisplay with the BANNER as its delta, and a
+//    naive forwarder would stream "Please run /login · API Error: 401 …" to the client as
+//    a normal answer, exactly the silent-error case C-1 exists to prevent.
+//    detectTuiUpstreamError() classifies a WHOLE message, so it cannot be run per-delta.
+//    Instead we HOLD BACK the first `holdbackChars` characters. The default detector only
+//    ever fires on a message of <= 100 chars (TUI_ERR_MAX_LEN — real banners are 69 and 73),
+//    so once the TRIMMED accumulation EXCEEDS 100 chars the final text cannot be a banner by
+//    that detector's own length rule, and releasing is safe. An answer that never exceeds the
+//    holdback is simply delivered whole at terminal — i.e. exactly today's buffered
+//    behaviour, gates and all.
+//    THE GUARANTEE HAS TWO HALVES, both required — neither alone is sufficient:
+//      (i)  Nothing is emitted for a message until its trimmed accumulation exceeds the
+//           detector's max banner length. This is what keeps the FIRST message of a turn
+//           safe: a banner-length message can never clear the holdback.
+//      (ii) Once a message boundary follows an emit (`restartedAfterEmit`), push() stops
+//           emitting ENTIRELY for the rest of the turn — a SECOND message (e.g. an
+//           auth-failure banner rendered mid-turn, after tool-using prose already streamed)
+//           gets zero bytes forwarded, not just a fresh holdback of its own. finalize() then
+//           refuses the whole turn (SSE error frame, no cache) precisely because the first
+//           message's bytes are unretractable and unverifiable against T. Without this half,
+//           (i) alone only protects the FIRST message per turn — see F1.
+//    ⚠️ Soundness is w.r.t. the DEFAULT detector. An operator who REPLACES it via
+//    CLAUDE_TUI_ERROR_PATTERNS with a pattern that can match a longer message must raise
+//    OCP_TUI_STREAM_HOLDBACK past their longest banner; server.mjs warns at boot. That is the
+//    one case (i) does not cover — (ii) still applies regardless. Even past both, the
+//    terminal gate still refuses to cache a banner and still ends the stream on an SSE error
+//    frame rather than finish_reason:"stop" — the holdback is the first of two layers, not
+//    the only one.
+//
+// 2. MESSAGE SCOPING (keeps `concat === T` the RIGHT assertion).
+//    The transcript's T is extractLatestAssistantText() — the LAST text-bearing assistant
+//    entry, not every assistant entry. A tool-using turn therefore has TWO messages
+//    (prose → tool_use → answer) and T is only the second. So the assembler scopes to the
+//    CURRENT message_id: when a new message_id appears and NOTHING has been emitted yet,
+//    the held text is DISCARDED — the transcript is about to discard it too, so this keeps
+//    us byte-identical to the buffered path instead of streaming prose the buffered path
+//    would have dropped. When a new message_id appears AFTER we have already emitted, the
+//    bytes are gone and cannot be retracted: finalize() then reports !ok and the caller
+//    fails the turn loudly (SSE error frame, no cache, counted on /health). Fail-loud is
+//    the correct posture — a proxy that silently serves text the transcript disagrees with
+//    is the exact class of bug ALIGNMENT.md exists to prevent.
+// Sentinel for "no message seen yet". Deliberately not null/undefined — see the constructor.
+const NO_MESSAGE_YET = Symbol("no-message-yet");
+
+export class TuiDeltaAssembler {
+  constructor({ holdbackChars = DEFAULT_HOLDBACK_CHARS, detectError = detectTuiUpstreamError } = {}) {
+    this.holdbackChars = holdbackChars;
+    this.detectError = detectError;
+    this.emitted = "";   // bytes ALREADY written to the client — unretractable
+    this.pending = "";   // held back, not yet written
+    this.released = false;
+    // NOT null: a payload may legitimately carry message_id === null, and if the sentinel were
+    // also null the FIRST such payload would compare equal to it, register no boundary, and
+    // leave `messages` at 0 — which used to disarm the restartedAfterEmit guard below entirely.
+    // A unique object is === to nothing a JSON payload can produce, so the first fire ALWAYS
