[gabe4coding]

Intent

Add the code-mode / shall verification loop to retroagent. Before each commit, the agent runs /simplify, then /code-review, then scripts/test --all, then /verify. /verify uses a project recipe, .claude/skills/verify/ (SKILL.md and drive.py). It runs this checkout's kb and the plugin hooks (headless claude -p --plugin-dir .) against a temporary sandbox data repo, never the real one.

Drift

• The plan said "no drive script": the agent was to run each command by hand. If one command left out KB_CONFIG, it reached the real data repo. So drive.py came back. It sets KB_CONFIG and KB_ROOT (and HOME for kb runs) for each run, and it refuses a folder that has no sandbox marker.
• No TUI drive for the decide mod. The recipe uses claude plugin validate . and claude plugin test ., and the hook run shows the mod load in pluginLog.
• The /code-review fixes (about 20 lines) changed code. I did not run /simplify and /code-review again for them. I checked them inline and with the drive.

Decisions

• The loop order and the "start again at step 1" rule are the same as in code-mode and shall. The efficiency review said that a restart after each small fix costs a full review. I kept the rule as you asked for it.
• A hook session keeps the real HOME, because the Claude login needs it. So kb-hint in a hook run can use your local embedding model (method: semantic). A kb run gets the sandbox HOME.
• drive.py reuses tests/fixtures.py (init_remote, clone, make_claude_tree, make_codex_tree, git) and kb.adapters.common.iter_records.
• Door: two-way. The change adds dev tooling and rule text only. No src/kb/ code, hook or skill changes.

Evidence

Before: main has no .claude/skills/verify/ (git ls-tree -r --name-only main .claude prints nothing). /verify has no recipe and nothing to drive.

After: real drive.py runs on 2026-10-13. Each run used one sandbox under $TMPDIR/ra-verify-*.

new → first sync in the sandbox:

2026-10-13T15:20:11Z verify: 3 sessions, 0 summaries, 1 redactions, pushed

kb <dir> -- find flaky:

page     2026-10-01 project demo-app           demo-app · …lock and rerun (5e6f7a8b) ## Open threads - Retry the «flaky» fetch test (5e6f7a8b) [kb page demo-app]
e94ad30f 2026-10-06 claude demo               ↳55555555 Explore tests · Find «flaky» tests [turn 1]
55555555 2026-10-06 claude demo               Fix flaky motion test · Fix the «flaky» test in tests/motion.spec.ts [turn 1]

hook <dir> "Run exactly this Bash command: sh -c 'echo \"ModuleNotFoundError: No module named yaml\" >&2; exit 1'" ($0.0028):

• SessionStart (kb-hook): Project page of demo-app: kb page demo-app (1 open thread, 2 errors → fixes). ... · - [s-000001] Rules · Wait for CI with a watch, not a sleep loop · 1 retro suggestion waits for the user to accept or reject ...
• PostToolUseFailure (kb-hint): retroagent: seen before → make checkfails: ModuleNotFoundError no module named yaml → install the dev extras withpip install -e .[dev] (kb summary 1a2b3c4d)
• pluginLog: hooks module retroagent@inline loaded ..., $.process.run (retroagent): bin/kb exited 0 in 150ms (the decide mod's kb decide --json), PreToolUse:Bash ["bin/kb-guard"] ... status 0 (90ms), PostToolUseFailure:Bash ["bin/kb-hint"] ... status 0 (204ms)
• data.log and data.remoteLog did not change (auto_sync: false).

kb-guard deny, from a run of the first version of drive.py (15:12, before the review changes; the session code is the same apart from the log file names):

PreToolUse:Bash hook error: retroagent: only the user answers `kb decide` questions, in their own terminal. Tell the user which answer you suggest and why; do not run it.

SessionStart sync with new --config '{"auto_sync": true, "debounce_minutes": 0}', then kb <dir> -- status: last sync: 2026-10-13T15:14:29Z (the first sync of new ran at 15:14:06).

Guards: kb <dir> on the temp folder itself → not a sandbox: /private/var/folders/.../T. kb <dir> --timeout 0.01 -- status → "exit": "timeout" with the evidence JSON. The drive found this bug: before the fix, --timeout went to kb as an argument.

scripts/test --all: 1726 passed in 81.06s.

Reading map

Config:

• .claude/skills/verify/seed/pages/suggestions.json, .claude/skills/verify/seed/pages/decisions.json (sandbox seed data, made up)

Error handling:

• .claude/skills/verify/drive.py:71 (raise SystemExit for a folder that is not a sandbox), :84-90 (kb timeout), :150 and :182 (session kill timer), :159 (BrokenPipeError), :168 (JSONDecodeError)

Stack

Step 1/1. Previous: none. Next: none.

Story

Add the verification loop (/simplify, /code-review, checks, /verify) and a /verify recipe for retroagent.

#	Step	Evidence	PR
1	Loop rule in CODING_STANDARDS.md and docs/development.mdx. /verify recipe: .claude/skills/verify/SKILL.md, drive.py (new, kb, hook, clean) and the seed pages.	Real drive.py output for the kb and hook probes. scripts/test --all passes.	this PR (done)

🤖 Generated with Claude Code