Merge remote-tracking branch 'origin/master' into codex/time-context-plugin

2026-07-14 17:26:20 +08:00
parent a9d74932b1 fb126e164a
commit 498de7f042
562 changed files with 4439 additions and 12572 deletions
@@ -24,7 +24,7 @@ Tool-call presentation remains tool-owned. A tool's `presentCall` and `presentRe

 Permission handling is an answerer on the [user-approval seam](2026-07-06-approval-seam.md), not an ask-every-tool policy in ACP. An `approval/request` for a bridge-owned agent with a call id becomes `session/request_permission` on that agent's editor session, with one-shot allow/reject choices. Foreign or call-less requests delegate; a missing or failed answerer remains fail-closed. The plugin that asks—such as a pre-execute policy or bash escalation—owns the decision to ask.

-The bridge advertises ACP config options instead of session modes. When `ctx.permission` is composed, it exposes one `permission` select whose options come from the deployment's preset table; the shipped presets are `workspace-write` and `danger-full-access`, and each bundles a sandbox mode with an approval policy. The current value comes from `PermissionService.current()`, including the derived switch-away-only `custom` state when the effective knobs match no preset. `session/set_config_option` validates through `PermissionService.set()` and writes the chosen preset through to both owning knob events. An open-turn switch appends immediately; an idle switch is overlaid in the response and anchored at the next turn start. Until that anchor it is memory-only and a crash reverts to the durable fold. ACP session modes are deliberately not modeled because config options are the forward protocol surface. Runtime model selection remains outside this decision; `AcpConfig.model` is connection-wide.
+When `ctx.permission` is composed, the bridge exposes one `permission` select from the deployment's preset table. The shipped `workspace-write` and `danger-full-access` presets each bundle a sandbox mode with an approval policy; unmatched effective knobs produce the switch-away-only `custom` state. `session/set_config_option` validates through `PermissionService.set()` and writes both owning knob events. A switch during an open turn appends immediately; an idle switch is overlaid in responses and anchored at the next `agent/prompt-submit`, before request assembly. Until then it is memory-only, so a crash restores the durable fold. ACP session modes are not modeled because config options are the forward protocol surface; `AcpConfig.model` remains connection-wide.

 The bridge also provides the ACP-backed `UserInteractionProvider`: `ask_user_question` requests become form elicitations on the owning session. Select, multi-select, option descriptions, and custom-answer override semantics are preserved.

@@ -50,7 +50,7 @@ Editors can create, load, prompt, cancel, render, ask, and reconfigure multiple

 The bridge deliberately does not implement session list/delete/resume/close capabilities, MCP passthrough, additional directories, image/audio/embedded-resource prompts, runtime model selection, plans, slash commands, usage updates, editor filesystem delegation, or the ACP terminal execution sub-protocol. The feature checklist records these as unsupported rather than silently accepting them.

-An idle config selection is truthful in the live response but not durable until the next turn anchors it. Crashing before that boundary loses the pending selection; this is the cost of keeping session events turn-enclosed and replay-safe.
+An idle config selection is truthful in the live response but not durable until the next `agent/prompt-submit` anchors it inside the open turn. Crashing before that boundary loses the pending selection; this is the cost of keeping session events turn-enclosed and replay-safe.

 ## Verification

@@ -24,33 +24,33 @@ Three decisions, each elaborated in its own section below:

 `ToolRegistry` gains a schemastery-validated config (`static Config`), its first: `mode: 'native' | 'code' | 'both'`, default `'native'`. A deployment flips it from `cordis.yml` (`tools: { mode: code }`) — no code edit, per the no-hardcoded-tunables convention.

-**Wire tool list = the registry's contribution before cooperative assembly.** The registry feeds assembly through a mode-aware provider: `'native'` contributes every capability visible to that assembly scope, `'code'` contributes only `run_code`, and `'both'` contributes both. Because [`PromptAssembly.tools` is the single source the loop's request header snapshots](../../../../packages/core/system-prompt/src/index.ts), the final presentation is logged and reconstructable. The reserved transport is not a capability: it lives outside global/scoped registration and restriction layers, cannot be registered or shadowed there, and cannot be named by `ctx.tools.restrict()`. The mode governs this provider's input to assembly; other direct `systemPrompt.tools()` providers own their schemas, and the trusted assembly waterfall owns the returned wire list.
+**Wire tool list.** The registry contributes visible capabilities in `'native'`, only `run_code` in `'code'`, and both in `'both'`. The final `PromptAssembly.tools` list is logged in the request header. `run_code` is a reserved presentation transport outside registration and restriction layers; direct prompt providers and the assembly waterfall remain responsible for their own contributions.

 **Interaction with `toolOrder`, stated up front:** a configured `systemPrompt.toolOrder` naming native capabilities rejects every assembly under `mode: 'code'`, because those names are outside that mode's wire-validation universe. This is correct behavior, not a bug: a deployment using Code Mode updates its order config or drops it.

-**The SDK prompt section.** Under `'code'` and `'both'` the registry registers one lazy prompt section (`tools:sdk`, in the 100–199 tool-guidance order band) whose thunk regenerates, for each assembly scope, a TypeScript declaration of every visible end-capability tool plus fixed usage instructions. It uses the same visibility resolver as lookup and execution, so scoped grants and shadows appear while restricted globals disappear; the reserved `run_code` transport itself is excluded. The thunk emits tools in lexicographic name order, so an unchanged visible set produces byte-identical text.
+**SDK prompt section.** In `'code'` and `'both'`, the lazy `tools:sdk` section in the tool-guidance order band renders TypeScript declarations plus fixed usage instructions for the scope's visible capabilities. It shares lookup and execution visibility, excludes `run_code`, and sorts tools lexicographically for byte-stable output.

 **Assembly ownership.** `run_code` and `tools:sdk` enter the trusted `system-prompt/assemble` waterfall as normal assembly inputs. A scoped `tools:sdk` section may shadow the global default before dispatch, and a listener may remove or replace either contribution. The waterfall's returned assembly is final, so whoever changes these inputs owns preserving a viable Code Mode protocol when the deployment expects Code Mode to remain usable; no restoration pass overrides deliberate composition.

-**Codegen.** A pure `jsonSchemaToTs(schema)` module inside `dsh-tools` (sibling of `json-schema.ts` — `schemas()` and the SDK are two projections of the same store) maps the JSON-Schema subset the `defineTool` DSL emits (object/string/number/boolean/array, `properties`, `required`, string `enum` → literal union, nested objects, array `items`, `description` → JSDoc) to a TS type literal. It is **total**: any construct outside that subset (`$ref`, `oneOf`/`anyOf`, `integer`, future MCP shapes, …) degrades to `unknown` without throwing. Because `ToolSchema.name` is an arbitrary string, the SDK is declared as one object constant — `declare const tools: { "some-mcp-tool"(args: …): Promise<string>; bash(args: …): Promise<string>; … }` — quoted keys make every name reachable with no sanitization or alias-collision logic. Typing is advisory (the runtime executes type-stripped JS); the instructions say so.
+**Codegen.** `jsonSchemaToTs()` maps the `defineTool` JSON-Schema subset to TypeScript, carries schema descriptions into JSDoc, and degrades unsupported constructs to `unknown`. The SDK exposes tools as quoted object keys, supporting arbitrary names without aliases or collisions. Typing is advisory because the runtime strips types before execution.

 ### The run_code tool and the dispatch bridge

 Under `'code'` and `'both'` the registry owns `run_code` as a reserved presentation transport with one required parameter, `{ code: string }`. It is represented by a normal `ToolDefinition` for dispatch but stays outside the filterable capability layers, so restrictions cannot accidentally remove Code Mode's only entry point. Calls traverse the complete tool pipeline — `tools/pre-execute` → monotonic guards → `tools/execute` around dispatch → `tools/post-execute` → immutable `tools/result` notification — exactly like native calls; a permission plugin can inspect the program text before it runs, and final-result observers see the normalized outer outcome. Its `execute(args, exec)`:

-1. **Builds the bindings**: the bridge owns a **run-scoped `AbortController`** whose signal follows `exec.signal` (an outer cancel propagates in) and which the bridge itself fires the moment the run settles for any reason — completion, program exception, `computeMs`/`maxWallMs` expiry, worker exit. For every visible capability tool, the binding is an async function that (a) checks the run signal before and after, (b) **JSON-normalizes the argument** — a `JSON.parse(JSON.stringify(args))` round-trip, rejecting that one call with a descriptive `Error` when the value does not survive (`BigInt`, circular structures) — because the seam's structured-clone boundary is wider than JSON while the session log accepts only JSON, (c) awaits its turn on the **per-run serialization queue** (below), (d) calls `this.execute({ callId, name, arguments, agent: exec.agent, parent: exec.token, signal: runSignal })` with a deterministic sub-id `` CallId(`${exec.callId}:code:${n}`) ``, (e) appends a `tool/code-dispatch` session event, and (f) maps the result: success → the text-block contents joined as a `string` (non-text blocks become placeholders), `isError` → **the binding rejects** with an `Error` carrying the result text. The child's readonly `parent` is only the outer execution's frozen, property-free token, so commit-style observers can correlate outcomes without receiving a mutation path into the live `run_code` wrapper. Every sub-call still traverses the full pipeline under its own immutable identity and registry-assigned token. The run signal, rather than the bare outer one, lets budget expiry abort an in-flight sub-tool instead of orphaning it. Rejection gives programs ordinary `try/catch` and `Promise.all` failure semantics rather than a bespoke result envelope.
+1. **Build bindings.** One run-scoped signal follows outer cancellation and is aborted whenever the run settles. Each visible tool binding JSON-normalizes its arguments—rejecting lossy values before dispatch—waits on the serialization queue, executes with a deterministic call id and the outer token as `parent`, and logs `tool/code-dispatch`. Successful text becomes a string and non-text blocks become placeholders; tool errors reject the binding promise. Every sub-call retains its own immutable execution identity and traverses the full tool pipeline.
 2. **Runs the program**: `ctx.codeRuntime.run({ program: args.code, bindings: [{ global: 'tools', functions }], signal: runController.signal })`. The runtime receives the run-scoped signal, not only the caller's outer signal, so any way the outer run settles also aborts work inside the runtime.
-3. **Surfaces the outcome — after reaching quiescence.** When `ctx.codeRuntime.run()` settles, whether by fulfillment or rejection, the bridge fires the run-scoped abort (cancelling any in-flight sub-dispatch and abandoning queued-unstarted ones), then **awaits the dispatch queue's drain before returning or propagating**, per the dispose-to-quiescence rule in [defensive patterns](../../../defensive-patterns.md): an aborted in-flight sub-call still settles and logs its `isError` `tool/code-dispatch` event *inside* the open turn, and nothing can append after `run_code` settles. A successful result then returns one text block — the captured console/stdout output followed by the rendered return value (if any) — plus a `meta` payload (capped logs, dispatch count) for presentation. A fulfilled run with `result.error` throws a `CodeRunFailedError extends HarnessError` (`code: 'CODE_RUN_FAILED'`, message = the error kind and text plus captured logs so the model can self-correct); a backend rejection propagates through the same registry error boundary. Both become structured `isError` tool results.
+3. **Settle after quiescence.** When the runtime settles, the bridge aborts outstanding work and drains the dispatch queue before returning. Success returns captured output and presentation metadata. A runtime failure becomes `CodeRunFailedError`; backend rejection uses the registry's normal error boundary. Both produce structured error results, and no sub-call can append after `run_code` settles.

-**Sub-call `additionalContext` is suppressed, deliberately.** A `tools/post-execute` hook may attach `additionalContext` to a call; for loop-dispatched calls the loop buffers those and appends each as a `context/message` only after the step's `tool/result`s, preserving call/result adjacency. A sub-dispatch result's `additionalContext` has no such safe outlet from inside a running `run_code`: injecting immediately would land a `context/message` between the parent's `tool/call` and its `tool/result` (breaking the adjacency the buffering exists to protect), and `PostToolDecision.additionalContext` is singular where a program may produce many. The MVP therefore drops sub-call `additionalContext`, pinned by a test and stated in the hooks bridge's docs; the follow-up (a plural context channel or loop-level sub-dispatch buffering) is deferred until a real hook needs it through Code Mode.
+**Sub-call `additionalContext` is omitted.** Injecting it during `run_code` would break parent call/result adjacency, while one program can produce many contexts. Supporting it requires a plural channel or loop-level sub-dispatch buffer.

-**Concurrency: serialized, enforced by the binding.** The bindings are async, so a model writing `Promise.all([tools.a(…), tools.b(…)])` starts both immediately — concurrent dispatch would be the default, while the tool contract carries no concurrency-safety metadata (the open parallel-execution TODO). Each `run_code` invocation therefore owns a dispatch queue and every binding call chains onto it, so even `Promise.all` executes the underlying `ctx.tools.execute()` calls one at a time in submission order; when the run settles, queued-but-unstarted dispatches are abandoned. Lifting this per tool remains tied to tools declaring themselves concurrency-safe.
+**Concurrency is serialized.** Each run owns a dispatch queue, so even `Promise.all` executes tool calls in submission order. Settlement abandons queued calls that have not started. Parallelism requires per-tool concurrency-safety metadata.

 **Presentation.** `run_code`'s render intent is decided here per the [render-intent RFC](../../implemented/architecture/2026-07-02-tool-render-intent-union.md): `presentCall` → a `generic` card, `kind: 'execute'`, title = the program text, `rawInput` = the same program text; `presentResult` → a `generic` card whose content is the captured output (from `meta`). The program is the title because ACP execute cards reliably render that field while some clients omit body and raw-input content. This is not a `terminal` card: that card's semantics are "a shell command in a working directory", which a program is not.

 ### Observability: `tool/code-dispatch`

-Each sub-dispatch appends one session event, declared by `dsh-tools` via `SessionEventMap` declaration merging (the map is merge-extensible for exactly this; `todo/write` is the log-only precedent): `tool/code-dispatch` with `{ parentCallId, subCallId, name, arguments, isError, resultSummary }` — `arguments` being the bridge's JSON-normalized value, the very one dispatched, so the append cannot fail on payload shape. It is log-only — `deriveEventMessage()` ignores unknown event types by design, so sub-calls never re-enter model context — but persistence and UIs get every call. As a log event it carries JSDoc prose but **no `@mode` tag** (that vocabulary belongs to cordis bus events; the persistence-catalog generator hard-errors on one) and lands in the regenerated `docs/persistence-catalog.md`; appends happen inside `run_code`'s execution, so the turn-enclosure invariant is satisfied by construction. A `run_code` execution arriving without `exec.agent` (the loop always supplies it; direct programmatic calls may not) still runs and simply skips event logging, exactly as the `ToolExecution` contract allows.
+Each sub-dispatch appends a log-only `tool/code-dispatch` event containing parent and child call ids, tool identity, normalized arguments, and result summary. It remains outside model history but available to persistence and UIs. Appends occur inside the open `run_code` turn. Direct executions without an agent still run but cannot log the event.

 ### The code-runtime seam

@@ -63,7 +63,7 @@ Each sub-dispatch appends one session event, declared by `dsh-tools` via `Sessio
 - `CodeRunFailure = { kind: 'exception' | 'timeout' | 'abort' | 'worker-exit'; message: string }` — orthogonal outcomes reported independently per [defensive patterns](../../../defensive-patterns.md); a timed-out run is not an exception, an abort is not a timeout.
 - Two readonly backend descriptors, informational not gating: `language` (what the program must be written in — `'typescript'` for the shipped backend; a Python backend would say so, and pair with its own SDK generator on the presentation side) and `isolation` (`'worker-thread'` for the shipped backend; `'process'`, `'container'`, … for future ones). `dsh-tools` requires `language === 'typescript'` in the MVP — its codegen emits TS — and fails the assembly loudly otherwise, the same misconfiguration idiom as `toolOrder` violations (as when `mode` is non-native with no `ctx.codeRuntime` loaded at all).

-Per explicit-over-implicit at seams, the request spells out everything the runtime acts on; defaulting (timeouts, caps) is the implementation's validated config, never a hidden `??` inside `run()`. Consumption uses the loop's optional-backend idiom: Cordis has no optional injection — every `inject` entry gates activation — so a static `inject` on the registry would hold `ctx.tools` (and every tool plugin behind it) hostage to a code runtime existing even under `mode: 'native'`; instead the registry reads `ctx.get('codeRuntime')` at use time, exactly as `agent-loop` consumes `sessionPersistence`, with absence failing loud in the provider thunk. The seam has concrete divergence on both axes: the worker-thread substrate can be replaced by a container or microVM implementation, and the TypeScript language contract can be paired with a language-specific SDK and runtime. `dsh-tools` consumes only the interface and tests against a trivial in-repo fake, exactly the interface/implementation/consumer shape of the bash template.
+Requests contain every runtime input; implementations own validated timeout and cap defaults. The registry looks up the optional runtime only when Code Mode is assembled, so native mode does not depend on one. Missing or language-incompatible runtimes fail loudly. Alternate substrates or languages can replace the implementation behind the same seam, paired with the appropriate SDK generator.

 ### The worker-thread runtime

@@ -73,36 +73,27 @@ Per explicit-over-implicit at seams, the request spells out everything the runti
 2. **Spawn one fresh `Worker` per run** from the package's own bootstrap module: `env: {}` (truly empty — stronger than the scrubbed-env rule for spawned commands), `resourceLimits` from config, `stdout`/`stderr` captured into `logs` rather than inherited. No pooling and no cross-run state: a program's world dies with its worker, which keeps runs reconstructable from the log alone and makes state bleed unrepresentable.
 3. **Execute** in the bootstrap: the stripped program becomes the body of an `AsyncFunction` whose parameters are the binding globals and a capturing `console` shim, so top-level `await` and `return` work and the program's completion value is the run's `value` (structured-cloneable values cross as-is; anything else is replaced by its `util.inspect` rendering, documented).
 4. **Bridge bindings over the message port**: each binding function in the worker posts `{ id, global, name, args }` and awaits the reply; the host validates the name against the request's bindings, invokes, and replies `{ id, ok, value }` or `{ id, ok: false, message }` (a host-side binding rejection becomes a program-side rejection). The worker-side namespace objects are built null-prototype via `defineProperty`, so a binding named `__proto__`, `constructor`, or `toString` is an ordinary own property, not a prototype collision. Unknown names, duplicate ids, and post-settlement messages are rejected or ignored — the port protocol assumes a hostile peer, because the peer runs model code.
-5. **Enforce caps — two independent budgets, because the peer is hostile.** The compute budget (`computeMs`) meters the worker's *measured busy time* via `worker.performance.eventLoopUtilization()` polling — not host-side "is an RPC pending" bookkeeping, which a program defeats by firing an un-awaited call at a slow tool and then spinning hot while the host thinks it is waiting. Measured busy time cannot be gamed: a hot loop accrues it whether or not a dispatch is in flight, and a program genuinely awaiting a slow tool accrues none, so a long-running `bash` sub-call still does not kill an innocent run. The wall ceiling (`maxWallMs`) never pauses for anything and backstops what busy-time cannot see (a program awaiting a promise nobody will resolve). Budget expiry, `signal` abort, and run completion all funnel into `worker.terminate()`, which ends hot synchronous loops too (measured; this was `node:vm`'s unfixable gap); the failure reports which budget fired. Heap overflow surfaces as the worker's OOM exit → `error.kind: 'worker-exit'`. Log and value sizes are capped by config, truncation marked in-band. All caps are validated config fields with defaults (`computeMs: 60_000`, `maxWallMs: 600_000`, `maxLogBytes: 65_536`, `maxValueBytes: 32_768`, `maxOldGenerationSizeMb: 512`), changeable from `cordis.yml`.
+5. **Enforce independent budgets.** `computeMs` meters worker busy time, allowing slow awaited tools without excusing a hot loop. `maxWallMs` bounds total elapsed time, including unresolved waits. Expiry, cancellation, and completion terminate the worker. Heap exits and truncation are reported explicitly; compute, wall, heap, log, and return-value caps are validated configuration.
 6. **Dispose to quiescence**: the service's own disposal terminates in-flight workers and *awaits* their exits before resolving, per [defensive patterns](../../../defensive-patterns.md).

