Scenario
post-merge-retrospective instructs the agent to confirm which pipeline gates fired, using tests/skill-activation-audit.sh "(this repo)" plus .claude/autodev-state/in-progress.jsonl. In a consuming project (where the kit is installed, not developed) that audit script does not exist — it lives in the kit's own test suite, not the user's repo. As a result the retro's "Missed skill activations" table comes out as a column of "no evidence / script does not exist" every single time.
Separately, several pipeline gates (e.g. alignment-check, requesting-code-review) leave no durable activation artifact at all, so even reading in-progress.jsonl does not reliably confirm whether they ran. The retro therefore cannot distinguish "gate was skipped" from "gate ran but left no trace."
Impact
- "Missed activation" findings are unreliable — the retro can't tell skipped-vs-unrecorded.
- Users are told to run a script that isn't present in their repository (confusing, looks broken).
Recommendation
- Don't point the retro at a project-local script. Ship the activation-audit capability with the kit itself (script under the plugin's own path, or a retro step that reads only
.claude/autodev-state/*.jsonl).
- Make every pipeline gate append a compact activation record (skill, phase, timestamp, pass/fail) to one state file under
.claude/autodev-state/, so the retro has a single source of truth.
- Have the retro degrade gracefully ("activation log unavailable") instead of emitting "script does not exist" rows.
Scenario
post-merge-retrospectiveinstructs the agent to confirm which pipeline gates fired, usingtests/skill-activation-audit.sh"(this repo)" plus.claude/autodev-state/in-progress.jsonl. In a consuming project (where the kit is installed, not developed) that audit script does not exist — it lives in the kit's own test suite, not the user's repo. As a result the retro's "Missed skill activations" table comes out as a column of "no evidence / script does not exist" every single time.Separately, several pipeline gates (e.g.
alignment-check,requesting-code-review) leave no durable activation artifact at all, so even readingin-progress.jsonldoes not reliably confirm whether they ran. The retro therefore cannot distinguish "gate was skipped" from "gate ran but left no trace."Impact
Recommendation
.claude/autodev-state/*.jsonl)..claude/autodev-state/, so the retro has a single source of truth.