| name | codex-productivity-auditor |
| description | Audit a project and recent Codex session behavior to recommend Codex productivity setup across AGENTS.md, skills, hooks, MCP/plugins, subagents, local environment actions, automations, memories, and workflow docs. Use when the user asks how to make repeated steering second nature or asks what Codex setup is missing. |
Codex Productivity Auditor
Recommend the smallest durable Codex setup changes that make repeated user
steering automatic without adding friction.
Workflow
- Inspect project setup:
AGENTS.md, .codex/config.toml,
.codex/agents/, skills, hooks, workflow docs, scripts, and tests.
- Run deterministic session analysis:
scripts/transcript_signal_scan.py --project <project-or-path>
- Load only the references needed for the recommendation:
- session signals:
references/session-analysis-rubric.md
- owner placement:
references/owner-surface-rubric.md
- hooks, automations, MCP/plugins, or always-loaded rules:
references/friction-risk-rubric.md
- subagent/model choices:
references/model-routing-rubric.md
- signal vocabulary:
references/signal-taxonomy.md
- Check official OpenAI Codex docs when product behavior may drift.
- Recommend at most five setup changes, plus what not to configure yet.
Fixed Decisions
- Raw Codex JSONL under
~/.codex/sessions and
~/.codex/archived_sessions is the canonical session-analysis source.
- Do not use qmd or Obsidian for scoring or recommendations; use them only for
explicit human-readable lookup or semantic discovery requests.
- For repeated local run/test/log command friction, prefer Codex local
environment actions over hooks or automations.
- Do not hand-write local-environment action schema. Create the first action
through Codex app settings, inspect the generated project
.codex artifact,
then edit only with that schema in hand.
Recommendation Shape
Top Recommendations
1. <setup change>
Evidence: <signal counts and pointers>
Owner: <surface>
Change: <exact action>
Confidence: high|medium|low
Validation: <command/proof>
Not Recommended Yet
- <feature>: <why>
Loopholes Checked
- <loophole>: <fix or downgrade>
Keep the answer compact. Do not paste long transcript excerpts; cite counts,
JSONL file paths, line numbers, and command evidence.
