What happened
On PR #12 (#12), the review agent flagged customer-research in agents/prioritize.md frontmatter as a missing dependency 4 times across 4 review iterations. The human author dismissed it. The retro agent then filed issue #13 as a bug, triage confirmed it, and the code agent produced PR #16 — an incorrect fix that removed the skill reference. ralphbean closed PR #16 and issue #13, explaining customer-research is an intentional extension point provided by target repos at org level. The retro agent filed issue #18 proposing an inline YAML comment annotation, but #18 was closed as a duplicate of #17. PR #19 (#19) implemented #17 by adding general skill resolution documentation to AGENTS.md section 6, but did not carry forward the inline annotation from #18.
What could go better
The general documentation in AGENTS.md section 6 explains the multi-source skill resolution chain, but an agent reading agents/prioritize.md would need to also consult AGENTS.md to understand that customer-research is intentionally absent from this repo. The inline annotation proposed in issue #18 was a defense-in-depth measure that provides context at the point of contact. It was lost when #18 was closed as a duplicate. Confidence is moderate — the AGENTS.md fix may be sufficient on its own, but the inline annotation adds a second layer of protection with minimal cost.
Proposed change
Add an inline YAML comment in agents/prioritize.md frontmatter annotating customer-research as an extension point:
skills:
- customer-research # extension point: provided by target repos, not built into this repo
Apply the same annotation pattern to any future extension point skill references in agent frontmatter files. This provides context at the point where agents encounter the reference, not only in a separate guidance document.
Validation criteria
Over the next 3 PRs that touch agent definition files with cross-repo skill references, the review agent should not flag extension point skills as missing dependencies. The inline comment should be sufficient context even if the agent does not read AGENTS.md section 6.
Generated by retro agent from #19
What happened
On PR #12 (#12), the review agent flagged
customer-researchinagents/prioritize.mdfrontmatter as a missing dependency 4 times across 4 review iterations. The human author dismissed it. The retro agent then filed issue #13 as a bug, triage confirmed it, and the code agent produced PR #16 — an incorrect fix that removed the skill reference. ralphbean closed PR #16 and issue #13, explainingcustomer-researchis an intentional extension point provided by target repos at org level. The retro agent filed issue #18 proposing an inline YAML comment annotation, but #18 was closed as a duplicate of #17. PR #19 (#19) implemented #17 by adding general skill resolution documentation to AGENTS.md section 6, but did not carry forward the inline annotation from #18.What could go better
The general documentation in AGENTS.md section 6 explains the multi-source skill resolution chain, but an agent reading
agents/prioritize.mdwould need to also consult AGENTS.md to understand thatcustomer-researchis intentionally absent from this repo. The inline annotation proposed in issue #18 was a defense-in-depth measure that provides context at the point of contact. It was lost when #18 was closed as a duplicate. Confidence is moderate — the AGENTS.md fix may be sufficient on its own, but the inline annotation adds a second layer of protection with minimal cost.Proposed change
Add an inline YAML comment in
agents/prioritize.mdfrontmatter annotatingcustomer-researchas an extension point:Apply the same annotation pattern to any future extension point skill references in agent frontmatter files. This provides context at the point where agents encounter the reference, not only in a separate guidance document.
Validation criteria
Over the next 3 PRs that touch agent definition files with cross-repo skill references, the review agent should not flag extension point skills as missing dependencies. The inline comment should be sufficient context even if the agent does not read AGENTS.md section 6.
Generated by retro agent from #19