feat: improve the Markdown conversation export; shrink the component test module graph

.agents/docs/sessions-surface.md

@@ -22,15 +22,30 @@ this page is the full text of the rules summarised there.
   `buildConversationMarkdown` (`packages/shared/src/conversation-markdown.ts`),
   NOT `buildReplayPromptFromHistory` — that one is the agent-facing replay
   prompt and its budget behaviour is load-bearing for CLI resume; keep the two
-  separate. The copy targets ~20k estimated tokens AND 50k characters (both
+  separate. The copy targets ~60k estimated tokens AND 150k characters (both
   bounds, because CJK is ~1 token/char) and gets there by degrading tool
-  output, terminal output, then thinking, oldest turns first.
-  **Message text is never trimmed** — a conversation whose prose alone exceeds
-  the budget returns `overBudget` instead of cutting it. Whatever was trimmed
-  must reach the toast (`describeCopiedConversation`); silent truncation reads
-  as "I copied everything".
+  results, terminal output, then collapsing tool calls to per-turn counted
+  summaries, and only then capping thinking — oldest turns first.
+  **Message text is never trimmed and thinking is never dropped** — the export
+  exists to move context to another agent, so prose-shaped content outranks
+  everything the agent produced around it. Thinking degrades by capping (head +
+  tail, middle elided) so the receiving conversation always inherits some of the
+  reasoning; a conversation whose prose alone exceeds the budget returns
+  `overBudget` instead of cutting it. Whatever was trimmed must reach the toast
+  (`describeCopiedConversation`); silent truncation reads as "I copied
+  everything". The one edit allowed on prose is heading level: bodies routinely
+  contain `#`/`##` headings that would outrank the `##` turn headings and
+  destroy the outline, so `demoteMarkdownHeadings` shifts them down two levels
+  outside fenced blocks. Turn headings carry a round number, local time, and —
+  for assistant turns — the recorded model and effective working time; speaker
+  names appear only when the conversation has more than one human in it.
+  The trim notice lives in the header blockquote, not the footer, so a reader
+  learns the transcript is incomplete before reading it rather than after.
   Message fork menus copy the inclusive history prefix independently of ACP fork
-  capability. Missing boundaries fail; generating replies are marked as partial.
+  capability. Missing boundaries fail. A reply that was still generating is marked
+  through the builder's `incompleteFinalResponse` option, so the warning lands in
+  the header blockquote with the trim notice rather than as a trailing line —
+  callers must not append their own note after the transcript.
   Attachment bytes are not exported. The session header copies all history.
   Header "Open in" / "Copy Path" launchers live here; shared launcher/path
   helpers are `../../lib/session-path-launchers.ts`,

.agents/notes/implemented/feature/2026-09-10-conversation-markdown-export-format.md

