# 60 — Cycle (cli-jaw (.ts))

> Part of [00_plan.md](00_plan.md) · Goal `68727b6d-d01` · **Status: ✅ DONE (104.7 deferred with rationale)**

## Target
- **Repo / lang:** cli-jaw (.ts)
- **Severity:** P1/P2
- **Gate command:** `npm test + npx tsc --noEmit`

## Gaps in scope
104.1-.18 session/active locks, model evidence + Korean i18n + legacy-pro reject, code-mode navigate-to-conversation, composer resolver-verified targets, attachment filename-verify (not count-only), gemini/grok runtime capability+model probes

> Note: 104.3 (watcher lock) + 104.19 (AX CDP fallback) were done in Cycle 5.

## Build log (cli-jaw branch `feat/webai-parity-100-260625`)

- **104.6 — Korean-locale model menu** — ✅ DONE — cli-jaw `21e115a0`
  - `normalizeModelPickerText` made Unicode-aware (`\p{L}\p{N}`) so Korean labels survive (they
    were stripped to '' and collided); Korean added to `CHATGPT_MODEL_OPTIONS`, `modelChoiceFromText`,
    the model-pill button pattern, the menu-open regex, `modelLabelPattern`. Tests BWAI-MODELI18N-001/002.
- **104.15 — `wrapError` field preservation** — ✅ DONE — cli-jaw `b0961de6`
  - A plain error-like object with a string `errorCode` keeps its fields instead of flattening to
    `internal.unhandled` (fallback still overrides). Tests BWAI-WRAPERR-001/002.
- **104.1 + 104.2 — session-store lock PID-liveness + deadline-aware active** — ✅ DONE — cli-jaw `0c47cdf5`
  - `isStaleLock` now reclaims a crashed holder's lock immediately (PID-liveness) vs acquiredAt-age only;
    `isSessionActive` extracted + deadline-aware (expired sessions inactive). Tests BWAI-SLOCK-001/002.
- **104.5 — model-selection evidence (P1)** — ✅ DONE — cli-jaw `efcce949`
  - `ChatGptModelSelectionEvidence` + `createModelSelectionEvidence`, returned from `selectChatGptModel`,
    + `modelSelection` on `WebAiSessionRecord`/`updateSessionResult`, persisted after send-path
    `createSession`. Test BWAI-MODELSEL-001.
- **104.7 — legacy-pro reject** — ⏸️ DEFERRED. cli-jaw's `isLegacyProModelLabel` rejects effort-pill
  labels (Standard/Extended Pro/Heavy); agbrowse's rejects legacy GPT-5.x-Pro *model rows* — different
  purposes. Matching agbrowse risks breaking cli-jaw's effort handling; needs deeper model-flow review.

- **104.20 — occurrenceIndex on interactive refs** — ✅ DONE — cli-jaw `768386d6`
  - per-(role,name) occurrence counts in both ref builders → 0-based `occurrenceIndex` on `ElementRef`.
    Test BWAI-OCCUR-001.
- **104.22 — observation-bundle element-ref filter + id/basis** — ✅ DONE — cli-jaw `4d285c80`
  - `isElementRef` (`@?e\d+`) accepts bare `e3`, rejects `@x`; emits `observationId`/`targetId`/`basis`.
    Test BWAI-OBSBUNDLE-001.

- **104.11 — zip scan chronological order** — ✅ DONE — cli-jaw `ce3ac51a`
  - `orderedConversationMessages` (create_time/update_time then mapping order) used by both zip scans →
    the resolved zip is the most-recent one deterministically. Test BWAI-ZIPORD-001.
- **104.16 — typed code-mode error + zip write-guard** — ✅ DONE — cli-jaw `afa61347`
  - `codeWebAi`/`extractCodeArtifacts` throw typed `code-mode.vendor-unsupported`; `downloadAndSaveZip`
    write wrapped → `code-artifact:write-failed`. Test BWAI-CODEMODE-001.

- **104.21 — contract-audit shared 7-feature contract + excludeNames** — ✅ DONE — cli-jaw `05112f52`
  - audit against `editorContractForVendor(vendor).semanticTargets` (catches responseFeed/streamingIndicator
    drift) + honours `excludeNames` (a "search" textbox no longer false-satisfies composer); forked
    `CONTRACTS` map removed. Existing test reworked + new excludeNames case.

- **104.13 — attachment filename-verify (P1)** — ✅ DONE — cli-jaw `7ab682b3`
  - `buildAttachmentReadyExpression` verifies the SPECIFIC uploaded files appear in chips
    (name/stem in text/attrs), not chip-count only; threaded from the single-file upload path.
    Test BWAI-ATTACH-001.

