Files
ocp/docs/adr/0007-tui-interactive-mode.md
T
1b324968f4 feat(tui): real SSE streaming via claude's MessageDisplay hook (OCP_TUI_STREAM, default off) (#159)
* feat(tui): real SSE streaming via claude's MessageDisplay hook (OCP_TUI_STREAM, default off)

Backlog #2. TUI-mode `stream:true` turns can now emit real SSE `delta.content` chunks as
`claude` generates them, instead of buffering the turn and replaying it with
streamStringAsSSE. Opt-in: with OCP_TUI_STREAM unset/0 the spawn argv, the SSE bytes and
the cache behaviour are byte-for-byte unchanged (asserted by test).

This PR does NOT mirror any cli.js function, so no `cli.js:NNNN` citation applies, and per
CLAUDE.md's hard requirement #1 that is stated explicitly here rather than left implicit:

  - We consume claude's OWN `MessageDisplay` hook surface AS EMITTED — forwarding, not
    inventing. No new endpoint, no fabricated protocol, no new field.
  - The TUI spawn is OCP-owned surface: ADR 0007 owns it, not cli.js.
  - The SSE wire shapes are the OpenAI chat/completions streaming spec, adopted by ADR 0006.
    Every frame emitted here (role chunk, content-delta chunk, stop chunk, `[DONE]`, and the
    post-header {error:{message,type}} frame) is COPIED from callClaudeStreaming, the -p path.

/health gains additive fields only (streamEnabled + 4 counters) — same grandfathered B.2
rationale as the existing tui block (ADR 0006). Existing keys are untouched.

`claude` fires MessageDisplay per rendered block, handing the hook the RAW MARKDOWN SOURCE
of an incremental delta on stdin. The hook is registered with `--settings` on the ordinary
interactive spawn (no -p, no --bare) — verified to leave the billing pool alone.

Sink: a static sh hook script appends each payload to `<streamDir>/<session_id>.jsonl`; OCP
polls that file and forwards deltas as SSE. The per-session-id keying is MANDATORY, not an
optimization — OCP_TUI_MAX_CONCURRENT defaults to 2, so two claude panes already run at
once and a shared sink would splice one client's deltas into another's stream.

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
reaches the pane through its own env (OCP_TUI_STREAM_FILE) and derives from the session-id,
which a pre-booted pane fixes at boot.

The hook is SYNCHRONOUS (forceSyncExecution: claude blocks on it), so the script writes and
exits: one `cat` append, nothing else. Measured p50 7.2ms / p90 14.7ms per fire, ~50ms across
a whole turn — noise against a 6-10s turn.

It remains the terminal-turn signal, the source of the returned/cached text T, and the input
to the honesty gates. The delta stream is a low-latency MIRROR, never a replacement:

  - the truncation gate (C-2) and auth-banner gate (C-1, issue #133) run BEFORE anything is
    committed or flushed, unchanged;
  - at end of turn the streamed bytes are asserted against T. Equal -> serve. A strict PREFIX
    of T -> top up from the transcript so the client still receives exactly T (counted).
    NOT a prefix -> REFUSE the turn: SSE error frame, no cache, no success, streamDivergences++.
    Serving text the transcript disagrees with is the failure class ALIGNMENT.md exists to
    prevent, so this fails loud rather than degrading quietly;
  - only T is ever cached — never the concatenated deltas.

The auth banner needs prevention, not just detection (SSE deltas cannot be un-sent), so the
first OCP_TUI_STREAM_HOLDBACK (100) chars are withheld: the default banner detector cannot
match a message longer than 100 chars, so releasing past that provably cannot leak a banner.
A custom CLAUDE_TUI_ERROR_PATTERNS has no such bound — OCP warns at boot.