+    // registers as message 1, whatever its message_id is (or isn't).
+    this.messageId = NO_MESSAGE_YET;
+    this.deltas = 0;     // hook fires seen
+    this.messages = 0;   // distinct message_ids seen
+    this.restartedAfterEmit = false;
+  }
+
+  // All hook bytes for the CURRENT message (emitted + still held).
+  get full() { return this.emitted + this.pending; }
+
+  // Feed one MessageDisplay payload. Returns the text to emit NOW, or null (held back).
+  push(payload) {
+    const delta = payload && typeof payload.delta === "string" ? payload.delta : "";
+    const mid = payload ? payload.message_id : null;
+    if (mid !== this.messageId) {
+      this.messageId = mid;
+      this.messages++;
+      if (this.emitted === "") {
+        this.pending = "";        // safe: the transcript will drop this message too
+      } else {
+        // A boundary while bytes are ALREADY out is unrecoverable, full stop — the count of
+        // messages seen so far is irrelevant. The old `else if (this.messages > 1)` guard was
+        // the sole reason a null-message_id first payload could disarm F1: it left `messages`
+        // at 0, so the real boundary evaluated 1 > 1 === false and never armed. The invariant
+        // is "a boundary occurred while emitted !== ''", and that is exactly what this says.
+        this.restartedAfterEmit = true; // unrecoverable — finalize() will refuse the turn
+      }
+    }
+    this.deltas++;
+    // F1: once a message boundary has followed an emit, the turn is ALREADY unrecoverable —
+    // finalize() will refuse it (see restartedAfterEmit above). `this.released` stays true
+    // from the FIRST message's release and, uncorrected, lets every later message's deltas
+    // stream straight through unfiltered — exactly the auth-banner-mid-turn leak this class
+    // exists to prevent. Stop emitting HERE, permanently, for the rest of the turn: there is
+    // nothing left to gain from continuing to forward bytes for a turn that will be refused,
+    // and every byte forwarded now is one more the client cannot be told to un-see.
+    if (this.restartedAfterEmit) return null;
+    if (!delta) return null;
+
+    if (this.released) {
+      this.emitted += delta;
+      return delta;
+    }
+    this.pending += delta;
+    // Release only once the TRIMMED accumulation is past the banner detector's reach.
+    // detectTuiUpstreamError() trims before measuring length (TUI_ERR_MAX_LEN is a trimmed-
+    // length bound), so gating release on the UNTRIMMED pending.length let a run of >
+    // holdbackChars whitespace trim down to "" — detectError("") sees nothing to classify,
+    // returns null, and release fires with the holdback never having actually screened
+    // anything. Trimming here keeps both sides of the check talking about the same string.
+    if (this.pending.trim().length > this.holdbackChars && this.detectError(this.pending) == null) {
+      const out = this.pending;
+      this.pending = "";
+      this.released = true;
+      this.emitted += out;
+      return out;
+    }
+    return null;
+  }
+
+  // Reconcile against the AUTHORITATIVE transcript text T. Call only AFTER the truncation
+  // and auth-banner gates have passed. Returns:
+  //   { ok:true,  tail, exact }  — tail is the remaining text to emit (may be ""). `exact`
+  //                               is concat(deltas) === T; when false we still serve exactly
+  //                               T, having topped up from the transcript, and the caller
+  //                               counts a topUp.
+  //   { ok:false, ... }          — what we already emitted is NOT a prefix of T. The client
+  //                               holds bytes the transcript disagrees with; the caller must
+  //                               NOT cache and must end the stream on an SSE error frame.
+  finalize(T) {
+    const text = typeof T === "string" ? T : "";
+    const full = this.full;
+    if (!text.startsWith(this.emitted)) {
+      return { ok: false, tail: null, exact: false, emitted: this.emitted.length, transcript: text.length };
+    }
+    return {
+      ok: true,
+      tail: text.slice(this.emitted.length),
+      exact: full === text,
+      emitted: this.emitted.length,
+      transcript: text.length,
+    };
+  }
+}

