run-workflowlisted

# run-workflow — the cross-host dynamic-workflow engine A deterministic orchestration engine (a TS fork of pi-dynamic-workflows, MIT) that decomposes a task, fans it out across isolated sub-agents, verifies, and returns findings — with a positional journal so a re-run replays instead of re-spending. It is the **runtime of the roster's rule**: *serialized writing, parallel reading*. Read `docs/design/dynamic-workflows.md` before changing it. **The one invariant:** the engine is **pure and never writes the ledger**. It fans out read-only sub-agents and RETURNS the surviving pins as JSON. **You** — the agent invoking this skill — write them with the `ledger_add_pin` MCP tool, one at a time. Fan-out is read; the write stays serialized. ## When to use - A task too broad for one context window: codebase-wide finding/audit, multi-perspective review, adversarial red-team of what was elected, or a dependency-ordered build. - Not for: a single edit, or electing a decision (only the human interview elects a to-be). ## Topologies (`--topology`) - **`phase1-finding`** (read-only, default) — multi-modal sweep + loop-until-dry + adversarial verify. Returns surviving **pins** `{file,line,kind,summary}` → write each with `ledger_add_pin`. - **`challenger-verify`** (read-only) — red-team each elected oracle under distinct lenses (unfalsifiable / inconsistent / unsatisfiable / unstated-assumption / ignores-fan-out). Needs `--args-file` with `{"oracles":[{"id","statement"}]}`. Retu