  - BANNER, before/after the spawn change: `Sonnet 4.6 with low effort · Claude Max` both,
    including on the pane the server itself spawns. Never `API Usage Billing`. Transcript
    entrypoint stays "cli". --settings is not a --bare-class flag.
  - --settings MERGES with <HOME>/.claude/settings.json rather than clobbering it (the
    user-level settings' `env` block still reached the hook), so the isolated-HOME settings
    story (permissions / additionalDirectories) survives.
  - EXACTNESS: 8/8 varied prompts (short, long, markdown, code fence, multilingual, JSON,
    table, unicode) byte-exact vs transcript T, streamed AND buffered. 0 top-ups,
    0 divergences over 15 streamed turns.
  - TTFT: buffered delivers NOTHING until the turn ends (TTFB == total, 7.5-15.8s). Streamed
    sends headers at ~25ms (heartbeat covers the pre-first-delta silence) and first content
    mid-generation, e.g. markdown 7.9s first chunk / 12.8s total; long 9.7s / 17.4s.
  - CONCURRENCY DEMUX: two concurrent streamed turns (ALPHA/BRAVO), tui.inflight peaked at 2,
    each read its own session-keyed transcript, ZERO cross-contamination.
  - AUTH-BANNER GATE under streaming, both layers: a short banner-like turn reached the client
    as 0 content chunks + an SSE error frame (never emitted); a long one was streamed but still
    ended on an error frame, not finish_reason:"stop", and was not cached.
  - DISCONNECT mid-turn: pane torn down and semaphore slot released within 1s (info-logged,
    not booked as a model error).
  - THINKING: not leaked. Opus 4.8 + xhigh turns carry a thinking block with a signature but
    `thinking:""` (the reasoning text is not persisted in interactive mode), both
    MessageDisplay text-extraction sites in the 2.1.207 bundle filter type==="text", and no
    reasoning prose appeared in any delta; concat===T held exactly on the single-message turn.
  - npm test: 282 passed, 0 failed (was 267 on main; +15).

The transcript keeps only the model's LAST assistant message. A turn where the model narrates
before calling a tool therefore has two messages, and T is only the second. If the narration
exceeds the holdback it has already been streamed and cannot be retracted -> the turn is
REFUSED. Reproduced live: Opus narrated 475 chars before a Bash call. The assembler discards
a prior message's text when nothing has been emitted yet (so short narration is handled
correctly and stays exact), and raising OCP_TUI_STREAM_HOLDBACK above the narration length
rescues the turn — verified on that exact transcript: holdback>=500 -> served, exact=true.
Documented in README and ADR 0007; this is why streaming is opt-in and off by default.

ADR 0007 line 59 ("no real token streaming — deliberate") is amended, not silently
contradicted.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(tui): re-integrate streaming onto the warm-pane pool (#158) — install the hook at BOOT

Rebasing backlog #2 (streaming) onto #158 (warm pane pool) is not a textual merge: #158
split the monolithic runTuiTurn into bootTuiPane + runTuiTurn, and streaming had patched
the monolith. Re-integrating it in the OLD shape would have compiled, passed every existing
test, and been WRONG.

The bug that shape would have shipped: the sink was derived at TURN time from a streamDir
argument. But a POOLED pane is pre-booted long before any request exists — so on a pool HIT
runTuiTurn never cold-boots, no hook was ever registered on that pane, and the turn would
silently serve BUFFERED. Every miss streams, every hit does not; no error, no failing test.
The operator sees "streaming does nothing in production" and has nothing to grep for.

Fix — install the hook where the pane is born:
  - bootTuiPane({ streamDir }) registers the MessageDisplay hook at spawn and returns the
    pane's own sink (pane.streamFile), keyed by the pane's own --session-id. The hook script
    and settings file are STATIC (one pair per streamDir); the only per-turn thing is the
    sink path, and it is fixed at boot. So nothing request-specific is baked into a spawn.
  - runTuiTurn reads pane.streamFile — never recomputes it — so a warm pane and a cold pane
    stream through byte-for-byte the same path.
  - server.mjs threads the same streamDir into the pool's bootPane closure, so pre-booted
    panes carry the hook too. TUI_STREAM/TUI_STREAM_DIR now declare before the pool needs them.

Three regression guards added (test-features.mjs), and the third was MUTATION-TESTED: with
the fix reverted to the turn-time shape it fails ("the pooled pane's deltas must reach the
client"), with the fix in place it passes. A guard nobody has watched fail is not a guard.

/health: pool + stream* fields are now a union — the shape assertion asserts CONTAINMENT of
the seven grandfathered keys plus an exact added-set, so a future field that silently
REPLACED an original key cannot pass.

Class B (ADR 0007, OCP-owned TUI spawn) — cli.js does NOT perform this operation.
npm test: 313 passed, 0 failed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VqgWJcjxrjjL9L9SkpZyXR

* fix(tui): close the streaming auth-banner leak + 6 further review findings (PR #159)

Independent review (Iron Rule 10) found a HIGH bug by EXECUTING the code, not reading it.
All seven findings fixed. F1 and F3 were merge-blocking.

