Alignment Workflow Reference

The alignment phase starts after Qwen Code has a candidate implementation. Use it to create a tight loop: run the selected reference agent and Qwen Code, compare traces, patch the target, and rerun only the failing scenario.

Trace Inputs

Expected raw capture layout:

.repro-runs/<scenario>/
  reference/
    http.jsonl
    command.stdout
    command.stderr
    command.exit
    state-before/state-manifest.json
    state-after/state-manifest.json
    state-diff/state-diff.md
  qwen/
    http.jsonl
    command.stdout
    command.stderr
    command.exit

Use capture scripts from $agent-reproduce-feature for raw capture, or use run_pair_capture.sh for simple non-interactive shell scenarios. Set REPRO_REFERENCE_AGENT=codex or REPRO_REFERENCE_AGENT=claude-code with the paired runner to capture reference-agent state automatically.

Normalization

normalize_trace.py reads mitm JSONL output and emits stable JSON:

• request method and URL path
• JSON request body summary
• message role order and brief content hashes
• tool/function names
• schema required fields
• response status code

It intentionally drops:

• timestamps
• authorization and cookie headers
• provider request IDs
• full message text unless needed for a hash

Diff Triage

High priority:

• missing request entirely
• wrong endpoint family
• missing tool/function schema
• incompatible required fields or enum values
• slash command not routed to the same behavior class
• state changes that prove the feature writes config, memory, permissions, or another user-visible local store

Medium priority:

• prompt role ordering differences
• terminal output phrasing differences
• streaming versus non-streaming if users can observe it
• unexplained state changes that plausibly affect future runs

Low priority:

• timestamps, IDs, token counts
• harmless wording differences
• extra target-side metadata ignored by the provider

Iteration Loop

1. Pick the highest-priority unexplained mismatch.
2. Patch only the likely owner module in Qwen Code.
3. Run the focused test/smoke path.
4. Capture only the affected scenario again.
5. Refresh the reference state diff if the suspected mismatch involves local state.
6. Normalize and compare again.

Stop when the target behavior is compatible and remaining differences are either irrelevant or explicitly documented.