managed-workflow

# Bounded managed workflows Use this for remote Claude tasks with independent pieces and explicit acceptance tests. Use `wasm-agent` for local isolated work, and `ruflo-workflows` for persisted approval gates or native workflow scripts. Do not turn a simple task into remote fanout just because tools are registered. 1. Inspect `managed_agent_capabilities`. Registration, configured credentials, provider beta access, tool authorization and remote health are separate facts. A configured credential is not proof of access. Never infer remote readiness. 2. The agent mod provides `/agent-mod capabilities` and the workflow mod provides `/wf-mod managed`. These commands list tools without making provider calls. Dynamic creation through the agent mod is off by default. Existing user authorization takes precedence; do not introduce another confirmation when the task already authorizes bounded execution. 3. Respect the host's `managedWorkflows`, `maxBudgetCents` and `allowInlineAgents` settings. Do not disable policy to bypass a refusal. Default maximum is 2500 US cents. Inline agents are off by default, requiring predefined agents whose tools were reviewed separately. In flight work can exceed a provider budget. This is local creation admission, not a global organization spending ceiling or a policy for externally created sessions. 4. Record task scope, source commit, output contract, deadline, budget, allowed network hosts and independent acceptance test before creation. Do not put keys or MCP tokens in prompts. Never give remote workers recursive Ruflo swarm launch, release or broad secret access. 5. Create with `dynamicWorkflows: true`, `budgetCents`, exact `allowedHosts`, and appropriate `predefinedAgents`. Save resource IDs. Send one scoped task asking for inventory, parallel work, independent reproduction and synthesis. Workers share files; require separate worker outputs and serialize patches. 6. Inspect open runs, pauses, failures and every worker thread with `managed_agent_events`. A run ending completed never proves correctness. Observation timeout does not stop remote work. Interrupt requests a pause; terminate confirms session deletion. Clean up only owned resources. 7. Independently test artifacts against the original source and acceptance contract. Persist only verified evidence through the repository's memory/RVF integration, binding source identity, reproduction commands and real usage. Never store credentials or promote an unverified transcript as learned truth. The console's Mods page shows the agent mod's creation policy, cap, inline setting and counters. It does not establish cloud run state. This skill grants no additional tool, network, credential, memory or release authority. ## Verified learning and console use Use the console Workflows, Cost, Approvals and Learning pages to inspect managed runs. Snapshots are staleable local observations and never authorize actions. Use `managed_agent_control` for fresh policy checked inspect/interrupt/terminate; the host must separately allow the exact session. Never pass an approval flag. Before a repeated task, ask `managed_agent_learning` for a recommendation in the exact task/policy/evaluator scope. It is advisory and cannot expand budget, worker permissions, network access or default models. After independent host verification, record the closed outcome and host Ed25519 signature. Unknown or unsigned observations stay quarantined. Never request the verifier private key. Use `index`/`recall` only for authenticated training evidence; preserve reported backend/embedding failures. Use `evaluate` on a frozen, disjoint heldout manifest with three or more trials before treating a strategy as eligible. Eligibility never grants promotion, deployment or configuration writes. After the host verifier has tested and signed outcomes, use managed_agent_learning action=cycle with a closed cycle object: scope, at most 24 evidence records, optional index, maxP95Ms and evaluation (manifest, baselineId, candidateId). One atomic batch precedes bounded native indexing and advisory evaluation. Replays preserve attempt identity. Never place commands, keys, approval or promotion flags in a cycle. A failed later phase can leave the signed batch persisted; inspect status and retry the same evidence rather than inventing new attempts. This is an explicit feedback step, not automatic task execution.