feat(desktop): add conversation transcript variant for focus-mode agent sessions

[baxen]

Category: feature

User Impact: In the full-cover focus view (any host that pins the conversation variant) the observer transcript now reads like a conversation instead of a dense activity feed — the human's prompt as a right-aligned bubble, the agent's reply as unboxed prose at a comfortable measure, reasoning behind a "Thinking…/Thought for Ns" disclosure, the plan as a live checklist, and session/status noise receded to quiet dividers.

Problem: The transcript had two presentations — the dense default activity feed and compactPreview — both tuned for a narrow panel. Read at full width they are a scan surface, not something you read a turn in.

Solution: A third, purely additive conversation variant, threaded through ManagedAgentSessionPanel and AgentSessionThreadPanel as an optional prop with a width-based default.

Rhythm

Item	conversation treatment
Human prompt	Right-aligned bubble, filled bg-muted/60, capped at max-w-[85%], author label + avatar, timestamp/context footer revealed on hover or keyboard focus. Unclamped — focus mode shows the whole prompt. Channel context stays behind the existing prompt-context affordance, not inline.
Agent message	Full-width unboxed prose in a centered max-w-3xl reading column, leading-relaxed, markdown/code through the existing shared Markdown renderer.
Thought	Collapsed <details> disclosure labelled "Thinking…" while streaming and "Thought for Ns" once the turn moves on. Auto-opens while streaming, collapses when the turn produces its next item. A reader who toggles it takes over — stream transitions can't yank it back.
Plan	Checklist card that mutates in place (completed / in progress / pending rows plus an "N/M complete" caption). Falls back to markdown for adapters that send free-form plan text.
Session boundary, turn setup, lifecycle status	Quiet centered ConversationDivider hairlines.
Error, permission	Unchanged loud treatment (destructive / amber), just respaced.
Tool items	Deliberately untouched — same ToolActivity presenter as default.

Spacing is gap-8 between turns and gap-3 within one. The timestamp preference is still respected.

Constraints honoured

• Existing variants render byte-for-byte identical. Enforced by a captured-markup fixture, not by inspection: AgentSessionTranscriptList.conversation.baseline.json was produced by mounting the same transcript on unmodified code, and I re-verified it against main's tip in a clean throwaway worktree before committing. The test fails on any drift in default or compactPreview output.
• Tool items untouched, so the parallel tool-chain-card slice merges cleanly.
• Drawer/shell untouched.

Notes for review

• Variant selection is caller-pinned only. AgentSessionThreadPanel defaults to default and only renders conversation when a host explicitly asks for it (Slice A's cover drawer does). The earlier 640px width heuristic and its module were removed in review: an automatic wide-pane mode is a separate product decision, and swapping the whole transcript presentation as a reader drags a resize handle across a threshold is not a decision to make implicitly.
• Thought durations are derived from the transcript itself, not from new wire fields. A thought ends at the next leaf item of any kind — tool call or plan update, not just the next assistant message — because reasoning should recede as soon as the agent acts on it; folding only on the next message would pin reasoning open across a long tool run. This is documented in agentSessionConversationMeta.ts and pinned by a thought → tool → message test.
• AgentSessionTranscriptChrome.tsx is a straight extraction of prompt/divider/setup chrome out of AgentSessionTranscriptList.tsx, which was approaching the 1000-line ratchet. Its one behavioural change is import.meta.env?. (optional chaining) so the module is importable from node render tests — the same pattern already used in features/onboarding/devFreshOnboarding.ts.

Review round (changes requested → addressed)

1. Programmatic-toggle echo in the thought disclosure — fixed. <details> fires toggle when React changes open, so auto-opening a streaming thought was recorded as a reader choice and pinned the disclosure open forever; reasoning never receded once the agent acted. Only a toggle that disagrees with the last rendered state can have come from the reader. The regression test injects the echo while streaming, then rerenders with a following item and asserts the disclosure collapsed — it is the only test that fails when the guard is reverted (verified). A second test proves a genuine reader toggle still survives later items.

2. Byte-for-byte fixture now covers the contract it claims. The baseline transcript contains every renderable item kind — prompt with prompt-context and setup lifecycle, assistant message, thought, plan, tool, lifecycle status, error, permission — across two sessions so a session-boundary divider is forced. Both legacy variants were recaptured from pre-change main (074561233) in a clean throwaway worktree. compactPreview's intentional omissions are captured in the fixture and therefore also protected. A companion test asserts the baseline really renders each kind, so a fixture that silently stopped covering one cannot keep passing while protecting nothing.