lib/tui/transcript.mjs

@@ -73,6 +73,17 @@ export function isTerminalLine(obj) {
 // transcript holding one logical exchange). If a future warm-pool ever reuses a
 // session WITHOUT a fresh session-id / clear, earlier-turn text could leak — that
 // author must add user-line scoping here. See spec §7.2.
+//
+// STATUS (warm pool, lib/tui/pool.mjs — the "future warm-pool" this note anticipated):
+// the pool does NOT reuse sessions, so the precondition above still holds and no
+// user-line scoping was added. Each pooled pane is booted with its OWN fresh
+// randomUUID() --session-id (bootTuiPane) and is SINGLE-USE: it serves exactly one turn
+// and is then killed and replaced. One session still means one logical exchange, so the
+// last assistant entry is still that request's answer.
+// The warning therefore stands UNCHANGED for anyone who later wants a pane to serve a
+// SECOND turn (or to reset one with /clear and reuse it): that is a leak, and it needs
+// user-line scoping HERE before it can be safe. Do not relax pool.mjs's single-use rule
+// without doing that work first.
 export function extractLatestAssistantText(events) {
   let text = "";
   for (const ev of events) {
@@ -256,11 +267,21 @@ export function detectTuiUpstreamError(text, patternsRaw = process.env.CLAUDE_TU
 // Resolution: pass an explicit `transcriptPath` (used by unit tests), OR pass
 // `home` + `sessionId` to resolve by glob each poll (production) — the transcript
 // file does not exist until the turn starts, so resolution happens inside the loop.
-export async function readTuiTranscript({ transcriptPath: p, home, sessionId, wallclockMs = 120000, pollMs = 250 }) {
+// `abortSignal` (optional): when it fires, stop waiting and throw TuiAbortError. The one
+// caller that passes it is the STREAMING TUI path, which ties it to the client's socket:
+// a client that disconnects mid-turn should not leave the pane running (and the caller's
+// concurrency slot held) until the turn or the 120s cap ends. runTuiTurn's finally does the
+// teardown. Omitted => the loop is byte-for-byte the pre-streaming loop.
+export async function readTuiTranscript({ transcriptPath: p, home, sessionId, wallclockMs = 120000, pollMs = 250, abortSignal = null }) {
   const deadline = Date.now() + wallclockMs;
   let lastText = "";
   let lastEntrypoint = null;
   while (Date.now() < deadline) {
+    if (abortSignal && abortSignal.aborted) {
+      const err = new Error("tui_aborted: client disconnected before the turn completed");
+      err.name = "TuiAbortError";
+      throw err;
+    }
     const resolved = p || findTranscriptPath(home, sessionId);
     if (resolved && existsSync(resolved)) {
       const events = parseTranscriptLines(readFileSync(resolved, "utf8"));

models.json

@@ -2,13 +2,21 @@
   "$schema": "./models.schema.json",
   "version": 1,
   "models": [
+    {
+      "id": "claude-opus-5",
+      "displayName": "Claude Opus 5",
+      "openclawName": "Claude Opus 5 (via CLI)",
+      "reasoning": true,
+      "contextWindow": 200000,
+      "maxTokens": 64000
+    },
     {
       "id": "claude-opus-4-8",
       "displayName": "Claude Opus 4.8",
       "openclawName": "Claude Opus 4.8 (via CLI)",
       "reasoning": true,
       "contextWindow": 200000,
-      "maxTokens": 16384
+      "maxTokens": 64000
     },
     {
       "id": "claude-opus-4-7",
@@ -16,7 +24,7 @@
       "openclawName": "Claude Opus 4.7 (via CLI)",
       "reasoning": true,
       "contextWindow": 200000,
-      "maxTokens": 16384
+      "maxTokens": 64000
     },
     {
       "id": "claude-opus-4-6",
@@ -24,7 +32,15 @@
       "openclawName": "Claude Opus 4.6 (via CLI)",
       "reasoning": true,
       "contextWindow": 200000,
-      "maxTokens": 16384
+      "maxTokens": 64000
+    },
+    {
+      "id": "claude-sonnet-5",
+      "displayName": "Claude Sonnet 5",
+      "openclawName": "Claude Sonnet 5 (via CLI)",
+      "reasoning": true,
+      "contextWindow": 200000,
+      "maxTokens": 64000
     },
     {
       "id": "claude-sonnet-4-6",
@@ -32,7 +48,7 @@
       "openclawName": "Claude Sonnet 4.6 (via CLI)",
       "reasoning": true,
       "contextWindow": 200000,
-      "maxTokens": 16384
+      "maxTokens": 32000
     },
     {
       "id": "claude-haiku-4-5-20251001",
@@ -40,12 +56,12 @@
       "openclawName": "Claude Haiku 4.5 (via CLI)",
       "reasoning": false,
       "contextWindow": 200000,
-      "maxTokens": 8192
+      "maxTokens": 32000
     }
   ],
   "aliases": {
-    "opus": "claude-opus-4-8",
-    "sonnet": "claude-sonnet-4-6",
+    "opus": "claude-opus-5",
+    "sonnet": "claude-sonnet-5",
     "haiku": "claude-haiku-4-5-20251001"
   },
   "legacyAliases": {

