A2A swarm: register hosted Maestro peers and route through Platform discovery

[haasonsaas]

Summary

Teach hosted Maestro to act as a native Platform A2A peer: register and heartbeat its Agent Card, child-agent skills, capacity, current objectives, and policy labels through Platform, then prefer Platform discovery and ledger state when delegating to other remote Maestro instances.

Parent: evalops/platform#2175
Related: #72
Research-backed plan PR: evalops/platform#2387

Why it matters

The Platform side can now prove remote delegation, callback replay, restart recovery, remote subagent evidence, parent result hydration, and targeted task control. Maestro still needs a first-class hosted loop so the product behavior does not depend on ad hoc peer URLs or manually curated staging records.

What to build

• Hosted-mode registration/heartbeat path that publishes Maestro's A2A endpoint, Agent Card freshness, supported extensions, child-agent backed skills, capacity, current objective ids, and policy labels to AgentService.
• A Platform-backed peer discovery client that selects peers by capability/skill/task class/workspace/capacity rather than direct URLs.
• A spec-native A2A client path for message:send, message.taskId resume, push callback setup, artifact polling, and /tasks/{id}:control child-lane steering.
• Fail-open local/offline behavior when Platform is unavailable, while hosted mode records clear degraded-state telemetry.
• Tests for disabled local mode, successful hosted registration, stale heartbeat refresh, peer discovery by skill/capacity, remote task resume, and child-lane control request shaping.

Acceptance criteria

• Two hosted Maestro instances can register with Platform and discover each other without private process coupling.
• A hosted origin Maestro can delegate through Platform-discovered A2A, resume the same task by taskId, and send a targeted child-lane control operation.
• The target Maestro advertises child-agent skills with required context grants, approval policy, max autonomy, artifact expectations, and task-class gates.
• Missing Platform config does not break local/offline Maestro.
• Hosted degraded states are visible in logs/telemetry rather than silently falling back to unsafe direct peer calls.

[haasonsaas]

Landed the first hosted Maestro A2A peer slice via #485 and mirrored/internalized it via https://github.com/evalops/maestro-internal/pull/2097 (internal merged 2026-05-23, merge commit 9b3c14d7).

Verified evidence:

• Public PR test: add local A2A swarm smoke #485 merged successfully.
• Internal PR #2097 merged after addressing the review thread about requiring observed heartbeats before declaring peers ready.
• Live local swarm proof on this host with two real Rust maestro-control-plane instances passed on the final internal PR head: two peers registered, both heartbeat-backed (heartbeatReceivedAgents=2), Platform discovery selected both peers, two remote A2A tasks completed, same-task resume completed, and policy-denied discovery returned zero candidates.
• CI on the merged internal head: pr-checks success, public-release-mirror success, build-and-publish success; coverage and rust-hosted-conformance intentionally skipped by the final smoke-only CI plan.

Leaving this issue open because the remaining acceptance criteria still include targeted child-lane control, target restart/recovery coverage, and richer hosted degraded-state/progress-artifact telemetry beyond this first registration/discovery/delegation/resume/policy slice.

[haasonsaas]

Latest hosted Maestro A2A swarm tranche landed and is mirrored publicly.

Merged:

• Mirror A2A swarm work graph proof #499, public A2A swarm work-graph proof mirror, merge commit 2abf98ac3fe5d7729c8a050b56c6ee76bd6d55d9
• Mirror A2A release readiness sync #508, public release-readiness mirror sync for the final internal base, merge commit fb3696980286375b020bbdf728372abbd1608798
• evalops/maestro-internal#2099, internal implementation/proof, merged at 2026-05-23T07:09:23Z with merge commit 24d65f3a4d8ef8117a3aa233b2e45c7eeb8e9fd7

Proof on the final internal head 2e9505ff1ea62c400860b00fc2137903a6cce3cb:

• npm run test:fast -- test/agent/swarm-executor.test.ts: 38 tests passed.
• npm run smoke:a2a-local-swarm: two local Rust control-plane peers registered and heartbeat-backed; two remote A2A tasks completed; 10 authenticated push notifications received; 2 ledger workGraphs recorded; same-task resume completed; policy-denied discovery returned zero candidates.
• #2099 required checks all passed: coverage, pr-checks, public-release-mirror, rust-hosted-conformance, scenario-replay, review-thread guard, build/publish, and Socket checks.

This closes the next local hosted-proof slice for push callbacks plus remote child workGraph evidence. Keeping the issue open for the remaining broader milestones: promote this into staging/live gates and turn the proven control plane into peer-visible child-lane control semantics.