# WP9 — Round 5 close-out Diff-level plan for the close-out cycle. Evidence tables are filled at WP9; the STEPS below are fixed now so the phase is executable rather than improvised (A-gate round 2, blocker 7). ## 0. Executable steps 1. **Fill §1-§2** from the goalplan's `capturedEvidence` — no re-derivation from memory. 2. **Promote:** `git mv devlog/_plan/260726_oracle_chase4 devlog/_fin/260726_oracle_chase4` (the directory is gitignored, so the follow-up `git add -f devlog/_fin/...` is mandatory). 3. **Sync counts** in `structure/str_func.md`. The gate is ALREADY red from this round's docs — measured at the round-2 audit: ``` FAIL devlog/ count drift: doc 553 files/77835 lines vs actual 562 files/79855 ``` Every implementation phase re-runs `npm run docs:counts` and syncs its own rows (`web-ai/`, `test/unit/`, `test/integration/`, per-file line rows). WP9 syncs the final `devlog/` row LAST — not because `_plan` → `_fin` changes the file count (it does not; the same 9 files are simply counted under a different directory), but because writing this close-out's own content changes the line total. Rows to touch: `devlog/`, plus any `web-ai/*.mjs` line row the implementation phases moved. 4. **Fresh gates, in this order:** `npm run test:unit`, `npm run docs:drift`, `npm run docs:counts`, then scoped `npx tsc --noEmit -p tsconfig.checkjs-dom.json` filtered to touched files. Playwright smokes are NOT run and NOT claimed green. 5. **Push:** `git push origin dev`, then `git rev-parse dev` and `git rev-parse origin/dev` must match; record both. 6. **Close the FSM** through D and `cxc loop validate`. ## 1. Outcome WP1-WP8 each reached D with an independent-reviewer PASS. Each row below is a full PABCD cycle, not a batched edit. WP9's own result is recorded in §3-§4. | WP | Rows | Commit | Result | |----|------|--------|--------| | WP1 | triage | `b799188^..35d747c` + `efe1c95` | 9 docs / 2783 lines; 11 audit rounds, 33 blockers folded. G7/G12 REOPENED as code debt after the reviewer measured the live predicate; G28 REDESIGNED after reading `e7526efa` | | WP2 | G7/G9/G12/G8 | `8635ad7` | `readChatGptStreamingState` returns `{strength, evidence}` with the full upstream `86d1fb2b` grammar declared body-local, plus `57d4a7af`'s optional `Edit` and the live-trace re-entry rule. Poll loop: stop-probe first, streaming = strong only, `minStableMs` 5000 when weak-active | | WP3 | G11 | `1b52e5d` | `readAssistantSnapshotSources` — one document-order union across both sources, per-node dedup, DOM-following filter, realm-safe `FOLLOWING`, `ok` flag; wholesale partial rejection and ok-gated fallback | | WP4 | G28 | `35eeb25` | Upstream `e7526efa`'s marker list and leading-marker semantics ported and adapted — `DR_TOOL_CALL_MARKERS` with our own `normalizeDeepResearchReportText` and a contextual check after the length floor, rather than upstream's `trimBeforeFirstAnswer` + bare `startsWith`. Improved past upstream with a whole-token boundary after the reviewer showed a legitimate "Used Tools in Modern Oncology" report was swallowed | | WP5 | G13c | `f59d963` | `emptyShellHostKind` with exact hostname sets, one terminal dot stripped before `www`, and `x.com` gated on an exact `/i/grok` path; only graced providers are re-probed, and `graced` never leaks into the public verdict | | WP6 | G81b | `d1d4398` | Ownership is conferred from a single OBSERVED element id read off the specific More locator — never a selector — and forwarded through both post-More resolutions | | WP7 | G25 | `edce15e` | zh implemented via canonical label sets with THREE consumer-specific projections, because one shared projection could not preserve three historically different predicates | | WP8 | G16 | `661e625` | Opt-in `ensureChatSurface`; chat no-op checked BEFORE the legacy throw, inserted before `selectChatGptModel` whose surface guard rejects Work, warning surfaced to the caller | | WP10 | composer-menu null verdict | see §3.2 | Post-push follow-up. Ran the integration suite for the first time in three rounds and fixed the crash it exposed on the plain tool-selection path | ### 1.1 What the reviewer caught that the implementation missed Across the implementation phases WP2-WP8, mutation testing repeatedly exposed assertions that were green against a deliberately broken implementation. The recurring failure was the same: asserting on source SHAPE (`src.toContain(...)