model-escalation

Advisory model escalation based on task outcome tracking

SKILL.md

GitHub

Model Escalation Skill

Tracks task outcomes and advises model upgrades when failures are detected. Advisory-only — the orchestrator makes the final decision (R010).

Escalation Path

haiku → sonnet → opus

Trigger Conditions

Condition Action
2+ failures with same model for same agent type Advise escalation
3+ consecutive failures across any agent type Advise global escalation
Sustained success after escalation Advise de-escalation

Thresholds

Parameter Default Description
failure_threshold 2 Failures before escalation advisory
consecutive_threshold 3 Consecutive failures for global advisory
cooldown_tasks 5 Successes before de-escalation advisory

Cost Guard

  • Advisory includes estimated cost multiplier
  • De-escalation suggested after sustained success at higher tier
  • Cost tracked per session via PPID-scoped temp file

Architecture

PostToolUse (Task) → task-outcome-recorder.sh
  Records: agent_type, model, success/failure, error_summary
  Storage: /tmp/.claude-task-outcomes-$PPID (JSON lines, max 50)

PreToolUse (Task) → model-escalation-advisor.sh
  Reads outcomes → counts failures → advises escalation via stderr
  Advisory only — never blocks, never modifies tool input

Advisory Format

--- [Model Escalation Advisory] ---
  Agent type: {agent_type}
  Current model: {current_model}
  Recent failures: {count}/{threshold}
  ⚡ Recommended: Escalate to {next_model}
  Cost impact: {multiplier} per task
---