picc-recap

Claude Code-style away summary for pi — a faithful port of Claude Code's services/awaySummary.ts + hooks/useAwaySummary.ts.

After you step away and the agent has settled, a short, dim "while you were away" line appears. It is generated by a small/fast model from your last ~30 messages, using Claude Code's exact prompt.

This is intentionally a different thing from @zhcsyncer/pi-recap. pi-recap produces a one-line recap + title (JSON), renames the session, and syncs a multiplexer. picc-recap has no side effects — just the dim summary, like Claude Code.

Install

pi install npm:@ladbabynpm/picc-recap

or from a git bundle / locally:

pi -e ./picc-recap

For local development the live config is a junction in ~/.pi/agent/extensions/picc-recap.

Behavior

• Trigger (auto): when the agent settles (agent_settled), a timer starts. If you stay idle for idleAfterSettleMs (default 5 min, like Claude Code's BLUR_DELAY_MS) with no queued work, the summary is generated. Typing anything cancels it.
• Trigger (manual): /away generates immediately (with a cancellable loader).
• Model: a small/fast model by default (Claude Code's getSmallFastModel()), with thinking effectively off (small maxTokens). Configurable.
• Display: a TUI widget below the editor, dim, prefixed with Claude Code's ※ marker (U+203B, REFERENCE_MARK). Cleared as soon as you type or the agent starts.
• Dedup: at most one summary per user turn (Claude Code's hasSummarySinceLastUserTurn).
• Not compaction: the summary is shown as a widget and persisted as a custom entry. It is never injected into the LLM context.
• TUI only: headless modes (print, json, rpc) are no-ops, matching Claude Code's terminal requirement.

Prompt (verbatim from Claude Code)

Session memory (broader context):
<session memory, if any>

The user stepped away and is coming back. Write exactly 1-3 short sentences. Start by
stating the high-level task — what they are building or debugging, not implementation
details. Next: the concrete next step. Skip status reports and commit recaps.

The memory block is included only when present.

Config

Read from (project overrides global; project only read when the project is trusted):

$PI_CODING_AGENT_DIR/extension-data/picc-recap/config.json
<project>/.pi/extension-data/picc-recap/config.json

{
	"enabled": true,            // master switch
	"auto": true,               // generate after idle (vs. /away only)
	"manualCommand": true,      // enable /away
	"idleAfterSettleMs": 300000, // 5 minutes
	"model": "small-fast",     // "current" | "small-fast" | "provider/modelId"
	"fallbackToCurrentModel": true,
	"recentMessageWindow": 30,  // Claude Code: 30
	"maxTokens": 300,
	"sessionMemoryPaths": []    // optional best-effort memory file(s)
}

Model

model accepts:

• "current" — the active session model.
• "small-fast" (default) — picks a small/fast model from the registry (heuristic on ids containing haiku/flash/nano/mini/small/fast/lite/instant), preferring the current provider. Falls back to the current model if none match.
• "provider/modelId" — an exact registry lookup (e.g. "Claude/claude-haiku").

Deltas from Claude Code

Pi differs from Claude Code in a few ways that cannot be replicated exactly:

• No terminal-focus detection — Claude Code triggers on terminal blur + 5 min. Pi has no focus events, so the trigger is an idle-after-settle timer of the same duration.
• No per-session memory file — Claude Code reads {projectDir}/{sessionId}/session-memory/summary.md. Pi has no equivalent; set sessionMemoryPaths to read specific files as "broader context" (best-effort, omitted if missing — matching Claude Code's null path).
• No "thinking disabled" knob in pi-ai's complete() — emulated with a small/fast model and a small maxTokens.
• Display is a TUI widget rather than a persistent transcript message, so it clears when you act (arguably closer to the "while you were away" intent).

Development

npm install
npm run lint && npm run typecheck && npm run test

Single test: npx vitest run test/awaySummary.test.ts