 ### Trust posture

-The worker runtime is **containment, not a security boundary**. Model code in the worker can reach Node globals — `fetch`, `process` (with an empty env), dynamic `import()` of built-ins — so a deliberately adversarial program has ambient authority comparable to what the harness's own `bash` tool grants every model turn: `dsh-bash-local` runs arbitrary model-written commands with the host filesystem, network, and a scrubbed-but-populated environment. One asymmetry runs the other way: `worker.terminate()` ends the thread, not OS processes a program may have spawned via `node:child_process` — weaker than `bash-local`'s process-group kill for direct children (equivalent for double-forked daemons, which survive both); the wall-clock ceiling bounds the worker itself, and orphan cleanup is the same deployment-level concern it is for bash. Code Mode is gated where bash is gated — `tools/pre-execute`, where permission/sandbox plugins veto or approve the program — and adds containment bash does not have: empty env, heap caps, hard termination of the program itself, and a separate isolate. A `node:vm` executor with no containment would need explicit unsafe acknowledgement; imposing that ceremony on the better-contained worker while bash needs none would be posture theater. A deployment that needs a hard boundary (untrusted multi-tenant input) needs it for bash too; that is a future `isolation: 'container'` backend, and the `isolation` descriptor lets deployments distinguish backends.
+The worker runtime provides containment, not a security boundary: model code can reach Node APIs and has authority comparable to the bash tool. `worker.terminate()` stops the thread but not OS processes it spawned. Code Mode uses the same `tools/pre-execute` policy gate as bash and adds an empty environment, heap limits, a separate isolate, and hard termination of the program itself. Deployments that need a hard multi-tenant boundary need a container-class backend for both code and bash; the runtime's isolation descriptor lets them distinguish that backend.

 ### What the model sees

-The `tools:sdk` section carries the `.d.ts` plus fixed instructions: the program is the body of an async TypeScript function (erasable syntax only — no `enum`/namespaces; type annotations are advisory); call tools as `await tools.name(args)` (quoted access for exotic names); a failed tool call **rejects** with an `Error` carrying the tool's error text — catch it to handle and continue; calls run **sequentially** even under `Promise.all`; emit results via `return` and/or `console.log`, and only that curated output returns to the context — intermediate tool results never do. That last line is the payoff the whole design serves: output-side context cost becomes the model's own editorial decision. On the input side the `.d.ts` is not free — for a large tool surface it can rival the native JSON schemas it replaces (and `'both'` pays for the two side by side) — but it is prefix-stable, so provider prefix caching amortizes it; the win is workload-dependent and the RFC claims no more.
+The SDK instructs the model to write an async erasable-TypeScript body, call tools through `await tools.name(args)`, catch rejected tool calls when needed, and return or log only the output that should re-enter context. Calls remain sequential even under `Promise.all`. The declaration prefix can be as large as native schemas, especially in `'both'`, but remains stable for provider caching.

 ## Consequences

-The design consists of the `dsh-code-runtime` interface package, the `dsh-code-runtime-worker` backend, and the `dsh-tools` presentation and dispatch integration.
-
-Shipped surface:
-
- **The seam**: `packages/code-runtime/` — `@deepseek-ai/dsh-code-runtime` (abstract `CodeRuntime`, the vocabulary above, `ctx.codeRuntime`) and `@deepseek-ai/dsh-code-runtime-worker` (the worker-thread backend, every cap a validated config field). Rows in the service map, capability-seams graph, config catalog, and cordis catalog.
- **The registry surface**: `ToolRegistry`'s `mode` config, mode-aware wire contribution, lazy `tools:sdk` section and reserved `run_code` transport, `jsonSchemaToTs`/`renderToolsSdk` (exported), the dispatch bridge and `CodeRunFailedError`, and the `tool/code-dispatch` log event (declaration-merged into `SessionEventMap`, regenerated into the persistence catalog; `run_code` in the tool catalog).
- **The composed surface**: the `tools` config forwards through `agent-core` and both app packages (`stdio-agent`, `acp-agent`); `demo:code-mode` boots each UI example's `code-mode.cordis.yml` overlay (the worker runtime + `mode: 'code'` over the base tree); every program sub-dispatch resolves the same scoped capability view and re-enters the complete tool pipeline with an immutable link to its enclosing transport execution.
- **Interactions inherited by deployments**: a `toolOrder` naming native tools rejects every assembly under `'code'` (update or drop the order config when switching modes); restrictions can hide end capabilities but cannot remove the registry-owned presentation transport, while assembly listeners may rewrite the final model-visible surface and own its protocol integrity; sub-call `additionalContext` is dropped by the bridge (a plural context channel is deferred until a real hook needs it through Code Mode); sub-dispatch stays serialized until tools can declare concurrency safety — the same metadata the native parallel-dispatch TODO waits on.
+Deployments switching to `'code'` must update any native-only `toolOrder`. Assembly listeners own the integrity of any rewritten protocol surface. Sub-dispatch remains serialized, and the bridge does not propagate per-call `additionalContext` until those contracts are designed for Code Mode.

 ## Testing

-What the suites pin, per tier:
-
- **Unit — worker runtime** (real workers, no mocks): output/value capture and log-source attribution; error kinds (exception incl. non-erasable syntax, abort, worker-exit under OOM); the two budgets from both sides (a hot loop behind an un-awaited pending dispatch dies at `computeMs` busy time; a program idling on a slow binding outlives `computeMs` and dies only at `maxWallMs`); binding-bridge hostility (junk/forged port traffic incl. non-object messages and forged `log`/`done` cap bypass attempts, unknown names, duplicate ids, post-settlement replies, `__proto__`/`constructor`/`toString` binding names); structured-clone fallback and cap truncation; `env` emptiness verified from inside a program; dispose-awaits-exit. A real-load-path e2e runs the BUILT package under plain `node` so the worker entry resolves both unbuilt (tsx) and built — the published-artifact guard from [docs/testing.md](../../../testing.md).
- **Unit — registry integration**: the codegen table (DSL subset, quoted names, `unknown` degradation, byte-identical determinism); provider contribution per mode (`'native'` capabilities, `'code'` exactly `[run_code]`, `'both'` capabilities + `run_code`); reserved-name, restriction, scoped shadowing, authoritative assembly transformation, and `toolOrder × mode` invariants; missing-runtime / wrong-language loud failures; full-pipeline and opaque parent-token behavior for sub-dispatches; serialization non-overlap (a probe tool records enter/exit under `Promise.all`); abort aborting the in-flight sub-dispatch and abandoning queued ones; binding rejection on `isError` and on JSON-unrepresentable arguments; `CodeRunFailedError` → structured `isError` carrying kind + logs; `tool/code-dispatch` payloads (JSON-normalized arguments identical to what dispatched); `deriveMessages()` ignoring the event; sub-call `additionalContext` suppression; HMR safety.
- **e2e (with-key, self-skips)**: a real model under `mode: 'code'` composes two bash calls in one program (`examples/coding-agent/tests/code-mode.e2e.ts`) — every logged `request/header` carries exactly `[run_code]`, the dispatch events land under the parent call, the file the program wrote exists, and the final answer is the curated output.
- **Snapshot (keyless replay)**: goldens for a `run_code` turn under `'code'` and `'both'` (`code-mode-turn`, `both-mode-turn`), each its own header-pinning class — the SDK section text, the collapsed header tool list, the dispatch events, and the result card are committed and replayed.
+- **Worker runtime:** Real-worker tests cover output and value capture, failure kinds, compute and wall budgets, hostile binding traffic, empty environment, structured-clone fallback, output caps, and disposal to quiescence. A built-package test runs the worker entry under plain Node.
+- **Registry integration:** Tests cover code generation, all presentation modes, reserved-name and restriction rules, scoped visibility, authoritative assembly rewrites, `toolOrder`, runtime compatibility failures, full-pipeline sub-dispatch, parent-token correlation, serialization, cancellation and queue drain, JSON normalization, error propagation, log events, omitted `additionalContext`, and HMR cleanup.
+- **With-key e2e:** A real model composes two bash calls in one program; the test verifies the collapsed request header, correlated dispatch events, resulting file, and curated answer.
+- **Snapshot:** The `code-mode-turn` and `both-mode-turn` fixtures pin the SDK section, header tool list, dispatch events, and result card.

 ## Alternatives considered

@@ -132,7 +123,7 @@ What the suites pin, per tier:

 **Registry scope growth.** `dsh-tools` absorbs codegen, a tool, a bridge, and an event. Contained by module boundaries inside the package (`ts-types.ts`, `code-mode.ts` beside `schema.ts`/`json-schema.ts`/`presentation.ts`) and by the seam: everything substrate-shaped lives behind `ctx.codeRuntime`.

-**Structured-clone limits at the binding boundary.** The seam's clone boundary admits values JSON does not (`Date`, `Map`, `BigInt`), and the session log accepts only JSON — left unhandled, a sub-call could execute and then fail at `tool/code-dispatch` append time. Closed by the bridge's JSON-normalization step (§ the dispatch bridge): what does not survive the round-trip rejects that binding call before dispatch, so every executed sub-call is loggable by construction. The seam itself keeps the wider structured-clone contract (it is about the port, and stated so a future binding producer cannot discover it in production); consumers with stricter payload needs enforce them at their own boundary, as the bridge does. Non-text sub-result content is reduced to placeholders — a known MVP limitation, recorded in the SDK instructions.
+**Structured-clone values can exceed JSON.** Tool bindings therefore JSON-normalize arguments before dispatch, ensuring every executed call can be logged. The lower-level runtime keeps its wider port contract, while stricter consumers validate at their boundary. Non-text sub-results become placeholders.

 **Serialized-only sub-dispatch.** `Promise.all` gains no wall-clock parallelism yet, only fewer round-trips; models may over-expect. The instructions state it; lifting it is tied to the same concurrency-safety metadata the native parallel-dispatch TODO needs.

@@ -6,7 +6,7 @@ Status: implemented

 The ACP bridge lets each tool own its call rendering via `presentCall`/`presentResult` (see [tool-call UI presentation](../../implemented/feature/2026-06-14-acp-agent-client-protocol.md) and `packages/core/tools`). For `bash` we surface the exact command as the `tool_call` title, the model's `description` as a content text block, `kind: 'execute'`, and the completed output wrapped in a fenced ` ```console ` text block.

-That is a correct, capability-free baseline, but not how the reference editors render a *terminal* tool at its best. An editor like Zed has a dedicated terminal tool-call card — a header showing the working directory, the command as the label, the command output rendered as a terminal, and an exit-status pill — but it only builds that card when the `tool_call` carries terminal metadata (below). With a plain text block the output appears as static markdown and there is no cwd header. (Zed also HIDES `rawInput` for `kind: 'execute'`, which is why the command IS the title — both reference adapters do the same. The human-readable description rides as a separate content block above the card; note this is a DELIBERATE divergence — claude-agent-acp DROPS the description in terminal mode and renders only the card — we keep the summary visible alongside.)
+Reference editors render terminal metadata as a dedicated card with cwd, command, live-style output, and exit status; plain text loses that structure. The command is the title because execute cards hide raw input, while the human-readable description remains a separate block above the card.

 ## Key finding: agent-executed terminals use a `_meta` convention, NOT `terminal/create`

@@ -34,7 +34,7 @@ An earlier draft put the full algorithm (the retention walk, token-summing, text

 ### Auto-compaction runs on `agent/pre-step`, a dedicated surface-mutation seam

-Compaction is a **surface mutation**, not a request transform — and that distinction is the seam it belongs on. The loop's request lifecycle, per step, is: assemble the system prompt → open the step → derive the message history from the surface → run the `agent/request` waterfall → call the model. An earlier cut wedged compaction into the `agent/request` waterfall, which forced two problems: (1) the loop had already derived `messages` from the *stale* surface, so the listener had to mutate the surface and then *re-derive* and overwrite `request.messages` — a double-derive whose only purpose was to undo the premature first derive; and (2) `agent/request` also carries downstream-injected context a listener might have added to `request.messages`, which compaction cannot act on (it can only compact the surface), inviting the confusion of measuring tokens compaction can't shed.
+Compaction mutates the session surface, so it runs before the step opens and before messages are derived. `agent/request` remains a call-config transform and never needs to rebuild history after a surface change.

 The fix is a dedicated loop seam, **`agent/pre-step`** (`@mode serial`), fired by the loop *after* system assembly and *before* the step opens (`step/start`):

