docs(adr): Phase 7 Amendment 1 — supersede PR-B with ephemeral-home + ISOLATION contract (#66)

Co-merging ADR 0014 Amendment 1 (4-layer Solution 1) + ADR 0002 Amendment 9 (Provider ISOLATION contract). Reviewed by 2 fresh-context opus subagents per Iron Rule 10. Second review verdict APPROVE after 6 citation-discipline fold-ins applied. PR-B outer-bwrap archived to branch phase-7-pr-b-outer-bwrap-snapshot.
2026-07-21 21:15:10 +00:00 · 2026-05-29 10:00:38 +10:00
parent 3551921f55
commit d67ba3d675
2 changed files with 841 additions and 2 deletions
@@ -9,6 +9,8 @@
 > **Note on numbering.** Sequence is 1, 3, 4, 5, 6, 7 — Amendment 2 was never written. The reserved slot was originally planned for a separate `maxConcurrent` ratification, but that content was folded into Amendment 1 (the retroactive contract-sync amendment) at filing time and the gap was not backfilled. The gap is intentional and load-bearing — no missing content; do not renumber Amendments 3+ to close it (cross-references to Amendment N from other docs would silently break).
 > **Forward-pointer:** Amendment 9 (2026-05-29) — Provider `ISOLATION` Contract for Multi-Tenant Spawn Isolation — is located at the **end of this file** (after § Sources), not in this Amendments block. The placement is documented in Amendment 9's editorial note; the substance is the addition of an OPTIONAL `ISOLATION` named export to provider plugin modules, consumed by `lib/sandbox/manager.mjs` (per ADR 0014 Amendment 1) to compose per-spawn ephemeral-home + per-provider isolation primitives. Co-merge with ADR 0014 Amendment 1.
 ### Amendment 8 — 2026-05-26: Permit `quotaStatus()` direct-API access (READ-ONLY exemption) for plan-usage probes (D79–D80 — Phase 5)
 - **Context:** ADR 0012 (Phase 5 charter) opens 2026-05-26 to port OCP's plan-usage probe (`ocp/server.mjs:842-1109`) into `lib/providers/anthropic.mjs:quotaStatus()`. The probe calls `POST https://api.anthropic.com/v1/messages` directly with an OAuth bearer and parses `anthropic-ratelimit-unified-*` response headers. This violates the plugin contract's implicit assumption that ALL provider interaction goes through `spawn` (the binary CLI). `ALIGNMENT.md` Rule 2 (provider-CLI-as-authority) further constrains plugins to operations the provider CLI itself performs. The OCP-derived plan-usage probe satisfies neither of these — it bypasses `claude -p` and hits the public API directly. **Without an explicit exemption Amendment, D80 is unalignable.**
@@ -210,3 +212,469 @@ Every provider plugin exports an object conforming to:
 - OLP v0.1 spec §4.2 (Plugin-based provider system, including the v1.0 Provider contract definition)
 - OCP ADR 0003 (`models.json` as SPOT) — informs the "static enumeration, not filesystem scan" loading model
 - OCP ADR 0005 — the context paragraph references OCP's `server.mjs` reaching 1667 lines at one provider; the plugin architecture is the structural response to that complexity scaling N×
 ---
 ### Amendment 9 — 2026-05-29: Provider `ISOLATION` Contract for Multi-Tenant Spawn Isolation (Phase 7, ADR 0014 Amendment 1 co-merge)
 > **Editorial note.** Per the existing "amendments most-recent-first" convention near the top of this file, Amendment 9 logically slots between Amendment 8 and the original body. It is physically located at the file's tail (after § Sources) to honor the constitution's "append, do not rewrite" discipline for this addition — the rationale is that the contract surface added here is large enough (a structured per-provider sub-export, not just a hint-bag field) that an in-line edit of the § Decision body would constitute a rewrite of the v1.0 contract listing rather than an amendment over it. Future readers consulting the amendment-history block at the top of the file will find a stub forward-pointer to this section.
 >
 > The amendment is otherwise a peer of Amendments 1–8 (same `###` heading depth, same shape).
 #### Context
 The OLP spawn pipeline currently treats every provider as a plain `child_process.spawn` of the provider's CLI binary with a homogeneous env block and the server process's working directory. This works on a single-tenant developer laptop. It does **not** work on the family-LAN PI231 deployment (multi-key, multi-caller, single OS user) and is a hard blocker for the cloud rollout described in `docs/plans/cloud-deployment-family.md` § 5 — both for the reasons captured in the 2026-05-27 incident memory at `~/.cc-rules/memory/projects/olp/incident_2026_05_27_spawn_cli_security.md` (OAuth-token exfiltration, codex `shell` tool real execution, cross-tenant filesystem read leakage).
 The parallel ADR 0014 Amendment 1 retires the **outer-bwrap PR-B approach** — which initialized `@anthropic-ai/sandbox-runtime` `SandboxManager` once at server startup and wrapped every provider spawn through a global namespace — and replaces it with a **per-spawn ephemeral-home + per-provider isolation primitives** architecture. The new shape of `lib/sandbox/manager.mjs` is no longer a thin wrapper around `wrapSpawn()`; it is an orchestrator that, on each spawn, asks the provider plugin *what isolation primitives this provider needs*, composes them, and hands the spawn a ready-to-execute environment.
 The thing the orchestrator asks for is the subject of this amendment: the **Provider `ISOLATION` contract**.
 #### The interaction surface this amendment governs
 ```text
                  ┌──────────────────────────────────────┐
                  │ server.mjs handleChatCompletions     │
                  │   → executeHopFn                     │
                  │   → provider.spawn(irRequest, ...)   │
                  └────────────────────┬─────────────────┘
                                       │
                                       ▼
                  ┌──────────────────────────────────────┐
                  │ lib/sandbox/manager.mjs              │
                  │   prepareIsolatedEnvironment(        │
                  │     provider,                        │
                  │     { keyId, reqId, ... }            │
                  │   )                                  │
                  │     ↓ reads provider.ISOLATION       │
                  │     ↓ mkdtemp ephemeralRoot          │
                  │     ↓ mkdir requiredHomePaths        │
                  │     ↓ symlink/copy credentialMounts  │
                  │     ↓ compose ephemeralEnvOverrides  │
                  │     ↓ wrap args via toolHardening    │
                  └────────────────────┬─────────────────┘
                                       │
                                       ▼
                  ┌──────────────────────────────────────┐
                  │ child_process.spawn(bin, args, {     │
                  │   env: composedEnv, cwd: epRoot, ... │
                  │ })                                   │
                  └──────────────────────────────────────┘
 ```
 The provider plugin is the **authority** for what isolation primitives are needed. The provider knows what env var its CLI honors for credential lookup (`HOME`, `CODEX_HOME`, `VIBE_HOME`, …). The provider knows whether the CLI has an inner sandbox that must be permitted to clone user namespaces. The provider knows the cross-tenant read protection regime it ships under. The orchestrator's job is purely composition; it must not know that "for codex, use `CODEX_HOME`" — that knowledge belongs in `lib/providers/codex.mjs`.
 This is the same separation-of-concerns principle that has governed every prior amendment to this ADR: provider-specific knowledge lives in the provider file; the orchestrator stays generic. Amendment 7's `doctorChecks()` followed it (per-provider repair recipes); Amendment 8's `quotaStatus()` followed it (per-provider probe authorities); this amendment follows it for isolation primitives.
 #### Decision — add OPTIONAL `ISOLATION` named export to the Provider plugin module
 Each provider plugin module (`lib/providers/<name>.mjs`) MAY export, in addition to the default-exported provider object, a named const `ISOLATION` describing the isolation primitives the orchestrator should compose for spawns of this provider. The shape is:
 ```javascript
 export const ISOLATION = {
  ephemeralEnvOverrides: ({ ephemeralRoot, keyId, reqId }) => ({ /* env var map */ }),
  credentialMounts: [ [srcAbsPath, dstRelativeToEphemeralRoot], ... ],
  requiredHomePaths: [ /* dirs to mkdir empty under ephemeralRoot */ ],
  hasInnerSandbox: boolean,
  crossTenantReadProtection: 'tool-suppression' | 'inner-sandbox' | 'none',
  recommendedDeploymentTier: 'shared-os-user' | 'per-os-user' | 'separate-vm',
  toolHardeningArgs: (existingArgs) => modifiedArgs,  // optional
 }
 ```
 The export is **optional**. A plugin that omits `ISOLATION` continues to spawn under the legacy unsandboxed shape exactly as it does today — see § Backward compatibility below. The opt-in surface is consistent with Amendment 7's `doctorChecks()` treatment (additive, no breakage for plugins that haven't been touched).
 The remainder of this amendment specifies each field's semantics, default-when-absent behavior, validation rules, and authority citations. The three currently-shipped providers' concrete declarations are specified in § Per-provider concrete instances.
 #### Field specification
 ##### 1. `ephemeralEnvOverrides({ ephemeralRoot, keyId, reqId }) → { [envVar]: string }`
 **Type and semantics.** A pure (no-side-effect, no-fs-touch) function that, given the orchestrator's composed context (`ephemeralRoot`: absolute path to the spawn-scoped temp dir; `keyId`: the OLP key identity from `lib/keys.mjs` driving the request; `reqId`: the per-request UUID), returns a flat object of environment variables that the orchestrator will merge into the spawn env. The returned env vars are how the provider CLI is steered to read its credentials from the ephemeral root rather than the server process's actual home directory.
 **Why a function and not a static object.** Because `ephemeralRoot` is generated per-spawn by `mkdtemp` and is not known at plugin load time. Because `keyId` and `reqId` are not known until the request arrives. A static object cannot carry the dependency on these values; a function carries it cleanly.
 **Purity contract.** The function MUST be referentially transparent w.r.t. its argument object: identical input arguments yield identical output env maps. It MUST NOT read the filesystem, spawn subprocesses, or mutate the input arguments. It MUST NOT close over module-level mutable state. This contract is what makes the spawn pipeline auditable: a reviewer reading `provider.ISOLATION.ephemeralEnvOverrides({ ephemeralRoot: '/tmp/x', keyId: 'k1', reqId: 'r1' })` can know the full env mutation without running the system.
 **Default behavior when absent.** When `ISOLATION` is absent or `ISOLATION.ephemeralEnvOverrides` is missing, the orchestrator MUST emit no environment overrides for that provider — `child_process.spawn` runs with `process.env` (possibly modified by other contract layers such as the existing `spawn()` method's env cleanup, ADR 0009 Amendment 1's `--system-prompt` injection, etc.). This preserves Phase 6c / pre-Phase 7 behavior exactly.
 **Validation rules.** At plugin load (in `validateProvider` or a sibling `validateIsolation` helper):
 - If `ISOLATION` is defined and `ephemeralEnvOverrides` is defined, it MUST be a function. A non-function value (e.g., a static object) is a load-time error.
 - The function is NOT invoked at load time — its return shape is not validated until first spawn. Load-time invocation would require synthetic dummy arguments and would couple the validator to the orchestrator's argument shape (which itself may evolve under future ADR 0014 amendments).
 - First-spawn invocation MUST validate the return value is a plain object whose values are all strings. Non-string values (numbers, booleans, undefined) MUST cause the spawn to abort with a clear error rather than coerce silently — the env block crosses a kernel boundary and silent coercion is a footgun.
 **Authority citation requirement.** Each env var returned must correspond to a documented credential-resolution lookup in the underlying provider CLI. For example, `HOME` is a POSIX convention for credential lookup (well-established, no citation needed beyond the POSIX umbrella). `CODEX_HOME` is documented (primary) at https://developers.openai.com/codex/config-reference (2 occurrences verified 2026-05-29: `$CODEX_HOME/profile-name.config.toml` and `$CODEX_HOME/log` path templates), with secondary corroboration at https://developers.openai.com/codex/auth/ (2 occurrences in the credential-storage section: `auth.json under CODEX_HOME`). `VIBE_HOME` is documented at https://docs.mistral.ai/mistral-vibe/terminal/configuration (3 occurrences verified 2026-05-29: descriptive sentence "Override the location with the `VIBE_HOME` environment variable", canonical `export VIBE_HOME="/path/to/custom/vibe/home"` example, and an enumeration of files/directories `VIBE_HOME` affects). The provider plugin author MUST cite the underlying CLI's env-var documentation in the plugin file's header (the same place existing CLI-flag citations live, per Rule 1 of `ALIGNMENT.md`).