F1 (HIGH) — the auth-banner holdback was bypassed after the first release.
  TuiDeltaAssembler.released was set once and never reset at a message_id boundary, so the
  holdback + detectError predicate guarded only the FIRST message of a turn. In production's
  own configuration (OCP_TUI_FULL_TOOLS=1, where multi-message tool-using turns are the norm):
  the model narrates past the holdback before a tool call -> released; credentials expire
  mid-turn -> claude renders the 401 as ordinary assistant TEXT as a NEW message -> push()
  took the `if (this.released)` branch and handed the banner verbatim to the client. That is
  precisely the silent-error case the C-1 gate exists to prevent. Detection survived (the
  turn was still refused at finalize) but PREVENTION did not.
  Fix: once a message boundary follows an emit, the turn is already unrecoverable — finalize()
  will refuse it — so push() now emits NOTHING further for the rest of the turn.
  Second hole in the same predicate: detectTuiUpstreamError() trims before applying its
  <=100-char rule, so 101 whitespace chars trimmed to "" -> detector had nothing to classify
  -> returned null -> release fired having screened nothing. Release now gates on the TRIMMED
  length, so both sides of the check talk about the same string.

F2 — the "provably safe" claim in stream.mjs, ADR 0007 and README was unsound as written.
  Restated with both required halves: (i) nothing is emitted until the trimmed accumulation
  exceeds the detector's max banner length, AND (ii) no emission at all once a message
  boundary follows an emit. Half (i) alone only ever covered a turn's first message.

F3 (blocker) — prepareStreamHook was write-if-missing, so md-hook.sh could never be updated
  OR repaired: a host that booted once under an older version was stuck on that HOOK_SCRIPT
  forever, and a non-atomic write interrupted mid-flight left a TRUNCATED script that
  existsSync() called fine — on a hook claude BLOCKS on synchronously. Now written
  unconditionally via tmp+renameSync (the pattern already used by ensureTuiCwdTrusted).

F4 — the two spawn paths differed for non-streaming requests: the pool installed the hook
  whenever OCP_TUI_STREAM was on (correct — a pre-booted pane cannot know what request it will
  serve), but the cold path gated it on this turn's onDelta. So one stream:false request got
  --settings on a pool HIT and not on a MISS: two spawn argvs for the identical request, on
  this project's billing-classification surface. Both paths now gate on TUI_STREAM alone;
  whether the sink is POLLED remains correctly gated on onDelta.

F5 — pool._drop() killed the pane but orphaned its sink file; the reap tick drains the whole
  pool, so sinks accumulated with no GC path. Now removed best-effort on every drop path.

F6 — /health counters did not measure what they documented: streamTurns was incremented only
  AFTER the honesty gates, hiding exactly the turns an operator most wants to see (and making
  streamDivergences/streamTurns a meaningless ratio); streamDeltas counted every fire while
  claiming to count forwarded ones. Counters and docs now agree.

F7 — total hook failure was silent: zero fires per turn still yields ok:true/exact:false and
  a normal, fully-buffered answer. Only streamTopUps moved, which the code itself calls
  benign. Added streamZeroDeltaTurns (+ a tui_stream_zero_deltas warning) to separate "the
  hook is dead" from "one fire was dropped".

Tests: 316 passed, 0 failed (was 313). Every new guard MUTATION-TESTED — with each fix
reverted the guard named for it fails, and passes with the fix restored:
  - drop the restartedAfterEmit guard   -> 2 failed (incl. the strengthened old test)
  - revert trim() in the release gate   -> 1 failed
  - revert F3 to write-if-missing       -> 1 failed
The pre-existing test "new message_id AFTER an emit" asserted finalize().ok === false but
never checked what push() RETURNED — so it passed while F1 was live, documenting the leak
instead of catching it. Strengthened to assert the emission, not just the verdict.

Class B (ADR 0007, OCP-owned TUI spawn) — cli.js does NOT perform this operation.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VqgWJcjxrjjL9L9SkpZyXR

---------

Co-authored-by: dtzp555 <dtzp555@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-07-14 08:19:24 +10:00

30 KiB
Raw Blame History

ADR 0007 — TUI Interactive Mode (subscription-pool bridge)

Date: 2026-05-31 Status: Accepted — amended by PR-4 (entrypoint hardening), PR-B (observability + concurrency), PR-C (env-token auth + defunct-reaping), PR-D (credential-isolated home — corrects PR-C) Deciders: project maintainer Authority: claude CLI v2.1.158 interactive mode — verified live on the test host that sessions launched without -p / --output-format carry cc_entrypoint=cli (subscription pool), not cc_entrypoint=sdk-cli (Agent SDK credit pool). Mechanism verified on cli.js v2.1.104; live-confirmed on v2.1.158.


