verification-loop

name	verification-loop
description	Verify DeepSearch changes with scoped diff review, import checks, pytest, optional live tests, and artifact hygiene.
disable-model-invocation	true

Verification Loop

Use this after implementing a feature, fixing a bug, or preparing a PR.

Phase 1: Scope Review

git diff --stat
git diff --name-only

Confirm the changed files match the task. Generated reports, logs, .env, databases, and runtime artifacts should not appear unless explicitly requested.

Phase 2: Import Check

For Python source changes:

PYTHONDONTWRITEBYTECODE=1 uv run python -m compileall -q openjiuwen_deepsearch server

Skip this for docs-only changes and state that it was skipped.

Phase 3: Targeted Tests

Run tests closest to the changed area:

uv run pytest tests/path/to/test_file.py

Examples:

• Query understanding -> tests/algorithm/query_understanding/
• Report generation -> tests/report/
• Server manager/API -> tests/server/
• User feedback -> tests/user_feedback_processor/

Phase 4: Broad Non-Live Tests

For changes spanning multiple areas:

uv run pytest -m "not llm"

Phase 5: Live Tests

Only run live LLM/search tests when the user requested it and credentials are configured:

RUN_LLM_TESTS=1 uv run pytest -m llm

Phase 6: Final Diff Review

git diff -- AGENTS.md CLAUDE.md .claude
git diff

For non-AI-guidance tasks, review all changed files. Check that public behavior has tests and docs when needed.

Report Format

Verification Report
Scope:     PASS
Import:    PASS / SKIPPED
Targeted:  PASS / SKIPPED
Broad:     PASS / SKIPPED
Live:      PASS / SKIPPED
Diff:      PASS
Status:    READY