Add an A2A v1.0 task server and dual-stack Hyrule Handoff intake

[Svaag]

Parent architecture issue: AS215932/agent-core#14

Related reliability work: #29

Problem

Engineering currently consumes an NOC-specific LHP-v1 fetch path and rejects the equivalent SOC path. Intake is pull-oriented and coupled to origin-specific transport details, while the intended Engineering workflow is shared across agents.

Implementation

• Add an A2A v1.0 HTTP+JSON task server using the official Python SDK.
• Publish an Agent Card with a prepare_guarded_change skill and require https://servify.network/a2a/extensions/hyrule-handoff/v1.
• Persist A2A tasks and the caseId / handoffId / source-loop / GitHub issue correlation in PostgreSQL.
• Validate size bounds, schema, authorization, source revision, canonical payload hash, evidence references, and redaction policy before durable acceptance.
• Store the highest accepted sourceRevision; reject regressions and reject a conflicting hash at the same revision.
• Create or reconcile one loop:candidate issue per handoff. Never apply loop:approved automatically.
• Before authorized execution, reconcile the latest CaseHandoff revision, GitHub labels, and required Knowledge context. Refuse stale or mismatched work.
• Report draft PRs, proposed changes, execution evidence, and bounded run output as A2A Artifacts.
• Implement the Hyrule task-state mapping from the parent issue. COMPLETED means the Engineering deliverable exists, not that the origin case is verified or resolved.
• Support client polling in the first release. Keep push notifications behind a feature flag until the production TLS endpoint is ready.
• Retain the existing LHP-v1 intake under a feature flag for shadow comparison and rollback.
• Emit task, revision-conflict, authorization, rejection, retry, and latency metrics/traces without logging secrets or raw untrusted payloads.

A2A status events may wake the Reliability Governor, but they must not authorize execution. #29 remains the owner of persistent wake/reconciliation semantics; this issue owns A2A server and transport integration.

Acceptance criteria

• Valid NOC and SOC profile requests are durably accepted and correlated to exactly one candidate issue.
• Duplicate submissions and service restarts do not create duplicate tasks or GitHub issues.
• Stale revisions, same-revision hash conflicts, incompatible extensions, bad authentication, and oversized payloads are rejected and observable.
• A task waits in INPUT_REQUIRED until the human authorization and domain prerequisites are present.
• A draft PR is returned as an Artifact and maps to change-planned/completed-deliverable semantics only.
• Engineering cannot write the originating case's verified or resolved state.
• Shadow comparison demonstrates A2A/LHP semantic parity and the feature flag restores LHP-only intake.

[hyrule-engineering-loop]

Reliability Governor Decision

• role: staff_sre_autonomous_operations
• decision: needs_human
• next_loop: human
• handoff_contract: human_review
• source: knowledge
• intent: secret / tier 4
• capability: none
• knowledge: current / A0
• reasons: secrets are not explicitly allowed; legal work is not explicitly allowed

{
  "affected_assets": [
    "AS215932/engineering-loop"
  ],
  "affected_customers": [],
  "affected_services": [],
  "allowed_paths": [],
  "authority_text_hash": "e8f44d017eb63b1efa851ee11bff4c8929b4898accc1fcf5ce34455a98d1664f",
  "blast_radius": "credential plane",
  "controlled_loops": [
    "engineering",
    "noc",
    "knowledge"
  ],
  "created_at": "2026-07-21T19:45:38.572022+00:00",
  "denial_reasons": [
    "secrets are not explicitly allowed",
    "legal work is not explicitly allowed"
  ],
  "expected_paths": [
    "secrets/"
  ],
  "forbidden_paths": [],
  "governor_name": "Reliability Governor",
  "governor_role": "staff_sre_autonomous_operations",
  "handoff_contract": "human_review",
  "intent_type": "secret",
  "issue_id": "AS215932/engineering-loop#40",
  "issue_number": 40,
  "issue_text_hash": "82af26f02daed2a2395c310ad5d560e7d437e7de75300694f4243fe654c23240",
  "knowledge_authority_level": "A0",
  "knowledge_context_pack_id": "ctx_622833469ab101620d895a206c6027b0",
  "knowledge_export_version": "local-20260710201321",
  "knowledge_status": "current",
  "labels_to_add": [
    "loop:needs-human"
  ],
  "labels_to_remove": [
    "loop:intake",
    "loop:candidate",
    "loop:approved",
    "loop:needs-context",
    "loop:knowledge-gap"
  ],
  "lhp": null,
  "matched_capability": null,
  "next_loop": "human",
  "policy_rules": [
    "deny sensitive domains unless a capability explicitly allows them"
  ],
  "record_id": "0020b5d17f2b8ec41b35",
  "repo": "AS215932/engineering-loop",
  "required_checks": [],
  "risk_tier": 4,
  "rollback_plan": "Use the rollback/revert procedure specified in the request.",
  "routing_decision": "needs_human",
  "schema_version": "reliability-governor.cdr.v1",
  "source": "knowledge",
  "storage_path": "/var/lib/engineering-loop/reliability-governor/as215932-engineering-loop-40-0020b5d17f2b8ec41b35.json",
  "verification_method": "Use the issue-specified verification: tests/checks/evidence named in the request."
}