pi-quality-gates

Quality gates for pi coding-agent: a settle-time rework loop (verification fails → the agent is forced to fix before finishing) plus instant lint injection after edits (diagnostics land in the tool result, next turn). Ships as a pi extension package; the core is host-agnostic so other pi hosts (e.g. pi-virtual-employee) can reuse it.

Install

pi install npm:pi-quality-gates

Or for local development: pi -e /path/to/pi-quality-gates/index.ts.

What it does

1. settle-gate (agent_before_settle)

When a conversation ends after edit/write-class tool calls, the gate runs your verification battery:

• Fail → forced rework: returns {continue:true} with a custom_message boundary entry listing the failing checks, so pi runs one more model turn to fix. Capped at maxReworks (default 2); after that it passes through with an explicit "work is NOT verified" message so the agent reports honestly instead of thrashing.
• Pass → quiet: nothing appended, zero context tax.

2. lint-inject (tool_result)

After an edit-class tool, a fast diagnostics battery (tsc/lint/check-shaped commands) runs and its findings are appended to the tool result content — the model sees the error in its next turn instead of at commit time. Clean runs stay silent.

Command discovery

PI_VERIFY_CMD (comma-separated supported) → <cwd>/verify.sh → <cwd>/.pi/verify.sh. None found → the extension is a no-op (projects without gates are untouched).

Discipline

• The gate fails open, verification failure is what triggers rework. Extension errors never wedge a host loop (mirrors pi's own fail-safe: a tool_call handler exception blocks the tool). Runner timeouts are enforced per check (editPerCheckMs, default 10s) because pi has no handler timeout upstream (#9910).
• Output is tail-truncated so a screaming test suite can't flood the context.
• No redacted content is ever reconstructed — only command output from your own repo.

Environment

Tests

npm test          # 17 unit tests (settle decision table + runner budgets + command discovery)
node test/smoke-rework.mjs   # end-to-end against bundled pi 1.3.0 (RPC + scripted provider)

The smoke asserts the rework loop end-to-end (4 provider requests: 1 write + 2 reworks + 1 pass-through) and the control run without the extension (2 requests, no rework) — the failure half of the contract.

Design notes

The design rationale (upstream API survey, pi-lens/pi-permission-system precedents, upstream issue constraints) lives in the pi-desktop repo: docs/design/pi-quality-infrastructure.md.