v1.4.0

Orca-Code-Review v1.4.0

Adds an optional precision-filter stage between the engine's raw findings and the merge gate.

New

• precision-filter input (default true) — when enabled, engine findings pass through two soft-fail stages before the gate:
  • L1: a deterministic filter that verifies each finding's existing_code snippet against the reviewed commit, re-homing or dropping findings whose snippet does not match the claimed path, then deduping by normalized content.
  • L2: an LLM judge (independent vendor from the reviewer) that clusters findings by root cause, drops clusters whose confidence falls below judge-threshold, and keeps one representative per surviving cluster.
• judge-model input (default deepseek/deepseek-v4-pro) — the model used by the L2 judge stage. Should differ from the reviewer model.
• judge-threshold input (default 0.5) — the L2 confidence gate (0–1). Lower keeps more findings; raise to be stricter.

Compatibility

The precision-filter: "true" default changes visible output for consumers on @v1. To preserve pre-v1.4 raw-engine output, set precision-filter: "false" in the workflow inputs. Both L1 and L2 are soft-fail — errors in either stage keep the prior stage's findings intact and never abort the review.

v1.3.0

Orca-Code-Review v1.3.0

Backward-compatible minor: adds a wall-clock guard on the review engine.

New

• timeout-minutes input (default 20). Wraps ocr review in GNU timeout --kill-after=10s: SIGTERM at N minutes, SIGKILL 10s later if the engine ignored the term. On timeout, the check fails closed with a distinct wall-clock timeout after Nm error — separate from the existing no usable result fail-close, so log readers can immediately tell which mode tripped and whether to bump the timeout or investigate the engine.
• Docs: README inputs row + a commented # timeout-minutes: "20" in the example workflow.

Behavior change

Consumers whose reviews previously took 20+ minutes will now fail closed at 20 minutes instead of running to completion (or eventual engine failure). The error message tells them to bump timeout-minutes in their workflow. Motivation: a large-PR + slow-model combo was observed running 43 minutes before failing closed on subtask timeouts; a hard ceiling makes the failure mode predictable (fail fast, clear message) and stops burning quota on runs that were going to fail anyway.

Notes

• Accepts decimals (e.g. "0.5" = 30 seconds) for testing.
• In exhaustive mode each engine pass has its own budget, so worst-case whole-review wall time = timeout-minutes × 3.
• No input/output signature changes beyond the new timeout-minutes; existing workflows keep working with the 20-minute default.

v1.2.1

Docs/hygiene: rename example workflow file from workflows/orcarouter-code-review.yml to workflows/orca-code-review.yml to match the shipped brand (Orca-Code-Review). Path text in README, RELEASE, and action.yml header updated to match. No action-code changes; existing installs at the old filename continue to work.

v1.2.0

Orca-Code-Review v1.2.0

Backward-compatible feature release.

New

• 👀 reaction on review requests. The action now posts a eyes reaction on the PR body (and on the trigger comment for a maintainer's /orca-code-review comment), so requesters get an immediate visible acknowledgement that the bot noticed the trigger. Static one-shot; failures are non-fatal (a reaction outage never blocks a review).

Rename

• Comment trigger /orcarouter-review → /orca-code-review in the shipped example workflow, README, SECURITY docs, and action input-description prose. Matches the shipped brand (Orca-Code-Review) and cleans up legacy naming from the pre-rebrand period.

Compatibility

• The trigger string is enforced by the consumer's own workflow file (startsWith(github.event.comment.body, '...')), not by action code. Existing workflows using /orcarouter-review keep working — the string in their if: gate still matches. If you copy the updated example workflow, the new trigger is /orca-code-review.
• No input/output changes; no engine-behavior changes.

Orca-Code-Review v1.1.0

Orca-Code-Review v1.1.0

Backward-compatible release. All new inputs default to their prior behavior —
existing workflows keep working with no changes.

Review quality

• Default review now runs a single gpt-5.5 tier; the cheap→strong cascade
  remains available and configurable via the recipe (#3)
• Calibrated P1/P2 severity and tightened precision rules to cut false
  positives on large PRs (#2)
• P3 severity tier added; review summary now pinned to the top of the PR
  description (#4)
• Feeds the repo's own conventions (AGENTS.md / CLAUDE.md / CONTRIBUTING.md)
  to the engine so project-specific choices aren't flagged as defects (#5)

Configurability (new inputs, all optional)

• auto-review-authors, max-diff-kb (512), max-diff-files (300),
  on-oversized-diff (fail), settings (true), report (true)
• Dashboard-driven settings, quiet mode, server-side rubric override, and an
  exhaustive review loop
• Retry/backoff proxy, oversized-diff guard, per-run report-back, and an
  edit-in-place summary comment

Security & hardening

• API key is passed via environment, never on argv (avoids process-list /
  CI-log leakage) (round-1 review)
• Fork-PR auto-review spend gate; author gate still applies when settings are
  disabled; policy-block stays fatal in exhaustive mode
• Project-conventions doc is read from the base revision only and gated to
  the default branch, framed as untrusted read-only data — a malicious
  AGENTS.md on a PR head cannot weaken the review or alter severity tags (#5)
• Proxy hardening: client-disconnect safety, upstream timeout, idempotency,
  crash/hang fixes; merge-gate integrity; settings authority
• Control-plane URLs preserve gateway sub-paths (self-hosted deployments)
• Added SECURITY.md: threat model, pull_request_target boundary, data flow,
  and key-hygiene guidance

Orca-Code-Review v1.0.0

First public release. An AI pull-request reviewer for GitHub Actions, powered by the OrcaRouter gateway. A cheap model screens every push; once it's clear of serious issues, a strong model does the final pass. Findings post inline, tagged by severity, with an optional gate that blocks the merge.

What you get

• Cost-tiered cascade — cheap screens, strong confirms, auto-promoted per PR. The router's DSL recipe picks the models; the action names none.
• Severity-tagged findings ([P0]/[P1]/[P2]), with untagged output failing safe.
• Inline comments plus one rolling summary edited in place — no comment pile-up.
• Merge gate that fails on P0/P1 (mark it required to enforce).
• Dashboard-driven settings, oversized-diff guard, and public-repo spend controls.
• Optional guardrail/firewall layer and privacy-safe run reporting (counts only — never your code).

Quickstart

- uses: Continuum-AI-Corp/orca-code-review@v1
  with:
    orcarouter-api-key: ${{ secrets.ORCAROUTER_API_KEY }}

Full setup in the README. Review engine: Open Code Review (Apache-2.0).