models.schema.json

@@ -0,0 +1,65 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://github.com/dtzp555-max/ocp/blob/main/models.schema.json",
+  "title": "OCP models.json",
+  "description": "Schema for models.json, the single source of truth for model metadata (ADR 0003). Enforced in CI by test-features.mjs, which validates models.json against this file using the repo's own validateJsonSchema (lib/structured-output.mjs) — no new dependency. ⚠ ONLY A SUBSET OF DRAFT 2020-12 IS ENFORCED: type, required, const, enum, additionalProperties (boolean and schema forms), items, minItems, maxItems, anyOf/allOf/oneOf, $ref, nullable. Anything else — minimum, maxLength, pattern, uniqueItems, propertyNames, not — is SILENTLY IGNORED by that validator, so adding one here buys nothing and misleads the next reader. Add such a constraint as a test in test-features.mjs instead. NOTE: referential integrity (every aliases/legacyAliases target must exist as a models[].id) is likewise not expressible here and is covered by separate tests.",
+  "type": "object",
+  "required": ["version", "models", "aliases", "legacyAliases"],
+  "additionalProperties": false,
+  "properties": {
+    "$schema": {
+      "type": "string",
+      "description": "Relative path to this file. Editors use it for completion; CI does not read it."
+    },
+    "version": {
+      "const": 1,
+      "description": "Schema version of this document. Bump only alongside a shape change and an ADR."
+    },
+    "models": {
+      "type": "array",
+      "minItems": 1,
+      "description": "Every model OCP advertises on /v1/models, newest first. Order is load-bearing: ocp-connect uses model_ids[0] as a primary fallback.",
+      "items": {
+        "type": "object",
+        "required": ["id", "displayName", "openclawName", "reasoning", "contextWindow", "maxTokens"],
+        "additionalProperties": false,
+        "properties": {
+          "id": {
+            "type": "string",
+            "description": "Canonical CLI model id, passed verbatim to `claude --model`. Must match the id in the compiled CLI registry."
+          },
+          "displayName": {
+            "type": "string",
+            "description": "Human label. Also used as the OpenClaw agent alias by scripts/sync-openclaw.mjs."
+          },
+          "openclawName": {
+            "type": "string",
+            "description": "Label written into OpenClaw's model registry."
+          },
+          "reasoning": {
+            "type": "boolean",
+            "description": "Whether OpenClaw should treat the model as reasoning-capable."
+          },
+          "contextWindow": {
+            "type": "integer",
+            "description": "Advertised context window. GLOBAL side effect: MAX_PROMPT_CHARS derives from max(contextWindow) x 3 across ALL entries (lib/prompt.mjs derivePromptCharBudget), so raising this on ONE model raises the truncation ceiling for EVERY model. See ADR 0009. Also feeds OpenClaw's compaction budget."
+          },
+          "maxTokens": {
+            "type": "integer",
+            "description": "Advertised output cap. OCP does not enforce it; it is propagated to OpenClaw (via setup.mjs / scripts/sync-openclaw.mjs) where it bounds request and compaction budgets."
+          }
+        }
+      }
+    },
+    "aliases": {
+      "type": "object",
+      "description": "Short name -> canonical models[].id. Client-addressable, so a repoint changes routing for every request using the alias. Cache keys resolve these before hashing (server.mjs cacheModel).",
+      "additionalProperties": { "type": "string" }
+    },
+    "legacyAliases": {
+      "type": "object",
+      "description": "Retired ids kept resolvable for backward compatibility -> canonical models[].id. Also client-addressable and also resolved in cache keys.",
+      "additionalProperties": { "type": "string" }
+    }
+  }
+}

ocp

@@ -622,7 +622,12 @@ cmd_restart() {
       self_r="${BASH_SOURCE[0]}"
       while [[ -L "$self_r" ]]; do self_r="$(readlink "$self_r")"; done
       script_dir="$(cd "$(dirname "$self_r")" && pwd)"
-      DISABLE_AUTOUPDATER=1 nohup node "$script_dir/server.mjs" >> "$HOME/.ocp/logs/proxy.log" 2>&1 &
+      # env -u strips test-only key-store redirection vars (A4): if the invoking shell had
+      # NODE_ENV=test + OCP_DIR_OVERRIDE exported (e.g. from a debugging session), this manual
+      # fallback would otherwise inherit them and start the daemon against a scratch/empty key
+      # store — a silent auth outage in AUTH_MODE=multi. The plist/systemd paths strip these via
+      # plist-merge's NEVER_PRESERVE; this covers the one direct-launch path OCP controls.
+      DISABLE_AUTOUPDATER=1 env -u NODE_ENV -u OCP_DIR_OVERRIDE nohup node "$script_dir/server.mjs" >> "$HOME/.ocp/logs/proxy.log" 2>&1 &
     fi
     sleep 3
     if curl -sf --max-time 5 "$PROXY/health" > /dev/null 2>&1; then

