Reference
14 commands and 8 tools, generated from the programs themselves rather than written down, so this page cannot describe one that does not exist.
dayshift hands the broker a job from a shell, a script or a Makefile. Build it
with make -C dayshift-cli-cpp install; global options --url
and --json go before or after the command.
-, or a pipe)So a script can tell a failed job from a broker that is down from one that said no, because each has a different fix.
The environment wins; ~/.dayshift/config supplies defaults for a shell that set nothing.
submit when --runner is not given (default claude)
The MCP server gives an agent the broker as tools, over stdio. It offers five tools
that only read. The three that start or change work exist only when
DAYSHIFT_MCP_ALLOW_WRITES=1 is set. Pausing dispatch is never a tool, because
that stops every caller.
A conversation's turns, oldest first, each with its prompt and answer. Every job opens a conversation; its conversation_id is on the job.
conversation_id string, required max_chars integer — Longest answer to return inline (default 20000); a longer one is cut and says so One job: status, prompt, and its answer (when done) or its error (when failed). An error says whose fault it was — the job's own limits, or the runner being broken — which decides whether a retry can help.
id string, required — A job id, e.g. job-20260921T153012Z-17 max_chars integer — Longest answer to return inline (default 20000); a longer one is cut and says so Jobs, newest first, as summaries (id, status, runner, service, a prompt excerpt). Use dayshift_get_job for one job's full answer. A caller token sees only its own jobs; an operator token sees all.
status queued | running | done | error | canceled service string — Only jobs submitted under this service label runner string host string — Only jobs that ran on this host; dayshift_status lists them limit integer — How many (default 20) The broker's state: whether dispatch is open, suspended or held by a resource gate (and why), how many jobs are running and queued, and each runner's health. Check this before submitting if work seems stuck.
Wait for a job to finish, up to timeout_seconds, then return it as dayshift_get_job would. If it is still going when the time is up, you get its current state with still_running: true — call again to keep waiting. The job itself is never affected by giving up.
id string, required timeout_seconds integer — How long to wait (default 120) max_chars integer — Longest answer to return inline (default 20000); a longer one is cut and says so Cancel a job that has not started yet. A running job cannot be canceled from here.
id string, required Add a turn to a conversation. The runner resumes the same session, so it remembers the earlier turns. One turn at a time: while the previous turn is still running the broker refuses (turn_in_flight).
conversation_id string, required prompt string, required model string effort string Hand the broker a job: an agent CLI run with this prompt, queued behind the broker's resource limits. Returns at once with the job id and conversation id — it does NOT wait. Use dayshift_wait for the answer. workdir is a path on the machine that RUNS the job, not yours: by default the broker's host; with labels, whichever connected host has them (e.g. ["home-src"] for the workstation's ~/src). A job no connected host can take waits, and says why.
prompt string, required runner string — Runner id (default claude); dayshift_status lists them workdir string — Directory on the broker's host to run in model string — Only for a runner that accepts_model; otherwise the broker refuses effort string — Only for a runner that accepts_effort allowed_tools string[] — Override the runner's default tool allowlist timeout_seconds integer labels string[] — Labels the host must have; dayshift_status lists each agent's labels