| name | session-review-patterns |
| description | Patterns for reviewing kiro-cli session transcripts. Use when analyzing sessions for protocol compliance, performance issues, or crew improvements. |
Session Review Patterns
What to Look For
Protocol Compliance
- Did workers emit structured DONE/BLOCKED/FAILED signals?
- Did orchestrator verify results before proceeding?
- Was verification actually run (not just claimed)?
- Were assumptions surfaced in DONE signals?
Performance Issues
- Context overflow (agent forgets earlier instructions mid-session)
- Repeated failures without strategy change (retry amplification)
- Orchestrator doing worker work (role bleed)
- Workers asking open-ended questions instead of researching
Crew Design Issues
- Agent prompt too long (> 80 lines = split signal)
- Skills triggering when irrelevant (vague descriptions)
- Tool/rule contradictions (allowedCommands permits what prompt forbids)
- Dead config (hooks in base crew.yaml ignored in full mode)
Hook Gotchas
Only agentSpawn reliably deploys in full-mode projects (those with crews/ directory). Other hooks (stop, postToolUse) must be defined in the themed crew YAML under the archetype, not in the base crew.yaml.
Subagent Limitations
When spawned as subagent, these tools are NOT available:
grep, glob, code, web_search, web_fetch, use_aws, todo_list
If a session shows an agent failing because it tried to use web_search as a subagent — that's a crew design issue, not an agent issue.
Red Flags in Transcripts
- "I'll try again" without changing approach → retry amplification
- Agent produces output for wrong audience → role confusion / split needed
- 3+ consecutive agent decisions without user input → rubber-stamp guard should fire
- "Done!" with no verification evidence → hallucinated success
- Agent modifies .kiro/agents/ or .kiro/steering/ → deniedPaths violation