 Inventing an env var the provider CLI does not actually honor (e.g., setting `MISTRAL_HOME=...` when no such env var exists) is a Rule 2 violation and is unalignable per Rule 4 of `ALIGNMENT.md`.
 ##### 2. `credentialMounts: [ [srcAbsPath, dstRelativeToEphemeralRoot], ... ]`
 **Type and semantics.** An array of `[src, dst]` tuples describing how the server process's real on-disk credential artifacts (OAuth tokens, API keys, refresh artifacts) are made available inside the ephemeral home. The orchestrator iterates this list and, for each tuple, ensures `<ephemeralRoot>/<dst>` resolves (via symlink, copy, or bind-mount depending on platform and constraints) to the data at `<src>`.
 The mount strategy is a property of the orchestrator, not the provider — `lib/sandbox/manager.mjs` decides between symlink (cheapest, on macOS and unconfined Linux), copy (when crossing a namespace boundary that breaks symlinks), and bind-mount (under a future bwrap-equipped path). The provider only declares the source-destination correspondence.
 **Why this is a list, not a function.** The mounts are static per-provider: anthropic always mounts `~/.claude/.credentials.json`, codex always mounts `~/.codex/auth.json`. A function form would invite plugin authors to compute mount paths from per-request state, which would be a security hazard (per-request mount lists are harder to audit at code-review time). Forcing the static form makes the credential surface visible by `grep ISOLATION lib/providers/*.mjs`.
 **Default behavior when absent.** Empty mount list — the spawn sees no credential files in its ephemeral home. For most providers this means authentication fails and the spawn errors out cleanly; the orchestrator MUST log a clear "no credentialMounts declared" message before allowing the spawn to proceed, since the most common cause is "plugin author forgot to declare the mount."
 **Validation rules.**
 - Each entry MUST be a 2-tuple (length-2 array). Single-element entries or 3+-tuples are load-time errors.
 - `srcAbsPath` MUST be an absolute path (starts with `/`). Relative paths or `~/`-prefixed paths are load-time errors — the plugin author must call `os.homedir()` explicitly. Rationale: `~/` expansion semantics vary between Node and shells and would silently break under the per-spawn ephemeral home (where `HOME` is rewritten).
 - `dstRelativeToEphemeralRoot` MUST NOT start with `..` (no parent-directory escape) and MUST NOT be absolute (no `/etc/passwd` overlay attempts). Both are load-time errors. The orchestrator's path-composition (`path.join(ephemeralRoot, dst)`) is the *only* path-resolution step that touches the destination — the validation forbids constructions that could escape `ephemeralRoot` even before composition.
 - `srcAbsPath` MAY refer to a path that does not exist at plugin-load time. The orchestrator's mount step does a `existsSync(src)` check at spawn-time and logs a "credential source missing" warning rather than failing the spawn — this is consistent with the existing `auth.path` field behavior in the Provider contract (an absent credential file is an auth condition, not a load-time error).
 - Two mounts with the same `dst` is a load-time error (no implicit ordering or override).
 **Authority citation requirement.** Each `srcAbsPath` MUST correspond to the credential location documented by the underlying provider CLI. For anthropic: `~/.claude/.credentials.json` is the OAuth artifact per `claude` CLI docs (already cited by the plugin's `auth.path` field). For codex: `~/.codex/auth.json` per https://developers.openai.com/codex/auth/. For mistral: `~/.vibe/.env` per https://docs.mistral.ai/mistral-vibe/terminal/configuration. Plugin authors MUST cite the same authority as the `auth.path` field they already declare — the citations should be consistent.
 ##### 3. `requiredHomePaths: [ /* relative paths */ ]`
 **Type and semantics.** An array of relative paths (e.g., `['.claude', '.claude/logs']`) that the orchestrator MUST `mkdir -p` under `ephemeralRoot` before any `credentialMounts` are processed and before the spawn begins. These are directories the provider CLI expects to exist in `HOME` and will fail or behave incorrectly if they're absent (e.g., logging directories that the CLI doesn't auto-create).
 **Why a separate field from `credentialMounts`.** Some providers expect empty directories — not mounted credential files — at certain paths. Treating "empty directory" as a mount with src=null would muddle the validation rules for `credentialMounts`. A dedicated list is cleaner.
 **Default behavior when absent.** Empty list — only the directories implied by `credentialMounts[i].dst` (their parent dirs, created by `mkdir -p` during the mount step) exist under `ephemeralRoot`. For most providers this is fine.
 **Validation rules.**
 - Each entry MUST be a relative path string. Same anti-escape rules as `credentialMounts[i].dst`: no leading `..`, no absolute paths.
 - Entries MAY overlap with `credentialMounts[i].dst` parent paths (no error; orchestrator's `mkdir -p` is idempotent).
 - Duplicate entries are not an error (idempotent), but the linter / future CI grep should flag them as a code smell.
 **Authority citation requirement.** None directly required for the path values themselves — these are typically convention (e.g., `.claude` mirrors the CLI's expected `$HOME/.claude` layout). However, if a plugin declares a `requiredHomePaths` entry that does not correspond to any documented CLI behavior, the plugin's header comment should explain *why* the directory must exist (observed behavior, error message from CLI, etc.). Speculative directories ("just in case the CLI wants this") are a Rule 2 violation — only directories whose absence is known to cause CLI failure should be listed.
 ##### 4. `hasInnerSandbox: boolean`
 **Type and semantics.** A boolean flag declaring whether this provider's CLI spawns its own internal sandbox boundary during normal operation. The orchestrator uses this flag to decide whether the outer isolation primitives need to be loosened to permit nested sandboxing (e.g., allow `clone(CLONE_NEWUSER)` syscalls, permit `bwrap` to nest).
 **Why a boolean and not an enum.** "Has inner sandbox or not" is the discriminator the orchestrator needs. The *kind* of inner sandbox (bwrap, sandbox-exec, seccomp-only) is a detail the orchestrator does not need to compose against — it just needs to know whether to relax the outer profile. If a future provider requires per-sandbox-flavor handling, this field can be widened to an enum in a subsequent amendment.
 **Default behavior when absent.** Treated as `false`. This is the safer-by-default value — outer isolation stays at its strictest setting. A provider that actually has an inner sandbox but forgets to declare it will fail at spawn time (inner-bwrap attempts denied by outer profile); the failure mode is loud and obvious, which is the desired behavior.
 **Validation rules.** MUST be a literal `true` or `false`. Truthy/falsy coercion (e.g., declaring `1` or `'yes'`) is a load-time error — booleans are the documented type and coercion would silently change the orchestrator's composition decision.
 **Authority citation requirement.** A `hasInnerSandbox: true` declaration MUST cite the CLI's documented or observed inner-sandbox behavior in the plugin header. For codex, the citation is `openai/codex#16018` (the GitHub issue documenting `codex exec` invoking bubblewrap internally) plus https://developers.openai.com/codex/concepts/sandboxing (the official docs page describing the `--sandbox` flag and `read-only` default). For a hypothetical future provider, the citation is whatever CLI doc or observed-behavior transcript establishes the inner sandbox.
 ##### 5. `crossTenantReadProtection: 'tool-suppression' | 'inner-sandbox' | 'none'`
 **Type and semantics.** A discriminated string declaring the regime under which this provider's spawn is protected against cross-tenant filesystem reads. The three values correspond to the three regimes observed in the 2026-05-27 prior-art / incident analysis (see incident memory § 6):
 - `'tool-suppression'` — the provider's CLI exposes no filesystem-reading tools to the model during the spawn, because OLP suppresses them at the request level. For anthropic, this is achieved via ADR 0009 Amendment 1's `--system-prompt` injection combined with the absence of `--tools` flags: the model has no shell, no file-read, no bash, no Read/Write/Edit primitives. The cross-tenant read surface is closed at the prompt-engineering layer; OS-level isolation is a defense in depth but not the primary regime.
 - `'inner-sandbox'` — the provider's CLI has tool execution (e.g., codex's `shell` tool, which actually runs commands) but the CLI's own inner sandbox prevents the tool from reading paths outside its declared allow-list. For codex, the inner bwrap sandbox enforces `--sandbox read-only` by default (per https://developers.openai.com/codex/concepts/sandboxing), so even though the model can call `shell`, the shell's reads are confined to the inner namespace. The cross-tenant read surface is closed at the inner-sandbox layer.
 - `'none'` — no protection regime is currently established for this provider. The model may have tools that read files, and there is no inner sandbox blocking those reads. Operationally this means the provider should NOT be enabled in a multi-tenant deployment until a regime is established. The orchestrator MUST log a WARN at server boot when a provider with `crossTenantReadProtection: 'none'` is enabled in a deployment with >1 active OLP key — observability, not enforcement (see Rule 4 compliance below).
 **Why a discriminated enum, not a free-form string.** The orchestrator and the operator dashboard both consume this field. Free-form values would require every consumer to perform string-matching against a moving target. The enum locks the consumer surface; future regimes are added by amending this list in a subsequent ADR 0002 amendment.
 **Default behavior when absent.** Treated as `'none'`. Safer-by-default in the WARN sense (operators get the WARN log) but NOT in the security sense (no protection is actually applied). This is intentional: the orchestrator cannot fabricate a protection regime the plugin hasn't implemented; the WARN nudges the plugin author to declare honestly.
 **Validation rules.** MUST be one of the three enum values literally. Any other string is a load-time error. The orchestrator MUST log the field's value at server startup so operators can audit the protection picture across providers at a glance.
 **Authority citation requirement.**
 - `'tool-suppression'` declarations MUST cite the suppression mechanism (e.g., for anthropic: ADR 0009 Amendment 1 § "--system-prompt" + the absence-of-tools posture documented at the incident memory § 6.1).
 - `'inner-sandbox'` declarations MUST cite the CLI doc or observed behavior establishing the inner sandbox (e.g., for codex: `openai/codex#16018` + https://developers.openai.com/codex/concepts/sandboxing).
 - `'none'` is the safer default and requires no citation but MUST be accompanied by a header-comment TODO documenting what regime is expected to be established when the provider transitions from Candidate to Enabled (or earlier if the provider is enabled in a multi-tenant context).
 ##### 6. `recommendedDeploymentTier: 'shared-os-user' | 'per-os-user' | 'separate-vm'`
 **Type and semantics.** A discriminated string giving operators a deployment-topology recommendation for this provider in a multi-tenant context. The three values express increasing degrees of operator-side isolation:
 - `'shared-os-user'` — the OLP server process runs as a single OS user, and multiple OLP keys share that user. Protection against cross-tenant leakage rests entirely on the provider's `crossTenantReadProtection` regime + the orchestrator's ephemeral-home composition. This is the recommended posture for providers where `crossTenantReadProtection` is `'tool-suppression'` AND `hasInnerSandbox: false` (i.e., the model has no filesystem-touching tools at all).
 - `'per-os-user'` — each OLP key (or each tenant) should map to a separate OS user, with file-permission-level isolation between tenants. The recommended posture for providers with `crossTenantReadProtection: 'inner-sandbox'` — the inner sandbox protects against accidental leakage from the model's tools, but a sandbox-escape (e.g., a CVE in bubblewrap, a misconfigured inner profile) would expose the OS-user filesystem; per-OS-user isolation adds defense in depth.
 - `'separate-vm'` — the provider should not be co-located with any other tenant on the same VM. The recommended posture for providers with `crossTenantReadProtection: 'none'` AND/OR ones where the operator has reason to distrust the inner sandbox's quality. Practically this means the provider should not be enabled in OLP's family-LAN deployment unless the family-LAN host runs only this tenant.
 **Why a recommendation and not a hard policy.** The orchestrator and OLP runtime cannot *enforce* OS-user separation or VM separation — those are properties of the host operator's deployment topology. This field is informational: it surfaces in `/health.providers.<name>.isolation` (a Phase 7 addition planned in a follow-up amendment) and in the dashboard, so operators making deployment decisions have the per-provider recommendation visible. Operator override is the expected normal path: a deployment that knowingly accepts the risk of running an `'separate-vm'` provider in a shared-user context is acceptable, just observable.
 **Default behavior when absent.** Treated as `'separate-vm'` — the safest recommendation in the absence of declared analysis. The WARN log emitted for missing `ISOLATION` blocks (see Rule 4 compliance below) covers operator visibility.
