AgentResponse.chunk_type field not renamed to message_type after backend PR #795 (workbench-ui Agent mode hangs)

[miolamio]

Summary

The TypeScript client's AgentResponse interface and the agentStreaming() switch still use the legacy chunk_type field, but the backend renamed it to message_type on 2026-04-13 in trustgraph-ai/trustgraph PR #795 ("Add agent explainability instrumentation and unify envelope field naming", commit d2751553).

Effect: every released trustgraph-flow from v2.3.1 onward (including v2.3.21 and all of v2.4.x) emits agent stream frames as {message_type: "thought" | "observation" | "answer" | …, content: …}. The client's switch on resp.chunk_type falls through the default case, so callbacks (think, observe, answer) are never invoked. Consumers like workbench-ui:1.8.2 get a stream of frames that silently produce no UI updates — the Assistant Agent / Graph RAG / Document RAG modes hang on "Processing chat message…" forever even though the backend completes the request with end_of_dialog=true.

Affected files

• src/models/messages.ts:130-154 — AgentResponse declares both chunk_type (the active field used in the switch) and message_type?: \"chunk\" | \"explain\" (the old envelope distinguisher). They now collide semantically because the backend uses message_type for the full set thought | action | observation | answer | final-answer | explain | error.
• src/socket/trustgraph-socket.ts:1400, 1406, 1426 — agentStreaming():
  • error check: `if (resp.chunk_type === "error" || resp.error)`
  • explain branch: `if ((resp.chunk_type === "explain" || resp.message_type === "explain") …)`
  • main switch: `switch (resp.chunk_type) { case "thought": … case "answer": case "final-answer": … }`

Backend evidence

Backend trustgraph-base/trustgraph/messaging/translators/agent.py:

def encode(self, obj: AgentResponse) -> Dict[str, Any]:
    result = {}
    if obj.message_type:
        result[\"message_type\"] = obj.message_type
    ...

Backend trustgraph-base/trustgraph/schema/services/agent.py:

@dataclass
class AgentResponse:
    message_type: str = \"\"  # \"thought\", \"action\", \"observation\", \"answer\", \"explain\", \"error\"
    ...

Grep across trustgraph-ai/trustgraph master: chunk_type is absent from all .py sources; only present in two design-doc translations under docs/tech-specs/{he,sw}/*.md.

Reproduction

1. Deploy trustgraph/trustgraph-flow:2.3.21 (post-rename) + trustgraph/workbench-ui:1.8.2 (bundles @trustgraph/client@1.7.x pre-rename).
2. Open Workbench → Assistant → switch mode to Agent → send any question.
3. Browser shows "Processing chat message…" indefinitely. Browser DevTools WS frames show 100+ messages each with {message_type: ...} but no chunk_type field. Server-side gateway logs show the agent completed with end_of_dialog=true.

Suggested fix

Rename chunk_type → message_type consistently across AgentResponse and agentStreaming(), and drop the now-redundant message_type?: \"chunk\" | \"explain\" slot from AgentResponse (the values fold into the same union).

Minimal diff (verified locally — full test suite (66 tests) passes, typecheck clean):

```diff
--- a/src/models/messages.ts
+++ b/src/models/messages.ts
@@ export interface AgentResponse {

• // Streaming response format (new protocol)
• chunk_type?: "thought" | "action" | "observation" | "answer" | "final-answer" | "explain" | "error";

• // Streaming response format — unified envelope field (trustgraph PR #795)
• message_type?: "thought" | "action" | "observation" | "answer" | "final-answer" | "explain" | "error";
  message_id?: string;
  …

• // Explainability fields
• message_type?: "chunk" | "explain";

• // Explainability fields (carried when message_type === "explain")
  explain_id?: string;
  explain_graph?: string;
  explain_triples?: Triple[];
  }
  ```

```diff
--- a/src/socket/trustgraph-socket.ts
+++ b/src/socket/trustgraph-socket.ts
@@

•  if (resp.chunk_type === \"error\" || resp.error) {
  

•  if (resp.message_type === \"error\" || resp.error) {
     error(resp.error?.message || \"Unknown agent error\");
     return true;
   }
  

•  if ((resp.chunk_type === \"explain\" || resp.message_type === \"explain\") && resp.explain_id && resp.explain_graph) {
  

•  if (resp.message_type === \"explain\" && resp.explain_id && resp.explain_graph) {
     …
   }
  

•  switch (resp.chunk_type) {
  

•  switch (resp.message_type) {
     case \"thought\": …
     case \"observation\": …
     case \"answer\":
     case \"final-answer\": …
     case \"action\": …
   }
  

```

Workaround

We are running a locally-built @trustgraph/client with the rename and a custom workbench-ui:1.8.2-msgtype image rebuilt from sources — happy to upstream as a PR if it'd help.

Refs

• Backend rename: Agent explainabilty improvements trustgraph#795 / d2751553 (2026-04-13)
• Last trustgraph-flow tag still emitting chunk_type: v2.3.0