research-loop（研究循环）

Research Loop Compatibility Skill

This is the Claude Code compatibility mirror. The primary skill lives at:

.codex/skills/research-loop/

Use this mirror only when the runtime is Claude Code. Keep the same safety contract:

1. Treat the current target repo HEAD as primary truth.
2. Treat MemPalace, prior handoffs, and expert memory as secondary context.
3. Use native web search only for short, attributed fresh context when needed.
4. Do not require Ollama API keys or Ollama native web search.
5. Run one bounded micro-step per cycle.
6. Write source changes into the workflow run's source_changes/ repo-root mirror.
7. Write the executor manifest with python3 -m autonomous_research_workflow.cli write-executor-manifest.
8. Do not mark empirical progress unless runnable code or checkable result artifacts were produced.
9. Run the protocol hygiene checks in protocol hygiene before choosing a micro-step.
10. Run the innovation-frontier checks in innovation frontier so current assets calibrate the loop without trapping it.
11. Stop only when Submission Advisor reports score >= 100 and level = submission_ready, executor completed, and delivery completed.

Phase Order

Run the 15-phase loop described in phases:

1. source_intake
2. mempalace_context
3. council_knowledge
4. expert_forum_coverage
5. literature_refresh
6. knowledge_publish
7. advisor_design
8. execution_packet
9. executor_run
10. advisor_reflection
11. openspace_retrospective
12. next_cycle_handoff
13. star_office_status
14. overwatcher_check
15. github_publish

Native Executor

Phase 9 remains the key agent-native phase:

• Read the execution packet.
• Read the source clone.
• Generate the smallest useful runnable research asset.
• Write only to source_changes/.
• Run lightweight checks when possible.
• Write executor_manifest.json.

Use executor-template as the output contract.

Protocol Hygiene

Before the advisor or executor treats a repo as submission-ready, check for:

• positioning drift between README claims and later submission/package locators
• synthetic or structural claims being mislabeled as empirical results
• tracked reproducibility artifacts that rewrite timestamps or other volatile fields
• CI gates that omit confirm scripts or fail to check dirty diffs
• premature dual-venue framing before empirical results exist
• misleading names such as result_quality_* for input-quality analysis
• weak contamination checks being promoted into strong evidence

Use protocol hygiene as the detailed rubric.

Innovation Frontier

Current assets are the evidence floor, not the imagination ceiling. Before selecting a micro-step, check whether the loop is overfitting to incomplete benchmarks, toy rows, inherited file names, or prior advisor backlog. Use innovation frontier to keep at least one evidence-linked but asset-transcending research route alive.

Completion Gate

Only stop when all are true:

• Submission Advisor score is at least 100.
• Submission Advisor level is submission_ready.
• Executor manifest status is completed.
• Delivery status is published or explicitly local_only for dry runs.