 **Validation rules.** MUST be one of the three enum values literally. Any other string is a load-time error.
 **Authority citation requirement.** The plugin author MUST cite the basis for the recommendation in the plugin header — typically a short paragraph reasoning about the combination of `hasInnerSandbox` and `crossTenantReadProtection` for this provider. The reasoning is not a CLI authority citation (the underlying CLI does not declare deployment topology); it is an OLP-side analysis. The expected citation form is `# isolation rationale: <2-3 sentences> (cf. ADR 0014 Amendment 1 § <relevant section>)`.
 ##### 7. `toolHardeningArgs: (existingArgs) => modifiedArgs` (OPTIONAL)
 **Type and semantics.** An OPTIONAL pure function that, given the plugin's `spawn()` method's CLI args (the array passed to `child_process.spawn`), returns a (possibly modified) args array with additional tool-hardening flags inserted. The orchestrator calls this hook after the plugin's `spawn()` constructs its args but before the actual `child_process.spawn` invocation.
 **Purpose.** Some providers expose CLI flags that suppress or restrict the model's tool access at the per-spawn level (e.g., `--disallowedTools` on `claude`, or `--sandbox read-only` on `codex`). These flags are the *enforcement mechanism* corresponding to the `crossTenantReadProtection` *declaration*. Splitting the declaration (a static field) from the enforcement (a function that mutates args) keeps the contract auditable while letting the enforcement evolve as the underlying CLI's flag set changes.
 **Why this is OPTIONAL.** For providers where `crossTenantReadProtection: 'tool-suppression'` is achieved entirely via the `spawn()` method's existing args construction (e.g., the existing anthropic.mjs `--system-prompt` injection), no separate hardening step is needed — the field can be omitted. For providers where the orchestrator needs to inject additional flags atop the plugin's base args, the field provides the hook.
 **Default behavior when absent.** No args modification — the plugin's `spawn()` method's args are passed through to `child_process.spawn` unchanged. This is the current Phase 6c behavior for anthropic and is appropriate when the `spawn()` method already encodes the hardening.
 **Validation rules.**
 - If declared, MUST be a function.
 - First-spawn invocation MUST validate the return value is an array of strings. Non-array or non-string-element returns abort the spawn (silent coercion is unsafe at the kernel boundary).
 - The function MUST be referentially transparent — same input array yields same output array (no module-level state, no fs reads).
 - The orchestrator MUST NOT pass the args by reference in a way that the function could mutate the original `existingArgs`. The hook receives a defensive copy; returning a fresh array is required.
 **Authority citation requirement.** The injected flags MUST be documented CLI flags of the underlying provider. Inventing a `--disable-tools` flag that the CLI does not support is a Rule 2 violation. For codex, citing https://developers.openai.com/codex/concepts/sandboxing § `--sandbox` is sufficient. For anthropic, the existing ADR 0009 Amendment 1 citation covers the tool-suppression mechanism.
 #### Per-provider concrete instances
 The three currently-shipped providers declare `ISOLATION` as follows. Each declaration MUST be present in the corresponding plugin file before that provider can be enabled in any multi-tenant deployment (see § Rule 4 compliance and § Backward compatibility for the transition path).
 ##### anthropic
 ```javascript
 // lib/providers/anthropic.mjs
 //
 // isolation rationale: Anthropic Claude reaches OLP via stream-json transport
 // without a tool surface (ADR 0009 Amendment 1's --system-prompt injection
 // suppresses env-block, file tools, bash, and Read/Write/Edit). The model
 // has no documented mechanism to read files during the spawn. Cross-tenant
 // read protection is achieved at the prompt-engineering / CLI-flag layer.
 // The OS-level isolation primitives (HOME redirect + ephemeral credential
 // mount) add defense in depth against future CLI changes that might
 // re-introduce a tool surface.
 //
 // Authority: @anthropic-ai/claude-code v2.1.150 § --system-prompt
 //   (ADR 0009 Amendment 1 + incident memory § 6.1 establishes the
 //   tool-suppression mechanism); HOME env conventional POSIX behavior.
 export const ISOLATION = {
  ephemeralEnvOverrides: ({ ephemeralRoot, keyId, reqId }) => ({
    HOME: ephemeralRoot,
    // CLAUDE_CONFIG_DIR is NOT honored as of v2.1.150 — the CLI reads from
    // $HOME/.claude/.credentials.json. Redirecting HOME is the documented
    // mechanism. The keyId / reqId arguments are unused here but received for
    // signature consistency with codex's overrides.
  }),
  credentialMounts: [
    // OAuth artifact location. Authority: existing anthropic.mjs `auth.path`
    // field — `~/.claude/.credentials.json` is the documented OAuth artifact.
    // The orchestrator resolves the absolute src path via os.homedir() at
    // load time (the plugin file shows the literal `join(homedir(), ...)`).
    [/* resolved at load: */ '<homedir>/.claude/.credentials.json',
     '.claude/.credentials.json'],
  ],
  requiredHomePaths: [
    '.claude',
    // No observed behavior requires additional dirs; CLI creates session logs
    // under .claude/ on demand. If future CLI versions add a mandatory pre-
    // existing subdir, add it here with an observed-behavior comment.
  ],
  hasInnerSandbox: false,
  crossTenantReadProtection: 'tool-suppression',
  recommendedDeploymentTier: 'shared-os-user',
  // toolHardeningArgs omitted — the existing spawn() method's args already
  // encode the --system-prompt suppression (ADR 0009 Amendment 1).
 };
 ```
 **Authority pin for the anthropic ISOLATION declaration:**
 - `--system-prompt` mechanism: ADR 0009 Amendment 1 + incident memory `~/.cc-rules/memory/projects/olp/incident_2026_05_27_spawn_cli_security.md` § 6.1
 - `HOME` env redirect: POSIX convention; `claude` CLI v2.1.150 observed to read `~/.claude/.credentials.json` via HOME (verified by the PR-B PI231 spike, confirmed by ADR 0014 Amendment 1's HOME-override verification task)
 ##### codex
 ```javascript
 // lib/providers/codex.mjs
 //
 // isolation rationale: OpenAI Codex's `codex exec` exposes a shell tool that
 // actually executes commands during the spawn (incident memory § 3.2). The
 // CLI provides its own inner bubblewrap sandbox (`--sandbox read-only` by
 // default per https://developers.openai.com/codex/concepts/sandboxing) that
 // confines shell tool reads/writes. The orchestrator's outer isolation
 // composes with the inner sandbox: HOME-equivalent redirect via CODEX_HOME
 // (per https://developers.openai.com/codex/config-reference) plus per-spawn
 // ephemeral credential mount. hasInnerSandbox: true so the outer profile is
 // relaxed to permit inner bwrap's user-namespace clone.
 //
 // Authority: openai/codex#16018 (inner bwrap behavior);
 //   https://developers.openai.com/codex/concepts/sandboxing (--sandbox flag);
 //   https://developers.openai.com/codex/config-reference (CODEX_HOME);
 //   https://developers.openai.com/codex/auth/ (~/.codex/auth.json path).
 export const ISOLATION = {
  ephemeralEnvOverrides: ({ ephemeralRoot, keyId, reqId }) => ({
    // CODEX_HOME overrides the base config / credential dir. Docs:
    // https://developers.openai.com/codex/config-reference and
    // https://developers.openai.com/codex/auth/
    CODEX_HOME: `${ephemeralRoot}/.codex`,
    // HOME also redirected for codex's own bubblewrap-internal HOME lookup
    // (the inner sandbox inherits parent HOME unless overridden).
    HOME: ephemeralRoot,
  }),
  credentialMounts: [
    // Auth artifact location. Authority: existing codex.mjs `auth.path` field
    // (Codex CLI reference § Authentication, plus
    // https://developers.openai.com/codex/auth/ canonical pin).
    [/* resolved at load: */ '<homedir>/.codex/auth.json',
     '.codex/auth.json'],
  ],
  requiredHomePaths: [
    '.codex',
    // Inner bwrap may create additional state under .codex/. If observed
    // behavior shows the CLI failing on absent subdirs, add them here.
  ],
  hasInnerSandbox: true,
  crossTenantReadProtection: 'inner-sandbox',
  recommendedDeploymentTier: 'per-os-user',
  toolHardeningArgs: (existingArgs) => {
    // If the operator has not explicitly passed --sandbox, inject the
    // documented read-only default. Per
    // https://developers.openai.com/codex/concepts/sandboxing the default
    // posture is `read-only`; this hardening hook makes the default explicit
    // at the spawn args level so a future CLI default change does not
    // silently weaken the isolation.
    if (existingArgs.some(arg => arg === '--sandbox' || arg.startsWith('--sandbox='))) {
      return existingArgs;
    }
    return [...existingArgs, '--sandbox', 'read-only'];
  },
 };
 ```
 **Authority pin for the codex ISOLATION declaration:**
 - `CODEX_HOME`: https://developers.openai.com/codex/config-reference (retrieved 2026-05-29)
 - `~/.codex/auth.json`: https://developers.openai.com/codex/auth/ (existing `auth.path` citation in codex.mjs)
 - Inner bwrap behavior: `openai/codex#16018` plus https://developers.openai.com/codex/concepts/sandboxing
 - `--sandbox read-only`: https://developers.openai.com/codex/concepts/sandboxing § "Sandboxing modes"
 ##### mistral
 ```javascript
 // lib/providers/mistral.mjs
 //
 // isolation rationale: Mistral Vibe ships at OLP Phase 7 with no known
 // equivalent to Anthropic's Phase 6c --system-prompt tool suppression and
 // no known inner sandbox. The IR-level normalization shipped at D8 does not
 // suppress tools at the CLI layer. Cross-tenant read protection is therefore
 // 'none' — the provider should not be enabled in a multi-tenant deployment
 // until a regime is established. The declaration here exists so the
 // orchestrator can compose ephemeral-home credential isolation (which still
 // works) while the operator sees a clear WARN that the tool-side protection
 // is not in place.
 //
 // Authority: TBD — a spike task tracked at Phase 7 follow-up (see Open
 //   Questions section below) will verify Vibe CLI's tool surface and inner
 //   sandbox posture against https://docs.mistral.ai/mistral-vibe/terminal/.
 //   Until that spike lands, this declaration documents the current honest
 //   state per ALIGNMENT.md Rule 3 (Match the Implementation): no protection
 //   is encoded because none has been established.
 export const ISOLATION = {
  ephemeralEnvOverrides: ({ ephemeralRoot, keyId, reqId }) => ({
    // VIBE_HOME is documented at
    // https://docs.mistral.ai/mistral-vibe/terminal/configuration as the
    // env var that overrides the default ~/.vibe/ base directory
    // (3 occurrences verified 2026-05-29: descriptive sentence,
    // canonical export example, and an enumeration of files/dirs the
    // variable affects). Task #4 PI231 spike verifies observed CLI
    // behaviour matches the documented contract.
    VIBE_HOME: `${ephemeralRoot}/.vibe`,
    HOME: ephemeralRoot,
  }),
  credentialMounts: [
    // ~/.vibe/.env per existing mistral.mjs `auth.path` field, sourced from
    // https://docs.mistral.ai/mistral-vibe/terminal/configuration.
    [/* resolved at load: */ '<homedir>/.vibe/.env', '.vibe/.env'],
  ],
  requiredHomePaths: [
    '.vibe',
  ],
  hasInnerSandbox: false,
  crossTenantReadProtection: 'none',
  recommendedDeploymentTier: 'separate-vm',
  // toolHardeningArgs omitted — no CLI hardening flag is currently known for
  // Vibe. The Phase 7 spike will revisit.
 };
 ```
 **Authority pin for the mistral ISOLATION declaration:**
 - `VIBE_HOME`: https://docs.mistral.ai/mistral-vibe/terminal/configuration (3 occurrences verified 2026-05-29: descriptive sentence "Override the location with the `VIBE_HOME` environment variable", canonical `export VIBE_HOME="/path/to/custom/vibe/home"` example, and the enumeration of files/directories `VIBE_HOME` affects).
 - `~/.vibe/.env`: same source (existing `auth.path` citation in mistral.mjs).
