Priorities

The ranked work queue for the autonomous improvement loop. The architecture-review pass (the architect) owns this file: each run it turns the roadmap plus an internal scan (gaps in the services → agents → workflows lifecycle, API coherence, drift, tech debt, test and DX friction) into a single ordered list — highest-value first — and links each item to a tracking issue. The hourly continuous-improvement pass works the top item whose issue is still open. So the architect decides what, and the increment loop builds it.

Reading / editing. An item is done when its linked issue closes (the increment that builds it adds Closes #<issue>). Roadmap phase (Now → Next → Later) is the primary ordering; internal findings are interleaved by value, not kept in a separate list. The human can reorder this list — or the issues — at any time to redirect the loop; direction always wins.

Off-limits to the loop (the architect proposes these as notes, never as queue items the loop can auto-merge): brand/positioning copy, breaking public-API changes, architectural rewrites. Those go to the human.

Work queue (ranked)

1. Trace agent runs with RunInfo OpenTelemetry spans (#3362) — the highest-value Next-phase operability gap now that the getting-started contract shipped in CI: scheduled and looping agents need run/session IDs, model attempts, tool calls, delegation, errors, cancellation, and terminal status visible as traces without leaking sensitive payloads by default. This keeps the agent harness operable in the same runtime developers already deploy.
2. Add human-in-the-loop pause and resume for agent workflows (#3329) — after memory compaction, tool-aware streaming, durable Ask/StreamAsk checkpoints, and the getting-started harness shipped, close the next workflow-operability gap: durable pending-input states that let humans approve or provide context and then resume the same service/agent/workflow runtime.
3. gateway/a2a: multiple typed skills per agent card (#3342) — the A2A gateway now has the full task lifecycle (send/stream/get/cancel/resubscribe, push config, multi-turn, input-required); the remaining interop gap is skill granularity. Card only ever advertises one synthetic "chat" skill with services flattened into tags. Let an agent advertise N typed skills and route per skill, so domain-routing agents expose their real capabilities over A2A through the gateway instead of a hand-rolled handler.

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