Summary

End-to-end smoke test that validates the full pipeline with a real Claude CLI: message → ChannelBridge → ClaudeCliRuntime → claude -p --output-format stream-json → NDJSON streaming → AgentDeliveryResult. Gated behind LIVE=1 environment variable so CI skips it by default.

Context

All middleware components are implemented and unit-tested (367 tests across 16 test files). This is the first test that exercises a real CLI subprocess rather than mocked streams. It validates that the argument building, NDJSON parsing, session resumption, and event extraction work end-to-end with actual Claude CLI output.

CLIs are pre-authenticated on the developer machine — no API key env var checks needed. The CLI handles its own auth.

Acceptance Criteria

Given the middleware is implemented and ChannelBridge works (PR #33)
When a LIVE=1 smoke test is executed with a pre-authenticated claude CLI
Then a test message is sent through ChannelBridge → ClaudeCliRuntime →
  claude --print --output-format stream-json
And a coherent text response is received via NDJSON streaming
And the session ID is captured and a follow-up message resumes the session
And the test setup (beforeAll) unsets CLAUDECODE from process.env to prevent
  nesting rejection when tests run from a Claude Code terminal
  (ref: https://github.com/anthropics/claude-agent-sdk-python/issues/573)
And the test is skippable in CI (gated behind LIVE=1 environment variable)

Architecture

Test Structure

src/middleware/__smoke__/claude-live.test.ts

Uses vitest with a describe block gated by process.env.LIVE === "1". When LIVE is unset, all tests in the file are skipped via describe.skipIf.

Test Flow

1. Setup (beforeAll):

   • Delete CLAUDECODE from process.env to prevent nesting rejection
   • Create a ChannelBridge with provider: "claude" and an in-memory SessionMap
   • No MCP server needed for this smoke test (pure CLI → response)
   • No API key env var checks — CLI is pre-authenticated

2. Test 1: Single-turn response:

   • Send a simple message (e.g., "What is 2+2? Reply with just the number.") via ChannelBridge.handle()
   • Assert: result.payloads.length > 0
   • Assert: result.run.text contains a coherent response
   • Assert: result.run.sessionId is a non-empty string
   • Assert: result.run.aborted === false
   • Assert: result.run.durationMs > 0

3. Test 2: Session resumption:

   • Send a follow-up message using the same channel/user identifiers
   • The SessionMap should return the session ID from Test 1
   • Assert: response is received (session resumed, not rejected)
   • Assert: result.run.sessionId matches the first test's session ID

4. Teardown (afterAll):

   • Restore original CLAUDECODE value if it was set

Key Design Decisions

• No MCP server: This test validates the CLI runtime pipeline, not MCP tools. ChannelBridge passes empty mcpServers when no gateway URL/token is configured (or the test uses a minimal config that skips MCP setup).
• CLAUDECODE unset: When running from a Claude Code terminal, CLAUDECODE=1 is inherited. claude -p refuses to run inside another Claude Code session. The test must unset this env var.
• No API key checks: CLIs are pre-authenticated on the developer machine. The test assumes the CLI is installed and authed — if not, the test fails with a clear CLI error.
• Timeout: Set a generous timeout (60s per test) — real CLI invocations are slow.
• Deterministic prompt: Use a simple factual question to get a short, predictable response.

Related

• Part of a 4-CLI smoke test suite: Implement LIVE=1 smoke test for Claude CLI runtime #34 (Claude), Implement LIVE=1 smoke test for Gemini CLI runtime #36 (Gemini), Implement LIVE=1 smoke test for Codex CLI runtime #37 (Codex), Implement LIVE=1 smoke test for OpenCode CLI runtime #38 (OpenCode)

Dependencies

• ChannelBridge orchestrator (PR feat(middleware): implement ChannelBridge orchestrator #33) ✅

Estimate

~120 LoC test file