[Phase 6] Research Review (8-reviewer panel)

[jeremymanning]

Summary

Eight specialist reviewers + a generalist vote on accept / minor_revision / full_revision / reject.

What this phase does

Dispatches the generic research_reviewer plus all 7 specialist reviewers (idea_quality, creativity, implementation_correctness, implementation_completeness, code_quality_research, data_quality_research, filesystem_hygiene). Each writes a review record under projects/<PROJ-ID>/reviews/research/<author>__<YYYY-MM-DD>__research.md with frontmatter (verdict:, score:, artifact_hash:).

Project lifecycle stages handled

• research_complete
• research_review

Dependencies

• Depends on: [Phase 5] Implementation #49
• Blocks: [Phase 7] Advancement & Verdict Routing #51

Agents in this phase

• sub-issue: research_reviewer ([Agent · Phase 6] research_reviewer #68)
• sub-issue: research_reviewer_idea_quality ([Agent · Phase 6] research_reviewer_idea_quality #73)
• sub-issue: research_reviewer_creativity ([Agent · Phase 6] research_reviewer_creativity #70)
• sub-issue: research_reviewer_implementation_correctness ([Agent · Phase 6] research_reviewer_implementation_correctness #75)
• sub-issue: research_reviewer_implementation_completeness ([Agent · Phase 6] research_reviewer_implementation_completeness #74)
• sub-issue: research_reviewer_code_quality_research ([Agent · Phase 6] research_reviewer_code_quality_research #69)
• sub-issue: research_reviewer_data_quality_research ([Agent · Phase 6] research_reviewer_data_quality_research #71)
• sub-issue: research_reviewer_filesystem_hygiene ([Agent · Phase 6] research_reviewer_filesystem_hygiene #72)

Manual test plan (phase-level smoke test)

1. Stage a project at the entry stage of this phase (see "stages handled" above)
2. python -m llmxive run --project <PROJ-ID> --max-tasks N (N = number of agents in this phase)
3. Verify the project transitions through every stage listed above
4. Run each agent's manual test plan (linked via the sub-issue checklist)
5. Verify state ends at the entry stage of the next phase, or human_input_needed/blocked on legitimate failure (never silently stuck)

Acceptance criteria (phase-level)

• Every agent sub-issue passes its acceptance criteria
• Phase-level smoke test passes end-to-end on a fresh project
• No silent shortcuts — agents that can't do their work MUST fail loudly (write human_input_needed.yaml or emit verdict: failed), not mark tasks [X] with placeholder content
• All artifacts written by this phase pass schema validation (where applicable)
• Run-log entries record outcome, started_at, ended_at for every agent invocation