Merge remote-tracking branch 'origin/master' into codex/rfc-subagent-background-tasks
Adopts #185 (dsh-timeout: clampTimeout/deadline/timeoutOf drive bash run() timeout classification; runBash loses its own timer) and #108 (ask_user_question) across the task-runtime rework: bash-local keeps the BashProcess handle shape with master's deadline mechanics, tool catalogs/expectations carry both the task_* and ask-user tools, and generated docs are regenerated on the union.
This commit is contained in:
Yichen Jiang
128 files changed
+6732
-332
No files matched your search
+4
-1
@@ -13,7 +13,6 @@ Generated by `pnpm run gen-rfc-index` from the RFC tree — never edit by hand;
|
||||
| [Code Mode — the model writes TypeScript against the tool registry](proposed/feature/2026-06-15-code-mode.md) | 2026-06-15 |
|
||||
| [Pre-tool input rewrite — a consistent design](proposed/feature/2026-06-30-pre-tool-input-rewrite.md) | 2026-06-30 |
|
||||
| [Claude Code and Codex subagent backends (out-of-process delegation to external coding agents)](proposed/feature/2026-07-07-claude-code-and-codex-subagent-backends.md) | 2026-07-07 |
|
||||
| [Repeat-tool-call guard plugin](proposed/feature/2026-07-08-repeat-tool-guard.md) | 2026-07-08 |
|
||||
|
||||
### Simplification
|
||||
|
||||
@@ -55,6 +54,7 @@ Generated by `pnpm run gen-rfc-index` from the RFC tree — never edit by hand;
|
||||
| [Compaction as a capability seam (abstract contract + basic backend)](implemented/feature/2026-06-18-compaction-capability-seam.md) | 2026-06-18 |
|
||||
| [Subagent capability seam](implemented/feature/2026-06-21-subagent-capability-seam.md) | 2026-06-21 |
|
||||
| [ACP subagent backend (out-of-process delegation)](implemented/feature/2026-06-22-acp-subagent-backend.md) | 2026-06-22 |
|
||||
| [Ask-user question capability](implemented/feature/2026-06-25-ask-user-question.md) | 2026-06-25 |
|
||||
| [The `todo_write` tool — model task list as event-sourced session state](implemented/feature/2026-06-29-todo-write-tool.md) | 2026-06-29 |
|
||||
| [dsh-hooks-claude + dsh-hooks-codex — the Claude Code / Codex hook bridges](implemented/feature/2026-06-30-hook-bridges.md) | 2026-06-30 |
|
||||
| [dsh-hook-protocol — the shared Claude Code / Codex hook wire-protocol core](implemented/feature/2026-06-30-hook-protocol-lib.md) | 2026-06-30 |
|
||||
@@ -63,6 +63,7 @@ Generated by `pnpm run gen-rfc-index` from the RFC tree — never edit by hand;
|
||||
| [Subagent lifecycle enrichment — lastAssistantMessage (observe-only)](implemented/feature/2026-06-30-subagent-observe-enrich.md) | 2026-06-30 |
|
||||
| [Explicit model-facing tool order](implemented/feature/2026-07-06-explicit-tool-order.md) | 2026-07-06 |
|
||||
| [Background subagent tasks](implemented/feature/2026-07-08-background-subagent-tasks.md) | 2026-07-08 |
|
||||
| [Repeat-tool-call guard plugin](implemented/feature/2026-07-08-repeat-tool-guard.md) | 2026-07-08 |
|
||||
|
||||
### Simplification
|
||||
|
||||
@@ -124,6 +125,8 @@ Generated by `pnpm run gen-rfc-index` from the RFC tree — never edit by hand;
|
||||
| [Prompt variables and tool-guidance ownership](implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md) | 2026-07-05 |
|
||||
| [Every LLM request is reconstructable from the session log](implemented/architecture/2026-07-05-reconstructable-requests.md) | 2026-07-05 |
|
||||
| [Subagent provider-lifecycle events — `subagent/provider-added` / `subagent/provider-removed`](implemented/architecture/2026-07-05-subagent-provider-lifecycle-events.md) | 2026-07-05 |
|
||||
| [A shared timeout/deadline primitive, with hard-kill left to each capability](implemented/architecture/2026-07-06-timeout-deadline-library.md) | 2026-07-06 |
|
||||
| [Tool-call timeout policy as a plugin](implemented/architecture/2026-07-07-tool-call-timeout-policy.md) | 2026-07-07 |
|
||||
|
||||
### Process
|
||||
|
||||
|
||||
@@ -0,0 +1,98 @@
|
||||
# RFC: A shared timeout/deadline primitive, with hard-kill left to each capability
|
||||
|
||||
Status: implemented
|
||||
|
||||
## Problem
|
||||
|
||||
Timeout handling was drifting apart across the tool-bearing capabilities, and the divergence was not superficial — it was the same logic re-implemented three ways, each with its own subtle correctness burden.
|
||||
|
||||
- **bash** ([packages/bash/bash-local/src/run.ts](../../../../packages/bash/bash-local/src/run.ts)) had a full, correct timeout inside the process plumbing: a config-clamped `timeoutMs`, two independent triggers — a `killTimer` for the timeout and an `onAbort` listener for upstream cancellation — each calling one `kill()` closure that escalates SIGTERM→grace→SIGKILL on the process group, and two orthogonal outcome booleans (`timedOut`, `aborted`) latched independently.
|
||||
- **web_fetch** ([packages/web/web-fetch-local/src/provider.ts](../../../../packages/web/web-fetch-local/src/provider.ts)) had a correct but *hand-rolled* timeout: it constructed an `AbortController`, wired `setTimeout(() => controller.abort(new WebError(…, 'WEB_FETCH_TIMEOUT')))`, manually added and removed the upstream-signal listener, cleared the timer in a `finally`, and recovered the timeout reason from `signal.reason` in a `translateAbortOrNetwork` helper because the reader surfaces a bare `AbortError`.
|
||||
- **web_search** ([packages/web/tool-web/src/search.ts](../../../../packages/web/tool-web/src/search.ts)) had **no timeout at all**: `WebSearchRequest` ([packages/web/web/src/types.ts](../../../../packages/web/web/src/types.ts)) carries no `timeoutMs` field, and each provider's `search()` only forwards `exec.signal`. (web_search stays untimed here — see Consequences.)
|
||||
|
||||
Each new external-process or network tool re-derived the same four things — clamp the requested value, start a timer, fuse the timeout with upstream cancellation, and distinguish "timed out" from "cancelled" on the way out — and the fusion and reason-recovery are exactly the parts that are easy to get subtly wrong (web_fetch's `signal.reason` dance is evidence). At the same time, the *termination* each performs is irreducibly different: bash kills an OS process group (work runs in a child process, outside this runtime, reachable only by signal), while web aborts an in-process `fetch` (undici tears down the socket). There is no single mechanism that can stop all of them.
|
||||
|
||||
The two reference agents surveyed converged on the same split. Codex models "what will end this exec early" as one value (`ExecExpiration`, an enum fusing timeout and a cancellation token) whose `wait_with_outcome()` returns `TimedOut | Cancelled`, while the actual `kill_process_group` lives outside it — and that abstraction is reused *only* across the exec family, with MCP, model-stream, and guardian each keeping their own bespoke `tokio::time::timeout`. Claude Code shares nothing: bash and ripgrep each own a private SIGTERM→SIGKILL kill and distinguish timeout from cancellation by throwing distinct error types, while file I/O has no timeout. Both confirm the boundary drawn here: the timing-and-classification half is worth sharing within a family of like-terminated operations; the termination half is not shareable and stays in each capability.