ocp-connect

@@ -122,11 +122,32 @@ provider = {
     "models": []
 }
 
-# Model metadata mapping (prefix match for versioned IDs like claude-haiku-4-5-20251001)
+# Model metadata mapping. Prefix match on the model FAMILY (claude-opus / -sonnet /
+# -haiku), not a pinned version. A version-pinned prefix like "claude-sonnet-4"
+# silently misses "claude-sonnet-5" and falls through to the non-reasoning /
+# 8k-output default (PR #152 review) — every future Sonnet/Opus/Haiku bump would
+# re-trip it. Family prefixes classify any versioned ID correctly with no per-model
+# edit. (ADR 0003: models.json is the SPOT for model existence; /v1/models does not
+# expose reasoning/maxTokens, so family classification stays here.)
+# maxTokens values are the floor of the CLI registry's max_output_tokens.default over the
+# family members CURRENTLY IN models.json (opus 64000; sonnet 32000, because sonnet-4-6 is
+# 32000 while sonnet-5 is 64000; haiku 32000) — NOT over the whole registry family. The
+# registry also holds claude-opus-4-0 / -4-5 at 32000, which this opus row would over-advertise
+# 2x. What keeps that unreachable is NOT "the ids come from models.json" — the ids reaching
+# get_model_meta have two sources that are not this checkout: /v1/models is fetched from a
+# user-supplied base_url, so it is the REMOTE server's models.json (possibly a different OCP
+# version or a fork), and on a JSON parse failure the ids come from the hardcoded list a few
+# lines above instead. The invariant that actually holds is that no released OCP has ever
+# served a legacy opus id. Adding one to models.json breaks it and this row must drop to 32000.
+# Family prefixes cannot distinguish versions, so a floor is used deliberately:
+# under-advertising caps a client lower than the model allows, which is safe, whereas
+# over-advertising would promise capacity a member of the family does not have.
+# models.json is authoritative per ADR 0003; this table only exists because /v1/models does
+# not expose maxTokens.
 model_meta = {
-    "claude-opus-4": {"name": "Claude Opus (OCP)", "reasoning": True, "maxTokens": 16384},
-    "claude-sonnet-4": {"name": "Claude Sonnet (OCP)", "reasoning": True, "maxTokens": 16384},
-    "claude-haiku-4": {"name": "Claude Haiku (OCP)", "reasoning": False, "maxTokens": 8192},
+    "claude-opus": {"name": "Claude Opus (OCP)", "reasoning": True, "maxTokens": 64000},
+    "claude-sonnet": {"name": "Claude Sonnet (OCP)", "reasoning": True, "maxTokens": 32000},
+    "claude-haiku": {"name": "Claude Haiku (OCP)", "reasoning": False, "maxTokens": 32000},
 }
 
 def get_model_meta(mid):
@@ -134,6 +155,13 @@ def get_model_meta(mid):
     for prefix, meta in sorted(model_meta.items(), key=lambda x: -len(x[0])):
         if mid.startswith(prefix):
             return meta
+    # Unknown id. 8192 is the GLOBAL MINIMUM max_output_tokens.default across all 17 records in
+    # the CLI 2.1.220 registry (distinct values: 8192 / 32000 / 64000). The two records holding
+    # it — claude-3-5-haiku and claude-3-5-sonnet, both default:8192, upper:8192 — are precisely
+    # the ones that reach this line, because "claude-3-5-*" matches none of the family prefixes
+    # above. Raising this to a family-table value would over-advertise those two by 4x. Keep 8192:
+    # under-advertising caps a client lower than a model allows, which is safe; over-advertising
+    # promises capacity that does not exist.
     return {"name": mid + " (OCP)", "reasoning": False, "maxTokens": 8192}
 
 for mid in model_ids:
@@ -178,11 +206,11 @@ config.setdefault("agents", {})
 config["agents"].setdefault("defaults", {})
 config["agents"]["defaults"].setdefault("models", {})
 