@@ -46,15 +46,13 @@ messages = session.deriveMessages()   ⟵ single derive, reflects the compaction
 request  = waterfall agent/request    ⟵ pure request transform (hooks, model switch)
 ```

-This makes the layering correct *by construction*: compaction mutates the surface, the loop derives **once** from the result (no double-derive), and at `pre-step` the assembled `messages` do not yet exist — so a listener structurally *cannot* see or be expected to act on downstream-injected context. `agent/request` reverts to a pure request transformer. Firing the seam **before** `step/start` (not inside the open step) is load-bearing for crash-safety: compaction's log-only `compact/*` records and its replacement node land *outside* any step, so the honest log structure a crash leaves (a dangling `compact/start` sitting before the synthetic `turn/end` that turn-repair appends) holds without a half-open step to reconcile. The seam is `serial` (awaited, in registration order), not `parallel`: a listener mutates the surface as a side effect — there is nothing to transform or return — and serial isolates listeners from each other so two surface-mutating listeners can never interleave their `session.append`s. Cordis `serial` does bail early if a listener returns a bail value, so `agent/pre-step` listeners are typed/documented to return `void` and must not use that bail channel as a semantic veto surface.
-
-This **amends** the original RFC's claim of "NO changes to `dsh-agent-loop`; compaction is a pure plugin." That claim was load-bearing for a wrong design — reusing `agent/request` was the mistake. Per the pre-release "foundation over blast radius" stance, adding the correct seam (one event declaration in `dsh-agent`, one awaited emit in the loop) beats preserving a no-change boast that locked in the double-derive.
+The loop derives messages once after `agent/pre-step`. Running before `step/start` keeps compaction records outside any half-open step, simplifying crash repair. The seam is awaited and serial so surface mutations cannot interleave; listeners return `void` and do not use Cordis bail values as vetoes.

 ### Retention is turn-agnostic; tool-pairing balance is the only structural guard

 Auto-compaction fires before **every** step, not once per turn. This is **load-bearing for runaway-turn survival**: a tool-heavy ReAct turn appends an `assistant/message` + a `tool/result` per step, so the surface grows *within* a turn. A single turn can grow past the window on its own (a "runaway turn") — and the only moment to rescue it before the next model call overflows is the next step's `pre-step` checkpoint. Gating compaction to a turn's first step (or, worse, retaining the whole in-flight turn verbatim) re-opens exactly the hole compaction exists to close: the harness would die when compaction is most needed.

-So retention does **not** protect the in-flight turn, and turn boundaries play no role in it. `compactIfNeeded` walks the surface nodes tail→head, summing per-node token estimates, and retains the smallest tail-run of **whole units** whose total reaches `retainTokens`; everything older is compacted (head-anchored — see below). A *unit* is either a whole closed step (its `assistant/message` plus its `tool/result`s) or a single no-step node (a pre-step `user/message`, inter-step `steering/message`, or injection `context/message`). The walk rounds toward retaining *more*: when the raw token cutoff lands mid-step, it extends the retained side head-ward until the cut before the retained node is **tool-pairing balanced**. The single structural guard is therefore **tool-pairing balance** — a region's edges are balanced cuts on the *surface* (no unanswered `tool-call` crosses either edge), so a compacted region never splits a step's tool-calls from their `tool/result`s (which would produce a transcript every provider rejects). The check is decided over the surface linked list, **not** the log's `step/*` markers: a compaction lands a replacement node at a high log seq whose surface position is the head, so a log-position scan mis-reads its neighbours — `dsh-session` exports `isToolPairingBalanced(nodes, events, beforeSeq)` for the surface-anchored check. `compactRegion` enforces it strictly, throwing on a boundary that would split a step.
+`compactIfNeeded` retains the smallest tail of whole surface units whose estimated size reaches `retainTokens` and compacts older nodes. A unit is a complete closed step or one no-step message. If the token cutoff lands inside a step, retention expands until the cut is tool-pairing balanced. Balance is checked on surface order, not log sequence, because replacement summaries have new sequence numbers at old surface positions. `compactRegion` rejects boundaries that split a tool call from its result. The in-flight turn receives no special retention.

 A runaway turn thus compacts exactly like any other history: its early *closed* steps get summarized while its recent steps stay verbatim. When the only compactable content left is an un-splittable open tail step (its tool-calls have no results yet), compaction declines (`null`) and retries once that step closes.

@@ -62,7 +60,7 @@ A runaway turn thus compacts exactly like any other history: its early *closed*

 ### Head-anchoring: one auto checkpoint, always at the head

-`compactIfNeeded` always anchors the compacted range at the surface **head** (`nodes[0]`). After a first compaction lands a summary node at the head, the *second* compaction's range starts at that summary node and re-summarizes it together with the steps accumulated since — so the surface holds **at most one** auto-generated checkpoint, always at the head, re-consolidated each cycle (the backend's checkpoint-merge prompt makes this a cheap incremental merge — see below). This is *why* `CompactionResult.shadowedRange` is a **surface-position span, not a numeric seq interval**: after a replace lands a fresh high-seq summary node at an older range's position, `start` can be numerically **greater** than `end`. The range is resolved positionally (index into the ordered node list and slice), and `shadowedSeqs` is the authoritative set in surface order. (Manual `compactRegion` may target any aligned mid-range and so *can* leave several checkpoints; the checkpoint framing does not claim everything after it is recent.)
+Auto-compaction always starts at the surface head, merging the prior checkpoint with newly compacted history so only one automatic checkpoint remains. `shadowedRange` is therefore positional rather than a numeric sequence interval: a newer summary sequence may occupy an older surface position. `shadowedSeqs` records the authoritative surface order. Manual mid-range compaction may leave multiple checkpoints.

 ### Approximate convergence invariant

@@ -85,7 +83,7 @@ compact/end      → log-only. Releases the lock (carries `error` on a recoverab

 ### Checkpoint framing + incremental merge (backend-private)

-The landed `user/message` is not the raw summary: the backend wraps it in a checkpoint preamble (so a resuming model reads it as established background, not a fresh request) and `<compacted-summary>…</compacted-summary>` tags. The tags make a prior checkpoint detectable on the next cycle, and the summarization prompt then instructs the model to *merge it in place* (preserve still-true facts, drop stale) rather than re-summarize verbatim — a cheap incremental merge that needs no extra log/event machinery. The raw, unframed summary stays on the `compact/summary` provenance event. This framing is entirely a **backend HOW decision** — the contract only promises "a single replace `user/message` carries the (possibly framed) summary; the raw summary lives on `compact/summary`." A template or remote backend may frame differently or not at all.
+The basic backend wraps the summary as established checkpoint context and tags it for incremental merging on the next cycle. The raw summary remains on `compact/summary`. Framing is backend policy; the seam promises only that one replacement user message carries the possibly framed summary.

 ### Blocking via a log-recorded lock, plus a crash/recoverable failure taxonomy

@@ -121,7 +119,7 @@ Two failure paths, both documented:

 ## Testing

- **Unit** (`dsh-compact-basic`): the whole-unit retention walk, the convergence-invariant throw, both failure paths (`compact/end` with/without `error`), head-anchoring producing a non-monotonic `shadowedRange`, decline-on-open-tail, crash-orphan inertness, and the **runaway-turn regression** — a single oversized open turn compacts its early closed steps (proven to fail on the layer-2 protection it replaced). Driven through the real `dsh-invariants` plugin and the real Loader/inject path.
- **Loop** (`dsh-agent-loop`): `agent/pre-step` fires once per step, after `turn/start` and before `step/start`, awaited; a surface mutation in a `pre-step` listener lands outside the step and is reflected in the single derived request.
- **With-key e2e** (`examples/coding-agent`): a real model + real bash session with a lowered `contextWindow`/`retainTokens` triggers compaction mid-session; the test verifies the WORLD (a `compact/start…end` pair landed, the surface shrank, the agent still completed the task after compaction). This is compaction's first real-world exercise and the runaway-survival net.
- **Snapshot (deferred, named gap)**: a full-transcript snapshot of a runaway-turn compaction is NOT yet possible — `dsh-llm-replay` derives one model call per `(turn, step)` from `assistant/chunk` events, but the summarization call records no `assistant/chunk`s and carries no `sessionId` (it binds to the anonymous cursor and claims a non-existent extra script). Covering it needs net-new replay infrastructure (record/replay an interleaved summarization call) and is scheduled as a follow-up rather than discovered mid-build.
+- **Unit:** Real Loader and invariant plugins cover whole-unit retention, convergence failure, both `compact/end` outcomes, head anchoring, open-tail refusal, inert crash orphans, and compacting closed steps inside one oversized open turn.
+- **Loop:** Tests pin one awaited `agent/pre-step` per step between `turn/start` and `step/start`; a surface mutation there lands outside the step and appears in the single derived request.
+- **With-key e2e:** A real model and bash session with lowered limits triggers compaction, records a complete `compact/start…end` pair, shrinks the surface, and finishes the task.
+- **Snapshot gap:** Runaway-turn compaction cannot yet replay because the summarization call records no `assistant/chunk` events or `sessionId`; interleaved summarization-call replay remains follow-up work.
@@ -37,7 +37,7 @@ A new package group `packages/subagent/`:

 ### The primitive: async `start → SubagentRun`

-A provider exposes `start(request) → Promise<SubagentRun>`. Promise fulfillment is the publication/readiness and provider-to-caller ownership boundary: for an in-process backend the child is already published in `ctx.agents`, and for ACP the remote session already exists. `SubagentStartRequest.signal` is the single cancellation channel before and after readiness; `SubagentRun` carries the terminal `result` and a `dispose()` method that cancels remaining work and awaits quiescence. The transport-neutral verb is **`start`**; "spawn" is reserved for the in-process `dsh-subagent-spawn` backend's identity, not the service verb. A rejected start cleans provider-owned partial resources and emits neither subagent lifecycle event.
+A provider exposes `start(request) → Promise<SubagentRun>`. Fulfillment publishes a ready child and transfers its run handle to the caller. One signal covers cancellation before and after readiness; `dispose()` cancels remaining work and awaits quiescence. A rejected start cleans partial resources and emits no lifecycle event. `start` is transport-neutral; `spawn` names only the fresh in-process backend.

 ### Two kinds of optional capability, discovered two ways

@@ -46,7 +46,7 @@ A provider exposes `start(request) → Promise<SubagentRun>`. Promise fulfillmen

 ### Fork vs. fresh are separate backends, not a flag

-Rather than a `context: 'fresh' | 'fork'` request field, the distinction is the provider's identity: `dsh-subagent-spawn` (fresh, isolated, own system prompt) and `dsh-subagent-fork` (seeded from the parent's log) are two registered providers. You pick behavior by picking a provider — consistent with the registry being the selection mechanism. The fork backend seeds only a **balanced, completed-turn prefix** of the parent log: at tool-execute time the parent's turn is open (it holds the `assistant/message` and the dangling spawn `tool/call` with no `tool/result`), and seeding that raw prefix would give the child an unbalanced turn that the [invariants](../../../../packages/support/invariants/src/index.ts) trace replay rejects.
+Fresh and forked children are separate providers, not a request flag. `dsh-subagent-spawn` starts an isolated child; `dsh-subagent-fork` seeds a balanced prefix containing only completed parent turns. The in-flight turn is excluded because its subagent call has no result yet and cannot form valid replay history.

 ### Child isolation and the parent log

@@ -54,7 +54,7 @@ Each subagent runs in its **own `Session`** (own id, `parentSession` lineage), p

 ### Synchronous collect (first cut)

-The `dsh-tool-subagent` consumer passes its execution signal into the start request, awaits the ready run's `result`, and returns the child's final output as the tool result, blocking the parent's turn until the child finishes. A `try/finally` always `dispose()`s the run, so no success, failure, or cancellation path leaks an idle child/session. A non-`completed` stop reason maps to an `isError` result rather than returning partial output as success. Steering (`sendMessage`) is part of the contract but intentionally unused in this consumer.
+`dsh-tool-subagent` passes its execution signal to `start()`, awaits the child result, and disposes the run in `finally`. Non-completed outcomes become error results rather than successful partial output. This foreground consumer does not use the run's optional steering method.

 ### Provider selection is config, not model-facing

@@ -62,7 +62,7 @@ The `dsh-tool-subagent` consumer passes its execution signal into the start requ

 ## Testing

-The seam is tested through the real cordis Loader / export path, not a hand-built `ctx.plugin` mount (which bypasses `unwrapExports` and cannot catch a broken export shape — [postmortem 0001](../../../postmortem/0001-acp-default-export-drops-inject.md)); the registry pins HMR-safety, duplicate-name rejection, and start-time capability rejection; the nested-agent snapshot scenarios replay keyless in the default gate ([per-session snapshot replay](../testing/2026-06-22-subagent-snapshot-replay.md)); in-process backends carry real-loop unit tests plus a with-key e2e.
+The seam is tested through the real Cordis Loader/export path, which catches the export-shape failure described in [postmortem 0001](../../../postmortem/0001-acp-default-export-drops-inject.md). Registry tests cover reload safety, duplicate names, and start-time capability rejection; nested-agent scenarios replay keylessly through [per-session snapshot replay](../testing/2026-06-22-subagent-snapshot-replay.md); in-process backends also have real-loop unit tests and a with-key e2e.

 ## Consequences

@@ -70,4 +70,3 @@ The seam is tested through the real cordis Loader / export path, not a hand-buil
 - **Blocking the parent turn.** Synchronous collect holds the parent's `runStep` open for the child's full duration. This is acceptable for the first cut; **background / poll / spill semantics are deferred to a future redesign that unifies long-running-tool handling across subagents AND bash** (a sub-agent and a long `bash` background task pose the same "the model started something slow, how does it collect later" problem, and should share one mechanism rather than each inventing its own).
 - **Live progress.** This cut surfaces only lifecycle + final result; a per-chunk child→parent update stream is deferred with the background redesign.
 - **ACP client surface.** Proxying `fs`/`terminal` from the ACP child back to the parent (a shared-workspace mode) is future work; the first cut advertises neither, so the child self-serves in its own process.
- **Snapshot coverage of nested agents.** The snapshot tier (`pnpm run test:snapshot`) replays a recorded session through `dsh-llm-replay`. It was built single-session: a single GLOBAL positional cursor (the Nth `llm/stream` call serves the Nth recorded entry) and a harness that harvested a single session log file. A subagent runs as a *second* agent with its own session log, so a parent→child scenario needed per-session-keyed replay plus harvest-all-logs and plural-session-id plumbing — self-contained infrastructure orthogonal to the backends, scheduled as a dedicated stacked follow-up rather than folded into the in-process-backends PR. That follow-up has **landed**: see [Per-session snapshot replay for nested agents](../../implemented/testing/2026-06-22-subagent-snapshot-replay.md). Replay now keys each call by its calling session (`GenerateOptions.sessionId`) and binds live sessions to recorded scripts by first-call order; the harness harvests every log; and two nested scenarios (`subagent-spawn`, `subagent-multi`) replay keyless in the default gate. In-process subagents remain covered by real-loop unit tests and a with-key e2e in addition to the snapshot tier.
@@ -32,17 +32,15 @@ The child is a separate process, so it inherits an environment. Credential-shape

 ## Testing

-Designed at every tier the backend touches, per the root AGENTS.md rule that a new capability shape names its coverage at every tier at plan time:
-
- **Keyless unit/integration** (`subagent-acp.spec.ts`): spawns a scripted mock ACP server subprocess (`tests/mock-acp-server.ts`) and drives it through the real backend over real ACP stdio. Coverage includes the prompt round-trip and output accumulation; every StopReason mapping; cancellation through the required request signal and through disposal; already-aborted and cancel-races-ahead-of-newSession starts; a torn pipe after cancellation settling `aborted`; permission auto-answer under both policies; non-message updates; nonexistent-command startup failure with process reaping; provider HMR; and the namespace export shape.
- **With-key e2e** (`subagent-acp.e2e.ts`): the harness drives ITSELF — the backend spawns the real `acp-agent` example process and a real model in that child answers a prompt (PONG) and does real file work (writes `proof.txt`, verified on disk). Self-skips without `DEEPSEEK_API_KEY`. This is the "talk to our own process" smoke and the out-of-process analogue of the in-process spawn e2e.
- **Snapshot**: deferred as `TODO(acp-subagent-replay)`. An ACP child is a distinct replay shape — each child is its own PROCESS with its own single-agent replay (booted under `DSH_SNAPSHOT=replay` with its own sessions-root + fixture), unlike the in-process per-session keying that [the per-session replay RFC](../testing/2026-06-22-subagent-snapshot-replay.md) added. The keyless mock-server tests give deterministic coverage of the backend in the meantime; the snapshot follow-up would record the parent driving a real-but-replayed ACP child.
+- **Keyless unit/integration:** A scripted ACP subprocess exercises real stdio for prompt/output flow, every stop-reason mapping, signal and disposal cancellation (including pre-abort, pre-session race, and torn-pipe cases), both permission policies, ignored non-message updates, missing-command cleanup, provider reload, and namespace exports.
+- **With-key e2e:** The backend spawns the real ACP example; its model answers `PONG`, writes `proof.txt`, and the parent verifies the file.
+- **Snapshot gap:** Each ACP child is a separate process with its own replay session, unlike in-process per-session replay. Deterministic mock-server coverage exists, while `TODO(acp-subagent-replay)` tracks parent replay against a replaying child.

 ## Alternatives considered

-### Why not the 0.28.x SDK bump?
+### Why stay on SDK 0.25.1?

-The plan proposed bumping `@agentclientprotocol/sdk` 0.25.1 → 0.28.x for the new fluent `acp.client()` / `ActiveSession.nextUpdate()` API. Validating that against the code (the AGENTS.md "RFC is a proposal, not golden truth" discipline) reversed the decision: the backend only needs `ClientSideConnection` + `ndJsonStream` + `PROTOCOL_VERSION` + the `Client`/`Agent`/`StopReason` types, **all present and non-deprecated in 0.25.1**. The fluent API and `unstable_forkSession` that motivated the bump are never used here, so the "cleaner client code" benefit did not materialize. Worse, 0.28.x **deprecates both** `ClientSideConnection` AND `AgentSideConnection` (it wants all callers on the fluent builders), which turns the `no-deprecated` lint red across the entire existing ACP layer — 33 usages including the server bridge this backend has no business rewriting. That cross-cutting connection-API migration is its own change, not baggage for "add an ACP subagent backend". So the bump was reverted and the backend is written against 0.25.1 (the plan's own fallback clause: "if the bump proves disruptive, fall back to `ClientSideConnection` (0.25.1), which is sufficient"). Migrating the whole ACP layer to the fluent API on a later 0.28.x bump is a worthwhile standalone follow-up.
+The backend needs only `ClientSideConnection`, `ndJsonStream`, `PROTOCOL_VERSION`, and the client protocol types, all supported in 0.25.1. The 0.28 fluent API would require migrating both client and server connection classes across the ACP layer without improving this backend, so that upgrade remains a separate change.

 ### Why not a persistent child process?

@@ -32,7 +32,7 @@ claude-code V1's item is `{ content, status, activeForm }`; later (V2) it grew i

 ### Single owner — no swarm machinery (YAGNI)

-The list belongs to the ONE agent session that called the tool (`exec.agent.session`); a non-agent caller is rejected. There is deliberately no shared/multi-owner scope, no capability seam (interface/impl/consumer), no scope resolver, and no delta protocol. The harness does have subagents, and a shared cross-agent list is conceivable — but building that now means designing for a form the product does not yet have. The whole-list-replace + single-owner shape is what claude-code V1, opencode, and codex all ship; if a shared list is ever needed, the on-log representation would change to per-item deltas (so concurrent writers can't clobber each other) and a scope resolver would choose the target log. That is a future RFC, not speculative scaffolding today.
+Each list belongs to the calling agent session, and non-agent calls are rejected. There is no shared scope, resolver, or delta protocol. Cross-agent lists would require per-item log deltas and explicit scope selection, so they remain a separate future design.

 ### Validation: the cheap middle

@@ -33,11 +33,11 @@ The CC bridge's `ask` result is a real permission path, not a terminal bridge de

 ### Context source is always the plugin (the mislabel guard)

-`agent.inject()` defaults a missing `MessageSource` to `{ kind: 'user' }` — which would record plugin-injected context as if the user had typed it. So every bridge `inject()` and every `HookContext` passes an explicit `{ kind: 'plugin', plugin: 'hooks-claude' | 'hooks-codex' }` source. A test asserts the resulting `context/message.source` is the plugin, never `user`.
+`agent.inject()` defaults a missing `MessageSource` to `{ kind: 'user' }`, so every bridge `inject()` and `HookContext` passes `{ kind: 'plugin', plugin: 'hooks-claude' | 'hooks-codex' }`. Unit coverage pins the resulting `context/message.source` as the plugin rather than the user.

