| name | self-recovery-limits |
| description | Canonical registry of self-recovery limits for subagents: how many attempts are allowed for self-fix / hypotheses / review and what to do after exhaustion. |
| alwaysApply | true |
Self-Recovery Limits (Canonical Registry)
A single source of numeric limits for subagents. Profiles reference this registry; the numbers in the profiles
duplicate the table and must not diverge from it. When the limit is exhausted, the agent does NOT continue
“blindly” — it switches to the specified action (bug-report or escalation to the orchestrator/user).
| Context | Who | Limit | What after exhaustion |
|---|
| Self-fix in the implementation's own code | developer-code | 2 attempts | create bug-report.json → STOP (orchestrator → Debugger) |
| Red gate: the reason for the "green" state before prod code is not found | scenario-coder | 2 attempts | create bug-report.json → STOP |
| Fixing a technical test error | tester | 3 attempts | create bug-report.json (when the runtime defect is not obvious) → STOP |
| Hypothesis investigation cycle | debugger | 5 hypotheses (+3 by agreement with the orchestrator → max 8) | escalation to the user (escalated_to_user) |
| BLOCK iterations of artifact review | reviewer / orchestrator | 3 iterations | escalation to the orchestrator/user |
| Review of the debugger's local fix (scope=debug) | reviewer / orchestrator | 1 iteration | escalation to the orchestrator/user |
| bug → fix → bug again cycle on the same artifact | orchestrator | 2 cycles | escalation to the orchestrator/user |
depends_on: []
