Compare commits

...

1 Commits

Author SHA1 Message Date
Codex 1f1c0b5aba Update architect priority queue after lifecycle hooks
Harness (E2E) / Harnesses (mock LLM) (push) Waiting to run
Harness (E2E) / Provider harnesses (live LLM conformance) (push) Waiting to run
Lint / golangci-lint (push) Waiting to run
Run Tests / Unit Tests (push) Waiting to run
Run Tests / Etcd Integration Tests (push) Waiting to run
2026-06-28 00:29:19 +00:00
+21 -8
View File
@@ -19,15 +19,28 @@ redirect the loop; direction always wins.
items the loop can auto-merge): brand/positioning copy, breaking public-API
changes, architectural rewrites. Those go to the human.
## Now (ranked)
## Later (ranked)
1. **Execution lifecycle hooks & metadata** (#2980) — before/after-tool, retry,
and failure hooks; first reconcile the issue with the shipped `AgentWrapTool`,
structured refusal reasons, `RunInfo`, OpenTelemetry spans, durable resume, and
A2A streaming task lifecycle, then close it or scope only a CI-verifiable gap
that those primitives cannot express. Roadmap → resilience/operability; now
ranked first because streaming shipped and this is the remaining open Now-phase
reliability seam for long-running agents and workflows.
1. **Agent memory compaction and retrieval controls** (#3209) — add explicit,
store-backed summarization/retrieval behavior for long-running agents so
durable, streaming, observable runs do not become context-growth demos.
Roadmap → memory management; ranked first because the Now/Next reliability
work has shipped and long-lived agents need bounded, recoverable context before
deeper autonomous workflows can be trustworthy.
2. **Human-in-the-loop pause/resume for agent runs** (#3210) — persist a paused
approval/intervention point and resume safely through the existing
guardrail/run-inspection path. Roadmap → human-in-the-loop pause/resume; ranked
second because it turns guardrails from one-shot refusals into an operable
workflow primitive without changing the public architecture.
3. **x402 spend caps and live facilitator conformance** (#3211) — enforce paid-tool
budgets and add credential-gated live facilitator checks. Roadmap → x402 paid
remote tools; ranked after memory and pause/resume because it hardens a narrower
paid-tool seam but is important for trust once agents can act for longer.
4. **A2A push notifications and multi-turn task support** (#3212) — extend the A2A
gateway/client path beyond streaming lifecycle updates into push and multi-turn
task state. Roadmap → A2A push notifications and multi-turn tasks; ranked after
local run operability because interop depth compounds best once the local run
lifecycle is bounded and inspectable.
_Seeded by Claude Code from the roadmap + open issues; thereafter maintained by the
architecture-review pass._