v0.15.3

Changed

• CI: the Gourmand job reports as Code Quality (Gourmand) / Code Quality (Gourmand) like the rest of the fleet, so the org ruleset can require it.
• Constitution is now a v1.18.0 manifest: purpose, CLI and workflow contract,
  external API, agents and ignored paths; fleet and profile rules apply by
  reference.
• Constitution validation pinned to v1.18.0 via constitution.yml, replacing
  the unpinned validation job in ci.yml.
• Dependabot auto-merges GitHub Actions minor and patch updates.

Fixed

• Reviewers no longer flag a version, tag or release as nonexistent. Every
  Dependabot bump to a major newer than the model's cutoff (checkout@v7,
  setup-python@v7, ...) drew a "tag does not exist" finding, repeated on
  each rebase.
• examples/gourmand.yml pinned gourmand.yml@v0.5.0 and was a bare job
  fragment. It is now a complete workflow on pull_request at the current
  release, and a test fails any example that pins a different release than
  the one it ships in (crunchtools/constitution#22).

v0.15.2

Fixed

• Replies with trailing content no longer abort an agent. Before this fix, GLM-5.3 replies that followed the findings array with another array or with prose failed as Extra data. Consecutive JSON values are now merged, and trailing text is ignored.

v0.15.1

Fixes the review job failing with PermissionError writing the step summary (#61, #62). Bump callers to @v0.15.1.

v0.15.0

Added

• Incremental re-review (#57). On a push to an open PR, review.yml
  passes --incremental: gatehouse reviews only the commits since its last
  complete review, fetched through the compare API with nothing checked
  out. It falls back to the whole PR for a first review, a force-push or
  rebase, when the last review had an unfinished agent, or when the compare
  fails. Files the PR does not touch are cut from the range, so merging the
  base branch in does not review the base's changes. The review summary
  states the scope.
• Serving model and usage per agent (#58). Each agent logs
  agent=… model=… prompt=… completion=… reasoning=… cost=$… fallback=yes|no
  to stderr, followed by a total line. In CI the same table goes to the job's
  step summary, and --usage-json PATH writes it as JSON.
• Fallback is visible. When a fallback model served any agent, the posted
  review says so, e.g. "3 of 8 agents served by google/gemini-3.1-flash-lite".

Changed

• Findings on a line outside the PR diff are not posted, and the summary
  counts them. GitHub rejects a whole review over one such comment.
• The posted review is pinned to the head commit that was reviewed, and a
  review in which every agent finished carries a hidden
  <!-- gatehouse review complete --> marker; incremental review starts
  from the newest one.
• call_model returns a Completion (text, requested and serving model,
  usage) instead of the bare text.

v0.14.0

LOW findings were 26% of the threads triage required an answer to, and a
30-finding sample from four crunchtools PRs held 18 noise, 6 re-raises of
answered threads and 1 with invented evidence (#56).

Added

• Findings with invented evidence are dropped, at every severity. The
  quoted evidence must appear in the diff (or the styleguide/constitution
  agents were given), ignoring whitespace, line-number prefixes and ...
  elisions. mcp-trentina#258 quoted def _cases(split: str, csv_bytes: bytes): for code that reads -> list[Case]:.
• Answered threads are not re-raised. On a PR, a MEDIUM or LOW from the
  same agent within 5 lines of a thread someone replied to is dropped.
  HIGH and CRITICAL always post: next to a "fixed in" thread they may be a
  regression.
• At most 5 advisory LOWs per review, highest confidence first. LOWs
  from the --required-low-agents list (Bug Hunter and Security Scan by
  default) are never capped.
• The review summary counts each kind of dropped finding.
• Each finding comment carries a hidden agent and confidence marker, so
  replies can be joined to confidence before deciding whether LOW needs a
  higher threshold.
• required_low_agents input on triage.yml and review.yml (default
  Bug Hunter,Security Scan), and --required-low-agents on the CLI. Give
  both workflows the same list.

