Limits

Every limit Ciresk CI applies to a request, a tree, a log, an artifact, a job and a worker, with its default and what you get when you reach it. A limit named with a policy field (max_tree_entries) is your organization's to change with ciresk ci policy set <field>=<value> (needs ci:policy.write); ciresk ci policy get shows what applies to you. Where a field says "at most", that is the cell's ceiling: the policy cannot go above it. The other limits are the same for every organization.

Limit	Default	Scope	When it is reached
Request body	1 MiB	one request	`413` `payload_too_large`. Bytes go through a presigned upload, never in the body.
Request body read time	30 s	one request	`408` `request_timeout`.
Inline blob in tree sync (`max_inline_blob_bytes`)	64 KiB, at most 1 MiB	one blob	A larger blob goes through a presigned upload; `ciresk ci` picks the way by itself.
Tree entries (`max_tree_entries`)	200,000, at most 1,000,000	one tree	`413` `tree_too_large`; `ciresk ci` refuses before it uploads and names `.ciresk/ignore`.
Tree size (`max_tree_bytes`)	2 GiB	one tree	`413` `tree_too_large`.
Blob (`max_blob_bytes`)	2 GiB	one file of a tree	`413` `blob_too_large`.
Tree sync requests (`trees_requests_per_minute`)	600 a minute, at most 6,000	one credential, counted on each API server	`429` `rate_limited` with `retry_after`; `ciresk ci` waits and retries.
Digests per tree sync request	10,000	one request	`413` `tree_missing_too_many`; split the request.
Objects from dev checkpoints per request	20,000 objects, 1 GiB, 64 directory levels	one job naming a `tree_root`	`413` `checkpoint_too_large`; sending the same root again continues.
Log chunk	64 KiB	one upload from a worker	`413` `log_chunk_too_large`; the worker splits its chunks below it.
Log per task (`max_log_bytes`)	64 MiB	one task attempt	`413` `log_chunk_too_large`; the log stops there and the task keeps running.
Log read per call	64 KiB, at most 256 KiB	one `ciresk ci logs` or `ci__job_logs` call	The answer is cut and says so; ask for the next window.
Artifact (`max_artifact_bytes`)	1 GiB	one artifact	`413` `artifact_too_large`.
Task timeout (`max_task_timeout_secs`)	3,600 s	one task	A task's `timeout` above it is cut to it; a task that runs past its timeout ends `timed_out`.
Concurrent jobs per organization (`max_concurrent_per_tenant`)	twice the fleet's slots plus 4, at least 32	queued, leased and running jobs	`429` `concurrency_limited` with `retry_after` 5: the job is refused, not queued. `ciresk ci` waits and submits again up to three times, then exits 5.
Concurrent jobs per actor (`max_concurrent_per_actor`)	the organization's	one credential and actor	`429` `concurrency_limited`, `scope: actor`.
Concurrent jobs per session (`max_concurrent_per_session`)	half the organization's, at least 4	one CLI or MCP session	`429` `concurrency_limited`, `scope: session`. Not a policy field: it follows the organization's.
Webhooks per organization	20	the organization	`422` `webhook_limit`.
Unauthenticated requests per address	60 a minute	one IP address, counted on each API server	`429` `rate_limited`.
Open long-polls per credential	32	one credential, counted on each API server	`429` `rate_limited`; close a wait before opening another.
Lease TTL	300 s	one task on a worker	The lease expires, the task is requeued, and the worker's next report gets `worker_lease_expired`.
Lease renewal	every 30 s	one worker	A worker renews each lease with its heartbeat, so a lease survives about ten missed renewals.

The concurrency caps follow your fleet unless your policy sets them: "the fleet's slots" are the slots your workers' pools can grow to plus the online slots of workers no pool reports, so work past a full fleet waits in the queue instead of being refused. ciresk ci budget shows concurrent: {used, limit}, and an agent sees a warning when every slot of it is in use.

A limit "counted on each API server" is enforced by every server of the cell on its own, so the cell as a whole may let a little more through. Budgets, the cost limits on a session, a day and a month, are on How it works.