Skip to content

testing_plan.md

Latest commit

 

History

History
49 lines (44 loc) · 2.9 KB

testing_plan.md

File metadata and controls

49 lines (44 loc) · 2.9 KB
Raw

Plan: make --topic/--skill/--rule orthogonal to --mode in cc_lint.py

  • Drop the cross-exclusivity in _main(): keep --topic/--skill/--rule mutually exclusive with each other (existing argparse group, unchanged), but remove --mode's membership in the extra num_exclusive assert, so any of {--topic, --skill, --rule, default-inferred} can be combined with any of {one_shot, session, stateless}.

  • Generalize the incremental branch (_process_file_incrementally / _build_incremental_messages), used when --mode is session/stateless, to build message chunks from whichever "what" was actually specified, instead of only the filename-inferred topic:

    • --topic (or default): unchanged — one chunk per H1 section across all of topic_info["rules"]'s files.
    • --rule: hmarsele.extract_rule_from_file()'s text, split into H1 sections when it contains more than one (whole-file rule spec), else a single chunk (line-anchored spec already extracts one section).
    • --skill: a single, non-decomposed chunk equal to the existing /{skill} {file_path} slash-command string — kept as-is because a skill invocation is a command for Claude Code's own skill loader, not declarative rule prose to split.

  • Leave the one_shot path's prompts/execution untouched: the existing topic/skill/rule/default branches in _process_file() that call _build_prompt() + _run_claude_code() (subprocess cc wrapper) keep their current prompt construction and behavior exactly as today; only their reachability changes (no longer blocked from coexisting with a non-default --mode elsewhere).

  • Factor the H1-splitting core out of file I/O: split _extract_h1_sections(rule_file) into a pure list[str]-based helper plus a thin file-reading wrapper, so the same splitting logic can run over extract_rule_from_file()'s in-memory string (for --rule) without writing a temp file.

  • Tests in linters2/test/test_cc_lint.py, covering --mode × {topic, skill, rule, default} (the 3×4 = 12 combinations from task 2), each in two flavors against the same input file:

    • fake: mock claude_agent_sdk.ClaudeSDKClient (mirrors Test_PromptSequencer_execute in dev_scripts_helpers/ai/test/test_cc_lib.py) for session/stateless, and mock the subprocess call for one_shot; assert the right chunks/messages get built and dispatched.
    • live: pytest.importorskip("claude_agent_sdk") + class-level @pytest.mark.skip(reason="Run manually: makes a real Claude Agent SDK call and costs tokens"), cheapest model, assertIn-style loose checks — mirrors Test_PromptSequencer_execute_end_to_end in the same file. Also update cc_lint.README.md's "Invariants"/architecture section, which currently documents "exactly one action mode is active" — that invariant is being intentionally removed.