 - **Open spike (Phase 7 follow-up, Task #4):** verify *observed CLI behaviour* matches *documented behaviour* — (a) Vibe CLI actually honours the documented `VIBE_HOME` env var during spawn; (b) Vibe CLI's tool surface (shell, file-read, etc.) during a `vibe --prompt` spawn; (c) any CLI sandbox or tool-suppression flag. Findings may transition `crossTenantReadProtection` from `'none'` to `'tool-suppression'` or `'inner-sandbox'` if a hardening regime is discovered. The spike is verification-grade, not authority-pin work.
 #### Backward compatibility
 A plugin that does NOT export `ISOLATION` continues to work exactly as it does today. The orchestrator's `prepareIsolatedEnvironment(provider, ctx)` function MUST detect the absence of `provider.ISOLATION` (or the absence of any individual field within it) and fall through to the legacy unsandboxed code path for that spawn. The legacy path is:
 - No ephemeral root created
 - No env overrides
 - No credential mounts
 - `cwd: process.cwd()` (the server's working directory)
 - `env: process.env` (composed with whatever the plugin's `spawn()` method's existing env logic produces)
 This is the same behavior as Phase 6c. No provider plugin is broken by Amendment 9's landing.
 Plugins MAY adopt `ISOLATION` incrementally: a plugin that wants the credential-mount benefit but has not yet analyzed its cross-tenant tool surface MAY declare `crossTenantReadProtection: 'none'` and `recommendedDeploymentTier: 'separate-vm'` (the safer-by-default values). The orchestrator will compose the credential isolation correctly; the WARN log nudges follow-up.
 #### Rule 4 compliance (ALIGNMENT.md)
 ALIGNMENT.md Rule 4 states: "Unalignable plugins / fields are deleted, not feature-flagged." This amendment introduces an OPTIONAL contract field, which on its face could be read as "feature-flagging" isolation. The reading is wrong, and the distinction is important enough to spell out:
 - Amendment 9 does NOT introduce an `ISOLATION` feature flag that operators or plugins toggle on/off. The field's presence/absence describes **the provider's truthful isolation posture** at a point in time. A plugin without `ISOLATION` declares (implicitly) that no analysis has been done and the safer-by-default treatment applies.
 - The OPTIONAL nature is purely transitional. Existing plugins ship without it; they continue to spawn (in their existing single-tenant developer-laptop posture). The orchestrator's WARN log surfaces the absence to the operator at server boot. An operator running a multi-tenant deployment with un-declared plugins is operating off-recommendation but not blocked.
 - A plugin that declares `ISOLATION` with values the orchestrator cannot honor (e.g., a `credentialMounts` entry pointing at a path that does not exist, or an `ephemeralEnvOverrides` function that returns non-string values) MUST fail at first spawn — the orchestrator does not silently fall back to the no-ISOLATION path. This is the Rule 4 enforcement vector: a *broken* declaration is unalignable and surfaces loudly; a *missing* declaration is the safer transitional state.
 The WARN at server boot is observability, not enforcement. It reads approximately:
 ```
 [WARN] provider "<name>" does not declare ISOLATION; spawns will run
       under legacy unsandboxed shape. Recommended in multi-tenant
       deployments: declare ISOLATION per ADR 0002 Amendment 9.
 ```
 Operators in single-tenant developer deployments may safely ignore the WARN. Operators in multi-tenant deployments should treat it as a Phase 7 follow-up task.
 #### Interaction with prior amendments
 - **Amendment 1 (`maxSpawnTimeMs`).** Independent. The spawn-timeout enforcement lives inside each plugin's spawn drain loop; the orchestrator's ISOLATION composition happens *before* the spawn, so the two amendments compose without conflict.
 - **Amendment 3 (`cacheable`).** Independent. The cache layer decides whether to call the orchestrator at all; once the orchestrator is reached, ISOLATION composition is orthogonal to cacheability.
 - **Amendment 4 (`contractVersion`).** Independent. `contractVersion: '1.0'` plugins MAY add an `ISOLATION` export under Amendment 9 without bumping the contract version — `ISOLATION` is an additive named export, not a v1.0 contract surface change. A future Provider contract v1.1 may promote `ISOLATION` to a required field (forcing all enabled plugins to declare); that decision is deferred to a future amendment, gated on the Phase 7 follow-up findings.
 - **Amendment 6 (`maxConcurrent` runtime enforcement).** Independent. The semaphore acquire happens before the orchestrator's `prepareIsolatedEnvironment`; the release happens after the spawn drains. ISOLATION composition is bracketed by the semaphore, not entangled with it.
 - **Amendment 7 (`doctorChecks()`).** Adjacent. A future plugin may add an `<provider>.isolation_declared` doctor check that reports whether `ISOLATION` is declared and whether its referenced credential paths resolve. The check is OPTIONAL per Amendment 7's framework and is appropriate for `olp doctor` operator UX.
 - **Amendment 8 (`quotaStatus()` direct-API exemption).** Independent. The quota probe runs outside the spawn pipeline (direct HTTPS from server process); it does not interact with `ISOLATION` composition.
 #### Companion ADR
 This amendment is the companion governance piece for **ADR 0014 Amendment 1** (the Phase 7 architectural shift from outer-bwrap PR-B to per-spawn ephemeral-home + per-provider primitives). ADR 0014 Amendment 1 describes the orchestrator's composition algorithm and the rationale for retiring the outer-bwrap approach; ADR 0002 Amendment 9 (this section) describes the contract surface the orchestrator reads.
 The two amendments are reviewed and merged together as a single coupled commit (Iron Rule 11 — minimum reviewable unit per layer). Reviewing them separately cannot verify producer-consumer alignment: the orchestrator's algorithm is meaningless without the contract it consumes, and the contract is meaningless without the orchestrator's composition discipline.
 #### Tests
 Test coverage for Amendment 9 lands as a new Suite in `test-features.mjs` co-merged with ADR 0014 Amendment 1's `lib/sandbox/manager.mjs` refactor. The suite covers:
 1. `validateProvider` (or `validateIsolation` helper) rejects each documented invalid shape: non-function `ephemeralEnvOverrides`; non-2-tuple `credentialMounts` entries; `dst` paths starting with `..` or absolute; non-boolean `hasInnerSandbox`; out-of-enum `crossTenantReadProtection`; out-of-enum `recommendedDeploymentTier`; non-function `toolHardeningArgs`.
 2. The legacy code path: a fake provider without `ISOLATION` spawns under the existing shape unchanged. Existing Phase 6c tests for anthropic continue to pass.
 3. The ephemeral-home composition path: a fake provider declaring a minimal `ISOLATION` block has its env overrides applied and its credential mount resolved into a `mkdtemp`-created ephemeral root.
 4. First-spawn return-shape validation: `ephemeralEnvOverrides` returning non-string values aborts the spawn loudly; `toolHardeningArgs` returning a non-array aborts the spawn loudly.
 5. Per-shipped-provider declaration smoke: each of `anthropic`, `codex`, `mistral` declares an `ISOLATION` block; each block's `credentialMounts[i][0]` (when resolved against the running user's `homedir()`) matches the plugin's `auth.path` field.
 The full test list is captured in ADR 0014 Amendment 1's PR-B-revised test suite specification.
 #### Open questions (Phase 7 follow-up)
 1. **Mistral Vibe tool surface and inner sandbox.** The mistral plugin's `ISOLATION` declares `crossTenantReadProtection: 'none'` honestly. A spike task is required to determine whether Vibe CLI exposes any tool surface and/or any sandbox flag; findings update the declaration. Tracked at the Phase 7 work plan.
 2. **HOME-only providers vs CODEX_HOME-style providers.** The current contract assumes credential redirection happens via env-var rewriting (`HOME` or `<PROVIDER>_HOME`). A future provider that hardcodes its credential path (no env override) would be unable to honor the contract and would need a different isolation strategy (e.g., bind-mount of the literal path). This is not a current problem (all three shipped providers honor env overrides) but should be tracked for future inclusion ADRs.
 3. **Promoting `ISOLATION` to required at contract v1.1.** Once all enabled providers declare `ISOLATION`, a future contract-version bump may promote the field from OPTIONAL to REQUIRED. The decision is gated on operational experience after PI231 + cloud deployment — see ADR 0014 Amendment 1 for the rollout milestones.
 4. **Per-spawn vs per-key ephemeral root.** This amendment specifies per-spawn ephemeral roots (one `mkdtemp` per `provider.spawn` call). A future optimization may cache ephemeral roots per-key (one ephemeral root per OLP key identity, reused across spawns) to reduce mkdtemp / mount overhead. The contract surface here is compatible with either strategy; the choice is an orchestrator implementation detail.
 5. **Cleanup discipline.** The orchestrator is responsible for `rm -rf`-ing the ephemeral root after the spawn drains. The cleanup mechanism (synchronous vs deferred, error vs success path symmetry) is specified in ADR 0014 Amendment 1, not here. This amendment notes the dependency for completeness.
 #### Authority citations summary
 | Field | Authority |
 |---|---|
 | `ephemeralEnvOverrides` (general) | POSIX `HOME` convention; per-provider env-var documentation cited per declaration |
 | `credentialMounts` (general) | Each plugin's existing `auth.path` field citation |
 | `requiredHomePaths` (general) | Observed CLI behavior; no speculative entries (Rule 2) |
 | `hasInnerSandbox` (general) | CLI doc or observed-behavior transcript |
 | `crossTenantReadProtection` (enum) | OLP-side analysis based on prior-art search in incident memory `~/.cc-rules/memory/projects/olp/incident_2026_05_27_spawn_cli_security.md` § 4 + § 6 |
 | `recommendedDeploymentTier` (enum) | OLP-side analysis; ADR 0014 Amendment 1 § Deployment topology |
 | `toolHardeningArgs` (function) | Documented CLI flags of the underlying provider; no invented flags (Rule 2) |
 | anthropic `--system-prompt` tool suppression | ADR 0009 Amendment 1 + incident memory § 6.1 |
 | codex `CODEX_HOME` | https://developers.openai.com/codex/config-reference + https://developers.openai.com/codex/auth/ |
 | codex inner bwrap | openai/codex#16018 + https://developers.openai.com/codex/concepts/sandboxing |
 | codex `--sandbox read-only` default | https://developers.openai.com/codex/concepts/sandboxing § "Sandboxing modes" |
 | mistral `VIBE_HOME` and `.vibe/.env` | https://docs.mistral.ai/mistral-vibe/terminal/configuration |
 #### Procedural mechanism
 - **Iron Rule 11 (Incremental Diff Review)** — Amendment 9 (governance, ADR 0002) and ADR 0014 Amendment 1 (orchestrator architecture) land as a single coupled PR. Reviewing them separately cannot verify consumer-producer alignment.
 - **Iron Rule 10 (Code Review)** — independent fresh-context reviewer per `CLAUDE.md` hard requirement #3. The reviewer MUST open each cited authority URL (Codex config-reference, sandboxing docs, Mistral configuration docs, the incident memory) and confirm the citation in the review comment.
 - **`ALIGNMENT.md` Rule 1 (Cite First)** — every per-field design choice is cited above. Every per-provider concrete instance is cited to the underlying CLI authority.
 - **`ALIGNMENT.md` Rule 2 (No Invention)** — no invented env vars, no invented CLI flags. The mistral `crossTenantReadProtection: 'none'` declaration is the explicit honest acknowledgment that no protection regime has been established, rather than invention of one.
 - **`ALIGNMENT.md` Rule 4 (Unalignable Plugins / Fields Are Deleted)** — see § Rule 4 compliance above for the explicit reasoning that OPTIONAL `ISOLATION` is not "feature-flagging" but rather "honestly transitional."
 - **`ALIGNMENT.md` Amendment Procedure** — this section (Amendment 9) is the PR-required citation of evidence (the 2026-05-27 incident memory, the ADR 0014 PoC spike report at `/tmp/sandbox-spike/report.md` on PI231) and the structural amendment of the Provider contract documented in this ADR's § Decision.