|
||||
|
||||
## Decision
|
||||
|
||||
`@deepseek-ai/dsh-timeout` lives under `packages/util/` (peer to `dsh-brand`) and owns the *timing and classification* half of timeout; the *termination* half — the hard kill — stays in each capability's implementation. It is a library of pure functions, **not** a cordis service or plugin: it takes no `ctx`, registers nothing, holds no cross-call state, and emits no events. There is deliberately no central "timeout service" that would have to know how to stop every capability's work — that knowledge is exactly what a microkernel keeps out of shared layers, and what Codex's exec-only `ExecExpiration` scope demonstrates.
|
||||
|
||||
### The library surface
|
||||
|
||||
Three functions plus one reason type:
|
||||
|
||||
```ts ignore-check
|
||||
/** The internal reason attached to a timeout abort, so consumers can classify it after the fact. */
|
||||
export class TimeoutReason extends Error {
|
||||
override name = 'TimeoutReason'
|
||||
|
||||
constructor(readonly code: string, readonly timeoutMs: number) {
|
||||
super(`${code} after ${timeoutMs}ms`)
|
||||
}
|
||||
}
|
||||
|
||||
/** Validate/fill a caller's optional positive hint from the backend's default, then cap at its max. */
|
||||
export function clampTimeout(
|
||||
requested: number | undefined,
|
||||
def: number,
|
||||
max: number,
|
||||
name = 'timeoutMs',
|
||||
): number
|
||||
|
||||
/**
|
||||
* Build a deadline signal that aborts on upstream cancellation OR on timeout,
|
||||
* with the timeout carrying a `TimeoutReason`. `timeoutMs <= 0` means "no
|
||||
* timeout" (background tasks): forward only the upstream signal, arm no timer.
|
||||
* The returned object's `[Symbol.dispose]` clears the timer — `using` for a
|
||||
* scope-lifetime consumer, a manual call for an event-lifetime one.
|
||||
*/
|
||||
export function deadline(
|
||||
upstream: AbortSignal | undefined,
|
||||
timeoutMs: number,
|
||||
code: string,
|
||||
): { signal: AbortSignal; [Symbol.dispose](): void }
|
||||
|
||||
/** Recover the TimeoutReason from an aborted signal (or error); `code` scopes the match to this deadline's timer. */
|
||||
export function timeoutOf(x: AbortSignal | { reason?: unknown }, code?: string): TimeoutReason | undefined
|
||||
```
|
||||
|
||||
`deadline` is `AbortSignal.any([upstream, <timeout controller>])` with three things the standard library does not give: a typed, identifiable `TimeoutReason` on the timeout abort (native `AbortSignal.timeout()` yields a fixed `TimeoutError`, indistinguishable across timeout kinds), an internal `timeoutMs <= 0` "no timeout" sentinel for backend-owned background work, and a `Symbol.dispose` cleanup that works with both `using` and manual disposal. `AbortSignal.any` is a Node ≥ 20 primitive; it is the single mechanism that fuses two abort sources into one, adopting the reason of whichever fires first. External request hints validate as positive finite numbers via `clampTimeout` before they reach `deadline`; `0` is not a model-/plugin-facing "disable timeout" value. When `timeoutMs <= 0` and no upstream signal is present, `deadline()` returns a never-aborting signal plus a no-op disposer so callers keep one call shape. `TimeoutReason` is an internal classification reason: providers translate it into seam-specific public errors or result fields before returning to callers. `timeoutOf`'s optional `code` scopes classification to the caller's own deadline: when the `upstream` is itself a deadline (a future `tools/execute` middleware arming a per-call deadline), `AbortSignal.any` preserves the outer `TimeoutReason` if it fires first, and an unscoped match would misreport the outer timeout as the inner capability's own; scoping to `code` reads a foreign timeout as an ordinary upstream cancel.
|
||||
|
||||
### The division of labor
|
||||
|
||||
| Concern | Owner |
|
||||
|---|---|
|
||||
| Validate request hint and clamp default/max | `dsh-timeout` (`clampTimeout`) — pure arithmetic plus the shared positive-finite request contract |
|
||||
| Arm timer, abort on deadline, carry reason, fuse with upstream cancel | `dsh-timeout` (`deadline`) |
|
||||
| Clear the timer | `dsh-timeout` (`[Symbol.dispose]`) |
|
||||
| Classify the first abort reason after abort | `dsh-timeout` (`timeoutOf`) |
|
||||
| **Actually terminate the work** | the capability's implementation |
|
||||
| The default/max *values* | the capability's config |
|
||||
| The timeout `code` string | the capability (`WEB_FETCH_TIMEOUT` ≠ `BASH_TIMEOUT`) |
|
||||
|
||||
The signal only *notifies*; termination is always the listener's job, and the listener differs by capability. bash writes its own `addEventListener('abort', kill)` because the OS process lives outside this runtime and nothing else will kill it; web hands `d.signal` to `fetch` and undici tears down the socket. This is why file read/write/edit take **no** `timeoutMs`: a local syscall is best-effort-abortable at most, a timeout could not force `fsync`/`rename` to stop, and adding one would be an implicit default that violates explicit-over-implicit. Both reference agents leave file I/O untimed for the same reason.
|
||||
|
||||
### How each capability consumes it
|
||||
|
||||
- **web_fetch** — the tool stays validate-and-forward; the provider's hand-rolled controller + `setTimeout` + manual listener + `finally` + `signal.reason` recovery is replaced by provider-owned `deadline`/`timeoutOf`. A pre-aborted upstream signal still throws `WEB_ABORTED` up front; otherwise `fetch` runs against the fused `d.signal`, and `translateAbortOrNetwork` classifies a thrown error by the signal (`timeoutOf` → `WEB_FETCH_TIMEOUT`, else aborted → `WEB_ABORTED`, else network → `WEB_PROVIDER_ERROR`). The public error-code contract is unchanged, and `TimeoutReason` never crosses the web seam as the public error.
|
||||
- **bash** — `resolve()` stays a pure request-to-spec step: it clamps with `clampTimeout(request.timeoutMs, config.timeoutMs, config.maxTimeoutMs, 'bash-local: request.timeoutMs')` and carries `request.signal` through unchanged. Foreground `run()` owns the timeout: `using d = deadline(spec.signal, spec.timeoutMs, 'BASH_TIMEOUT')`, then `runBash` receives only `d.signal`. `runBash` no longer owns any timer — it listens for abort and runs its existing SIGTERM→grace→SIGKILL process-group kill, and its `SpawnSpec`/`SpawnOutcome` no longer carry `timeoutMs`/`timedOut`/`aborted` (the executor classifies from the deadline signal instead). `run()` computes `timedOut = timeoutOf(d.signal, 'BASH_TIMEOUT') !== undefined` and `aborted = d.signal.aborted && !timedOut`, so the public seam booleans (`BashRunResult.timedOut`/`aborted`) are mutually exclusive — the shared deadline reports the cause that first cut the command short, and the `code` scope keeps a nested outer deadline from being misread as bash's own timeout. Background `start()` creates no deadline and forwards only the upstream signal, so background tasks stay timeout-free; a task's killed-vs-completed status reads its own `spec.signal.aborted`.
