Skip to content

v0.2.4 — Corrupted MCP Arg Circuit-Breaker, Search Perf, Stress Benchmark & README Overhaul

Choose a tag to compare

@INK-Q INK-Q released this 21 Nov 04:20
· 19 commits to main since this release

Highlights

Fixed

  • Circuit-breaker for corrupted call_mcp_tool nested arguments across all 4 MCP tools (apply_edits, batch_search, read_slices, run_checks):
    • Detects when Antigravity's untyped call_mcp_tool argument parser collapses nested edit/query/slice/command objects into booleans or null (e.g., edits: [true], [false, true], [null], or hunks: [true]).
    • Immediately returns an actionable circuit-breaker message instructing the caller not to retry the same payload and to fall back to parallel built-in tools—eliminating multi-minute retry loops observed in live sessions.
    • Prevents uncaught TypeError crashes on null/primitive array entries and transparently accepts stringified JSON arrays.
  • Read-only git shell chaining guard: Rejects standalone shell operators (&&, ||, ;, |) in batch_search git queries with guidance to split into separate queries.

Improved

  • Search & fuzzy-path performance: Added a zero-allocation ASCII fast path in fuzzy-match.ts, switched hot-path filename sorting from ICU localeCompare to deterministic code-unit comparison, and replaced regex matchAll scanning of rg summary output with direct \0 index scanning.
  • Rules & benchmark diagnostics:
    • Updated rules/AGENTS.md Rule 3 (keep all images out of >150k-token main sessions to prevent prompt-cache invalidation), Rule 5 (corrupted-arg circuit-breaker fallback), and Rule 6 (swarm worker brief guidance).
    • Added agyCorruptedArgCalls, agyCorruptedArgWastedSec, agyCorruptedArgWastedOutTokens, and swarmWorkerSessions metrics to dist/benchmark.mjs.
  • Reproducible 12-scenario stress benchmark (npm run bench:stress): Covers 3,000–6,000 file monorepos, rg + pure-Node fallback, 5×5,000-line / 40-file outlines + ReDoS resilience, 50-file atomic & Phase-2 disk rollback / bestEffort / 2 MB noisy verify, 12-way run_checks + 5 MB log flood + grandchild process-tree kill, 400-request soak under a strict 128 MB V8 heap cap, 20-attack read-only git red-team checks, and all 14 outline formats + 4-byte emoji UTF-8 truncation safety.

Documentation

  • Reorganized English and Simplified Chinese READMEs (#4, thanks @GoodCoder666): Added badges, KPI summary table, bottleneck breakdown, hybrid execution Mermaid diagram, collapsible install/example/env sections, expanded FAQ, and Star History chart.

简体中文更新摘要

  • 全 4 工具接入 call_mcp_tool 嵌套参数序列化损坏熔断与容错:精准识别 edits: [true] / [false, true] / [null] 等底层序列化损坏,直接返回明确熔断提示引导立即回退并行原生工具(消除实测中反复重试白烧十余分钟的问题),同时修补 queries/outlines/hunks 收到 null/原始值可能崩溃的隐患,并兼容 JSON 字符串化数组参数。
  • 大规模搜索性能优化与只读 Git 链式拦截:fuzzy-match 新增零分配 ASCII 快速路径、热路径移除昂贵 ICU localeCompare、rg 汇总输出改为直接 \0 扫描,并拦截只读 git 查询中的独立 &&/||/;/| 链式符。
  • 规则、基准与 12 场景全链路压测:Rule 3 扩展至任意图片(防 >150k 长对话缓存清零)、Rule 5 补入熔断指引、Rule 6 优化 Swarm Worker 协作提示;benchmark 新增参数损坏耗时与 Swarm Worker 覆盖率指标;新增 npm run bench:stress 12 场景端到端压测套件。
  • README 全面改版(#4,感谢 @GoodCoder666):中英文 README 新增状态徽章、核心指标速览表、混合架构 Mermaid 图、折叠示例与更完整的 FAQ。

Full Changelog: v0.2.3...v0.2.4