Context

On 2026-05-14 Anthropic announced (effective 2026-06-15) a billing split that routes requests by cc_entrypoint:

cc_entrypoint value Billing pool
cli Pro/Max subscription pool
sdk-cli Agent SDK credit pool (~$20/mo on Pro = easily exhausted)

OCP's existing path (claude --output-format stream-json -p) sets cc_entrypoint=sdk-cli. After 2026-06-15 every OCP request will draw from the Agent SDK pool rather than the subscription.

The structural response: add an opt-in mode that drives a real interactive claude session (no -p, no --output-format), which carries cc_entrypoint=cli and therefore bills against the subscription. The response text is read from claude's native JSONL transcript instead of from stdout.

This is a personal-use A-path feature (single-user, single-subscription host). It is not a multi-tenant isolation layer.

Source-verified entrypoint mechanism (PR-4 amendment)

Claude CLI's main() calls a startup function (t$A in the compiled bundle) that sets process.env.CLAUDE_CODE_ENTRYPOINT only if unset to:

(argv has -p/--print/--init-only/--sdk-url  OR  !process.stdout.isTTY) ? "sdk-cli" : "cli"

The billing header reads cc_entrypoint = process.env.CLAUDE_CODE_ENTRYPOINT ?? "unknown". The "unknown" branch is dead code for any real main() spawn — the startup function always sets a value on unset env. The real risk is not "unknown": it is a lost TTY (e.g. stdout redirected or a non-PTY spawn) silently flipping the self-classification to "sdk-cli" and drawing from the metered pool.

cc_entrypoint is one of ~6 upstream run-mode signals. The dominant discriminator is the system-prompt identity block ("official CLI" vs "Claude Agent SDK"), which is driven by genuine interactivity (no -p, no --output-format, real PTY) and is overridable by no env var. This is the real reason the tmux/no--p approach works: the spawn is genuinely interactive, not just labelled as such.


Decision

Add CLAUDE_TUI_MODE=true as an opt-in flag in server.mjs.

How it works

  1. Each request spawns a fresh tmux session running claude --model <M> --session-id <UUID> --strict-mcp-config --disallowedTools 'mcp__*' (no -p, no --output-format).
  2. The spawn result is checked immediately: if tmux new-session returns a non-zero exit status (or a falsy result), the request is aborted with tui_spawn_failed: tmux session not created before the boot sleep. This is the spawn/PTY gate — OCP must not issue a billing request without a verified interactive session.
  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"). 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)

CLAUDE_CODE_ENTRYPOINT on the spawn env is managed by resolveTuiEntrypointEnv(env, mode) (exported from lib/tui/session.mjs, pure, testable). The function always deletes any inherited value first so a stray env var from OCP's own parent process can never leak in and mislabel the billing header. Then:

OCP_TUI_ENTRYPOINT Behaviour
cli (default) Sets CLAUDE_CODE_ENTRYPOINT=cli deterministically — subscription-pool classification. Honest only because the spawn is a genuine interactive PTY (tmux pane, no -p, stdout not redirected, new-session verified).
auto Deletes the key → claude self-classifies via t$A (TTY → cli). Use to observe/diagnose the real TTY-derived value.
off Leaves the env exactly as inherited — diagnostics / honesty audit only.

Governing rule (verbatim): OCP may make a true value deterministic; it may never assert a value the spawn's real state contradicts. When it cannot make the claim true (e.g. cannot guarantee a PTY), it fails/drops the request — it does not force the signal.

This is why the spawn/PTY gate (step 2 above) is load-bearing for mode="cli": if new-session fails, there is no PTY, so asserting cli would be dishonest. Abort rather than lie.

OCP never suppresses the billing header (anti-fingerprinting: we do not mask the spawn).

2026-06-15 verification protocol

Run one quiesced canary request in TUI-mode and watch the Agent SDK credit balance (not the request header). If the balance drops, the subscription pool is unreachable via spawn. Per the constitution (ALIGNMENT.md), the response is to drop the Anthropic provider rather than escalate spoofing.

Version caveat: mechanism verified on cli.js v2.1.104 + live on v2.1.158. Re-verify after any major cli.js upgrade.

Default behaviour is unchanged

When CLAUDE_TUI_MODE is unset (the default), no code path touches callClaudeTui or runTuiTurn. upstreamCall === callClaude and streaming uses callClaudeStreaming — byte-for-byte identical to the pre-TUI code path.

Kill-switch

Unset CLAUDE_TUI_MODE (or set it to any value other than "true") → stream-json path restored immediately on next restart.

