deep-discoverylisted

# Deep discovery (Workflow-driven) The heavyweight discovery fan-out for **L3/L4** changes that touch many seeds (models / services / tables). It maps each seed in parallel and consolidates one blast radius, instead of running one `impact-mapper` at a time. ## When to use - **Explicit invocation only**, **L3/L4 only.** For L2 use `start-task`'s single `impact-mapper`; for L0/L1 a self-trace. Decline politely and point there otherwise. - Drives the `Workflow` tool (~15× tokens) — invoking it is the opt-in. - **Reuse the cache first:** honor `.claude/groundwork/impact/<slug>.md` staleness (see `${CLAUDE_SKILL_DIR}/../../guidelines/working-memory.md`) — do not re-run if the seeds are unchanged. - **If `Workflow` is unavailable**, fall back to spawning a single `impact-mapper` (the `start-task` flow) and say so. ## How to run 1. List the seeds (models, services, tables, symbols) and pick a `<slug>` from the primary seed. 2. Run the Workflow in `${CLAUDE_SKILL_DIR}/workflow.md`: parallel `impact-mapper` per seed (+ a `grounded-researcher` if an external API is involved) → dedup into one blast radius + ranked hotspots → verify the dynamic / unresolved edges. `log()` the seed count. 3. Write the consolidated map to `.claude/groundwork/impact/<slug>.md` and refresh the `BASE_COMMIT` / `SEEDS` header (working-memory layer). ## Output One blast-radius map · ranked hotspots · checked dynamic-edge findings · a ranked read-list · open questions. Workers run via `agen