mirror of
https://github.com/EveryInc/compound-engineering-plugin.git
synced 2026-06-19 15:41:46 +02:00
8351c4daaf
ce-plan bailed out of the workflow on investigative or analytical questions, treating them as non-planning tasks and answering ad hoc (or, for software-topic questions, over-planning them into a plan document the user never asked for). Both contradicted the skill's own "when directly invoked, always plan" rule. Add an answer-seeking disposition to the universal-planning path: when the deliverable is an answer rather than a saved plan, ce-plan states a brief, right-sized plan-of-attack as working scaffold, executes it (research and synthesis only, never code), and delivers the answer in chat with no plan file. Plan-seeking tasks (trips, study plans, events) still produce their saved plan as before; a cross-contamination guard keeps the two flows from bleeding into each other. - Narrow the universal-planning exit gate to genuinely trivial single-fact lookups; multi-step, retrieval, or synthesis questions no longer bail. - Route by task-type, not topic, in Phase 0.1b: a question that merely references code, repos, or APIs is answer-seeking, not software work, while build/modify/refactor requests stay on the implementation path. - Add veil-of-value guidance so answers surface question-domain thinking and hide skill machinery, while never hiding caveats that affect trust in the answer. - Sync the ce-plan skill doc with the two-dispositions mechanic.