Autodev Lev

Autodev Lev

Heartbeat-driven autonomous development loop using Lev primitives. Unlike /autodev-loop (which uses Claude Code CronCreate), this runs through lev loop autodev with sleep-based pacing — staying in-process with full state continuity between ticks.

Invocation

/autodev-lev                           # Start with defaults (5m heartbeat)
/autodev-lev 10m                       # Custom interval
/autodev-lev --budget=50k              # Token budget cap
/autodev-lev --max-ticks=5             # Stop after 5 ticks
/autodev-lev --until="all specs pass"  # Semantic exit condition
/autodev-lev status                    # Show queue and config
/autodev-lev --dry-run                 # Scan only, show what would execute

Architecture

┌─────────────────────────────────────────────────────────┐
│  Top-Level Orchestrator (you)                           │
│  Manages SDLC stages: SCAN → PLAN → EXEC → VALIDATE    │
└──────────────┬──────────────────────────────────────────┘
               │ /autodev-lev 5m
               ▼
┌─────────────────────────────────────────────────────────┐
│  Heartbeat Loop (lev loop autodev)                      │
│                                                         │
│  ┌──────┐  sleep(5m)  ┌──────┐  sleep(5m)  ┌──────┐   │
│  │SCAN  │────────────→│SCAN  │────────────→│SCAN  │   │
│  └──┬───┘             └──┬───┘             └──┬───┘   │
│     │ entities?           │ entities?          │ none   │
│     ▼                     ▼                    ▼        │
│  ┌──────┐             ┌──────┐           all_done      │
│  │ EXEC │             │ EXEC │                          │
│  └──┬───┘             └──┬───┘                          │
│     │                     │                              │
│     └─ runSdlcLoop()     └─ runSdlcLoop()               │
│        (token spend)        (token spend)                │
└─────────────────────────────────────────────────────────┘

Key Difference from autodev-loop

Protocol

Phase 0: Pre-flight

1. Check lev loop for entity queue
2. Determine interval, budget, exit conditions
3. Execute lev loop autodev with flags

Phase 1: Scan (zero cost)

• Filesystem glob across configured surfaces (.lev/pm/plans/, docs/specs/)
• Parse frontmatter for priority, lifecycle state, fitness functions
• Validation gates alignment: load .lev/validation-gates.yaml, report gate status
• Skip blocked/deferred entities
• Sort by priority (P0 > P1 > P2 > P3 > P4)

Phase 2: Execute (token spend)

• Run runSdlcLoop() on discovered entities
• Uses prompt-stack for composition (default: sdlc-exec-validate)
• Uses exec profiles for execution (default: sdlc.flowmind.exec)
• Respects maxWorkers for concurrent execution
• Emits LevEvents for observability

Phase 3: Exit Check

After each tick, check exit conditions in order:

1. fail_fast — any failure this tick + --fail-fast flag
2. circuit_breaker — 3 consecutive ticks with zero successes
3. max_ticks — hit --max-ticks limit
4. budget_exhausted — cumulative tokens >= --budget
5. until_met — semantic condition satisfied
6. all_done — no entities remaining in queue

Phase 4: Sleep

Sleep for --interval duration. Process stays alive, state preserved. SIGINT/SIGTERM triggers graceful shutdown after current tick completes.

Configuration

Config lives in plugins/core-sdlc/config.yaml under autodev.heartbeat:

autodev:
  heartbeat:
    interval: 5m
    max_workers: 3
    max_entities: 5
    max_ticks: 0          # 0 = unlimited
    budget_tokens: 0       # 0 = unlimited
    stack: sdlc-exec-validate
    profile: sdlc.flowmind.exec
    circuit_breaker_threshold: 3

Overridable via CLI flags (flags win over config).

Skill Routing

When invoked as /autodev-lev:

# 1. Parse args
INTERVAL="${1:-5m}"
FLAGS="${@:2}"

# 2. Delegate to lev loop autodev
lev loop autodev --interval=$INTERVAL $FLAGS

When the orchestrator (top-level agent) uses this skill:

1. Start: lev loop autodev --interval=5m --max-ticks=10 --budget=100k
2. Monitor: Watch stdout for tick summaries
3. Interrupt: SIGINT to stop gracefully
4. Status: lev loop autodev status for queue snapshot

Integration with Prompt Stacks

Uses existing prompt-stack plugin for entity composition:

• sdlc-exec-validate — default: execute entity + validate gates
• sdlc-deepen-plan — deep plan decomposition
• sdlc-hygiene — hygiene scan
• Custom stacks via --stack=<id>

FlowMind will absorb this when it settles. Until then, prompt stacks are the steering layer.

Exit Reasons

Example Session

$ lev loop autodev --interval=5m --max-ticks=10 --budget=50k

⚡ Tick 1 | 4 entities | interval 5.0m
  → Processed: 2 | Succeeded: 2 | Failed: 0 | 45200ms
  💤 Sleeping 5.0m...

⚡ Tick 2 | 2 entities | interval 5.0m | budget 12000/50000
  → Processed: 2 | Succeeded: 1 | Failed: 1 | 38100ms
  💤 Sleeping 5.0m...

⚡ Tick 3 | 1 entities | interval 5.0m | budget 28000/50000
  → Processed: 1 | Succeeded: 1 | Failed: 0 | 22300ms
  💤 Sleeping 5.0m...

⚡ Tick 4 | 0 entities | interval 5.0m | budget 35000/50000
  ✓ No entities remaining — all done.

─── Autodev Session Complete ───
  Ticks:      4
  Processed:  5
  Succeeded:  4
  Failed:     1
  Duration:   15.2m
  Exit:       all_done

Anti-Patterns

• Don't use CronCreate — that's the old way. Heartbeat > cron for state continuity.
• Don't set interval < 1m — burns tokens. If you need faster, use lev loop run directly.
• Don't skip --budget on long runs — set a budget to prevent runaway spend.
• Don't ignore circuit_breaker — 3 failures = something structural is wrong. Investigate.

Dependencies

• @lev-os/orchestration/entities — discoverPlans, loadLoopConfig
• plugins/core-sdlc/src/workflows/sdlc-loop.ts — runSdlcLoop
• plugins/prompt-stack — prompt composition (until FlowMind absorbs)
• core/exec — execution engine with semaphore + token tracking

Auto-Nudge + Stale Detection Heuristics (from OMX + Clawhip)

Enhanced idle/stale detection beyond basic pane_last_active:

Stale Session Detection (from Clawhip cw-04)

• Content hashing: Hash tmux pane output before/after intervals. No change = stale.
• Keyword windows: Aggregate keyword hits (error, complete, stuck) over N-second windows. Burst = event.
• Parent process tracking: If the parent process of a tmux session dies, the session is orphaned.
• Configurable stale_minutes: Per-session stale threshold (default 30min from Clawhip).

Auto-Nudge (from OMX omx-09)

• Proceed-intent detection: OMX detects when an agent is waiting for input vs genuinely stuck.
• Stall signature matching: Match against known stall patterns (repeated output, token exhaustion, rate limit).
• Nudge escalation: nudge → checkpoint → escalate → EXIT_SIGNAL (progressive, not immediate kill).

Source: .lev/pm/parity/clawhip.yaml (cw-04), .lev/pm/parity/omx.yaml (omx-09), tribunal items 18+24 (UNANIMOUS AGREE)