An agent's session, from token to a person's decision Diagram

An admin runs mainworthy token new --agent reviewer --can read,produce. The secret is shown once; configure the agent with it. Add --project to restrict a new actor to one project. Tokens use the actor’s grants.
Connect over stdio with mainworthy mcp, which forwards to the running server, or use the authenticated HTTP endpoint /mcp. Supply MAINWORTHY_TOKEN to the CLI, or a Bearer token over HTTP.
The agent sees ten tools. Listing a tool does not grant permission to call it. Pass project unless exactly one project is available.

Start with impact for a change or changed paths. Keep the case hashes and cursor. Use get for case bodies and find for text, folder or tag searches.
Use get with as: "md" for Markdown and unless for hashes you already hold. Use impact with ledger: false for a fresh review.

propose names the base version of an existing case. For an ordinary, nonconflicting agent edit, loose adopts it by rule; standard and strict leave a question for a person.
A conflicting edit follows the conflict clause even under loose. The agent may withdraw its own conflicting proposal, but cannot accept it. Policy changes ask a human admin.
The decision path below illustrates accepting an open proposal under standard. A rejection leaves the case unchanged. Other questions have their own answers.

claim records the cases reviewed as C-1@hash, even with no findings. Findings raise questions; an already known finding returns its existing question and answer.
report records observed results for a target or a planned run. Agent failures raise questions. Passing results do not each require a decision.
Agent results never move a commit’s gate. They can count in a planned run only when the run asks for agents and the pinned case admits them.

A person with answer capability reads the question and current evidence in the queue, then decides. The accepted edit keeps the agent’s authorship. Strict requires a reason for an ordinary adoption.
since uses the saved cursor to return changed case summaries and settled questions. Read the changed case with get; if more is true, call since again with its new cursor.
An agent with read and produce can offer advice on failures and findings where policy allows it. Check advised, decided or forbidden before treating an answer as settled.