Research-only mode (R17 strict enforcement)
When user message has NO action verb (правь, делай, фикси, добавь, реализуй, fix, build, implement, edit) — you are in RESEARCH mode.
What you DO
- Read code via Read, Glob, Grep, Serena
find_symbol,get_symbols_overview - Fetch external docs via Context7 MCP, WebFetch, WebSearch
- Use Explore subagent for wide search (delegates to fresh context, doesn't pollute main)
- Synthesize findings into structured text answer
What you DO NOT
- ✗ Edit any file
- ✗ Write any file
- ✗ Run state-changing commands (git, npm install, alembic, etc.)
- ✗ Make commits
- ✗ Modify git state
Format of answer
Short summary (2-3 sentences):
...
Key findings (bulleted):
- Finding 1 (file:line)
- Finding 2 (file:line)
...
Proposed next steps (R4 — propose, don't ask):
- I would do X first because Y. Stop me if different direction.
- Then Z. Then W.
If you want me to do <X> — say so explicitly.
When research uncovers serious problem
- DO NOT secretly fix it — that breaks research-only contract
- DO record finding in
docs/agent/06_known_issues.mdif explicitly broken - DO propose fix in your text answer, but do not execute
Subagent delegation
For wide research (>5 files, >3 web searches) — delegate via Agent tool with subagent_type="Explore":
- Subagent gets isolated context, returns ≤200-word summary
- Main context stays clean (R7 budget)
- This is mandatory for repo-wide audits, not optimization
Anti-patterns
- ✗ "Research" the bug, then quietly fix it without permission
- ✗ Ramble for 1000 words because "research = thorough"
- ✗ End with "Should I do X?" → use R4 propose format instead