- name
- debug
- description
- Use when: parallel adversarial bug-hunt fan-out or pre-commit audit; scouts write files, orchestrator model triages. NOT a diff-review scout (/diff-review) or plan critique (/critique).
- user-invocable
- true
- argument-hint
- <repo> [scope] [extra prompt]
- allowed-tools
- ["Read","Glob","Grep","Bash","Write"]
- effort
- medium
# Debug scout pipeline
Formalizes the genomics bug-hunt pattern: **scouts write, orchestrator model judges, operator steers at ship boundaries.** (The file-bus pipeline that once wrapped this — `/orchestrate`, consolidation, commit-slice — was retired 2026-09-02; the parent session triages `docs/audit/` inline.)
## Roles
| Role | Who |
|------|-----|
| **Operator** | Human — goals, approval, irreversible apply |
| **Orchestrator model** | Frontier parent (Opus, etc.) — dispatch, triage, propose fixes |
| **Scout** | GPT-6 Astra low in a read-only codex sandbox (`--backend codex --effort low`) — audit files only |
## Flow
```bash
J="just -f ~/Projects/agent-infra/justfile"
$J adversarial-debug-scout ~/Projects/genomics recent \
--backend codex --effort low --prompt "cardinal rule + RESEARCH_ONLY"
$J debug-until-dry ~/Projects/genomics & # or: wave loop until no new confirmed bugs
# orchestrator model reads docs/audit/*-debug-handoff.md — validates SUSPECT, fixes confirmed
# findings inline, notes rejections in the memo; operator approves tier-1/2 apply.
```
**arc-agi multi-round audits:** before round 2+ scouts, run `just -C ~/Projects/arc-agi audit-ledger skip-fixed-md` and prepend the markdown block to each scout prompt. Ledger: `loop/audit/ledger.jsonl`. After fixes: `just -C ~/Projects/arc-agi audit-ledger seed-from-git`.
## Rules
1. Scouts run **read-only only** (`agent-infra/scripts/debug_scout.py`) — no edits, no commit. Pass
`--backend codex --effort low`: the script's `cursor` backend still defaults to Composer 2.5,
retired 2026-10-07 (llmx and the cursor-model guard refuse it); `--backend cursor` needs an
explicit `--model grok-4.7-*`.
2. Output → `{repo}/docs/audit/YYYY-MM-DD-debug-{run}-{slug}.md`
3. **Orchestrator model** reads the handoffs and triages inline — scouts do not implement fixes.
5. **Operator** approves tier-1/2 apply and taste calls.
6. Check **scientific/conceptual** claims, not just code bugs (see `agent-infra/scripts/debug_scout_prompt.md`).
7. Scouts **verify when cheap**: git log, targeted pytest; skip costly/live/mutating → SUSPECT + hand off.
## vs other skills
| Skill | Role |
|-------|------|
| `/debug` (this) | Adversarial scout fan-out only |
| `/diff-review` | Structural line-format scout on source batches |
| `/critique` | Cross-model design/plan review |
Critique before large changes: `/critique model` on the handoff or plan packet.
View on GitHub