A2a Test

A2A Test Skill

A2A (Agent-to-Agent) tests verify real runtime behavior against live LLM backends. They are interactive CLI tests using the aw CLI tool — not automated TS scripts.

Test plans

All a2a test procedures are documented as markdown:

• packages/loop/test/a2a/loop-tests.md — Loop-level tests per runtime (AiSdkLoop, ClaudeCodeLoop, CodexLoop, CursorLoop)
• packages/agent/test/a2a/agent-tests.md — Agent-level end-to-end tests across all providers
• packages/agent/test/a2a/messaging.md — Async messaging pipeline tests (batching, interleaving, state transitions)

How to run

A2A tests are run manually using the unified aw CLI (packages/agent-worker/src/cli/index.ts):

# 1. Add an agent with a specific runtime (daemon auto-starts)
aw add test-agent --runtime ai-sdk --model anthropic:claude-haiku-4-5-20251001
aw add test-agent --runtime claude-code --model sonnet
aw add test-agent --runtime codex
aw add test-agent --runtime cursor
aw add test-agent --runtime mock   # test without API keys

# 3. Send messages and observe behavior
aw send test-agent "Reply with exactly: HELLO_A2A_TEST"
aw read test-agent     # View responses
aw log --json          # View debug events
aw state test-agent    # View agent state

# 4. Clean up
aw rm test-agent       # Remove agent
aw daemon stop         # Stop daemon

Test flow

For each runtime, verify in order:

1. Preflight — Agent added successfully (aw add, daemon auto-starts)
2. Simple prompt — Send a trivial prompt, verify text response contains marker
3. Event structure — aw log --json entries have correct type/shape
4. Result structure — run_end has durationMs > 0, usage tracking (where supported)
5. Status transitions — aw log -f shows idle → processing → idle
6. Cancel — aw rm during processing terminates cleanly
7. Tool calls — tool_call_start/tool_call_end pairing (where supported)

Unit tests (separate from a2a)

Unit tests are mock-based and run with bun test:

bun test packages/loop/test/
bun test packages/agent/test/

Loop APIs under test

Each loop class exposes:

• preflight() → PreflightResult — env/config check
• run(prompt) → LoopRun — streaming LoopEvent + .result: Promise<LoopResult>
• cancel() — abort in-flight run
• status — "idle" | "running" | "completed" | "failed" | "cancelled"

Event types

• text — text output
• thinking — reasoning/chain-of-thought
• tool_call_start — tool invocation begins (name, callId, args)
• tool_call_end — tool invocation completes (name, callId, result, durationMs, error)
• error — error occurred
• unknown — unrecognized event from provider

Runtime capability matrix

Verdicts

When running a2a tests, record results as pass/fail/skip in the test matrix at the bottom of each markdown file.