@@ -1,6 +1,6 @@
 # ADR 0014 — Sandbox-Runtime Integration for Multi-Tenant Provider Spawning
-**Status:** Accepted (PR-A — deps + doctor + ADR only; PR-B/C/D pending)
+**Status:** Accepted (PR-A shipped; PR-B shipped pending PI231 Suite 44 validation + HTTP-path activation debug; PR-C/D pending) — **see Amendment 1 (2026-05-29): PR-B's outer-bwrap approach is superseded by the per-spawn ephemeral-home + per-provider ISOLATION contract architecture. PR-C/D are reframed; the substantive decision moves into Amendment 1.**
 **Date:** 2026-05-28
 **Phase:** Phase 7
@@ -186,7 +186,7 @@ curl -X POST http://127.0.0.1:4567/v1/chat/completions \
 ```
 Additional criteria:
- `SandboxManager.initialize()` is called once at startup (per ADR 0014 § 5 singleton decision, TBD in PR-B ADR amendment)
+- `SandboxManager.initialize()` is called once at startup (singleton shape follows `@anthropic-ai/sandbox-runtime` v0.0.52 `dist/sandbox/sandbox-manager.js` SandboxManager export, where `reset()` is a process-wide operation; see § 5 Open question 1 for the per-provider config concern still to be resolved)
 - p95 latency overhead of wrapping ≤ 200ms measured over 50 warm requests
 - `checkSandboxAvailability().available === true` reported in `/health.sandbox` after PR-B rolls out
 - All existing Suite 41 tests continue to pass (stream-json transport unaffected)
@@ -289,3 +289,374 @@ These were confirmed empirically or inferred from the library source during the
 ## Status transitions
 - 2026-05-28 — Created. Status: Accepted for PR-A scope. PR-B/C/D pending operational prereqs.
 - 2026-05-28 — PR-A shipped (commit `07d9c8a`).
 - 2026-05-28 — PR-B implementation shipped (commit chain `d0dcd28` → `2864275` → `497b255` → `b1e24b7` → `3551921`). Status: shipped pending PI231 Suite 44 validation + HTTP-path activation debug. `OLP_SANDBOX_DISABLED=1` emergency disable installed (b1e24b7) because the in-process MITM proxy interaction with OLP's HTTP request handler suppressed claude stdout on the HTTP path while the same wrap script produced output when invoked directly from a manual shell. Prod is currently running with `OLP_SANDBOX_DISABLED=1` set.
 - 2026-05-29 — **Amendment 1 — supersede PR-B outer-bwrap with ephemeral-home + per-provider ISOLATION contract.** See § Amendment 1 below. PR-B's `lib/sandbox/manager.mjs` outer-bwrap implementation is archived to branch `phase-7-pr-b-outer-bwrap-snapshot` and superseded; PR-C is reframed as inner-sandbox preservation under the new architecture; PR-D is reframed as the README "Security Model" section. `lib/sandbox/doctor.mjs` is preserved unchanged.
 ---
 # Amendment 1 — Supersede PR-B outer-bwrap with ephemeral-home + per-provider contract (2026-05-29)
 - **Date:** 2026-05-29
 - **Status:** Accepted (governance only — the implementation refactor lands in subsequent PRs per ALIGNMENT.md Rule 1 / Iron Rule 11)
 - **Author:** project maintainer (with AI drafting assistance)
 - **Reviewer:** independent fresh-context reviewer per Iron Rule 10 — pending at this draft
 - **Scope:** This amendment supersedes the implementation strategy of PR-B (the outer-bubblewrap-wrap of `claude` CLI shipped in commits `d0dcd28` → `b1e24b7`). It does NOT supersede the multi-tenant security gap analysis in § 1 of the original ADR, nor the four-tier authority citation list (§ 7), nor `lib/sandbox/doctor.mjs` (preserved unchanged). It DOES supersede the PR-B implementation, the PR-C scope ("wrap codex spawn in the same outer-bwrap pattern with `enableWeakerNestedSandbox`"), and PR-D's framing as "documentation update for cloud rollout unblock".
 ---
 ## A1.1 — Why the substitution is forced (the four forcing reasons)
 PR-B as designed (outer-bwrap wrapping of the `claude` CLI spawn, with the OLP server initializing `SandboxManager` once at boot and every spawn routed through `wrapWithSandbox`) was shipped on 2026-05-28 and disabled on the same day via the `OLP_SANDBOX_DISABLED=1` env-var gate after the HTTP-path activation regression appeared on PI231 (the manual-shell wrap produced claude stdout; the OLP HTTP-request-handler wrap produced none). The 2026-05-28/2026-05-29 follow-up investigation found that the HTTP-path failure was not the whole story — even if the in-process MITM proxy lifecycle issue were debugged, four independent and load-bearing reasons forced the architecture away from outer-bwrap entirely. Each is cited to its primary authority below.
 ### A1.1.1 — Forcing reason #1: Anthropic's stated design intent for `@anthropic-ai/sandbox-runtime`
 The PR-B design used `@anthropic-ai/sandbox-runtime` to wrap the `claude` CLI from the outside. Anthropic's published design intent for the library is the opposite direction of containment: the library is for sandboxing what Claude Code itself *triggers* (tool calls, MCP servers, sub-processes spawned during model execution), not for wrapping Claude Code from outside.
 **Primary citation:** https://www.anthropic.com/engineering/claude-code-sandboxing — "Claude Code sandboxing" engineering blog. The post describes Claude Code's *internal* use of the sandbox-runtime library: the model emits a `tool_use` Bash call → Claude Code wraps the resulting `/bin/sh -c <…>` in a sandbox via `SandboxManager.wrapWithSandbox()` before spawning. The blog also notes the library "can be used to sandbox arbitrary processes, agents and MCP servers" — i.e., it is general-purpose, not Claude-Code-internal-only. **Our reading:** Anthropic's documented and demonstrated usage is *inner-wrap by Claude Code*; outer-wrap of `claude` itself is not documented in the blog and not shown in the post's example invocations. **This is a project-design judgment based on the absence of outer-wrap precedent, not a "don't do this" statement from Anthropic.** The architectural concerns enumerated below (MITM proxy lifecycle, semver leverage) stand on their own merits regardless of how Anthropic frames the library's intended usage.
 **What this means for PR-B's design:** wrapping `claude` from outside with the same library is "out-of-distribution" usage. The library was not designed for, tested against, or documented for the outer-wrap case. Two concrete consequences observed in PR-B:
 1. **MITM proxy collision.** The library starts a per-process local MITM proxy on Linux to inspect HTTPS traffic for allowlisted domains. When the same library is invoked again from inside the sandboxed process (e.g., for any sub-spawn `claude` might do), a second MITM proxy attempt collides. The PR-B implementation never reached this case because it disabled before tripping it, but the architecture invites the collision.
 2. **Inner-sandbox conflict** (see A1.1.3 for codex, but the principle applies generally). Any CLI that itself uses the same library to sandbox its own tool calls is *expected* by Anthropic to be the *holder* of the sandbox, not the *content* of one. The library's `enableWeakerNestedSandbox` option exists precisely to acknowledge this — but only as a partial mitigation.
 The Anthropic design-intent reason is not a "won't work" reason. The PR-B outer-bwrap path did work for the smoke case (manual-shell invocation produced output). The reason is a *don't-do-this* reason: OLP would be the only known user of the library in the outer-wrap configuration, taking on the maintenance burden of a usage pattern Anthropic doesn't test, doesn't document, and doesn't owe semver discipline for. The library is `^0.0.52`. A future minor version bump could break OLP's outer-wrap path without warning. Aligning OLP's use of the library with Anthropic's documented design intent restores semver leverage.
 ### A1.1.2 — Forcing reason #2: The `~/.claude.json` upstream "closed as not planned" — permanent maintenance treadmill for outer-bwrap
 Anthropic's `claude` CLI persists session state (most notably MRU project list, last-used model, telemetry session id) to `~/.claude.json` using a **non-atomic** write pattern: write-in-place rather than write-temp-then-rename. When `~/.claude.json` is on a read-only mount (e.g., `--ro-bind ~/.claude` in a bwrap config), the CLI either silently degrades (some state lost) or outright fails depending on which write path is hit on a given invocation.
 The OLP outer-bwrap path in PR-B mounted `~/.claude` read-only initially, hit this exact issue (the 2026-05-28 fold-in commit `2864275` "allow read ~/.claude + skip wrap under test mock" added the read allowance), and then on the further `b1e24b7` commit the operator had to widen the mount again. The fold-in commit message explicitly notes: "~/.claude is NOT in denyRead. It contains the spawn's own OAuth credentials — claude CLI must read its own auth file to function. Denying read here causes 'Not logged in' failures."
 **Primary citation:** https://github.com/anthropics/claude-code/issues/29250 — upstream issue requesting atomic-write semantics for `~/.claude.json`. Status: closed `not_planned` by GitHub's inactivity bot on 2026-03-31 (`closed_by: github-actions[bot]`, comment text: "Closing for now — inactive for too long"), labeled `duplicate` and `stale`. **No upstream Anthropic maintainer comment in the thread articulates any policy rationale** — the only non-bot comments are the reporter's own follow-ups and a third-party-tool plug. The closure does NOT establish "won't fix" as Anthropic policy; at most it establishes that the issue has not received maintainer engagement and that Anthropic considers it a duplicate of some other tracked report. The forcing argument for ephemeral-home over outer-bwrap rests on the architectural-cost analysis (the maintenance-treadmill description below), not on an alleged upstream policy posture.
 **What this means for the outer-wrap maintenance treadmill:** Every future addition of `claude`-CLI-owned state files (telemetry, cache directories, session locks, MCP registration files, etc.) is, by upstream policy, free to use any write pattern the maintainers prefer. The outer-bwrap pattern requires OLP to track each of these additions and add corresponding `--ro-bind` / `--rw-bind` / write-allowlist entries — forever — because the CLI does not give OLP an enumerable contract surface for "files I will write to." The maintainer-time cost is a permanent recurring tax.
 A non-outer-wrap approach that gives `claude` a fresh, ephemeral home directory inverts this: `claude` is free to invent any state file under its $HOME with any write pattern it chooses; OLP never tracks the list. The treadmill goes away. This is the load-bearing case for Solution 1 even setting aside the codex inner-sandbox issue below.
 ### A1.1.3 — Forcing reason #3: Codex inner-bwrap conflict (multi-provider forcing function)
 The PR-C plan in the original ADR was to wrap the `codex` spawn in the same outer-bwrap pattern as PR-B, with `enableWeakerNestedSandbox: true` set on the `SandboxManager.initialize()` call to allow codex's own internal bubblewrap sandbox to function inside OLP's outer bubblewrap sandbox.
 Empirical investigation (2026-05-29 PI231 prep — to be confirmed in Task #4) and published codex CLI behaviour both indicate this nested-sandbox path is structurally fragile:
 **Primary citation:** https://github.com/openai/codex/issues/16018 — upstream codex CLI issue. The issue body documents that codex's bwrap-based default sandbox **fails outright** in environments lacking unprivileged user namespaces — the reporter quotes the error `bwrap: No permissions to create new namespace, likely because the kernel does not allow non-privileged user namespaces`. The issue is a **feature request by the reporter** asking codex to "suggest or automatically fall back to an alternative supported backend when available"; **the issue body itself does NOT contain the string `danger-full-access` and does NOT document an existing automatic fallback to it**. The codex `--sandbox danger-full-access` mode is a documented *manual* opt-out (https://developers.openai.com/codex/concepts/sandboxing § "Sandboxing modes"). Whether codex automatically degrades into it under nested-bwrap failure — or whether the spawn aborts outright — is an empirical question slated for Task #4 PI231 spike verification.
 In other words: wrapping codex in OLP's outer bwrap, *if* the outer bwrap is configured with sufficient capability to allow the inner clone, requires giving the outer sandbox more capability than the security boundary should grant. *If* it is configured to a tighter, safer capability set, codex's inner-bwrap initialization fails (the documented failure mode per the linked issue). Whether codex then aborts the spawn or silently degrades to `danger-full-access` is empirically open (Task #4); either outcome is undesirable. The strict-additive-isolation invariant (outer-bwrap + inner-bwrap = composed isolation) does not hold for codex under this configuration: either OLP gives up outer-isolation strength to admit the inner clone, or codex's inner isolation breaks in some manner.
 **What this means as a multi-provider forcing function:** OLP is by constitution (ADR 0001 § Mission) a multi-provider proxy. The outer-bwrap architecture cannot cover codex without a security regression. The structural response is to abandon outer-bwrap as the foundational architecture and adopt a strategy that is *compatible* with each provider's own native isolation (claude's lack of inner sandbox vs codex's `--sandbox read-only` inner sandbox). This is what Solution 1 does — see A1.2 below.
