Skip to main content

debug

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).

Source facts

Repository
markusstrasser/skills
Last source activity
October 18, 2026 at 15:02
Detected SKILL.md language
English
Stars
0
Forks
0

Install options

The review-first prompt is selected by default. You can switch to a direct command or download a local copy.

Review the source files

Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.

Showing SKILL.md

SKILL.md
Source instructions · Read-only preview
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