-# Build alias map (prefix match)
+# Build alias map (family prefix match — version-agnostic, see model_meta note)
 alias_prefixes = {
-    "claude-opus-4": "Claude Opus",
-    "claude-sonnet-4": "Claude Sonnet",
-    "claude-haiku-4": "Claude Haiku",
+    "claude-opus": "Claude Opus",
+    "claude-sonnet": "Claude Sonnet",
+    "claude-haiku": "Claude Haiku",
 }
 
 for mid in model_ids:
@@ -196,8 +224,13 @@ for mid in model_ids:
 
 # Handle primary/backup
 if priority == "1":
-    # OCP as primary — pick the best model (prefer sonnet for daily use)
-    primary_model = provider_name + "/claude-sonnet-4-6" if "claude-sonnet-4-6" in model_ids else provider_name + "/" + model_ids[0]
+    # OCP as primary — pick the best model (prefer the latest Sonnet for daily use,
+    # tracking the `sonnet` alias default in models.json; fall back across versions).
+    _sonnet_pref = ["claude-sonnet-5", "claude-sonnet-4-6"]
+    primary_model = next(
+        (provider_name + "/" + m for m in _sonnet_pref if m in model_ids),
+        provider_name + "/" + model_ids[0],
+    )
     config["agents"]["defaults"].setdefault("model", {})
     config["agents"]["defaults"]["model"]["primary"] = primary_model
     # Keep existing fallbacks

package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-claude-proxy",
-  "version": "3.21.0",
+  "version": "3.25.0",
   "description": "OCP (Open Claude Proxy) — use your Claude Pro/Max subscription as an OpenAI-compatible API for any IDE. Works with Cline, OpenCode, Aider, Continue.dev, OpenClaw, and more.",
   "type": "module",
   "bin": {

scripts/doctor.mjs

@@ -59,6 +59,16 @@ export async function runDoctor(opts = {}) {
   // of recommending a downgrade against a stale hardcoded value.
   let latestVersion = opts.mockLatest;
   if (!latestVersion) {
+    // Issue #173: `git show origin/main:...` reads the LOCALLY CACHED remote ref. Without a
+    // fetch first, a machine that hasn't pulled since the last release sees latest == current
+    // and reports noop — new releases were invisible everywhere except the machine that cut
+    // the tag (live repro: Oracle VM, 2026-07-17). Fetch before comparing; on failure
+    // (offline, auth, timeout) fall through to the cached ref — the pre-existing behavior.
+    if (!opts.skipNetwork) {
+      try {
+        execSync(`git -C ${ocpDir} fetch --tags --quiet`, { stdio: ["pipe", "pipe", "pipe"], timeout: 15000 });
+      } catch { /* offline → compare against cached origin/main, as before */ }
+    }
     try {
       const out = execSync(`git -C ${ocpDir} show origin/main:package.json 2>/dev/null`, { stdio: ["pipe", "pipe", "pipe"] }).toString();
       const remotePkg = JSON.parse(out);

scripts/lib/plist-merge.mjs

@@ -8,6 +8,19 @@
 //
 // No new dependencies — regex-based, plist <key>X</key><string>Y</string> shape
 // is stable enough for our hand-written templates in setup.mjs.
+//
+// SECURITY DENYLIST (A4): keys that must NEVER be carried into a service unit, even when a
+// prior unit already contained them. OCP's key store honors OCP_DIR_OVERRIDE only when
+// NODE_ENV === "test" (keys.mjs). If BOTH somehow reached a daemon's environment, the server
+// would open a scratch/empty key store instead of ~/.ocp/ocp.db — in AUTH_MODE=multi a silent
+// total auth outage. The preservation rule below ("keys only in EXISTING are kept verbatim")
+// is exactly a vector for that: a unit that once carried these test-only vars would otherwise
+// survive every setup re-run. So we strip them from the preserved set unconditionally. This is
+// defense-in-depth: setup.mjs's own template never injects them, so the only way they enter is
+// preservation, and this closes it. (The residual path — a hand-rolled `node server.mjs` with
+// both vars exported — is out of any launcher's reach; keys.mjs's loud "NOT the default" log is
+// the backstop there.)
+export const NEVER_PRESERVE = new Set(["NODE_ENV", "OCP_DIR_OVERRIDE"]);
 
 // Note: setup.mjs XML-escapes all injected values before writing (via xmlEscape()),
 // so raw `<` / `>` / `&` never appear in plist <string> bodies — the [^<]* regex below is safe.
@@ -36,7 +49,7 @@ export function mergePlistEnv(existing, template) {
 
   const preserved = {};
   for (const [k, v] of Object.entries(existingEnv)) {
-    if (!KNOWN.has(k)) preserved[k] = v;
+    if (!KNOWN.has(k) && !NEVER_PRESERVE.has(k)) preserved[k] = v;
   }
   if (Object.keys(preserved).length === 0) return template;
 
