phase2/a2aagent-reconciler — full reconciler + AgentCard compile + helm CRD (S3)

[Pal Lakatos-Toth (pallakatos)]

Summary

Phase 2 S3 — full A2AAgent reconciler. Closes §14.6 column 4 (A2A 1.2 + AP2 — schema → AgentCard publication path). Mirrors S2's compile-and-publish pattern: spec → wire-format AgentCard JSON → ConfigMap. Router-side /.well-known/agent.json mount + JWS signing + trust-store informer wiring deferred to S7 (phase2-conditions-ssa-leader).

Responsibility boundary (AGT)

AzureClaw owns: CRD schema, K8s reconciliation, ConfigMap distribution, helm/drift detection.
Upstream Microsoft agentmesh v3.1.0 from crates.io (unmodified) remains the policy authority.
No fork of AGT. The vendored vendor/ directory contains only AgentMesh transport (npm SDK + relay/registry from amitayks) and is unrelated.

What ships

• controller/src/a2a_agent.rs — CRD struct (group/version/kind/shortname a2a), spec sub-types (A2aSigningKey, TrustThresholds, FederationPeer, PolicyRefs), status. signingKeys[*] shape mirrors inference-router::a2a::agent_projection::A2aAgentSigningKeySpec 1:1 — wire bytes traverse controller → ConfigMap → router with no transformation.
• controller/src/a2a_agent_compile.rs — pure compile_agent_card(spec, namespace, name) → Value + version_hash (sha256 prefix). 6 unit tests.
• controller/src/a2a_agent_reconciler.rs — full reconciler. Field manager azureclaw-controller/a2aagent, finalizer azureclaw.azure.com/a2aagent-cleanup, ConfigMap a2aagent-{name}-card. 7 unit tests.
• controller/src/crd_validations.rs — 4 CEL admission rules (signingKeys non-empty, EdDSA-only, productionMode⇒https, federation in-cluster/external mutual exclusion) + 5 admission tests.
• controller/src/helm_drift.rs — third helm-drift gate (a2aagent).
• deploy/helm/azureclaw/templates/crd-a2aagent.yaml — generated by dumper.
• docs/security-audits/2026-04-27-phase2-a2a-reconciler.md — full audit (§0 reuse map of ~14 seams, STRIDE, OWASP A2A, explicit out-of-scope list, two sign-offs).

New slice added: S3.5 phase2-a2a-gateway-component

ADR-0001 ("A2A 1.0 ingress — single gateway, router never publicly exposed") is Accepted. Phase 1 shipped 8 of its 9 implementation steps; only step #4 (the gateway binary itself) was missing. Without it, only outbound A2A is usable end-to-end. S3 ships the AgentCard data; S3.5 ships the public edge that lets inbound A2A 1.2 federation actually receive traffic.

Added to:

• docs/implementation-plan.md §8 scope item 2a.
• plan.md slice list (between S3 and S4).

Tests

• +16 controller unit tests (6 compile + 7 reconciler + 5 admission/drift).
• Full controller suite: 193 passing (was 177 after S2). Workspace total unchanged otherwise (router 595, integration 26, all green).

Gates (BASE_REF=origin/dev)

Gate	Result
cargo fmt --all -- --check	✅
cargo build --all	✅
cargo test --workspace	✅ 193 controller / 595 router / 26 integration
cargo clippy --all-targets -- -D warnings	✅
ci/check-loc.sh	✅
ci/no-stubs.sh	✅
ci/no-custom-crypto.sh	✅
ci/security-audit-required.sh	✅
ci/no-null-provider-prod.sh	✅
ci/a2a-module-isolation.sh	✅
ci/vendored-patch-audit.sh	✅
Helm CRD drift (a2aagent)	✅
CLI npm run typecheck	✅
CLI npm run lint	✅ (26 pre-existing warnings, 0 errors)

Merge

Same cadence as S1/S2: squash + admin merge once Rust CI is green; container scan deferred to the dev → main integration cadence.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com