Intent → verb map
The compact, measured capability map an AI reaches for at the moment of need — plain intent on the left, the exact SEIF verb on the right, plus the multi-step workflow recipes.
This is the affordance layer in usable form: state a need in plain language, find the verb that satisfies it. It is the same compact map SEIF surfaces to an agent at session start — the surface that lifted a cold model's verb-reach from 0.50 to 1.00 (measured). It is built for the AI working beside you, but it reads fine for a human too.
For any need not listed here:
seif <group> --help. The long tail is real — roughly sixty command groups exist. You are not meant to memorize them; you are meant to reach the right one on demand.
Need → verb
| You want to… | Reach for |
|---|---|
| Track / watch a peer agent's progress; what's queued or done | seif relay board (status) · seif relay watch (live) · seif relay poll (once) |
| Pick up the next queued task as executor | seif relay next |
| Hand finished work back to the orchestrator | seif relay handback (dispatch one with seif relay brief) |
| Remember a durable fact / learning across sessions | seif gov memory add |
| Recall a past decision — "where did we decide X" / history | seif memory query |
| Find code or docs without reading whole files | seif context query (refs) · seif context retrieve (slices) |
| Get a one-shot workspace health report | seif doctor |
| Repair broken cycle / session coherence | seif heal |
| Defer something so the next cycle won't forget | seif gov pending create then register |
| Open / close a unit of work (cycle) + closure ritual | seif gov cycle |
| Avoid clobbering a file another agent may edit | seif claim |
| Classify sensitivity before sharing | seif classify |
| Put a human gate before a risky action | seif confirm-action |
| Get a second / third AI opinion; resolve disagreement | seif debate |
| Append an amendment to a SEIF artifact (no re-sign) | seif gov amend |
| Author a design memo + coherence node | seif gov memo |
Workflow recipes
Some needs are multi-step — what matters is the order, not just one verb.
An orchestrator dispatches work to an executor
seif relay brief <slug> # write the brief
seif relay pointer # share the 2-line chat pointer
watch seif relay board/poll # arm the return path
seif relay close <slug> # on return, close itDispatch and track through the inbox — never offload the transport to the human.
An executor picks up and returns work
seif relay next # claim the oldest queued brief
# …do the work; seif relay heartbeat if it runs long
seif relay handback <slug> # hand it back (seif relay release if you can't finish)Safely edit a file other agents may touch
seif claim check <path> # is it claimed?
seif claim acquire <path> # take the lease
# …edit…
seif claim release <path> # release itRun a unit of work end to end
seif gov cycle open c<N>-<slug>
seif gov cycle set-manifest
# …work; seif gov cycle decision-add as decisions land
seif gov cycle close-ritual <slug> # checkpoint→audit→quality→meditate→absorb→seal→seedPersist a learning (raw writes to .seif/memory are gated)
# compose a scratch file, then:
seif gov memory add <name> --content <path>Where this comes from
The map is not hand-waving — it is a measured, shipped surface. The reach numbers, the method, and the honest caveats are on the Just-in-time expertise page. For the exact, always-current command list and every flag, see the CLI reference, generated directly from the binary.