Skip to content

v3.6.0

Choose a tag to compare

@its-miso its-miso released this 18 Oct 19:12
· 431 commits to main since this release

3.6.0 is about closing the loop. Reviews used to be fire-and-forget: the action published findings and never learned whether a suggestion helped. This release makes findings actionable — line-exact suggestions you can apply in one click, and a usefulness signal (a 👎) that keeps a rejected prescription from coming back. Operators can stamp every model request with a per-run correlation id so proxy cost logs group one run's calls, and repositories can teach the reviewer bounded, repo-local skills.

Underneath, it's a hardening release for the write path: only proven-trusted identities may touch the action's comments, webhook event hints must agree with the signed payload, and requirement-trace stops manufacturing blockers the model never substantiated.

If any of the first three features matter to you, 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-extra-headers: |
      x-litellm-session-id: ${{ github.run_id }}

Interpolate github.run_id in the calling workflow — a literal GitHub expression here fails at action load time.

- uses: misospace/pr-reviewer-action@v3
  with:
    repository-skills: true

Findings you can act on

  • Findings carry apply-ready suggestions (#1101, closes #762): a finding may now include an exact line/range replacement, rendered as a native one-click suggestion block on GitHub (degraded cleanly on Forgejo), plus a concise fix prompt for coding agents. A 👎 from a trusted collaborator suppresses that exact prescription on later runs — range-exact, keyed to the published text, withhold-only: verdict, severity, and thread handling are untouched, and the action's own reactions never count. Untrusted reactions stay telemetry. (#1110, closes #1107: model-authored text can no longer forge a suggestion opener, and one shared predicate decides whether a finding carries a prescription.)

Trace a run, teach the reviewer

  • ai-extra-headers rides on every provider request (#1092, closes #1037): multi-line Name: value pairs merged into every lane — primary, smart, fallback, specialist — and both wire formats. Malformed lines, duplicates, and collisions with transport headers (Authorization, x-api-key, accept-encoding, …) are rejected at startup. The input is treated as a secret end-to-end: values are redacted from config dumps and masked out of provider error and response surfaces, even when an endpoint echoes a header value back in an HTTP 200 body (#1094). Fingerprinting is name-only for ephemeral correlation headers (-session-id, traceparent, …), so a per-run id never invalidates a cached review; any other header participates by name plus a one-way digest of its value.
  • repository-skills opts into repo-local guidance (#1081): discover bounded .agents/skills/*/SKILL.md files from the trusted PR base. In native-loop mode the model reads relevant skills on demand; with tools off, repository-skill-path includes one operator-selected skill as a bounded corpus addendum. Never reads PR-head content, disabled on fork PRs, and enabled state feeds the precheck config hash so a re-review can't silently reuse an incompatible run.

Only trusted hands touch a comment

  • Forgejo owner identity is proven, not assumed (#1097, fixes #1093): ownership now binds a canonical user id — Authorized Integration audience or a reaction self-write probe — and every unproven state fails closed as before. A clean re-review also no longer publishes an empty "Inline findings…" review; cleanup anchors on the sticky comment's own publication checkpoint instead.
  • GitHub sticky upsert only patches your own comments (#1100, closes #1098): a foreign-authored marker comment is never overwritten, even when it's the newest match; when no owned target exists, unattributable comments fail closed to creation.
  • Trusted owner resolves once per publish run (#1109, closes #1099): the Forgejo reaction self-write probe fires once instead of twice per run.
  • Superseded advisory reviews minimize instead of piling up (#1129, closes #1114): bot-owned COMMENTED reviews left behind by incomplete re-runs are collapsed — presentation-only, idempotent, and Forgejo no-ops. Destructive cleanup (inline retirement, verdict dismissal, thread resolution) stays behind the unchanged complete-replacement gate.

The signed body is the event

  • Forgejo event hints must match the signed payload (#1103, closes #1024): a replayed delivery with a valid signature but a swapped event-header hint could previously mint a fresh review; hints are now honored only when compatible with the signed action.
  • Check-update kinds are shape-anchored (#1106, closes #1102): check_run / check_suite / status classifications must match the signed body's own shape, closing header-driven reclassification into schedulable check_update — without dropping real Forgejo deliveries that carry no pull-request fields.

