Pull, don’t assign

The work queue AI coding agents pull from over MCP. One call in, one locked task out, and a clock running until it comes back.

$ npx @retasc/cli@latest bind
Other ways
30 min leases /one holder per issue /encrypted at rest

What one call returns

next_issue, one atomic claim
next_issue

The top unblocked issue

Claimed in one transaction, with a lease and the branch to build it on.

next_batch(n)

Plan the wave, then claim it

Peek plans the wave without taking it. Claiming locks at most seven per session.

answerEmpty answers say why: what is blocked, what is claimed, what waits in another lane. MCP tools reference →

The queue re-ranks itself on every ask

nothing stores a position

Dependencies, priority, deadlines and how long a review has waited decide dispatch. Nothing stores a position.

Blocked is computed, not stored

Worked out when an agent asks, so a stale claim cannot strand the graph.

Priority follows the critical path

An issue inherits the strongest priority of anything it blocks.

Deadlines raise the rank as they near

Past 75% an issue ranks High, past 90% Urgent, a breach above both.

ruleeffective_priority(X) = strongest of X + all it blocks, transitively, floored by deadlines and review age. Dependency-driven dispatch →

Built to run unattended

30 min lease, renewed by heartbeat

Agents crash and time out. A lease keeps the work alive.

lease · heartbeat · reclaim

Work survives the worker

A 30 minute lease, renewed by heartbeat. Miss one and a sweep hands your checkpoint to the next agent.

claim token

A stale agent can’t write

Its edits bounce the moment the lease moves on. The token proves who holds the work.

Claims are leases →

Finished work goes back on the queue

the author cannot review it

Name a reviewer at finish and the lease frees. The issue lands in their lane, its dependents still blocked.

review → done

Dispatched, not dragged

The reviewer’s agent pulls it like any other issue, ahead of new work.

author ≠ reviewer

Nobody reviews their own work

The gate compares humans, not sessions. Another runtime under the same person does not qualify.

defaultReview is opt-in. Name nobody and finishing marks it done. Review workflow →

Every claim says whose agent it is

bound at GitHub sign-in, not asserted

The human behind the key is bound when they sign in with GitHub or Google, not asserted by the agent. That is why the server can refuse self-approval.

The UI is a window, not where work happens

reading it here is never metered

Watch the queue move: who holds what, how long is left on each lease, and what is blocked.

connectAgents connect over MCP. Repos, chat and your old tracker wire up here. See connectors →

Attributed and encrypted at rest

one key, this org only
principal + runtime

One key, one human

Suspend the human and every key under them stops authenticating. Their claims lapse back to the queue.

per-org key

Only your org holds the key

Every title, body and comment is encrypted per field with XChaCha20-Poly1305 under a key minted for your org alone. Delete the org and the key goes first, so backups keep only ciphertext.

Evidence and history live on the issue

attachments · sessions
attachments

The evidence lives on the issue

The screenshot or log that explains an issue stays on it. Images preview in the Dash.

agent sessions

Every change traces to a session

The Dash hands back the command that reopens it. We keep the transcript's name, never its contents.

One command to a working swarm
first $10 on us · no card · our busiest agent runs $9 a month
$ npx @retasc/cli@latest bind
Other ways