- **104.10 — code-extract navigates to the target conversation (P1)** — ✅ DONE — cli-jaw `437526cb`
  - `resolveConversationUrl` + `shouldNavigateForExtraction` (origin + convo-id compare); extract
    navigates to `chatgpt.com/c/<id>` first → `code-extract.navigation-failed` on error; untrusted ref
    rebuilt from id. Tests BWAI-CODENAV-001/002.

- **104.17 — copy scroll-jump suppression** — ✅ DONE — cli-jaw `41081655`
  - copy evaluate patches `scrollIntoView`→noop + `focus`→preventScroll around the copy, restored in
    finally. Browser-only behavior → verified by tsc + no-regression (no unit test); GPT-Pro live gate covers it.

- **104.14 — composer CDP `Input.insertText` for large prompts** — ✅ DONE — cli-jaw `cae7b825`
  - `insertTextLikeProvider` mirrors agbrowse: after the `insertText` override and before the keyboard
    fallback, an optional `getCdpSession()` routes large prompts through CDP `Input.insertText`,
    detaching in `finally`. `getCdpSession?` added to `VendorEditorAdapterOptions`. Exported for test.
    Tests BWAI-COMPOSER-CDP-001/002.
- **104.12 — composer accepts resolver-verified send/composer targets** — ✅ DONE — cli-jaw `2c2b22c6`
  - send-path resolved `composer.fill`/`send.click` targets but discarded them (resolved path dead).
    `findComposerCandidate` short-circuits to a resolved composer selector; `submitPromptFromComposer`
    tries `clickResolvedSendButton` (plain click → forced retry) before the scan, then Enter;
    `clickEnabledSendButton` timeout configurable; `adapter.submitPrompt(submitOptions)` merges per-call.
    `chatgpt.ts` threads resolved selectors into a `liveAdapter` + upload-aware `sendButtonTimeoutMs`
    (45s w/ attachments, 20s else). Tests BWAI-COMPOSER-RESOLVED-001..005.
- **104.18 — pollWebAi per-tick conversation-drift + tab-crash guards** — ✅ DONE — cli-jaw `d4dc46d7`
  - `captureAssistantResponse` runs an optional `driftCheck()` per tick (bails as `conversation-mismatch`
    when the held tab's `/c/<id>` differs from baseline — ignores the expected fresh-chat none→id), and
    wraps the tick so a page-death returns a recoverable `tab-crashed`; non-page-death errors propagate.
    `poll()` maps `result.drift` → WebAiOutput, marks session `crashed`. `WebAiStatus`/`WebAiSessionStatus`
    gain the new states; `WebAiOutput` gains `recoverable?`. Tests BWAI-POLL-DRIFT-001..003.
- **104.8 — runtime capability-probe engine + vendor arrays** — ✅ DONE — cli-jaw `0bb92f5d` + `5ea6af26`
  - new `capability-probe.ts` (port of `capability.mjs`): `defineCapability`/`runCapabilities`/
    `worstCapabilityState` + shared `probeHostMatches`/`probeFirstVisibleSelector`. Vendor arrays in
    dedicated `gemini-capabilities.ts`/`grok-capabilities.ts` (gemini-live.ts is already >500 lines;
    separate module also avoids a live↔capabilities cycle): 6 probes each (active-tab/composer/model-alias/
    upload/copy/streaming). `geminiStatus`/`grokStatus` attach `capabilityProbes`+`capabilityState`
    (named distinct from the DECLARATIVE `capabilities`/201 rows). Tests BWAI-CAP-ENGINE-001..007,
    BWAI-VCAP-G01..G04/K01.
- **104.9 — Gemini/Grok model capability probe + fallback ladder** — ✅ DONE — cli-jaw `a561f835`
  - `geminiModelCapabilityProbe`/`grokModelCapabilityProbe`: read-only, report ok (already active) /
    warn (selectable, not active) / fail (unavailable) / unknown (no model; Gemini deep-think=tool).
    `closeGeminiModelMenu`/`closeGrokModelMenu` restore the menu via Escape — selection never mutated.
    Tests BWAI-MODELPROBE-G01..G05, K01..K03. (fail-not-selectable is a 5s live path → Cycle-12 gate.)

**Gate (Cycle 6 complete):** full cli-jaw `npm test` → **4833 tests, 4815 pass, 0 fail** (one unrelated
orchestrate state-machine test is flaky across the suite — passes in isolation/on re-run; zero orchestrate
files touched this cycle); tsc 0.

### Cycle-6 disposition
All P1/P2 shared-module divergences ported except **104.7** (legacy-pro reject) — DEFERRED with rationale
above (cli-jaw's `isLegacyProModelLabel` targets effort-pills, not legacy model rows; matching agbrowse
risks cli-jaw's effort handling). 104.7 carried forward as a standalone follow-up, not a blocker.

## Verification
Per-item gates above; A-phase audit (Cycle 1) confirmed these as line-diff divergences.