Skip to content

CHANGELOG.md

Latest commit

 

History

History
103 lines (78 loc) · 5.94 KB

CHANGELOG.md

File metadata and controls

103 lines (78 loc) · 5.94 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.4.1] - 2026-06-22

Fixed

  • kimi_query no longer breaks when the CLI is installed but not logged in. The previous routing preferred the CLI whenever the binary merely existed, ignoring auth state, so kimi_query errored against an unauthenticated kimi CLI instead of using the configured API. It now prefers the direct API whenever a key is configured (a contextless query needs no codebase), falling back to the CLI only when no key is present

Changed

  • kimi_status now reports the Kimi Code API configured state first (the backend that actually serves kimi_query/kimi_verify), then the CLI install/version/auth state — making clear that a "not authenticated" CLI does not affect the API tools

Documentation

  • Added "Two backends: API vs CLI" and "Guidelines for agents" sections to the READMEs, documenting which tool uses which backend and what each requires
  • Documented the ~/.local/bin PATH gotcha that causes command not found: kimi after install

[0.4.0] - 2026-06-18

Added

  • API mode (src/kimi-api.ts) — call the Kimi Code API directly, no Python CLI required. Reads the key from $KIMICODE_API_KEY or ~/.kimi/config.toml (section-aware, prefers the coding provider, supports ${VAR} interpolation) and sends the KimiCLI/1.0 User-Agent the endpoint requires
  • kimi_verify tool — Kimi as an independent third-party verifier. Cross-check a fix, diff, claim, or plan from a second model. Context-driven: callers must pass self-contained material since Kimi has no access to the session/repo. API-backed, so it works without the CLI
  • kimi_query API fallback — uses the direct API when the CLI isn't installed
  • KIMI_SHARE_DIR env var support in session-reader.ts and kimi-runner.ts — overrides the default ~/.kimi data root, matching the same env var the upstream Kimi CLI recognizes. Default behavior unchanged

Documentation

  • Brand sync with current Kimi Code docs (2026-06): replaced "Kimi K2.5" references in README.md, README_zh.md, and package.json with kimi-for-coding (the official model ID, auto-upgraded on the backend)
  • Pricing table removed: replaced the hardcoded Moderato/Allegretto/Allegro/Vivace table with a link to kimi.com/code; subscription tiers and quotas change too often to maintain in README
  • Added CLI Invocation Reference section to both READMEs clarifying the exact flags the MCP server uses (--print -p) and explicitly noting that kimi analyze subcommand does not exist

0.3.0 - 2026-03-04

Added (Phase 1: Context Caching)

  • Session Cache Manager — automatic caching of Kimi sessions per working directory
    • CacheManager class with LRU eviction, TTL expiration, and change detection
    • Git commit hash detection for automatic invalidation on code changes
    • Fallback to file mtime hashing for non-git repositories
    • Concurrent warmup deduplication (prevents duplicate session creation)
  • New MCP Tools
    • kimi_cache_status — view cache statistics, hit rates, and entry details
    • kimi_cache_invalidate — manual cache invalidation (single or all)
  • Enhanced kimi_analyze
    • use_cache parameter (default: true) — enable automatic session caching
    • Automatic cache hit/miss indicators in response
    • Automatic retry on invalid cached sessions
  • Enhanced kimi_runner
    • sessionId returned in KimiResult for cache tracking
    • extractSessionId() function parses session ID from Kimi output
  • Configuration
    • KIMI_CACHE_DEBUG environment variable for debug logging
    • Configurable maxSize (default: 10) and maxAgeMs (default: 30min)
  • Documentation
    • Comprehensive test requirements in TEST_REQUIREMENTS.md
    • 25+ test cases covering functional and non-functional requirements

Performance Improvements

  • Cache hit latency: ~10s vs ~60-120s for cache miss (6-12x faster)
  • Token cost reduction: Subsequent queries reuse cached context
  • Session reuse: Up to 256K tokens of context retained between calls

0.2.0 - 2026-03-03

Added

  • detail_level parameter for kimi_analyze and kimi_resume (summary/normal/detailed)
  • max_output_tokens parameter for all tools (default: 15000, hard truncation safety net)
  • include_thinking parameter (default: false — saves 10-30K tokens per call)
  • Structured output prompt engineering — Kimi returns concise markdown reports
  • Token Economics documentation section in both READMEs
  • Kimi Code reviewer documentation and workflow examples

Changed

  • Thinking blocks now excluded by default (previously always included)
  • Output truncated at clean markdown boundaries when exceeding budget
  • Tool descriptions updated to reflect budget-controlled output

0.1.0 - 2026-03-03

Added

  • Initial MCP server with 4 tools:
    • kimi_analyze — deep codebase analysis (architecture, audit, refactoring)
    • kimi_query — quick programming questions without codebase context
    • kimi_list_sessions — list existing Kimi sessions with metadata
    • kimi_resume — resume previous sessions (up to 256K token context)
  • Kimi CLI subprocess management with stream-json parsing
  • Session metadata reader for ~/.kimi/sessions/
  • Configurable timeouts (10 min for analysis, 2 min for queries)
  • Thinking mode support (--thinking flag)
  • Bilingual documentation (English + 繁體中文)
  • CONTRIBUTING.md with development workflow guide