feat(reflex): runtime guardrails — command/edit guard + Pi hook + dataset export

[bearmug]

Summary

First implementation of the repo-arch reflex runtime guardrails PRD. Adds a deterministic policy layer for shell commands and file edits, with an event log, test-plan generator, and Pi auto-mode hook.

What changed

New reflex subcommands

repo-arch reflex check-command <shell-command>   → Allow / Ask / Block
repo-arch reflex check-edit <file-path>           → risk + signals + warnings
repo-arch reflex test-plan <file-path>             → recommended tests/checks
repo-arch reflex install                           → install Pi pre-action hook
repo-arch reflex events [limit]                    → event log summary
repo-arch reflex dataset                           → export training dataset

Modules (src/reflex/)

Module	Purpose
classify-command.ts	Deterministic Allow/Ask/Block for shell commands
classify-edit.ts	Risky file edit detection (patterns + fragile-map scoring)
test-plan.ts	Test recommendations based on co-change history
events.ts	Append-only JSONL event log
install-pi.ts	Pi pre-action hook generator + settings update
dataset.ts	Export events + synthetic examples as training data
index.ts	Public exports

Command guard examples

[BLOCK] risk=critical  rm -rf /home
[ASK]   risk=medium   sudo apt update
[ALLOW] risk=low       git status

Edit classification examples

[CRITICAL] package-lock.json  signals: lockfile(1)
[HIGH]     tsconfig.json     signals: config(0.9)
[LOW]      src/utils.ts       no high-risk signals

Build order (from PRD)

• ✅ classify-command
• ✅ classify-edit
• ✅ test-plan
• ✅ install-pi
• ✅ events
• ✅ dataset
• ⬜ train (teacher → student classifier — next step)
• ⬜ session ingest (learn from Pi session failures — future)

Notes

train.ts and session-ingest.ts remain TODOs — wired in the build order but not yet implemented.