Durable backlog

An agent loop that dies mid-task should not lose the task. This recipe adds NullTickets — a headless task tracker backed by SQLite — so work survives crashes, retries are policy, and multiple agents can pull from one rail without stepping on each other.

Ingredients#

• NullTickets (nullhub install nulltickets, or build from source)
• One or more NullClaw instances as workers
• No orchestrator — this is the deliberately simple baseline. Add NullBoiler later if you need routing policy.

1. Start the tracker#

Via NullHub it is already supervised. From source:

git clone https://github.com/nullclaw/nulltickets.git
cd nulltickets
zig build
zig build run -- --port 7700 --db tracker.db

Defaults are --port 7700 and --db nulltickets.db. Check it:

curl http://127.0.0.1:7700/health
curl http://127.0.0.1:7700/openapi.json

That second call matters: NullTickets is self-describing. It serves its own OpenAPI 3.1 schema at /openapi.json (and /.well-known/openapi.json) precisely so agents can bootstrap themselves against it. Task-creation and pipeline endpoints are all documented there.

2. Define a pipeline#

Pipelines are finite-state machines you declare over the REST API: states, transitions, terminal states. A minimal one is queued → in_progress → done with a failed branch. Attach a retry policy — max_attempts, retry_delay_ms, and optionally a dead-letter stage for runs that keep failing.

Tasks can also declare dependencies (DAG edges): a blocked task is excluded from claiming until its upstream tasks reach a terminal stage.

3. Run the claim loop#

The tracker contract is three moves, and any runtime that speaks it can be a worker:

1. Claim — POST /leases/claim with the agent's role. The lease has a TTL and returns a bearer lease token; heartbeat it while working.
2. Report — append progress events to the run's timeline; attach artifacts.
3. Transition or fail — move the task to its next stage, with optimistic checks (expected_stage, expected_task_version) so two agents can't finish the same task twice.

Double-claiming is prevented at the SQL layer (BEGIN IMMEDIATE transactions), lease tokens are stored as SHA-256 hashes, and a worker that vanishes simply lets its lease expire — the task returns to the pool. That is the durability: no supervisor process, just state that outlives everyone.

The practical NullClaw wiring — prompts and patterns for running the loop with nullclaw as executor — lives in the tracker repo's nullclaw.md integration guide.

4. Watch the queue#

curl http://127.0.0.1:7700/ops/queue

Per-role queue stats, made for dashboards and orchestrators. NullHub also proxies the tracker's shared key-value store into its dashboard — a namespaced KV with SQLite FTS5 full-text search that doubles as cross-session agent memory.

When one queue and role-based claiming stop being enough — fan-out, multi-stage routing, worker capacity — graduate to the multi-agent workflow.