feat(PyAutoBrain): hygiene conductor — phase 3 perf mode

[Jammy2211]

Overview

Phase 3 (final core phase) of the PyAutoBrain hygiene conductor (P1 scaffold #88, P2 modes #90). Bring the perf mode live — developer-loop cost: slow unit tests, slow integration-mode workspace scripts, and import time — completing all five modes. Keeps the phase-2 shape: cheap read-only pre-scan + delegate; the conductor never runs heavy audits and never mutates a repo. Umbrella: PyAutoMind/issued/hygiene_agent.md.

Plan

• import-cost pre-scan — time python3 -c "import <pkg>" per library in a subprocess (best-effort; the conductor stays bash and never imports the JAX stack itself). Report per-library import seconds; libraries over threshold are the actionable signal; not-importable → advisory.
• consult Heart for heavy signals — slow tests/scripts are already tracked by PyAutoHeart's script_timing / test_run legs; perf surfaces those (via vitals / pyauto-heart) rather than re-running them, and routes the slow items. ("measurement lives in Heart; hygiene acts.")
• delegate fixes — slow function → /refactor or /bug; JAX-adaptation surfaced as a recommendation, never an automatic rewrite.
• extend --json perf row from staged to real rows; flip perf live in the conductor + veneer (nothing then remains staged); default worklist ranks the perf import signal alongside tidy debris.

Detailed implementation plan

Affected Repositories

• PyAutoBrain (primary)

Work Classification

Library (Brain-source)

Implementation Steps

1. agents/conductors/hygiene/hygiene.sh — add prescan_perf (subprocess import timing per library, best-effort, threshold flag) + a Heart-timing consult (surface script_timing/test_run findings); make perf a live mode (drop the staged branch); extend the --json row.
2. agents/conductors/hygiene/AGENTS.md — flip perf from staged → live; document the import-pre-scan + Heart-consult + delegate.
3. skills/hygiene/hygiene.md — remove the "perf staged" note (nothing staged now).
4. tests/test_hygiene_conductor.py — extend hermetic tests for the perf row shape (it is no longer staged).

Key Files

• agents/conductors/hygiene/hygiene.sh — the conductor to extend.
• agents/_common.sh — consult_vitals_verdict / resolve_heart for the Heart consult.

Charter guard

The import timing runs in a subprocess — the conductor never imports JAX/the science stack into the Brain process (stdlib/bash only).

Testing

pyauto-brain hygiene perf emits real import timings + surfaced Heart findings (human + --json); nothing reports staged; PyAutoBrain tests/ green.

Deferred (NOT this phase)

A standing PyAutoHeart import_time (or cli_noise) leg — adding Heart legs is a PyAutoHeart change; keep phase 3 PyAutoBrain-only and note the leg as a follow-up if the import pre-scan proves worth promoting to a tracked Heart signal.

Autonomy

--auto, effective supervised (header supervised ∧ feature-medium cap safe). Parks at ship sign-off. Merge/close human.

Original Prompt

Click to expand starting prompt

Phase 3 (final): bring the hygiene conductor's perf mode live — dev-loop cost (slow unit tests, slow integration-mode scripts, import time). Keep phase-2 shape (pre-scan + delegate; never mutate). import-cost pre-scan times python3 -c "import pkg" per library in a subprocess (conductor never imports JAX); consult Heart's script_timing/test_run for slow tests/scripts and route to refactor/bug; JAX-adapt is a judgement recommendation. Flip perf live (nothing staged after). Defer a standing Heart import_time leg. Done when perf emits real timings + Heart findings, nothing staged, tests green.