 ### Adding context is not a veto — delegate, then fold

-A hook that only attaches `additionalContext` (no block/deny) is NOT a decision the bridge should return on its own: returning `allow`/`accept` from a waterfall listener WITHOUT calling `next()` short-circuits every later `agent/prompt-submit` / `tools/post-execute` listener, so a policy/sandbox plugin registered after the bridge would never see the prompt. So on the context-only path each bridge **delegates via `next()`** and then **folds** its `additionalContext` onto the downstream decision (`concatContext`). The fold differs by seam because the two Decision unions differ: `tools/post-execute` — a downstream `block`/`accept` both carry an `additionalContext` field, so the bridge context rides along either way (a downstream block wins AND keeps the context; a downstream accept keeps its content rewrite and gains the context). `agent/prompt-submit` — a downstream `allow` gains the bridge context (and keeps its own content rewrite / additionalContext), but `PromptDecision.block` carries no context field, so a downstream block drops the bridge context — which is correct: a blocked prompt never reaches the model, so context attached to it is moot. Only a real `deny`/`block` from the hook itself short-circuits. Tests assert a later listener can still block a prompt a context-only hook allowed, and that both contexts survive when the downstream also adds one.
+A context-only hook must call `next()` and then fold its `additionalContext` into the downstream decision; returning allow or accept directly would bypass later policy listeners. Post-tool block and accept decisions both preserve added context. Prompt allow preserves it, while prompt block drops it because the prompt never reaches the model. Only an explicit hook denial or block short-circuits the waterfall.

 ### CLAUDE_PROJECT_DIR defaults to the session workspace

@@ -49,7 +49,7 @@ The config is parsed ONCE at load; a read/parse failure logs and registers nothi

 ### Where hooks run, and where their config comes from

-Two different cwds, kept distinct on purpose. The hooks **themselves** run in the agent's **session workspace**: for the agent-scoped points the bridge threads the session's `cwd` (`session/new.cwd`, on the session header) to `runHook` as the process working directory, so a hook's `pwd` / relative-file read / marker write operates in the user's project tree, not the server's launch directory. The **config path**, by contrast, is **process-level**: `configPath` is resolved and parsed once at load against the process launch cwd, so a single `hooks.json` applies to the whole process — there is no per-session config discovery that reads a project-local `hooks.json` from each `session/new.cwd` (`TODO(per-session-hook-config)`). This is an honest limitation of the current cut: the example `cordis.yml` documents that its `./hooks.json` is process-level, not per-project.
+Hooks run in the agent's session workspace, so relative paths target the user's project. `configPath` is resolved once against the process launch cwd and applies to every session. Per-session project-local discovery remains deferred under `TODO(per-session-hook-config)`.

 ## Deferred compatibility gaps

@@ -57,7 +57,7 @@ Two different cwds, kept distinct on purpose. The hooks **themselves** run in th
 - **Stop loop-guard** (`TODO(stop-loop-guard)`). Claude Code supplies `stop_hook_active` and overrides a hook after eight consecutive blocks; Codex supplies `stop_hook_active` but documents no equivalent cap. Both bridges always report `false`, so a Stop hook that unconditionally blocks force-continues every step — a hook author must self-limit until state tracking lands.
 - **Hook `continue:false` (hard halt).** A hook can ask to halt the whole run (CC/Codex `continue:false`); the shared merge folds it into `MergedHookOutcome.stop`/`stopReason`, but no bridge acts on it (`TODO(hook-continue-false)`) — the interception seams have no "hard-halt the agent" primitive yet (a Decision blocks/steers a single point, not the run). Deferred with the loop-guard work; the halt request is recorded in the `hook/result` log, and the hook keeps its per-point effect (decision/context) meanwhile.
 - **Config discovery.** The path is explicit in `cordis.yml` and process-level (see above); the full multi-layer CC/Codex precedence walk, per-session project-local discovery, and the trust/hash model are not reimplemented (`TODO(per-session-hook-config)`).
- **Session-start / subagent-start context is best-effort, not gated (`TODO(session-start-gating)`).** `agent/session-start` is a synchronous emit and the bridge runs its hook on a detached `.then`, so the injected `additionalContext` is not guaranteed to land before the first turn reaches the model — a slow hook can miss the first request (the context then arrives as a later injection). `subagent/start` is emitted only after child publication, so the bridge can capture the live in-process child synchronously, but the result driver may queue the prompt as that same readiness boundary resolves and a short-lived child can finish before the detached hook injects. Making startup context a gated/awaited primitive is a loop-level change deferred to the interception seams; the contract is "injected as soon as the hook resolves", not "before the first request". The bridge tests do NOT wait on the injection where they assert the guaranteed-timing behavior, so they document the real (best-effort) timing rather than masking it.
+- **Session-start / subagent-start context is best-effort (`TODO(session-start-gating)`).** Both hooks run detached from startup, so their context is injected when ready but may miss the first request or a short-lived child. Guaranteeing first-request delivery requires an awaited startup seam.

 ## Alternatives considered

@@ -65,4 +65,4 @@ Two different cwds, kept distinct on purpose. The hooks **themselves** run in th

 ## Consequences

-The bridges are thin and readable standalone: the correctness-critical halves (matcher semantics, exit-code contract, merge precedence) live in the shared `dsh-hook-protocol`, so each bridge is just config-parse + payload-build + outcome-map. Each is covered at per-file 100% — config-parse branches as unit tests, and the seam mappings end-to-end through the REAL loop + REAL `dsh-bash-local` + REAL shell scripts from a temp `hooks.json` (a scripted mock MODEL is the only stand-in), plus a real-Loader export-shape guard so a stray default export can't silently drop `inject`. Because the seams already carry typed Decisions, a future native plugin needs none of this bridge machinery — it returns a Decision directly.
+Matcher semantics, exit-code handling, and merge precedence live in `dsh-hook-protocol`; each bridge only parses config, builds dialect payloads, and maps outcomes. Per-file coverage includes config branches plus end-to-end mappings through a real loop, `dsh-bash-local`, and shell scripts, while a real-Loader smoke guards the package export shape. Native plugins bypass the wire protocol and return typed decisions directly.
@@ -23,8 +23,8 @@ A new `packages/hooks/` group with `hook-protocol` as a pure library. It owns fo

 ## Alternatives considered

-**One parameterized engine.** A single engine parameterized by a full `dialect` descriptor was considered and rejected. The payload construction and decision mapping are where the dialects genuinely diverge (different field names, different supported outputs, CC's env/substitution); folding those into a data-driven descriptor would make the *bridge* logic indirect — a reader of `dsh-hooks-claude` would have to chase a descriptor to see what payload it sends. Keeping the truly-identical primitives shared (matcher, codec, runner, merge, events) and letting each bridge write its own straightforward payload+mapping keeps each bridge readable standalone, at the cost of a little duplication in the payload shape. The primitives are the part where duplication would actually be dangerous (a divergent matcher or exit-code rule is a correctness bug); the payload is the part where explicitness beats sharing.
+**One parameterized engine.** Rejected because payload construction and decision mapping genuinely differ by dialect. Matchers, codecs, execution, merge rules, and events remain shared; each bridge keeps its payload and mapping explicit so its wire behavior is readable in place.

 ## Consequences

-The two bridge plugins become thin: parse the config file, pick a matcher mode, build the per-event payload+env, call `runHook` + `mergeHookOutputs`, map the outcome to a Decision, and append `hook/*`. The protocol's correctness-critical halves (matcher semantics, exit-code contract, merge precedence) live in one tested place — `hook-protocol` ships with heavy unit tests (matcher per-mode, codec per exit-code/field, runner plumbing with a stub executor, merge precedence, the `hook/*` helpers) at per-file 100%. Input rewrite (`updatedInput`) is parsed but not honored (the deferred [pre-tool-input-rewrite RFC](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md)); a bridge logs+warns on it. The package is a library, so it has no `cordis.yml` load path of its own — its real-load-path coverage comes through the bridge plugins that consume it.
+Each bridge parses config, builds its dialect payload, invokes the shared runner and merge logic, maps the decision, and appends `hook/*`. Protocol tests cover every matcher mode, exit-code and codec field, runner plumbing, merge precedence, and audit helper at per-file 100%; bridge tests exercise the library's real load path. `updatedInput` is parsed but only logged and warned until the [input-rewrite proposal](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md) lands.
@@ -20,7 +20,7 @@ The canonical surface separates transformable policy, around-dispatch control, a

 ### The tool pipeline gives each phase one kind of authority

-Every call follows one ordered pipeline: `tools/pre-execute` → monotonic guards → `tools/execute` → core dispatch → `tools/post-execute` → `tools/result`. The registry reads each caller-owned input field once, materializes `arguments` as detached lossless JSON in one recursive pass, and snapshots `ToolExecutionInput` into a pipeline execution with its own opaque token. Identity fields and deeply frozen arguments are immutable for the whole pipeline, and a nested call's `parent` contains only the enclosing execution's token rather than its live object. Optional `signal` is the only operational field an around-dispatch wrapper may add, replace, or remove, and the complete object freezes before final observers run. This identity contract prevents a policy listener from silently changing what the log, UI, and tool body believe ran.
+Every call follows `tools/pre-execute` → guards → `tools/execute` → dispatch → `tools/post-execute` → `tools/result`. The registry snapshots caller input, materializes and freezes arguments, and assigns an opaque token. Nested calls carry only the parent token. Identity remains immutable; only `signal` may change around dispatch. The log, UI, and tool body therefore agree on what ran.

 - **`tools/pre-execute`** is the extensible waterfall gate. Its `PreToolDecision` allows, denies, or asks. Deny skips `tools/execute` and core dispatch. Ask resolves through the optional approval seam: only `allowed-once` continues through guards and dispatch; rejection, cancellation, an unavailable channel, a missing approval service, or an agent-less call becomes a normalized denial. Every outcome still reaches post-policy and final observers.
 - **`ctx.tools.guard()`** installs synchronous scope-aware policy after the whole pre-execute waterfall. A guard may deny or abstain, never force-allow, so listener ordering cannot resurrect an operation that a final invariant forbids.
@@ -34,7 +34,7 @@ Core dispatch and the tool body sit inside normalization boundaries, so tool, li

 ### Three load-bearing loop decisions

-1. **Always open the turn first; a fully-blocked batch is a zero-step `rejected` turn; every veto is recorded as `prompt/blocked`.** `prompt-submit` fires AFTER `turn/start`, per message. A batch whose every prompt is blocked does NOT skip the turn — it opens a zero-step turn that closes with `rejected`. This one move resolves three problems at once: (1) turn-enclosure holds (every event has an open turn to live in); (2) the durable `turn/end` is appended and the ACP bridge settles normally off it (mapping `rejected`→`cancelled`) instead of hanging; (3) the block reason is a durable in-turn fact. Independently, each individual veto appends a `prompt/blocked` session event (the original `content`, `source`, and `reason`) in place of the `user/message` the prompt would have become — necessary because a MIXED batch (one prompt blocked, another allowed) does NOT end `rejected`, so the boundary reason alone would silently lose the blocked prompt on replay. An `allow`'s `additionalContext` is `inject()`ed into this now-open turn.
+1. **Open the turn before prompt policy.** A fully blocked batch becomes a zero-step `rejected` turn, preserving enclosure and giving ACP a durable terminal event. Every veto also records `prompt/blocked` with the original prompt and reason, so mixed batches retain blocked inputs. Allowed `additionalContext` is injected into the open turn.

 2. **Post-tool `additionalContext` is buffered and appended AFTER all `tool/result`s.** `content`/`feedback` shape the result `execute()` returns, but `additionalContext` is a SEPARATE `context/message`, and a single step can carry multiple tool calls. Appending context right after each result would interleave `result(c1) → context → result(c2)` and break tool-call/result adjacency. So `execute()` surfaces `additionalContext` on its `ToolExecutionResult`, and the loop buffers every per-call context for the step and appends them as `context/message`(s) only after every `tool/result` is appended.

@@ -42,7 +42,7 @@ Core dispatch and the tool body sit inside normalization boundaries, so tool, li

 ### Pre-tool input rewrite is a separate consistency decision

-`PreToolDecision` is allow/deny/ask only — **no `arguments` rewrite**. Output replacement is safe because `tool/result` is logged after execution from the final result. Input rewrite is different: `assistant/message` (model history) and `tool/call` (the audit record) are logged before `ToolRegistry.execute()`, while ACP and tool presentation read those arguments. The registry therefore seals the materialized arguments before `tools/pre-execute`; no listener or test shim can mutate them in place. An honest rewrite must update history, audit, presentation, and execution as one unit before that identity is created, which belongs to the separate [pre-tool input-rewrite proposal](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md) and its loop-side `TODO(pre-tool-input-rewrite)`.
+`PreToolDecision` cannot rewrite arguments. History and the audit call are logged before execution, and ACP presentation reads the same input, so the registry seals arguments before policy. A valid rewrite must update history, audit, presentation, and execution before identity is created; that contract belongs to the [input-rewrite proposal](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md).

 ### Boundaries

@@ -24,7 +24,7 @@ class SessionStore extends Service {

 `boundary` is the inclusive source event `seq` to copy through. When omitted, it defaults to the source session's current last event; on an empty source, omitted `boundary` creates an empty child. Fork-specific validation only checks that the requested boundary exists and is a `turn/end`. The selected prefix is then deep-cloned into the child seed. The child inherits the source session's `cwd`, stamps `parentSession` to the source id, and sets `seedLength` to the copied prefix length. When `childSessionId` is omitted, `SessionStore` generates one using its existing id policy.

-The boundary rule is structural: an empty selected prefix is forkable, and any non-empty selected prefix must end at `turn/end`, regardless of the turn-end reason (`completed`, `aborted`, `error`, `disposed`, `max-tokens`, `interrupted`, or a future merge-extensible reason). A boundary that is not an existing event seq, is not a safe integer, or does not point at `turn/end` is rejected with a typed `SessionForkError` code. Broader session-log sanity remains in the existing invariant/repair layers: `dsh-invariants` checks turn enclosure and richer event ordering in dev, while persistence repair handles the valid crash-tail case of a final interrupted turn. The API also classifies non-live source ids (`SESSION_NOT_FOUND`), stale `Session` object references whose id is live on a different instance (`SESSION_NOT_LIVE`), duplicate requested child ids (`SESSION_ALREADY_EXISTS`), and invalid boundary values (`INVALID_BOUNDARY`).
+An empty prefix is forkable; any non-empty boundary must be a safe existing sequence at `turn/end`, regardless of reason. Typed errors distinguish missing sources, stale objects, duplicate child ids, and invalid boundaries. Broader log validation and crash repair remain with their existing owners.

 ## Alternatives considered

@@ -12,7 +12,7 @@ A workflow capability family at `packages/workflow/` in the bash seam shape (int

 ### The script contract (Claude Code-compatible)

-A workflow call is two parts: a `meta` JSON parameter (the identity block — `name`, `description`, optional `whenToUse`/`phases`; the field vocabulary matches Claude Code's meta block) and a `script` — a plain-JS body with top-level `await`, ending in `return <json-value>`. Meta is DATA, never code: the engine shape-validates it and evaluates no script text to obtain it (a body still opening with a CC-style `export const meta` statement is rejected with a pointed message). The body sees exactly: `agent(prompt, {label, phase, schema, model})`, `parallel(thunks)`, `pipeline(items, ...stages)` (NO cross-stage barrier; `(prev, item, index)` callbacks), `phase(title)`, `log(message)`, and `args`. CC semantics are preserved where they matter to script authors: a failed child resolves `null` (scripts `.filter(Boolean)`); an ordinary stage throw nulls the ITEM and skips its remaining stages. CC's determinism bans (`Date.now()`/`Math.random()`/argless `new Date()` throwing) are NOT enforced — they exist for CC's journaling/resume, which this cut defers — so a CC-authored BODY runs unchanged (its meta header moves into the parameter) while scripts written here may freely read the clock.
+A workflow call contains JSON `meta` (`name`, `description`, and optional `whenToUse`/`phases`) and a JavaScript `script` body with top-level `await` that returns a JSON value. Metadata is validated as data and never evaluated. The body receives `agent(prompt, options)`, `parallel(thunks)`, `pipeline(items, ...stages)`, `phase(title)`, `log(message)`, and `args`. Pipeline stages receive `(prev, item, index)` with no cross-stage barrier; failed children and ordinary stage errors resolve the affected item to `null` and skip its remaining stages. Claude Code's determinism restrictions are deferred with journaling, so compatible bodies may use clock and randomness after moving their meta header into the parameter.

 One deliberate strictness DIVERGENCE from CC: hook misuse — unknown or deferred options (`effort`/`isolation`/`agentType`), malformed arguments, schemas outside the supported subset, tripped caps, seam start failures — throws a `WorkflowError` with `fatal: true`, and the combinators RE-THROW fatal errors instead of nulling the item. Without this, a typo'd option dissolves into a `null` indistinguishable from a child failure — the accepted-then-ignored failure mode this repo bans. One addition: the tool's `args` parameter is a JSON OBJECT (a bare list is wrapped as a field) so the wire schema stays honest.

@@ -22,15 +22,17 @@ One deliberate strictness DIVERGENCE from CC: hook misuse — unknown or deferre

 ### The engine (dsh-workflow-workerthread): one worker thread per run

-**Trust premise (governs every engine decision below)**: workflow scripts are MODEL-WRITTEN — the same trust level as the model's existing bash access — so the engine defends against BUGGY scripts, never hostile ones. In scope: `result` never rejects, no unhandled rejections from dropped hook promises, loud rejection of values JSON cannot carry, fatal-vs-null hook discipline, cancellation that always frees the caller. Out of scope, deliberately: adversarial values (throwing/spinning accessors, proxies with hostile traps, prototype forgery, `prepareStackTrace` hijack) AND Node-API escape from the script's context — the vm context shares object machinery with its surrounding realm, so a script can reach the `Function` constructor (`globalThis.constructor.constructor`) and from it `process` and every Node builtin; the absent globals are API surface, not containment, and a worker thread is NOT a security boundary (an escapee holds process-wide privileges — Node's permission model is per-process). Worker-side code MAY run script code while reading script values, and that is accepted: a synchronous spin costs the script its OWN thread (terminated at the post-cancel grace), never the host loop, so containing error VALUES would be cost without a threat model. Genuine sandboxing (isolated-vm, a separate process) remains an engine swap behind the seam, not incremental defenses here.
+**Trust premise**: workflow scripts have the same trust as the model's bash access. The engine contains buggy scripts and guarantees settled results, JSON-safe values, and cancellation quiescence; it does not defend against hostile code. A vm context and worker thread are not security boundaries: a script can escape to Node APIs with process-wide authority. Sandboxing requires a separate-process or isolated-vm engine behind this seam.

-**Why node:worker_threads**: one run uses one unpooled worker because a workflow run is already heavyweight relative to thread startup. The script runs in a vm context inside the worker, keeping the script-visible surface to the hook contract instead of exposing a bare worker realm, while `agent()` bridges by message-port RPC to I/O-bound child loops on the host. This keeps `start()` from blocking the host on the script's synchronous slice, makes the post-cancel deadline end in a real `worker.terminate()`, and gives cross-thread values a serialization boundary by construction. isolated-vm was rejected for its maintenance state, required `--no-node-snapshot` consumer flag on Node ≥ 20, and node-gyp fallback.
+**Why `node:worker_threads`**: each run gets one unpooled worker. A vm context limits the documented script surface, while message-port RPC bridges `agent()` to host-side child loops. The worker prevents synchronous script work from blocking the host, provides a serialization boundary, and permits forced termination after cancellation. `isolated-vm` was rejected because of its maintenance state and deployment requirements.

-Host-side meta validation and body pre-parsing preserve the seam's synchronous errors, and private enum-keyed payload maps define the wire protocol. Pending async starts, published child records, one host cancellation signal, worker-death reaping, result precedence, and disposal quiescence preserve the subagent run contract across that wire; the [agent-scope runtime-design RFC](../architecture/2026-07-12-agent-scope-runtime-design.md#workflow-children-are-pending-starts-or-published-records) owns those race algorithms. Coverage uses an in-process `MessageChannel` for worker-side logic that main-process V8 coverage cannot see and separately proves the built `lib/worker.cjs`—a second tsdown entry sanctioned by the `"./worker"` subpath export—under plain Node in the built-bin smoke gate.
+The host validates metadata and parses the body before publication. Private enum-keyed payload maps define the wire protocol; pending starts, published child records, one cancellation signal, worker-death reaping, result precedence, and disposal quiescence preserve the subagent run contract across it. The [agent-scope runtime-design RFC](../architecture/2026-07-12-agent-scope-runtime-design.md#workflow-children-are-pending-starts-or-published-records) owns those race algorithms.

-**Meta as data, never evaluated**: the meta block reaches the seam as a plain JSON request field (the tool's schema-validated `meta` parameter) and the engine only shape-validates it, every violation named. This is a host-isolation invariant, not a convenience: evaluating a meta literal host-side — even one contractually "pure", in an empty timed vm context — hands script-controlled getters a host stack with no timeout the moment the result is READ, defeating the exact spin isolation the worker thread buys.