|
||||
|
||||
## Consequences
|
||||
|
||||
- `runBash`'s outcome no longer independently latches `timedOut` and `aborted`; a timeout and a user abort racing before process close now report a single first-abort cause instead of both being true. The uniform SIGTERM→grace→SIGKILL kill is unchanged, and the seam type `BashRunResult` keeps both booleans (now mutually exclusive), so `dsh-tool-bash`'s result rendering is untouched.
|
||||
- `SpawnSpec.timeoutMs` and `SpawnOutcome.timedOut`/`aborted` were removed rather than kept as always-zero/always-false vestiges: with `runBash` owning no timer and the executor owning classification, they were read nowhere. This is the one deviation from the literal proposal shape (which passed `timeoutMs: 0` into `runBash`); an always-0 field read by nothing is dead weight under the per-file coverage gate.
|
||||
- web_fetch shed its bespoke controller/timer/listener/reason-recovery; the classifier now keys off the deadline signal (`timeoutOf` + `aborted`) rather than the thrown error's shape, which is robust across both the request-phase reject-with-reason and the read-phase bare-`AbortError`.
|
||||
- `AbortSignal.any` and `using`/`Symbol.dispose` enter the repo for the first time here (Node ≥ 24 baseline, already met).
|
||||
|
||||
Out of scope, named to mark the boundary: `web_search` can gain an optional model-facing `timeout_ms` once its tool-schema/snapshot coverage is planned; future ripgrep-backed fs discovery tools can consume the same provider-owned deadline shape once they exist; a `tools/execute` waterfall middleware could arm a default deadline for every tool call by driving `exec.signal` — that would be a plugin that *consumes* this library and still only notifies, the hard kill remaining each capability's job.
|
||||
|
||||
## Alternatives considered
|
||||
|
||||
**A unified timeout *plugin* / `ctx.timeout` service.** Rejected on microkernel grounds. A service that could stop any tool's work would have to understand every capability's termination mechanism (process-group SIGKILL, socket teardown, syscall-boundary checks) — the "kernel knows too much" the architecture forbids. Codex's `ExecExpiration` is scoped to the exec family precisely because the kill it drives (`killpg`) is process-family-specific; MCP and model-stream keep their own. There is no coherent middle layer that owns termination for everything, so the shared piece can only be the pure timing/classification half — a library, not a service.
|
||||
|
||||
**Per-tool ad-hoc timeout, no shared code (the prior status quo, and Claude Code's choice).** Rejected because it was already producing divergence and duplicated correctness burden: web_fetch hand-rolled the exact controller/reason logic that future network/process-backed tools would each have to re-derive, and the fusion + `signal.reason` recovery are the error-prone parts. Claude Code tolerates full duplication; this repo has a single shared abort channel (`exec.signal` on every `execute`) that makes a small shared primitive strictly cleaner, so the cost/benefit differs.
|
||||
|
||||
**A `withTimeout(promise, ms)` wrapper instead of a signal factory.** Rejected because racing a promise against a timer resolves the *tool-call* promise on deadline without stopping the underlying work — the child process or fetch socket leaks on. Handing out a signal and requiring the capability to listen is what forces a real termination path to exist. This mirrors the "dispose must reach quiescence, not just request it" defensive rule.
|
||||
|
||||
**Keep bash's two independent triggers (`killTimer` + `onAbort`) rather than fusing.** Rejected for the convergence goal: fusing into one `deadline` signal removes bash's bespoke timer and gives every capability one shape. The trade-off is that bash's `timedOut`/`aborted` booleans become first-abort classifications rather than independent facts that can both be true when timeout and user abort race before process close. That is acceptable because the result reports the cause that first cut the command short; the termination action stays the same uniform SIGTERM→grace→SIGKILL kill. Note the deliberate non-alignment with Codex: Codex forks its kill by outcome (timeout → immediate SIGKILL; cancel → SIGTERM + 50 ms grace → SIGKILL), whereas the fused signal drives one uniform `kill()` for both, matching Claude Code's unified bash kill. Splitting the kill by `timeoutOf` is possible later if a need appears; there is none now.
|
||||
@@ -0,0 +1,110 @@
|
||||
# RFC: Tool-call timeout policy as a plugin
|
||||
|
||||
Status: implemented
|
||||
|
||||
## Problem
|
||||
|
||||
The [timeout/deadline RFC](2026-07-06-timeout-deadline-library.md) extracted the timing-and-classification primitive into `@deepseek-ai/dsh-timeout`, but timeout policy was still attached to individual capabilities and model-facing schemas. `bash` exposed `timeoutMs`; `web_fetch` exposed `timeout_ms`; `web_search` had no model-facing timeout even though providers already honor `exec.signal`; a future grep/glob tool would either import the timeout library directly or invent its own timeout policy. That is the wrong authoring shape for a plugin SDK: a tool author should normally forward `exec.signal` to the implementation it calls, and deployment policy should decide the budget.
|
||||
|
||||
At the same time, not every timeout in the repo is a model-facing tool-call budget. Hooks execute command hooks by calling `ctx.bash` directly, not through `ctx.tools.execute()`, and the `bash` model tool multiplexes foreground execution, background start, background polling, and hook reuse through the same backend. Moving every timeout into a tool plugin in one step would conflate those paths and risk breaking hook timeout semantics.
|
||||
|
||||
## Decision
|
||||
|
||||
Tool-call timeout is a policy that applies only to model-facing tool execution, in three parts:
|
||||
|
||||
- `@deepseek-ai/dsh-timeout` remains the shared library that owns `deadline()` and `timeoutOf()`.
|
||||
- `@deepseek-ai/dsh-tools` has an around-dispatch waterfall, `tools/execute`, between `tools/pre-execute` and `tools/post-execute`.
|
||||
- `@deepseek-ai/dsh-timeout-policy` reads each tool's declared `timeoutMs` from the registry and wraps a call that has one by deriving a new `exec.signal`.
|
||||
|
||||
The execution pipeline is:
|
||||
|
||||
```text
|
||||
ctx.tools.execute(exec)
|
||||
-> tools/pre-execute
|
||||
-> tools/execute
|
||||
-> registry dispatch (the base next())
|
||||
-> tool.execute(args, exec)
|
||||
-> thrown tool errors normalize to ToolExecutionResult
|
||||
-> tools/post-execute
|
||||
```
|
||||
|
||||
The default behavior is conservative: a tool that declares no `timeoutMs` receives no `TOOL_TIMEOUT` deadline from the plugin.
|
||||
|
||||
### The `tools/execute` around seam
|
||||
|
||||
`@deepseek-ai/dsh-tools` declares a `tools/execute` waterfall whose base `next()` is the dispatch-with-normalization thunk — the same inner `try`/`catch` that turns a thrown tool (or unknown tool) into an `isError` `ToolExecutionResult`. A listener receives `(exec, next)`: it calls `next()` to delegate to dispatch (returning its result, optionally wrapped) or returns a replacement result to short-circuit dispatch. The whole pipeline still sits inside `execute`'s outer try/catch, so a throwing listener becomes an `isError` result, never a turn failure.
|
||||
|
||||
That the catch is the base `next` — not something outside the waterfall — is load-bearing: when a provider sees the timeout signal and throws its own upstream-abort error, registry dispatch first converts it to a normal error result, and only then can `timeout-policy` replace the final result with `TOOL_TIMEOUT`.
