refactor: collapse the duplicate stacks (PR 3/3)

[EtienneLescot]

Last of three. #167 and #168 have merged, so this no longer stacks on anything — it has been rebased onto release/1.8.0 and stands alone. 55 files, −2,165 net.

This is the behaviour-changing set, so every claim here was re-verified against the code rather than taken from the review.

Verification

Gate	Result
tsc --noEmit	clean
biome check src electron	clean (9 pre-existing warnings, untouched files)
vitest --run	96 files / 1108 tests pass

One provider stack, and two OAuth fixes

llm-call.ts opened with "no LangChain dependency" and hand-rolled a 736-line multi-provider fetch/SSE client — OpenAI-compat, Anthropic, Codex and Copilot dialects — while the live agent loop went through LangChain the entire time. Its only two remaining callers were one-shot prompt→text calls: chat compaction and caption translation. Both now use createOpenScreenChatModel(...).invoke(), which already knew every provider.

Removes llm-call.ts, codex-session.ts (a bespoke chatgpt.com Responses dialect reachable only from it) and agent-provider-capabilities.ts, whose reasoning-effort matrix was duplicated inside deep-agent/.

Copilot is fixed, not just preserved. I flagged this as the functional risk of deleting llm-call.ts. On inspection the risk ran the other way: chat-model.ts used the PAT directly with a comment saying the runtime-token refresh "can land as a follow-up" — so the main agent path was the broken one, and only the two side-features had it right. chat-model.ts now exchanges the PAT for the short-lived runtime token, honours the base URL that exchange returns, and sends the editor-identifying headers the endpoint requires.

Codex had the same bug, pointed the other way — caught in review. accountId was threaded from both new call sites into OpenScreenChatModelConfig and then read by nobody, while the deleted llm-call.ts:542 set chatgpt-account-id from it and called that header required by chatgpt.com/backend-api. So the reroute silently broke caption translation and compaction for Codex accounts. tsc cannot see this: an optional field that nothing reads still typechecks. The header is now wired at the same shared factory, so every caller benefits.

Two more review findings in the same area:

• caption-translate built a model per batch. For Copilot that is a PAT → runtime-token exchange against api.github.com on every iteration; at batch size 40 a 400-segment transcript paid for ten of them. Built once, above the loop.
• The surviving capability matrix was not the equivalent duplicate the deletion assumed. isOpenRouterReasoningModel strips the anthropic/ slug prefix but not openai/, and both vendor matchers are anchored at the start — so openrouter + openai/gpt-5 silently got no reasoning wiring at all. The deleted file stripped both and its deleted test asserted it. Fixed, and the tests for the module that survived are added back.

extractDelta became messageContentToText and moved to chat-model.ts: it was exported from deep-agent/service.ts, whose top level pulls langchain and the whole agent-tools graph — a lot to import for ten lines of content flattening. Provider aliases (claude → anthropic, gemini → google) are now normalized once at the createOpenScreenChatModel entry point, which also gives normalizeProviderId a caller again instead of leaving it dead.

Schema and store

Four document fields — agent, preview, export, history — were declared, defaulted, seeded in every test fixture, written on every edit, and read by nothing. preview.revision was bumped at eight sites and consumed only by tests asserting the bump happened. Gone, along with the ten schemas behind them.

timelineOperationSchema stays: the review called it dead, but AxcutTimelineOperation has 19 call sites. Only operationSchema, which extended it with suggestion ops nothing dispatches, goes.

Old projects are unaffected — documentSchema is not strict, so zod strips the now-unknown keys.

Seven useTimeline actions had no callers. editClip looked alive because it had its own test suite; the live Edit Clip modal actually calls tl.updateClipSourceRange + tl.updateClipCrop, so the action and its tests both went. useTimeline.ts drops 1192 → 970 lines.

A follow-up sweep caught what the first pass missed: migrate.ts still built the agent envelope on every migrated project (shipped code, not a fixture), and two fixtures still carried the removed fields. They survived for a reason worth knowing when reading the gate table above: tsconfig.json excludes **/*.test.ts(x) and its include is ["src", "electron"], so no test file and nothing under tests/ is typechecked — and vitest transpiles without checking.

showTrimWaveform

Removing TimelinePaneBody left this boolean declared, defaulted, persisted, migrated and history-tracked across five files with no reader and no UI able to set it. Removed, along with the settings.timeline.* keys in all 13 locales — the separate top-level timeline namespace (FloatingInspector, V4Timeline) is untouched.

Old projects keep working: migrate spreads input.editor wholesale into legacyEditor and legacyEditorSchema is z.object({}).passthrough(), so a stored showTrimWaveform rides along inertly rather than failing the parse. Pinned with a test rather than assumed.

Deferred, with reasons

Three items from the review are not here. Each needs something this PR cannot provide.

• The native inspector channel (live.rs InspectorParams + 31 setNativeParam calls). My review said RightPanes' panes were unmounted; that was wrong — FloatingInspector.tsx:45 mounts them, so all 31 call sites are live. Removing the second write path means proving every one of those keys survives the scene push, which needs the app driven control-by-control. Not something to ship on a green typecheck.
• 601 lines of "dead" CSS. LeftPanel.tsx:108 does styles[palette] — computed class access that static grep cannot see. No test would catch a wrong deletion; it would just silently break visuals. Pure styling is not worth that risk.
• ModalShell → radix, LangButton → Popover. Carried over from PR 2 for the same reason: replacing hand-rolled focus/Escape/outside-click on eight live dialogs needs visual verification.

The remaining Rust cleanups (app.rs, the compositor.rs resize path, run_standalone, PreviewTransport) are mechanical and cargo check-verifiable, but belong with the inspector work since they touch the same files.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: b54f5658-6731-4c5e-9299-f28d53a803ce

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch chore/ponytail-3-refactor

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}