When issue discovery first succeeded after an outage, one pass created 225 forge tasks in 5 seconds (triage 103, feasibility 49, tdd 63, implement 10) and all were claimed at once. Each prepared a full worktree clone. The data volume filled: 180 worktree prepare failed ... no space left on device, 6 more tasks parked mid-checkout after spending 6.1M tokens, and 1417 JetStream ARCHIE_SPANS write errors. Side effects of the full disk: the gateway memory curator failed writing its state, and the ARCHIE_TASKS file store logged "could not locate msg block" on later recovery. The disk-full window lasted about 80 minutes.
Nothing bounds how many tasks one intake pass creates, nor how many worktrees prepare concurrently, and worktree prepare does not check free space before cloning.
Done when
Intake caps tasks created per pass (a backlog drains over several passes) and the cap is a control-plane setting.
Worktree prepare checks free space on the work volume first and waits or parks with a clear "disk low" reason instead of failing mid-clone.
Verified by a test that a pass over N open issues creates at most the cap.
## What is wrong
When issue discovery first succeeded after an outage, one pass created 225 forge tasks in 5 seconds (triage 103, feasibility 49, tdd 63, implement 10) and all were claimed at once. Each prepared a full worktree clone. The data volume filled: 180 `worktree prepare failed ... no space left on device`, 6 more tasks parked mid-checkout after spending 6.1M tokens, and 1417 JetStream `ARCHIE_SPANS` write errors. Side effects of the full disk: the gateway memory curator failed writing its state, and the ARCHIE_TASKS file store logged "could not locate msg block" on later recovery. The disk-full window lasted about 80 minutes.
Nothing bounds how many tasks one intake pass creates, nor how many worktrees prepare concurrently, and worktree prepare does not check free space before cloning.
## Done when
- Intake caps tasks created per pass (a backlog drains over several passes) and the cap is a control-plane setting.
- Worktree prepare checks free space on the work volume first and waits or parks with a clear "disk low" reason instead of failing mid-clone.
- Verified by a test that a pass over N open issues creates at most the cap.
Fixed in 52b5740c. Discovery: one poll now creates at most intake_per_pass tasks (default 20, scheduling policy setting with a dashboard field); issues past the cap are found again on the next poll, and issues that already have a task do not use the cap. Per-task disk work: after #1797 clone/worktree prepare only runs inside process, after the task holds a dispatcher slot, so nothing needed moving. Disk guard: before cloning, a task waits (holding its slot) until the work volume has 5 GiB free, checking every 15s for up to 10 min, then parks as transient with a 'disk low' reason (transient parks now retry with backoff per #1804). Tests: table-driven cap test (failed to build before the change) and a wait/recover/stays-low test for the guard. task check passes.
Fixed in 52b5740c. Discovery: one poll now creates at most intake_per_pass tasks (default 20, scheduling policy setting with a dashboard field); issues past the cap are found again on the next poll, and issues that already have a task do not use the cap. Per-task disk work: after #1797 clone/worktree prepare only runs inside process, after the task holds a dispatcher slot, so nothing needed moving. Disk guard: before cloning, a task waits (holding its slot) until the work volume has 5 GiB free, checking every 15s for up to 10 min, then parks as transient with a 'disk low' reason (transient parks now retry with backoff per #1804). Tests: table-driven cap test (failed to build before the change) and a wait/recover/stays-low test for the guard. task check passes.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
What is wrong
When issue discovery first succeeded after an outage, one pass created 225 forge tasks in 5 seconds (triage 103, feasibility 49, tdd 63, implement 10) and all were claimed at once. Each prepared a full worktree clone. The data volume filled: 180
worktree prepare failed ... no space left on device, 6 more tasks parked mid-checkout after spending 6.1M tokens, and 1417 JetStreamARCHIE_SPANSwrite errors. Side effects of the full disk: the gateway memory curator failed writing its state, and the ARCHIE_TASKS file store logged "could not locate msg block" on later recovery. The disk-full window lasted about 80 minutes.Nothing bounds how many tasks one intake pass creates, nor how many worktrees prepare concurrently, and worktree prepare does not check free space before cloning.
Done when
Fixed in
52b5740c. Discovery: one poll now creates at most intake_per_pass tasks (default 20, scheduling policy setting with a dashboard field); issues past the cap are found again on the next poll, and issues that already have a task do not use the cap. Per-task disk work: after #1797 clone/worktree prepare only runs inside process, after the task holds a dispatcher slot, so nothing needed moving. Disk guard: before cloning, a task waits (holding its slot) until the work volume has 5 GiB free, checking every 15s for up to 10 min, then parks as transient with a 'disk low' reason (transient parks now retry with backoff per #1804). Tests: table-driven cap test (failed to build before the change) and a wait/recover/stays-low test for the guard. task check passes.