|
||||
|
||||
### The `timeout-policy` plugin
|
||||
|
||||
The plugin is `@deepseek-ai/dsh-timeout-policy`, a zero-config function/namespace plugin (`name` / `inject` / `apply`) in the `packages/timeout/` group. The per-tool budget is DECLARED on the tool, not on this plugin: a `ToolDefinition` carries an optional `timeoutMs`, which the owning tool plugin sets from its own config. `dsh-tool-web`, for example, resolves `fetchTimeoutMs` / `searchTimeoutMs` (default 30000) onto the `web_fetch` / `web_search` definitions:
|
||||
|
||||
```yaml
|
||||
- id: timeout-policy
|
||||
name: '@deepseek-ai/dsh-timeout-policy'
|
||||
- id: tool-web
|
||||
name: '@deepseek-ai/dsh-tool-web'
|
||||
config:
|
||||
fetchTimeoutMs: 30000
|
||||
searchTimeoutMs: 30000
|
||||
```
|
||||
|
||||
Keeping the tool name out of this plugin's config is deliberate: a budget keyed by a free-text tool name could be mistyped (`web_fech`) and then silently apply to nothing. Declaring `timeoutMs` on the tool makes that failure class structurally impossible — the enforcer reads `ctx.tools.get(exec.name)?.timeoutMs`, and `exec.name` is the tool being dispatched, so the lookup always resolves and there is no unknown-name path to warn or throw about. `timeoutMs` is validated positive-finite by `defineTool` at definition time. For a tool that declares a budget the listener arms `deadline(exec.signal, timeoutMs, 'TOOL_TIMEOUT')`, swaps the derived signal onto `exec` for the downstream dispatch, restores the caller's own signal afterward, and returns a structured `TOOL_TIMEOUT` result when `timeoutOf(d.signal, 'TOOL_TIMEOUT')` matches. A tool with no declared budget delegates unchanged.
|
||||
|
||||
Signal replacement is by **in-place mutation of `exec.signal`**, not by passing a new object to `next()`. Cordis's waterfall `next()` ignores any arguments handed to it and re-invokes downstream listeners with the shared payload array (`vendor/cordis/src/events.ts`), so the documented cordis idiom — mutate the shared object, then delegate — is the only mechanism that reaches dispatch. The plugin restores `exec.signal` to the caller's original in a `finally` so `tools/post-execute` never sees this plugin's (possibly already-aborted) deadline signal.
|
||||
|
||||
`timeout-policy` owns both uses of the `TOOL_TIMEOUT` code: the internal deadline code passed to `deadline()`/`timeoutOf()` (scoped so a nested outer deadline reads as an ordinary cancel) and the structured tool-result error code. Its replacement result is:
|
||||
|
||||
```ts ignore-check
|
||||
function toolTimeoutResult(callId: CallId, timeoutMs: number): ToolExecutionResult {
|
||||
return {
|
||||
callId,
|
||||
content: [{ type: 'text', text: `Error: tool call timed out after ${timeoutMs}ms` }],
|
||||
isError: true,
|
||||
error: { name: 'ToolTimeoutError', code: 'TOOL_TIMEOUT' },
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
This is a cooperative deadline. It does not kill arbitrary work by racing the tool promise; the tool or the capability it calls must honor `exec.signal` and reach quiescence. Declaring `timeoutMs` therefore MEANS "this tool is cooperative with `exec.signal`", which the plugin README states as its contract.
|
||||
|
||||
No new session event is needed for reconstructability: `TOOL_TIMEOUT` is the final model-facing `tool/result` for that call, so the existing session log already records the content and structured `{ name, code }` error the next model request sees.
|
||||
|
||||
### Existing tool adaptation
|
||||
|
||||
`web_fetch` and `web_search` are migrated. `dsh-tool-web` keeps ownership of their model-facing schemas, and those schemas expose no timeout knob: `web_fetch` dropped its `timeout_ms` parameter to match the reference-agent shape, and `web_search` stays query-only. The tool bodies do not import `@deepseek-ai/dsh-timeout`; they forward `exec.signal` to `ctx.web`.
|
||||
|
||||
`dsh-web-fetch-local` keeps a provider-level timeout (`timeoutMs`/`maxTimeoutMs`) as a large resource backstop for direct `ctx.web.fetch()` callers and misconfigured deployments; it owns no model-facing timeout. When a `TOOL_TIMEOUT` signal reaches the fetch provider first, provider-scoped classification treats it as upstream `WEB_ABORTED`, and the outer `tools/execute` wrapper replaces the final tool result with `TOOL_TIMEOUT`. A shipped web-tool deployment configures the provider backstop above the `timeout-policy` budget so the tool-call policy normally wins for model calls.
|
||||
|
||||
`bash` stays on the current backend timeout path. `dsh-tool-bash` continues to expose `timeoutMs` and `run_in_background`; `dsh-bash-local` continues to use `@deepseek-ai/dsh-timeout` for `BASH_TIMEOUT`; hook bridges continue to call `runHook()` and pass `timeoutMs` through `ctx.bash`. This keeps foreground/background/hook behavior stable.
|
||||
|
||||
`read`, `write`, `edit`, `todo_write`, `bash_output`, and `bash_kill` do not opt into tool-call timeout: they are local filesystem or short registry/session operations where a deadline would be best-effort only or unnecessary.
|
||||
|
||||
A future model-facing grep/glob tool can be implemented on top of `ctx.bash` without importing `@deepseek-ai/dsh-timeout`: it forwards `exec.signal` to `ctx.bash`, and declares its own `timeoutMs` (from its plugin's config) for the enforcer to apply. If bash-local's backend timeout becomes a problem for such a tool, the bash seam can later add a caller-owned-deadline mode; that is outside this cut.
|
||||
|
||||
## Alternatives considered
|
||||
|
||||
**Name the plugin `tool-timeout`.** The literal RFC name matched the `gen-tool-catalog` completeness guard's `packages/*/tool-*` glob, which requires every match to register a model-facing tool. This plugin registers none — it is a `tools/execute` wrapper — so a `tool-*` name would either fail `verify-tool-catalog` or force a misleading boot entry. The package is `@deepseek-ai/dsh-timeout-policy` in a new `packages/timeout/` group; the cordis.yml `id` can still be `timeout-policy`.
|
||||
|
||||
**Keep per-tool timeout handling only.** This was the shape for `bash` and `web_fetch`, and it matches Claude Code and Codex for shell commands. It loses for web-style tools because every new timeout-capable tool must choose validation, cap semantics, docs, snapshots, and classification. The plugin centralizes policy and classification while leaving each tool's schema focused on business input.
|
||||
|
||||
**Move all timeout policy out of bash-local immediately.** Cleaner long-term — bash-local would become a pure subprocess executor and all callers would own their deadlines. It loses as the first step because hooks call `ctx.bash` directly and the bash model tool has foreground/background semantics that are not the same tool-call lifetime. Keeping `BASH_TIMEOUT` preserves those paths while tool-call timeout proves itself on simpler tools.
|
||||
|
||||
**Use a global default budget for every tool.** Convenient, but it surprises tool authors: any tool that accidentally runs longer than the global budget would start failing once the plugin loads. A per-tool declared budget makes adoption deliberate.
|
||||
|
||||
**Expose a model-facing `timeout_ms` override.** Claude Code's `WebFetch`/`WebSearch` and Codex's web tools keep timeout out of the model-call shape. A model override would make timeout part of prompt semantics and force schema/argument-stripping rules into `timeout-policy`. Web timeout stays deployment policy only.
|
||||
|
||||
**Let `timeout-policy` match tool arguments itself.** A rule engine such as "disable timeout when `bash.run_in_background` is true" would make the policy plugin know tool-specific argument semantics. Avoided by not migrating bash to tool-call timeout.