Requirement-trace blocks less, explains more

  • Process/risk notes stop forcing CHANGES_REQUESTED (#1096, fixes #1095): advisory sections (Dependencies, Notes, Risk, Caveats, Limitations, Rollout) no longer enter the requirement ledger; test-execution outcomes classify as non-blocking external state; and an unmet requirement is disclosed in the rendered trace instead of being synthesized into blocking findings the model never substantiated. Blocking findings now come only from the model's own substantiated violations. Pinned by zero-inference golden replays, including the false-CHANGES_REQUESTED shape from miso-gallery#543 now approving.
  • Repair-pass outcome is visible (#1105, fixes #1044): the bounded requirement-trace repair pass records its outcome (skipped_by_gate, no_claims, errored, merged, …) on the run artifact and renders it. "The reviewer reported no trace" and "the repair pass ran and returned nothing" used to be indistinguishable; on a degraded primary they no longer are.

Provenance you can check

  • Related-code snippets carry their provenance (#1104, closes #1030): every related-code snippet is stamped committed_source or sanitized_source with a redaction count, on a harness-constructed header line — never inside the fence, where repository content could forge the annotation. Related code reads the working tree, so snippets carry revision: null and literal-claim authorization fails closed there until revision binding exists.

Groundwork: measuring the tool loop

  • tool_stop_reason lands in the metadata marker (#1122, refs #847): every harness run records why its tool loop stopped — the production signal #847 needs to compare tool-budget exhaustion across request sizes. The marker key appends last; unset serializes byte-identical. Also faster CI: semantic-eval fixture integrity is memoized and bad corpora fail fast before evaluation (#1111, closes #1078).

Upgrading

Pin to v3.6.0, 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 — and the review fingerprint — unchanged when unset. Treat ai-extra-headers values like API keys: they never appear in logs, artifacts, or error text.
  • Requirement-trace users (requirement-trace: true): one behavior changed in the direction of blocking less — advisory/process sections and test-execution outcomes no longer produce blockers, and unmet requirements are disclosed rather than synthesized into findings (#1095). Strict provenance checks on real code criteria are unchanged.
  • The metadata marker gains tool_stop_reason, appended last. Consumers parsing marker fields positionally should read it as an optional trailing key; when unset, the marker is byte-identical to 3.5.x.
  • Reviews may now carry suggestion blocks, and a trusted collaborator's 👎 suppresses that exact prescription on later runs. Suppression never touches the verdict, so a downvoted suggestion can't downgrade an approval.

Every input's current contract lives in docs/inputs.md.

Full changelog: v3.5.0...v3.6.0

Generated changelog (release-please)

Features

  • #1037: send ai-extra-headers on every AI provider request (#1092) (a640068)
  • #847: record tool_stop_reason in the metadata marker on every harness run (#1122) (c9a0c0e)
  • deps: update dependency astro (5.18.2 → 7.4.0) (#1091) (4e22d0d)
  • findings: actionable suggestions and durable usefulness feedback (#1101) (6db0b06)
  • reviewer: stamp provenance on related-code corpus snippets (#1104) (041c68d)
  • review: opt-in trusted-base repository skills (#1081) (154ead3)

Bug Fixes

  • #1024: bind Forgejo event hints to signed payload action (#1103) (b2fafa0)
  • #1037: mask configured extra-header values on provider response surfaces (#1094) (20d3dc9)
  • #1044: surface the requirement-trace repair-pass outcome in the artifact and rendered trace (#1105) (47fa8cf)
  • #1093: trusted Forgejo owner identity; no empty inline reviews on clean re-review (#1097) (82d4cbc)
  • #1098: GitHub sticky upsert only patches trusted-owner comments (#1100) (2e7adfb)
  • #1099: resolve trusted owner once per publish run (#1109) (37fccd2)
  • #1102: bind check update event kinds to signed payload shape (#1106) (82c862f)
  • #1107: reserve model suggestion openers and unify feedback precheck (#1110) (09a05ed)
  • #1114: minimize superseded advisory reviews on incomplete re-runs (#1129) (2376d50)
  • ci: follow the fork workflow rename (05fe106)
  • ci: follow the fork workflow rename (99836fd)
  • requirement-trace: stop process/risk notes forcing CHANGES_REQUESTED (#1095) (#1096) (584ca94)

Performance Improvements

Documentation