You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Jobs. The main model hands a scoped job to any configured co-worker with one tool call, dispatch (who, job, scope, optional context and wait); no sub_agent: true is
needed. The job runs the same scratch agent, consent, scoped registry and ask_main a
sub-agent gets, and its step is filed under the active task, rendered as job by /task. scope is required: an array, a bare path or a comma list; "." is the whole workspace,
accepted only while nothing else runs. A scope that overlaps a running job or an assigned
step, or falls outside the co-worker's max_scope, is refused; one co-worker may run several
jobs, all counted against sub_agents.max_concurrent. The hand-back arrives as a later
message or, with wait: true, as the tool's own result.
The main model steers its jobs.jobs status|wait|reply|extend|stop: wait for one job
or for any, answer a pending question, clear a stall, end it. A job's question, hand-back and
stall line reach the main model as job <id> … lines through the same queue a mid-task
message uses, so it reacts on its next turn; a headless be-code run keeps taking turns on
them (up to ten) instead of ending with a job still out.
A stalled job wakes the main model, not the operator. A supervisor watches every job:
after sub_agents.stall_after seconds (180) with no tool call or output the job is marked stalled and the main model is told — job 2.3 (big) has been quiet for 3m: 4 tool calls, last read_file 3m ago. jobs extend 2.3 to give it more time, jobs stop 2.3 to end it, jobs reply 2.3 <text> to steer it, or jobs status 2.3 to look. — and decides. The clock pauses
while the job waits on a question, the shared server lane, a consent prompt, a write approval
or editor review, or a tool call in flight. A third stall stops the job (blocked: stalled)
and the operator is told in the transcript; a headless run stops at the second. sub_agents.max_runtime (1800) caps running time, waits excluded (blocked: ran out of time); 0 disables either.
/agents lists every job with its state (running, waiting (tool: shell), asking, stalled, done, blocked, stopped). A job's step closes blocked: session ended at exit
and blocked: interrupted at the next start after a crash; /task assign refuses a job id;
a scheduled event's turn refuses dispatch and the mutating jobs actions.
The e2e suite gains a dispatch scenario: a job handed to a scripted co-worker, its
hand-back, and a stall the main model clears with jobs extend.
Anthropic and Gemini work as online main models. Anthropic's current models refuse any temperature, so none is sent to api.anthropic.com; Gemini 3 returns a thought signature on every tool call (extra_content) and demands it back on later requests, which the client now carries, stores with the session and echoes; and a request error after a native tool call has already worked no longer downgrades the session to embedded tool calls. The Gemini preset carries its 1,048,576-token window.