deer-flow/frontend/src/components/workspace/AGENTS.md

Workspace components (frontend/src/components/workspace/)

The signed-in workspace UI: sidebar, composer, artifact panel, settings dialogs, the Capability Center and plugin pages. Product contracts: capability center, full-stack plugins. Message rendering rules are in the frontend source guide.

Verify

cd frontend
pnpm check
pnpm test
pnpm catalog:sync   # after a change to the capability catalog

Requests and dialogs

• A dialog aborts its requests when it closes, and ignores a callback that arrives for an earlier user, skill or thread. A late response must not write into a dialog that now shows something else.
• A write is never retried automatically. Model import, plugin backend calls and deletion show the failure and let the user retry.
• A change to a connection's destination clears its credential draft. A key typed for one endpoint must not be sent to another. A blank key keeps the saved one; an explicit removal sends an empty key.
• Thread deletion is hosted outside the virtualized sidebar lists. The dialog keeps its thread snapshot and retry state when a partial deletion removes the row. onDeleted runs after both the remote and the local step succeed.
• Project memory promotion compares the saved instructions. Disable it while the editor has unsaved changes. After promotion, refetch the project details so the editor sees the new instructions.
• A browser download handoff does not prove that the file was saved. Do not report an export as stored.
• Read an error body as unknown, and narrow each field before use. A malformed body keeps the typed HTTP failure.

Capability Center

• Installed, enabled, credential-configured and authorization-verified are four states. Keep them apart in the UI.
• Provider identity comes from catalog metadata. Do not infer a provider or a brand icon from a server display name; a custom server name never selects a brand.
• The catalog source is capabilities/builtin.json in the harness. The demo snapshot is generated from it, and a unit test compares the two.
• Agent plugin selection uses installation IDs. null means all and [] means none. The dialog submits only the selections that changed since it opened, so a concurrent update elsewhere survives. The selection is a runtime choice, not an authorization policy.
• MCP secrets and raw settings stay in the administrator editor. Other views read the safe status projection. A plugin icon is a bounded PNG data URL in presentation metadata and never enters transport parameters. Remote URLs and SVG are not rendered.
• Skill diagnostics check YAML syntax only. Keep the scope note visible, and do not interpret parser exception wording in the UI.
• pluginPages lists enabled runtime entries. A catalog entry does not imply activation.

Composer references

• A label grants no read authority. The backend checks every referenced skill against the user's registry and the agent allowlist.
• Render a label as DOM text, never as HTML.
• Only a successful discovery may flatten references. While discovery is pending or failed, keep the IDs and block a send that carries references.
• Commit mention text before you restore focus, in the same tick. A delayed callback would overwrite what the user typed in between.
• The web slash picker lists composer commands only. Skills are selected with @. Typed slash text is still submitted as a normal message, which the backend and the IM channels rely on.

Delimited preview

• CSV parsing runs in the worker only, with bounded input, cancellation of stale work and a timeout. Parsing on the main thread freezes the workspace on a large file.
• The parser is told the record separator explicitly. Auto-detection is corrupted by a newline inside an incomplete quoted field of a truncated sample.