E2E Testing Guide

How to run the Qwen Code CLI end-to-end — from building the bundle to inspecting raw API traffic. Use when unit tests aren't enough and you need to verify behavior through the full pipeline (model API → tool validation → tool execution).

Setup

Which binary to use

• Reproducing bugs: use the globally installed qwen command — this matches what the user ran when they filed the issue.
• Verifying fixes: build first (npm run build && npm run bundle), then run node dist/cli.js — this tests your local changes.
• Runtime-only checks (fastest): npm run dev -- "<prompt>" <flags> — runs TS source via tsx, no build. Use build && bundle + node dist/cli.js only when the shipped artifact itself matters. (<qwen> below can be npm run dev --.)

Running against a real model

Headless auth comes from ~/.qwen. Force a known-good model with --auth-type + --model:

<qwen> "your prompt" --auth-type openai --model deepseek-v4-flash \
  --approval-mode yolo --output-format json

Gotcha: --model alone won't switch providers — --auth-type (openai/anthropic/qwen-oauth/gemini/vertex-ai) does. Omit it and the run falls back to the default provider and dies on its missing key.

Isolating runtime artifacts

QWEN_RUNTIME_DIR=<dir> redirects qwen's runtime output — tmp/, debug/, and projects/<sanitized-cwd>/... (chat recordings, auto-memory, history) — into <dir> instead of ~/.qwen. Config (settings.json, OAuth tokens, commands/) still reads from ~/.qwen, so real auth and provider config work without any setup.

Use when repeated test runs would clutter your real chat history or auto-memory. Skip when the bug you're reproducing depends on the user's actual history or runtime state — that is the repro.

QWEN_RUNTIME_DIR=/tmp/test-1/runtime <qwen> "prompt" ...

Run modes

Headless Mode

Run the CLI non-interactively with JSON output (<qwen> = qwen or node dist/cli.js per above):

<qwen> "your prompt here" \
  --approval-mode yolo \
  --output-format json \
  2>/dev/null

--output-format json emits one JSON array (all messages, flushed at end of turn) — filter with jq '.[] | …', never a bare jq 'select(…)'. (--output-format stream-json instead emits NDJSON, one object per line.) Element types:

• type: "system" — init: tools, mcp_servers, model, permission_mode
• type: "assistant" — model output: content[].type is text, tool_use, or thinking
• type: "user" — tool results: content[].type is tool_result with is_error
• type: "result" — final output with result text and usage stats

Filter with jq — lead with .[] to enter the array, e.g. tool-result errors: ... 2>/dev/null | jq '.[] | select(.type=="user") | .message.content[] | select(.is_error)'

Interactive Mode (tmux)

Use when you need to verify TUI rendering, test keyboard interactions, or see what the user sees. Headless mode is simpler when you only need structured output.

Launching

tmux new-session -d -s test -x 200 -y 50 \
  "cd /tmp/test-dir && <qwen> --approval-mode yolo"
sleep 3  # wait for TUI to initialize

Sending prompts

Split text and Enter with a short delay — sending them together can cause the TUI to swallow the submit:

tmux send-keys -t test "your prompt here"
sleep 0.5
tmux send-keys -t test Enter

Waiting for completion

Poll for the streaming indicator to disappear instead of blind sleeping. The footer placeholder Type your message is always rendered — don't grep for that or the loop exits on iteration 1 while the model is still working. The status line esc to cancel is present only while the model is producing output:

for i in $(seq 1 60); do
  sleep 2
  tmux capture-pane -t test -p | grep -q "esc to cancel" || break
done

Capturing output

tmux capture-pane -t test -p -S -100   # -S -100 = 100 lines of scrollback

Limitations

• Key combos: tmux send-keys cannot reliably send all key combinations. C-?, C-Shift-*, and function keys with modifiers are unsupported or unreliable. For these, use the InteractiveSession harness in integration-tests/interactive/ or test manually.
• Visual artifacts: capture-pane captures the final rendered frame, not intermediate states. Flicker, tearing, or brief blank frames cannot be detected this way.

Cleanup

tmux kill-session -t test

Inspecting

Inspecting Raw API Traffic

When debugging model behavior (wrong tool arguments, schema issues), enable API logging to see the exact request/response payloads:

<qwen> "prompt" \
  --approval-mode yolo \
  --output-format json \
  --openai-logging \
  --openai-logging-dir /tmp/api-logs

Each API call produces a JSON file (can be 80KB+ due to full message history). The bulk is in request.messages (conversation history). Trimmed structure:

{
  "request": {
    "model": "coder-model",
    "messages": [
      { "role": "system|user|assistant", "content": "...", "tool_calls?": [...] }
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "tool_name",
          "description": "...",
          "parameters": { ... }      // schema sent to the model
        }
      }
    ]
  },
  "response": {
    "choices": [
      {
        "message": {
          "role": "assistant",
          "content": "...",          // text response (may be null)
          "tool_calls": [
            {
              "id": "call_...",
              "function": {
                "name": "tool_name",
                "arguments": "..."   // raw JSON string from the model
              }
            }
          ]
        }
      }
    ]
  }
}

Structured-output calls (those requesting a JSON schema, e.g. side queries via BaseLlmClient.generateJson) deliver the schema as a synthetic tool named respond_in_schema under request.tools[0] — not under response_format, which is null for OpenAI-compatible providers. The model's structured reply lands in tool_calls[0].function.arguments instead of message.content. Text-mode calls have no tools and use message.content.

Token Usage Stats

Use scripts/token-stats.py to summarize token usage across recent API logs:

python3 .qwen/skills/e2e-testing/scripts/token-stats.py 20  # last 20 requests

Shows input, cached, and output tokens per request with cache hit rates. Useful for verifying prompt caching behavior or investigating unexpected token counts.

Test harnesses

MCP Server Testing

For testing MCP tool behavior end-to-end, read references/mcp-testing.md. It covers the setup gotchas (config location, git repo requirement) and includes a reusable zero-dependency test server template in scripts/mcp-test-server.js.

Mock OpenAI Server

For driving the CLI through scenarios that are hard to provoke against a real model — specific error codes, malformed tool calls, deterministic multi-turn loops, controlled usage blocks — read references/mock-openai-server.md. It covers when to reach for a mock vs --openai-logging, how to point the CLI at it, and patterns for specializing the zero-dependency template at scripts/mock-openai-server.js.

Tips

• Use interactive (tmux) mode when the bug involves permission prompts, slash commands, or keyboard interactions. Headless mode has no TUI — these don't exist there.
• Use interactive (tmux) mode for hang-related issues. Headless mode produces no output when the process stalls, giving you nothing to work with.
• Use --approval-mode default when testing permission rules. yolo bypasses rule evaluation entirely — it can't test whether a rule matches.