Repository navigation
v3.5.0
3.5.0 is about lanes. Until now every model lane — primary, smart, fallback, specialists — was effectively a clone of the primary's request shape, and "does this need a human?" lived nowhere outside the model's prose. This release makes each lane independently configurable, makes structured output on the Anthropic wire honest about what it actually enforces, and adds a first-class human-review triage that can route to the right reviewer deterministically.
It's also the biggest groundwork release yet for the self-hosted controller: the Forgejo ingress, the durable review queue, and the Tangled read seam all landed as libraries. None of it is wired into the Action runtime — Action users should notice exactly nothing there.
If your lanes run against different backends, this is the release to pin:
- uses: misospace/pr-reviewer-action@v3
with:
ai-base-url: ${{ vars.AI_BASE_URL }}
ai-api-key: ${{ secrets.AI_API_KEY }}
ai-model: ${{ vars.AI_MODEL }}
ai-response-format: json_schema # base default
ai-smart-response-format: json_schema # smart lane keeps structured output
ai-fallback-response-format: off # fallback model rejects the parameter
ai-specialist-reasoning-effort: high # specialists think harder than primaryModel lanes stop being clones
- Response format is lane-aware (#1040, closes #1009):
ai-primary-response-format,ai-smart-response-formatandai-fallback-response-formatoverrideai-response-formatper lane; unset lanes inherit, so single-value workflows stay request-byte compatible. The motivating shape: a primary behind a proxy that rejectsresponse_formatno longer forcesoffonto every other lane, and the reviewer-requested smart re-review (#721) can opt into richerjson_schema. - Reasoning effort is profile-aware, on both wire families (#1049, closes #998):
ai-reasoning-effortgains per-lane overrides (ai-primary-…,ai-smart-…,ai-specialist-…,ai-fallback-…). OpenAI-compatible requests sendreasoning_effort; Anthropic requests sendoutput_config.effort— but only under that lane's own*-anthropic-effort-supportassertion, never inherited from another lane. The fallback lane inherits the base effort on purpose (#965: availability recovery, never quality escalation) yet asserts its own endpoint, becauseai-fallback-base-urlmay be a different backend entirely. - Anthropic structured output got truthful (#1047): previously
ai-response-format: json_schemaonai-api-format: anthropicwas accepted and then silently omitted — operators believed the schema was enforced when it was only prompted. Now Anthropic-format requests carry no structured-output parameter unless the lane's*-anthropic-structured-output-supportopt-in is set, the run logs what is actually enforced, and if a wrong assertion draws a 400 naming the parameter, the request retries once without it so the review degrades to prompted JSON instead of failing. - Specialists survive model misbehavior: streamed Anthropic overruns are detected instead of accepted, completion-limit replies with zero leads are retried, malformed no-lead responses classify as parser errors (retryable) rather than fatal, overrun-recovery retries are reported only when actually adopted, and the specialist completion budget is now a real input —
deep-review-max-tokens, default16384, independent ofai-max-tokens— enforced across every surface. - Smart-tier limits are yours to set (#1032, closes #1020):
smart-tool-max-roundsandsmart-tool-loop-wall-clock-secbound only the post-review escalation rerun, leaving the primary loop untouched. Empty inherits the previous scaling behavior.
Ask a human, deterministically
- Opt-in triage (#1051, closes #1003):
review-triage: truemakes the reviewer answer a tri-state "does this PR need specialized/owning human review?" —not_required,required, orunknown, fail-closed tounknownwhen the answer is missing or malformed. It's advisory only: it never moves the verdict, findings, or any tool/access gating, and with the input off the outputs stay empty — "did not run" is never confused with "no work". - Deterministic routing (#1005): with
review-triage-routing-mode: request, a triage ofrequiredadditively requests reviewers. The model can only pick logical target ids declared in the trusted base-ref.github/pr-reviewer-routing.yml; the action maps those ids to concrete users/teams, requests are idempotent and never remove existing reviewers, and an unknown target or any uncertain state mutates nothing. Never on forks.advisorymode publishes the routing plan without touching the forge, for safe dogfooding. (GitHub-only; Forgejo degrades torequest_failed.)