+The engine exposes an in-process `MessageChannel` test path because main-process V8 coverage cannot see worker execution.

-**Value boundary**: values leaving the script (meta, hook options, schemas, the return value) go through `materializeFromRealm` — a plain recursive walk that rejects loud everything JSON cannot carry (exotic prototypes, functions, symbols, cycles, sparse arrays, non-finite numbers, nested `undefined`), copying via `Object.defineProperty` so a `"__proto__"` key becomes a data property, never a prototype mutation; getters are read ordinarily and their RESULT crosses (a throwing read fails loud) — which is also what makes every later postMessage hop total. Values entering the realm (`args`, `agent()` results, hook promises and failures, combinator arrays) are handed over directly as worker-realm values — the script is trusted, so outer prototypes are not a leak; `args` rides the `workerData` structured clone (the caller-isolation copy) and is cloned once more so a script scribbling on it cannot mutate the session's init object. Hook failures are `WorkflowError`s built OUTSIDE the script's context: the combinators recognize fatality by `instanceof` against the engine's own class (unforgeable from the script), and the script-visible consequence — in-script `instanceof Error` is `false` for hook errors; branch on `e.name`/`e.code` — is documented in the engine README. Realm functions (stages, thunks) are called, never materialized. Thrown script values are rendered by a total renderer (stack → message → `String()`, fixed label if rendering throws), so `result` cannot reject. Caps (`maxConcurrentAgents` auto = `min(16, max(1, availableParallelism() - 2))`, `maxTotalAgents` 1000, `maxItemsPerCall` 4096) and timeouts are validated Config, not literals.
+**Meta is data**: the schema-validated `meta` field reaches the seam as JSON and is only shape-validated. The host never evaluates a metadata literal, which would let script-controlled accessors run outside the worker's isolation.
+
+**Value boundary**: `materializeFromRealm` copies outbound values and rejects functions, symbols, nested `undefined`, exotic prototypes, cycles, sparse arrays, and non-finite numbers. Data-property copies make `"__proto__"` safe; getters are read normally and a throwing getter fails loudly. `args` crosses through `workerData` and is cloned again before exposure. Realm functions are invoked rather than copied, and thrown values use a total renderer so `result` cannot reject. Hook errors are host-realm `WorkflowError`s, so scripts branch on `name` or `code` rather than `instanceof Error`, as documented in the engine README. Concurrency, total-agent, item, timeout, and grace limits are validated config.

 ### The consumer (dsh-tool-workflow)

@@ -44,6 +46,10 @@ An output schema makes a schema-valid committed capture mandatory for successful

 `StructuredOutputSchema` is the raw enforceable JSON-Schema subset in `dsh-tools` (single-string `type`, `properties`/`required`/`additionalProperties`, `items`, scalar `enum`/`const`), and unsupported keywords fail loudly because that wire data becomes the capture tool's parameters verbatim. The [agent-scope runtime-design RFC](../architecture/2026-07-12-agent-scope-runtime-design.md#structured-output-commits-only-authoritative-outcomes) owns the assembly, commit, guard, and terminal-stop correctness algorithms.

+## Testing
+
+Worker-side logic runs through an in-process `MessageChannel` so V8 coverage measures it. Unit tests cover script helpers, fatal and nullable failures, JSON boundaries, caps, cancellation, child ownership, and structured output through real loops. A built-bin smoke runs the separately bundled `lib/worker.cjs` under plain Node, a with-key e2e drives real child agents, and model-facing workflow behavior is snapshot-covered through its owning example.
+
 ## Deferred (documented non-goals of this cut)

 - **Background collection** (start tool → run id → completion notice → collect), designed alongside bash/subagent background unification.
@@ -65,8 +71,8 @@ An output schema makes a schema-valid committed capture mandatory for successful
 - **`SchemaSpec` as the outputSchema type**: the author-facing DSL cannot express what arrives as data and cannot be validated against without conversion loss.
 - **A schema-object library (zod, or the repo's schemastery) for the structured-output subset**: the schema is wire data — plain JSON that crosses the vm realm boundary in `agent({schema})` and lands verbatim in the forced tool's parameters — exactly where live schema objects cannot sit; consuming raw JSON Schema at runtime would need a third-party converter on top (zod core only emits JSON Schema, not the reverse), and it would put a second schema language beside schemastery's config role.
 - **ajv for value validation**: it validates FULL JSON Schema, so the subset gate — the module's actual point, since every accepted keyword must be one the harness enforces — would remain hand-written regardless; it compiles validators through `new Function`; and it would be dsh-tools' first runtime dependency, all to replace the ~70-line value walker while the path-qualified, every-violation error reporting stays custom either way.
- **Provider JSON mode (`response_format: {type: json_object}`) instead of the forced capture tool**: the official API guarantees valid JSON, not schema-conforming JSON (no `json_schema` type; the docs' own guidance is to validate client-side, with the schema riding in the prompt), so both walkers survive untouched and only the capture-tool mechanics could go — at the cost of tools during a structured child's run (whether `response_format` composes with tool calling is undocumented), the in-turn validation retry (`ToolArgsError` keeps recovery inside the turn; a JSON-mode empty body — a documented failure mode — ends the turn, and the only recovery is the re-prompt loop this design rejects), and a new per-adapter `LlmCallConfig` surface. The accepted upgrade path is strict TOOL schemas (provider-side constrained decoding on tool parameters) when available: the same forced tool and subset gate, with the gate narrowed to the provider's strict subset.
+- **Provider JSON mode instead of the capture tool:** it guarantees valid JSON, not schema conformance, and its interaction with tool calling is unclear. The capture tool preserves in-turn validation retries. Provider-side strict tool schemas can later narrow the accepted subset without changing this design.

 ## Consequences

-The harness gains CC-compatible script orchestration: fan-out plans live in a rerunnable artifact instead of the parent context, and `outputSchema` yields an authoritative structured child result across native and Code Mode presentation. The cost, bounded by the trust premise, is a worker thread per run (~tens-of-ms spin-up), every hook crossing a message port as RPC, and a termination-path `agentsStarted` that degrades to the host-observed count; in exchange `start()` never blocks the host, a post-cancel grace ends in a real `worker.terminate()`, and the value boundary is serialization by construction. A worker thread is still not a security boundary — scripts share the model's trust level, and actual sandboxing requires an isolated-vm/separate-process engine behind the seam. The fatal-vs-null strictness divergence from CC means a CC-authored script that relies on option typos dissolving to `null` behaves differently, preserving the repo's no-accepted-then-ignored rule. Consumers must hold the run handle for control (`cancel`/`dispose`); observers get data snapshots only, so no listener can extend a run's lifetime or corrupt another's view.
+Fan-out plans now live in rerunnable scripts, and `outputSchema` provides authoritative structured child results. Each run pays worker startup and message-port RPC costs, but host startup stays non-blocking, cancellation can terminate the worker, and serialization enforces the value boundary. Worker threads are not a security boundary. Invalid options fail rather than degrading to Claude Code's `null`; consumers retain control through the run handle while observers receive snapshots only.
@@ -49,44 +49,45 @@ The `escalation-rejected` twin ends in `{"outcome": "rejected"}` instead: nothin

 #### The seam: mechanism and policy split

-After request validation and a successful `approval/asked` append, the answerer phase always resolves to a closed `ApprovalOutcome` — `allowed-once` / `rejected` / `cancelled` / `unavailable`. `ApprovalRequest` is a readonly same-process contract, so the service borrows its routing identity and cancellation signal instead of copying the record or capturing a parallel callback bundle. It dispatches the `approval/request` waterfall, races the request signal (abort settles `cancelled`; a late answer is discarded, never double-audited), contains a throwing answerer as `unavailable`, normalizes a rogue non-vocabulary return to `unavailable`, and lands the log-only audit pair `approval/asked`/`approval/decided` (paired by the branded `ApprovalRequestId`) on the request agent's session log. Request acceptance and either pre-commit audit append may still reject; returning a decision that could not be logged would violate the pair. Session owns post-commit observer containment, so a callback failure cannot turn an authoritative audit append into a rejected request or suppress the matching event. Grants are one-shot by definition: `allowed-once` authorizes the single asked-about action, never a class of future ones, and the service stores nothing between requests. `request()` also throws before appending anything when the agent's session has no open turn — the audit pair must be turn-enclosed, the turn being the durable log's commit/replay boundary (a bare event between turns is dropped as crash tail on reload); every ask path runs mid-turn already, and idle asks are a deferred design.
+After validation and an `approval/asked` append, `request()` resolves to `allowed-once`, `rejected`, `cancelled`, or `unavailable`. The service borrows the readonly request, runs the answerer waterfall, races cancellation, and normalizes thrown or invalid answers to `unavailable`. It then appends the matching `approval/decided`, paired by `ApprovalRequestId`.

-Answerers are the policy, and they are `approval/request` waterfall listeners. The waterfall buys exactly what the seam needs: with zero listeners the dispatch falls through to the caller-supplied default — `unavailable`, so fail-closed needs no configuration and no code in any deployment; a listener that recognizes the request's agent answers by returning an outcome without calling `next()` (the decision slot is single-occupancy, first answer wins — the same documented semantics as the `fs/write-intent` gate); a listener that does not recognize the agent MUST delegate via `next()` so another answerer or the default gets the question; and listeners dispose with their owning fiber, so an unloaded UI plugin degrades the next ask to `unavailable` instead of leaving a dangling channel. Registration order across sibling plugins is not load-order deterministic (the loader starts siblings concurrently), so a deployment composes ONE terminal answerer and reserves `prepend` listeners for decide-or-delegate gates.
+Both audit events must be inside an open turn; acceptance or a pre-commit append failure rejects the request. Post-commit observers are contained by the session. `allowed-once` grants only the requested action, and the service retains no grant state.

-`ApprovalRequest` carries the asking `agent` (routes the question; receives the audit events), the `toolName`, the optional exact `callId`, the asker's human-readable `reason`, and the optional `signal`. The caller retains ownership and honors the readonly contract for the duration of `request()`. The vocabulary is deliberately self-contained — it names the tool-call by the `CallId` brand from `dsh-llm` and never imports `dsh-tools` — because `dsh-tools` depends on `dsh-user-approval` (the ask routing) and a `ToolCallView` import would close a package cycle. It deliberately does NOT carry tool arguments: a UI answerer attaches the prompt to the already-streamed tool call via `callId` instead of re-rendering the call.
+Answerers are `approval/request` waterfall listeners. A listener returns an outcome for an agent it owns and calls `next()` otherwise. With no answerer, the default is `unavailable`; unloading a UI therefore fails closed without leaving a channel. Because sibling registration order is not deterministic, a deployment composes one terminal answerer and uses `prepend` only for decide-or-delegate gates.
+
+`ApprovalRequest` carries the agent, tool name, optional `callId`, reason, and signal. The agent routes both the prompt and audit events. The request uses `dsh-llm`'s `CallId` without importing `dsh-tools`, avoiding a package cycle. Tool arguments are omitted because UI answerers attach to the already-rendered call.

 #### Ask routing in dsh-tools

-`ToolRegistry.execute()` resolves an `ask` decision through the seam before the shared deny path: `allowed-once` proceeds to guards and dispatch, and the three non-grants deny with distinct reasons — "the user rejected…", "…was cancelled", "…no approval channel is available" — so the model can tell a human "no" from an absent channel. The seam is consumed opportunistically (`ctx.get('approval')`, the `tool-bash`/`agent-loop` pattern), not statically injected: with no ApprovalService, or after one unmounts, the next ask fails closed without gating the registry's fiber. An agent-less execution also fails closed — without an agent there is no session to audit to and no UI to route to.
+`ToolRegistry.execute()` sends `ask` through the approval seam before the deny path. Only `allowed-once` proceeds; rejection, cancellation, and an unavailable channel produce distinct model-visible reasons. The registry looks up the optional service per call, so an absent or unloaded service fails closed without gating the registry fiber. Agent-less execution also fails closed because it cannot be routed or audited.

 #### The per-session policy tier

-The seam also owns the session-scoped approval policy — the approval knob of the two-knob per-session switching design ([the sandbox RFC](2026-07-06-sandbox.md) § Per-session modes is the pattern's home: one log-only event per knob, a pure fold, THE write path, ACP config-option advertisement, and turn-anchoring). `ApprovalPolicy` is `'ask' | 'never'`, and `effectiveApprovalPolicy(events) ?? Config.policy` (default `'ask'`) decides every request BEFORE any interactive answerer: the service resolves a `'never'` session to `'rejected'` INSIDE `request()`, before dispatching the waterfall at all — no listener registration, including a later `prepend`, can sit ahead of it — while `'ask'` dispatches unchanged and falls through to fail-closed `'unavailable'` when nobody answers. Visibility follows the switching design's two layers with one asymmetry: the prompt section states ONLY `'never'` (deterministic, availability-independent — "you will be prompted" would overclaim in a composition with no answerer, and absence under a logged header is exactly how the narrator reads `'ask'` back), the narrator injects at most one coalesced notice per switch, and the audit pair still lands on every ask, including the policy's auto-rejections.
+The seam owns the session policy `'ask' | 'never'`, following the switching contract in the [sandbox RFC](2026-07-06-sandbox.md). The effective session or config policy is applied before answerers: `'never'` rejects inside `request()`, while `'ask'` dispatches and falls through to `unavailable` when unanswered. The prompt states only deterministic `'never'`; the narrator reports switches, and every request still receives its audit pair.

 #### The ACP answerer

-The bridge registers the first real answerer: it resolves the owning session through its existing `WeakMap<Agent, sessionId>` reverse map, issues `session/request_permission` with the request's `callId` as the `toolCall` reference and the one-shot options `allow_once`/`reject_once`, and maps the response — selected `allow-once` → `allowed-once`, any other selection → `rejected` (an unknown optionId from a non-conforming client never grants), client `cancelled` → `cancelled`. A request for a foreign agent — or one without a `callId`, since the protocol prompt must attach to a tool call — delegates via `next()`. A rejected RPC (client gone mid-prompt) propagates to the service, which contains it as `unavailable`. Whether a call ASKS at all is policy — a hook or `tools/pre-execute` plugin returning `ask` — never the bridge's own judgment.
+The ACP bridge finds the owning session, sends `session/request_permission` for the `callId`, and maps one-shot allow, reject, and cancel responses to the seam vocabulary. Unknown selections never grant. Foreign agents and requests without a `callId` delegate via `next()`; RPC failure becomes `unavailable`. The bridge answers requests but does not decide which calls require approval.

 The answerer routes through the bridge's reverse-map ownership seam described by [the ACP support RFC](../../implemented/feature/2026-06-14-acp-agent-client-protocol.md), implementing the per-session permission ownership required by [the multi-session RFC](../../implemented/feature/2026-06-14-acp-multi-session.md).

 #### Audit, and what the model sees

-`approval/asked` / `approval/decided` are log-only session events (the `hook/invoked`/`hook/result` precedent): durable, replayable, never in the model transcript. The model's entire view of an approval is the tool result the asker derives from the outcome — reconstructability holds because that result is an ordinary logged `tool/result`. One `decided` lands per `asked`, whatever the outcome, including an already-aborted signal (settled `cancelled` without dispatching), a contained answerer failure, or a session observer that throws after either event is already appended.
+`approval/asked` and `approval/decided` are durable log-only events. The model sees only the asker's logged `tool/result`. Every accepted request appends one matching decision, including cancellation and contained answerer failures.

 #### Entities and dependencies

-One package, no cycles: `dsh-user-approval` peers on `cordis`, `dsh-session` (event-map merge + append), `dsh-agent` (the `Agent` type), `dsh-llm` (`CallId`, via `dsh-brand`). `dsh-tools` and `dsh-acp` each peer on it; the escalation phase's asker lives in `dsh-tool-bash` (see [the sandbox RFC](2026-07-06-sandbox.md) § Escalation), so the sandbox family keeps its ZERO-edge relation (the executor contributes the per-call override mechanism, and transport seams never ask humans questions). The seam is one package, not the capability-seam three: the service body (dispatch + audit) has no replaceable implementation — the replaceable part is the answerer listeners, and those live with their owners (the bridge; future terminal UIs; test scripts). `@cordisjs/plugin-capability` stays orthogonal (a static grant registry answers "is this already authorized", not "ask the user now"), and `subagent-acp`'s child-side `permission` auto-answer is untouched — routing a child's approvals to the parent session is deferred (§ Deferred).
+`dsh-user-approval` owns the fixed dispatch-and-audit mechanism; `dsh-tools` asks and `dsh-acp` answers. Replaceable answerers remain listeners in their channel-owning plugins, so a three-package capability split would add an empty implementation layer. Sandbox executors remain transport-only, and static capability grants remain separate from interactive approval.

 ### Testing

-Unit tier: the service's outcome branches (fail-closed default, first-wins slot, delegation, containment, rogue-value normalization, abort-before and abort-during with late-answer discard, fresh ids, fiber-disposal degradation), scoped routing, post-append observer throws on both audit events, and the policy tier (both values × dispatch/decide, a `'never'` decision unbypassable even by an answerer prepended AFTER the service, audit pair intact) in `dsh-user-approval`; the ask routing matrix (grant dispatches; three non-grant reasons pinned verbatim; unmounted and agent-less degrades; the registry's own exhaustiveness backstop against a non-conforming stand-in) in `dsh-tools`; the answerer (wire shape of the prompt, outcome mapping, unknown-option conservatism, foreign-agent and call-less delegation) driven through a real bridge + scripted client in `dsh-acp`.
-
-Snapshot tier: the harness accepts scripted permission answers (`permissionAnswers` in a scenario's `input.json`, consumed FIFO; an unscripted prompt answers `cancelled`, fail closed). The seam's wire is recorded end to end in the sandbox example's suite: both escalation branches drive `session/request_permission` through this seam over scripted answers (grant and rejection), and the recorded `mode-switching` scenario pins the `'never'` prompt sentence and the policy-switch notice ([the sandbox RFC](2026-07-06-sandbox.md) § Testing).
+- **Unit/integration:** cover first-wins delegation, fail-closed defaults, malformed and throwing answerers, cancellation races and late-answer discard, audit pairing despite observer failures, unbypassable `'never'`, distinct tool-denial reasons, and ACP per-session routing/outcome mapping.
+- **Snapshot:** script permission answers through both sandbox escalation branches and pin the `'never'` prompt plus policy-switch notice. Hook-produced asks without a composed answerer remain covered as fail-closed denial.

 ## Deferred

 - **`allow_always` grant storage** — honoring a persistent grant means designing storage, scope identity (call? path? prefix? session? time window?), and revocation; until designed, only the one-shot options are advertised ([the sandbox RFC](2026-07-06-sandbox.md) § Escalation records the open scope question).
- **A recorded hook-driven `ask` through a composed answerer** — the human-prompt wire is recorded through the sandbox example's escalation branches. The hook matrix's `hook-cc-pretool-ask` pins the no-ApprovalService fallback denial, while the hook-producer-plus-answerer composition remains on the unit tier.
+- **A recorded hook-produced ask with a composed answerer** — escalation records the human-prompt wire, while the current hook fixture pins the no-service denial; their combined producer/answerer path remains unit-covered.
 - **Routing a child agent's approvals to the parent session** — `subagent-acp`'s child auto-answers its own `permission` requests; surfacing them to the parent's editor is its own design.

 ## Alternatives considered
@@ -100,14 +101,10 @@ Snapshot tier: the harness accepts scripted permission answers (`permissionAnswe

 ## Consequences

-The implemented contract is pinned by the suites in Testing:
-
- With an ApprovalService and an answerer composed, a hook's `ask` reaches a human and `allowed-once` dispatches the tool; every other outcome denies with its distinct reason.
- A `'never'` session auto-rejects every ask without prompting anyone, states the policy in its prompt, and narrates switches (the shared switching mechanics are pinned in [the sandbox RFC](2026-07-06-sandbox.md)).
- Every unanswerable path fails closed to `unavailable`: no service, no listener, a foreign or agent-less request, a throwing answerer, a rogue return value, or a dead client connection.
- Every `request()` routes through its readonly agent identity and lands exactly one `approval/asked`/`approval/decided` pair on that agent's log, replayable and invisible to the model transcript; post-append observer failures cannot split the pair.
- Prompts route per-session through the bridge's ownership map; one session's prompt can never reach another session's editor.
- A deployment with no ApprovalService emits no approval prompt or approval audit events and denies every `ask` request.
+- Only `allowed-once` dispatches an asked-about action; absent, rejected, cancelled, or failed answer paths deny.
+- Session ownership routes prompts, policy, and audit events without crossing editor sessions.
+- Accepted requests append one durable audit pair; the model sees only the resulting tool result.
+- A deployment without the service emits no approval prompt or audit events and denies every `ask` at the tool boundary.

