Inspecting state¶
board is the source of truth the web app and IDE read. It returns the
project's live state as JSON and exposes the actions the panels' buttons call,
so the IDE stays a thin view over the engine. verifier reads and labels the
gate results that back a verdict; mcp-context exposes the same state to
agents over MCP.
board¶
With no verb, board prints the whole project document: runs, learnings,
automations, recipes, workspaces, scheduler, and an errors map for any
section that failed to compute. Each section is computed independently — one
broken source does not blank the board.
Verbs¶
| Verb | Arguments | What it does |
|---|---|---|
show |
[<run_id>] |
Whole project document, or one run |
run |
<run_id> |
One run's card as markdown, plus its open workspace |
runs |
[--query TEXT] [--limit N] |
Paged run list; whitespace-separated terms are ANDed |
merge |
<run_id> |
Merge a finished run's workspace |
discard |
<run_id> |
Discard a run's workspace |
race-keep |
<winner> [--discard ID ...] |
Merge the winner and discard the runners-up |
stop |
<run_id> |
Soft-stop a running run |
kill |
<run_id> |
SIGTERM, then SIGKILL after two seconds |
resume |
<run_id> |
Resume a cost-paused run |
retry |
<run_id> |
Spawn a recovery for a failed run (see below) |
gate |
approve\|reject <inbox_id> |
Resolve a review-gate row, with --note |
page |
<key> [--tab T] [--arg k=v] |
One IDE page as data |
node |
<run_id> <node_id> |
One node's detail (choose a --view) |
steer |
<run_id> --role R --severity S --text T |
Inject a steering message into a live run |
answer |
<run_id> --text T [--for Q ...] |
Answer a question the run asked |
checkpoints |
<run_id> |
Per-node git checkpoints for a run |
restore |
<run_id> <node> |
Restore the run's worktree to a checkpoint |
start |
--kickoff <path> |
Launch a run from a kickoff file, detached |
dismiss |
<run_id> [--undo] |
Hide a run from the inbox without stopping it |
not-needed |
<run_id> [--undo] |
Close a run as not needed |
Global options¶
| Flag | Meaning |
|---|---|
--home <dir> |
Project home (default $MINI_ORK_HOME or ./.mini-ork) |
--json |
Emit JSON (the default shape for most verbs) |
--shell |
Project to the IDE shell's keys only: version, project, home, generated_at, header, runs, counts, errors |
--view <name> |
Node detail view: stream, output, prompt, telemetry, learning, changes, overview, artifacts, harness |
--offset N |
Stream view: start at this line |
--limit N |
runs page size (default 50, max 200) |
--query TEXT |
runs full-text filter |
--force |
retry: bypass the retryable=false gate; race-keep: supply the operator's merge approval |
--dry-run |
retry: return the hint only; race-keep: print the plan, change nothing |
--ack-change |
retry: acknowledge the hint's needs_change block |
--lane <alias>=<lane> |
retry: pin a lane for the spawned recovery (repeatable) |
--undo |
dismiss / not-needed: reverse the close |
--note TEXT |
gate: reason attached to the decision |
start mirrors the corresponding run flags: --kickoff, --recipe,
--workspace worktree|in-place, --set K=V (repeatable), --run-id <id>,
--init, --dry-run.
Examples¶
mini-ork board --json | jq '.runs[0]'
mini-ork board runs --query "changelog" --limit 10
mini-ork board node <run-id> implement --view stream --offset 40
mini-ork board gate approve <inbox-id> --note "scope verified"
mini-ork board retry <run-id> --dry-run
mini-ork board start --kickoff ./kickoff.md --recipe code-fix
retry refuses to spawn when the failure hint says the code — not the
environment — must change, unless --ack-change is passed; --force bypasses
the retryable=false gate. Prefer retry over re-running from scratch: it
carries the run's evidence forward.
Exit codes
0 success, 1 failure, 2 usage. A malformed verb or a missing required
positional is a usage error, not a failed run.
verifier¶
Inspect and label gate results — the calibration ground truth for the oracle.
mini-ork verifier list [--run <run_id>] [--unannotated] [--limit N] [--json]
mini-ork verifier annotate --result-id <vr-…> --kind <false_positive|false_negative>
[--annotator NAME] [--notes TEXT]
| Flag | Meaning |
|---|---|
--run <run_id> |
Restrict to one run |
--unannotated |
Only results with no operator label |
--limit N |
Maximum rows (default 50) |
--json |
Emit JSON instead of a table |
--result-id |
The vr-… id printed by list |
--kind |
false_positive or false_negative |
--annotator |
Who labelled it (default $USER) |
--notes |
Free-text rationale |
Annotations are how the abstain gate is calibrated. Label a PROVEN verdict that
was wrong a false_positive; a REFUTED verdict on a correct change a
false_negative.
version¶
For the engine commit and the home's migrated version, use mini-ork doctor.
mcp-context¶
Runs a stdio MCP server exposing runs, learnings, cost and the lane map to
agents — for example inside Zed. By default only read-only tools are
exposed. --control (or MO_MCP_CONTROL=1) additionally exposes
list_recipes, start_run, run_status, wait_for_run, stop_run and
certify. The server runs until stdin closes.
Related¶
- Automations — scheduled runs shown on the board
- Recovery — checkpoints and restore in depth
- The web app and IDE — the panels these verbs feed
- Verification and gates — reading a verdict