mdmagnuson-creator/dynamic-reassignment

Dynamic Reassignment Skill

⟳ When specialists fail, try alternatives before escalating.

Transient failures (rate limits) retry with backoff. Persistent failures (bad output, crashes) trigger reassignment to alternative agents. Checkpoints preserve progress across switches.

Triggers

• Agent delegation failure
• Rate limit during task execution
• Context overflow during implementation
• Verification contract failure
• "fallback", "reassign", "alternative agent"

Bundled Resources

• data/fallback-chains.yaml — Default fallback chains for all task types

---

Fallback Chain Lookup

When delegating a task, determine the fallback chain:

1. Read ~/.config/opencode/data/fallback-chains.yaml for toolkit defaults
2. Check project.json → agents.fallbackChains for project overrides
3. Merge chains (project takes precedence) unless override: true
4. Determine task type from files and description:
   • *.tsx, *.jsx → react-component
   • *_test.go → go-tests
   • *.spec.ts in e2e/ → playwright-tests
   • *.test.tsx → react-tests
   • Dockerfile* → docker
   • Default → generic

---

Failure Detection

Use verification contracts to detect failure:

| Outcome | Detection | Action | |---------|-----------|--------| | Success | All verificationContract.criteria pass | Task complete | | Verification failure | One or more criteria fail | Try alternative | | Rate limit | Error contains "429", "rate limit", "quota" | Retry with backoff | | Context overflow | Agent reports context limit | Checkpoint → fresh session | | Crash/error | Agent throws exception | Try alternative |

---

Rate Limit Handling

Rate limits are transient — retry same agent with exponential backoff:

⟳ Rate limited, retrying in 30s... (1/3)

Backoff schedule: 30s → 60s → 120s (3 retries max)

After 3 retries exhausted:

1. If alternatives exist → switch to next in chain
2. If no alternatives → escalate to user

---

Alternative Selection

When a specialist fails with non-transient error:

1. Create checkpoint (per session-log skill)
2. Look up fallback chain for task type
3. Find next alternative not already tried
4. Show status:

   ⟳ Switching to go-tester (react-tester failed: verification failed)
   

5. Delegate to alternative with checkpoint in prompt
6. Record attempt in chunk.json → reassignment.attempts[]

Try ALL alternatives in the chain before escalating.

---

Reassignment State

Track reassignment attempts in chunk.json:

{
  "reassignment": {
    "taskId": "US-003",
    "taskType": "react-tests",
    "currentAgent": "jest-tester",
    "attempts": [
      {
        "agent": "react-tester",
        "startedAt": "2026-02-28T10:00:00Z",
        "endedAt": "2026-02-28T10:15:00Z",
        "outcome": "verification_failed",
        "error": "3 tests failed: DarkModeToggle click handler",
        "retryCount": 0
      }
    ],
    "checkpointRef": "chunk.checkpoint"
  }
}

---

Context Overflow Handling

Context overflow needs fresh context, not retry:

1. Create checkpoint immediately
2. Do not retry same agent (won't help — context is full)
3. Show status:

   ⟳ Context limit reached, starting fresh session with checkpoint
   

4. Delegate same task to same agent with fresh context + checkpoint
5. If fresh session also hits limit: task is too large — escalate with suggestion to break down

---

Escalation Protocol

When ALL alternatives are exhausted, escalate to user:

═══════════════════════════════════════════════════════════════════════
                    TASK REQUIRES YOUR ATTENTION
═══════════════════════════════════════════════════════════════════════

Task: US-003 - Add dark mode toggle

Attempted:
  1. react-dev (primary) — Failed: verification failed (3 tests)
  2. developer (alternative) — Failed: context overflow
   
Checkpoint saved with:
  - 3 completed steps
  - 2 pending steps
  - 1 decision recorded

Options:
  [R] Retry with different approach (describe what to try)
  [M] Fix manually (I'll provide the checkpoint context)
  [S] Skip this task for now
  [A] Abandon PRD and start fresh

> _
═══════════════════════════════════════════════════════════════════════

User responses:

• R (Retry): Prompt for different approach, restart with primary agent
• M (Manual): Output full checkpoint context for copy-paste
• S (Skip): Mark story as skipped, continue to next story
• A (Abandon): Stop PRD execution, preserve state for later

---

Reassignment Flow Summary

Task Delegation
       │
       ▼
Primary Agent Executes ──────┐
       │                     │
       ▼                     │
Verification Contract        │
       │                     │
   ┌───┴───┐                 │
   │       │                 │
SUCCESS  FAILURE             │
   │       │                 │
   │    Rate limit? ──YES──► Retry with backoff (max 3)
   │       │                     │
   │      NO                  Exhausted?
   │       │                     │
   │    Context overflow? ─► Fresh session with checkpoint
   │       │
   │    Create checkpoint
   │       │
   │    Alternative available?
   │       │
   │    ┌──┴──┐
   │   YES   NO
   │    │     │
   │ Switch  ESCALATE
   │    │
   │    └────────────────────┘
   │
   ▼
Task Complete