Home strategy

Superseded by the PR-D amendment below for the env-token case. As of PR-D, TUI_HOME is computed by resolveTuiHome(): when CLAUDE_CODE_OAUTH_TOKEN is set (and OCP_TUI_HOME is unset) the default is a credential-free scratch home, not the real home. The descriptions below remain accurate for the no-env-token case and the explicit OCP_TUI_HOME override case.

  • Real-home (default when NO env token, OCP_TUI_HOME unset): claude runs with the operator's own ~/.claude/ — shared credentials, existing onboarding, no OAuth fork risk. ensureTuiCwdTrusted seeds the trust record for the scratch cwd in the real ~/.claude.json (atomic write).
  • Scratch-home opt-in (OCP_TUI_HOME=<path>, no env token): a dedicated HOME that symlinks ~/.claude/.credentials.json from the real home (token is never copied) and seeds a stripped ~/.claude.json (no project history, trusts only the scratch cwd). Caveat: claude rewrites .credentials.json on OAuth token refresh, replacing the symlink with a regular file — this forks the credentials. Use this legacy symlink mode only with a dedicated OAuth or for ephemeral testing. (The PR-D env-token mode avoids this caveat entirely — no credentials file to fork.)

Working directory

TUI_CWD = OCP_TUI_CWD || $HOME/.ocp-tui/work (dedicated scratch cwd). Transcripts land under <HOME>/.claude/projects/<encoded-cwd>/ — a stable, single location separate from the operator's real project histories. The directory is created automatically on first request.

MCP hard-disable

--strict-mcp-config (no --mcp-config argument) prevents account-attached managed MCP servers from connecting. Belt-and-braces: --disallowedTools 'mcp__*' blocks any MCP tool invocation even if a server were somehow loaded. Built-in tools (Bash, Read, etc.) are left enabled on the A-path (single-user, acceptable).

Session namespace

All tmux sessions use the prefix ocp-tui-. The prefix-scoped reaper (reapStaleTuiSessions) kills only ocp-tui-* sessions, never olp-tui-* or any other prefix. A stale-session cleanup runs once at OCP boot when TUI_MODE is on.


SECURITY — PROMINENT WARNING

TUI-mode is SINGLE-USER / SINGLE-OPERATOR ONLY.

claude runs as the OCP process owner with full filesystem access regardless of HOME setting. Home selection is not user isolation. If OCP is serving multiple users or guest API keys:

  • A guest prompt would run claude with the operator's filesystem access.
  • An adversarial prompt could exfiltrate files, run shell commands, or exhaust the subscription.

Never enable CLAUDE_TUI_MODE=true on an OCP instance that serves untrusted callers or multiple users.

The B-path (multi-tenant isolation) requires:

  1. --tools "" (no built-in tools)
  2. Per-key ephemeral HOME (isolated credentials + no cross-key project pollution)
  3. Sandbox runtime (e.g. @anthropic-ai/sandbox-runtime)

B-path is deferred and is not implemented in this ADR. Until B-path lands, TUI-mode must only be enabled on a personal single-user OCP.


Observability and concurrency (PR-B amendment)

Date: 2026-06-10 Status: Accepted — amends ADR 0007. Motivation: the post-PR-A code audit, findings C-4 (P1) and C-5 (P1).

C-4 — independent concurrency bound for the TUI path

The global MAX_CONCURRENT gate lives in spawnClaudeProcess() (the -p / stream-json path). callClaudeTui() never calls spawnClaudeProcess — it calls runTuiTurn(), which cold-boots a full interactive claude inside a fresh tmux session. So the TUI path had no concurrency bound: N concurrent TUI requests spawned N simultaneous cold-boot tmux+claude processes. On a small host (e.g. a Pi 4 serving a family) a burst of ~5 is an OOM risk and also multiplies subscription rate-limit pressure.

PR-B adds an independent limiter for the TUI path (lib/tui/semaphore.mjs, TuiSemaphore):

  • OCP_TUI_MAX_CONCURRENT, default 2. Rationale: a TUI turn is heavy — a per-request cold-boot of tmux+claude plus up to CLAUDE_TUI_WALLCLOCK_MS (120 s) of wallclock — so a small host cannot run many at once. 2 is the conservative default that keeps a Pi-class host alive under a family burst while still allowing some overlap. It is deliberately not the same knob as MAX_CONCURRENT (default 8): the two pools have different shapes (a stream-json spawn is cheap and fast; a TUI turn is a heavy cold-boot + long wallclock), so coupling them would mis-size one of the two paths.
  • Queue, don't reject. The limiter queues (awaits a slot), mirroring the spirit of MAX_CONCURRENT — requests are not dropped on contention. To bound memory against a runaway client, the wait queue itself is capped (maxQueue, default 32× the limit); when the queue is full run() rejects with tui_queue_full, surfaced as a 503 — deterministic backpressure rather than silent OOM.
  • Slot released in a finally. TuiSemaphore.run(fn) releases the slot in a finally, so any throw — PR-A's honesty gates (tui_wallclock_truncated, tui_upstream_error), a tui_paste_not_landed, or a tui_spawn_failed — can never leak a slot.

