dynamic-workflowslisted

# Dynamic Workflows Generate and run OpenClaw-native multi-agent orchestration programs. You write a `.prose` program for the task, validate it with OpenProse, then execute it through the host's OpenProse skill/command surface. Use direct session spawning only as a bounded fallback for small plans. The user speaks natural language; you handle the orchestration. A finding is kept only after it survives an independent _refute_ pass — that gate is what separates `verified` from `partial` in the final synthesis. ## Success contract A correct run produces four artifacts: 1. A pattern choice with a one-line reason. 2. A complete `.prose` program (or bounded direct-session plan) in which every agent is a standard role (✓) or a `custom_*` role (⚠) listed for approval. 3. Validation evidence: `prose compile <file>` output, or manual validation against the 5 checks in Step 3 (indent / inputs / agents / unique vars / synthesis-ending). 4. A final synthesis that separates _refuted_ (discarded) from _survived_ (verified) findings, with partial/blocked branches labeled. OpenClaw contract: prefer OpenProse because it keeps intermediate branch output inside workflow state and returns only the final synthesis to the user context. Direct `sessions_spawn` fallback is bounded (see Resource limits for the cap); do not use it for loops, recursive search, large pipelines, tournaments, or 6+ branch fan-out. If the task needs file edits, use a git branch or equivalent checkpoint be