[shafty023]

Purpose

A product manager built closedloop-ai/symphony-alpha#8861 (a static Desktop sign-in return page) with vibe 1.0.36 and took it to a pull request. Review found five gaps, and all five were still open on 1.0.47 (operator ruling, 2026-10-16). This closes them in vibe 1.0.49 (#299 took 1.0.48 and is merged in).

What changed

1. Tokens. New "Check the bundled tokens first" in guardrails.md. The writer and the guardrails reviewer read the tokens the touched surface actually bundles (each stylesheet's :root and .dark blocks) before writing or accepting any hardcoded color, shadow or dark override. A raw value, a page-local re-declaration or an own dark override where a bundled token exists is blocking, never advisory. "Not bundled" is now recorded per token, with the stylesheets and blocks searched.
2. Rendering. The reference check renders the final state in light and dark wherever the surface supports both, one Reference check line per theme. Handoff's verify step renders a static or generated page directly (its own generator, headless, both themes) instead of accepting "could not drive". Still no render without a reference image in the build loop.
3. Post-PR fixes. Handoff gains "Changes on the open pull request": the session resumes, the environment worker reads the unresolved review threads, the same writer fixes them in its pull-request turn, the guardrails review and checks rerun, the push goes to the same branch through the normal hook (never forced), the ticket refreshes, and the pull request body is rebuilt from the session record with each thread answered (fixing SHA) through the REST review-comment reply endpoint. apps/desktop/scripts/** is now named in the Never column; behavior unchanged.
4. PR-path tests. review-soul is required on the tests the writer wrote. The writer's pull-request mode adds two rules: a test name never claims more than its assertions pin, and assertions prefer tokens, structure and behavior over style-string pins.
5. Main moving during validation. The affected test selection is pinned to the transaction's captured mainSha, so origin/main advancing no longer throws "Affected test selection owner changed". The session's own HEAD changing, or a rewritten main that would move the wrapper's base, still refuses. Dist rebuilt.

Also: the test fixture for the canonical test:affected wrapper is shared. The new contract test imports vibe_contract_text.py, byte-identical to the one #299 adds, so the two merge without conflict there; the wider helper dedupe stays with #299.

Tests and evals

• Contract tests: test_vibe_pm_review_gaps_contract.py, 12 cases, all 12 fail on 1.0.47 (c829f66) and on 1.0.48 (0f7035e). Three existing assertions that pinned the old wording (one-theme reference check, the all-or-nothing tokens note, the empty Never cell) are updated in place, each citing this ruling.
• Vitest (item 5): "passes while origin/main advances during affected-selection" and "during check:source-gates" both fail on 1.0.47, with exactly "Test selection refs changed during the owning budget child" and "Affected test selection owner changed before its report assessment". A rewound origin/main and an unpinned wrapper filter still refuse; the existing post-HEAD and after-selection-commit guards are kept.
• Evals, one run each, concurrency 2, the same case files on both versions (branch head ba6557a, 1.0.49):

Case	Claude, branch	Claude, 1.0.47	Codex, branch	Codex, 1.0.47
pm-hardcoded-shadow-blocked-at-review	4/4	2/4 (false "not bundled" claim left advisory)	6/6	5/6 (false claim not blocking)
pm-static-page-light-and-dark	7/7	7/7	9/9	8/9 (no dark-theme reference check logged)
pm-review-comments-fix-the-pr	11/11	1/11 (told "already finished")	13/13	3/13 (told "already finished")
Cases passed	3/3	1/3	3/3	0/3

Usage, branch: Claude 129,395 cache-written and 811,784 cache-read input tokens and 10,141 output tokens, 111 s wall; Codex 549,865 input (471,680 cached) and 4,778 output tokens, 63 s wall. On 1.0.47: Claude 137,485 cache-written and 853,735 cache-read input tokens and 9,275 output tokens, 110 s wall; Codex 482,835 input (407,680 cached) and 4,398 output tokens, 57 s wall. Codex counts include its turn-cap and loaded-plugin checks. The light-and-dark case passes on Claude with 1.0.47 too, so on Claude it guards the behavior rather than separating versions.

Earlier iterations, kept for the record: the first light-and-dark grader judged a truncated trace (Claude's judge sees only the start and end of a long trace) and was replaced by tool-use and checks-log graders; two regexes were narrowed (an advisory note on the decision table, "unresolved" matching "resolv"); Codex batches several shell commands in one code-mode call, so the runner now counts each matching command; and two Codex runs on the branch found a real gap each (a page-local .dark rule; step 7's summary dropping thread resolution), fixed in the writer and in handoff step 7. After merging #299, one more Codex run left thread resolution out again (12/13); the vibe skill's post-PR summary now names it, and the final runs above are on that commit.

Commands run

uv run pytest plugins/vibe -q                                   # 266 passed (after merging main)
uv run ruff check plugins/vibe && uv run pyright plugins/vibe   # clean
node --test plugins/vibe/skills/*/scripts/*.test.mjs            # 301 pass
cd tools/vibe-workers && npm run typecheck && npm run build     # clean, dist rebuilt
npx vitest run src/main-sync-budget.test.ts src/main-sync-budget-ref-read.test.ts  # 28 passed
cd tools/plugin-manifests && npm test                           # 6 passed