This limiter has zero effect when TUI_MODE is off: callClaudeTui is never reached, so the semaphore is never entered. The default stream-json path is untouched.

C-5 — operator-visible drift surface on /health (additive)

The tui_entrypoint_mismatch warning only reached journald. After the 2026-06-15 flip, a silent sdk-cli drift (the documented top risk in this ADR — a lost TTY flipping the self-classification to the metered Agent SDK pool) would drain metered credits invisibly. PR-B adds a tui block to the /health JSON response so an operator can poll it:

tui: {
  enabled:              <TUI_MODE>,
  entrypointMode:       <OCP_TUI_ENTRYPOINT>,   // cli | auto | off
  lastEntrypoint:       <last observed cc_entrypoint, e.g. "cli", or null>,
  entrypointMismatches: <count of cli-expected-but-got-other turns>,
  inflight:             <current concurrent TUI turns>,
  queued:               <turns waiting for a slot>,
  maxConcurrent:        <OCP_TUI_MAX_CONCURRENT>
}

lastEntrypoint is recorded and entrypointMismatches incremented inside callClaudeTui in the same mismatch branch that already emits the journald warning (via recordTuiEntrypoint). inflight / queued / maxConcurrent come from the C-4 semaphore. When TUI_MODE is off the block still appears with enabled:false (cheap, harmless) so the response shape is stable for consumers regardless of mode.

ALIGNMENT authorization for the /health change

/health is a grandfathered B.2 endpoint under ADR 0006, frozen at its v3.16.4 behaviour. ALIGNMENT.md's grandfather provision states: "Any change to the contract (request shape, response shape, semantics) of a grandfathered B.2 endpoint is treated as a new authorization request and requires either a behaviour-preserving refactor PR or its own ADR."

This amendment is that authorization. The argument:

  • The change is additive: it adds one new top-level field (tui) containing only new sub-fields. No existing /health field is changed, renamed, removed, or re-typed, and no existing semantics change. Existing /health consumers (the dashboard, ocp-connect, monitoring) read the fields they already read and are unaffected — the change is behaviour-preserving for them, which is exactly the bar the grandfather provision sets for a non-ADR contract change.
  • The TUI observability surface is an intrinsic part of the TUI feature whose authorizing authority is this ADR (0007), not a brand-new B.2 endpoint. We are not adding a new B.2 endpoint or a new method (which would each require their own fresh ADR under the New Class B endpoint procedure) — we are extending the response of an existing grandfathered endpoint with fields that report state owned by an ADR-0007 feature. ADR 0007 is the natural home for that authority, and this amendment records it explicitly.
  • cli.js does not perform this operation — /health is OCP-owned (Class B), so no cli.js citation applies; the citation is this ADR + ADR 0006 (grandfathered B.2) per ALIGNMENT.md's Class B citation requirement.

OCP_TUI_MAX_CONCURRENT summary

Env var Default Meaning
OCP_TUI_MAX_CONCURRENT 2 Max concurrent interactive TUI turns. Independent of CLAUDE_MAX_CONCURRENT (the stream-json path). Excess turns queue (bounded); a full queue yields a 503.

Authentication + defunct-reaping (PR-C amendment)

Date: 2026-06-13 Status: Accepted — amends ADR 0007. Motivation: the PI231 production incident — TUI-mode returned Please run /login · API Error: 401 for days; re-login never stuck.

How the TUI claude authenticates

The spawned interactive claude obtains its OAuth bearer in one of two ways, in this order of preference:

  1. CLAUDE_CODE_OAUTH_TOKEN in env (PREFERRED). If the env var is set on the OCP process, buildTuiCmd adds CLAUDE_CODE_OAUTH_TOKEN=<shq-escaped token> to the pane command's env prefix. claude then authenticates via this long-lived token and never touches the credentials-refresh path. This is the stable mode — it is exactly how the oracle and Mac-mini hosts already run (and how server.mjs's own getOAuthCredentials() takes the same env at highest precedence). cli.js is not the authority here: this is a Class B, OCP-owned TUI spawn — see the Class B citation below.
  2. <HOME>/.claude/.credentials.json (FALLBACK). When the env var is unset, claude falls back to the credentials file and its short-lived access token, renewing via the single-use refresh token.

