Harness Inspect¶
Two modes. Choose based on available time and the trigger:
| Mode | Scope | When |
|---|---|---|
| Health check | Quick pass | Monthly; after any structural change, make target rename, or harness file edit |
| Full audit | Thorough assessment | Quarterly; after an agent incident, major refactor, or new AI tool adoption |
If the health check surfaces issues, escalate to a full audit.
Instructions¶
Mode selection¶
Check for an argument before proceeding:
| Argument | Action |
|---|---|
health-check |
Run the health check directly — skip to the Health check section below |
full-audit |
Run the full audit directly — skip to the Full audit section below |
| (none) | Ask the user: "Which mode — health-check (quick pass, monthly) or full-audit (thorough assessment, quarterly)?" Do not assume a default. |
Before starting (optional baseline)¶
Load references/before-after-probe.md to record current agent behavior before making changes. Run again after fixes to measure impact.
Health check¶
Load references/health-check.md and run the prompt twice independently (without referencing your first pass on the second run). Compare pass/fail verdicts; resolve any disagreements. Report pass/fail for each of the ten questions and list action items only — no full analysis.
Full audit¶
Load references/full-audit.md and run the prompt twice independently (without referencing your first pass on the second run). Compare verdicts; resolve disagreements. For repeat runs, fill in the inspection mode and paste the prior verdict table — the agent will produce a delta automatically.
Processing results¶
| Finding type | Action |
|---|---|
| Rot | Fix immediately — stale guidance actively misleads agents |
| Critical gaps | Address before the next agent task in the affected area |
| Quick fixes | Address in the current session — they compound if deferred |
| Token optimizations | Address when convenient; reduces initialization cost over time |
| Larger improvements | Add to project backlog with the assessment date as context |
Before acting on any finding immediately (rot, quick fixes), write a specific plan — file, change, reason. If your tool supports a planning mode (e.g., Claude Code /model opusplan), use it for the audit run itself, then switch to execution mode for fixes.