 Costs and accepted limits:

@@ -117,8 +114,6 @@ Costs and accepted limits:

 ## FAQ

-Behavioral and usage questions only — every "why not X?" design question lives in [Alternatives considered](#alternatives-considered), whose job is exactly that.
-
 - **What happens in a deployment with no answerer at all (headless, CI)?** Every ask falls through the empty waterfall to `unavailable` and the tool call denies with the "no approval channel is available" reason. Fail-closed is the zero-listener default, not a configuration.
 - **Can a grant persist — "always allow this"?** No. `allowed-once` authorizes the single asked-about action and the service stores nothing between requests; `allow_always` is deliberately not advertised until grant storage is designed (§ Deferred).
 - **What does the model see of an approval?** Only the tool result the asker derives from the outcome — the audit pair never enters the transcript. The three non-grant reasons are distinct, so the model can tell a human "no" from a dismissed prompt from a missing channel.
@@ -4,7 +4,7 @@ Status: implemented

 ## Problem

-The order of the tool list a model call carries — `request/header.tools` on the session log and `GenerateOptions.tools` on the wire — was an emergent artifact: the tool registry returns schemas in registration order, the system-prompt assembly concatenates providers in registration order, and the loop logged and dispatched the result verbatim. Registration order is plugin load order, and plugin load order is a race: the cordis loader imports every `cordis.yml` entry concurrently, so which tool plugin registers first depends on module-import completion timing. The plugin dependency relation cannot rescue this — it is a partial order under which independent tool plugins (e.g. `tool-subagent` vs `tool-todo`) are incomparable, so both interleavings are legal linearizations. This stopped being theoretical when a CI runner resolved the race differently from every recording machine: snapshot goldens pinned one permutation of `request/header.tools`, the `node 22.18` CI leg produced the other, and 5/5 snapshot tests failed on a diff that was pure array reordering. Tool order is part of the request bytes (prompt-cache stability, potentially model behavior) and, since the reconstructability contract, part of the durable session log — it must be a decision, not a residue.
+Model-facing tool order followed plugin registration order, which depends on concurrent module loading for otherwise independent plugins. That race produced different request headers in CI and snapshot recordings. Because order affects request bytes, caching, and the durable header, it needs an explicit deterministic policy.

 ## Decision

@@ -17,7 +17,7 @@ The system-prompt assembly owns the canonical model-facing tool order, exactly w
 - The list must contain the rest entry exactly once and no duplicate names.
 - When `toolOrder` is unset, the canonical order is plain lexicographic name order (code-unit comparison, locale-independent), so determinism requires no configuration.

-The policy is applied where the list is born: `assemble()`, before the `system-prompt/assemble` waterfall. The assembly canonicalizes the tools it collects from providers the same way it sorts sections by their `order` field — on the initial assembly, killing the registration-order entropy at its source. The waterfall therefore starts from one deterministic list; when a listener leaves that order intact, the loop's `EpochHeader`, the `request/header` event, the deep-frozen request, and the dev invariant's cross-check inherit it with no new loop change.
+`assemble()` canonicalizes provider tools before the `system-prompt/assemble` waterfall, removing registration-order variance at its source. The waterfall starts from this deterministic list; unchanged order then flows into the request header, frozen request, and reconstruction checks without loop-specific ordering logic.

 Scope is deliberately narrow: this fixes the REGISTRATION-ORDER race, not plugin behavior. A `system-prompt/assemble` listener may still add, remove, or rearrange tools — same as it may edit sections after their sort — and owns the determinism of what it emits; the waterfall contract already demands deterministic listeners (the reconstructability invariant would catch a listener that diverges between build and replay).

@@ -38,7 +38,6 @@ Config plumbing follows the `persona` precedent, and `toolOrder` sits beside it:

 - Every registry-built assembly starts with a deterministic tool order on every host; absent an expert listener that deliberately changes it, every `request/header` event and model request inherits that order. The CI-vs-local registration-order flip is structurally gone, and the default is lexicographic.
 - The initial `PromptAssembly.tools` is canonical, so waterfall listeners start from the model-facing order; provider registration order is observable nowhere before that cooperative seam.
- The snapshot suite's single pinned request-header fixture (`text-turn`) carries the new canonical tool order; every other ACP snapshot keeps the header bulk scrubbed as `{{system}}`/`{{tools}}`, per the pinned-header design.
 - A pure tool reordering between steps is representable only as a `request/header` `'fallback'` snapshot (the name-keyed `ToolsDelta` cannot express it); with a stable canonical order such reorders no longer occur in practice, so the fallback path stays a safety valve.
 - The `toolOrder` key rides the app → `agent-core` → `SystemPrompt` forwarding chain, so deployments set it next to `persona` in the app config; `dsh-llm` and the agent loop are untouched.
 - A misspelled or unloaded tool name in `toolOrder` fails the turn at prompt assembly, not the boot: the loop assembles inside the turn (after `turn/start`, before `step/start`), so the rejection reaches the turn's outer catch — the turn closes balanced with an `error` reason carrying the message, `agent/error` mirrors it, no step opens, no `request/header` is logged, no request reaches the adapter, and the agent returns to idle. Every turn fails identically until the config is fixed; the process itself stays up (matching the repo rule that explicit config references must not be silently ignored — the enforcement point is the assembly because no earlier universal moment exists).
@@ -46,4 +45,4 @@ Config plumbing follows the `persona` precedent, and `toolOrder` sits beside it:

 ## Testing

-Unit tests on `dsh-system-prompt` pin the ordering semantics (lexicographic default, listed/rest placement, unknown-name rejection at assembly, reserved tool-name rejection, stable handling of shared names, provider-order independence), the pre-waterfall contract (listeners observe the canonical list; a listener-appended tool is not re-sorted), and each invalid-list rejection at load. Loop-level tests assert the `request/header` fold carries the canonical order for scrambled registration orders (identical across permutations), that a configured `toolOrder` reaches both the logged header and the dispatched deep-frozen request, that the frozen loop-built envelope survives to the adapter, and that an unregistered `toolOrder` name fails the turn with a balanced `error` `turn/end`, an `agent/error`, no step, no logged header, and no dispatched request. Forwarding is asserted at every level that exposes the key (`dsh-agent-core`, `dsh-stdio-agent`, `dsh-acp-agent`). The snapshot tier replays all scenarios while only the pinned `text-turn` header carries the full canonical tool list; non-pinning fixtures continue to compare through `{{tools}}`.
+System-prompt tests cover lexicographic default order, listed/rest placement, provider-order independence, shared names, invalid lists, unknown or reserved names, the canonical pre-waterfall list, and the rule that listener-added tools are not re-sorted. Loop tests pin identical logged and dispatched order across registration permutations, forwarding through agent-core and both apps, deep-frozen requests, and balanced turn failure with no step, header, or adapter call for an unknown configured name. Snapshot replay keeps the full canonical list only in the pinned `text-turn` header; other fixtures continue to use `{{tools}}`.
@@ -38,30 +38,13 @@ The swap is invisible to every consumer of `ctx.bash`: the bash tools, hook comm

 Misconfiguration fails loud: `mode` outside the closed vocabulary is rejected at plugin load, and a host with no usable backend throws the structured `SANDBOX_UNAVAILABLE` — at `confine()` before the command ever spawns — rather than degrading to unconfined execution. `runnerCommand` on `dsh-sandbox-local` is the operator's explicit assertion of a bwrap-compatible runner (chain and probes skipped); it doubles as the deterministic fake-runner seam for keyless tests.

-What the model then experiences: denied file effects come back as result facts with a `[sandbox: file access denied under <mode> mode]` marker plus standing instructions not to retry around them; under a confining executor the schema offers `sandbox_permissions` + `justification` for the one-approval escalated retry (validated strictly wider than the session's effective mode at execution); the system prompt deliberately does NOT state the sandbox mode — the model learns the boundary from the marker (which names the mode) when it hits it, instead of preemptively refusing work a standing declaration discourages. What an ACP editor experiences: one `Permissions` config-option select per session (advertised when the `dsh-permission` preset layer is composed; each preset bundles a sandbox mode and an approval policy and writes through to both knob events — a knob state outside the table derives a switch-away-only `custom` current), switchable at runtime; a sandbox switch simply changes what subsequent commands may do, while an approval-policy switch to `'never'` is stated in the prompt and narrated.
-
-The product path, concretely (the escalation arc is verbatim from the recorded `escalation-approved` scenario; the denial leg is pinned on the real-kernel e2e tier):
-
-```
-tool/result   … [sandbox: file access denied under read-only mode]   ← the write RAN; the kernel refused it
-tool/call     bash {"command": "printf 'escalated\n' > escalated.txt && cat escalated.txt",
-                    "sandbox_permissions": "workspace-write",
-                    "justification": "the user asked to write escalated.txt in the workspace"}
-  → the editor is prompted on this very call (session/request_permission through the approval seam); Allow once
-tool/result   "escalated" — THIS call ran under workspace-write and its result facts say so; the session stays read-only
-```
-
-Reject instead and nothing executes: the result is the verbatim `the user rejected escalating this command to "workspace-write"`, and the teaching makes that final — no re-ask.
+Denied file effects return a `[sandbox: file access denied under <mode> mode]` marker and instructions not to work around the denial. A confining executor adds paired `sandbox_permissions` and `justification` fields for one approved retry that must be strictly wider than the session's effective mode. A grant widens only that retry; rejection executes nothing, returns `the user rejected escalating this command to "<mode>"`, and permits no re-ask. The prompt does not announce sandbox mode, avoiding preemptive refusal. When `dsh-permission` is composed, ACP exposes one `Permissions` select whose presets write both knob events; unmatched knobs appear as switch-away-only `custom`. Only a switch to the deterministic `'never'` approval policy is stated in the prompt and narrated.

 ### Design detail

-#### Grounding — verified against the code
+#### Scope grounding

- Runtime OS subprocesses exist at exactly two sites: the `ctx.bash` seam's single spawn (`packages/bash/bash-local/src/run.ts`; hook commands flow through `ctx.bash`, so bash confinement covers them transitively) and `subagent-acp`'s child agents (`packages/subagent/subagent-acp/src/run.ts`) — the second consumer that makes a shared seam due rather than preemptive under the [capability seams RFC](../architecture/2026-06-13-capability-seams.md)'s "don't split preemptively" rule.
- Everything else executes inside the harness process (fs is in-process `node:fs`, web is in-process `fetch`, every `ToolDefinition.execute()` closes over `ctx`): an OS sandbox wraps `execve` and cannot wrap an in-process function call, so "sandbox any tool" is policy at each tool's seam, never a mechanical transport change.
- `tools/pre-execute` (`allow`/`deny`/`ask`) exists, with `ask` serviced by [the approval seam](2026-07-06-approval-seam.md); the fs intent gates are version guards with no mode input yet.
- `dsh-bash`'s request/spec split (`BashExecRequest` → `resolve()` → `BashExecSpec`) carries per-call fields the way escalation needs — `owner` is the template: request-optional, spec required-but-nullable, carried verbatim — and the result types already speak `SandboxMode`, so a per-call policy field adds no dependency edge.
- The pinned-header snapshot design means a schema/description change churns at most one pinning fixture per suite, and the escalation fields are advertised only under a sandboxing executor — so they live in exactly one pinned header, the acp example suite's `permission-switching` fixture.
+OS subprocess confinement applies to the bash executor, including hook commands, and later to ACP subagent children. Filesystem, web, and other tools execute in-process and require policy at their own seams; an argv wrapper cannot confine a function closing over `ctx`. The existing bash request/spec split carries per-call overrides, while `tools/pre-execute` and the approval seam own the human decision.

 #### The seam: `ctx.sandbox`

@@ -75,33 +58,33 @@ Left open, for the phase that needs them: whether network restriction arrives as

 #### Local backends and the shipped launcher

-`dsh-sandbox-local` selects BY PLATFORM, once per lifetime, and caches the verdict: each platform names its runner chain, a chain of one is selected directly — probing arbitrates between candidates, and a sole candidate leaves nothing to arbitrate — and a chain of several is probed FUNCTIONALLY in preference order (build and enforce a real profile, never `--version` — a present-but-unusable `bwrap` must fail its probe). Linux: `bwrap` first (its mount profile is closest to the mode vocabulary: whole tree read-only, fresh `/dev`+`/proc`, `workspace-write` adds an ephemeral `/tmp` and rebinds the workspace root; deliberately no `--unshare-pid` and no network claim), else the npm-distributed `landlock-run` Landlock launcher. darwin: `sandbox-exec` speaking a Seatbelt (SBPL) profile — allow-default with `(deny file-write*)` plus write allow-lists, every granted root canonicalized because Seatbelt matches resolved paths (`/tmp` IS `/private/tmp`) — unprobed, the sole candidate. A platform with no chain fails closed at `confine()`; an unprobed runner that turns out unusable fails closed at EXECUTION instead — it refuses to run the command, and every wrap carries `runnerFailureSignatures` (the runner's own error prefix, which also matches the shell's runner-not-found message) so the consumer classifies that as a SANDBOX failure, never a task failure: on either path the command neither runs unconfined nor slips through as a plain failure. A non-empty `runnerCommand` config is the operator's assertion of a runner that fully enforces the bwrap-shaped profile — chain and probes skipped; it doubles as the deterministic fake-runner seam for keyless tests. It is not exempt from fail-closed execution: its wrap carries argv0-scoped outer-shell failure shapes (`exec: <argv0>: not found`, `<argv0>: No such file or directory`, `<argv0>: Permission denied`) as its runner-failure dialect, so a missing or unexecutable configured runner classifies as a sandbox failure like every other rung — never as a failing command, and never as a denial.
+`dsh-sandbox-local` selects one platform runner per provider lifetime and caches the verdict. Linux functionally probes `bwrap` then Landlock; macOS uses Seatbelt. Unsupported platforms and unusable runners fail closed. Each wrap carries backend-specific denial and runner-failure signatures so `dsh-bash-sandbox` can distinguish a denied file effect from a broken sandbox. `runnerCommand` skips selection as an operator assertion of a bwrap-shaped runner, but missing or unexecutable commands still classify as sandbox failure and never run the payload unconfined.

 The launcher is a ~300-line C program (plain C11 over the raw Landlock UAPI — no libraries beyond a statically linked musl, so the audit surface is that one file plus the kernel's stable syscall contract): `--ro <path>` / `--rw <path>` grants, `--`, the wrapped argv; it installs the ruleset on itself and `exec`s (rulesets are inherited across `execve`, and it sets `no_new_privs` before restricting); `--probe` enforces a maximal ruleset in a short-lived child and exits 0 only when the kernel actually enforces; launcher failures exit 125 without exec'ing.

-The launcher lives in its own repository and reaches the harness as the npm package family [`node-addon-landlock-run`](https://www.npmjs.com/package/node-addon-landlock-run) (the per-platform-package pattern of `node-addon-require-builtin` and esbuild): an entry package — `dsh-sandbox-local`'s one runtime dependency — plus per-platform binary packages selected at install time by npm's `os`/`cpu` fields. The entry package owns the launcher's CLI contract end to end (`launcherPath()` resolution with a never-existing fallback, the functional `probe()`, `grantArgs()` flag spelling), versioned together with the binary so probe-report parsing can never drift against it; the harness keeps only the policy side, `landlockProfileArgs()` mapping the mode vocabulary to grants. Native-only per-architecture builds, pack gates (binary presence, executability, ELF architecture), and the byte-pinned publish rehearsal are that repository's release pipeline; this repo's Landlock CI legs install the published family from the registry — the true consumer path — and prove real-kernel confinement through it.
+The Landlock launcher ships through [`node-addon-landlock-run`](https://www.npmjs.com/package/node-addon-landlock-run), with platform binaries selected by npm. That package owns path resolution, probing, and CLI flags; the harness maps sandbox modes to grants. Versioning the entry point with its binaries keeps probe parsing and launch syntax aligned.

 FIXME: Revisit the separate-repository boundary and try to maintain the launcher source and its platform package family inside this monorepo, so the native release surface and harness contract evolve together.

-Profile parity is honest rather than identical: under Landlock, `read-only` grants `--ro /` plus `--rw /dev/null` (the node, not `/dev` — the host's `/dev/shm` is a persistent shared tmpfs), and `workspace-write` grants the HOST `/tmp` where bwrap's is ephemeral; under Seatbelt, `read-only` likewise grants only the `/dev/null` literal, and `workspace-write` grants the host `/tmp` plus the per-user darwin temp dir (`os.tmpdir()` — the platform's real temp area for mkstemp-family tools; omitting it would deny what the mode promises). Every wrap carries the rung's denial dialect (`denialSignatures`: EROFS text under bwrap, EACCES under Landlock, EPERM under Seatbelt) so consumers match the active backend rather than a cross-runner union. Enforcement is honest per ABI level: an older kernel enforces the subset its ABI governs (path truncate is ungoverned before ABI v3), the probe's report line distinguishes the cases, and every confined result carries the structured `enforcement: 'full' | 'partial'` fact — refusing partial enforcement would deny the fallback to precisely the older-kernel hosts that need it. The bwrap and Seatbelt profiles govern every promised file effect by construction, so their passing probes always report `full`.
+Backend profiles share the mode contract but differ in necessary host grants. Landlock and Seatbelt allow only `/dev/null` in read-only mode; workspace-write also permits their required host temp roots. Each wrap carries backend-specific denial signatures. Landlock reports partial enforcement on older ABIs that cannot govern every operation, while successful bwrap and Seatbelt profiles report full enforcement.

 #### The bash consumer

-`dsh-bash-sandbox` extends `LocalBashExecutor` (spawn mechanics, process-group kills, spill files, background tasks, credential scrub inherited verbatim) and hands `ctx.sandbox` the exact `['bash', '-c', command]` argv it is about to spawn. A sandbox denial is a RESULT FACT, not an error: the command RAN and the kernel refused a file operation, so `result.sandbox.denied` is orthogonal to `exitCode`/`signal`. Classification is conservative text inference over the collected stderr tail against the WRAP's own dialect, so a backend is never credited with a denial text its kernel does not speak (bare EPERM under a Linux runner names non-file boundaries the mode vocabulary does not govern); the known residual false positive is non-sandbox text in the active dialect (an ssh auth failure under Landlock, a refused `kill` under Seatbelt), and a structured runner signal wins once one exists. A RUNNER failure is the opposite of a denial and outranks it in classification (a runner's error text can itself contain denial words): the wrap's `runnerFailureSignatures` matching a failed run means the sandbox broke and the command NEVER RAN — the foreground path re-throws it as the structured `SANDBOX_UNAVAILABLE` error (the late twin of the confine-time throw, carrying the runner's first stderr line), a settled background task stamps `sandbox.runnerFailed` and `bash_output` renders its own marker — so a broken sandbox can never read as a failing command.
+`dsh-bash-sandbox` reuses local process execution and asks `ctx.sandbox` to wrap the exact bash argv. A kernel denial is a result fact independent of exit status and is inferred only from the selected wrap's stderr dialect. Runner failure outranks denial because it means the command never ran: foreground calls throw `SANDBOX_UNAVAILABLE`, while settled background tasks set `sandbox.runnerFailed` for `bash_output`. This keeps broken confinement distinct from both task failure and an enforced denial.

 The model's view is result facts only: the static tool description explains the denial marker (`[sandbox: file access denied under <mode> mode]`), encourages attempting commands that may be denied, and forbids retrying around a denial; when the escalation fields are advertised, a denied result additionally carries the escalation hint itself, so the sanctioned same-turn retry is prompted at the decision point rather than depending on the model recalling the description (§ Escalation). No prompt section states the sandbox mode (§ Per-session modes).

 #### Escalation: one approved wider retry after a denial

-The seam level is mechanism only. `BashExecRequest` carries `sandboxMode?: SandboxMode`, an explicit per-call policy input; `BashExecSpec` carries it required-but-nullable (the `owner` pattern: a forgotten field is a visible `undefined`, and `resolve()` is the one explicit defaulting step); `BashExecutor` exposes the capability fact `get sandboxMode(): SandboxMode | undefined` — `undefined` in the base class, the configured mode in `SandboxBashExecutor` — so the tool layer can advertise only what the mounted executor honors: composition truth, not configuration. The seam honors ANY explicit mode, including a narrower one; the wider-only ladder is escalation policy and lives in the tool. A non-sandboxing executor (`dsh-bash-local`) carries the field verbatim and confines nothing — the field reaching it means the caller bypassed the tool's gate, and its honest behavior stays unconfined execution, not a guess at enforcement it does not have.
+`BashExecRequest.sandboxMode` is an optional per-call input; resolved specs make the field explicit. `BashExecutor.sandboxMode` advertises whether the mounted executor can honor it, so only a confining composition exposes escalation. The seam accepts any explicit mode; the tool owns the wider-only escalation rule. Non-sandboxing executors remain honestly unconfined.

