[nplez1]

	Files	Added	Deleted	Net
Test	5	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​513	0	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​513
Prod	19	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​1542	0	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​1542

ELI5

When you come back to a long-running agent session after being away, you currently have to scroll through everything it did to understand where things stand. This PR adds a "Session summary" strip to the agent detail view in Activity: click it open and you get a short written brief — what the agent is trying to do, what it has finished, what it is doing right now, anything blocking it, and a timeline of the highlights — with a "Since you last looked" section at the top so you see what is new first.

What Changed

Before: Checking in on an agent meant scrolling its terminal output — slow, and only shows a narrow slice of what happened.

After: The Activity thread detail pane has a collapsed "Session summary" strip with a one-line hint (e.g. "12 events to catch up on"). Expanding it shows the summary: current activity, goal, plan, completed work, blockers, decisions, and a timestamped timeline whose entries point back to the transcript message they came from. Collapsing it remembers what you have seen, so the next visit leads with what is new.

Mechanism: The summary is a cached "materialized view" over the session's own transcript file (the JSONL the agent CLI writes), folded into a structured ledger by an LLM call. Key properties:

• It only runs while you read it. Opening the strip starts the fold; closing it cancels in-flight work. Nothing runs for sessions nobody opens — no polling, no background LLM spend. Cheap non-LLM facts (state, timestamps, event counts) are tracked at all times so the hint line and staleness are free.
• Long sessions stay cheap and resumable. The fold advances a cursor through the transcript in bounded chunks (extracted in parallel, merged deterministically). Closing mid-fold keeps completed chunks; reopening resumes where it stopped instead of starting over. Early history survives transcript trimming because it is folded early.
• A failed fold never corrupts anything. Ledger state is atomic per fold; an unparseable model reply keeps the last good summary and simply retries next time.
• The model is user-selectable through the existing Settings → Git → Source Control AI pickers (agent, model, thinking level, custom command) — the same configuration text generation already uses. No new settings surface.

Also included: the pane's UI follows the design system (tokens and existing primitives only), and the ledger cache lives in userData as a rebuildable cache.

Why

The problem is not "summarize the transcript" — it is "what changed since I last looked". That reframe drove the design: entries are append-only and counted, so a per-session last-seen cursor can lead with what is new.

Alternatives considered:

• Full re-summarization on demand: simpler, but cost grows with session length and there is no delta view. The chunked cursor design costs one pass over only the unseen part.
• A rolling "old summary + new delta" rewrite: bounded cost, but long sessions silently lose early detail. The ledger keeps facts updatable in place and an append-only timeline, so history is never rewritten.
• Background/continuous summarization: freshest, but pays LLM cost for sessions nobody reads. Laziness costs only a short wait on open, mitigated by showing the deterministic skeleton immediately while the fold runs.
• Having each agent summarize itself: unreliable across agent CLIs and costs the agent's own context. A separate observer reads the transcript and works for every agent.

Linked Issue

None — fork-local feature work (no external issue exists; happy to attach one if the maintainer wants tracking).

Visual Proof

Pending — a new strip and panel in the Activity detail pane; rendered before/after captures will be attached once the UI validation pass runs. Layout: one collapsed row below the session header (chevron + "Session summary" + hint), expanding to a scrollable brief with labeled sections and a timeline (kind badge, text, #message-index anchor).

Testing

• Automated tests added/updated: 26 tests across src/main/session-summary/ covering ledger merge semantics (facts updated in place, timeline append-only), fold chunking/caching/atomic failure/cancellation, service lifecycle (open → fold → push, close cancels, reopen resumes from cursor, session-identity change resets the ledger, failed fold keeps last good ledger), and the settings→fold-params contract.

• pnpm tc clean, pnpm test green for the new suites plus the touched register-core-handlers suite, pnpm run check:code-quality:changed gate green, pnpm format applied.

• Platform: macOS. Manual rendered testing pending (noted above). No Windows-specific paths or processes are touched beyond the existing cross-platform spawn utilities (spawnSourceControlAgent, shared child-process).

• I manually tested these changes locally

• Automated tests added/updated, or explained why not below

AI Disclosure

Authored with the pi coding agent (model: xiaomi/mimo-v2.6-pro) under human direction; design decisions reviewed and chosen by the human.

Review

Agent skill upstream boundary

• Not applicable, or this change follows docs/reference/agent-skill-sharing-upstream-boundary.md and copies or mechanically translates no upstream skill-installer source, tests, fixtures, registry entries, path tables, comments, or documentation.

Notes

• Remote SSH: transcripts are local-FS today; sessions whose transcript cannot resolve degrade to a "no readable transcript yet" panel — no error, no crash. The verdict vocabulary and execution-host boundary are untouched.
• Status store: the summarizer is a subscriber/reader only; it adds no polling and no subprocess on the status path (per docs/reference/agent-status-store.md).
• Performance: zero background LLM work; fold runs only on explicit panel open with bounded chunks and a per-fold cache. Transcript parsing is local and happens only on open.
• Security/privacy: the ledger cache is derived from the session transcript and stored under userData with the same posture as existing scrollback/status caches; it is treated as rebuildable and never authoritative.
• Mobile / cross-platform: untouched. Process spawning reuses the vetted spawnSourceControlAgent path (WSL/Windows-aware) and path.join for paths.

Checklist

• This PR is small and focused
• I explained what changed and why (ELI5, the user-facing before/after, the mechanism, and why over the alternatives)
• Before/after screenshots or videos attached for UI changes, or N/A with reason — pending rendered pass (see Visual Proof)
• Self-reviewed for correctness, security, and performance
• Cross-platform, SSH/remote, and path/shortcut impact considered (or N/A)
• pnpm lint, pnpm typecheck, pnpm test, and pnpm build pass (or CI will cover; local preferred) — typecheck/tests/changed-lines gate pass locally; full lint and build left to CI