curiositech/windags-mutator

WinDAGs Mutator

Receive failure information and quality vectors. Diagnose the problem on four dimensions. Follow the escalation ladder. Apply the smallest mutation that fixes the problem. Log everything. Escalate to human when automated approaches are exhausted.

DECISION POINTS

Primary Mutation Selection Matrix

Given failure classification (Cognitive, Decomposition, Temporal), select optimal mutation type:

IF System=omission AND Temporal=transient
  → loop_back (simple retry)
  → Cost: 1x, Latency: +retry_timeout

IF System=omission AND Temporal=persistent
  → replace_node (different model/skill)
  → Cost: 1x, Latency: +model_switch_time

IF System=crash AND Decomposition=node
  → replace_node OR add_node (gap fill)
  → Cost: 1x-2x, Latency: +mutation_time

IF System=crash AND Decomposition=topology
  → restructure (add_node + remove_node sequence)
  → Cost: 2-4x, Latency: +decomposition_time

IF System=byzantine AND Cognitive=confident_incorrect
  → Cross-family ensemble review → replace_node if confirmed
  → Cost: 2x (review cost), Latency: +review_time + model_switch

Escalation Level Decision Tree

L1 (Fix Node): IF first_failure OR system=crash_recovery
  → Try: different prompt, model tier, or alternative skill
  → Threshold: <3 retries on same node

L2 (Diagnose Structure): IF pattern_matches_decomposition_signature
  → Signatures: granularity_mismatch, semantic_gap, method_explosion
  → Classify on all 4 dimensions

L3 (Generate Alternative): IF retry_count >= 3 OR decomposition_cycles >= 2
  → Apply Polya strategies: restate, simplify, specialize, analogize
  → Generate new DAG variant

L4 (Fix Topology): IF coordination_failure OR L3_mutations_failed
  → Conway-informed restructuring, intermediary agents
  → Trade-off: higher complexity for better isolation

L5 (Human Escalate): IF all_automated_exhausted OR budget_exceeded
  → Package: failure_trace + decomposition_history + recommendations

Circuit Breaker Override Logic

IF node_breaker=OPEN
  → route_around: skip this node, use alternative path
  → Cost: potentially 2-3x if no direct alternative

IF skill_breaker=OPEN AND alternative_skill_available
  → skill_substitution: use backup skill for same task
  → Cost: 1.2x (skill switching overhead)

IF model_breaker=OPEN
  → model_failover: use different provider (Claude→GPT→Gemini)
  → Cost: 1x, Latency: +provider_switch_time

IF multiple_breakers=OPEN
  → escalate_immediately: skip lower escalation levels

FAILURE MODES

1. Rubber Stamp Escalation

Symptom: Every failure immediately escalates to L5 without attempting lower levels. Detection: IF escalation_level=5 AND previous_levels_attempted=0 Fix: Enforce BC-FAIL-002 ordering. Check escalation ladder prerequisites before each jump.

2. Mutation Thrashing

Symptom: Applying contradictory mutations in rapid succession (add_node then remove_node on same position). Detection: IF mutation_cycle_count > 3 per wave AND mutations_cancel_each_other Fix: Implement mutation debouncing. Wait for wave completion before applying inverse mutations.

3. Byzantine Blindness

Symptom: Confident-incorrect failures treated as simple crashes instead of Byzantine. Detection: IF confidence>0.8 AND downstream_failures_cascade AND system_layer≠byzantine Fix: Add confidence-accuracy calibration check. Route high-confidence wrong answers through cross-family ensemble.

4. Circuit Breaker Cascade

Symptom: Opening one breaker triggers opening others unnecessarily. Detection: IF multiple_breakers_open_simultaneously AND failures_unrelated Fix: Verify BC-FAIL-004 independence. Each breaker must track separate failure domains.

5. Saga Compensation Amnesia

Symptom: Failed mutations leave DAG in inconsistent state without compensation. Detection: IF mutation_failed AND compensating_action=null AND saga_type=COMPENSATABLE Fix: Execute reverse compensation immediately. Log limitations honestly if full restoration impossible.