The token MUST be set explicitly in buildTuiCmd because tmux does not forward the parent process's environment to the pane (verified live 2026-06-01 — the same reason the whole env is delivered as an env prefix). A token sitting in the OCP process env is invisible to the pane unless buildTuiCmd re-emits it.

Why the fallback path corrupts (the PI231 incident)

When the env token is absent, every per-request spawn drives claude through the credentials.json refresh path. OAuth refresh tokens are single-use / rotating: a refresh consumes the old refresh token and writes a new one. The per-request kill-session teardown can race / interrupt claude mid-rotation, and over many spawn+kill cycles the refresh token ended up an empty string — at which point renewal is impossible and the host returns a permanent 401. Re-login writes a fresh token, but the next spawn re-corrupts it. Proof the env-token fix works: on the broken PI231 host, CLAUDE_CODE_OAUTH_TOKEN=<oat01 token> claude -p ... returned a real answer despite the corrupt credentials.json (control without the env token = 401).

Operator guidance: set CLAUDE_CODE_OAUTH_TOKEN on any TUI-mode host. The credentials.json fallback is retained only for hosts that intentionally rely on it; it is not recommended for a long-running TUI deployment.

Security note: with the token in the pane command, it is visible in ps. This is acceptable for the single-user A-path (it mirrors the existing plaintext-token practice for server.mjs), and the multi-user B-path is already refused at boot (CLAUDE_TUI_MODE=true + AUTH_MODE=multi is a hard FATAL), so a guest can never reach this spawn.

Defunct <claude> reaping

The connected leak: 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 returns the instant the server forks the pane). Node can therefore never waitpid()/reap it — a SIGKILL still needs the parent (the tmux server) to reap. kill-session destroys the session but leaves the pane's claude (and its grandchildren) as <defunct> zombies that only the server reaps; over 30 days on PI231 this accumulated to 25 defunct <claude> (a live tmux kill-server dropped it 25→3).

The node-reachable action that actually reaps — rather than merely re-signalling — is to stop the tmux server: on server exit the kernel reparents survivors to init (PID 1), which reaps them. reapStaleTuiSessions therefore, after killing our own ocp-tui-* sessions, issues kill-server only when no foreign session of any prefix remains (coexistence: never disrupt a co-hosted olp-tui-* instance). This runs at boot (existing) and now on a 15-min periodic interval gated on TUI-mode and on the TUI path being idle (inflight === 0 && queued === 0) so a live turn's pane is never torn down. Residual: a request whose pane is created in the narrow window between the idle-check and kill-server would fail cleanly via the existing honesty gates (rare; documented in the server comment).

ALIGNMENT authorization (Class B)

Both changes are Class B (OCP-owned TUI spawn). cli.js does not perform either operation — there is no cli.js analogue for "how the TUI pane authenticates" or "reaping tmux-server-owned zombies"; this surface is authorized by this ADR (0007) per ALIGNMENT.md's Class B citation requirement. No Class A wire surface, no endpoint shape, no alignment.yml blacklist token, and no models.json entry is touched.


Credential-isolated home for env-token auth (PR-D amendment)

Date: 2026-06-13 Status: Accepted — amends ADR 0007. Corrects the PR-C rationale and the original "Home strategy" section's scratch-home caveat. Motivation: PR-C's env-token passing alone did not fix the PI231 401. Decisive live evidence (claude 2.1.104, PI231):

Condition Result
env token passed + a broken ~/.claude/.credentials.json present 401 (Please run /login · API Error: 401)
env token passed + credentials.json moved aside works (real answer)

Corrected root cause

Interactive claude PREFERS ~/.claude/.credentials.json over the CLAUDE_CODE_OAUTH_TOKEN env var. A stale/corrupt credentials.json therefore shadows the env token. (This is unlike -p mode, where the env token wins — which is why server.mjs's own getOAuthCredentials() is unaffected and why PR-C's premise looked sufficient.) So passing the token (PR-C, buildTuiCmd) is necessary but insufficient: the TUI claude must additionally run in a HOME that has no credentials.json, so the env token is the only credential and is authoritative.

This also fixes the original incident at the root, more completely than PR-C claimed: with no credentials.json in the home, claude never runs the token-refresh path at all, so the single-use refresh token can never be rotated — and therefore never corrupted — by the spawn+kill-session cycle. The 25-zombie / empty-refresh-token failure mode becomes structurally impossible, not merely avoided.