`) instead of on BEHAVIOUR through the public path. Representative cases: - **WP8** is the clearest. The send-path tests stopped before completion, so deleting the user-facing warning entirely left all 15 tests green. Only a harness that runs `sendWebAi` through to `status: 'sent'` can observe `result.warnings`. - **WP4** shipped a predicate that swallowed a legitimate report titled "Used Tools in Modern Oncology" — found by the reviewer constructing the adversarial fixture, not by mutation. - **WP6** ownership was proven only once a test observed the real `evaluate` payload rather than the resolver's inputs. WP1 is docs-only and has no such finding; its 11 audit rounds were factual corrections to the triage, including reopening G7/G12 as code debt. Two other standing rules were re-confirmed the hard way: `page.evaluate` serializes a function BODY and not its module bindings, so browser-context functions must declare every constant inside themselves and be proven by a real transport round trip; and a test that fails by TIMEOUT proves nothing, so a mutation must fail on its assertion to count. WP10 adds a third, and it is the sharpest one of the round: **an inherited "known broken" label is a claim, not a fact.** Three rounds restated "Playwright smokes fail for missing Chromium" without once testing it. The suite ran fine, and it had been hiding a real production crash since round 4. A gate nobody runs is indistinguishable from a gate that passes — and strictly worse, because it buys false confidence. ## 2. Final gap-matrix state The backlog is empty. Every row from the round-3 matrix (`_fin/260724_oracle_chase2/040_gap_matrix.md`, G1-G34) and every follow-up row opened in rounds 3-4 (G1b, G13b, G13c, G81b, issue #81) now has a terminal disposition. No row remains in "Deferred" limbo. ### 2.1 Rows this round closed | Row | Disposition | Where | |-----|-------------|-------| | G7 | **Closed** — anchored grammar with heading prefix, decimals, unit aliases, compound and worded durations | WP2 `8635ad7` | | G8 | **Closed** — activity split into strong/weak/none strata | WP2 `8635ad7` | | G9 | **Closed** — live-trace re-entry rule adopted; busy/shimmer deliberately not adopted (no false positive to fix) | WP2 `8635ad7` | | G11 | **Closed** — wrapperless answers correlated at snapshot acquisition | WP3 `1b52e5d` | | G12 | **Closed** — optional `Edit` affordance inside the anchored predicate | WP2 `8635ad7` | | G13c | **Closed** — provider-aware empty shell behind a hydration grace | WP5 `f59d963` | | G16 | **Closed as opt-in** — the objection was never to the capability, only to doing it silently; it now requires `--normalize-surface` | WP8 `661e625` | | G25 | **Closed** — zh via canonical label sets, 5530-string differential across six consumers showed zero behavioural change for en/ko | WP7 `edce15e` | | G28 | **Closed** — tool-call placeholder captures rejected, with a whole-token boundary that upstream lacks | WP4 `35eeb25` | | G81b | **Closed** — submenu ownership from the observed triggering row | WP6 `d1d4398` | ### 2.2 Rows already terminal before this round Carried forward unchanged; listed so the matrix is complete rather than partial. | Rows | Terminal state | Closed in | |------|----------------|-----------| | G1, G2, G3, G4 | **Implemented** — streaming scope hardening | round 3 | | G5, G6 | **Implemented** — terminal evidence binding | round 3 | | G13, G14, G15 | **Implemented** — interstitial hardening | round 3 | | G20, G22, G27 | **Implemented** — Sol CLI/effort | round 3 | | G29 | **Implemented** — durable conversation URL | round 3 | | G30 | **Implemented** — CDP disconnect recovery | round 3 | | G34 | **Implemented** — strict env parsing | round 3 | | G19, G21, G24, G26, G32, G33 | **Covered** — no action, path:line proof in the round-3 unit | round 3 | | G23, G31 | **Not-applicable** — API routing and serve-owned tabs do not exist in a browser-only, shared-page architecture | round 3 | | G1b | **Implemented** — stop-probe scoping migrated to the remaining callers | round 4 | | G13b | **Implemented** for challenge/login — Grok/Gemini interstitial wiring | round 4 | | G17, G18 | **Implemented** — Work-surface detection, G18 narrowed to the model-mutation boundary | round 4 | | issue #81 | **Fixed** — ChatGPT connector selection against the current popover DOM | round 4 | | G10 | **Permanently not ported** — superseded upstream by `67da293a`; re-affirmed this round | round 3, unchanged | Upstream `steipete/oracle` remains at `6009d4ad` (`chore(changelog): open 0.16.2`, 2026-07-23), newest tag `v0.16.1`. Round 3 closed the real `1146107..6009d4ad` delta; round 4 was the first