v0.7.14 — Scene-driven tool descriptions & integration improvements
What's Changed
Scene-Driven Tool Descriptions
All 12 MCP tool descriptions and 5 skill/command descriptions rewritten with "Use when:" trigger conditions. This significantly improves Claude Code's auto-invocation rate — tools are now selected based on matching scenarios, not just functionality keywords.
MCP Integration
- Decision rules: 7 intent→tool mapping rules in MCP instructions (e.g., "who calls X?" →
get_call_graph, not grep) get_ast_node: Newinclude_testsparameter, test callers filtered by default withtest_callers_hiddencount- JSON safety: Error responses now use
serde_jsoninstead of manual string escaping
Hook Improvements
- Pre-edit hook: Grep-based fallback detects containing function for body edits (previously only triggered on function signature lines)
- User-prompt hook: Backtick symbol extraction for mixed Chinese+code prompts (e.g., "修改 `parse_code` 函数")
CLI UX
show --refs/show --impactshort aliases (previously required--include-refs/--include-impact)- Help text updated to show short forms as primary
Breaking Changes
overview,trace,depsnow exit code 1 (was 0) when no results found. This is semantically correct but may affect external scripts.
Token Budget
Fixed plugin context cost: ~3,454 tokens (1.4% of 1M context) — unchanged from previous version.
Full Changelog: v0.7.13...v0.7.14
Full Changelog: v0.7.13...v0.7.14
Full Changelog: v0.7.13...v0.7.14