# .github/pr-reviewer-routing.yml (read from the base ref)
targets:
platform:
users: [alice]
teams: [platform-team]
security:
teams: [security]Provenance you can check
- Source-claim blockers verify against the exact head (#1033, closes #1016): a finding whose blocker rests on repository source is now verified against the source at the reviewed revision, not a stale or sanitized stand-in.
- Provenance rides through tool output (#1029, closes #1015): source evidence the model sees through native tools carries the same committed-vs-sanitized provenance discipline as the corpus, so a citation the reviewer makes is checkable against what it actually read.
- Non-code criteria stop falsely blocking (#1062): requirement-trace acceptance criteria that demand no code change no longer block approval.
- Publish fails closed on bad data (#1079, #1050): malformed requested-reviewers responses abort the publication path rather than guessing, and superseded managed inline comments are retired safely instead of piling up on long-lived PRs.
- No-diff runs can't claim coverage (#1031, closes #1026): partitioned reviews fail closed when the no-diff path would have reported coverage it never gathered, and outcomes are bound to the plan round that produced them.
- Fingerprints stay honest: changing the effective triage routing mode (#1005) or leaving a fallback effort unset now changes the review fingerprint, so config edits trigger a fresh review instead of riding a stale skip.
Forgejo
- CI status discovery works on Forgejo v16 (#1046): the review job no longer times out against the v16 API.
authorized_integrationauth fixed (#1045), plus a dedicatedforgejo-tokeninput so the Forgejo token can differ from the GitHub-compatible one, and aforgejo-auth-methodinput to pick between token and OIDC authorized-integration modes.
Toward the self-hosted controller
Shipped as libraries, exercised by tests, consumed by controller work ahead — the normal Action runtime doesn't touch any of it yet, and that's deliberate.
- Forgejo ingress (#1019, closes #730): webhook signature verification before parse (fail closed), native-vocabulary projection, and missed-webhook reconciliation — the daemon's front door, with the contract details in
docs/architecture/self-hosted-forgejo-ingress.md. - Durable review queue (#1021, closes #731): 8-state job records, settle/debounce windows that collapse bursts, at-most-one-active-generation per PR, head supersession, lease/heartbeat dispatch with bounded retry — the scheduler's brain. (
docs/architecture/durable-review-queue.md) - Tangled read seam (#1013, #1012, closes #585, #586): pull metadata normalized into the platform seam, and the latest pull round fetched with patch-blob diff/files normalized into review inputs. Groundwork — Tangled review support is not complete yet.
- One forge capability for Python tooling (#1039, closes #1041): every Python tool's reads and writes route through
pr_reviewer.forge— one URL/token resolution, a redirect-safe transport that never copies theAuthorizationheader onto a redirect, and an HTTPS-only base-URL guard that raises before any byte hits the wire. Thegh apierror-echo surface is gone with it. - Reproducible eval lanes (#1022, closes #967): heterogeneous model/profile lanes for the eval harness, so A/B results across different models reproduce.
Upgrading
Pin to v3.5.0, or the floating v3 tag, which now follows this build. Immutable-version and SHA pins must be bumped deliberately.
Nothing breaks and no input changed meaning. The honest notes:
- All new inputs are opt-in with empty/false defaults, and leave request bytes unchanged when unset. The one to be deliberate about: on Anthropic-format lanes,
ai-response-format: json_schemawas never actually sent before — it only looked like it was. If you want native enforcement, setai-anthropic-structured-output-support(and the lane variants) only after confirming the endpoint acceptsoutput_config.format. - Specialist completions now default to a
16384cap (deep-review-max-tokens), independent ofai-max-tokens; deep-review runs may spend more tokens per role. - Requirement-trace users: non-code acceptance criteria no longer block approval (#1062) — one behavior changed in the direction of blocking less.
Every input's current contract lives in docs/inputs.md.