@@ -0,0 +1,98 @@
+# Rank thinking above tool detail in the Markdown export
+
+Status: implemented
+Translation: pending
+
+## Abstract
+
+Copy as Markdown is how a conversation reaches another agent, machine or
+workspace when native forking is unavailable, but its degradation ladder was
+ordered as if the export were a transcript: thinking was dropped one level
+before tool calls were touched, and the floor spent up to 120 characters per
+call on bold, mid-word-truncated command strings whose results were already
+gone. Prose headings inside message bodies also outranked the `## User` /
+`## Assistant` headings, so long exports lost their turn structure entirely.
+Thinking is now capped rather than dropped, tool calls collapse to one counted
+per-turn summary before that happens, body headings are demoted, turns carry
+round numbers and provenance, and the budget rose to 150k characters / 60k
+tokens. The character caps and the raised budget are product judgements tuned
+against synthetic fixtures, not measurements of real sessions.
+
+## Decision
+
+The export's purpose sets the value order. Its job is to move a conversation's
+CONTEXT somewhere else, so prose-shaped content — message text, proposed plans,
+and the agent's reasoning — outranks everything produced around it. The old
+ladder encoded the opposite for thinking.
+
+Four changes in `packages/shared/src/conversation-markdown.ts`:
+
+- `LevelConfig.includeThinking: boolean` became `thinkingCap: number`. Thinking
+  degrades through the same `clampMiddle` as tool results, so the head and the
+  conclusion survive and the receiving conversation always inherits some of the
+  reasoning. A binary drop was rejected: it is the one content class where
+  losing the middle is clearly better than losing the whole.
+- Tool calls collapse one level earlier than thinking is capped, and collapse to
+  a single `<details>` per turn with counts by `toolName`. Simply swapping the
+  two old levels was considered and is strictly worse: the old floor's per-call
+  lines were themselves a large share of the budget, so aggregating first is
+  what buys the room that keeps most sessions off levels 5 and 6 altogether.
+- `demoteMarkdownHeadings` shifts ATX headings in message bodies down two levels
+  outside fenced blocks. This is an edit to prose, which the file's own
+  invariant forbids, and the invariant was narrowed rather than waived: no
+  characters are removed, only `#` markers added. The alternative — abandoning
+  headings as turn delimiters — loses the document outline that makes a long
+  paste navigable.
+- Turn headings carry a round number, local `MM-DD HH:mm`, and for assistant
+  turns the recorded model and effective working time; a rule opens each new
+  round; the header blockquote carries the time range, repo/branch, models, and
+  the trim notice. The notice moved from footer to header because its reader is
+  often another agent, which needs to know the transcript is incomplete before
+  reading it. Speaker names reach turn headings only when the conversation has
+  more than one distinct `userId`; repeating one name through a solo session is
+  noise, and the caller supplies the `userId → name` map because `@lody/shared`
+  cannot resolve it.
+
+The same reasoning later absorbed the "last response was still generating"
+warning that #558 appended after the document: it is a caller-localized string
+passed as `incompleteFinalResponse` and rendered in the header block. Two notices
+about the transcript's completeness sitting in different places, one of them
+after the content it qualifies, is the defect this change set exists to remove.
+
+`describeTrim` states that collapsing drops results and terminal output rather
+than incrementing the per-block tallies. A collapsed call never reaches the
+block renderer, so those counters stay at zero; faking them to make the notice
+accurate would have made the stats lie instead.
+
+The budget moved from 50k/20k characters/tokens to 150k/60k. The old ceiling
+predates current context windows and was the reason ordinary sessions reached
+the floor at all. Secret redaction still covers terminal output and tool results
+only, and deliberately not message text: a user forking a conversation wants
+the material they pasted into it to travel with them.
+
+`buildReplayPromptFromHistory` is untouched. It is the agent-facing replay
+prompt, its budget behaviour is load-bearing for CLI resume, and it keeps its
+own `thinkingOmitted` semantics.
+
+## Evidence and limits
+
+`packages/shared/tests/conversation-markdown.test.ts` covers the new behaviour
+at 24 assertions, including: heading demotion leaving fenced `# comment` lines
+intact and clamping at h6; collapse reached with thinking untouched; thinking
+capped with `characters elided` and the block still present at the floor; round
+numbering with exactly one rule per new round; model and `4m12s` working time
+derived from `endedAt - timestamp - permissionWaitMs`; speaker names present
+with two `userId`s and absent with one; and the trim notice preceding the first
+turn. The squeezed-budget tests drive `maxChars` directly rather than building
+fixtures large enough to trip the real ceiling.
+
+`pnpm --filter @lody/shared --filter @lody/components run typecheck` passes.
+Verification is fixture-based: no real session was exported and pasted into
+another agent, so the claim that the new header and per-turn metadata help a
+receiving model is reasoned, not measured. The 2000/500 thinking caps and the
+150k/60k budget are starting values chosen against those fixtures and should be
+revisited against real session data.
+
+`.agents/docs/sessions-surface.md` holds the full rule text; the binding
+one-liner stays in `packages/components/src/components/sessions/AGENTS.md`.
+No Spec owns this surface today, so none was updated.

