feat: add quirk:full-code-review

.github/workflows/test.yml

@@ -10,6 +10,9 @@ permissions:
 jobs:
   pytest:
     runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.9", "3.12"]
     # several tests make throwaway git commits, which fail on a runner with no identity
     env:
       GIT_AUTHOR_NAME: ci
@@ -21,8 +24,16 @@ jobs:
 
       - uses: actions/setup-python@v5
         with:
-          python-version: "3.12"
+          python-version: ${{ matrix.python-version }}
 
-      - run: pip install pytest
+      - run: pip install pytest pygments pyyaml
+
+      # the kernel-sandbox tests skip when no sandbox works, so fail here rather than skip silently
+      - name: Install bubblewrap and verify it can isolate
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y bubblewrap
+          sudo sysctl -w kernel.apparmor_restrict_unprivileged_userns=0 || true
+          bwrap --unshare-pid --die-with-parent --ro-bind / / --dev /dev --proc /proc /bin/true
 
       - run: pytest -q

BUGS.md

@@ -133,3 +133,25 @@ entries' IDs; manual edits to fix typos are fine.
 - **Severity**: medium
 - **Proposed fix**: Skip link destinations and autolinks when applying the underscore guard, as scenario-example.tsx now does.
 
+## BUG-15: SKILL.md and orchestration.md disagree on which dispatches carry the mutation-probe line
+- **Observed**: 2026-10-18
+- **File**: skills/full-code-review/SKILL.md:318
+- **Description**: SKILL.md Step 4 says to state the mutation-probe line in every dispatch, but orchestration.md's 'Every specialist prompt also says' paragraph says the coordinator adds 'mutation probe: enabled' only to the three test lenses, and only with consent. A coordinator can't tell whether every dispatch needs a mutation-probe line (disabled for the non-test lenses) or only the test lenses do. Doc-only. Found while fixing the gaps the MR !722 re-review exposed.
+- **Severity**: low
+- **Proposed fix**: Pick one rule: either every dispatch states the probe status (enabled or disabled), or only the test lenses get the line. Then make both files say the same thing.
+
+## BUG-16: full-code-review lets the coordinator skip the lens fan-out on small diffs
+- **Observed**: 2026-10-19
+- **File**: skills/full-code-review/SKILL.md:292
+- **Description**: In a headless smoke-corpus run (claude -p "/quirk:full-code-review --local --no-open", Sonnet 5.5 default, 2026-10-19), all 5 smoke cases reviewed the 2-9 line diff inline instead of dispatching lenses and wrote no report.json (one used the ReportFindings tool, one wrote a Markdown report). One also overrode a pre-exported REVIEW_OUT_DIR with its own mktemp despite the ${REVIEW_OUT_DIR:-...} default in Setup. SKILL.md never says the fan-out runs regardless of diff size, so the smoke corpus can't run without a prompt override and real tiny-diff reviews silently drop to one inline pass.
+- **Severity**: medium
+- **Proposed fix**: State in SKILL.md (Setup + Step 4) that the lens fan-out and report.json are mandatory at any diff size with no inline shortcut, and that a pre-set REVIEW_OUT_DIR is kept; add a test pinning that wording.
+- **Resolved**: 2026-10-19. Wording alone (fan-out at any size, keep REVIEW_OUT_DIR) fixed the out-dir override but still left 3 of 5 plain-prompt runs skipping, now citing "headless run on a one-line diff". Closing those excuses by name (one Agent per lens; skip only on run_when, policy, or failed dispatch; diff size, headless, no forge, or cost is never a skip reason) brought all 5 to a full fan-out, report.json in the caller's dir, and a passing smoke_check. Residual lens-packing is BUG-17.
+
+## BUG-17: full-code-review coordinator occasionally packs several lenses into one agent
+- **Observed**: 2026-10-19
+- **File**: skills/full-code-review/SKILL.md:298
+- **Description**: In the 2026-10-19 plain-prompt smoke run after the BUG-16 fix, 1 of 54 lens dispatches (correctness case) combined comments, db and observability into one agent despite SKILL.md Step 4 saying one Agent per lens. A packed agent shares one context across checklists, so a lens can get a shallower pass than its own dispatch would give, and run.models then reports per-lens models for a single dispatch. Earlier rounds also packed lenses (test-adequacy case, 2 multi-lens agents).
+- **Severity**: low
+- **Proposed fix**: Re-run the smoke corpus a few times to measure the rate; if it persists, have the coordinator check its dispatch count against policy.json's runnable lenses before Step 5, or record per-lens agent ids in run so a packed dispatch is visible in the report.
+

