feat(web): show a fork preview dialog before forking a conversation

[junmo-kim]

Summary

Forking a conversation currently gives no indication of what will happen:
which messages stay in the original session and where the new session
starts. This adds a confirmation dialog that renders the transcript
boundary before the fork executes.

• Selecting Fork on any message now opens a preview dialog instead of
  forking immediately.
• The dialog shows the last turns kept by the original session above a
  highlighted "Fork point" divider, and the selected message quoted as the
  first message of the new session below it. The boundary semantics mirror
  the hub's existing prefix selection (selectForkTranscriptPrefix).
• Confirm runs the existing fork flow; cancel leaves the session untouched.
• Localized (en / zh-CN). Rewind behavior is unchanged.

Desktop	Mobile

Testing

• New unit tests: web/src/lib/forkPreview.test.ts (boundary slicing,
  turn merging/truncation, current-fork semantics) — 5 cases.
• New e2e: e2e/fork-preview.spec.ts drives the production
  ForkPreviewDialog via the standalone fixture page (scratchlist
  pattern): rendering, boundary marker, cancel/confirm callbacks, zh-CN
  localization. RED was verified before the implementation commit.
• Full suite green locally: web vitest 2850 passed, root bun run test
  pass, bun typecheck clean, Playwright 52 passed.

[github-actions]

Findings

• [Major] Fork still requires the old generic confirmation before the new preview opens - the added handler only runs after MessageActions confirms, producing two dialogs. Evidence: web/src/components/SessionChat.tsx:584. Suggested fix: invoke the preview callback directly from the Fork action and remove the fork-specific generic ConfirmDialog.
• [Major] The preview reverses historical-fork semantics - the hub and native Codex fork exclude the selected boundary, but the dialog says the new session starts with it. Evidence: web/src/components/AssistantChat/ForkPreviewDialog.tsx:51. Suggested fix: label the prefix as copied into the child and the selected message as excluded/the cutoff.
• [Major] Pending queued rows are rendered as part of the fork preview even though the hub never copies invokedAt == null rows. Evidence: web/src/lib/forkPreview.ts:55. Suggested fix: skip blocks whose invokedAt === null and add a queued-row test.

Summary

Review mode: initial

Three Major issues found in the confirmation flow and preview/backend semantic parity.

Testing

Not run (automation; PR-provided code was not executed). Coverage gap: no integration test enters the preview through MessageActions, and no preview test covers queued/uninvoked rows.

HAPI Bot

[github-actions]

Findings

• [Major] Failed fork requests close the preview before completion and surface no error — setForkPreviewRequest(null) unmounts the dialog before the API call, while the click handler discards the returned promise. The added pending UI therefore never renders, and backend/network failures become an unhandled rejection with no retry context. Evidence: web/src/components/SessionChat.tsx:569.
  Suggested fix:

  const executeForkConversation = useCallback(async (messageLocalId?: string) => {
      setHistoryActionPending(true)
      try {
          const result = await props.api.forkConversation(props.session.id, messageLocalId)
          await navigate({ to: '/sessions/$sessionId', params: { sessionId: result.sessionId } })
          setForkPreviewRequest(null)
      } finally {
          setHistoryActionPending(false)
      }
  }, [navigate, props.api, props.session.id])

  Make the dialog await onConfirm and render caught errors, matching ConfirmDialog.

• [Major] Current-tail forks render a historical cutoff that does not exist — current fork actions pass no messageLocalId, so the hub copies the complete transcript and boundaryText is null. The dialog still renders “Fork point” and always says the child resumes before “this message,” which is false for this common path. Evidence: web/src/components/AssistantChat/ForkPreviewDialog.tsx:41.
  Suggested fix:

  {previewKind === 'historical' ? <ForkBoundary text={boundaryText} /> : null}
  <p>{t(previewKind === 'historical'
      ? 'forkPreview.below'
      : 'forkPreview.currentTail')}</p>

  Return an explicit 'current' | 'historical' kind from buildForkPreview rather than inferring it from nullable text.

• [Minor] Attachment-only messages disappear from the preview — fork actions are available for user messages without text, and the API accepts attachment-only messages, but the preview extractor reads only block.text. A selected attachment-only boundary is shown with no quoted message, and copied attachment-only turns are omitted. Evidence: web/src/lib/forkPreview.ts:28.
  Suggested fix:

  if (block.kind === 'user-text') {
      const attachments = block.attachments?.map(({ filename }) => filename).join(', ')
      text = [block.text, attachments].filter(Boolean).join(' ')
  }

Summary

Review mode: follow-up after new commits

The three findings from the prior bot pass are no longer present. The current head has two Major correctness/UX regressions and one Minor preview-parity gap.

Testing

Not run (automation; PR-provided code was not executed). GitHub integration check passed; the main test check was still in progress when reviewed. Missing coverage: a SessionChat-level failed-fork case, current-tail dialog semantics, and attachment-only preview input.

HAPI Bot

[github-actions]

Findings

• [Major] The empty-prefix state can misrepresent a paginated fork — the preview is built only from the currently loaded message window, but keptTurns.length === 0 is rendered as “The new session starts empty.” When the selected boundary is the oldest loaded message and hasMoreMessages is true, the hub still copies all older invoked messages. Evidence: web/src/components/SessionChat.tsx:1362, web/src/components/AssistantChat/ForkPreviewDialog.tsx:54.
  Suggested fix:

  // Thread through whether older copied history is outside this window.
  const prefixMayContainUnshownContent = props.hasMoreMessages
  
  {keptTurns.length === 0 ? (
      <div>{t(prefixMayContainUnshownContent
          ? "forkPreview.noTextPreview"
          : "forkPreview.emptyPrefix")}</div>
  ) : null}

• [Minor] The new Playwright coverage is not run by PR CI — .github/workflows/test.yml invokes only terminal-wrap-fidelity.spec.ts and composer-copy.spec.ts, so this added spec never executes in automation. Evidence: e2e/fork-preview.spec.ts:1.
  Suggested fix:

  - run: bun run test:e2e -- terminal-wrap-fidelity.spec.ts composer-copy.spec.ts fork-preview.spec.ts

Summary

Review mode: follow-up after new commits

The previous bot findings are addressed. Two fresh-head issues remain: one preview-correctness bug for paginated history and one CI coverage gap.

Testing

Not run (automation): Bun is unavailable in this review environment. GitHub integration check passed; the main test check was still pending when reviewed.

HAPI Bot

[github-actions]

Findings

No high-confidence issues found.

Summary

Review mode: follow-up after new commits

The previous HAPI Bot findings are addressed on the current head. The full PR diff was reviewed; no Blocker/Major/Minor/Nit findings met the confidence threshold. Residual risk: the pagination-specific copy wording and the complete SessionChat → preview dialog → fork API wiring are not directly exercised by the added tests.

Testing

Not run locally (automation/review policy). GitHub integration and test checks pass at the reviewed head.

HAPI Bot

[github-actions]

Findings

No high-confidence issues found.

Summary

Review mode: follow-up after new commits

The full current diff was reviewed. The commits since the prior HAPI Bot pass only change the fork-point badge and confirm-button foreground colors. Residual risk: the complete SessionChat → preview dialog → fork API wiring and pending/error behavior are not directly exercised; the added browser test mounts the dialog in an isolated fixture.

Testing

Not run locally (automation/security policy). Current-head CI typecheck, targeted Playwright (including fork-preview.spec.ts), and root tests passed. The separate CLI runner integration job failed in unchanged CLI tests; no CLI or integration source files are changed by this PR, so it is not attributed to this diff.

HAPI Bot