@@ -72,7 +85,7 @@ export function mergeSystemdEnv(existing, template) {
   const KNOWN = new Set(Object.keys(templateEnv));
 
   const preservedLines = Object.entries(existingEnv)
-    .filter(([k]) => !KNOWN.has(k))
+    .filter(([k]) => !KNOWN.has(k) && !NEVER_PRESERVE.has(k))
     .map(([k, v]) => `Environment=${k}=${v}`);
   if (preservedLines.length === 0) return template;
 

scripts/lib/snapshot.mjs

@@ -2,7 +2,7 @@ import { mkdirSync, writeFileSync, readFileSync, copyFileSync, existsSync, readd
 import { join } from "node:path";
 
 export function writeSnapshot({ homeDir, fromCommit, fromVersion, toVersion, extraFiles = [] }) {
-  const ts = new Date().toISOString().replace(/\.\d+Z$/, "Z");
+  const ts = formatSnapshotTimestamp(new Date());
   const root = join(homeDir, ".ocp", `upgrade-snapshot-${ts}`);
   mkdirSync(root, { recursive: true });
 
@@ -48,7 +48,10 @@ export function listSnapshots(homeDir) {
   return readdirSync(root)
     .filter(name => name.startsWith("upgrade-snapshot-"))
     .map(name => ({ name, path: join(root, name), mtime: statSync(join(root, name)).mtimeMs }))
-    .sort((a, b) => a.name.localeCompare(b.name));
+    .sort((a, b) => {
+      const chronological = parseSnapshotTimestamp(a.name) - parseSnapshotTimestamp(b.name);
+      return chronological || a.name.localeCompare(b.name);
+    });
 }
 
 /**
@@ -107,9 +110,21 @@ export function gcSnapshots(homeDir, opts = {}) {
 }
 
 function parseSnapshotTimestamp(name) {
+  // Both legacy ISO names and Windows-safe names are supported.
   // upgrade-snapshot-2026-05-11T08:30:00Z → epoch ms
+  // upgrade-snapshot-2026-05-11T08-30-00Z → epoch ms
   const m = name.match(/upgrade-snapshot-(.+)$/);
   if (!m) return 0;
-  const t = Date.parse(m[1]);
-  return Number.isFinite(t) ? t : 0;
+  const raw = m[1];
+  const t = Date.parse(raw);
+  if (Number.isFinite(t)) return t;
+  const iso = raw.replace(/(T\d{2})-(\d{2})-(\d{2})Z$/, "$1:$2:$3Z");
+  const portable = Date.parse(iso);
+  return Number.isFinite(portable) ? portable : 0;
+}
+
+function formatSnapshotTimestamp(date) {
+  // Windows forbids ':' in directory names. Replacing only the time separators
+  // preserves chronological lexical order and keeps the timestamp readable.
+  return date.toISOString().replace(/\.\d+Z$/, "Z").replace(/:/g, "-");
 }

scripts/upgrade.mjs

@@ -17,6 +17,20 @@ import { existsSync, copyFileSync } from "node:fs";
 import { writeSnapshot, listSnapshots, readSnapshot, gcSnapshots } from "./lib/snapshot.mjs";
 import { DEFAULT_PORT } from "../lib/constants.mjs";
 
