v0.9.4

Added

• Agent install guide in Settings key-creation dialog: Creating an agent API key from Settings used to dead-end on the raw key with no setup guidance, while onboarding had a full 5-client install guide. The guide is extracted into a shared AgentInstallGuide component (single source of truth for both surfaces) and rendered inline in the create-key success state with the freshly created key embedded, so users can connect their agent right where they got the key. Onboarding refactored to consume the same component (behavior-preserving). (#295)

Changed

• Elaboration flow simplified — unified round generation, optional roundUuid, idea-level resolve: chorus_pm_start_elaboration now generates any round (first, follow-up, or appended-after-resolution) and is callable both while elaborating and after the Idea is resolved; appended rounds set isAppended=true and keep the Idea resolved so an in-flight proposal is never re-blocked. chorus_answer_elaboration's roundUuid is optional — auto-locates the single active (pending_answers) round. chorus_pm_validate_elaboration is re-scoped to a pure idea-level resolve action: takes only ideaUuid (no roundUuid), gated on idea:admin, requires human confirmation (except YOLO), with precondition "≥1 round and every round answered" — never mutates round status. A round's only active states are now pending_answers → answered; validated/needs_followup are legacy-only. New ElaborationRound.isAppended column (DDL-only migration); UI surfaces a "Follow-up" badge on appended round cards. Skills across all four surfaces (Claude Code, Codex, OpenClaw, standalone) make the loop explicit: re-enter start_elaboration when answers derive new questions or the human raises a concern at the resolve gate; resolve only once the loop has settled. (#296, #297)
• Task edit UI's acceptance-criteria editor aligned with Task Draft: The real Task edit form previously rendered AC as a single legacy Markdown textarea while the Task Draft panel used a structured rows editor — the two surfaces had drifted apart. The structured AC editor is extracted into a shared controlled component so both panels render the same UI by construction. Real-task edits persist through the existing replaceAcceptanceCriteria service via a small extension to updateTaskFieldsAction. Client-side change detection ensures the destructive replace runs only when the criteria set actually changed, so verification marks survive non-AC edits. (#299)

Plugin

• Claude Code plugin → 0.9.4 and Codex plugin → 0.9.4: Lockstep version bump for the elaboration flow simplification. All four skill surfaces (Claude Code, Codex, OpenClaw, standalone) updated with the explicit ask→answer loop guidance, validate reduced to ideaUuid-only, and the appended-round semantics. (#296, #297)
• OpenClaw plugin → 0.5.3 (skills 0.9.4): Independent package version bump tracking the elaboration flow simplification. (#296)