v2.14.0

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

docs: prepare release v2.13.0

v2.13.0

GitHub 이슈 분석과 LLM 리뷰 코멘트 대응을 위한 두 개의 새로운 standalone 스킬을 추가하여 오픈소스 관리 워크플로우를 afc 파이프라인에 통합합니다.

Highlights

• /afc:issue — GitHub 이슈 URL/번호 하나로 제목, 본문, 첨부 이미지(멀티모달), 코멘트를 분석하고 코드베이스 관련 파일을 탐색합니다. 이슈를 Bug/Feature/Question/Enhancement로 분류하고 적절한 다음 스킬(/afc:debug, /afc:spec, /afc:auto)을 제안합니다.
• /afc:resolve — PR에 달린 CodeRabbit, GitHub Copilot, Codex 등 LLM bot의 리뷰 코멘트를 자동 수집하고 VALID/NOISE/DISCUSS로 분류합니다. VALID 항목은 자동 수정, DISCUSS 항목은 사용자 판단을 요청하며, 확인 후 일괄 커밋합니다.

What's Changed

Added

• /afc:issue skill — GitHub 이슈 심층 분석 및 구조화된 분석 문서 생성 (.claude/afc/issues/)
• /afc:resolve skill — LLM bot 리뷰 코멘트 수집, 분류(VALID/NOISE/DISCUSS), 자동 수정 및 커밋
• UserPromptSubmit 인텐트 라우팅에 issue/resolve 키워드 패턴 추가
• 4개의 새 ShellSpec 테스트 (영문/한글 키워드 라우팅)

Changed

• Spec 디렉토리에 날짜 접두사 추가

Breaking Changes

없음

Full Changelog: v2.12.0...v2.13.0

docs: prepare release v2.12.0

v2.12.0

Fix critical skill path resolution bug and separate init/setup responsibilities for reliability.

Highlights

• New /afc:setup skill — Dedicated command for managing the AFC routing block in ~/.claude/CLAUDE.md. Run once after install, again after updates. Idempotent.
• Fixed stale cache bug — Skills were referencing ${CLAUDE_PLUGIN_ROOT} which Claude Code doesn't substitute in SKILL.md files. Replaced with ${CLAUDE_SKILL_DIR}/../.. across all 14 skill files.
• Init is now project-local only — /afc:init no longer touches ~/.claude/CLAUDE.md. Cleaner separation following industry conventions (git init, terraform init).

What's Changed

Added

• /afc:setup skill for global ~/.claude/CLAUDE.md configuration

Changed

• Init/Setup separation: init = project-local, setup = global config
• ${CLAUDE_PLUGIN_ROOT} → ${CLAUDE_SKILL_DIR}/../.. in all SKILL.md files
• Doctor now prints version header and plugin root path
• Fix suggestions correctly route to /afc:setup vs /afc:init

Fixed

• Skills running from stale cached plugin versions
• Init trying to execute nonexistent afc-init.sh script
• Doctor example output mismatched actual script output

Breaking Changes

None. /afc:init still works for project setup. Global CLAUDE.md management is now available via /afc:setup.

Full Changelog: v2.11.0...v2.12.0

docs: prepare release v2.11.0

v2.11.0

Skill Advisor system that dynamically invokes auxiliary skills at phase boundaries, plus a project-wide shift from static keyword matching to LLM semantic evaluation across all 16+ skills.

Highlights

• Skill Advisor: The auto pipeline now has 5 checkpoints (A–E) at phase boundaries that evaluate whether auxiliary skills (ideate, consult, architect, security, analyze, test, qa, learner) would add value. Uses LLM semantic judgment with 1–5 scoring — not keyword counting. Budget-controlled (max 5 per pipeline run) with 3 execution modes: Transform, Enrich, Observe.

