v2026.4.11

Added

• A2A server -- expose any agent as an A2A (agent-to-agent protocol) server via initrunner a2a serve role.yaml. InitRunnerWorker routes through the executor so audit logging, retries, budget enforcement, and output processing all work. Multi-turn context is preserved across requests via task storage. Outbound delegation uses the new A2AInvoker (JSON-RPC message/send with polling, timeout, env-based header resolution, and policy checks). Delegate tools gain mode: "a2a" alongside the existing inline and mcp modes. New [a2a] install extra (fasta2a, uvicorn, starlette)
• Daemon retry policy -- triggers can retry failed runs with exponential backoff. Configure via spec.guardrails.retry_policy: { max_attempts: 3, backoff_seconds: 5 }. Per-attempt usage is recorded; one audit record per trigger fire. Structured error classification (ErrorCategory enum) drives retry-vs-fail decisions
• Circuit breaker -- daemon triggers trip a circuit breaker when the provider is unhealthy. Configure via spec.guardrails.circuit_breaker: { failure_threshold: 5, reset_timeout_seconds: 300 }. Thread-safe state machine (closed/open/half-open) with audit events on state transitions
• Persistent budget counters -- daemon daily and weekly token/cost counters now survive restarts. On startup, the daemon queries the audit log for spend since the last reset boundary and restores the counters. BudgetSnapshot dataclass carries the serialized state
• Real-time cost visibility -- daemon run output now shows USD cost alongside token counts. Dashboard agent detail page includes a BudgetProgressBar component. SSE usage events carry cost for flows and teams
• Timezone-aware budget resets -- budget_timezone guardrail (default "UTC") controls when daily/weekly cost counters reset. Uses ZoneInfo for correct DST handling

Changed

• README redesigned -- opening paragraph and structure rewritten to lead with the execution story (chat -> autonomous -> daemon -> autopilot) instead of "YAML-first AI agent platform." New sections: "One file, four modes," "Agents that learn" (consolidation story), "Security is config, not plumbing" (structured around why each layer exists), "Cost control" (USD budgets). Removed "Why InitRunner" section and feature-count marketing copy. zh-CN and ja translations updated to match
• cli/run_cmd.py split into cli/run_cmd/ package -- _command.py (main run command), _dispatch.py (mode dispatch), _sensing.py (intent sensing), _starters.py (starter catalog), _validate.py (pre-flight validation). No public API changes
• cli/_helpers.py split into cli/_helpers/ package -- _context.py, _display.py, _resolve.py, _console.py. No public API changes
• dashboard/schemas.py split into dashboard/schemas/ package -- per-domain modules (agents.py, audit.py, builder.py, cost.py, flow.py, ingest.py, mcp.py, memory.py, skills.py, system.py, team.py). Re-exports preserved in __init__.py

Fixed

• test_flow_stream_passes_audit_logger failed without dashboard extras -- added pytest.importorskip("fastapi") guard to TestDashboardAuditWiring
• test_reply_fn_called_with_output failed after timezone support -- all DaemonRunner test mocks now set budget_timezone = "UTC" on guardrails
• tests/test_run_dispatch.py lint failure -- reformatted line exceeding 100-char limit