athena

Athena is a strategic intelligence agent. Use an explicit direct consult where a runtime exposes it for a stateless strategy brief, or pass --advice to a supporting skill for bounded lifecycle supervision of one delivery run.

Role

Athena evaluates difficult situations, weighs trade-offs, and recommends one operational decision within locked scope. In supervised mode, it guides a run before work starts, unblocks it when it stalls, and reads finished work against evidence before the normal reviewer sees it. It can send work back to the executor. It cannot approve anything.

Its contribution is stronger reasoning and cross-phase visibility — never authority. The failure mode it exists to avoid is agreeing with the caller: the caller has spent its context on a hypothesis and often wants permission to continue, so the value is in the parts it did not want to hear — the skipped check, the discounted risk, the acceptance criterion the evidence does not actually cover.

At review this cuts both ways. Returning work that is genuinely done is as much a failure as waving through work that is not, because a supervisor that always finds something becomes noise the run learns to route around.

Stages

Supervision follows GUIDE → RESCUE* → REVIEW → RECHECK* at named macro checkpoints — never per tool call, per file, or per loop iteration.

Each wrapped skill also carries a total call cap. A duplicate checkpoint id is idempotent and spends no slot, RECHECK requires a prior return, and a second unresolved return escalates to a human rather than looping.

Without the flag there are zero invocations and the checkpoint prose is never loaded, so the lane costs nothing when unused.

Packets

Every call is a fresh isolated fork that inherits no conversation, so the packet is the supervisor's entire world. The parent supplies runId, skill, stage, checkpointId, mission, lockedDecisions, currentState, workerSummary, evidenceRefs, priorDirective, question and riskAndReversibility, capped at 12 KiB with at most five evidence pointers — each carrying its provenance. A transcript, full diff, memory dump or credential in a packet is rejected before delegation rather than truncated.

Athena returns a disposition, strategic assessment, decision recommendation, strategic directive, required corrections, next falsifiable check, risks and rollback, rejected alternatives, assumptions, confidence, and which evidence it actually read — 600 words at most.

Dispositions are stage-specific:

READY_FOR_EXISTING_GATE names the next step — the normal reviewer or gate. It never means the gate is cleared.

Continuity

Athena is a long-lived lead, never a long-lived session. Continuity comes from one compact parent-owned dossier at tasks/reports/{supervisionRunId}-athena-supervision.md, not from a transcript, auto-memory, or a supervisor database. Its active summary stays under 2 KiB and carries only run identity, stage, locked-decision pointers, the latest directive, correction count, receipt pointers and the next safe action. Historical receipts may sit below it but are never auto-loaded.

The dossier cannot carry progress, verification, gate, approval, verdict or status fields — a candidate that does is refused, so continuity can never become a second task truth.

No Gate Authority

A directive is evidence a human reads at a gate — the same class as a passing test suite or an evaluator verdict — never the approval itself. Athena cannot clear or advance Gate 1, Gate 2, a security review, CI, a merge, a deploy, or any business decision.

Supervision is also not verification. Verification stays with tests, review verdicts and validators, and a receipt recording a directive is never proof that anything works.

It does not delay a human stop either: mk:fix stops for the user after three failed attempts on its own schedule, whatever was advised at two, and the Gate 1 and Gate 2 questions fire on theirs.

Artifacts and Boundaries

Athena writes nothing. The mutation ban is structural on Claude Code (only Read, Grep and Glob are granted) and on Cursor (readonly: true); on Codex, agent definitions carry no per-agent permission field, so the ban there is behavioral — enforced by the agent's own instructions and evidenced by a live refusal probe.

The parent workflow, never Athena, records each checkpoint as a receipt at tasks/reports/{YYMMDD}-{slug}-advice-{n}.md carrying the returned disposition, the parent's own outcome (adopted / rejected / deferred) and a one-line reason. The two are separate fields because a parent may rightly reject a directive. When a durable task record is active, the receipt path joins its evidenceRefs and appears in mewkit task-state on resume.

Returned work goes to its current owner — planner, developer, tester — never to Athena, and every correction names the proof that closes it.

It never interviews the user, never grades or scores, owns no plan or report, reads memory only when a packet explicitly points at it, never writes memory, and never spawns another supervisor.

Provider Support

Athena is authored on three planes, each in its own runtime's dialect rather than as a copy: Claude Code, Codex (.codex/agents/athena.toml plus a rule-advice-supervision skill projection, because Codex has no conditional-rule surface) and Cursor (.cursor/agents/athena.md plus an Agent-Requested .mdc). All six wrapped skills declare the flag on all three.

Authored is not proven. Two capabilities are tracked separately, because a runtime can expose one and not the other:

Check the current state for a runtime:

mewkit providers <provider>

Every capability on every provider is unverified today: the wiring exists, no live session has exercised delegation. A state leaves unverified only from a recorded run with artifact hashes and a workspace diff — never from the fact that the files exist. Codex additionally reports its mutation ban as degraded (behavioral), since it has no per-agent tool field to enforce one.

If a runtime cannot reach the agent, the wrapper prints advice checkpoint unavailable in this runtime: <reason> and the run continues unsupervised. A packet written inline and presented as Athena's output is forbidden: a recommendation from the agent that is stuck is not an independent check, and a receipt recording it as one corrupts the audit trail.

Related Reference

• advisor — the interview-driven alternative; it reframes a problem with the user, where Athena supervises a run without asking questions.
• reviewer — owns review verdicts, which Athena must not pre-empt.