.agents/notes/implemented/testing/2026-09-10-components-test-module-graph.md

@@ -0,0 +1,82 @@
+# Shrink the component test module graph
+
+Status: implemented
+Translation: pending
+
+## Abstract
+
+The `@lody/components` suite takes roughly ten minutes because Vitest isolates
+every one of its 446 files and re-evaluates each file's whole import graph;
+assertions themselves account for well under a fifth of the wall clock. Tracing
+`vite-node` showed a single menu test pulling 991 modules, almost none of them
+from `node_modules`, so four sources of accidental graph weight were removed: an
+unused `Calendar` re-export dragging `react-day-picker`, `date-fns/locale`
+barrel imports, an eagerly globbed directory of 324 icon SVGs, and two Vite
+plugins plus a catch-all `tsconfig` path mapping that are only needed by the
+production build. That test now loads 667 modules. The wall-clock effect is
+**not** established: run-to-run variance on the measuring host exceeded the
+effect being measured, and the target of a 20-second suite is unreachable
+without disabling isolation, which currently crashes on pending async work.
+
+## Decision
+
+Vitest's `collect` phase — importing a test file and evaluating its transitive
+graph — dominates this suite. Isolation means that work repeats per file, so the
+only lever available without touching isolation is the size of the graph itself.
+Four independent sources were found by ranking `DEBUG=vite-node:*` requests:
+
+- `src/ui/index.ts` re-exported `src/ui/calendar.tsx`, whose `react-day-picker`
+  dependency pulls `date-fns`, `date-fns-jalali` and `@date-fns/tz`. The
+  `Calendar` component had no callers anywhere in the repository. Deleted, along
+  with the `react-day-picker` dependency; this also removes 3.6 MB from the
+  product bundle.
+- Eight modules imported `{ enUS, zhCN }` from the `date-fns/locale` barrel,
+  which natively loads all 96 locales. They now deep-import
+  `date-fns/locale/en-US` and `date-fns/locale/zh-CN`, both declared exports.
+- `new URL(\`./files/${name}.svg\`, import.meta.url)` makes Vite eagerly glob the
+  whole icon directory — 324 SVG modules — into every consumer of `FileIcon`.
+  That is correct for a bundle, which has to emit the assets anyway, so the two
+  calls moved to `file-icons/asset-url.ts` and the test config aliases only that
+  module to a stub. Stubbing the whole `file-icons` entry was rejected because
+  `file-tree-virtual-rows.test.tsx` exercises its component cache for real.
+- The test config dropped `vite-plugin-top-level-await` (Node and Vitest support
+  top-level await natively; the loro WASM suites pass without it) and replaced
+  `vite-tsconfig-paths` with explicit aliases, because `tsconfig`'s catch-all
+  `"*": ["./*"]` mapping makes the plugin probe the filesystem for every bare
+  specifier in every module. `vite.config.ts` keeps both: the production bundle
+  does need TLA downlevelling, and its targets are not this file's concern.
+
+Disabling isolation is the only change that would reach a 20-second suite, since
+it turns 446 graph evaluations into one per worker. It was measured and rejected
+for now: with `isolate: false` a run succeeds up to six files and fails at seven
+regardless of which file is seventh, with tinypool reporting
+`Unhandled Rejection: Terminating worker thread`. Raising the worker heap to 8 GB
+does not help, so this is leaked pending async work that isolation currently
+contains, not memory pressure. 89 of the 446 files also use `vi.mock`, whose
+registry is shared once isolation is off. Both are per-file cleanup projects.
+
+## Evidence and limits
+
+Module counts are deterministic and were taken from `DEBUG=vite-node:*` on
+`tests/session-header-menu.test.tsx`: 991 requests before, 667 after, with
+`date-fns` modules down to 3 and `react-day-picker` to 0. Of the original 991,
+only 11 came from `node_modules`, confirming externalisation was already working
+and that the weight was first-party. jsdom was ruled out as a cause: a single
+file spends about 1.2 s in `environment`.
+
+**Wall-clock improvement is unverified.** Timings on the measuring host are not
+reproducible at the needed resolution — the same baseline configuration measured
+97 s cold and 128 s warm on an identical 112-file shard, a spread wider than any
+effect claimed here. Earlier single-file figures in this work were taken across
+config changes that invalidate the Vite cache and should not be treated as
+evidence. A trustworthy number needs repeated runs on a quiet host.
+
+The full suite never completed during this work; three attempts were terminated
+by the environment's background-task limit. Correctness was checked on a
+`--shard=1/4` run (112 files, 912 tests, all passing) plus the tests touching the
+changed modules, and `typecheck` and `lint` pass repository-wide. The suite also
+requires `NODE_ENV=test`: under `NODE_ENV=production` React resolves to a build
+that does not export `act`, and 154 files fail for that reason alone.
+
+Vitest 3.2.4 is pinned here; `fsModuleCache` and `vitest doctor`, both relevant
+to this problem, arrive in Vitest 4, which parts of the workspace already use.

