Autonomous Task Queue

Status: Open — no code exists yet; jackin❯ has reusable launch/isolation/status/telemetry primitives but no queue, task source, or dispatcher (Phase 4, Agent Orchestration Program)

Shipped baseline

jackin❯ has the primitives an autonomous queue would reuse, but not the queue itself. Manual jackin load already creates isolated role instances through the runtime launch pipeline, claims DNS-safe container names with an 8-character random instance id, writes instance manifests, prepares workspace mounts, launches the capsule, attaches the foreground session, and finalizes isolated worktrees on exit. Worktree isolation is per container: crates/jackin-isolation/src/materialize.rs materializes host worktrees under the per-container state tree, and crates/jackin-isolation/src/branch.rs derives jackin/scratch/<selector> branch names today. Agent runtime status is also partly shipped: Capsule owns the status state machine, publishes AgentStatusReport over the protocol, and combines screen rules, OSC markers, foreground process evidence, and cooperative reports. Usage telemetry has a landed Capsule slice: account quota snapshots persist in a Capsule-local Turso account_usage_snapshots table, and token-spend monitoring is described as shipped in the Token & Cost Telemetry research page, but the host-global per-instance usage_samples substrate remains open.

No source-backed task queue exists in code. The operator config has workspaces, role sources, auth, runtime, docker, git, and dirty-exit policy fields, but no top-level task_sources, no workspace task_sources, and no [workspaces.<name>.queue] schema. The CLI has load, console, status, usage, workspace/config/role management, and other operational commands, but no jackin queue command. Repository search found no queue dispatcher, TaskSource trait, task table, PR lifecycle table, queue scheduler, budget gate, or queue console panel. Treat this page as planned work only, with the shipped baseline above as reusable substrate.

Remaining work

1. Per-workspace queue config and source assignment (task_sources, [workspaces.<name>.queue]).
2. Durable task table and idempotent source polling integration, once the Task source abstraction and Persistent storage layer land.
3. Dispatch loop tied to console/runtime lifecycle, with pause/resume state; per-task instance launch reusing the existing load and isolation primitives.
4. First-prompt templating through task-source handlers, keeping user-controlled task body clearly separated from role instructions (prompt-injection surface).
5. Completion classification from runtime status, exit/finalization, and outcome markers (completed, failed, needs_review).
6. PR lifecycle metadata and manual gates for publish, ready-for-review, request-review, merge, and auto-merge.
7. CLI commands: jackin queue list|feed|retry|cancel|pause|resume|approve. None exist today.
8. Console queue panel with pending, in-flight, failed, completed, and gate-wait states.
9. Budget gating (budget_usd_per_day), blocked on the host-global per-instance usage-storage substrate — see Token & Cost Telemetry.

Full design rationale, config sketch, prior-art comparison (multicode), and open questions live in Autonomous Task Queue: Design Exploration.

Out of scope

• Multi-workspace or organization-wide scheduling.
• Automatic retry policies, exponential backoff, and jitter.
• Cross-task dependencies.
• Webhook-driven dispatch.
• Rich reviewer assignment, team routing, and policy-driven auto-merge.
• Background daemon mode independent of console/runtime supervision.

Related work

• Autonomous Task Queue: Design Exploration — problem framing, multicode comparison, recommended shape, open questions.
• Agent Orchestration Program
• Task source abstraction
• Persistent storage layer
• Agent runtime status
• Token & Cost Telemetry
• Per-mount isolation
• Idle runtime cleanup

Related files

• crates/jackin/src/cli.rs — current top-level CLI enum; would need Queue commands.
• crates/jackin-config/src/app_config.rs and crates/jackin-config/src/schema.rs — current config schema; would need task sources and queue blocks.
• crates/jackin-runtime/src/runtime/launch.rs — existing manual launch pipeline to reuse programmatically.
• crates/jackin-instance/src/naming.rs — current short DNS-safe container naming.
• crates/jackin-isolation/src/materialize.rs and crates/jackin-isolation/src/branch.rs — existing per-container worktree materialization and scratch branch naming.
• crates/jackin-agent-status/src/lib.rs and crates/jackin-protocol/src/agent_status.rs — shipped runtime-status signal to consume.
• crates/jackin-usage/src/telemetry_store.rs — shipped Capsule-local account quota store; not enough for queue budget gates.