[qduc]

Summary

Finishes the agent-harness adoption (#40). Every request with tools now runs one @qduc/agent-core turn. Streaming requests reply with SSE; stream: false requests reply with one JSON body. Both hand-written loops (toolsStreaming.js, toolsJson.js) and the AGENT_HARNESS_TOOL_LOOP flag are removed. On the Responses API, conversations now chain across messages again, with a stricter safety check than before.

Changes

One tool loop for both reply formats (448520d, 7f6bb13)

• The harness loop is the only tool loop. TurnEventSink handles persistence and passes client output to a TurnOutput:
  • SseTurnOutput writes the stream exactly as before.
  • JsonTurnOutput returns the shape the frontend already reads: choices[0].message, tool_events and _conversation. Failures return a 500 with error and tool_events.
• Gaps closed before removing the legacy code:
  • provider_stream: false is honoured. The frontend falls back to it for models that can't stream.
  • Malformed tool-call JSON gets the legacy message, which shows the model what it sent.
  • Responses API tool rounds chain within a turn. This is gated on the model's continuity capability, because chaining Chat Completions, Anthropic or Gemini would drop their history.
  • The Responses adapter now sends a chained request that has no assistant message. Before, it fell back to "last user message" and sent nothing.
  • The UI indents JSON tool results that arrive as text.
• Tool call indexes are numbered per model response. Two calls in one response could both get index: 0, in the stream, in persistence, and in the text-offset lookup.
• Legacy JSON-loop bugs found while recording its behaviour, fixed by the move:
  • A single tool call ran without the user, so the journal failed.
  • The iteration cap made one extra model call and stored empty content.
  • provider_stream: true serialized a raw Response object into the reply.
• executeToolCall and executeToolCallsParallel had no callers left and are removed.

Responses API chaining across messages (2c7ed36)

• A turn's first request continues from the previous turn's stored response and sends only the new user messages.
• The full-history request is kept aside. If the provider refuses the id (any 400 or 404 naming previous_response_id), the request is retried with it. Before, only invalid_value was retried.
• resumableResponseId is used by both the tool loop and the plain path. It chains only from the newest stored reply on the active branch, and only when:
  • that reply is final and has a resp_ id;
  • any tool rows after it answer its own calls;
  • only new user messages follow it.
• The old lookup could skip a newer reply that had no id, and the provider then lost that exchange.
• Only a completed turn keeps its response id. After a cap, cancel or failure the id is cleared, and a cancelled plain reply clears it too.
• The plain path sends every new user message after the stored reply, not just the last one.
• The Responses stream translator ends once. [DONE] arriving after response.completed wrote after the end and stored a good reply as an error.

Verification

• Backend: npm run typecheck and npm run lint are clean. 1275 tests pass.
• npm run test:provider-replay: 30/30.
• Frontend: typecheck and lint are clean. 532 tests pass.
• Streaming tool-loop expectations are the legacy loop's results for the same scripts, recorded before it was removed. The legacy JSON loop's results were recorded the same way and diffed; the differences are the fixes listed above.
• New route-level coverage:
  • JSON reply;
  • text before tool calls;
  • two calls in one response;
  • special characters in arguments;
  • a 200 KB tool output;
  • malformed SSE;
  • the idle timeout;
  • a second message in the same conversation, for every provider family over recorded traffic;
  • plain-path chaining;
  • plain cancel.
• Mutation checks: undoing each fix makes its test fail.

Known limits

• Not run against real providers or in a browser. Everything is tested against fakes and recorded traffic.
• While a Responses conversation is chained, a system-prompt change made mid-conversation doesn't reach the model. Legacy behaved the same.
• With store: false, each message makes one refused request before falling back to full history.
• jest.setup.js makes structuredClone realm-local for all backend tests. Under Jest's VM modules its copies used foreign prototypes, so engine continuity's isDeepStrictEqual checks failed. That doesn't happen in production.
• tool_orchestration_deep_dive.md still describes the legacy loops under a note.

---

Generated by Claude Code