|
||||
|
||||
**Use `tools/pre-execute` plus `tools/post-execute` instead of a new around seam.** A pre listener could arm a deadline and mutate `exec.signal`; a post listener could classify and replace. That loses because the deadline lifetime would cross two independent waterfalls: a call-id map, cleanup on every pre-deny/tool-throw/post-throw/dispose path, and ordering rules with every other listener. `tools/pre-execute` is also the allow/deny gate, not an execution wrapper. `tools/execute` gives the timeout one lexical scope: arm, delegate, classify, dispose.
|
||||
|
||||
**Use `Promise.race` to enforce timeouts for non-cooperative tools.** Rejected for the same reason as the timeout-library RFC: it returns control to the caller while the underlying process, fetch, or provider operation may still be running. The plugin only sends a signal; termination remains the implementation's responsibility.
|
||||
|
||||
## Consequences
|
||||
|
||||
- `@deepseek-ai/dsh-tools` gains an around-dispatch surface after the interception seams deliberately split pre/post tool hooks. Its contract is narrow — wrap registry dispatch, not replace the pre-gate or post-result policy — and the base `next()` is dispatch-with-normalization so a wrapper never sees a raw tool throw.
|
||||
- Multiple `tools/execute` listeners compose by ordinary Cordis waterfall order: a listener that calls `next()` wraps downstream listeners plus dispatch; one that returns without `next()` short-circuits them. A deployment combining timeout with a future retry/sandbox/metrics wrapper chooses semantics by registration order ("timeout covers the whole retry" vs "timeout covers each attempt").
|
||||
- Opt-in by declaration is a deliberate misconfiguration risk: a tool can declare a `timeoutMs` without honoring `exec.signal`, and that tool will not stop on timeout. The plugin contract states that declaring a budget means cooperative; the web tools prove the pattern on tools that already forward the signal.
|
||||
- During the transition `bash` and the migrated web tools use different timeout paths on purpose: `TOOL_TIMEOUT` is the model-facing tool-call budget, while `BASH_TIMEOUT` remains the bash backend timeout used by bash and hooks.
|
||||
- Deviation from the literal proposal, recorded per the implemented-RFC rule: the plugin package is `@deepseek-ai/dsh-timeout-policy` (not `tool-timeout`), signal replacement is in-place `exec.signal` mutation before `next()` (not `next({ ...exec, signal })`, which cordis ignores), and the per-tool budget is declared on the `ToolDefinition` (`timeoutMs`, set by the owning tool plugin from its config) rather than mapped by tool name in this plugin's config — so the enforcer is zero-config and a mistyped tool name is impossible. All three are described in `## Decision` above.
|
||||
@@ -0,0 +1,49 @@
|
||||
# RFC: Ask-user question capability
|
||||
|
||||
Status: implemented
|
||||
|
||||
## Problem
|
||||
|
||||
The agent sometimes cannot proceed safely from model inference alone: it needs the human to choose a path, confirm a risky/default action, or provide missing information. Before this change, the only way to get that answer was for the model to ask in assistant text and then stop, which broke the normal tool-call loop: the agent had no structured way to pause, no option metadata for UIs, no abort/error taxonomy, and no way for non-stdio front doors to present the question consistently.
|
||||
|
||||
This is a user-facing capability, but it also crosses package boundaries. A model-facing tool needs a provider-neutral request vocabulary; each UI surface needs to decide how to show and collect the answer; the agent loop should remain unchanged because a tool call already has the right async shape.
|
||||
|
||||
## Decision
|
||||
|
||||
Introduce `dsh-user-interaction` as the provider-neutral interface package for `ctx.userInteraction`, colocated with the model-facing consumer `dsh-tool-ask-user` under `packages/ui`. The grouping is intentional: asking a human is a UI-backed product affordance, not part of the providerless core spine. The seam still owns the stable request/answer/error vocabulary, while UI product surfaces provide the concrete provider that collects the answer. The tool registers `ask_user_question`, forwards `{ questions, agent, signal }`, and returns the provider-computed structured answers as the tool result.
|
||||
|
||||
The model-facing request vocabulary is deliberately aligned with the product-research schema: `ask_user_question({ questions: [{ id, question, header?, options?: [{ label, description? }], multi_select? }] })`. `id` is supplied per question and echoed in the result so a batch can be routed without relying on question text. `label` is both user-facing display text and the selected value returned to the model; there is no separate `value`, no `recommended`, no `allow_custom`, and no `desc` alias.
|
||||
|
||||
Providers return `{ answers: [{ id, selected, custom? }] }`. `selected` is always an array of selected option labels, so single-select and `multi_select` answers share one result shape. `custom` carries a free-text "Other" answer; optionless questions collect `custom` directly. When `custom` is present, it overrides any selected choices and `selected` is empty.
|
||||
|
||||
`UserInteractionError` extends `HarnessError`, so failures such as `NO_PROVIDER`, `ASK_ABORTED`, ACP cancellation, or missing session routing survive `ctx.tools.execute()` as machine-routable `{ name, code }` tool errors. This matches the structured-error taxonomy and lets the model or a wrapping plugin distinguish "user cancelled" from a generic thrown exception.
|
||||
|
||||
## UI mappings
|
||||
|
||||
`dsh-stdio-agent`'s in-package readline module renders each question, shows each option's `description` on the next line, supports comma/space-separated numeric choices for `multi_select`, accepts free-form custom answers, and rejects pending questions on abort, provider disposal, or stdin EOF. A batched request is asked in order and resolved as one answer object. The stdio provider serializes simultaneous requests with an internal queue so only one prompt owns stdin at a time.
|
||||
|
||||
`dsh-acp` provides the same seam for ACP sessions. It routes an ask request from the calling `Agent` through the bridge's `agent→sessionId` reverse map and calls ACP `unstable_createElicitation` with a session-scoped form for each question. Single-select options become a `choice` string enum; `multi_select` options become a `choice` array enum; optionless questions use a required `custom` text field. If the client returns both `choice` and non-empty `custom`, the custom answer wins. ACP `decline`/`cancel`, a missing answer, a missing session, and a client without elicitation support all become structured `UserInteractionError`s.
|
||||
|
||||
The ACP mapping deliberately uses elicitation, not `session/request_permission`. `request_permission` is still reserved for the separate permission gate: it is a yes/no-or-policy authorization protocol around tool execution. `ask_user_question` is a general information-gathering tool with optional free-form answers, so ACP form elicitation is the closer protocol fit. The bridge's session routing is shared with the future permission gate, but the user intent is different.
|
||||
|
||||
## Alternatives considered
|
||||
|
||||
**Assistant text followed by a stopped turn.** The model could ask the user in plain assistant text and then stop. That loses the structured option metadata, gives UIs no provider-neutral way to render a choice, and forces the next human answer to arrive as a new user prompt rather than as the result of the operation that needed the answer.
|
||||
|
||||
**Core-owned ask-user packages.** The first implementation split the seam and the model-facing tool across `packages/core` and `packages/ui`, but both names describe one UI-backed human-interaction affordance. The seam remains provider-neutral, but it is not providerless core infrastructure like sessions, tools, or the agent registry. Keeping `dsh-user-interaction` and `dsh-tool-ask-user` together under `packages/ui` makes the package map match the product boundary: apps and bridges provide the human-answer provider, and the stdio app opts into the model-facing tool.