• Intent-based evaluation everywhere: Replaced 26 static patterns (keyword tables, hard-coded file count thresholds, arbitrary retry caps) across all skills with semantic, context-aware evaluation. Examples: consult domain routing now identifies expertise gaps instead of counting keyword hits; review orchestration selects complexity mode holistically instead of counting files.

• Cross-skill consistency audit: Comprehensive analysis resolved 16 inconsistencies including contradictory memory management approaches, missing data flow injection steps, stale numeric references, and terminology mismatches between auto pipeline and standalone skills.

What's Changed

Added

• Skill Advisor system with 5 phase-boundary checkpoints, budget control, and 3 execution modes
• Expert Agent Routing to 8 domain experts via explicit agent ID lookup
• Advisor state persistence to .afc-state.json
• Dynamic template pattern support in consistency check
• Learner safety guardrails in auto pipeline Checkpoint E

Changed

• Intent-based evaluation replaces static keyword matching across 16 skill files
• Memory management aligned between auto and standalone clean (soft guidelines + semantic assessment)
• Critic loop standardized: 5-pass cap, consistent terminology and doc paths
• Error retry: deterministic vs transient classification before retry

Fixed

• Ideate Transform extraction targets corrected
• Security agent unreachable MEMORY.md write instructions removed
• QA_FINDINGS and complexity-analysis.md now properly injected into downstream phases
• Stale numeric thresholds removed from review cross-boundary verification
• Phase gate step count, architect subagent quoting, implement terminology, CONTRIBUTING.md state file reference

Full Changelog: v2.10.0...v2.11.0

docs: prepare release v2.10.0

2.10.0

Migrates the plugin from the legacy commands/ directory to the modern skills/ directory structure, aligning with Claude Code's recommended plugin conventions. Each skill now has its own directory, enabling co-located supporting files like templates and reference docs.

Highlights

• Skills directory structure: commands/<name>.md → skills/<name>/SKILL.md — each skill is now an independent directory that can bundle supporting files
• Template extraction: Spec and plan templates extracted to standalone files (spec-template.md, plan-template.md), making skill prompts more concise and templates independently maintainable
• Single-consumer doc relocation: nfr-templates.md moved into skills/spec/ — docs used by only one skill now live alongside it

What's Changed

Changed

• Migrated all 28 skills from commands/*.md to skills/*/SKILL.md
• Extracted spec template (72 lines) and plan template (118 lines) to supporting files
• Relocated docs/nfr-templates.md to skills/spec/nfr-templates.md
• Renamed .claude/rules/commands.md to skills.md
• Updated all scripts, tests, docs, and schemas for skills structure
• Aligned terminology from "command" to "skill" across all documentation

Migration Notes

• Non-breaking: All slash commands (/afc:spec, /afc:plan, etc.) work identically
• Plugin cache refreshes automatically on upgrade — no manual action needed

Full Changelog: v2.9.1...v2.10.0

docs: prepare release v2.9.1

v2.9.1

Fix for commands that asked "would you like to proceed?" instead of executing immediately.

What's Changed

Fixed

• Command execution reliability: Commands using model: haiku (doctor, init, checkpoint, resume, clean, principles) would prompt for confirmation instead of executing. Root cause: haiku is too conservative with multi-step instructions. All commands now use model: sonnet.
• AFC routing table wording: Clarified user-only skill descriptions to prevent models from describing commands instead of running them.

Full Changelog: v2.9.0...v2.9.1

docs: prepare release v2.9.0

v2.9.0

SDD+TDD 방법론 통합과 다국어 지원 강화 — 테스트 우선 개발을 파이프라인 수준에서 강제하고, 한국어 사용자를 위한 자연어 라우팅을 추가했습니다.