+// Post-flight acceptance predicate (issue #173). A health probe passes ONLY when the server
+// is authed AND actually serving the TARGET version. auth.ok alone is not enough: a stale
+// process holding the port answers auth.ok=true while still running the OLD code — exactly
+// what a nohup-fallback orphan did on 2026-07-17 (upgrade "succeeded", /health kept serving
+// 3.21.1). Comparing /health.version to the checkout target catches orphan-holds-port,
+// restart-didn't-take, and wrong-unit-restarted alike. `target` tolerates a leading "v"
+// (doctor reports "v3.22.1"; /health reports "3.22.1"); an empty/unknown target degrades to
+// the old auth-only check rather than blocking an otherwise-good upgrade.
+export function postFlightOk(body, target) {
+  if (body?.auth?.ok !== true) return false;
+  const want = String(target || "").replace(/^v/, "");
+  return !want || body?.version === want;
+}
+
 export async function runUpgrade(opts = {}) {
   const dryRun = !!opts.dryRun;
   const yes = !!opts.yes;
@@ -137,16 +151,22 @@ async function runFullUpgrade({ doctor, opts }) {
     if (!opts.mockExec) {
       const port = process.env.CLAUDE_PROXY_PORT || String(DEFAULT_PORT);
       let ok = false;
+      let lastSeen = null;
       for (let i = 0; i < 10; i++) {
         try {
           const out = execSync(`curl -sf --max-time 2 http://127.0.0.1:${port}/health`).toString();
           const body = JSON.parse(out);
-          if (body.auth?.ok === true) { ok = true; break; }
+          lastSeen = body.version;
+          if (postFlightOk(body, doctor.latest_version)) { ok = true; break; }
         } catch { /* retry */ }
         await new Promise(r => setTimeout(r, 1000));
       }
       if (!ok) {
-        phases.push({ name: "post-flight", status: "fail", message: "health did not return auth.ok=true within 10s" });
+        phases.push({
+          name: "post-flight", status: "fail",
+          message: `health did not return auth.ok=true AND version=${doctor.latest_version} within 10s`
+            + (lastSeen ? ` (last saw version=${lastSeen} — a stale process may still hold the port; check \`ss -ltnp\` / \`lsof -i\`)` : ""),
+        });
         throw new Error("post-flight failed");
       }
       execSync(`curl -sf --max-time 3 http://127.0.0.1:${port}/v1/models > /dev/null`);

setup.mjs

@@ -334,7 +334,7 @@ if (OPENCLAW_PRESENT) {
     `║  Aider / OpenClaw) at:                                       ║`,
     `║    http://${BIND_ADDRESS}:${String(PORT)}/v1${" ".repeat(Math.max(0, 47 - BIND_ADDRESS.length - String(PORT).length))}║`,
     `║                                                              ║`,
-    `║  See README § "Client Setup" for per-IDE instructions.       ║`,
+    `║  See docs/lan-mode.md for per-IDE client setup.              ║`,
     `║                                                              ║`,
   );
 }
@@ -390,7 +390,9 @@ if (!DRY_RUN) {
   // and "ocp-proxy" keeps the proxy invisible to that heuristic.
   const OCP_HOME = join(HOME, ".ocp");
   const ocpLogsDir = join(OCP_HOME, "logs");
-  if (!existsSync(ocpLogsDir)) mkdirSync(ocpLogsDir, { recursive: true });
+  // mode 0700: with `recursive`, this call can create ~/.ocp ITSELF on a fresh install, and
+  // without an explicit mode that parent lands at the umask default (world-listable 0755).
+  if (!existsSync(ocpLogsDir)) mkdirSync(ocpLogsDir, { recursive: true, mode: 0o700 });
 
   // Uninstall legacy service names if present (upgrade path)
   if (platform === "darwin") {

test-env.mjs

@@ -0,0 +1,26 @@
+// Imported FIRST by test-features.mjs, before keys.mjs, so this runs before anything can open
+// the key store. ESM hoists imports and evaluates them in order, so a `process.env.X = ...`
+// statement in the test's own body would run too late — hence a separate module.
+//
+// Why this exists: `npm test` used to write real, UNREVOKED api_keys rows into the operator's
+// live ~/.ocp/ocp.db (the same database the running server reads) — two per run, unbounded.
+// It also made the suite racy: two concurrent runs (e.g. review worktrees) shared one file, so
+// `listKeys()` could miss "test-user-1" and the `in` check would throw on undefined.
+import { mkdtempSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+
+export const TEST_OCP_DIR = mkdtempSync(join(tmpdir(), "ocp-test-"));
+
+// BOTH are required. keys.mjs honors OCP_DIR_OVERRIDE only when NODE_ENV === "test", so neither
+// var alone redirects anything — a stray OCP_DIR_OVERRIDE in a production env is inert without
+// NODE_ENV=test alongside it. (A daemon OCP launches never carries either: the service units and
+// the `ocp` restart fallback strip both — see plist-merge NEVER_PRESERVE / keys.mjs's comment.)
+process.env.NODE_ENV = "test";
+process.env.OCP_DIR_OVERRIDE = TEST_OCP_DIR;
+
+// Remove the scratch store on exit. Without this the fix would trade unbounded growth in
+// ~/.ocp/ocp.db for unbounded growth in $TMPDIR — better, but still litter.
+process.on("exit", () => {
+  try { rmSync(TEST_OCP_DIR, { recursive: true, force: true }); } catch { /* best effort */ }
+});