CLI command
jeden pursue
Turn a rough objective into a source-grounded autonomous contract, execution, independent review, verdict, and receipt.
Exact invocation#
jeden pursue "rough objective" [--json] [--cwd path] [--model name] [--allow-write] [--allow-command]
jeden pursue --request-file request.json --allow-write --allow-command [--json]
jeden pursue --status REQUEST_ID [--json]
jeden pursue --state REQUEST_ID [--json]
jeden pursue --resume-run REQUEST_ID [--json]Inputs and options#
- Supply either a non-empty rough objective or exactly one machine operation:
--request-file,--status,--state, or--resume-run. Machine operations cannot be combined with a positional objective.--stateprints every value the fleet database holds for the request (its immutablerequest,response,schema_version, execution options and preferences), so what was persisted can be checked without reading the database directly. Each machine operation printskey: valuelines (nested values as compact JSON, an absent one as-);--jsonprints the same value as one JSON document. - The workspace, model route, grants, JSON output, and maximum steps use the same options as
run. - Transcript Lake preference evidence is consulted only when command execution is granted; otherwise pursuit proceeds without that external executable.
- An immutable request has
schema_version: 1,request_id,initiative_id,objective, absolute canonicalcwd,evidence_refs, decimal-stringbudget_usd,allow_write,allow_command, and optionalrepositoriescontaining exact GitHub owner/name pairs. Submission needs CLI grants matching every requested grant. - Request values, completed stages and the inference ledger live in the database named by
JEDEN_DATABASE_URL(tablespursuit_values,pursuit_stages,pursuit_calls), keyed by request ID. Hosts share these values only when they use the same server database; a SQLite file remains local. Run artifacts and the owner lockowner.lockstay underJEDEN_PURSUIT_STATE_ROOT, or~/.jeden/pursuit. A request ID cannot be rebound to different input. Resume retains its original execution options and grants; it is not a way to enlarge authority. An absent URL or database refusal names the operation and cause rather than asking Stado or Skarbiec for a credential.
Output and state effect#
Runs canonical Pursuit stages through persistent planner/executor conversations and fresh read-only reviewers. Machine responses include request and initiative identities, state, actual known spend, errors, source revisions and contract/verdict/receipt paths. Completed stages are reused only for the same role and input digest. Independent acceptance is bound to clean source revisions observed before and after review; successful machine requests must also match remote main. Status reads retained state without starting another run. A completed request is returned rather than executed again.
Refusals and boundaries#
- A missing objective is refused as
pursue requires a taskby the CLI parser; the command itself also rejects an empty value aspursue requires a rough objective. - The run fails if Pursuit cannot establish its contract, execute it, or produce its verdict and durable receipt.
- Tool mutations remain subject to the same write and command grants as
run. - A second process cannot execute an owned request. A mutable tool call without a durable matching result is indeterminate and is never automatically replayed. Missing transcripts and incomplete model-call accounting retain their uncertainty.
- A machine request requires a concrete Brama model with current positive prices and token ceilings. Each model call reserves its catalog-priced upper bound before dispatch. Unknown usage retains the reservation; exhausted allocation prevents the next call. This meter does not authorize or account for external infrastructure, advertising or child-process spending.
- Canonical repositories must be direct children of the one workspace, use main, have one checkout and match their declared origin. Dirty, changed-during-review or unpushed source cannot become accepted delivery evidence. A rejected contract or verdict remains failed, not succeeded.
- A command line of the wrong form (an unknown command or option, a missing, unexpected or invalid argument, a missing confirmation) prints the error and the usage and exits
2; a refusal that retrying unchanged can help (a dependency down or not answering, a rate limit) exits69; any other refusal or failure exits1. A refusal a functionality declares ends withfix: …andcause: <operation>.<cause>, andjeden errors show <operation>lists every cause of that functionality.