 ### A1.1.4 — Forcing reason #4: `CODEX_HOME` exists and is the documented relocation lever
 The "ephemeral home directory per spawn" component of Solution 1 (A1.2 Layer 1) only works if each provider CLI offers a documented mechanism for relocating its state directory away from the default `$HOME` location. For `claude`, the standard `HOME` env var works (the CLI reads `~/.claude` as `$HOME/.claude`, and changing `HOME` relocates the lookup). For `codex`, the equivalent lever is the `CODEX_HOME` env var.
 **Primary citation:**
 - https://developers.openai.com/codex/config-reference — OpenAI's published codex CLI configuration reference. The page documents `CODEX_HOME` in 2 places (verified by independent fetch 2026-05-29): as the root of the per-profile config path (`$CODEX_HOME/profile-name.config.toml`) and as the default log directory base (`$CODEX_HOME/log`). The variable is the documented relocation lever for the codex state, configuration, and authentication directory away from the default `~/.codex`.
 - Secondary corroboration:
  - https://developers.openai.com/codex/auth/ — OpenAI's published codex CLI authentication reference. The page documents `CODEX_HOME` in 2 places (verified by independent fetch 2026-05-29), both in the credential-storage section: "file stores credentials in `auth.json` under `CODEX_HOME` (defaults to `~/.codex`)." Confirms `CODEX_HOME` is the credential-directory base.
  - https://codex.danielvaughan.com/2026/04/08/codex-cli-configuration-reference/ — third-party reference page that mirrors the documented behaviour, used as cross-reference for the reachability check.
 **What this means for Solution 1 feasibility:** All three Tier-D providers have a documented one-env-var relocation lever:
 - claude via `HOME` (POSIX convention)
 - codex via `CODEX_HOME` (citations above)
 - mistral via `VIBE_HOME` per https://docs.mistral.ai/mistral-vibe/terminal/configuration (3 occurrences verified 2026-05-29, including the canonical `export VIBE_HOME="/path/to/custom/vibe/home"` example and an enumeration of files/directories `VIBE_HOME` affects).
 The ephemeral-home approach is implementable today; it does not require upstream changes from any of Anthropic, OpenAI, or Mistral. Task #4 PI231 spike verifies *observed CLI behaviour* matches *documented behaviour* for each provider — this is verification-grade follow-up, not authority-pin work.
 ---
 ## A1.2 — The substitute architecture: per-spawn ephemeral home + per-provider ISOLATION contract
 The new architecture is layered. Each layer addresses a distinct attack surface, and each layer is independently reasoned about, independently reviewable, and independently revertible. The four layers, in order of containment depth:
 ### A1.2.1 — Layer 1: Per-spawn ephemeral home directory
 Every uncached `/v1/chat/completions` request (per-`keyId`, per-`reqId`) provisions a fresh ephemeral home directory at `/tmp/olp-spawn/<keyId>/<reqId>/home/`. The directory is created on the spawn path and torn down (best-effort) on response completion. The spawn process gets this directory passed in via a per-provider env-var override:
 - **anthropic** (`claude` CLI): `HOME=/tmp/olp-spawn/<keyId>/<reqId>/home`. The CLI's `~/.claude.json` and `~/.claude/` state writes go to the ephemeral location. No cross-request, no cross-tenant carry-over.
 - **openai** (`codex` CLI): `CODEX_HOME=/tmp/olp-spawn/<keyId>/<reqId>/home/.codex`. Codex's `~/.codex` state, auth artifacts, and config files go to the ephemeral location.
 - **mistral** (`vibe` CLI): `VIBE_HOME=/tmp/olp-spawn/<keyId>/<reqId>/home/.vibe` per https://docs.mistral.ai/mistral-vibe/terminal/configuration (documented env var, 3 occurrences verified at amendment time). Vibe's `~/.vibe/` state — `.env`, `agents/`, `prompts/`, `skills/`, `tools/`, `config.toml` — goes to the ephemeral location. Task #4 PI231 spike verifies observed CLI behaviour matches the documented contract.
 Layer 1 provides:
 - **No cross-tenant state carry-over** at the filesystem level. Two clients invoking anthropic concurrently get two separate `$HOME` directories; the CLI cannot read the other's `~/.claude.json`, recent-projects list, or session state.
 - **No accumulation of stale state** across requests. The MRU project list does not grow without bound. The telemetry session id is fresh per request.
 - **No outer-wrap maintenance treadmill.** When `claude` invents a new state file under `~/.claude.foo.json` next quarter, OLP does not need to update a `--ro-bind` list. The new file lives in the ephemeral home and goes away with the request.
 What Layer 1 does NOT provide:
 - It does not protect against the CLI walking *out of* its $HOME to read other paths (e.g., a model emitting a `Read` tool call on `/etc/passwd` or `~/.ssh/id_rsa`). For that protection, Layers 3 and 4 are needed.
 ### A1.2.2 — Layer 2: Symlinked credential files into the ephemeral home
 A fresh `$HOME` is empty. The CLI needs its OAuth credentials, API key, or equivalent auth artifact to function. Layer 2 provisions these by reading the operator-pinned credential location and symlinking the relevant file(s) into the ephemeral home at the location the CLI expects.
 Each provider plugin declares its credential paths in the ISOLATION block (see ADR 0002 Amendment pending). The runtime spawn pipeline reads this declaration, walks the list, and symlinks each entry from its real location (under the operator's real `$HOME`) into the ephemeral home. The symlinks are file-level, not directory-level, so the CLI sees its credential file but does not see the rest of the operator's `~/.claude/` or `~/.codex/` tree.
 Example (anthropic):
 - Real: `~/.claude/.credentials.json` (operator's actual OAuth credential)
 - Ephemeral: `/tmp/olp-spawn/<keyId>/<reqId>/home/.claude/.credentials.json` (symlink → real)
 Example (codex):
 - Real: `~/.codex/auth.json`
 - Ephemeral: `/tmp/olp-spawn/<keyId>/<reqId>/home/.codex/auth.json` (symlink → real)
 Layer 2 provides:
 - **Credential availability** without granting visibility into other state under the same provider directory.
 - **A narrow declared surface.** The provider plugin enumerates exactly which files matter. New CLI state files that are not declared do not get symlinked, and the CLI re-initializes them in the ephemeral home (which is exactly the Layer 1 behaviour).
 What Layer 2 does NOT provide:
 - It does not protect against the CLI walking out of its $HOME (see Layer 3).
 - It does not protect against the CLI's tool-use surface reading the symlink target's *containing directory* if the model emits a `Read` tool call with an absolute path that resolves around the symlink. For that, Layer 3 + Layer 4.
 ### A1.2.3 — Layer 3: Optional `sandbox-runtime` per-call `customConfig` for non-$HOME read protection
 For providers whose own inner sandbox does NOT exist or does not cover the OLP threat model (the `claude` CLI today is the leading example — claude has no inner sandbox; codex has `--sandbox read-only` by default but the protection scope differs), Layer 3 wraps the spawn in `@anthropic-ai/sandbox-runtime`'s `SandboxManager.wrapWithSandbox()` *per-call* with a `customConfig` argument tailored to the per-spawn ephemeral home.
 The key architectural difference vs PR-B's outer-wrap:
 - PR-B initialized `SandboxManager` once at server boot with a *global* config covering all providers.
 - Layer 3 calls `wrapWithSandbox()` *per spawn* with a *per-spawn* `customConfig` that names the ephemeral home as the allow-read root.
 The per-call `customConfig` shape:
 ```javascript
 {
  network: { allowedDomains: provider.ISOLATION.allowedDomains },
  filesystem: {
    denyRead: [
      // Operator's real $HOME — sandbox cannot read OTHER clients' OLP keys,
      // operator's SSH identity, other providers' tokens, etc.
      operatorHome,
      // Operator's known sensitive directories (defensive even though they
      // are already under operatorHome) — declared so a future refactor that
      // moves the operator home does not regress this protection.
      `${operatorHome}/.ssh`,
      `${operatorHome}/.gnupg`,
      `${operatorHome}/.olp`,
    ],
    // Layer 1 ephemeral home is the allow-read root for this spawn.
    // Layer 2 symlinked credentials live inside, so credential access works.
    allowRead: [ephemeralHomeForThisSpawn],
    allowWrite: [ephemeralHomeForThisSpawn, '/tmp'],
  },
 }
 ```
 Layer 3 is invoked **only when** the provider's `ISOLATION.hasInnerSandbox === false`. For providers with their own inner sandbox (codex via `--sandbox read-only`), Layer 3 is skipped to avoid the nested-sandbox conflict (A1.1.3).
 Layer 3 provides:
 - **OS-level deny of reads outside the ephemeral home and OLP-permitted paths.** A prompt-injected `cat /home/<operator>/.olp/keys/owner-key.json` or `cat /home/<operator>/.ssh/id_ed25519` hits a syscall-level deny.
 - **Per-spawn (not per-process) configuration.** Each request gets a fresh sandbox scope. Two concurrent spawns do not share a sandbox; the MITM-proxy collision and singleton-config-mutation hazards from PR-B disappear.
 What Layer 3 does NOT provide:
 - It does not protect against the CLI's *own* tool-use surface emitting destructive shell commands within the allowed write zones. For that, Layer 4.
 - Per-call `wrapWithSandbox()` has higher per-request latency than PR-B's once-at-boot pattern. The amortization budget is recovered by Layer 1's $HOME-as-cwd discipline keeping the sandbox config small and by ripgrep-based glob expansion being avoided (Layer 3 uses absolute literal paths throughout).
 ### A1.2.4 — Layer 4: Provider-specific tool hardening already in place
 This is already-shipped work, re-affirmed here as part of the layered model:
 - **anthropic Phase 6c `--system-prompt`** (commits `97e7d16` + fold-in `65f945c`). The system prompt is fully replaced at every spawn, suppressing the default tool descriptions that Claude Code would otherwise inject. Without tool descriptions, the model is highly unlikely to emit `tool_use` for `Bash`, `Read`, etc. even under prompt injection. See cc-mem `~/.cc-rules/memory/projects/olp/incident_2026_05_27_spawn_cli_security.md` § 5.
 - **codex `--sandbox read-only` default.** OLP's codex provider spawn passes `--sandbox read-only` as a fixed flag. Codex's own inner sandbox provides read-only-by-default tool isolation. The provider's ISOLATION block declares `hasInnerSandbox: true` so Layer 3 is correctly skipped.
 - **mistral.** TBD per Task #4 — the mistral provider's tool surface and inner-sandbox status need to be characterized.
 Layer 4 provides:
 - **Reduction of the *probability* of tool emission.** Layer 4 does not depend on OS-level enforcement; it works at the prompt layer. It is the cheap, fast, first-line defense. Layers 1–3 are the structural fallback when prompt-layer defenses are bypassed.
 ---
 ## A1.3 — The provider ISOLATION contract (named here; specified in ADR 0002 Amendment N)
 Each provider plugin declares an `ISOLATION` block on its module export. The fields are:
 | Field | Type | Meaning |
 |---|---|---|
 | `ephemeralEnvOverrides` | `(spawnCtx) => Record<string, string>` | Returns the env-var map to set for this spawn, given the spawn context (ephemeral home path, keyId, reqId). For anthropic: `{ HOME: spawnCtx.ephemeralHome }`. For codex: `{ CODEX_HOME: spawnCtx.ephemeralHome + '/.codex' }`. |
 | `credentialMounts` | `{ realPath: string, ephemeralPath: string }[]` | List of credential files to symlink from real → ephemeral. For anthropic: `[{ realPath: '~/.claude/.credentials.json', ephemeralPath: '.claude/.credentials.json' }]`. Provider declares; runtime symlinks. |
 | `hasInnerSandbox` | `boolean` | If true, Layer 3 is skipped to avoid nested-sandbox conflict. codex: true. anthropic: false. |
 | `crossTenantReadProtection` | `'tool-suppression' \| 'inner-sandbox' \| 'none'` | Self-declared label for what layer is providing the read-protection. Used by `/health.sandbox` to report the protection posture per provider. **The canonical enum is defined in ADR 0002 Amendment 9 § 5; this row mirrors it.** |
 | `recommendedDeploymentTier` | `'shared-os-user' \| 'per-os-user' \| 'separate-vm'` | Deployment tier the provider's current isolation posture is rated for. ADR 0006 risk-tier integration. **The canonical enum is defined in ADR 0002 Amendment 9 § 6; this row mirrors it.** |
 **This amendment names the contract but does NOT specify its full validation, lifecycle, or test discipline.** Those land in **ADR 0002 Amendment (pending)** — the Provider contract amendment that ratifies `ISOLATION` as a required field, defines `validateProvider`'s checks on it, and documents how `lib/providers/base.mjs` enforces declaration. Until that ADR amendment lands, the ISOLATION block is a forward-looking contract; the implementation refactor (Tasks #5–#8) is gated on the ADR 0002 amendment landing first.
