[ranxianglei]

Implements the kernel-owned output-side steering decision logic specified in #355 (owner-decided scope from billion-context#1093). The kernel decides; adapters land decisions onto wire fields — no protocol knowledge enters the kernel.

What's in it

New root module src/output-steering.ts (+ barrel exports, README/DESIGN docs):

1. Turn classifier — classifyTurn(messages): TurnKind
   Pure structural read of the last user message's block composition (new_user_ask / mechanical_continuation / error_continuation / unknown); no content regex. The wire→structure normalization contract is documented in the module JSDoc; each adapter maps its protocol's tool-result shapes into StructuralBlock (bili keeps its per-wire classifiers as thin normalizers).
2. Verbosity levels L0–L4 — VERBOSITY_LEVELS, verbosityDirective(level), renderSteeringBlock, applySteeringToPrompt
   Byte-stable directive texts (frozen baseline; tests assert the exact literals). L0 = no injection, default L2. Placement is replace-in-place or append-at-tail — never prepend (prefix-cache invariant); same-level re-application is a no-op (idempotence tested, incl. the compress-retry double-pass path).
3. Clamp-only effort routing — EFFORT_LADDER, clampEffortToFloor(current, floor = "low"): EffortValue | null
   Never raises; returns null when the value is absent / unrecognized / at-or-below the floor, so adapters inject nothing (client-absent fields are never injected; the kernel does not know protocol field names by design).
4. Entry point — decideOutputSteering(messages, config?): { turnKind, verbosityLevel, lowerEffort }
   Config { enabled, verbosityLevel, effortRouting }, default OFF; resolveOutputSteeringConfig tolerates malformed input with warnings. Per-provider route overrides remain an adapter concern.

Behavior deltas vs the bili reference (ranxianglei/billion-context#1096)

1. Unrecognized effort strings are left untouched instead of being forced to "low". The reference forced any non-low string down to low, which can raise unknown sub-floor values — a latent clamp-only violation. Known-ladder behavior is identical; xhigh was added to the ladder (the reference hit it via the force path — same outcome).
2. Numeric budget floors stay adapter-side (Anthropic thinking.budget_tokens → 1024, Google thinkingBudget → 128, dynamic -1 untouched): they are wire-field properties, not decision properties.
3. Default sentinel <acp_output_steering> differs from bili's <bili_output_steering>. When migrating, bili must pass its own sentinel to renderSteeringBlock / applySteeringToPrompt, otherwise sessions pinned under bili's block would accumulate two steering blocks (cache bust). Documented in the module JSDoc.

Verification (rebased onto current origin/master, v0.0.82 base b30723f)

• npm run typecheck — clean
• npm test — 904 pass / 0 fail, including the new 429-line regression suite that ports all applicable cases from the bili reference (byte-exact L1–L4 text anchors, idempotence/retry paths, clamp matrix, config resolution)
• npm run build — green
• New source files are prettier-clean (pre-existing repo-wide drift left untouched; CI has no format gate)

Follow-up after merge + release

Per #355: confirm npm view acp-kernel version shows the new release → billion-context bumps the exact pin → release → replace the duplicated decision logic in PR#1096 with kernel calls (wire-side landing + its own sentinel stay in bili).

Closes #355

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-23_output-steering (7d9487e)

Option A — Install from npm PR tag (recommended)

In your adapter project:

npm install acp-kernel@pr-356

Each push to this PR publishes a new version under the pr-356 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf acp-kernel-pr356.tgz
npm install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Reviewing PR #356 now: diff cleanliness check, full pre-flight (typecheck / test / build), and a regression review of existing behaviors touched by the diff. Findings will follow as a comment on this thread; small nits will be fixed directly on the branch.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review of PR #356 complete — findings below.

Diff cleanliness (first gate): PASS

• Single commit 7d9487e on current origin/master (b30723f, v0.0.82) — no stale base.
• 5 files, +778/−0, purely additive: src/output-steering.ts (new, 320 ln), tests/output-steering.test.ts (new, 429 ln), barrel exports in src/index.ts, one doc line each in README/DESIGN. No unrelated changes, no whitespace churn, no lockfile/generated churn, no version bump. New files are prettier-clean. Zero-runtime-deps preserved (the module has zero imports).