|
||||
|
||||
**ACP `session/request_permission`.** Permission requests are authorization around tool execution; `ask_user_question` is information gathering with optional free-form answers. Using permission for general questions would collapse two different product concepts and make the future permission gate harder to reason about.
|
||||
|
||||
**A loop-level pause primitive.** The agent loop already knows how to await a tool call and resume from a tool result. Adding a new loop special case would duplicate that async shape and make every loop implementation learn about a UI concern.
|
||||
|
||||
## Consequences
|
||||
|
||||
ACP elicitation is currently marked unstable in the SDK. The fallback is still structured: if a client does not implement it, the tool returns `ASK_FAILED` rather than hanging. A later ACP stabilization may rename or reshape the method; that migration should stay inside `dsh-acp` because the core `ctx.userInteraction` vocabulary is provider-neutral.
|
||||
|
||||
The feature gives the model a powerful pause primitive, so prompt guidance matters. The tool description tells the model to ask concise questions and use options when possible. Product policy can later wrap `tools/execute` to restrict when the tool is allowed, but the loop should not special-case it.
|
||||
|
||||
`dsh-user-interaction` and `dsh-tool-ask-user` both live in `packages/ui` because they form one product-facing human-interaction capability. `agent-core` does not load either the tool or a provider. `stdio-agent` opts into the seam, its readline provider, and the model-facing tool. `acp-agent` keeps only the `userInteraction` seam/provider by default: ACP elicitation support is still client-dependent, so an ACP leaf must opt into the model-facing tool deliberately once its client can complete elicitation requests.
|
||||
|
||||
## Testing
|
||||
|
||||
Unit coverage pins provider registration/disposal, duplicate-provider rejection, abort-before-provider, empty-question rejection, structured tool errors through `ctx.tools.execute()`, batched answers, multi-select answers, custom answers, and the model schema including the removal of `value`, `recommended`, `allow_custom`, and `desc`. `dsh-stdio-agent` tests cover option descriptions, queued requests, EOF/abort cleanup, optionless free-form input, invalid option reprompts, duplicate multi-select numbers, and batched question flows. ACP bridge tests drive a real in-memory ACP connection with the real `ask_user_question` tool and verify selected-option, custom-overrides-choice, multi-select, and optionless free-form elicitation paths continue the agent loop.
|
||||
@@ -0,0 +1,73 @@
|
||||
# RFC: Repeat-tool-call guard plugin
|
||||
|
||||
Status: implemented
|
||||
|
||||
## Problem
|
||||
|
||||
A model stuck in a loop re-issues the same tool call with byte-identical arguments — re-running a failing grep, re-reading an unchanged file, polling a command that already gave its answer — and each round trip burns tokens, wall-clock, and (for paid APIs) money without adding information. The harness has nothing that notices: the loop has no step budget, no plugin tracks call repetition, and the model only escapes when it happens to vary its own behavior. The failure mode is real and cheap to detect — [pi-repeat-tool-guard](https://github.com/Kingwl/pi-repeat-tool-guard) ships exactly this as a pi coding-agent extension: count consecutive identical calls and, past a threshold, append a `<system-reminder>` telling the model to stop repeating itself and change course.
|
||||
|
||||
The harness already has every seam the pi extension uses, and better ones: [the interception-seams RFC](2026-06-30-interception-seams.md) gives `tools/post-execute` a sanctioned way to attach model-facing context to a finished call, the loop buffers and injects that context with call/result adjacency preserved, and injected context is a logged `context/message` — so a native guard satisfies the model-visible ⟺ logged rule with no new session event. What was missing was only the plugin itself.
|
||||
|
||||
## 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 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.
|
||||
|
||||
- **`tools/post-execute` (waterfall)** — the one detection point. The listener receives `(exec, result)` together, so counting and reminder delivery need no cross-event pending map (the pi extension needs one only because its `tool_call`/`tool_result` hooks are separate events). It always delegates via `next()` and, when a threshold is hit, folds a reminder onto the downstream decision's `additionalContext` — the observe-and-enrich posture [the hooks bridges](2026-06-30-hook-bridges.md) already use, honoring the waterfall contract. Counting happens here rather than in `tools/pre-execute` because post-execute also runs for denied calls (`ToolRegistry.execute` routes a deny through the same pipeline), and a model hammering a denied call is exactly the loop worth breaking.
|
||||
- **`agent/prompt-submit` (waterfall)** — pure reset hook: delegate via `next()`, clear the submitting agent's chain. A user interjection changes the context; repetition across it is not a loop.
|
||||
- **`agent/status` (emit)** — on `disposed`, drop the agent's state, bounding the maps over harness lifetime.
|
||||
|
||||
### Detection semantics
|
||||
|
||||
The chain key is `(tool name, canonical arguments)`; a call identical to the previous tracked call increments the agent's consecutive counter, a different tracked call resets it to 1. Canonicalization is a deep key-sort plus `JSON.stringify`: `ToolExecution.arguments` is by construction the loop's `JSON.parse` output (or the raw string fallback for malformed argument JSON, which is itself a comparable value), so the pi original's bigint/circular/`undefined` handling has no inputs here and is deliberately dropped.
|
||||
|
||||
Two deliberate rules, both documented in [the package README](../../../../packages/guard/repeat-tool-guard/README.md) because they are behavior a reader would otherwise guess at:
|
||||
|
||||
- **Untracked calls are transparent to the chain.** A call excluded by `include`/`exclude` neither increments nor resets the counter, so `grep X → todo_write → grep X` still counts as two consecutive `grep X` when `todo_write` is excluded. This is what makes exclusion useful — bookkeeping tools interleaved into a loop must not launder it — and it is the pi extension's (undocumented) semantics, kept on purpose and written down.
|
||||
- **Calls without an agent are ignored.** A direct `ctx.tools.execute()` caller (tests, non-loop consumers) has no model to remind and no `AgentId` to key on.
|
||||
|
||||
### 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.
|
||||
|
||||
### Config
|
||||
|
||||
```yaml
|
||||
- id: repeat-tool-guard
|
||||
name: '@deepseek-ai/dsh-repeat-tool-guard'
|
||||
config:
|
||||
thresholds: [3, 5, 8] # default; consecutive counts that trigger a reminder
|
||||
include: [] # tool-name patterns to track; empty ⇒ all tools
|
||||
exclude: [todo_write] # tool-name patterns transparent to the chain
|
||||
argumentsPreviewChars: 500 # default; cap on arguments quoted in the detailed reminder
|
||||
```
|
||||
|
||||
`thresholds` is validated at load and throws on an empty list, a non-integer, a value below 2, or a duplicate — misconfiguration fails loud, replacing the pi original's silent fall-back to defaults. `include`/`exclude` entries support `*` wildcards. Patterns are predicates over whatever tools exist at call time, not references to a registry entry, so an entry matching no currently registered tool is NOT an error — unlike `toolOrder`'s referent check, `exclude: [mcp_*]` must stay valid in a deployment that loads no MCP tools.
|
||||
|
||||
## 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.
|
||||
|
||||
## Alternatives considered
|
||||
|
||||
- **Append the reminder into the tool result** (`accept` with replaced `content` — the pi extension's mechanism, which patches result content because that is the only channel its API offers) — rejected: it makes the logged `tool/result` lie about what the tool returned, and `additionalContext` exists precisely as the separate sanctioned channel for post-execute commentary, with loop-level buffering that preserves call/result adjacency.