WORKED EXAMPLES

Example 1: Transient Crash → L1 Retry Success

Initial State: Text analysis node crashes with "rate limit exceeded" Classification: System=crash, Temporal=transient, Decomposition=node Decision: Apply loop_back mutation (L1)

1. Classify: crash + transient → simple retry eligible
2. Check circuit breakers: node=CLOSED, skill=CLOSED, model=HALF_OPEN
3. Apply mutation: loop_back with 30s delay
4. Log MutationEvent: saga_type=RETRIABLE
5. Re-execute: SUCCESS
6. Quality gate: output passes schema validation

Novice miss: Would escalate immediately instead of trying simple retry. Expert catch: Recognizes rate limit as transient system issue, not skill problem.

Example 2: Persistent Byzantine → L3 Redecomposition

Initial State: Code review node reports "95% confident, looks good" but code has security vulnerability Classification: System=byzantine, Cognitive=confident_incorrect, Temporal=persistent Decision: Cross-family ensemble → L3 redecomposition

1. Detect: High confidence + downstream security test failures
2. L1 retry: FAIL (same confident-incorrect response)
3. L2 structure check: FAIL (edge dependencies correct, node skills correct)
4. Trigger cross-family ensemble: Claude→GPT review
5. Ensemble confirms: original output wrong, confidence miscalibrated
6. L3: Decompose security review as separate specialized node
7. Apply: add_node(security_specialist) + add_edge(code→security)
8. Re-execute: PASS (vulnerability caught by specialist)

Novice miss: Would trust high confidence, miss the Byzantine failure pattern. Expert catch: Calibrates confidence against historical accuracy, uses ensemble validation.

Example 3: Topology Failure → L4 Restructure with Trade-offs

Initial State: Data pipeline has circular dependency causing deadlock Classification: System=crash, Decomposition=topology, Temporal=persistent Decision: Conway-informed restructuring (L4)

1. L1-L3 attempts: FAIL (retries, node swaps, alternative decompositions all hit circular dependency)
2. L4 triggered: topology-level problem identified
3. Analyze: Conway's Law suggests communication structure mismatch
4. Trade-off decision: Add intermediary buffer nodes vs. merge conflicting nodes
   - Option A: +2 nodes, +communication overhead, -circular dependency
   - Option B: -1 node, +complexity, potential skill capability loss
5. Choose Option A: Better failure isolation despite higher cost
6. Apply: add_node(buffer1) + add_node(buffer2) + restructure edges
7. Circuit breaker update: original failing nodes → HALF_OPEN
8. Re-execute: SUCCESS with 2.1x cost increase

Novice miss: Would try to fix nodes individually, miss structural topology issue. Expert catch: Recognizes deadlock as Conway's Law problem, accepts cost trade-off for reliability.

QUALITY GATES

• [ ] Failure classified on all 4 dimensions (System, Cognitive, Decomposition, Temporal)
• [ ] Escalation level determined by ladder prerequisites, not random jumping
• [ ] Mutation type selected from decision matrix based on classification
• [ ] Circuit breaker states checked before applying any mutation
• [ ] MutationEvent logged with before/after structural diffs (BC-EXEC-002)
• [ ] Saga compensation strategy identified (COMPENSATABLE/PIVOT/RETRIABLE)
• [ ] Cross-family ensemble used for Byzantine + confident_incorrect cases
• [ ] Budget impact calculated (cost multiplier and latency addition)
• [ ] Re-execution triggered after successful mutation
• [ ] Quality vector validates fix effectiveness (scores improved)

NOT-FOR BOUNDARIES

Do NOT use this skill for:

• Initial DAG decomposition → Use windags-decomposer instead
• Quality vector computation → Use windags-evaluator instead
• Wave execution planning → Use windags-executor instead
• Learning from mutation patterns → Use windags-learning-engine instead

Delegate when:

• L5 escalation reached → Present to human with full context
• Mutation requires new skills → Route to skill-acquisition system
• Circuit breakers indicate systemic issues → Escalate to infrastructure-monitoring
• Budget constraints violated → Route to resource-management for approval