control-plane-continuity

Control-plane continuity

Use this skill when the operator asks for continued progress, a larger local-safe batch, or a fresh session needs to recover the continuity profile without rereading long project history.

Operating Frame

• Treat this as a control-plane profile, not a scheduler, executor, swarm, or unattended service.
• Prefer one useful local-safe slice over a long plan.
• Before a larger batch or worker discussion, use environment_runtime_health_status when available; if it returns safe-mode or stop-and-investigate, reduce surface or stop before adding work.
• Treat needs-evidence as an evidence request, not as stop-and-investigate. Stop only when the packet reports a real blocker, protected scope, unsafe mutation, or active runtime degradation.
• When watchdog safe-mode is on, keep the next step read-only or one small local-safe slice. Do not escalate to expensive/protected capabilities such as worker dispatch, pi-lens, web gateway, remote/offload, or publish until the runtime is stable or the operator makes a fresh explicit decision.
• Do not execute Pi TUI slash commands such as /watchdog:status through bash or worker shell. Those are operator input commands; for agent-readable runtime health, use environment_runtime_health_status and environment_dev_pressure_status.
• If the operator asks to validate runtime health, mentions watchdog, or pastes /watchdog:* commands as intent, pass that text to operator_intent_intake_packet; the packet can infer the read-only runtime-health route. Do not translate the pasted slash commands into shell commands.
• If the operator asks whether workers/subagents are ready or safe to use, pass that text to operator_intent_intake_packet; the packet can infer the read-only worker-readiness route. Readiness checks do not prepare a worker, dispatch a worker, or require budget; budget is required only when preparing a worker packet or dispatching.
• If the operator asks for the next local-safe slice, says there are no eligible tasks, or asks to seed/brainstorm the board, pass that text to operator_intent_intake_packet; the packet can infer the report-only brainstorm seed route. Preview candidates before asking the operator to choose, customize, or cancel.
• Convert free-form intent through operator_intent_intake_packet before asking broad follow-up questions or preparing workers.
• After intake, follow details.controlPlaneAction and details.executionPlan: if it is run-report-only-route, confirmationRequired=false, and executionPlan.executeWithoutTextualConfirmation=true, run the actual tools named in executionPlan.steps in order and summarize their returned details; do not answer from memory or free-form inference, and do not ask the operator to confirm a non-mutating packet. If any named tool is unavailable, report blocked_missing_tool with the missing tool name instead of synthesizing the packet.
• Ask a short interview only when the intake reports missing focus, validation, rollback, budget, or stop conditions.
• Use first-party packets before free-form judgment: operator_intent_intake_packet, project_intake_plan, structured_interview_plan, control_plane_profile_packet, local_batch_manifest_packet, context_watch_continuation_readiness, context_watch_local_slice_preview, local_continuity_loop_canary_packet, context_watch_checkpoint.
• If the operator explicitly authorizes read-only diagnostics, or says not to ask for confirmation, do not answer with a permission question such as "Quer que eu rode?". Run the relevant first-party read-only packets and return a compact decision.
• Delegate to workers only after the lower agent-run gates are green; batch intent never bypasses per-worker start/outcome gates.
• Stop on protected scope, unknown validation, unexpected git state, budget block, compact/reload pressure without checkpoint, repeated failure, or real operator/product ambiguity.