v2.0.17

Patch Changes

• cc82c81: Add the opt-in advisory ask_system1 consumer for the parent Agent Hub in operator and orchestrator modes. It batches choice, predicate, and ordinal questions through the existing System 1 service. Advice does not execute commands, grant permissions, change tiers, or close acceptance, assertion, plan, or review gates.

  Enable it only in the shared human-owned .ai/system1.json v2 document under consumers.agenticAsk. The installed default is off. recommended and advisory both require explicit remote-data consent; a preserved consumer config cannot re-enable a disabled System 1 feature. recommended instructs the Hub to consult first for suitable bounded semantic judgments, including small tasks, without expanding file scope, output capture, or process authority. Start a new Hub session after changing the mode.

  Source capture is explicit and fail-closed: selected repo paths and ranges, optional recorded Hub bash evidence only when allowToolOutputs is approved, and no file or output bodies in the tool result. Damage-control policy still governs local reads. Refusals, missing evidence, stale advice, and an exhausted session budget fall back to ordinary reading and research. Native workers and Hub-spawned peers do not inherit the tool.

  The Fleet System 1 communication viewer can show an agenticAsk projection of status, counts, latency, and usage without question text or answer values. npm run test:agentic proves the engineering integration with fake answers and no network; it does not claim semantic accuracy or measured savings.