packages/components/AGENTS.md

@@ -27,6 +27,17 @@ mobile surfaces. Background for the rules below:
   intentionally public or narrowly token-scoped DTO.
 - Renderer and worker builds that cannot use native top-level await must use
   `vite-top-level-await-fixed.ts`. Do not bypass its audited-version assertion.
+- Keep accidental weight out of the test module graph: Vitest re-evaluates a
+  file's whole import graph per test file. Do not re-export a heavy leaf from
+  `src/ui/index.ts`, deep-import `date-fns/locale/<tag>` rather than the barrel,
+  and keep the icon-asset `new URL` glob alone in
+  `src/components/icons/file-icons/asset-url.ts` so the test alias can replace
+  it. Rationale and measurements:
+  [module graph note](../../.agents/notes/implemented/testing/2026-09-10-components-test-module-graph.md).
+- `vitest.config.ts` deliberately omits `vite-plugin-top-level-await` and
+  `vite-tsconfig-paths`; `vite.config.ts` keeps both because the product bundle
+  needs them. Run the suite with `NODE_ENV=test` — a `production` value resolves
+  React to a build without `act`.
 - System theme state, persistence, and browser preference tracking are owned by
   `next-themes`. Keep Lody's wrapper focused on preview state, fixed VS Code theme
   application, and the Electron native-theme bridge.

packages/components/package.json

@@ -169,7 +169,6 @@
     "prosemirror-flat-list": "^0.7.1",
     "qrcode": "^1.5.4",
     "react": "catalog:",
-    "react-day-picker": "^9.8.0",
     "react-dom": "catalog:",
     "react-hook-form": "^7.60.0",
     "react-i18next": "^15.6.0",

packages/components/src/components/codex-reset/codex-reset-forecast-dialog.tsx

@@ -1,7 +1,8 @@
 import { useCallback, type ReactNode } from 'react';
 import { Trans, useTranslation } from 'react-i18next';
 import { formatDistance, type Locale } from 'date-fns';
-import { enUS, zhCN } from 'date-fns/locale';
+import { enUS } from 'date-fns/locale/en-US';
+import { zhCN } from 'date-fns/locale/zh-CN';
 import { ExternalLink, Loader2, TimerReset } from 'lucide-react';
 
 import { Badge } from '@/ui/badge';