3. Thought completion documented as intended (ruling: next leaf item of any kind) with a thought → tool (4s) → message (88s) test asserting the 4s duration.

4. Width heuristic removed; agentSessionTranscriptVariantChoice.ts and its tests are deleted.

5. Plan-in-place rerender assertion added: same plan id with advanced entries retains the same card DOM node while progress text changes 1/3 → 2/3.

6. Fixture pinned to UTC — this was a real bug in the previous head's CI. The captured markup embeds formatted timestamps (formatTranscriptTimestampTitle renders in the ambient zone), and the fixture had been captured at UTC-7, so Desktop Core failed in CI (UTC) on a 12KB diff whose only real content was 12:00:01 PM vs 7:00:01 PM — on 7927fa37a too, not just this round. The test now pins process.env.TZ = "UTC" before the transcript modules are imported (their Intl.DateTimeFormat instances are module-level constants that resolve the zone once at construction), the baseline was recaptured from pre-change main under that zone, and a guard assertion fails with a specific message instead of a markup dump if the ambient zone is not UTC. Verified passing under TZ=UTC, TZ=America/New_York, and TZ=Asia/Tokyo.

7. Fixture locale pinned too — same bug class as the timezone one, one variable over. AgentSessionTranscriptChrome.tsx formats the session-boundary date with a bare new Date(...).toLocaleString(), which is locale-sensitive as well as zone-sensitive: the fixture holds 6/14/2026, 7:05:00 PM, and under LANG=de-DE the same render produces 14.6.2026, 19:05:00. The TZ guard passed silently while this failed. Three things make the fix non-obvious: assigning process.env.LANG at runtime is a no-op (node resolves its default locale once at startup — TZ works only because Date reads it lazily); patching Intl.DateTimeFormat alone is insufficient because Date.prototype.toLocale* does not route through it; and the Intl.DateTimeFormat replacement must be a plain function, not an arrow, since the render path calls it with new. The guard now asserts resolved locale and a known toLocaleString() output alongside the zone. Non-vacuous: stripping the pin fails under de-DE. Verified green under UTC/en-US, UTC/de-DE, America/New_York/ja-JP, Asia/Tokyo/de-DE, and UTC with LANG unset.

Known follow-up (Slice C sequencing)

Slice C (#6536) is not merged yet, so the shared controlled-details hook it promises does not exist on origin. The echo guard here is a local implementation with the same semantics as C's useToolRunDisclosure (rendered-state ref + useLayoutEffect + ignore-agreeing-toggle), flagged in a code comment so the swap is mechanical on rebase. Still outstanding for the rebase: agentSessionConversationMeta.ts hardcodes segment.kind === "summary", which C replaces with "tool-run" — it must flatten segment.run.items, plus a thought → tool-run meta test.

Validation

At e6c994a69 (based on main tip 074561233):

• pnpm test — 5379 passing, 0 failing (full desktop suite, not scoped), and re-run green across five TZ/locale combinations
• pnpm typecheck — clean
• pnpm check (biome + px-text + pubkey-truncation) — findings identical to main's baseline (pre-existing terminal.css !important warnings and infos)
• pnpm check:file-sizes — clean
• All pre-push hooks green: desktop-check, desktop-typecheck, desktop-test, file-size-check
• GitHub CI green on e6c994a69 — every check SUCCESS or SKIPPED, including Desktop Core, which was the job the ambient-formatting-dependent fixture had been failing (polled to completion, not inferred from local green)

New coverage: 13 JSDOM rendering tests for the variant (including the recaptured baseline byte-identity assertion, its coverage guard, the echo-regression pair, and the plan-in-place rerender) plus the meta/checklist unit tests.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7927fa37a3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Ignore programmatic thought disclosure toggles

When a live thought mounts or isStreaming changes, React updates the <details> element's open property, which itself emits toggle; this handler then records that automatic state in userOpen. Consequently, the normal streaming transition pins userOpen to true, and the disclosure remains expanded when the next transcript item arrives instead of automatically collapsing as intended. Track actual summary interaction separately, or suppress toggle events caused by synchronizing open.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Select the variant using the rendered panel width

When a previously widened panel is restored and the channel viewport later narrows, widthPx retains the stored width while RightAuxiliaryPane clamps the rendered pane to calc(100% - 300px). For example, a stored 700px pane in an 800px channel area is actually rendered around 500px, but this check still selects conversation; because the single-panel breakpoint is only below 600px, the cramped pane remains in the layout this threshold was meant to avoid. Measure or clamp the effective panel width before resolving the variant.

Useful? React with 👍 / 👎.