feat: multi-work agent

README.md

@@ -41,6 +41,22 @@ Both installers use `~/.config/opencode`, respect `XDG_CONFIG_HOME`, and accept
 See [GPT configuration](examples/gpt.opencode.jsonc) and [custom model assignments](examples/custom.opencode.jsonc).
 
 For an installed plugin, use `./plugins/stanp-opencode` as the plugin package path. Set `lunaPriority` to `false` to avoid requesting priority pricing.
+Custom configurations require the six worker assignments. The legacy optional `coordinate`
+assignment is accepted for compatibility but deprecated and ignored; orchestration inherits
+`deep-lo`, or `good-lo` when the configured `deep-lo` model is unavailable.
+
+## Multi-work
+
+Select `multi-work` in a verified live Herdr root checkout to delegate local retrieval
+to `fast-lo` and independent work to `coordinate` children in new Herdr worktrees.
+It is primary-only and, like `coordinate`, uses the `deep-lo` model and effort,
+falling back to `good-lo` when `deep-lo` is unavailable. Clients viewing the
+same eligible location can see it; eligibility is revalidated before creation.
+
+Configure `experimental.subagent_depth: 2` at the root and worktree locations.
+The native `subagent` tool accepts `worktree: { base: "source-branch" }` for new
+coordinate tasks and retains native session/job handling. See [usage, permissions,
+failure handling, and verification limits](docs/multi-work.md).
 
 ## Playwright