hydra-orchestrate

Hydra Orchestrate

Principle

Keep Sol as implementation owner, verifier, and merge authority. Treat Hydra output as evidence to check, never as authority to accept.

Workflow

1. Call hydra_preflight for the target repository. Stop on any preflight error. State that account usage credits remain a manual check.
2. Call hydra_budget with {"action":"status"}. Treat it as local advisory evidence, never as an account-quota source. Missing or stale evidence must not block required quality-floor work, but it blocks optional delegation.
3. Build the context capsule below. Classify kind, scope, ambiguity, novelty, risk, write need, test need, parallelizability, context overlap, partitionability, and independent-judgment need. Call hydra_route with the capsule's exact byte count.
4. If the route is FAST, continue locally. Start no Claude job.
5. Start only plan sessions whose activation is immediate. Use hydra_start with the named model, effort, role, capsule, role-specific delta, and least-privileged access mode. Use budgetIntent.priority: "quality-floor" for the required lead and route-required immediate panel members. Use "opportunistic" for conditional or manually added sessions, with valuableWork: true and a truthful distinctInformationGain value.
6. Enforce a maximum three Claude sessions, maximum one Fable session, and the lower configured concurrency limit. Do not invent a fallback model.
7. Record each returned job ID. Check hydra_status until a terminal state, then call hydra_result. Use condition-based checks; do not hold a tool call open while Claude runs.
8. Use hydra_continue only when the same saved context is materially useful. Preserve the model, role, and access context.
9. Verify every accepted claim against local repository evidence. Open cited files, reproduce relevant commands, and run focused tests when safe. Sol alone decides whether to apply a worktree patch.
10. Call hydra_record_outcome only after the job is terminal. Supply taskType, scope, accepted and rejected claims, false positives, testsBefore, testsAfter, patch verdict, and user verdict when known.