/titan

Anti-Stall Engine

Purpose: Read this file when Titan detects stalled execution, needs the full L1-L5 recovery ladder, or must account for stall budgets and guardrail interactions.

Contents

• Stall detection
• Recovery cascade
• Budget tracking
• Recovery flowchart
• Guardrail integration

13-level recovery cascade ensuring Titan never stops before exhausting all options.

Stall Detection

A stall is detected when a phase cycle produces zero progress artifacts (no file changes, no test additions, no decisions recorded, no documents generated).

Detection threshold: 2 consecutive zero-progress cycles trigger the Anti-Stall Engine.

Stall Classification

Recovery Cascade

L1: Tactical Recovery (stay within current approach)

Budget: 5 attempts per phase, 10 per project.

1.1 Retry with Context (on first failure)

• Inject error message and failure context into agent prompt
• Add relevant code snippets, specify exact expected output format

1.2 Agent Swap (same agent fails twice)

• Switch to alternative agent: Builder↔Artisan · Sentinel↔Canon · Forge↔Builder · Scout↔Lens · Radar↔Voyager

1.3 Decompose Further (agent succeeds partially)

• Invoke Sherpa to break failing task into smaller pieces
• Target: each sub-task completable in a single agent call

L2: Operational Recovery (change methodology, same goal)

Budget: 3 attempts per phase, 5 per project.

2.1 Alternative Approach (L1 exhausted)

• Switch: TDD↔Prototype-first · Top-down↔Bottom-up · Monolithic↔Modular · REST↔GraphQL

2.2 Skip and Return (other non-blocked work exists)

• Mark blocked item as DEFERRED, continue with other items
• Return after other phase work completes (context may resolve block)

2.3 Scope Reduction (feature partially implementable)

• Reduce to MVP version, cut non-essential requirements
• Document cuts and add to backlog for future iteration

L3: Strategic Recovery (change direction)

Budget: 1 attempt per phase, 3 per project.

3.1 Phase Reorder (other phases can proceed)

• Postpone stuck phase, advance to non-blocked phase, return with more context

3.2 Scope Cut (feature blocking critical path)

• Remove entire features from roadmap (preserve core value proposition)
• Update SUCCESS_CRITERIA to reflect reduced scope

3.3 Architecture Pivot (architecture is root cause)

• Fundamental change: Monolith↔Microservices · SSR↔SPA · SQL↔NoSQL · Sync↔Event-driven
• Requires re-running ARCHITECT phase for affected components

3.4 Technology Swap (specific technology is bottleneck)

• Replace problematic library/framework via Magi evaluation + Forge prototype

L4: Degradation Recovery (deliver what's possible)

Budget: No limit (graceful degradation always acceptable).

4.1 Partial Delivery (most features work)

• Ship working parts, create detailed TODO docs for unfinished parts

4.2 Stub Implementation (architecture clear, implementation blocked)

• Create interfaces/stubs with TODO annotations, full type definitions, tests against stubs

4.3 Documentation-Only (implementation not feasible)

• Comprehensive design document, implementation guide, all decisions documented

L5: User Escalation (last resort)

Budget: 1 per project per L1-L4 cycle.

5.1 Single Focused Question

• Must be: Specific · Actionable · Minimal (one question) · Options-based (2-4 choices)

Budget Tracking

## Stall Budget Status
| Level | Phase Budget | Phase Used | Project Budget | Project Used |
|-------|-------------|------------|----------------|--------------|
| L1 Tactical | 5 | [N] | 10 | [N] |
| L2 Operational | 3 | [N] | 5 | [N] |
| L3 Strategic | 1 | [N] | 3 | [N] |
| L4 Degradation | ∞ | [N] | ∞ | [N] |
| L5 User | — | — | 1 | [N] |

Budget resets per phase for L1-L3. Project budget is cumulative across all phases.

Recovery Flowchart

Stall Detected → L1.1 Retry → L1.2 Swap → L1.3 Decompose
  → L2.1 Alt approach → L2.2 Skip+return → L2.3 Scope reduce
  → L3.1 Phase reorder → L3.2 Scope cut → L3.3 Arch pivot → L3.4 Tech swap
  → L4.1 Partial delivery → L4.2 Stub impl → L4.3 Docs-only
  → L5.1 Ask user ONE question → Response → Restart cascade
Each step: Success? → Continue. Fail? → Next level.

Guardrail Integration

Guardrail events (_common/AUTORUN.md § Guardrail Protocol L1-L4) map to Anti-Stall levels: L2 recovery failure → Anti-Stall L1, L3 pause → Anti-Stall L1-L2, L4 abort → immediate halt + rollback. Guardrail auto-recoveries do NOT consume Anti-Stall budget; only failures escalated to Titan do.