Assistant Agent / RAG modes hang in v1.8.2 against trustgraph-flow ≥ v2.3.1 (downstream of trustgraph-client#17) #141

Summary

In the shipped trustgraph/workbench-ui:1.8.2 image, the Assistant → Agent mode (and any flow that routes through agentStreaming()) hangs on "Processing chat message…" indefinitely when paired with any trustgraph/trustgraph-flow:>=2.3.1 backend.

Root cause is in the bundled @trustgraph/client — the JS switch operates on resp.chunk_type but the backend renamed that field to message_type in trustgraph-ai/trustgraph PR #795 (commit d2751553, 2026-04-13). Full analysis and a minimal patch are filed at trustgraph-ai/trustgraph-client#17.

Reproduction

1. docker compose with trustgraph/workbench-ui:1.8.2 + trustgraph/trustgraph-flow:2.3.21 (or anything ≥ v2.3.1).
2. Workbench → Assistant → mode Agent → "Who is X?" → Send.
3. Spinner stays forever. WS frames arrive (verifiable in browser DevTools); each carries message_type: \"thought\"|\"observation\"|\"answer\"|… and content: …. The bundled index-*.js does switch(g.chunk_type){case \"thought\": …; case \"answer\": case \"final-answer\": …}, so every chunk hits the default and produces no UI update. Backend completes the request with end_of_dialog: true.

CLI works fine (tg-invoke-agent) because it doesn't go through this client. Graph RAG / Document RAG UIs partially work (their explain branch already had a message_type fallback) but the agent-driven RAG paths exhibit the same hang.

Impact

Any deploy of upstream images released after 2026-04-13 makes the entire Agent UX unusable. Last commit on workbench-ui is 9b6541c1 from 2026-04-08 — five days before the backend rename, so no compatible release pair exists.

Fix path

The actual fix is one level down — update @trustgraph/client per trustgraph-ai/trustgraph-client#17 and ship a new workbench-ui release that picks it up. We have verified locally that bumping the client to the patched version (renamed chunk_type → message_type, AgentResponse fields unified) and rebuilding the SPA fully restores Agent mode (Thinking / Observation / Answer all rendered correctly, no console errors).

Happy to send a PR that bumps @trustgraph/client and rebuilds, once trustgraph-client is patched and a new version is published.

Workaround

Local trustgraph/workbench-ui:1.8.2-msgtype image rebuilt from sources with the client patched and linked via file:../trustgraph-client. Drops in transparently.

Refs

• Backend rename: Agent explainabilty improvements trustgraph#795 / d2751553 (2026-04-13)
• Last chunk_type-emitting backend tag: v2.3.0
• Upstream fix tracking: AgentResponse.chunk_type field not renamed to message_type after backend PR #795 (workbench-ui Agent mode hangs) trustgraph-client#17