 `SandboxBashExecutor.resolve()` stamps the effective mode — escalation grant > session override > configured default — so `run()`/`start()` read the spec, never the config. The `danger-full-access` branch, the confine call, and the result facts all key off the spec's mode, and the per-task facts map carries each task's mode alongside its wrap facts (`notifyTaskDone()` stamps from the map entry): one escalated call — foreground or background — reports the mode it ACTUALLY ran under while every neighbor keeps its own.

-The tool gate advertises two extra parameters exactly when `ctx.bash.sandboxMode` reports a confining mode at registration: `sandbox_permissions`, an enum of the closed escalation-target vocabulary — `workspace-write`/`danger-full-access`, every mode a session could ever escalate TO — and `justification`, required together with it. The enum is deliberately NOT cut down to the modes wider than the executor's DEFAULT: schemas are registry-global while the effective mode is per-session and switchable, so a default-relative ladder strands a session overridden NARROWER than the default (with a `danger-full-access` default and a `read-only` override it would advertise nothing at all — confined, but with no lever). Strict widening is instead enforced at EXECUTION against the call's effective mode (session override ?? executor default): a request that is not strictly wider fails closed with its own text and prompts no one. An escalating call resolves approval BEFORE anything executes — no `ctx.approval` composed, or no agent on the execution, fails closed with its own text; otherwise `ctx.approval.request({ agent, toolName: 'bash', callId, reason, signal })` with the audit-self-contained reason `escalate sandbox to ${mode}: ${justification}`, while the UI attaches the prompt to the already-streamed call (the command is visible there; the approval RFC's no-arguments rule holds). The four outcomes map to distinct results: `allowed-once` stamps `sandboxMode` onto the bash request and proceeds; `rejected`, `cancelled`, and `unavailable` each produce their own error text, so the model can tell a human "no" from a dismissed prompt from a missing channel. The grant is consumed by the very call that asked; nothing is stored.
+When a confining executor is mounted, `bash` advertises paired `sandbox_permissions` and `justification` fields. The schema exposes the full closed escalation vocabulary because effective mode is per-session; execution rejects any target that is not strictly wider than that call's effective mode. Approval resolves before execution. `allowed-once` stamps the granted mode onto only that request, while `rejected`, `cancelled`, `unavailable`, a missing approval service, or a missing agent all fail closed with distinct results. No grant is persisted.

-The tool description teaches — and a denied result itself prompts — the SAME-TURN flow when the fields exist: on a denial a wider mode would cure, escalate immediately in that turn by retrying the exact command once with `sandbox_permissions` (the narrowest mode that suffices) + `justification`, without detouring through chat to ask first — the approval prompt raised by the retry IS how the user consents. Never speculatively: an escalation is grounded in a real denial — normally the one the command just hit, up front only when the session already denied the same access — and a prompt stating approvals are disabled turns the exception off entirely; a rejected escalation is final for that command. Denial-grounding is deliberately model discipline plus human judgment, not harness bookkeeping — the human sees the exact command and justification on the prompt (see Alternatives for why hard-matching is rejected). No new session events anywhere: the attempt is an ordinary `tool/call` whose logged arguments carry the two fields, the decision is the approval seam's `approval/asked`/`approval/decided` pair, the outcome is an ordinary `tool/result` whose sandbox facts name the mode it ran under. The asker lives in `dsh-tool-bash`, NOT the executor: a transport seam has no `agent`, no `callId`, and no business asking humans questions.
+Escalation is a same-turn retry of the denied command with the narrowest sufficient `sandbox_permissions` and a `justification`; the approval prompt is the consent step. It must be grounded in an actual denial, except when the session already observed the same denied access, and a disabled or rejected approval ends that command. The retry, approval decision, and result use existing tool and approval events. `dsh-tool-bash` owns the ask because the executor seam has neither the agent nor call id required for user interaction.

-Left open, recorded for the phase that picks them up: what a grant's scope identity is beyond the sandbox mode — the exact call, a path, a command prefix, the session, a time window — the question `allow_always` grant storage must answer before that option can be advertised; how cancellation behaves while an approval prompt is pending; and how escalation is defined for `run_in_background` denials that arrive via `bash_output`.
+Left open, recorded for the phase that picks them up: what a grant's scope identity is beyond the sandbox mode — the exact call, a path, a command prefix, the session, a time window — the question `allow_always` grant storage must answer before that option can be advertised; and how escalation is defined for `run_in_background` denials that arrive via `bash_output`.

 #### Per-session modes: the session log as the store

@@ -122,11 +105,11 @@ interface SessionEventMap {

 Each owner exports the same three-piece kit: the event declaration, a pure fold (`effectiveSandboxMode(events)` / `effectiveApprovalPolicy(events)` — a `findLast`, typed to the domain's closed union), and THE write path (`setSandboxMode(session, mode)` / `setApprovalPolicy(session, policy)` — a switch IS its event; nothing mutates state out of band). No shared owner service, no generic facts map, no registry: a third knob copies the ~40-line pattern into its own package. Execution follows the fold on both sides — the bash tool's per-call stamp reads it as the middle rung of the § Escalation precedence chain, and the approval seam's `'never'` gate is [the approval RFC](2026-07-06-approval-seam.md)'s side of the same pattern.

-**Visibility is deliberately asymmetric between the knobs.** The SANDBOX mode is stated nowhere and its switches are not narrated: a standing "you are read-only" declaration teaches the model to refuse preemptively (observed live: sessions where the model would not even attempt a write it could have escalated), while the denial marker already names the mode the command ran under at exactly the moment the boundary matters — behavior, not belief, carries the state, and a switch simply changes what the next command does. The APPROVAL policy keeps both layers, because its failure mode is the opposite: an auto-rejected ask under `'never'` returns "the user rejected …" wording no behavior can disambiguate, so the prompt states `'never'` (and ONLY `'never'` — an `'ask'` promise is unknowable without asking, and absence under a logged header is how the narrator reads `'ask'` back), and an `agent/pre-step` narrator injects at most one coalesced notice per policy switch: idle flip-flops collapse to one notice at the next turn's first step, a net-zero round trip to none, and a mid-turn change is narrated no later than the next step. Its "last told" is in-memory with a log-derived fallback (the folded header's system text parsed against the closed candidate sentence; LAST occurrence wins, so a persona quoting it cannot shadow the real section), so restarts lose nothing; attribution is positional (a knob event after the log's last `request/header*` reads `changed by the user`, a drift with no such event reads `changed by the operator/config`).
+Sandbox mode is not narrated in the prompt; denial results report the mode when it matters, avoiding preemptive refusal based on a standing label. Approval policy is different: only `'never'` is stated because automatic rejection otherwise looks like a user decision. Policy-change notices are coalesced and delivered by the next pre-step, with log-derived fallback after restart. The notice source is inferred from event position: a knob event after the last request header is user-driven; unlogged drift is operator or config driven.

 **The editor surface** is protocol-native [Session Config Options](https://agentclientprotocol.com/protocol/session-config-options) — the spec's replacement for session modes (slated for removal in ACP v2), already SDK-typed. When `ctx.permission` is composed, the bridge advertises one `permission` select (category `mode`) in `session/new` and `session/load`; its options are the deployment's preset table, and its `currentValue` is `PermissionService.current()` over the session log plus composition defaults. The shipped `workspace-write` and `danger-full-access` presets each bundle a sandbox mode with an approval policy and write through to both domain setters; a knob combination outside the table is reported as switch-away-only `custom`. `session/set_config_option` validates and switches through the permission service, then returns the complete refreshed state (the spec contract).

-**Anchoring: turn-enclosure is the commit boundary.** The turn-enclosure contract makes a bare between-turns append invalid (the JSONL backend treats a post-`turn/end` tail as crash garbage; dev invariants throw). A switch while a turn is open appends immediately — openness read from the LOG (last boundary event is `turn/start`), not `agent.status`, which stays `running` between queued turns. An idle switch is held on the bridge's session record and anchored at the next turn's `agent/prompt-submit` — inside the turn, before anything in it assembles or executes, last write per knob, and OUTSIDE any `session/event` emit (appending from inside that feed reorders events for later-registered listeners — a bug the dev invariants caught live). Until anchored, the switch exists only in bridge memory: responses overlay it truthfully, and a crash before the next turn reverts it — `session/load` then reports the fold's truth, so the editor UI self-corrects rather than lies.
+**Turn enclosure is the commit boundary.** A switch during an open turn appends immediately. An idle switch remains pending on the bridge record and is appended at the next prompt submission, before assembly or execution; last write wins per knob. Openness comes from log boundaries rather than `agent.status`, and setters do not append from inside a `session/event` listener because that would reorder later observers. Until anchoring, responses overlay the pending value. A crash discards it, and reload returns the durable fold.

 #### In-process tools

@@ -136,10 +119,10 @@ FIXME: Revisit this tool-local boundary. The follow-up design needs to determine

 ### Testing

- Unit tier (no real runner anywhere): profile dialects, per-platform chain selection (sole candidate unprobed, no chain fails closed, multi-candidate probe order), verdict caching, the fail-closed end, probe-report parsing, and the launcher/`sandbox-exec` CLI contracts via fake runner scripts in `dsh-sandbox-local`; wrapping, policy hand-off, fact stamping, and runner-failure-outranks-denial classification (foreground throw, background `runnerFailed` fact) against a fake provider in `dsh-bash-sandbox`; the error's structured identity in `dsh-sandbox`. The escalation matrix spans the three bash packages: verbatim carry-through in `dsh-bash-local`, stamp/branch/per-task-facts in `dsh-bash-sandbox`, and the capability gate, `justification` pairing, fail-closed texts (pinned verbatim), and grant stamping in `dsh-tool-bash`. The switching surface pins the folds, the stamping precedence, the `'never'` gate, per-session section rendering, the full narrator matrix (cold start, coalescing, net-zero, resume drift with operator wording, positional attribution, persona-shadow hardening), and the bridge's advertisement gating, validation rejections, idle-vs-mid-turn anchoring (dev invariants mounted), and `session/load` reporting over a real two-process JSONL round trip.
- Keyless real-runner e2e, split along the seam and per rung: CI's `sandbox-e2e` matrix runs bwrap and Landlock on Linux (the Landlock leg once per architecture, each confining through the registry-installed launcher) and Seatbelt on macOS against real kernels, failing on a silent all-skip. World-proofs live in `dsh-sandbox-local` (denied writes absent on disk, workspace writes landing, temp-area grants pinned, kernel denial text matching the advertised dialect) and `dsh-bash-sandbox` (the through-`ctx.bash` consumer proofs, including denied-then-overridden-write-lands). This package's own publish path is rehearsed without publishing (`packed-install.e2e.ts`): `pnpm pack`, tarballs installed into a throwaway consumer with the launcher family resolving from the registry, plain `node` confining through the INSTALLED launcher — asserted executable apart, so a mode-stripped binary can never masquerade as a non-enforcing kernel. The switching surface has its own keyless e2e (the acp-agent example's `escalation.e2e.ts`): the real default `cordis.yml` tree advertises the one permission option, honors switches end to end, and rejects out-of-vocabulary values.
- With-key e2e (`examples/acp-agent/tests/escalation.e2e.ts`): real model + real runner + the REAL bridge answerer, world-verified — denied under `read-only`, escalates with justification, the scripted editor grants and the retried write lands on disk, while a rejected escalation leaves no write. Self-skips without `DEEPSEEK_API_KEY` or a usable runner (e2e.yml installs bubblewrap so it actually executes in CI).
- Snapshot tier (`examples/acp-agent/tests/acp.snapshot.ts`): the keyless config-option wire; the recorded permission-switching arc as the pinned header of its class — necessarily, since mid-session switches emit the `request/header-delta`s the uniformity guard licenses only in the pin — committing one `workspace-write`→`danger-full-access` preset switch (the `permission/preset` event written through to both knobs), the approval prompt-section delta and its "changed by the user" notice; and both recorded escalation branches over scripted `permissionAnswers` (grant runs under the granted `danger-full-access`; rejection executes nothing and pins the fail-closed text). Snapshot mode starts the shared example tree at `danger-full-access` so established fixtures remain runner-independent; the switching and escalation inputs explicitly select `workspace-write` before exercising the policy path. Deliberately absent: a fixture carrying a real DENIAL — denial stderr is the backend's dialect and would pin a fixture to its recording platform; the escalation prompts assert the prior denial instead, and the denial→marker path stays on the real-runner tiers above.
+- **Unit:** pin platform selection and profiles, fail-closed runner classification, per-call facts, escalation validation and outcomes, permission preset folding and write-through, narrator coalescing, ACP advertisement and validation, and turn-enclosed config writes.
+- **Keyless real-runner:** exercise bwrap, Landlock, and Seatbelt against real filesystem effects at provider and bash-consumer layers; packed-install coverage proves the registry launcher remains executable. The real ACP composition pins permission switching and rejects unknown presets. CI rejects a silent all-skip.
+- **With-key:** drive a real model, runner, bridge answerer, and disk effect through granted and rejected escalation; unavailable credentials or runners self-skip.
+- **Snapshot:** pin the permission config-option wire, preset and knob events, prompt deltas and notices, and both scripted approval branches. Snapshot mode starts unconfined so unrelated fixtures remain platform-independent; policy scenarios switch explicitly. Real denial stderr stays on platform tests because its dialect is runner-specific.

 ## Deferred phases

@@ -198,15 +181,13 @@ Costs and accepted limits:
 - **The model may over-ask.** Escalating without denial grounding, or picking `danger-full-access` where `workspace-write` suffices: the description steers and the enum forces the ladder, but the human prompt is the actual gate; the `approval/asked` reasons make over-asking auditable, and a `prepend` policy answerer can auto-reject patterns a deployment never wants.
 - **The advertised target set is static while the effective mode is per-session** (schemas are registry-global) — a session already at the widest mode is still offered the fields. Harmless by construction: the strict-wider check at execution, not the enum, is the safety boundary — a non-widening request fails with its own text and never prompts anyone.
 - **A granted escalation is not a working sandbox.** An unavailable backend still fails closed even for a granted escalation to a confining mode — at `confine()` when the platform has no chain or every probe fails, at execution when an unprobed sole runner refuses (classified as a sandbox failure, not a command failure) — while a granted `danger-full-access` run never touches the provider at all: there the grant, not the probe, is the authority.
- **An idle switch lives in bridge memory until the next turn anchors it.** A crash in that window reverts it (reported honestly on `session/load`), and a session that never runs another turn never persists it — accepted, with the loop-owned idle commit turn named as future work if durability becomes a requirement.
+- **An idle switch lives in bridge memory until the next prompt submission anchors it.** A crash in that window reverts it (reported on `session/load`), and a session that never submits another prompt never persists it — accepted, with a loop-owned idle commit turn left as future work if durability becomes required.
 - **The approval narrator's restart baseline parses prompt prose.** The closed candidate sentence is owned by the writing module itself, so a wording change is a coordinated writer+parser edit in one file; a session whose headers predate the section silently adopts the current policy without a notice.
 - **The approval section is still a dynamic prompt surface** (a `'never'` switch breaks provider prompt-prefix caching for that session). Accepted: policy switches are rare, and a model acting on a stale `'never'` is worse. The sandbox knob no longer touches the prompt at all.
 - **The model may hold a stale belief about the sandbox mode** (nothing announces a switch). Accepted deliberately: the next attempt's marker or success corrects it, and the observed failure mode of announcing — preemptive refusal — is worse than one wasted retry.

 ## FAQ

-Behavioral and usage questions only — every "why not X?" design question lives in [Alternatives considered](#alternatives-considered), whose job is exactly that.
-
 - **A command came back with `[sandbox: file access denied under read-only mode]` — did it fail?** It RAN, and the kernel refused a file effect: the denial is a result fact orthogonal to exit code. The teaching forbids retrying around it; the one sanctioned move is the same command retried once with an escalation request.
 - **How is a BROKEN sandbox told apart from a failing command?** Runner failure outranks denial in classification: a failed run matching the wrap's `runnerFailureSignatures` means the command NEVER ran — foreground re-throws the structured `SANDBOX_UNAVAILABLE` with the runner's stderr line, a background task stamps `sandbox.runnerFailed` and renders its own marker. A broken sandbox can never read as a failing command, and the command never runs unconfined.
 - **What happens on a platform with no backend — Windows today?** `confine()` throws the fail-closed `SANDBOX_UNAVAILABLE` and the command never spawns; `win32` is a reserved EMPTY chain, pinned by test to fail closed identically until a Windows runner fills it (§ Deferred phases).
@@ -214,7 +195,7 @@ Behavioral and usage questions only — every "why not X?" design question lives
 - **Does the sandbox restrict network or process visibility?** No — `SandboxMode` claims FILE effects only; the bwrap profile deliberately does not unshare pid, and no backend claims network. Whether network restriction becomes its own knob is left open in § The seam.
 - **Which tools actually run confined?** OS subprocesses through `ctx.bash` — the bash tools, and hook commands transitively. fs/web/todo execute in-process, where an `execve` wrapper is mechanically meaningless; their `read-only` semantics arrive with the cross-family deferred phase, and until then the contract says bash-only honestly.
 - **Does a granted escalation persist, or cover background tasks?** Neither: the grant is consumed by the very call that asked (foreground or background), that one call reports the mode it actually ran under, and every neighbor keeps its own. How escalation should be DEFINED for a background denial that only surfaces later via `bash_output` is left open in § Escalation.
- **When does an editor's mode switch take effect?** Mid-turn: appended immediately, honored by the very next call's stamp. Idle: held on the bridge's session record, anchored at the next turn's `agent/prompt-submit`, with N flips coalescing to at most one event (none if net-zero); a crash before anchoring reverts it and `session/load` reports the truth. The model is not told — its next command simply behaves under the new mode.
+- **When does an editor's mode switch take effect?** Mid-turn: appended immediately, honored by the very next call's stamp. Idle: held on the bridge's session record, anchored at the next `agent/prompt-submit` inside its open turn, with N flips coalescing to at most one event (none if net-zero); a crash before anchoring reverts it and `session/load` reports the truth. The model is not told — its next command simply behaves under the new mode.
 - **What survives a restart — and what if the operator changed the config default while the process was down?** Overrides replay from the session log (`effective = fold ?? config`), so a resumed session keeps its modes with zero catch-up machinery; a default that drifted offline changes behavior the same way a switch does (the approval policy, being stated, is additionally narrated with operator/config attribution).
 - **What does `enforcement: 'partial'` on a result mean?** The selected backend enforces the subset its kernel ABI governs — e.g. Landlock before ABI v3 does not govern path truncate — and says so structurally instead of refusing the host; the probe's report line distinguishes the cases. The bwrap and Seatbelt profiles govern every promised file effect by construction, so they always report `full`.

@@ -22,7 +22,7 @@ Because composition runs before the boundary snapshot, a composing listener's se

 ## Testing

-**Unit** — [interception.spec.ts](../../../../packages/core/agent-loop/tests/interception.spec.ts) pins compose-once across turns and steps (one composition, zero `request/header-delta`s), canonical prepend ordering, empty-prefix omission from the header, the frozen seed (in-place push throws), held-reference mutation immunity, and composition-precedes-pre-step with the seam receiving the composed value; [cancel.spec.ts](../../../../packages/core/agent-loop/tests/cancel.spec.ts) pins cancel/dispose landing inside the composition window and the discard-and-recompose stale-cache guard; dsh-session codec tests cover the `messagePrefix` fold/diff/apply arms (empty ≡ absent); dsh-invariants tests pin the `messagePrefix + derivation` equation; dsh-compact-basic tests pin that the pressure estimate counts the handed prefix. **Snapshot** — the acp-snapshot normalizer scrubs header prefixes to count-preserving `{{messagePrefix}}` tokens (unit-covered in dsh-acp-snapshot); header content itself is pinned per [the pinned-header scenario RFC](../testing/2026-07-06-pin-request-header-content-in-one-scenario.md), and the example tree loads no prefix contributor, so live goldens stay prefix-free. **e2e** — none prefix-specific: the seam is provider-independent and deterministic; the with-key cache measurement in [request-cache.e2e.ts](../../../../packages/core/agent-loop/tests/request-cache.e2e.ts) already proves the cacheable-prefix economics the design rests on.