Highlights

• TDD Guard: 구현 파일 작성 시 대응하는 테스트 파일이 없으면 자동 차단 — spec의 EARS 표기법, plan의 Test Strategy 섹션과 함께 SDD+TDD 워크플로우를 완성합니다
• Peer Mode (consult): /afc:consult에 동료 모드 추가 — 일방적 자문이 아닌 상호 대화형 브레인스토밍이 가능합니다
• 한국어 스킬 라우팅: "버그 수정해줘", "테스트 작성", "보안 점검" 등 한국어 자연어로 afc 스킬을 직접 호출할 수 있습니다

What's Changed

Added

• SDD+TDD methodology integration (EARS spec, Test Strategy, TDD Guard hook)
• Peer mode for /afc:consult — interactive thinking-together dialogue
• Reverse impact analysis in review command
• Auto-run /afc:spec from plan when spec.md is missing
• Project rules auto-generation in /afc:init (.claude/rules/afc-project.md)
• Korean language intent detection patterns
• Skill catalog fallback for unmatched prompts

Changed

• Intent detection hardened with catalog-based fallback
• Fork commands receive streamlined subagent context

Tests

• 396 examples, 0 failures (+63 new cases)

Full Changelog: v2.8.2...v2.9.0

v2.8.2

Changed

• Doctor fully script-driven: All 12 check categories now run in afc-doctor.sh — LLM only executes the script and prints output. Prevents model drift (auto-executing Fix commands, analyzing project code after health check)

Fixed

• grep -c with pipefail producing 0\n0 on multiline input — replaced with grep -q
• Fork-agent check crash when context: fork command has no agent: field

Full Changelog: v2.8.1...v2.8.2

v2.8.1

Fixed

• Plugin loading failure: Removed hooks field from plugin.json manifest — Claude Code auto-loads hooks/hooks.json, explicit reference causes "Duplicate hooks file" error

Full Changelog: v2.8.0...v2.8.1

v2.8.0

v2.8.0

This release introduces /afc:learner, a new pattern learning and rule promotion system that captures recurring user preferences and corrections during Claude Code sessions and promotes them to persistent project rules. It also includes targeted accuracy improvements across the implement, plan, review, and validate phases.

Highlights

• /afc:learner — When you keep telling Claude the same thing ("always use named exports", "never use var"), learner captures those patterns in the background and lets you review and promote them to .claude/rules/afc-learned.md as permanent project rules. Enable it once and corrections accumulate automatically.
• Smarter parallel mode selection — Implement phase now requires ≥3 genuinely independent tasks before spinning up parallel workers, reducing context-loss failures from unnecessary delegation.
• Plan path validation — File Change Map paths are now verified against the actual codebase at plan time, with auto-correction for near-matches and ambiguity prompts for multi-match cases.

What's Changed

Added

• /afc:learner command: pattern learning and rule promotion (review, status, reset, enable, disable)
• Background UserPromptSubmit hook for signal collection (async, non-blocking, 5s timeout)
• Korean language support for correction patterns (앞으로는, 항상, 절대, 쓰지마, 기억해)
• Secret redaction and queue cap (50 entries) in learner signal collection
• Learner queue notification at session start
• Learner health check category in /afc:doctor

Changed

• Implement parallel batch threshold raised to ≥3 tasks; config.gate empty-string guard added
• Plan File Change Map: auto-correct close matches, ambiguity prompt for multi-match, sibling directory boundary check
• Review: external dependency/test imports skipped in callee re-reads; standalone review loads spec context
• Expert consultation write scope explicitly limited to .claude/afc/ and .claude/agent-memory/
• Critic loop: stale evidence from prior passes is invalid unless re-verified in current pass
• Validate: loads context.md, verifies AC completeness before cross-checking
• Auto fast-path: SIDE_EFFECT_SAFETY criterion added at Step 3.6
• Spec research findings persist to .claude/afc/memory/research/ to survive compaction
• Triage risk classification considers change semantics, not just file type
• afc-auto-format.sh: format errors now emit stderr warning instead of silently discarding

Fixed

• Silent failure in async format hook when timeout occurs

Breaking Changes

None.

Full Changelog: v2.7.1...v2.8.0