pi-summarize

A pi extension package with two features, both driven by a cheap side model:

1. Live activity line. While the agent works, the working indicator shows one short, specific present-progressive line (≤ 12 words) describing the goal rather than the command, e.g. Checking why the sync test fails. Tool calls summarize immediately; text and thinking summarize at most every ~2 s. Subagents run their own copy of the extension and narrate themselves; the parent does not relay their progress. While a tool runs, the line shows elapsed time every 5 s (display only, not recorded), so long waits visibly move. Calls are cancelled when superseded or when the run ends, and failures or timeouts (10 s) fall back to a deterministic line such as Reading files.js… or Working….

   In the transcript, tool rows are hidden. Each step (one model turn that runs tools) leaves one breadcrumb, ◆ <summary>, where its tool rows were, recorded at turn_end from that step's live line (waiting up to 2.5 s for it). A tool-only step with the same goal as the previous breadcrumb merges into it; a step with its own lead-in text always gets its own breadcrumb. Ctrl+O (expand) brings back the tool rows and shows each summary's thinking tail and tool calls. With activity off, tools render normally.

2. Auto-title. After the first assistant reply in a session with no name, it sets a 2–4 word session name. It never overwrites a name you set, and it does not retry once it has named the session.

Each breadcrumb and each generated title is persisted with pi.appendEntry("pi-summarize", …). These custom entries are not sent to the LLM. Hosts can read them from the session JSONL.

// {"type":"custom","customType":"pi-summarize","data":{"kind":"activity","text":"Reading files.js…","at":1760000000000,"turn":2,"source":"fallback","toolCallIds":["t1"]}}
// {"type":"custom","customType":"pi-summarize","data":{"kind":"title","text":"Fix Sync Adapter","auto":true,"at":1760000000000}}

Host integration

A non-TUI host (such as pi-web) needs only two things: show the extension's setWorkingMessage text as the live line (falling back to Working… when it is cleared), and render the activity entries above, grouping each step's tools by toolCallIds. Subagents write their own pi-summarize activity entries into their session JSONL (the subagent outputFile).

Install

From npm:

pi install npm:pi-summarize

From git:

pi install git:github.com/bry-guy/pi-summarize

From a local checkout:

pi install /path/to/pi-summarize

Configuration

Set these in ~/.pi/agent/settings.json under the pi-summarize key, or override with environment variables (env wins):

{
  "pi-summarize": {
    "model": "anthropic/claude-haiku-5-5",
    "thinking": "low",
    "activity": true,
    "title": true
  }
}

The extension reads settings.json from $PI_CODING_AGENT_DIR (default ~/.pi/agent). Project-level settings are not read.

Try it

Run pi in the TUI and send a prompt that uses tools. The working line should change from the default spinner text to activity lines. After the first reply, the session name should be set (check /name or the session selector).

Non-interactive check:

cd /tmp && pi --session-dir /tmp/ps-sessions -p "list files in /tmp briefly"
grep '"customType":"pi-summarize"' /tmp/ps-sessions/*.jsonl

Development

npm test   # node --test on test/*.test.ts (Node ≥ 22.18, native TS type stripping)
mise run release   # tests, refuses an already-published version, then npm publish

• src/core.ts: config, fallback lines, debounced ActivityTracker, auto-title rules. No pi imports.
• extensions/index.ts: pi wiring (events, model call via ctx.modelRegistry.streamSimple, setWorkingMessage, appendEntry, setSessionName).
• test/core.test.ts: unit tests with an injected mock completion.
• test/extension.test.ts: loads the extension against a fake pi API and mock registry.

Limits

• Model calls use streamSimple, not complete(), because only streamSimple accepts a reasoning level. complete() is used in pi's own examples.
• Activity entries record toolCallIds but not fromEntryIds, because streaming events do not carry transcript entry ids.
• Each activity line costs one side-model call. A long run with text streaming makes roughly one call every 2 s. Use a small model.
• Only user-scope settings are read. The extension does not add a settings UI.