 Cross-reference: see ADR 0002 § Amendments for the pending Amendment N that codifies the ISOLATION block contract.
 ---
 ## A1.4 — Revised PR plan
 The original ADR's four-PR split (PR-A / PR-B / PR-C / PR-D) is restated as follows. PR-A is unchanged from its as-shipped state.
 | PR | Original scope | Amendment 1 scope | Status |
 |---|---|---|---|
 | **PR-A** | npm dep + `lib/sandbox/doctor.mjs` + `/health.sandbox` | **Unchanged.** Doctor preserved; `/health.sandbox` field preserved. | ✅ Shipped (commit `07d9c8a`) |
 | **PR-B** | Outer-bwrap wrap of anthropic spawn at boot-singleton level | **Superseded by Amendment 1.** Implementation archived to branch `phase-7-pr-b-outer-bwrap-snapshot`. New scope: refactor `lib/sandbox/manager.mjs` to the Layer 1 + Layer 2 + Layer 3 architecture (Tasks #5, #8). | ⛔ Superseded |
 | **PR-C** | Outer-bwrap wrap of codex spawn with `enableWeakerNestedSandbox: true` | **Superseded by Amendment 1.** Codex isolation now flows via Layer 1 ephemeral `CODEX_HOME` + Layer 4 `--sandbox read-only`. Layer 3 deliberately skipped (`hasInnerSandbox: true`). Codex-specific PR (Task #7) lands the ISOLATION block declaration; no outer-wrap. | ⛔ Superseded |
 | **PR-D** | Documentation update for cloud rollout unblock | **Reframed.** New scope: README "Security Model" section documenting the four-layer architecture, the deployment-tier mapping, and what the operator gets vs does not get at each tier. Task #10. | ♻ Reframed |
 The new effective PR-list:
 - **PR-B' (Refactor):** `lib/sandbox/manager.mjs` rewritten to expose `prepareIsolatedEnvironment(spawnCtx)` (Layer 1 + Layer 2) and `maybeWrapForReadProtection(spawnCtx, command)` (Layer 3 conditional). The `OLP_SANDBOX_DISABLED=1` env-var gate is preserved for 1-2 releases as belt-and-suspenders, then removed. Singleton bootstrap pattern is removed (per-spawn config eliminates the singleton's reason to exist).
 - **PR-C' (Wiring + Anthropic ISOLATION):** `server.mjs` calls `prepareIsolatedEnvironment` on the spawn path; `lib/providers/anthropic.mjs` declares its ISOLATION block (Task #6); negative-test confirmation via Task #9 PI231 E2E.
 - **PR-D' (Codex ISOLATION):** `lib/providers/codex.mjs` declares its ISOLATION block (Task #7); `hasInnerSandbox: true` skips Layer 3; codex inner sandbox preserved unmolested. Verified on PI231 (Task #9).
 - **PR-E' (README + Phase 7 close):** README "Security Model" section (Task #10) + `docs/plans/cloud-deployment-family.md` § 5 update + Phase 7 close per `CLAUDE.md release_kit.phase_rolling_mode`.
 The original PR sequence's load-bearing security gate (the negative test "in-sandbox `cat ~/.olp/keys/...` MUST fail") remains the acceptance criterion for the security-bearing PRs in the new sequence. The test itself transfers; only the wrap mechanism changes.
 ---
 ## A1.5 — What survives from PR-B (preserved)
 The following artifacts from the original PR-B implementation are preserved through Amendment 1:
 1. **`lib/sandbox/doctor.mjs` — preserved unchanged.** Pure preflight is still useful: it tells the operator whether the npm package is installed, whether the OS deps are present, and whether the platform is supported. Even though the architecture no longer relies on a boot-time `SandboxManager.initialize()`, the `/health.sandbox` field consumers (dashboard, monitoring scripts) expect a stable shape. Doctor stays.
 2. **`/health.sandbox` field — preserved.** Shape adjusts slightly: the `active` boolean shifts meaning from "SandboxManager.initialize() succeeded" (PR-B) to "Layer 3 is operational for at least one provider whose `ISOLATION.hasInnerSandbox === false`" (Amendment 1). The field's name and JSON path stay the same so downstream consumers (dashboard, Hermes self-check, monitoring) do not break. The per-provider isolation posture is exposed via a new `/health.sandbox.providers[<name>].crossTenantReadProtection` subfield sourced from each ISOLATION block.
 3. **`@anthropic-ai/sandbox-runtime` npm dependency — preserved.** Layer 3 still uses the library, but via per-call `wrapWithSandbox()` with `customConfig`, not via a once-at-boot `SandboxManager.initialize()`. The dependency line in `package.json` stays.
 4. **The four authority citations in original § 7 — preserved.** The library URL, the spike report URL, the cc-mem incident URL, and the cloud deployment plan URL are unchanged. Amendment 1 *adds* the four new primary citations enumerated in § A1.1 above.
 5. **The `OLP_SANDBOX_DISABLED=1` env-var gate — preserved for 1-2 releases, then removed.** Documented in A1.6 below.
 ---
 ## A1.6 — What disappears from PR-B (superseded)
 The following artifacts are removed by the PR-B' refactor (Task #5):
 1. **Outer-bwrap wrapping of the `claude` spawn.** The bwrap wrap goes away. `claude` runs directly (without bwrap shell-wrap) with its `HOME` set to the ephemeral location. Layer 3 wraps the *sub-spawn* shell when it is invoked, not the `claude` process itself.
 2. **EROFS-driven mount patches.** The fold-in commit `2864275` ("allow read ~/.claude") and the subsequent `~/.claude` rw promotion (Task #5 was filed against this) were both consequences of trying to outer-bwrap a CLI that writes non-atomically to its `$HOME`. Solution 1 gives the CLI its own fresh `$HOME` and the entire mount-patch problem disappears. Task #5 ("allowWrite ~/.claude rw promotion fix") is closed as obsolete by this amendment.
 3. **Boot-time `SandboxManager.initialize()` call.** Removed entirely. The library is loaded lazily per-spawn (with import memoization for performance — the import itself is cached after the first call; only the `wrapWithSandbox()` call is per-spawn).
 4. **The singleton config-at-boot pattern.** Removed. The `_initConfig`, `_active`, `_initialized` module-level variables in `lib/sandbox/manager.mjs` no longer represent a global sandbox state; the only module-level state retained is the import cache for the library.
 5. **The MITM proxy CA cert generated once at boot.** Per-call `wrapWithSandbox()` may regenerate per call (TBD on library v0.0.52 behaviour — Task #4 verifies). If per-call regeneration is too expensive, an alternative is a per-process MITM CA cached at first-use; the implementation detail is reserved to PR-B'.
 6. **The `enableWeakerNestedSandbox: true` flag plan.** Removed. Codex isolation does not run inside an OLP outer sandbox at all. `enableWeakerNestedSandbox` is irrelevant to Amendment 1's architecture.
 ### A1.6.1 — The `OLP_SANDBOX_DISABLED=1` env-var gate
 The env-var gate added in commit `b1e24b7` ("add OLP_SANDBOX_DISABLED=1 env-var emergency disable") is preserved through the Amendment 1 refactor as belt-and-suspenders. Its semantics under Amendment 1:
 - **PR-B world (current main, with the gate set in prod):** the gate skips `SandboxManager.initialize()` at boot. Prod is currently running with the gate set, which means PR-B's outer-bwrap path is not active — Layer 3 protection is also not active.
 - **Amendment 1 world (after PR-B' lands):** the gate skips Layer 3's per-call `wrapWithSandbox()` and reverts each spawn to a Layer 1 + Layer 2 + Layer 4 configuration. The CLI still gets an ephemeral `$HOME` with symlinked credentials, still gets the `--system-prompt` tool-description suppression for anthropic, still gets `--sandbox read-only` for codex. What is given up is the OS-level deny of reads outside the ephemeral home. This is a *meaningful* but not *catastrophic* degradation — the prompt-layer defense remains, and Layer 1's $HOME isolation still prevents the most common cross-tenant accident path.
 - **Sunset:** the gate is preserved for **1-2 releases** after PR-B' ships to give the operator a fast escape hatch if the Layer 3 per-call wrap regresses in production. After two clean releases with no operator escalation, the gate is removed in a subsequent ADR amendment or a clean PR citing this section as authority for the removal.
 The gate's behaviour is documented in README's Security Model section per PR-D' (Task #10).
 ---
 ## A1.7 — Reversibility
 Amendment 1 is reversible at the implementation layer:
 - **PR-B' refactor** is reversible by `git revert` of the refactor commit + restoring the snapshot from `phase-7-pr-b-outer-bwrap-snapshot`. The archive branch is pushed and persistent at:
  https://github.com/dtzp555-max/olp/tree/phase-7-pr-b-outer-bwrap-snapshot
 - **The `@anthropic-ai/sandbox-runtime` dependency** stays in `package.json`, so reverting does not require an `npm install`.
 - **The `lib/sandbox/doctor.mjs` module** is unchanged across the refactor, so reverting does not affect `/health.sandbox` shape.
 Amendment 1 itself, as a governance artifact, is reversible by a subsequent superseding amendment if the empirical foundation it rests on changes (e.g., if Anthropic publishes guidance endorsing outer-wrap use of `sandbox-runtime` and adds a contract for `~/.claude.json` write paths). ALIGNMENT.md § "Amendment Procedure" applies: such a future amendment would need to cite the new evidence.
 The archive-branch retention policy: the snapshot branch is kept indefinitely (no auto-delete) so a future maintainer investigating outer-bwrap-around-CLI as an architecture has a working reference point. The branch's HEAD commit matches commit `b1e24b7` (the last commit of the outer-bwrap implementation before the architecture pivot).
 ---
 ## A1.8 — Updated open questions (supersedes original § 5)
 The original § 5 listed five open questions all of which were specific to the outer-bwrap architecture. Amendment 1 supersedes those and lists the open questions for the new architecture:
 1. **Per-call `wrapWithSandbox()` latency.** PR-B amortized the MITM CA generation (100-500ms) across all spawns by initializing once at boot. Per-call wrap regenerates this if the library does not cache internally. Task #4 PI231 spike measures the actual per-call cost; if it exceeds the original ≤200ms p95 budget, an internal cache wrapper around the library is added in PR-B'. Decision reserved for PR-B'.
 2. **`vibe` (mistral) home-relocation env var.** Task #4 PI231 spike checks whether `vibe` honours `MISTRAL_HOME` / `VIBE_HOME` / similar. If yes, mistral's ISOLATION block declares it and mistral participates in Layer 1. If no, mistral falls back to Layer 4 (prompt layer) + Layer 3 (per-call wrap with `denyRead` on the operator's real home) only. The provider's `recommendedDeploymentTier` is set accordingly.
 3. **macOS coverage.** sandbox-runtime supports macOS via `sandbox-exec` (seatbelt profile). Layer 1 ephemeral home is OS-agnostic (just an env var). Layer 3 macOS path needs verification: does per-call `wrapWithSandbox()` with `customConfig` produce a per-spawn sandbox-exec profile, or does it re-use a singleton seatbelt profile? Task #4 PI231 spike is Linux-only; a parallel macOS verification is a Task #9 deliverable.
 4. **Symlink-vs-bindmount for credentials.** Layer 2 uses symlinks for credential mounting. An alternative is bindmounting the credential file into the ephemeral home (only available inside the Layer 3 wrap). The trade-off: symlinks work outside any sandbox context (so Layer 2 works even when Layer 3 is skipped, e.g., for codex); bindmounts are stronger isolation (the CLI cannot follow the symlink to discover the real path). Decision reserved for PR-B' implementation review.
