v2.14.0
Massive quality overhaul: all 31 skills and 12 agents refactored to Anthropic official best practices, plus a new Context Management Harness for proactive 1M context window control.
Highlights
- Context Management Harness: Phase-boundary compact recommendations, context budget monitoring (50%/70%/90% warnings), PostCompact context recovery, and advisor result accumulation across phases. Designed around the principle that 1M context = control headroom, not capacity.
- Skills & Agents BP compliance: Every skill reduced to under 500 lines with progressive disclosure. Every agent now has maxTurns, HITL rules, and explicit tool restrictions. 15 skills use dynamic context pre-fetching. 15 reference files created for extracted content.
- Hook performance: Early-exit patterns eliminate 95%+ of unnecessary I/O on file writes. TaskCompleted keyword tightened to reduce false positive API calls by ~80-90%.
What's Changed
Added
- Context Management Harness (phase-boundary compact, budget monitor, PostCompact recovery, CLAUDE_ENV_FILE caching)
- StopFailure hook for rate_limit/auth/server error detection
- Advisor to context.md accumulation at Skill Advisor checkpoints A-E
- SendMessage resume pattern for dependent task chaining
- Skill authoring guide, Agent authoring guide, Context harness docs
- Orchestration modes shared reference (Sequential/Batch/Swarm)
- validate:plugin npm script, marketplace features array
- README Development Docs section
Changed
- All 31 skills refactored (auto: 1156 to 437, implement: 407 to 213, review: 343 to 197 lines)
- All 12 agents refactored (maxTurns, HITL, effort, Write policy)
- 8 expert agents: effort: medium for thinking token savings
Fixed
- Hook early-exit optimization (spec-guard, tdd-guard)
- TaskCompleted false positive reduction
- PostToolUseFailure made async, plugin schema agents field removed
- Error handling in afc-sync-cache.sh, CLAUDE.md docs listing updated
Full Changelog: v2.13.0...v2.14.0