Changed

• Triage requires an answer to LOW findings only from required_low_agents.
  Other LOWs still post, as advisory. CRITICAL, HIGH and MEDIUM are required
  from every agent, as before.
• The Documentation agent no longer asks for docstrings on private or
  non-exported names, or beyond a file's one-line-docstring-plus-type-hints
  convention; its low tier is gone.
• Consistency Check findings must cite the in-repo code that sets the
  convention, as file:line, or are not reported.

v0.13.0

Changed

• The Constitution agent reviews amendments as amendments. A diff that
  changes the constitution file itself is reviewed for rules that contradict
  each other or the code, not against the text it replaces, which it can
  only ever contradict. crunchtools/petit#77 and #78 each drew CRITICAL
  findings for changing the rule they were changing. Other files in the
  same diff are still judged by the base branch's constitution, so a PR
  still cannot relax a rule and use the relaxation at once. The agent is
  also told it does not know today's date, which made it flag same-day
  Amended: lines as stale.

v0.12.0

Added

• .gatehouse-ignore: gitignore-syntax paths that are never reviewed.
  Matching files leave the diff and the file listing before any agent runs;
  on crunchtools/petit#61 that cuts each agent's diff from 263 KB of deleted
  fingerprint data to 3 KB. A change is skipped only when every path it
  touches matches, changes to the ignore file itself are always reviewed,
  and in CI the file comes from the base branch so a PR cannot hide its own
  changes. The posted review says how many files were skipped (when every changed
  file is ignored, nothing is posted).
• Under a trusted base (GATEHOUSE_CONTEXT_REPO/REF), the styleguide,
  ignore file and auto-discovered constitution come only from the base; a
  missing base copy no longer falls back to the working tree. An explicit
  --constitution path still wins, as before.

v0.11.0

Added

• retriage.yml and examples/gatehouse-retriage.yml: a reply now clears
  the triage check. Triage re-ran on pull_request_review_comment, but that
  run's checks do not count toward branch rules: on #50, four passing reply
  runs left the PR blocked until the pull_request_target run's triage job
  was re-run by hand. The new workflow_run listener does that re-run
  automatically, from the default-branch definition. Repos that require
  Gatehouse triage should add the example file.

Security

• Answers #47. Because reply-event checks do not count toward the
  ruleset, a fork that rewrites its copy of the workflow cannot satisfy
  Protect workflows or Gatehouse triage on a reply. The re-run that does
  count uses the base definition.

v0.10.0

Changed

• An agent that cannot finish now exits 2. It used to print a warning and
  count as zero findings, so an OpenRouter outage or a garbled reply passed
  as a clean review (the pre-commit hook included). Findings from the agents
  that did finish are still printed and posted, and the posted review names
  the ones that did not. --advisory, and so the default review.yml, still
  exits 0 (#41).

Fixed

• --stdin crashed on httpx.ReadTimeout over a large diff. Timeouts and
  dropped connections are now retried with the same backoff as a 429 (#41).
• review.yml failed on diffs containing escape sequences. gh pr diff
  refuses such a diff unless told otherwise, printed nothing, and the job
  reported No changes to review. before exiting 1. The diff is now fetched
  with --allow-escape-sequences into a file, and a failed fetch says so
  (#44).
• The workflow guard missed renames. It read gh pr diff --name-only,
  which lists only a renamed file's new path, so a fork could move a workflow
  out of .github/workflows/ (deleting it) and pass. The guard in
  examples/gatehouse.yml and in this repo now reads both filename and
  previous_filename from the pull-request files API. Found by Gatehouse
  reviewing its own fleet rollout (RT #1507); repos that copied the example
  should re-copy it.
• A reply could clear a failed workflow guard. The guard ran only on
  pull_request_target, so on a pull_request_review_comment run it reported
  skipped, and a required check counts skipped as passing. It now runs on
  both events, and posts its explanation only once.

v0.9.0

Changed

• Gatehouse calls OpenRouter, not the Gemini API. The default model is
  openai/gpt-6-luna, with google/gemini-3.1-flash-lite as an automatic
  fallback when it is rate-limited or down. On a 33-diff replay of real
  crunchtools changes, judged blind by two models from other vendors, Luna
  caught as many reintroduced bugs as gemini-2.5-flash while nearly all of
  2.5 Flash's findings on clean PRs were noise, at about a sixth of the cost
  per review (RT #1505).
• Every request requires zero data retention and forbids training on prompts.
• --model takes an OpenRouter slug.
• A reply wrapped as {"findings": [...]} is unwrapped instead of dropped.

Added

• OPENROUTER_API_KEY_FILE: read the key from a file (wins over the
  variable; warns when the file is group- or world-readable).

Removed

• Breaking: GEMINI_API_KEY is no longer read. Replace it with
  OPENROUTER_API_KEY in ~/.config/mcp-env/gatehouse.env and in the
  GitHub secret passed to review.yml. gatehouse exits 2 with a pointer
  when only the old key is set.