| name | deep-analysis |
| description | Use when the user asks to analyze, investigate, explain why something happens, or provide a grounded cross-file explanation without modifying code. |
Read-Only Deep Analysis
Answer the user's actual question through repository evidence. This workflow is read-only: do not edit files or silently turn diagnosis into implementation.
Method
- Restate the question narrowly.
- Inspect the smallest set of relevant files, tests, configuration, history, and generated artifacts.
- Rank plausible explanations by evidentiary strength.
- Separate direct evidence from inference and unknowns.
- Stop when additional inspection would not materially change the answer.
Use parallel exploration only for independent search surfaces. Do not delegate trivial lookups.
Evidence rules
- Cite concrete file paths and tight line references for material claims.
- Label conclusions as evidence, inference, or unknown.
- Do not present architectural preference as repository fact.
- Prefer current code and tests over stale prose.
- When sources conflict, explain the conflict rather than selecting one silently.
Output
Lead with the conclusion, followed by:
- ranked synthesis;
- strongest supporting evidence;
- material inferences;
- unresolved unknowns or limits.
If a next step would reduce uncertainty, recommend one discriminating read-only probe. Do not append a generic implementation plan.