documented NOOP, so this is the **second consecutive NOOP** and the third consecutive round anchored at `6009d4ad`. There is nothing left to port: the entire round was our own accumulated debt. ## 3. Gate evidence Run fresh at close-out, after the `_plan` → `_fin` promotion: ``` npm run test:unit -> Test Files 156 passed (156), Tests 1682 passed (1682), exit 0 npm run docs:drift -> All structure drift checks passed (164) npm run docs:counts -> All structure count checks passed (76) ``` The round opened at 152 files / 1529 tests, so it adds 4 files and 153 cases. WP10 (§3.1-§3.2) adds one case and re-runs the set: unit 156 files / **1683** tests, integration 21 files / **171** tests, `docs:drift` 164, `docs:counts` 76. The Chromium transport tests added in WP2/WP3/WP7 were each verified by the A-gate reviewer with `AGBROWSE_CHROMIUM_EXECUTABLE_PATH` pointed at the installed Chrome. ### 3.1 Correction — the "missing Chromium" baseline was wrong (WP10) Rounds 3-5 all recorded the Playwright integration smokes as unrunnable for a "missing bundled Chromium". That was never true. `playwright-core@1.58.2` resolves its default `executablePath()` to build **1208**, which is absent, but the cache holds working **1217** and **1228** Chrome-for-Testing builds. The repo already ships the escape hatch — `AGBROWSE_CHROMIUM_EXECUTABLE_PATH`, read by `test/integration/playwright-launch.mjs`. Pointed at 1228, the smokes run: ``` $ AGBROWSE_CHROMIUM_EXECUTABLE_PATH=".../chromium-1228/.../Google Chrome for Testing" \ npm run test:integration Test Files 21 passed (21) Tests 171 passed (171) ``` The cost of that unexamined claim was a real defect sitting in `dev` since round 4. Running the suite for the first time surfaced a crash on the ordinary tool-selection path — see §3.2. ### 3.2 The defect the unrun suite was hiding `evaluateComposerMenu` normalized a REJECTED `page.evaluate` to a no-verdict shape but not one that RESOLVES `null`, so a bare `null` reached `isComposerPlusMenuOpen` and threw `Cannot read properties of null (reading 'reason')` — through `selectChatGptComposerTools`, on the plain send path. Bisected to `6aeb656` (round 4's issue-#81 fix), which introduced the `result.reason` read. Verified by running the fixture on detached worktrees at `6aeb656` and at round 4's tip `e5623a0`: both fail identically, so round 5 did not introduce it and round 4 shipped it unseen for exactly this reason. Fixed by normalizing both no-verdict cases at the single evaluate wrapper. The regression test asserts the public `selectChatGptComposerTools` path against a page whose `evaluate` resolves `null`; reverting the fix reproduces the exact production `TypeError`. `npm run typecheck:checkjs-dom` is red at repo baseline (184 diagnostics) and stays that way. The bar this round holds itself to is **zero NEW diagnostics in touched files**, measured rather than asserted: the WP8 files (`chatgpt-work-picker.mjs`, `chatgpt.mjs`, `cli.mjs`) carry 25 diagnostics both at `HEAD` and on a `HEAD~1` worktree, and the new test file contributes none. ## 4. Push ``` $ git push origin dev To https://github.com/lidge-jun/agbrowse.git e5623a0..defb7cc dev -> dev $ git fetch origin dev && git rev-parse dev defb7ccd0369e8917a9b44051efda584b0907ba4 $ git rev-parse origin/dev defb7ccd0369e8917a9b44051efda584b0907ba4 ``` 20 commits, `e5623a0..defb7cc`: the round-4 push was the previous remote tip. No force-push, no PR, `dev` only. `defb7cc` is this close-out commit itself; the §4 block above is therefore recorded in the follow-up commit that carries the push evidence, since a commit cannot contain its own successor's SHA. ## 5. Next round Anchor for round 6: upstream `6009d4ad`, local `dev` at the WP10 commit. The clone at `/tmp/oracle-chase-260724` is still valid but lives in `/tmp`, so round 6 should re-fetch rather than trust it — and after two consecutive NOOPs the first question for round 6 is whether upstream has moved at all before any planning work begins. With the deferred backlog now empty, round 6 has no inherited queue. Its scope is whatever upstream ships next, plus anything the reviewer flagged this round as honestly-unreachable rather than covered. One concrete carry-over from WP10: `npm run test:integration` should be part of the standard gate set from round 6 on, with `AGBROWSE_CHROMIUM_EXECUTABLE_PATH` pointed at a present Chrome-for-Testing build. Pinning `playwright-core` to a version whose default build actually exists in the cache would remove the need for the override entirely.