fix(cli): rework session recap rendering and add blur threshold setting

[wenshao]

Summary

Follow-up to #3478. Aligns five pieces of recap behavior with Claude Code's /recap, including reverting the sticky-banner placement that #3478 introduced based on a misread of the original report.

1. Render inline in history (was: sticky banner above input)

Claude Code's recap is a regular history item that scrolls with the conversation, not a sticky banner pinned above the input. The sticky placement in #3478 came from misinterpreting the original bug report. Tear out the banner machinery (state in slashCommandProcessor, UIStateContext field, DefaultAppLayout / ScreenReaderAppLayout placement, the handleResume wrapper that cleared it on session switch, the layout-effect re-measure, mocks) and addItem the recap into history. AwayRecapMessage now renders as ※ + bold recap: + italic content, all dim — matching Claude Code's visual shape.

2. Configurable away threshold (default still 5 minutes)

5 minutes matches Claude Code's default. Add general.sessionRecapAwayThresholdMinutes (number, default 5) for users who briefly alt-tab often and don't want recaps piling up, or who want a shorter value for testing. Non-positive / unset values fall back to 5 — dropping the key behaves identically to before.

3. Prompt aligned with Claude Code

Replace the previous "exactly one sentence, ≤80 chars" rule with the looser "under 40 words, 1-2 plain sentences" budget Claude Code uses. The natural shape of "current task + next action" is two clauses; forcing them into a single sentence either crams them with a semicolon or drops the next action on complex sessions. Add a Chinese-budget note (~80 chars) since the English-only spec drifts off-budget for CJK users. Keep our <recap>...</recap> wrapping which protects against reasoning-model preambles leaking into the UI (we serve glm/qwen/etc. that emit thinking text; Anthropic models don't).

4. Slash-command and settings copy

Restore "one-line" in the slash-command description (Generate a one-line session recap now) and the settings description even though the prompt allows 1-2 sentences. This deliberately sets a tighter user expectation than the prompt guarantees, keeping the surface feel "glanceable". The internal design doc keeps the accurate "1-2 sentence" wording.

5. Dedup back-to-back auto-recaps with no new turns between

Reported case: leaving the terminal twice while a /review of one PR was still on screen produced two recaps both about that same review (slightly different wording from temperature-driven LLM variance). Two consecutive blur cycles, each over the threshold but with no new user activity in between, would each fire their own auto-recap and add two near-duplicate entries to history.

Add a shouldFireRecap gate before kicking off the LLM call:

• Need at least 3 user messages in history total (don't fire on a near-empty session).
• If a previous away_recap is already in history, need at least 2 new user messages since that one before another can fire.

Same shape as Claude Code's gate (their constants are Sc1=3 total user messages and Rc1=2 user messages since the last recap). Read history through a ref so this isn't in the effect's deps and the effect doesn't re-run on every message.

Intentional remaining differences

	Claude Code	qwen-code
Default for showSessionRecap	on (built-in cheap fast model makes ambient calls negligible)	off (we have no equivalent free fast model — if fastModel is unset, an ambient recap would fire on the user's main coding model)
Reasoning-tag stripping (<recap>...</recap>)	not present	retained (we serve models that emit thinking text)

Test plan

Automated checks:

• npm run typecheck — passes
• slashCommandProcessor, useResumeCommand, btwCommand, clearCommand vitest suites — 72/72 pass

Manual verification (run in a tmux session against the built CLI with general.showSessionRecap: true and fastModel set):

#	Scenario	Result	Evidence
1	/recap renders in history (not pinned above input)	✅	※ recap: … line appears immediately below the > /recap user echo, just like any other history item — not floating above the divider
2	Recap scrolls with the conversation (the core behavior change vs. #3478)	✅	After /recap, sending say hi briefly produced ✦ 嗨。 below the recap line, pushing recap upward; the input-box separator did not stay glued to recap
3	AwayRecapMessage shape: ※ gutter + recap: label + content	✅	※ recap: Setting up chat context … rendered with the asterism in a 2-col gutter; bold/italic styling not visible in tmux capture but layout matches the spec
4	Generated content is 1–2 sentences (not crammed into one)	✅	Sample output: "Setting up chat context and answering a quick arithmetic check. Next: send your actual coding task or question so we can start." — two clean sentences, task + next action
5	sessionRecapAwayThresholdMinutes honored	✅	With 0.1 (≈6 s), an 8-second focus-out / focus-in dance triggered the auto-recap; recap appeared inline in history without further user action
6	First focus-cycle after ≥3 user turns fires recap	✅	After 3 arithmetic prompts + 8 s blur/focus, ※ recap: 正在为本次聊天建立上下文... appeared once
7	Second focus-cycle with no new user turns is suppressed	✅	Repeating the 8 s blur/focus dance immediately, without typing anything, produced no new recap (history recap count stayed at 1 across 25 s of waiting)
8	Third focus-cycle after 2 new user turns fires again	✅	Sending two more arithmetic prompts, then another 8 s focus dance, produced a second ※ recap: (count went to 2)
9	Listener-leak suppression from #3478 still holds	✅	30 retries logged across followup-speculation + recap, 0 MaxListenersExceededWarning on screen or in ~/.qwen/debug/latest

tmux focus-event reproducer used for cases 5–8:

echo '{"fastModel":"<your-fast-model>","general":{"showSessionRecap":true,"sessionRecapAwayThresholdMinutes":0.1}}' \
  > /tmp/recap-test/.qwen/settings.json
tmux new-session -d -s recap-test -x 200 -y 50
tmux send-keys -t recap-test "QWEN_WORKING_DIR=/tmp/recap-test npm --prefix <repo> start" Enter
# wait for boot, type a few prompts, wait for the model replies
printf '\033[O' | tmux load-buffer - && tmux paste-buffer -t recap-test  # blur
sleep 8
printf '\033[I' | tmux load-buffer - && tmux paste-buffer -t recap-test  # focus
# the recap line should appear within a few seconds

[wenshao]

No issues found. LGTM! ✅ — gpt-5.4 via Qwen Code /review

[wenshao]

[Review Summary]
Overall, the architectural shift from a sticky UI banner to an inline history item is a great UX improvement, aligning well with Claude Code's behavior. The history-based implementation simplifies state management and removes the need for complex UI hooks. I've left a minor note regarding the TypeScript typings for handleResume.

— gemini-3.1-pro-preview via gemini cli review

[yiliang114]

LGTM