packages/components/src/components/icons/file-icons/asset-url.ts

@@ -0,0 +1,21 @@
+/**
+ * Asset URLs for the file-type icon SVGs.
+ *
+ * `new URL(<template containing a variable>, import.meta.url)` is how Vite emits
+ * an asset whose name is only known at runtime — but it gets there by eagerly
+ * globbing the whole directory into the module graph. That is 324 extra modules
+ * dragged in by every consumer of `FileIcon`, which is fine for a bundle (they
+ * have to be emitted anyway) and ruinous under Vitest, where the graph is
+ * re-evaluated for every test file.
+ *
+ * Keeping the two `new URL` calls alone in this module lets the test config
+ * alias it to a stub without also stubbing the extension mapping or the cached
+ * component factory, both of which real tests do exercise. Do not add anything
+ * else here, and do not inline these back into `index.tsx`.
+ */
+
+export const getFileIconUrl = (iconName: string): string =>
+  new URL(`./files/${iconName}.svg`, import.meta.url).href;
+
+export const getFolderIconUrl = (iconName: string): string =>
+  new URL(`./folders/${iconName}.svg`, import.meta.url).href;

packages/components/src/components/icons/file-icons/index.tsx

@@ -1,4 +1,5 @@
 import { useMemo, type ComponentType } from 'react';
+import { getFileIconUrl, getFolderIconUrl } from '@/components/icons/file-icons/asset-url';
 import {
   compoundExtensionMap,
   extensionMap,
@@ -70,14 +71,6 @@ interface FolderIconProps {
   className?: string;
 }
 
-// Create icon URL from icon name
-const getFileIconUrl = (iconName: string): string => {
-  return new URL(`./files/${iconName}.svg`, import.meta.url).href;
-};
-
-const getFolderIconUrl = (iconName: string): string => {
-  return new URL(`./folders/${iconName}.svg`, import.meta.url).href;
-};
 
 export const FileIcon = ({ filePath, className = 'h-4 w-4' }: FileIconProps) => {
   const iconUrl = useMemo(() => {

packages/components/src/components/mobile/mobile-acp-history-sheet.tsx

@@ -1,7 +1,8 @@
 import { useMemo, useState } from 'react';
 import { useTranslation } from 'react-i18next';
 import type { Locale } from 'date-fns';
-import { enUS, zhCN } from 'date-fns/locale';
+import { enUS } from 'date-fns/locale/en-US';
+import { zhCN } from 'date-fns/locale/zh-CN';
 import { formatDistanceToNow } from 'date-fns';
 import { AlertCircle, Check, Download, Loader2, RefreshCw, X } from 'lucide-react';
 import type { LocalProjectHistoryCatalogItem, LocalProjectHistoryProvider } from '@lody/shared';

packages/components/src/components/mobile/mobile-project-skills-sheet.tsx

@@ -2,7 +2,8 @@ import { useMemo, useState } from 'react';
 import { useTranslation } from 'react-i18next';
 import { AnimatePresence, motion } from 'framer-motion';
 import { formatDistanceToNow, type Locale } from 'date-fns';
-import { enUS, zhCN } from 'date-fns/locale';
+import { enUS } from 'date-fns/locale/en-US';
+import { zhCN } from 'date-fns/locale/zh-CN';
 import {
   AlertCircle,
   Boxes,

packages/components/src/components/sessions/AGENTS.md

@@ -57,9 +57,9 @@ changing those files.
 
 - Read receipts are gated on VISIBILITY, not on being mounted: keep the
   explicit per-surface `isVisible` prop.
-- Markdown copy uses `buildConversationMarkdown`, never `buildReplayPromptFromHistory`;
-  preserve prose and toast omissions. Fork-menu copy includes the selected message,
-  independent of ACP fork support.
+- Markdown copy uses `buildConversationMarkdown`, not `buildReplayPromptFromHistory`:
+  keep prose whole, cap thinking rather than drop it, toast every omission, and
+  include the selected message regardless of ACP fork support.
 - Read ACP capabilities via `useResolvedMachineMeta()` and selectors via
   `useSessionAcpSelectorContext()`; the controlled composer must not recompute
   selector options.