Pre-flight at PR head: all green

• Local: npm run typecheck clean · npm test 904/904 pass · npm run build green.
• CI: test (22), test (24), pr-validation, build-artifact — all success.

Spec conformance (#355) — verified against the spec AND the bili reference (PR#1096)

• classifyTurn: pure structural, no content regex, conservative unknown; rules match the spec and bili's four per-wire classifiers, incl. the Responses sets (outputs: function/custom_tool/local_shell/apply_patch *_call_output; neutrals: message/function_call/custom_tool_call/local_shell_call/apply_patch_call/reasoning; user-signal definition).
• Verbosity L1–L4 texts: byte-identical to the bili reference (lengths 123/262/696/392) — frozen-baseline requirement met; sessions pinned under these texts stay cache-stable through migration. L0 = no injection, default L2.
• Placement: tail-only, never prepends; same-level re-apply is a byte-stable no-op (incl. compress-retry double-pass); level change replaces in place without duplication; malformed unclosed sentinel consumed, not duplicated — all tested.
• Clamp-only routing: never raises; absent/unrecognized → null (adapters inject nothing); ladder minimal<low<medium<high<xhigh; fires iff mechanical_continuation.
• Config: default OFF; strict enabled === true; effortRouting on-by-default-when-enabled; malformed verbosityLevel falls back with warning while omitted takes the default silently; non-object → pure defaults, no warnings.

Disclosed deltas vs the bili reference — all three verified accurate

1. Unrecognized effort strings left untouched: confirmed bili forces any non-low/non-minimal string (OpenAI/Responses) or any non-low string (Anthropic) to "low" — coercing an unknown sub-floor value would be a raise, violating clamp-only. Kernel's leave-alone is the stricter reading; known-ladder outcomes identical (incl. xhigh→low).
2. Numeric budget floors stay adapter-side: confirmed present in bili's lowerEffort (Anthropic 1024, Google 128, dynamic -1 untouched) and correctly absent from the kernel.
3. Default sentinel differs: kernel default is `` (verified byte-wise) vs bili's own sentinel; the "pass your own sentinel on migration" note is in the module JSDoc.

Regression review — existing behaviors touched by this diff

Touched path	Carried behavior	Verdict
Barrel src/index.ts	existing public API surface	Preserved — purely additive exports, no name collisions (typecheck clean), no signature changes
processTurn pipeline, nudges, wire codec, prompts, block model, ref system	§7.3 kernel contracts	Untouched — new module is standalone, zero imports; no node ordering, config-schema, or state-shape changes
Existing defaults / output format	shipped behavior	Preserved — feature is default-OFF; nothing changes until a host opts in

Verdict: no regressions; this is a purely additive feature.

Minor find (non-blocking, reported in-thread per policy)

README's Status line still says "23 source modules" while master already carried 38 root-level src/*.ts files before this PR (39 after) — pre-existing doc drift, not introduced here; left untouched to keep the diff additive. Say the word if you want it fixed separately.

Merge gating

This is a feat: adding new public API surface → human-gated per AGENTS.md §7.4 (no auto-merge). And agents can't merge regardless — please merge yourself: #356

---

中文摘要:审查了 PR #356(输出侧 steering 决策逻辑入内核)——diff 纯增量、本地与 CI 全绿、对 bili 参考实现的三处行为差异均属实且更严格(clamp-only 无升档风险)、文案字节级一致、不触碰任何既有行为与内核契约,可以合并(feat 类按规则需人工合并)。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Addendum (second minor find, non-blocking): running npm install on any branch locally modifies package-lock.json by 2 lines — its "version" field still says 0.0.77 while package.json is at 0.0.82. This is pre-existing on master (release commits touch only package.json per the release workflow), so every local install shows that 2-line diff until reverted; it's expected churn, safe to discard, and I left it out of the review scope accordingly. Worth a one-line lockfile sync at some future release if you want to stop the recurring dirt.