pi-session-summary

A pi extension that dynamically maintains a one-line LLM-generated session summary, set as the session name so it appears in pi's status bar and /resume session list.

Model is auto-detected from available cheap models, in this order:

1. openai-codex/gpt-5.6-luna — if you're logged into ChatGPT/Codex: covered by the subscription, no per-token cost
2. gpt-5.4-nano, gpt-5.4-mini (direct OpenAI key, or via openrouter)
3. gemini-3.1-flash-lite, gemini-3-flash-preview
4. claude-haiku-4-5

Other openai-codex models are never auto-selected: the ChatGPT-OAuth backend rejects generic small models such as gpt-5.4-mini. If an auto-detected model fails with a provider error, it's skipped for the rest of the session and the next candidate is used (which may be a paid one — pin provider/model if you want to prevent that). A failure is reported once via a notification, even with the widget off.

Or configure the model explicitly (see below).

Install

pi install npm:pi-session-summary

Or add to settings.json:

{
  "packages": ["pi-session-summary"]
}

Commands

Command	Description
/summary:settings	Creates the global settings JSON file (~/.pi/agent/session-summary.json) with defaults if it doesn't exist, and shows instructions to edit it. Run /reload after editing.
/summary:update	Force an immediate summary update, bypassing the debounce timer.
/summary:clear	Reset the summary to the first line of the first user message, clearing all accumulated state.
/summary:cost	Show the summary model name, number of LLM calls, token usage, and cost breakdown for the current session.

Configuration

Create ~/.pi/agent/session-summary.json (global) or .pi/session-summary.json (project override). Project settings are merged on top of global settings, which are merged on top of defaults. Config is reloaded on session start/switch and /reload.

All fields are optional — only specify what you want to override:

{
  "provider": "openai-codex",
  "model": "gpt-5.6-luna",
  "debounceSeconds": 60,
  "maxTokens": 300,
  "resummarizeTokenThreshold": 40000,
  "showWidget": false,
  "verbose": false
}

Setting	Default	Description
provider	(auto-detect)	Model provider
model	(auto-detect)	Model ID
debounceSeconds	60	Min seconds between LLM calls
maxTokens	300	Max tokens for LLM response (not enforced by the openai-codex API; the prompt's single-line instruction is the effective cap there)
resummarizeTokenThreshold	40000	Token threshold for full re-summarize vs incremental update
showWidget	false	Show a belowEditor widget with summary, staleness, and compaction info
verbose	false	Show a notification whenever the summary changes

Upgrading from ≤1.0.2: if /summary:settings previously materialized "provider": "openai-codex", "model": "gpt-5.4-mini" into your settings file, that model is no longer accepted by Codex — change it to gpt-5.6-luna or remove both keys to re-enable auto-detection.