docs(priorities): refresh architect queue

2026-06-30 14:55:26 +00:00
1 changed files with 4 additions and 2 deletions
@@ -21,9 +21,11 @@ changes, architectural rewrites. Those go to the human.

 ## Work queue (ranked)

-1. **Make long agent runs resumable from checkpoints** ([#3449](https://github.com/micro/go-micro/issues/3449)) — the flow grader loop shipped in #3443 and the run-trace analyzer shipped in #3447, closing the previous top queue item (#3439). With flow durability and optimization now proving the workflow side, the highest-value remaining lifecycle seam is agent-run durability: long agent loops should persist enough progress to resume without replaying completed tool side effects. This aligns the services → agents → workflows runtime by giving agents the same operable recovery posture that flows already have, while staying scoped to a non-breaking, CI-verifiable checkpoint/resume contract.
+1. **Add scheduled cross-provider 0→hero conformance** ([#3454](https://github.com/micro/go-micro/issues/3454)) — the previous top Now/Next items shipped: verification/grader loops landed in #3443, run-trace optimization analysis landed in #3447, and durable agent resume was completed by #3452, closing #3449. With the core loop primitives now in place, the highest-value Now-phase gap is trust: the same scaffold → run → chat/tool → workflow path must stay true across supported providers on a schedule, with keyed runs gated and no-secret CI still useful. This keeps the services → agents → workflows lifecycle cohesive instead of letting provider behavior drift behind green unit tests.

-2. **Verification / grader loop for flows** ([#3435](https://github.com/micro/go-micro/issues/3435)) — founder-prioritized from the [loop-engineering](https://www.langchain.com/blog/the-art-of-loop-engineering) read: of the four loops (agent / verification / event-driven / hill-climbing), the harness provides the agent loop, event-driven flows, and the trace foundation for hill-climbing — but **verification** is the missing primitive. Add `flow.Verify(body, grader)` + `flow.LLMGrader(rubric)` that grade a step's output against a rubric and route failures back with feedback for a bounded retry, building on `flow.UntilLLM` and the existing retry/backoff. Closes the one gap between what we say (operable, trustworthy loops) and what the framework offers as a primitive.
+2. **Emit OpenTelemetry spans for agent RunInfo timelines** ([#3455](https://github.com/micro/go-micro/issues/3455)) — flows now have trace-oriented optimization feedback, and durable agent runs can resume, but the agent side still needs first-class operability in production traces. Translating `RunInfo` / run timeline events into spans closes a Next-phase observability seam across agent runs, tool calls, model calls, retries, failures, and checkpoint/resume events without changing public APIs.
+
+3. **Complete end-to-end chat and A2A streaming coverage** ([#3456](https://github.com/micro/go-micro/issues/3456)) — provider streaming conformance and A2A fallback work recently shipped, but the mission is one runtime where agents can operate as services, which means streaming must be dependable through the whole path: provider tokens → chat / `Agent.Chat` → A2A. This remains behind conformance and observability because it is a Next-phase depth item, but it is the next user-visible seam in the developer inner loop and interop story.

 _Seeded by Claude Code from the roadmap + open issues; thereafter maintained by the
 architecture-review pass._