|
||||
- **Count in `tools/pre-execute` with a pending-reminder map** (the pi two-phase shape) — rejected: post-execute alone sees `(exec, result)` together and also fires for denied calls, so one listener with no cross-event state covers strictly more attempts with less machinery.
|
||||
- **Escalate to `block` at the highest threshold** — rejected for the initial scope: a blocked call punishes legitimate identical repeats (polling a long-running terminal, re-checking a file the agent expects to change), and an advisory reminder keeps the model in control. Revisit with evidence; the decision shape (`PostToolDecision`) already supports it.
|
||||
- **A per-deployment external hook via the CC/Codex bridges** (a `PostToolUse` script) — rejected as the answer: it works for one deployment, but a shipped, unit-tested, `cordis.yml`-configurable plugin is the harness-native form, without per-call subprocess cost.
|
||||
- **A loop-level step or repetition budget in `agent-loop`** — rejected: "plugins, not loop changes"; a hard step budget is a blunter, orthogonal control that would need its own proposal.
|
||||
- **Fuzzy/near-identical detection** (normalized paths, similar-but-not-equal arguments) — rejected: exact match after canonicalization is cheap, deterministic, and explainable to the model; similarity thresholds invite false positives and need evidence before they earn complexity.
|
||||
- **Placing the package in `core/`** — rejected: core is the product spine; a behavioral guard is an optional leaf plugin, and the `todo/` precedent is a small dedicated group per plugin family.
|
||||
|
||||
## Consequences
|
||||
|
||||
- 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
|
||||
|
||||
- Compaction does not reset chains: a compacted history changes what the model sees, but the repetition risk usually survives compaction.
|
||||
- Escalating to `block` at a high threshold is not implemented; `PostToolDecision` already supports it if evidence arrives.
|
||||
- Subagent chains stay isolated per agent; no sharing mechanism exists until a concrete case appears.
|
||||
@@ -68,7 +68,7 @@ The replay plugin lives in its own package, `@deepseek-ai/dsh-llm-replay` (`pack
|
||||
|
||||
### Two subcommands, replay in the default gate
|
||||
|
||||
`pnpm run test:snapshot` runs replay (keyless) and is composed into the default `pnpm run test` gate so every PR gets the regression check (the main `vitest.config.ts` include stays narrow; the gate is `test && test:snapshot`). `pnpm run test:snapshot:record` requires `DEEPSEEK_API_KEY` (loaded from repo `.env` first), hits the real API, harvests the produced `session.jsonl` (the replay source AND the expected-log artifact), and `--update`s the stdout golden in one pass. Both forward a scenario filter. A missing fixture in replay **fails loud** with a "record first" message rather than self-skipping (the e2e self-skip rule is a CI-secret accommodation, not appropriate here — a committed-fixture test that silently vanishes is a coverage hole). A no-model scenario's `session.jsonl` simply has no `assistant/chunk` events (empty derived script); fail-loud still applies if a model call happens with no entry. An orphan-fixture guard test fails on a golden/fixture not referenced by any scenario (Vitest does not prune orphaned raw goldens), and a per-kind required-fixture guard asserts each scenario ships exactly the files its kind needs (`input.json` + `stdout.golden.jsonl` + `session.jsonl` for ALL scenarios — the harness passes `<dir>/session.jsonl` to `llm-replay` unconditionally, so even a no-model scenario needs its header-only fixture or `loadReplayScript()` fails; `replay.override.json` additionally for authored model scenarios).
|
||||
`pnpm run test:snapshot` runs replay (keyless) and is composed into the default `pnpm run test` gate so every PR gets the regression check (the main `vitest.config.ts` include stays narrow; the gate is `test && test:snapshot`). `pnpm run test:snapshot:record` requires `DEEPSEEK_API_KEY` (loaded from repo `.env` first), hits the real API, harvests the produced `session.jsonl` (the replay source AND the expected-log artifact), and `--update`s the stdout golden in one pass. Both forward a scenario filter. A missing fixture in replay **fails loud** with a "record first" message rather than self-skipping (the e2e self-skip rule is a CI-secret accommodation, not appropriate here — a committed-fixture test that silently vanishes is a coverage hole). A no-model scenario's `session.jsonl` simply has no `assistant/chunk` events (empty derived script); fail-loud still applies if a model call happens with no entry. An orphan-fixture guard test fails on a golden/fixture not referenced by any scenario (Vitest does not prune orphaned raw goldens), and a per-kind required-fixture guard asserts each scenario ships exactly the files its kind needs (`input.json` + `stdout.golden.jsonl` + `session.jsonl` for ALL scenarios — the harness passes `<dir>/session.jsonl` to `llm-replay` unconditionally, so even a no-model scenario needs its header-only fixture or `loadReplayScript()` fails; `replay.override.json` exactly for the scenarios whose table entry sets `overridden` — required with the flag, forbidden without it, because the harness forwards the sidecar purely on file existence and an unregistered stray would silently replace the derived script).
|
||||
|
||||
## Alternatives considered
|
||||
|
||||
|
||||
@@ -1,79 +0,0 @@
|
||||
# RFC: Repeat-tool-call guard plugin
|
||||
|
||||
Status: proposed
|
||||
|
||||
## Problem
|
||||
|
||||
A model stuck in a loop re-issues the same tool call with byte-identical arguments — re-running a failing grep, re-reading an unchanged file, polling a command that already gave its answer — and each round trip burns tokens, wall-clock, and (for paid APIs) money without adding information. The harness has nothing that notices: the loop has no step budget, no plugin tracks call repetition, and the model only escapes when it happens to vary its own behavior. The failure mode is real and cheap to detect — [pi-repeat-tool-guard](https://github.com/Kingwl/pi-repeat-tool-guard) ships exactly this as a pi coding-agent extension: count consecutive identical calls and, past a threshold, append a `<system-reminder>` telling the model to stop repeating itself and change course.
|
||||
|
||||
The harness already has every seam the pi extension uses, and better ones: [the interception-seams RFC](../../implemented/feature/2026-06-30-interception-seams.md) gives `tools/post-execute` a sanctioned way to attach model-facing context to a finished call, the loop buffers and injects that context with call/result adjacency preserved, and injected context is a logged `context/message` — so a native guard satisfies the model-visible ⟺ logged rule with no new session event. What is missing is only the plugin itself.
|
||||
|
||||
## Proposal
|
||||
|
||||
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 shape: one new leaf plugin package, `@deepseek-ai/dsh-repeat-tool-guard` at `packages/guard/repeat-tool-guard/`, opening a `guard/` group for loop-hygiene plugins (single-package groups have precedent: [the todo-write RFC](../../implemented/feature/2026-06-29-todo-write-tool.md) shipped `todo/tool-todo`). The plugin registers three listeners via `ctx.effect()` 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.
|
||||
|
||||
- **`tools/post-execute` (waterfall)** — the one detection point. The listener receives `(exec, result)` together, so counting and reminder delivery need no cross-event pending map (the pi extension needs one only because its `tool_call`/`tool_result` hooks are separate events). It always delegates via `next()` and, when a threshold is hit, folds a reminder onto the downstream decision's `additionalContext` — the observe-and-enrich posture [the hooks bridges](../../implemented/feature/2026-06-30-hook-bridges.md) already use, honoring the waterfall contract. Counting happens here rather than in `tools/pre-execute` because post-execute also runs for denied calls (`ToolRegistry.execute` routes a deny through the same pipeline), and a model hammering a denied call is exactly the loop worth breaking.