+[Interception tests](../../../../packages/core/agent-loop/tests/interception.spec.ts) pin compose-once reuse with no header deltas, prepend order, empty-prefix omission, immutability, and composition before pre-step; [cancellation tests](../../../../packages/core/agent-loop/tests/cancel.spec.ts) pin discard and recomposition. Session codec, invariant, and compaction tests cover header round trips, request reconstruction, and prefix-aware pressure accounting. Snapshot normalization preserves prefix counts, while the [pinned-header scenario](../testing/2026-07-06-pin-request-header-content-in-one-scenario.md) owns content and the default example remains prefix-free. No prefix-specific e2e is needed because the seam is deterministic and provider-independent; the with-key [request-cache e2e](../../../../packages/core/agent-loop/tests/request-cache.e2e.ts) covers its cache economics.

 ## Alternatives considered

@@ -10,7 +10,7 @@ The harness already has every seam the pi extension uses, and better ones: [the

 ## Decision

-The guard is a loop-hygiene plugin, not a model-facing tool: it never appears in the tool list, never vetoes or rewrites a call, and adds exactly one behavior — it watches each agent's stream of tool calls, counts runs of consecutive calls to the same tool with identical canonicalized arguments, and at configured run lengths injects an escalating advisory reminder telling the model to stop repeating itself, re-read the last result, and either change approach or conclude. The purpose is to break unproductive loops within a few wasted steps instead of letting them run to the turn's natural end — while leaving the decision (retry differently, gather more evidence, or finish) entirely with the model, so a legitimately repeated call is delayed by nothing and blocked by nothing.
+The guard is a loop-hygiene plugin, not a model-facing tool. It counts consecutive calls to the same tool with identical canonical arguments and injects advisory reminders at configured thresholds. It never delays, blocks, or rewrites a call; the model decides whether to retry differently or finish.

 The plugin is `@deepseek-ai/dsh-repeat-tool-guard` at `packages/guard/repeat-tool-guard/`, opening the `guard/` group for loop-hygiene plugins (single-package groups have precedent: [the todo-write RFC](2026-06-29-todo-write-tool.md) shipped `todo/tool-todo`). It registers three listeners and holds all state in plugin-local maps keyed by `AgentId` — the tool registry is a context-level singleton whose waterfalls interleave every agent's calls (subagents run on the same context), so per-agent keying is correctness, not polish.

@@ -29,7 +29,7 @@ Two deliberate rules, both documented in [the package README](../../../../packag

 ### Reminder delivery

-Reminders ride `additionalContext` (source `{kind: 'plugin', plugin: 'repeat-tool-guard'}` — the label is load-bearing per `HookContext`), never a `content` replacement: the `tool/result` event stays the tool's own output for audit, and the loop appends buffered context as `context/message`(s) after the step's results, which the session renders as the tagged synthetic-user envelope and derived history replays. Thresholds escalate: the first configured threshold gets a short "you are repeating yourself, analyze the previous result" nudge; each later threshold gets the detailed form naming the tool, the repeat count, and the canonical arguments (head-truncated at `argumentsPreviewChars`, default 500 — a looping `write`-sized payload must not ride into the next request unbounded; the chain key always compares the full canonical string), and stating that the calls made no progress. The pi original hardcodes the gentle text to the literal count 3; the guard keys it to `thresholds[0]`, fixing that bug in the port. When the downstream decision already carries `additionalContext` (a hook bridge on the same call), the guard concatenates content under its own `source` — a `HookContext` holds one `MessageSource`, and `source.kind` is what framing depends on.
+Reminders use `additionalContext` with the plugin source, preserving the original `tool/result`. The first threshold emits a short nudge; later thresholds include the tool, count, and a bounded argument preview while comparison still uses the full canonical string. Existing downstream context is concatenated under the guard's source because `HookContext` supports one source.

 ### Config

@@ -47,7 +47,9 @@ Reminders ride `additionalContext` (source `{kind: 'plugin', plugin: 'repeat-too

 ## Testing

-**Unit** — the suite drives a real agent loop against a scripted mock adapter (no network) and covers, at per-file 100%: counting/reset semantics (identical, different-tracked, untracked-transparent, prompt-submit reset, disposal cleanup, per-agent isolation), canonicalization (deep key-order insensitivity), threshold escalation including the `thresholds[0]` gentle-text rule, denied-call counting, no-agent transparency, wildcard escaping, config fail-loud cases, and both fold-onto-downstream paths (block and accept-with-replacement). **Snapshot** — the `repeat-tool-guard` scenario in the acp-agent example suite scripts five identical `todo_write` calls and pins both reminder tiers (gentle at the third, detailed at the fifth) as `context/message`s in the ACP transcript and the session log; the guard is loaded in the example's live tree (`cordis.yml`), inert for every other scenario (none repeats a call three times). The scenario is authored keyless (like `error-finish`/`cancel`): deterministically forcing a live model to repeat one call three times is not a stable recording. **e2e** — none: the plugin is provider-independent and deterministic, and the seam contracts it relies on are e2e-covered by their owners.
+- **Unit:** A real loop with a scripted adapter covers counting and reset rules, untracked transparency, disposal cleanup, per-agent isolation, canonical argument key order, escalation, denied calls, no-agent execution, wildcard escaping, invalid config, and downstream block or replacement decisions at per-file 100% coverage.
+- **Snapshot:** The keyless `repeat-tool-guard` scenario makes five identical `todo_write` calls and pins the gentle third-call and detailed fifth-call reminders in both ACP output and the session log. The plugin is loaded in the live example but remains inert in other scenarios.
+- **E2e:** None; the plugin is deterministic and provider-independent, and its seam contracts are covered by their owners.

 ## Alternatives considered

@@ -64,7 +66,6 @@ Reminders ride `additionalContext` (source `{kind: 'plugin', plugin: 'repeat-too
 - The reminder is advisory by design: idempotent polling patterns that repeat identical calls on purpose still receive nudges past the thresholds, and the pressure valves are config (`thresholds`, `exclude`) plus reminder text that explicitly allows finishing when enough evidence has been gathered. Each trigger costs reminder tokens on the next request; thresholds bound the frequency.
 - Chain state is in-memory only: a session resumed from persistence starts with a fresh chain, so a loop spanning a resume draws its reminders later than a live one — accepted, the guard is a heuristic nudge, not a logged invariant, and persisting counter state would buy little for real complexity.
 - When multiple post-execute producers attach context on one call, the fold concatenates under the guard's `source`; ordering between plugins follows listener registration order. The seam cannot represent mixed provenance — a limit inherited from `HookContext`, not owned by this plugin.
- Implementing the snapshot tier surfaced a hidden assumption in the suite kit: the fixture guard equated "authored model scenario" with "override-driven". The `Scenario` table now carries an explicit `overridden` flag, and the sidecar's presence is checked BOTH ways against it (an unregistered stray sidecar would silently replace the derived script) — the suite kit is stricter than it was before this plugin existed.

 ## Deferred

@@ -12,7 +12,7 @@ First, model-written registration must be validated where it happens: a malforme

 The toolset ships as [`@deepseek-ai/dsh-tool-cordis`](../../../../packages/cordis/tool-cordis/README.md) — a new top-level `packages/cordis/` group — and is demoed by [`examples/cordis-agent`](../../../../examples/cordis-agent/README.md). It gives the model three tools over the live cordis runtime it is running inside: inspect it, mount model-written plugins into it, dispose them again.

-The trust stance, stated once and threaded through the rest: the `node:vm` sandbox isolates the global context only — it prevents accidental global pollution, not malice — and the `ctx` a mounted plugin's `apply` receives is a whitelist façade that narrows the *surface* (framework internals withheld) but not the *privilege* of what it exposes. The verbs the façade does expose reach the real runtime: a mounted tool can shell out through `ctx.bash`, read the filesystem through `ctx.fs`, reach the network through `ctx.web`. Neither the sandbox nor the façade is a security boundary; handing the model this power is the point of the toolset. A deployment loads this plugin exactly as deliberately as it grants a bash tool — an opt-in capability in the app's `cordis.yml`, never a product default.
+The vm isolates accidental global pollution, and the context façade hides framework internals. Neither restricts the authority of exposed services: a mount can call `ctx.bash` to run commands with the host executor's privileges and can reach the real filesystem and web services. This is an opt-in development tool with bash-equivalent trust, not a security boundary or product default.

 ### The three tools

@@ -26,17 +26,17 @@ The trust stance, stated once and threaded through the rest: the `node:vm` sandb

 ### Sandbox semantics

-Mount code runs via `vm.createContext` + `runInContext`, wrapped as the body of an async function under a per-mount filename (`cordis-mount-<id>.js`). The vm gives the code a fresh realm: writes to `globalThis` stay inside the sandbox, and no Node API is handed in — capability access is *steered* toward the cordis services (`ctx.fs` for files, `ctx.web` for HTTP, `ctx.bash` for processes, the `ctx.timer` helpers for timing) rather than Node built-ins, so a well-behaved mount stays inspectable through `cordis_inspect` and disposable with its fiber. This is steering, not containment: consistent with the trust stance above, the small global surface keeps *honest* code on the cordis services but is not a security boundary — the host-realm helpers it exposes (`harness`, `console`, `btoa`) are reachable functions, so mount code that goes looking (through such a helper's `.constructor`, say) can still reach the host realm and Node itself, which is accepted because the `ctx` a mount ultimately receives is fully privileged anyway. The `vmTimeoutMs` config bounds only the synchronous portion of evaluation; an async body escapes the bound (also acceptable under the trust stance).
+Mount code runs as an async-function body in a fresh vm realm. Its documented surface steers file, network, process, and timer access through Cordis services so mounts remain inspectable and disposable. Host-realm helpers still make Node escape possible, consistent with the trusted posture. `vmTimeoutMs` bounds only synchronous evaluation.

 Sandbox globals are deliberately small: a tagged write-through `console` (`[cordis:<id>] …` on the host stdout/stderr, so a listener that fires long after the mount call still lands somewhere the user sees), the `harness.defineTool` / `harness.registerTool` registration pair, the encoding primitives fresh vm contexts lack (`btoa`/`atob` as host closures over `Buffer` — a sanctioned exception, `Buffer` itself is never exposed — plus `TextEncoder`/`TextDecoder`), and callable traps over the withheld Node APIs (`require`, `setTimeout`/`setInterval`/`setImmediate`/`clearTimeout`/`clearInterval`, `fetch`) that throw a redirect naming the cordis alternative. Only function-shaped globals are trapped; `process` and `Buffer` stay `undefined` so a `typeof` feature probe stays inert rather than detonating a throwing accessor.

-Three boundary mechanisms make model-written code behave correctly across the realm seam. **Dual-realm `instanceof`**: most objects sandbox code touches are host-realm (tool `args`, event payloads, service returns), so a plain `x instanceof Array` in the vm would silently be false — a per-sandbox prelude gives the vm realm's own constructors a `Symbol.hasInstance` that checks both the vm constructor and its host counterpart, patching only vm-realm globals. **Realm normalization of tool results**: objects built inside the vm carry the vm realm's `Object.prototype`, which the session log's append-time plainness check (`isJsonValue` in `dsh-session`, a prototype-identity comparison) rejects, so the sandbox's `harness.defineTool` JSON round-trips every `execute` return into the host realm — which also projects it onto exactly what the log durably stores — and then shape-checks it against the two `ToolExecuteReturn` forms, so a JSON-valid but wrong-shape return (a bare string, `{ content: 'ok' }`) fails that one call with a teaching error instead of entering the log as corrupt tool-result content. **A whitelist context façade**: the `ctx` a mounted plugin's `apply` receives is NOT the real context nor a pass-through proxy over it — it is a façade exposing only what a mount legitimately needs (`tools.register` marker-guarded, a read-only `tools.get`/`schemas`, `on`/`once`, `provide`, the timer helpers, and the services the plugin DECLARED in `inject`), with every framework-plumbing member (`root`, `parent`, `fiber`, `reflect`, `registry`, `extend`, `isolate`, `intercept`, `plugin`, `set`, `mixin`, …) denied with a teaching error. This closes an escape *class* rather than a single hole: a proxy that merely special-cased `ctx.tools` still handed back the raw context through `ctx.root`, `ctx.extend()`, or a service instance's `.ctx`, and mount code could then `ctx.root.tools.register({…})` to bypass the marker check and realm normalization — a raw vm-realm result then errors a real agent turn at the plainness check. The façade has no context-valued member to reach, and the one indirect leak (an injected-service method returning a `Context`) is rejected on the way back to sandbox code. Two narrower rules complete the surface. First, **service access requires an `inject` declaration**: reaching a service the mount did not declare is refused even when a global provider is live — otherwise a mount could depend on a provider cordis never sees, and unmounting that provider would neither park the consumer nor unwind the tools it registered, leaving a model-visible tool that fails only at execution time. Because the read is gated on the declaration, cross-mount `provide`/`inject` keeps its lifecycle guarantees (the plugin's own `inject` and the fiber's pending/active gating drive activation and unload); only the `apply`-time `ctx` surface is narrowed. Second, **`ctx.tools.get` returns a read-only schema view** (name/description/parameters), never the live `ToolDefinition` — handing back the definition would expose its `execute`, letting mount code call another tool directly and bypass `ToolRegistry.execute` and its pre/post-execute hooks and accounting; a mount that wants to invoke a tool must go through the registry, and one that wants to introspect gets the same view `schemas()` returns.
+Mount code crosses the vm boundary through three controls. Dual-realm `instanceof` recognizes both host and vm objects. `harness.defineTool` normalizes results into host-realm JSON and validates the `ToolExecuteReturn` shape before logging. The mounted plugin receives a whitelist context façade, not a raw or pass-through `Context`; framework plumbing and context-valued returns are rejected. Service reads require a declared `inject`, preserving Cordis activation and unload semantics. `ctx.tools.get` exposes only the schema view, so mounted code cannot bypass `ToolRegistry.execute` by calling a definition directly.

-Boundary errors are written around the mistakes models actually make (see [Consequences](#consequences) for how each was found), and the boundary normalizes rather than lectures wherever the input has exactly one meaning: schema `parameters` accept the JSON-Schema dialect models write by strong prior — the `{ type: 'object', properties, required: […] }` wrapper unwraps to the SchemaSpec DSL (the `required` array becoming per-property flags, at any nesting level), `type: 'integer'` maps to `number`, and `required: false` reads as optional — while genuinely meaningless input is rejected with the vocabulary enumerated (an unknown type lists the five valid ones; a non-boolean `required` names the rule). The remaining teaching errors: an unbalanced `});` closing gets the vm's offending source line plus a "code is a function body" reminder; TypeScript syntax gets the remove-annotations fix (detected on the failing line only, so an ` as ` inside a description string does not misfire); a forgotten `return` gets the two valid plugin forms; a Node built-in call gets the redirect to its cordis service; a tool-name collision on re-mount gets the unmount-first-then-remount recipe.
+The boundary normalizes unambiguous JSON-Schema forms into `SchemaSpec`, including object wrappers, `integer`, and optional fields. Invalid vocabulary fails with the accepted alternatives. Parse, TypeScript, missing-return, Node-API, and duplicate-tool errors include the relevant source line or corrective contract without narrating implementation internals.

 ### The dynamic group and mount lifecycle

-Every dynamic mount is a child of a single `cordis-dynamic` group fiber, itself a child of the `tool-cordis` plugin's fiber. The group exists so the mounts form one subtree: they are disposed as a unit, and disposing `tool-cordis` (HMR reload, config unload) cascades over every mount through the ordinary parent→child fiber lifecycle — no bespoke cleanup. Mounting settles before it reports: the returned fiber is `await()`ed, and a startup error (a throwing `apply`, a duplicate tool name, a duplicate service) disposes the fiber and surfaces as the tool error, so a failed mount never lingers. A settled fiber that is not active is a legal pending mount — cordis semantics for unsatisfied `inject` — kept mounted and reported with what it waits for. Everything the plugin registers is an effect on its fiber, so `cordis_unmount` is nothing but an awaited `fiber.dispose()`.
+All dynamic mounts are children of one `cordis-dynamic` group beneath the tool plugin, so ordinary fiber disposal handles reload and unload. Mounting awaits settlement; startup failure disposes the fiber before returning an error. A settled pending mount remains visible with its missing injections. `cordis_unmount` awaits the mount fiber's disposal.

 ### Cross-mount composition via provide/inject

@@ -44,7 +44,7 @@ Mounts relate to each other through ordinary cordis service semantics, with thei

 ### The generated API catalog

-`cordis_inspect what:"api"` and `what:"events"` answer from a machine-readable catalog generated at build time, never a hand-maintained table that would drift from the JSDoc it paraphrases. [`scripts/gen-cordis-api.ts`](../../../../scripts/gen-cordis-api.ts) reuses `collectServices` / `collectEvents` from [`scripts/gen-cordis-catalog.ts`](../../../../scripts/gen-cordis-catalog.ts) — the same AST walk that generates [the cordis service catalog](../../../cordis-catalog/services.md) and [events catalog](../../../cordis-catalog/events.md) — and emits `packages/cordis/tool-cordis/src/api-catalog.ts`, a committed, banner-commented data module. The artifact carries, per service, its key + one-line summary + raw method signatures; per event, name + `@mode` + signature + summary; the comment-stripped declarations of every exported type the service signatures reference (transitive closure — so a consumer sees that a bash run's `stdout` is `{ text, truncated }`, not a string); plus the curated inherited `ctx` surface shared with the cordis catalog generator. A type name declared in more than one package (each plugin's `Config`) is dropped as ambiguous, and an oversized declaration is truncated with a marker.
+`cordis_inspect` serves API and event data from a generated catalog rather than a duplicated table. The generator reuses the Cordis catalog AST scan and emits service summaries, signatures, event modes, referenced type declarations, and the inherited context surface. Ambiguous type names are omitted and oversized declarations are marked as truncated.

 Freshness is gated like every generated artifact: `pnpm run verify-cordis-api` (in `doc-sync`) regenerates in memory and fails on any diff, so a JSDoc edit that changes a public signature cannot ship without regenerating the catalog the model reads. At runtime the inspect tool intersects the catalog with the live runtime rather than dumping it: live catalogued services render summary + signatures, live services without a catalog entry (mount-provided ones) render name + owning fiber, catalogued services with no live provider are listed tersely, and the referenced type shapes follow.

@@ -78,7 +78,3 @@ The correctness investment therefore goes where it pays for every capability at
 ## Consequences

 The toolset is a deliberate opt-in with a fully-privileged `ctx`, so a deployment adopts it as consciously as a bash tool. Several facts follow that the tool descriptions warn the model about directly: a waterfall listener (e.g. `tools/pre-execute`) that returns without calling `next()` vetoes the chain, so a mounted listener can lobotomize the agent's own tool dispatch ([waterfall semantics](../../../cordis-primer.md#cordis-waterfall-semantics)); mount code runs inside a tool call of the current turn, so awaiting anything that resolves only after the turn deadlocks; `vmTimeoutMs` bounds synchronous evaluation only; and mounts do not survive session resume.
-
-The instructive boundary errors were not guessed — they were written against live self-design sessions in which a real model was asked to build itself coding tools. Those sessions surfaced the failure modes now mitigated: the model closed a returned plugin object with `});` and got only a bare `Unexpected token ')'` it retried blind; it hit a false-positive "this is TypeScript" hint because a description string contained the word "as"; it guessed a bash run's `stdout` was a string and burned six steps building throwaway debug tools to discover it is `{ text, truncated }`; and it wrote tool schemas in the JSON-Schema dialect (`type: 'integer'`, `required: false`, then the full wrapper) three rejections in a row — the rejection text itself pushing it from a nearly-correct DSL attempt back to raw JSON Schema. The fixes — source-line-plus-caret parse errors, line-scoped TypeScript detection, the type-shape closure in the API catalog, the redirect traps, and schema-dialect normalization in place of rejection — cut later sessions from dozens of tool calls with repeated errors to a first-try success on every capability, including a model that hit a Node-`setTimeout` trap and self-corrected to `inject: ['timer']` in one step.
-
-Coverage is named per tier: package unit specs drive the three tools through a real `ToolRegistry` on a real fiber tree (the mount success/failure family, vm isolation, dual-realm `instanceof`, realm normalization against the real `isJsonValue`, the SchemaSpec and raw-registration rejections, the Node-API traps, the cross-mount provide/inject matrix, catalog-backed `api`/`events` rendering, config validation, presenters, quiescent unmount, and the HMR cascade), a `MockAdapter` loop test proves a tool mounted in one step is dispatchable in the next, and the example carries a keyless Loader smoke plus a with-key smoke that world-verifies a live model mounting a listener, building its own tool, and composing two mounts. No snapshot scenario is added: the toolset ships in no ACP-served app, so it changes no editor-facing transcript, and its presenters are unit-tested pure functions — adding it to the ACP example solely for a golden would rewrite the pinned request-header tool set of every recorded scenario.