What happened
On PR #5957, the PR added a status-comment-id input to multiple reusable workflow files. The review agent ran 11 iterations (Aug 5–Aug 24) before discovering in run 12 that the prioritize workflow stage never receives this input — a consumer-completeness gap. The human reviewer (waynesun09's Aug 17 review) caught this much earlier. The agent's finding tag also changed between runs ([scope-alignment] → [consumer-completeness]), suggesting the agent stumbled onto the finding rather than systematically looking for it.
What could go better
The correctness sub-agent lacks specific guidance for tracing GitHub Actions reusable workflow input propagation. When a PR adds a new inputs: parameter to a reusable workflow, the agent should grep for all uses: references to that workflow and verify each caller passes the new input. This is analogous to the code-level consumer completeness in #795 but requires workflow-specific tracing (YAML uses: references, with: blocks, matrix expressions) rather than Go function call sites. The 11-run delay is strong evidence this pattern is not in the agent's current checklist — it was eventually found but likely by coincidence rather than systematic analysis. Confidence: high, based on the clear latency between when the code existed and when the finding appeared.
Proposed change
Add workflow-specific consumer-completeness guidance to the correctness sub-agent definition or the CI workflow correctness checks (#394) in fullsend-ai/agents. When the diff adds or modifies inputs: in a reusable workflow YAML file, the agent should: (1) identify all callers by grepping for uses: references to the modified workflow, (2) check whether each caller passes the new input in its with: block, and (3) flag any caller that omits the input. This is distinct from #795 (Go code functions) because the tracing mechanism and file format are different, and the failure mode is silent no-op rather than a compile error.
Validation criteria
The next review of a PR that adds a new input to a reusable workflow should include a consumer-completeness finding about missing callers on the first review run, not after 10+ iterations. Verify across 2 such PRs within 90 days.
Generated by retro agent from fullsend-ai/fullsend#5957
What happened
On PR #5957, the PR added a
status-comment-idinput to multiple reusable workflow files. The review agent ran 11 iterations (Aug 5–Aug 24) before discovering in run 12 that the prioritize workflow stage never receives this input — a consumer-completeness gap. The human reviewer (waynesun09's Aug 17 review) caught this much earlier. The agent's finding tag also changed between runs ([scope-alignment]→[consumer-completeness]), suggesting the agent stumbled onto the finding rather than systematically looking for it.What could go better
The correctness sub-agent lacks specific guidance for tracing GitHub Actions reusable workflow input propagation. When a PR adds a new
inputs:parameter to a reusable workflow, the agent should grep for alluses:references to that workflow and verify each caller passes the new input. This is analogous to the code-level consumer completeness in #795 but requires workflow-specific tracing (YAMLuses:references,with:blocks, matrix expressions) rather than Go function call sites. The 11-run delay is strong evidence this pattern is not in the agent's current checklist — it was eventually found but likely by coincidence rather than systematic analysis. Confidence: high, based on the clear latency between when the code existed and when the finding appeared.Proposed change
Add workflow-specific consumer-completeness guidance to the correctness sub-agent definition or the CI workflow correctness checks (#394) in
fullsend-ai/agents. When the diff adds or modifiesinputs:in a reusable workflow YAML file, the agent should: (1) identify all callers by grepping foruses:references to the modified workflow, (2) check whether each caller passes the new input in itswith:block, and (3) flag any caller that omits the input. This is distinct from #795 (Go code functions) because the tracing mechanism and file format are different, and the failure mode is silent no-op rather than a compile error.Validation criteria
The next review of a PR that adds a new input to a reusable workflow should include a consumer-completeness finding about missing callers on the first review run, not after 10+ iterations. Verify across 2 such PRs within 90 days.
Generated by retro agent from fullsend-ai/fullsend#5957