pi-follow-through

pi-follow-through is a Pi extension that checks whether an agent has left useful work unfinished.

After a run settles, it sends the recent request, tool calls, transcript, and final answer to TypeSafe's Jev model. The extension only sends a continuation prompt when Jev identifies a current user request, selects an exact evidence line from the final answer, classifies the work as incomplete, and passes the configured probability threshold. The extension validates every selected ID against the request it sent.

Install

pi install npm:pi-follow-through

Review the source before installing it. Pi extensions run with the permissions of the Pi process.

Set a TypeSafe API key in the environment that starts Pi:

export TYPESAFE_API_KEY=your_key_here

TYPESAFE_AI_API_KEY is accepted as a fallback. Restart Pi after changing the environment.

Configuration

The default threshold is conservative:

{
  "followThrough": {
    "threshold": 0.8,
    "includeToolData": true
  }
}

Put this in ~/.pi/agent/settings.json for all projects or in .pi/settings.json for one project. Project values override global values. threshold must be between 0 and 1. Set includeToolData to false to leave tool calls and tool-result text out of the TypeSafe request.

What it does

The extension runs after agent_settled, not after every model call or tool call. It asks Jev one probability question and three typed evidence questions: which user request is active, which exact final-output line shows unfinished work, and whether the work is complete, incomplete, blocked, or unknown. The probability must meet the configured threshold (default 0.80), and the other answers must cite IDs that exist in the submitted state. The extension never parses free-form model output as a citation.

The unfinished-evidence candidates come only from the current final assistant output. This is deliberately conservative: a run that does not explicitly expose unfinished work is not nudged automatically. Identical progress after a previous nudge is also suppressed.

When all gates pass, it sends this message:

Continue useful work that is still within the user's request. Check for unfinished requested work and complete it now; do not invent follow-up work. If the request is complete, or progress needs user input, permission, or an external event, stop and say so.

The evaluator prompt treats short confirmations such as "Yes" and "Continue" as part of the earlier request. It also treats an explicit statement that implementation or review remains incomplete as evidence of unfinished work. A previous nudge is not enough on its own. The latest response must show new, actionable progress or a concrete next step.

The extension only sends nudges in TUI and RPC modes. It skips print and JSON modes, failed or aborted runs, active async subagent workflows, and stale evaluator responses after a new run starts. A missing key, request error, or two-second timeout skips the nudge instead of interrupting Pi.

Data sent to TypeSafe

The request can contain user prompts, source code, command lines, tool arguments, and tool output. By default, the extension sends:

• the last eight user requests, capped at 8,000 characters;
• the last 20 tool calls, capped at 8,000 characters;
• the last 20 transcript entries, capped at 24,000 characters;
• the final assistant output, capped at 8,000 characters.

These caps keep the evaluation request below Jev's context window. Truncated fields are marked in the request. Do not use the extension for sessions whose contents you do not want to send to TypeSafe.

With includeToolData: false, tool calls and tool-result text are omitted.

Try it from a checkout

git clone [email protected]:Nabsku/pi-follow-through.git
cd pi-follow-through
pi -e ./extensions/follow-through.ts

The extension has no runtime dependency beyond the Pi extension API and the fetch implementation provided by the Pi runtime.

Run the local checks with:

npm install
npm test

Design notes

See docs/design.md for the event flow, failure behavior, and the reasons for the current limits.

License

MIT. See LICENSE.