| name | anti-hall-root-cause |
| description | Codex-native root-cause debugging discipline. Use when investigating a bug, failing test, broken config, or runtime behavior before proposing or applying a fix. |
anti-hall root cause for Codex
Apply the Iron Law:
No claim without evidence; no fix without a proven root cause.
Workflow:
- Capture the observed failure exactly: command, input, output, stack trace, affected file, or UI state.
- Verify the relevant code/config/data with tools before making factual claims.
- Trace from symptom to root cause. Do not patch the first suspicious line.
- If evidence is insufficient, add targeted instrumentation or request the exact missing repro/log.
- Apply the smallest fix that addresses the proven cause.
- Re-run the authoritative check this turn before saying fixed/passing.
Codex model routing:
- Ambiguous diagnosis, architecture, or safety-sensitive analysis:
gpt-5.6-sol
- Implementation once cause is proven:
gpt-5.6-terra
- Simple command execution / file lookup subtask:
gpt-5.4-mini (default; gpt-5.6-luna available when 5.6-era capability/cutoff matters) — gpt-5.3-codex-spark is a distinct, faster/less-capable model, ChatGPT Pro only
Do not copy raw noisy command output into the final answer. Summarize findings with evidence references.