pitwallCAUTION

The agent loop that queues your work, proves every PR, and leaves the merge to you.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: pitwall
description: Single entry point of the agent loop — reads the request (Thai / English), picks a playbook, then does it with evidence. Use when the user runs /pitwall:pitwall or says "ใส่คิว", "ลุย", "ออกรถ", "ทํา 1 งาน", "หยุดหลังจบงานนี้", "หยุดเดี๋ยวนี้", "merge ตัวไหนก่อน", "แก้บั๊ก", "เทสด้วยกัน", "เปิด task", "ผมย้ําอะไรบ่อย", "วางแผน", "สัมภาษณ์ผม", "แบ่ง phase", "ให้ทําเอง", "งานไหน agent ทําได้", "PR ไหนพร้อม" · repo values live in .claude/agent-loop.md
user-invocable: true
---

# pitwall

The skill holds no repo-specific values: everything that belongs to the repo (base, check commands, browser, human-only steps, tracker, plan, recipes for building states) lives in the profile `.claude/agent-loop.md` · no profile → only `init` and `doctor` run; every other request stops with "this repo has not set up pitwall yet — run `/pitwall:init`" · never guess values the profile should hold

**pitwall dir** = `${CLAUDE_SKILL_DIR}` (this skill's folder: `playbooks/`, `principles/`, `scripts/`) · every `playbooks/...`, `principles/...`, `scripts/...` path in this skill and in the playbooks is relative to it, and `<pitwall dir>` in a command means this path

## Rules you cannot skip

In the report, state which rule changed which decision

1. **Prove it against the real thing** — the real page / request / value, not a stand-in · typecheck passing does not mean the UI works
2. **A check the agent can run itself must never be handed to a human** — ⏳ is allowed only for "the 3 wait cases (⏳)" at the top of `playbooks/lap.md` (`1 human-only step while the user is away` · `2 needs the dev env after merge` · `3 environment unusable`), with a specific reason
3. **Ask everything once, up front; never mid-round** — questions that belong to the human are asked with AskUserQuestion as tick boxes, in the profile's user language, recommended option first, marked with the profile's recommended marker; gather every point of the task into a single round (before `/pitwall:drive` = briefing in `playbooks/queue.md`)
4. **Loop work uses the 3-step ladder** (next section) — every task that runs through the `queue` / `drive` / `lap` playbooks (including `/pitwall:lap` typed by the user) · work outside the loop that the user orders directly (edit files as told, answer questions) uses "Direct work" below
5. **Humans merge** — the agent builds, proves, opens the PR · never merge, force-push, or touch branches the profile forbids
6. **Lessons become structure first** — script / lint / test / CI > a step in a playbook > text in rules > memory · writing the same instruction a second time = it should be a check
7. **Record decisions when you make them** — a row in the task's `decisions.tsv` (state folder), not written back at the end

## The 3-step ladder (loop / autopilot)

When you hit a fork during a run that has no answer in state (`answers`):

| Step | Use when | What to do |
|---|---|---|
| 1 Try it | the answer can be seen by running something (what an endpoint returns, shape, behavior, layout, timing) | run / prototype in a throwaway branch, then decide from the result · put the result in the PR · **read-only, or on local / testnet only** — never write data on a shared backend, no transactions, never touch anything on the "Always park" list |
| 2 Default + record | reversible, and there is a default in the rules / playbook (port = follow the source, existing behaviour = keep as is, UI = closest to the existing one) | follow the default · a row in `decisions.tsv` · PR section **Decisions to confirm** (one line each: option taken / other options / where to change it) · continue |
| 3 Park | on the "Always park" list, or no safe default | park as usual (`playbooks/lap.md`) + keep the question to ask as tick boxes next round |

**Classify before parking or asking** (the one definition — `/pitwall:queue` and `/pitwall:lap` point here) · **lists first**: an item on "Always park" / "Never default either" is parked / asked even when it can be run — a prototype only attaches evidence to the question, never answers it · *observable* = a fact a run reveals (what X does: behavior, timing, output, what an endpoint returns), not a choice (which X it should be) → step 1, the prototype result goes with the decision, not a question for the human · *not decidable by experiment* (business rule, backend / third-party value, design with no design file or source, a plan open question) → park in the run / ask in the briefing · a plateau is not a stop: change method before parking (`playbooks/lap.md` §4)

**Always park** (never use a default): money / price / fee · permissions / security / user data · deletes / irreversible · backend / third-party values not yet provided (placeholder only where the plan allows)

**Never default either** ("Never guess about" in "Direct work") if briefing has not answered it → park: business rules (limit, expiry, eligibility), API contract shape, which variant / environment / tenant a feature applies to, new design with no design file or source to compare against · the default "UI closest to the existing one" applies only to UI that has an existing version / source to anchor to

## Direct work (outside the loop)

Work the user gives directly while present: **below 95% sure of the approach → stop and ask before acting** — state what you understood, what you are unsure of, the options, and the one you lean toward and why · ask when unsure which file / pattern to use, when the task reads two ways, before an architectural decision (new file, new state, new pattern), before a refactor, before deleting or rewriting working code

**Never guess about**: business rules (pricing, limits, eligibility) · API contracts (endpoints, payloads, response shapes) · design decisions (layout, spacing, colours) with no design reference · which variant / environment / tenant a feature applies to · whether to create a new file or change an existing one

At 95%: a clear bug with an obvious cause, a pattern that al

CAUTIONgrade C · trust 79/100 Install with care. The audit found things worth knowing before you trust its output.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (1)

4. Has `currentTask` → **don't touch the task**, tell the user: "will stop after `<id>` finishes (currently at step `<step>`)" — the current `/pitwall:lap` round works until the PR is opened, then sto

Gates applied: instruction_override, no_behavioural_pass.