 5. **Concurrent-spawn cleanup ordering.** The ephemeral home cleanup (rmdir at response end) must not race with a still-streaming spawn. The current plan: track per-`reqId` cleanup and only fire on the spawn's `exit` event. If a streaming abort leaves the spawn alive past the HTTP response, cleanup is deferred until `exit`. Tested in Task #9.
 6. **`/health.sandbox.providers` shape under Amendment 1.** Original `/health.sandbox` had a flat `{ available, active }`. Amendment 1 adds per-provider posture: `{ available, providers: { anthropic: { crossTenantReadProtection: 'tool-suppression', layers: ['L1','L2','L3','L4'] }, openai: { crossTenantReadProtection: 'inner-sandbox', layers: ['L1','L4'] } } }`. Exact shape ratified by PR-B'.
 7. **Dashboard `/dashboard` Security panel.** The dashboard currently has no security panel. Amendment 1 names the addition as a follow-up: render `/health.sandbox.providers` as a per-provider posture badge so the operator can see at a glance which providers are in `tool-suppression` vs `inner-sandbox` vs `none` mode. Out of Phase 7 scope; recorded for a future ADR.
 ---
 ## A1.9 — Authority citations (Amendment 1)
 Per ALIGNMENT.md Rule 1 (Cite First) and Iron Rule 12 (Pre-Brainstorm Prior-Art Search), every load-bearing claim in this amendment is cited to a primary source. The four forcing reasons are cited above in A1.1.1–A1.1.4; this section enumerates them in one place plus the supporting citations.
 **Forcing reasons:**
 1. **sandbox-runtime documented use-case is inner-wrap by Claude Code.**
   - https://www.anthropic.com/engineering/claude-code-sandboxing — "Claude Code sandboxing" engineering blog. Documents Claude Code's *internal* use of the library to wrap tool-spawn calls. The blog also notes the library "can be used to sandbox arbitrary processes, agents and MCP servers" — i.e., it is general-purpose, not Claude-Code-internal-only. **Our reading:** outer-wrap of `claude` itself is not the documented or demonstrated direction; OLP would be the only known user in that configuration. Project-design judgment, not an Anthropic prohibition.
 2. **`~/.claude.json` non-atomic write — upstream issue closed `not_planned` by inactivity bot.**
   - https://github.com/anthropics/claude-code/issues/29250 — upstream issue requesting atomic-write semantics. Status: closed `not_planned` by `github-actions[bot]` on 2026-03-31 (inactivity), labeled `duplicate`, `stale`. **No upstream Anthropic maintainer comment articulates a policy position**; the closure does not establish "won't fix" as policy. Forcing argument rests on architectural-cost analysis (permanent maintenance treadmill for outer-`--ro-bind`), not on alleged upstream policy.
 3. **Codex inner-bwrap conflict.**
   - https://github.com/openai/codex/issues/16018 — upstream codex CLI issue. Documents that codex's default bwrap sandbox **fails outright** in environments lacking unprivileged user namespaces. The issue is a feature request asking codex to add a fallback path; **the issue body does NOT document an existing automatic fallback to `--sandbox danger-full-access`**. Whether codex degrades to `danger-full-access` or aborts the spawn under nested-bwrap failure is empirically open (Task #4 deliverable). Either failure mode breaks the strict-additive-isolation invariant for outer-wrap of codex. This is the multi-provider forcing function regardless of which failure mode applies.
 4. **`CODEX_HOME` documented relocation lever.**
   - https://developers.openai.com/codex/config-reference — OpenAI codex CLI config reference (primary).
   - https://codex.danielvaughan.com/2026/04/08/codex-cli-configuration-reference/ — third-party reference (cross-reference for reachability).
 **Supporting citations (carried forward from original ADR § 7):**
 5. **`@anthropic-ai/sandbox-runtime` v0.0.52** — https://github.com/anthropic-experimental/sandbox-runtime
   - `dist/sandbox/sandbox-manager.js` — `SandboxManager.wrapWithSandbox(command, undefined, customConfig)` is the per-call wrap surface used by Layer 3. The third argument `customConfig` is the per-call override mechanism that makes Amendment 1's per-spawn config architecture implementable without library modification.
 6. **Internal evidence:**
   - **PR-B implementation chain** — commits `d0dcd28` → `2864275` → `497b255` → `b1e24b7` → `3551921`. The HTTP-path activation regression is documented in commit message `b1e24b7` and in `lib/sandbox/manager.mjs` § "OLP_SANDBOX_DISABLED env-var gate" comments.
   - **cc-mem incident memory 2026-05-27** — `~/.cc-rules/memory/projects/olp/incident_2026_05_27_spawn_cli_security.md` — the original multi-tenant gap and the prior-art search that established the ecosystem has no working solution.
   - **2026-05-28 PoC spike on PI231** — `/tmp/sandbox-spike/report.md` on PI231. Verdict was YELLOW (architecturally green, operationally blocked on apt deps). The follow-up 2026-05-29 PI231 prep work re-evaluates against the new architecture; results land in Task #4.
 7. **OLP governance:**
   - **OLP ALIGNMENT.md Rule 1** — Authority citation required for any provider-plugin / entry-surface / IR change. Amendment 1 amends governance only; the implementation refactor (PR-B') carries its own per-commit citations to the same primary sources enumerated above.
   - **OLP ALIGNMENT.md Rule 4** — Unalignable plugins are deleted. Mistral's potential lack of a home-relocation env var (open question 2 above) is *not* an alignability gap (mistral's CLI authority is unchanged); it is a deployment-tier classification, recorded in the provider's ISOLATION block.
   - **Iron Rule 10** — Independent reviewer required. This amendment's review is pending at draft time.
   - **Iron Rule 11** — Minimum reviewable unit. PR-B' is one PR (sandbox manager refactor); the anthropic ISOLATION block, codex ISOLATION block, server wiring, and README section are each separate PRs per the revised PR plan in § A1.4.
   - **Iron Rule 12** — Pre-brainstorm prior-art search. The four forcing reasons each satisfy the rule's "provider-specific authority check decisive" condition: Anthropic's blog post + upstream issue 29250 (for the anthropic side), and the codex issue 16018 + the OpenAI config reference (for the codex side).
 8. **OLP ADR cross-references:**
   - **ADR 0001 § Mission** — multi-provider proxy. Codex inner-bwrap conflict is the multi-provider forcing function.
   - **ADR 0002 (pending Amendment N)** — Provider ISOLATION contract specification. Amendment 1 names the contract; Amendment N specifies it.
   - **ADR 0006** — Provider Inclusion / Risk Tier. `recommendedDeploymentTier` in the ISOLATION block integrates with the risk tier framework.
   - **ADR 0009 Amendment 1 § Caveats #3** — "Sandbox-runtime still required for real multi-tenant deployment." Amendment 1 satisfies this caveat via Layer 3, not via outer-wrap.
   - **`docs/plans/cloud-deployment-family.md` § 5** — sandbox is a cloud rollout prerequisite. PR-E' updates this section to reflect that the layered architecture is the cloud prerequisite, not outer-bwrap.
 ---
 ## A1.10 — Consequences of Amendment 1
 ### Positive
 - **No outer-bwrap maintenance treadmill.** New `claude` CLI state files do not require OLP-side `--ro-bind` updates. Layer 1 absorbs them automatically.
 - **Multi-provider compatible.** Codex inner sandbox is preserved unmolested. The architecture works for both anthropic (no inner sandbox) and codex (has inner sandbox) without per-provider workarounds in the sandbox layer; the per-provider differences live in the per-provider ISOLATION block where they belong.
 - **Per-spawn isolation primitives.** Every request gets a fresh `$HOME`. Cross-tenant state carry-over at the filesystem level is structurally impossible, not "mitigated by careful denylist."
 - **Aligned with Anthropic's design intent.** OLP uses sandbox-runtime in the direction the library was designed for (sandboxing what the spawn triggers, not wrapping the spawn from outside). The library's semver discipline becomes leverage rather than risk.
 - **Reduced HTTP-path activation surface.** PR-B's regression was that the in-process MITM proxy lifecycle interacted with OLP's HTTP request handler. Per-call `wrapWithSandbox()` does not require an always-on in-process proxy; the failure mode goes away by construction. (To be confirmed empirically in Task #4 + Task #9.)
 - **Doctor and `/health.sandbox` continuity.** Operators and dashboard consumers see the same field at the same JSON path. Shape additions are additive, not breaking.
 ### Negative
 - **Per-call latency cost.** Per-call `wrapWithSandbox()` is more expensive than once-at-boot init+wrap. The mitigation is library-import caching and (if measured high) a sandbox-config cache keyed by the union of allowed-read paths. Empirical measurement in Task #4.
 - **New contract surface (ISOLATION block).** Each provider plugin now declares ISOLATION fields. This is incremental complexity in the Provider contract — ratified by ADR 0002 Amendment N. ADR 0002 amendment is on the critical path.
 - **Mistral declared `crossTenantReadProtection: 'none'`.** Vibe CLI has no Phase-6c-equivalent tool suppression and no known inner sandbox as of D8 ADR 0006 enablement. The mistral provider's `recommendedDeploymentTier` is therefore `separate-vm` per ADR 0002 Amendment 9 § Per-provider concrete instance, meaning mistral can run only in a dedicated VM rather than sharing the OS user with other providers. Not a regression vs status quo (mistral is not deployed today); reflects honest characterization of current state per ALIGNMENT.md Rule 3. Task #4 spike may discover a hardening regime, transitioning this tier upward.
 - **Symlink semantics edge cases.** Layer 2 symlinks credential files into the ephemeral home; some CLIs may resolve the symlink and write a sibling file in the *target* directory rather than the ephemeral location. Each provider's ISOLATION block should declare any such known behaviour; the runtime tests verify by examining the operator's real `$HOME` for stray writes after a test spawn.
 - **The `OLP_SANDBOX_DISABLED=1` env-var gate is preserved for 1-2 releases.** It remains a valid escape hatch — but as belt-and-suspenders rather than as load-bearing. Operators who rely on the gate after sunset will see a deprecation message before removal.
 ### Reversibility (governance level)
 - Amendment 1 is reversible by a superseding ADR amendment that cites new evidence overturning any of the four forcing reasons. The most likely overturning scenario: Anthropic publishes guidance endorsing outer-wrap of `claude` CLI plus an atomic-write contract for `~/.claude.json`. If that happens, the superseding amendment cites the new guidance and re-enables outer-wrap as an option (alongside, not replacing, the Solution 1 architecture).
 - The implementation-level reversibility is documented in § A1.7 above.
 ---
 ## A1.11 — Forward-looking pointer
 Amendment 1 is the governance layer. The implementation lands across Tasks #5–#10 (per the working task list at the time of this draft):
 - Task #4 — PI231 spike to verify `HOME` / `CODEX_HOME` env-var override behaviour (live, with the same `claude` and `codex` CLI versions OLP ships against).
 - Task #5 — Refactor `lib/sandbox/manager.mjs` to the Layer 1 + Layer 2 + Layer 3 architecture (PR-B').
 - Task #6 — Add ISOLATION block to `lib/providers/anthropic.mjs` (PR-C').
 - Task #7 — Add ISOLATION block to `lib/providers/codex.mjs` (PR-D').
 - Task #8 — Wire `prepareIsolatedEnvironment` into `server.mjs` spawn pipeline (folds into PR-C' or its own PR depending on diff size).
 - Task #9 — PI231 E2E validation of Solution 1 + close PR-B's load-bearing negative test ("in-sandbox `cat ~/.olp/keys/...` MUST fail") against the new architecture.
 - Task #10 — README "Security Model" section + cloud-deployment-plan § 5 update + Phase 7 close (PR-E').
 ADR 0002 Amendment N (Provider ISOLATION contract specification) is a co-merged ADR with PR-C'; it cannot land after the ISOLATION block reaches the codebase per ALIGNMENT.md Rule 2(c)'s spirit (no contract field without an authorizing ADR).
 ---
 ## A1.12 — Amendment status
 - **Drafted:** 2026-05-29 (this document).
 - **Reviewer:** independent fresh-context reviewer per Iron Rule 10 — pending.
 - **Implementation gate:** ADR 0002 Amendment N (Provider ISOLATION contract specification) must land before or together with PR-C' (the first ISOLATION-block-bearing provider plugin commit).
 - **Production gate:** PI231 E2E (Task #9) must pass the load-bearing negative test before the `OLP_SANDBOX_DISABLED=1` env-var gate is removed from prod startup.