Full changelog: v3.4.1...v3.5.0
Generated changelog (release-please)
Features
- eval: reproducible heterogeneous model/profile lanes for #967 (#1022) (492c743)
- forge: route Python tooling through a configured forge capability (#1041) (#1039) (c235b7d)
- ingress: self-hosted Forgejo webhook/auth/reconciliation adapter (#730) (#1019) (c16c85b)
- model: allow response-format overrides per model lane (#1009) (#1040) (20cab2d)
- model: independent ai-fallback-reasoning-effort override (#1053) (cd20d43)
- model: independent ai-fallback-reasoning-effort override (#1053) (51cad39)
- model: per-profile reasoning effort for OpenAI and Anthropic (#998) (#1049) (01dda91)
- queue: durable review-queue scheduling semantics (#731) (#1021) (e9582f5)
- tangled: fetch latest pull round and normalize patch-blob diff/files (#586) (#1012) (e7db913)
- tangled: normalize pull metadata into the platform read seam (#585) (#1013) (aa12fff)
- triage: add first-class human-review triage output (#1003) (#1051) (9cf25c8)
- triage: route required human review to trusted reviewer targets (#1005) (bcdda85)
- triage: route required human review to trusted reviewer targets (#1005) (3d4ec70)
Bug Fixes
- ci: repair dogfood regressions and retire obsolete Python test debt (#1060) (afd5013)
- classification: treat self-location readlink as trusted path bookkeeping (#1034) (#1042) (92424f9)
- config: preserve unset fallback effort in hash (c6630d7)
- deps: update dependency @types/node (24.19.1 → 24.19.2) (#1023) (e44a237)
- enforcement: confirm documentation trace rows only from their named target source (86229e0)
- enforcement: harden documentation scan per independent review (259de11)
- forgejo: Fix CI timeout on V16 API (#1046) (8350071)
- forgejo: Fix Forgejo authorized_integration mode (#1045) (6b38cbe)
- ingress: bind reconciliation jobs to the configured repository (#1025) (#1052) (d1cff53)
- model: declare review_triage_targets in strict OpenAI verdict schemas (aec5111)
- model: make Anthropic structured-output semantics truthful and recover usable verdicts (7a7e130)
- model: make Anthropic structured-output semantics truthful and recover usable verdicts (816927b)
- model: mask configured secrets before diagnostic truncation; specialist response-format via tierResponseFormat (f533d65)
- model: report the effective wire format after a 400 strip recovery; CodeQL cleanup (f183288)
- partition: fail-closed no-diff coverage and bind outcomes to plan round (#1026) (#1031) (55ecf00)
- precheck: fingerprint the effective triage routing mode (20ff2e2)
- publish: fail closed on a malformed requested-reviewers GET (#1079) (5738670)
- publish: safely retire superseded managed inline comments (#1050) (cc1fa07)
- requirement-trace: stop non-code acceptance criteria falsely blocking approval (#1062) (50bf601)
- requirement-trace: stop non-code acceptance criteria falsely blocking approval (#1062) (8f93a01)
- reviewer: carry source-evidence provenance through model-visible tool output (#1015) (#1029) (afe282f)
- specialists: apply overrun recovery to the combined-scout path (6fcb154)
- specialists: classify malformed/no-lead responses as parser (2bfa3ca)
- specialists: detect Anthropic overruns, raise cap (4abbfb1)
- specialists: detect streamed Anthropic overruns and raise the default cap to 16K (d60da4a)
- specialists: only report overrun recovery when the retry is adopted (d54ab2b)
- specialists: retry completion-limit responses with zero leads (5d23b68)
- tools: make smart-tier loop round and wall-clock limits settable (#1020) (#1032) (f5b3e11)
Performance Improvements
- ci: stop tests paying the real transport backoff (8b8f0d1)
- ci: stop tests paying the real transport backoff (4fd2a05)
Chores
- test: drop unreferenced dogfood-623 and test-evidence artifacts (496e2d0)
- test: drop unreferenced dogfood-623 and test-evidence artifacts (e1d1ff8), closes #1063
Documentation
- agents: align normative guidance with shipped v3 action (#1055) (34d68f1)
- ci: record the observed after-CI numbers (67d2f16)
- requirement-trace: pin non-code proof shapes in contract doc (bf6a2b2)