README.md

@@ -6,13 +6,24 @@ Version: **2026.7.9**
 
 ## What ships
 
-- **25 skills** under `skills/` covering TDD, systematic debugging, brainstorming (which produces a logic spec), deep-research/brainstorming exploration (with a `--involve` user-inclusion dial: checkpoints, mid-run steering, and idea-gate co-creation), divergent ideation, the two-document spec rubric (logic spec + tech spec) and plan writing, scannable-prose revision, code review, adversarial review (typed profiles, a promote/refute two-stage, and an evidence gate), parallel agent dispatch, guided bug/feature/code-change filing with provenance-marked artifacts, gating simplification behind a correctness check, and more.
+- **26 skills** under `skills/` covering TDD, systematic debugging, brainstorming (which produces a logic spec), deep-research/brainstorming exploration (with a `--involve` user-inclusion dial: checkpoints, mid-run steering, and idea-gate co-creation), divergent ideation, the two-document spec rubric (logic spec + tech spec) and plan writing, scannable-prose revision, code review, a full multi-lens pre-merge review for GitHub and GitLab, adversarial review (typed profiles, a promote/refute two-stage, and an evidence gate), parallel agent dispatch, guided bug/feature/code-change filing with provenance-marked artifacts, gating simplification behind a correctness check, and more.
 - **Typed artifacts** (added in 5.1.0) — surface-routing for observations Claude cannot act on. See below.
 
 ## Installation
 
 Install via the quirk-dev marketplace (`.claude-plugin/marketplace.json`).
 
+## Deep code review
+
+`/quirk:full-code-review --deep [minutes]` adds bounded area finders, cross-file
+contract checking, and fresh candidate verification to the standard review.
+The default budget is 60 minutes. Eligible runs ask before executing repros in
+throwaway copies; plausible findings stay visible but cannot be posted.
+See the [deep-mode runbook](skills/full-code-review/references/deep.md).
+
+`--since <ref>` reviews only the changes after `<ref>`, such as the fix commits made since
+the last review's head SHA. A delta review is report only: it loads no memory and posts nothing.
+
 ## Typed artifacts
 
 A discipline for routing deferred observations into structured markdown files instead of burying them in prose ("pre-existing", "out of scope", "skipped for brevity").

REVIEW.md

@@ -0,0 +1,5 @@
+---
+skip_globs: ["tests/fixtures/full-code-review/answer-keys/**"]
+---
+
+The answer keys are evaluation fixtures; reviewing them would leak expected defects to review agents.

TEST_BACKLOG.md

@@ -31,3 +31,10 @@ Reviewed every 2 weeks. Use `/quirk:artifacts:test-skip` to append.
 - **Edge cases to cover**: brainstorming step 9 reads logic-spec.md and writes logic.md with all six required sections; brainstorming halts at the user review gate; subagent-driven-development Step 2 resolves quirk:writing-specs and reads tech-spec.md when the complexity gate fires; executing-plans Step 0 does the same on the no-subagent path
 - **Priority**: P2
 
+## TEST-3: Two complete Sonnet deep-mode recall acceptance runs
+- **File under test**: skills/full-code-review/references/deep.md
+- **Test type**: e2e
+- **Reason skipped**: Sonnet runtime dispatch failed Anthropic OAuth refresh with invalid_grant. Historical eval-run1/eval-run2 were invalidated because they reused old lenses, supplied literal timings, seeded prompts, and claimed unvisited areas complete. No qualifying recall run has been established.
+- **Edge cases to cover**: After reauthenticating Anthropic, run every applicable lens and every area twice at 3605fd8777b60646bcad11aba08f83c2c858e6b9 with unseeded prompts, measured whole-run budgets, fresh per-claim checkers, and independent spot checks; score against the corrected answer key.
+- **Priority**: P1
+