Problem
When a workflow step declares a file_path output that references pre-existing, unchanged files (e.g., listing reference files as outputs), the quality gate triggers reviews on those files. This forces the agent to either run unnecessary reviews or use quality_review_override_reason to skip them.
Example
The initial_understanding step in the deepplan workflow outputs key_reference_files — a list of existing files the agent found relevant. These files weren't modified, but because they're declared as step outputs, .deepreview rules match against them and trigger review tasks.
Expected Behavior
Reviews triggered by the quality gate should only run on files that were actually created or modified during the step, not on pre-existing unchanged files passed through as references.
Possible Approaches
- Content hash comparison: Compare file content hashes before/after step execution; only review files whose content changed
- Output metadata: Allow step outputs to distinguish between "produced" and "referenced" files (e.g., a
reference_only flag on the output ref)
- Separate argument type: Use a different step argument type for reference-only file lists that doesn't trigger reviews
Context
Discovered during end-to-end testing of the DeepPlan workflow (PR #331). The key_reference_files output in the initial_understanding step triggered 5 review tasks on files like plugins/claude/skills/deepwork/SKILL.md that were merely listed as context, not modified.
Problem
When a workflow step declares a
file_pathoutput that references pre-existing, unchanged files (e.g., listing reference files as outputs), the quality gate triggers reviews on those files. This forces the agent to either run unnecessary reviews or usequality_review_override_reasonto skip them.Example
The
initial_understandingstep in thedeepplanworkflow outputskey_reference_files— a list of existing files the agent found relevant. These files weren't modified, but because they're declared as step outputs,.deepreviewrules match against them and trigger review tasks.Expected Behavior
Reviews triggered by the quality gate should only run on files that were actually created or modified during the step, not on pre-existing unchanged files passed through as references.
Possible Approaches
reference_onlyflag on the output ref)Context
Discovered during end-to-end testing of the DeepPlan workflow (PR #331). The
key_reference_filesoutput in theinitial_understandingstep triggered 5 review tasks on files likeplugins/claude/skills/deepwork/SKILL.mdthat were merely listed as context, not modified.