Repos change after you configure them. You add a wiki/ dir, a CHANGELOG, a Dependabot config — and the rationale stored in project.json slowly stops describing the repo you actually have. Re-detection re-runs the same rules and shows you the gap. It surfaces the diff; it does not act on it.

1. Run the diff. Say "re-detect this project" or run /setup --redetect. Either way the underlying call is:

   python3 scripts/project_config.py redetect .

   You get one block per target that moved, each naming the rule behind it:

   Re-detected my-app.
   Rules matching now: R-wiki, R-pii
   
   Proposed changes:
     [newly-detected] skill diataxis-author
         a rule now justifies this — refresh its rationale
         now: R-wiki — Found wiki/ dir -> diataxis-author manages Diataxis-shaped documentation.
   
   Suppressed — you already decided these, so re-detect will not re-suggest them:
     [newly-detected] skill pii-scrubber — declined on 2026-06-01T00:00:00Z (.envrc is direnv, not a secret)
   
   Nothing was changed. Re-run with --apply to refresh the detection rationale.
   

2. Read the suppressed list. Anything you declined at registration lands there instead of in the proposal, with your original reason attached. That list is how you check the harness is still honoring a past decision — if a skill you turned off shows up under Proposed changes rather than Suppressed, the override didn't record, and you should re-run the decline.

3. Decide. The run above changed nothing except the last_redetect_at stamp. Your options:

   • Keep the stored rationale — do nothing. The diff is informational.
   • Refresh the rationale — python3 scripts/project_config.py redetect . --apply. This writes auto_detected, rule_id, and rationale. It never touches enabled, so applying cannot turn a skill off.
   • Turn something off — that's an override, not a detection result. Record it through the register path: python3 scripts/project_config.py register . --disable <name>. From then on, re-detect suppresses suggestions about that target.

4. Confirm it converged. Re-run the plain command. A clean repo prints No changes — the stored config still matches what this repo looks like and exits 0.

The default run stamps last_redetect_at, because that field records when the scan last ran. For a look that touches the file not at all:

python3 scripts/project_config.py redetect . --dry-run

Use --format json for the machine-readable diff and read the exit code: 0 means the config matches the repo, 1 means changes were surfaced, 2 means re-detect couldn't run — the repo has no enablement block yet, or it's the harness source repo, which detection bypasses.

python3 scripts/project_config.py redetect . --format json

• Configure a new project — the first-session flow that writes the block this compares against.
• Project config — the change categories, the exit codes, and the enablement-block schema.
• Detection rules — the rules being re-run and what each attaches a rationale to.
• Auto-detect + auto-configure — why re-detection surfaces a diff instead of applying one.