Decision

When CLAUDE_CODE_OAUTH_TOKEN is set, the TUI claude runs in a credential-free scratch home by default:

  • resolveTuiHome({ realHome, configuredHome, envTokenSet }) (exported from lib/tui/session.mjs, pure) decides the home:
    • OCP_TUI_HOME set → that path (explicit override, back-compat — an operator who configured it keeps exactly that home).
    • else env token set<realHome>/.ocp-tui/home — a dedicated scratch home seeded with a minimal .claude.json (hasCompletedOnboarding=true + trust only the scratch cwd) and its own projects/ dir, and deliberately NO .credentials.json (no symlink, no copy).
    • else (no env token) → the operator's real home — byte-for-byte the pre-fix behaviour for hosts that intentionally rely on credentials.json.
  • prepareTuiHome(realHome, tuiHome, cwd, { envTokenMode }) gates the credential handling: in envTokenMode it creates the scratch projects/ dir and seeds the minimal trusted .claude.json but never creates the credentials symlink. runTuiTurn sets envTokenMode = !!CLAUDE_CODE_OAUTH_TOKEN && ehome !== realHome.
  • readTuiTranscript reads from the same home claude runs under (ehome), so transcripts land under <scratch home>/.claude/projects/ and findTranscriptPath globs them there — the home is threaded through consistently. (We chose scratch-HOME over CLAUDE_CONFIG_DIR: the binary supports CLAUDE_CONFIG_DIR, but it relocates the transcript root to <CONFIG_DIR>/projects/ rather than <HOME>/.claude/projects/, which would fork the transcript-resolution rule across modes for no benefit. The scratch-HOME lever reuses the existing, tested prepareTuiHome/ehome plumbing.)

This RESOLVES — not reintroduces — the scratch-home caveat

The original "Home strategy" section and PR-C's prepareTuiHome comment warned that scratch-home is unsafe because claude rewrites a symlinked .credentials.json on token refresh → forks/corrupts the OAuth credentials. That caveat does not apply to env-token mode: there is no credentials.json in the home to fork, and claude never refreshes (it uses the long-lived env token), so there is no rotation and no corruption. The fork risk was inherent to the symlink approach; removing the credentials file entirely removes the risk. The legacy symlink mode is retained only for an operator who explicitly sets OCP_TUI_HOME without an env token, and its caveat is preserved for exactly that path.

ALIGNMENT authorization (Class B)

Class B (OCP-owned TUI spawn). cli.js has no analogue for the TUI pane's auth/home strategy; authorized by this ADR (0007) per ALIGNMENT.md's Class B citation requirement. server.mjs is touched only to compute TUI_HOME via resolveTuiHome() (TUI wiring) and to surface the auth mode in the boot log — no Class A wire surface, no endpoint shape, no alignment.yml blacklist token, and no models.json entry is touched.


Consequences

Positive

  • After 2026-06-15, requests in TUI-mode bill against the Pro/Max subscription pool (cc_entrypoint=cli) rather than the Agent SDK credit pool.
  • Kill-switch is immediate (unset env var + restart); zero code change required.
  • Default stream-json path is untouched — no regression risk for existing deployments.

Negative / trade-offs

  • No token streaming: responses are buffered then replayed as chunked SSE. Clients see a delay then the full response arrives; real-time token streaming is not available in TUI-mode.
  • Billing unmeasurable until 2026-06-15: the cc_entrypoint=cli signal is verified, but the credit deduction from the correct pool cannot be confirmed until the billing split activates.
  • tmux dependency: the host must have tmux installed. CI / Docker images that lack tmux cannot use TUI-mode (the default stream-json path is unaffected).
  • Wall-clock cap: long Opus thinking turns may hit the 120 s cap. Increase CLAUDE_TUI_WALLCLOCK_MS if needed (no quiescence heuristic — the reader polls until terminal marker or cap).
  • Grey-area usage: running an interactive claude session headlessly to serve HTTP requests is not an officially documented use case. If Anthropic policy changes to block this pattern, OCP must fall back to the stream-json path (unset CLAUDE_TUI_MODE).

Coexistence

  • tmux prefix ocp-tui- is registered. Any co-hosted OLP test instance must use olp-tui-. Never run two TUI proxies on the same OAuth concurrently — stop one instance during integration testing.

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: trueclaude 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 610 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 S1S6 / T1T6 were validated live on the test host against claude v2.1.158.