|
||||
- **`agent/prompt-submit` (waterfall)** — pure reset hook: delegate via `next()`, clear the submitting agent's chain. A user interjection changes the context; repetition across it is not a loop.
|
||||
- **`agent/status` (emit)** — on `disposed`, drop the agent's state, bounding the maps over harness lifetime.
|
||||
|
||||
### Detection semantics
|
||||
|
||||
The chain key is `(tool name, canonical arguments)`; a call identical to the previous tracked call increments the agent's consecutive counter, a different tracked call resets it to 1. Canonicalization is a deep key-sort plus `JSON.stringify`: `ToolExecution.arguments` is by construction the loop's `JSON.parse` output (or the raw string fallback for malformed argument JSON, which is itself a comparable value), so the pi original's bigint/circular/`undefined` handling has no inputs here and is deliberately dropped.
|
||||
|
||||
Two deliberate rules, both documented in the package README because they are behavior a reader would otherwise guess at:
|
||||
|
||||
- **Untracked calls are transparent to the chain.** A call excluded by `include`/`exclude` neither increments nor resets the counter, so `grep X → todo_write → grep X` still counts as two consecutive `grep X` when `todo_write` is excluded. This is what makes exclusion useful — bookkeeping tools interleaved into a loop must not launder it — and it is the pi extension's (undocumented) semantics, kept on purpose and written down.
|
||||
- **Calls without an agent are ignored.** A direct `ctx.tools.execute()` caller (tests, future non-loop consumers) has no model to remind and no `AgentId` to key on.
|
||||
|
||||
### 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 already 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, 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 folds content following the shared-merge precedent in `dsh-hook-protocol`.
|
||||
|
||||
### Config
|
||||
|
||||
```yaml
|
||||
- id: repeat-tool-guard
|
||||
name: '@deepseek-ai/dsh-repeat-tool-guard'
|
||||
config:
|
||||
thresholds: [3, 5, 8] # default; consecutive counts that trigger a reminder
|
||||
include: [] # tool-name patterns to track; empty ⇒ all tools
|
||||
exclude: [todo_write] # tool-name patterns transparent to the chain
|
||||
```
|
||||
|
||||
`thresholds` is validated at load and throws on an empty list, a non-integer, a value below 2, or a duplicate — misconfiguration fails loud, replacing the pi original's silent fall-back to defaults. `include`/`exclude` entries support `*` wildcards. Patterns are predicates over whatever tools exist at call time, not references to a registry entry, so an entry matching no currently registered tool is NOT an error — unlike `toolOrder`'s referent check, `exclude: [mcp_*]` must stay valid in a deployment that loads no MCP tools.
|
||||
|
||||
### Testing
|
||||
|
||||
Coverage named at plan time, per tier: **unit** — counting/reset semantics (identical, different-tracked, untracked-transparent, prompt-submit reset, disposal cleanup, per-agent isolation), canonicalization, threshold escalation including the `thresholds[0]` gentle-text rule, config fail-loud cases, and the fold-onto-downstream-decision path, to per-file 100% like every `packages/*/*/src` file. **Snapshot** — one scripted-replay scenario where the model repeats a call to threshold and the reminder `context/message` appears in the transcript, pinning the model-visible text and its envelope (this is a transcript-surface change; the ACP snapshot suite is the tier that owns it). **e2e** — none: the plugin is provider-independent and deterministic, and forcing a live model to repeat a call three times is not a stable test; the seam contracts it relies on are already e2e-covered by their owners.
|
||||
|
||||
## Alternatives considered
|
||||
|
||||
- **Append the reminder into the tool result** (`accept` with replaced `content` — the pi extension's mechanism, which patches result content because that is the only channel its API offers) — rejected: it makes the logged `tool/result` lie about what the tool returned, and `additionalContext` exists precisely as the separate sanctioned channel for post-execute commentary, with loop-level buffering that preserves call/result adjacency.
|
||||
- **Count in `tools/pre-execute` with a pending-reminder map** (the pi two-phase shape) — rejected: post-execute alone sees `(exec, result)` together and also fires for denied calls, so one listener with no cross-event state covers strictly more attempts with less machinery.
|
||||
- **Escalate to `block` at the highest threshold** — rejected for the initial scope: a blocked call punishes legitimate identical repeats (polling a long-running terminal, re-checking a file the agent expects to change), and an advisory reminder keeps the model in control. Revisit with evidence; the decision shape (`PostToolDecision`) already supports it.
|
||||
- **A per-deployment external hook via the CC/Codex bridges** (a `PostToolUse` script) — rejected as the answer: it works today for one deployment, but a shipped, unit-tested, `cordis.yml`-configurable plugin is the harness-native form, without per-call subprocess cost.
|
||||
- **A loop-level step or repetition budget in `agent-loop`** — rejected: "plugins, not loop changes"; a hard step budget is a blunter, orthogonal control that would need its own proposal.
|
||||
- **Fuzzy/near-identical detection** (normalized paths, similar-but-not-equal arguments) — rejected: exact match after canonicalization is cheap, deterministic, and explainable to the model; similarity thresholds invite false positives and need evidence before they earn complexity.
|
||||
- **Placing the package in `core/`** — rejected: core is the product spine; a behavioral guard is an optional leaf plugin, and the `todo/` precedent is a small dedicated group per plugin family.
|
||||
|
||||
## Acceptance criteria
|
||||
|
||||
- `packages/guard/repeat-tool-guard/` exists, registers all listeners through `ctx.effect()`, and is loadable from a `cordis.yml` with the config above; the config catalog regenerates with its entry.
|
||||
- Invalid `thresholds` (empty, non-integer, `< 2`, duplicate) throw at plugin load.
|
||||
- Unit suite covers the semantics list above at per-file 100%; a snapshot scenario replays a threshold-crossing repetition and pins the reminder `context/message` in the transcript on macOS and Linux.
|
||||
- The reminder is reconstructable from the session log alone (it is an ordinary `context/message` with a plugin source — no new session event).
|
||||
- The package README opens with the plugin's purpose — an advisory loop-breaker that is not a model-facing tool, never blocks or rewrites a call, and only injects reminders — then documents the transparency rule, the per-agent keying, and the in-memory-only state; `doc-sync` is green.
|
||||
|
||||
## Risks
|
||||
|
||||
- **False positives on legitimately repeated calls.** Idempotent polling patterns repeat identical calls on purpose; the reminder is advisory and thresholds/`exclude` are the pressure valves, but a badly tuned deployment adds noise to the transcript. Mitigation: conservative defaults and the reminder text explicitly allowing "finish the task if enough evidence has been gathered".
|
||||
- **Reminder tokens are model-visible cost.** Each trigger appends a paragraph to the next request; thresholds bound the frequency, but a pathological agent can hit 3/5/8 repeatedly across different keys.
|
||||
- **State is in-memory only.** A session resumed from persistence starts with a fresh chain, so a loop spanning a resume gets 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.
|
||||
- **Multiple context producers on one call.** When a hook bridge and the guard both attach `additionalContext`, ordering follows listener registration order; the fold keeps both, but the combined envelope's readability depends on merge behavior that this RFC inherits rather than owns.
|
||||
|
||||
## Open questions
|
||||
|
||||
- Should compaction reset chains? A compacted history changes what the model sees, but the repetition risk usually survives compaction; the initial answer is no.
|
||||
- Should subagents inherit the parent's thresholds via config only, or ever share chain state? Per-agent isolation is the proposed default; sharing looks like a smell until a concrete case appears.
|
||||
Reference in New Issue
Block a user