v1.0.17

Fixed

• Scale the StrangeLoop graph recursion_limit with agent.loop.max_iterations so clarification rounds and plan reviews don't exhaust LangGraph's default budget and abort the loop mid-goal.
• Deduplicate manual clarification cards on late re-emits: a submitted widget is reused instead of mounting a second question card after the user has already answered.
• Tag captured interrupts with goal_id and drop relay-inbox entries lacking a matching goal_id, fixing the approve-tool infinite-loop where a stale interrupt from a cancelled prior goal leaked into the newly submitted goal.
• Route ask_user defer / failure outcomes to the retry sentinel in bypass mode (instead of the human relay), so bypass runs stay autonomous instead of stalling on an unanswered clarification interrupt.
• Degrade classifier escalate verdicts to instructive rejects in autopilot (no human attached), matching the deterministic escalate path and preventing indefinite interrupts with no one to answer them.

Changed

• Default InlineGateConfig.active_in_bypass to False: bypass mode now permits all gated tool calls by default. Set True to keep deny rules absolute even in bypass.
• Strip RFC references from middleware / config docstrings (no behavior change).