0.5.4 — compare agents from the skill and the runner

• /adv-review:review 140 --agent claude,codex: two or more agents, comma separated,
  review the pull request once each through the runner and the session reports the
  comparison. Never posts. A model may follow each agent after a colon,
  --agent claude:sonnet,codex:gpt-5.5; the runner's --agent takes the same form.
• /adv-review:review --compare claude.json codex.json: no review; results written
  earlier with --out put side by side.
• The runner takes the same list: node runner/review.mjs 140 --agent claude:sonnet,codex:gpt-5.5
  reviews once per agent at the same time and prints the comparison; --out results.json
  keeps results.<agent>.json per agent.

0.5.3 — --agent from inside Claude Code

• /adv-review:review 140 --agent codex (or gemini, copilot) in Claude Code runs the
  review with that CLI's agents through the runner and reports as usual. The skill finds
  the runner through $CLAUDE_PLUGIN_ROOT in an installed plugin.

0.5.2 — GitHub Copilot as a fourth agent

• --agent copilot: the GitHub Copilot CLI as a fourth agent, through copilot -p in
  silent mode with the answer asked for in the prompt and checked by the runner. Its tool
  allow-list enforces --no-run (git shell commands only, writes denied); the repository's
  own Copilot extensions stay out of each agent. Experimental, built against a stand-in
  CLI and not benchmarked: no Copilot plan.

0.5.1 — compare reviews by different agents

• runner/compare.mjs puts reviews of one commit by different agents side by side:
  what every agent found, what only some did, what one confirmed and another refuted,
  and the agents, minutes and cost of each run.

0.5.0 — the review with Claude, Codex or Gemini agents

Benchmarked through the runner on a four-PR sample of the fixture (#4, #7, #14, #11),
next to the 0.4.0 plugin pass on the same PRs. Claude agents on Sonnet: 4 of 4 planted
defects confirmed, 6 of 7 confirmed findings true, nothing on the clean PR. Codex on its
default model, gpt-5.6-sol: 4 of 4, 21 of 22, one pre-existing finding on the clean
PR. Gemini not run: no plan. See calibration/README.md.

• A standalone runner, runner/review.mjs, runs the same review with each agent as a
  headless CLI process in its own git worktree: --agent claude, codex or gemini,
  with --model choosing the model within the vendor. It prepares the pull request or
  branch, enforces each agent's answer shape (Claude and Codex through their CLIs,
  Gemini through the runner's own check and one retry), caps concurrency with --jobs,
  writes --out, posts with --post, and records what produced the result under
  runner. Codex and Gemini are experimental; Gemini has not been run against a real
  CLI. The runner is also the way to run the review from CI or with an API key.
• The skill works for agents without a Workflow tool: it tells them to run the runner.
• A capability matrix in the README says what each agent enforces.
• Agents' worktrees may start at the commit under review, not only on main; the
  prompt says so. Codex's sandbox keeps .git read-only, so the agent's own checkout
  cannot be relied on there.
• calibration/bench.sh runs a four-PR sample through the runner for one agent, and
  calibration/bench-table.mjs renders result folders as the table the README carries.
• Scorer: a ruling whose lines contain a finding's line outranks one merely within the
  window, and a ruling may carry finding, a regular expression on the title, so two
  findings on one line can be ruled apart.

0.4.0 — branch review, a same-site check, a harder calibration set

Calibrated on both sets with Sonnet agents: 17 of 18 planted defects confirmed, none
wrongly refuted; 33 of 33 confirmed findings true; no findings on the four clean PRs.
The one miss, D3, was named by the critic but not verified. See calibration/README.md.

• --branch [name] reviews a branch with no pull request, against --base or the
  remote's default branch. Committed work only; report only, nothing to post.
• A same-site check after triage: when separate groups sit at the same place in the
  code, one agent rules whether they are one defect or several. Only groups that share
  a place can merge; a failed check, or no room under the ceiling, leaves them as
  triage made them. The 0.3.0 calibration reported one path traversal twice.
• Fixed: a retried triage could take a run one agent past its ceiling. The skeptics
  are now counted from what is spent after triage, not before it.
• A second, harder calibration set: the fixture gains interest, webhooks and a read
  cache, and PRs #12 to #21 carry eight defects that need more than the changed lines
  to see, plus two clean PRs. The answer key labels each entry with its set; the scorer
  reports each set and takes --set.
• calibration/verify/verify.mjs shows every planted defect biting on its branch, from
  a clone of the fixture.
• The scorer marks a missed defect that the critic named among its gaps: seen, never
  verified, still a miss.
• calibration/run.sh takes branch:<ref> targets, and now stops each session it
  starts. It used to stop the shell around the session instead, so every review session
  of a pass kept running after it finished.

0.3.0 — a blind test is evidence, not a finding

Calibrated again on the same eleven PRs with Sonnet agents: all ten planted defects
confirmed, none wrongly refuted; 15 of 15 confirmed findings true; no findings on the
two clean PRs; 130 agents against 152. See calibration/README.md.

• A test that cannot fail is no longer a finding of its own. When it hides a defect in
  the diff, the defect is the finding and the blind test is one of its sites, named in
  the summary; when the code under it is right, nothing is reported. Finders, triage and
  the skeptics are each told so, and a blind test can no longer lead a root cause or set
  its severity. In the 0.2.0 calibration this class was every false positive and seven
  of the nine true findings nobody planted.
• --model <name> picks the subagents' model for one run.
• calibration/run.sh runs the calibration pass headless, one review per PR from a
  clone of the fixture, and results are kept per version under calibration/results/.

0.2.0 — calibrated

Calibrated: on the fixture, with Sonnet agents, all ten planted defects confirmed and
none wrongly refuted; 20 of 25 confirmed findings true after adjudication; three
findings on the two clean PRs, all false. See calibration/README.md.

• Every finding says whether the PR introduced the defect or it was already on the base
  commit (introducedByDiff from the finders, correctable by a skeptic; preexisting in
  the result). Pre-existing findings are reported apart and left out of the posted
  comment unless --include-preexisting is passed.
• A refuted finding that nobody showed false, where every skeptic was unsure, is marked
  dismissedAsUnsure and listed for a person to glance at.
• An agent that dies, or that answers about the wrong commit, is retried once while the
  ceiling has room. A retry is an agent spent; the result lists retried.
• Given no --agents, the run computes the smallest ceiling with room for every fixed
  angle and two of the scout's own, never under 20.
• --out <path> writes the run's result JSON to a file.
• A calibration set: a public fixture repository with ten planted defects across the
  eight generic angles and two clean PRs, the answer key in calibration/manifest.json,
  and calibration/score.mjs, which scores --out results for recall and precision.
• An eval case (evals/review-smoke) for claude plugin eval that runs one real review
  from a fresh clone of the fixture.