popup-studio-ai/bkit-claude-code/cc-version-analysis

CC Version Analysis — Claude Code CLI 버전 영향 분석 워크플로우

CC CLI 버전 업그레이드 시 bkit plugin에 대한 영향을 체계적으로 조사, 분석하고

개선 기회를 도출하는 전문 워크플로우 스킬.

Overview

이 스킬은 CC CLI의 새 버전이 출시되었을 때 다음을 자동화합니다:

1. Phase 1 (Research): CC 변경사항 심층 조사
2. Phase 2 (Analyze): bkit 아키텍처 영향 분석
3. Phase 3 (Brainstorm): Plan Plus 브레인스토밍으로 개선안 도출
4. Phase 4 (Report): 종합 영향 분석 보고서 작성

Agent Team 구성:

• cc-version-researcher: CC 버전 변경사항 외부 조사
• bkit-impact-analyst: bkit 내부 아키텍처 영향 분석
• report-generator: 최종 보고서 생성

HARD-GATE

<HARD-GATE> Do NOT skip any phase. Each phase produces artifacts that feed into the next. Do NOT generate the final report without completing Research and Analysis phases. Do NOT implement any ENH items — this skill is analysis-only. All documents MUST be written in Korean (한국어).

NEW (v2.1.16 errata learning — added 2026-05-20): Do NOT advance from Phase 1 → Phase 2 without completing the Phase 1.5 Raw Source Verification Gate. The cc-version-researcher agent's quantitative claims (bullet count, issue counts, file counts) MUST be cross-checked against ≥2 raw sources (GitHub release tag URL + raw CHANGELOG.md URL) before being adopted into the report. The v2.1.145 cycle leaked an unverified bullet count of 21 (actual: 20) and an unverified agent count of 36 (actual: 34) because this gate did not exist. If raw and agent disagree, raw wins. </HARD-GATE>

Invocation

/cc-version-analysis                    # Auto-detect: installed vs latest
/cc-version-analysis 2.1.78 2.1.85     # Specific version range
/cc-version-analysis to 2.1.85         # From current installed to target

Process Flow

┌─────────────────────────────────────────────────────┐
│                CC Version Analysis                   │
├─────────────────────────────────────────────────────┤
│                                                      │
│  Phase 0: Setup & Version Detection                  │
│  ├── Detect installed CC version (claude --version)  │
│  ├── Determine target version (args or latest)       │
│  ├── Create Task tracking structure                  │
│  └── Load previous analysis from memory              │
│                                                      │
│  Phase 1: Research (cc-version-researcher agent)     │
│  ├── Official docs (code.claude.com)                 │
│  ├── GitHub (anthropics/claude-code)                 │
│  │   ├── Releases & changelog                        │
│  │   ├── Issues (open & recently closed)             │
│  │   ├── PRs (merged in version range)               │
│  │   └── Commits (significant changes)               │
│  ├── npm registry (@anthropic-ai/claude-code)        │
│  ├── Technical blogs & community                     │
│  └── Output: CC Change Report (structured)           │
│                                                      │
│  Phase 1.5: Raw Source Verification Gate (MANDATORY) │
│  ├── WebFetch https://github.com/anthropics/         │
│  │   claude-code/releases/tag/v{to}                  │
│  ├── WebFetch https://raw.githubusercontent.com/     │
│  │   anthropics/claude-code/main/CHANGELOG.md        │
│  ├── Cross-check agent's bullet count vs raw count  │
│  ├── If mismatch → raw wins, record errata           │
│  ├── Spot-check ≥3 random Fixed bullets verbatim     │
│  └── Block Phase 2 if any mismatch unresolved        │
│                                                      │
│  Phase 2: Analyze (bkit-impact-analyst agent)        │
│  ├── Map CC changes → bkit components                │
│  ├── Identify ENH opportunities                      │
│  ├── File impact matrix                              │
│  ├── Philosophy compliance check                     │
│  ├── Test impact assessment                          │
│  └── Output: bkit Impact Analysis (structured)       │
│                                                      │
│  Phase 3: Brainstorm (Plan Plus methodology)         │
│  ├── Intent discovery (핵심 목표/리스크/기회)         │
│  ├── Alternative exploration                         │
│  ├── YAGNI review (각 ENH 필요성 검증)               │
│  ├── Priority assignment (P0~P3)                     │
│  └── Output: Prioritized ENH roadmap                 │
│                                                      │
│  Phase 4: Report Generation                          │
│  ├── Merge all phase outputs                         │
│  ├── Generate from template                          │
│  │   (cc-version-analysis.template.md)               │
│  ├── Save to docs/04-report/features/                │
│  ├── Update MEMORY.md (version history)              │
│  └── Output: Final Impact Report (Korean)            │
│                                                      │
└─────────────────────────────────────────────────────┘

Phase Details

Phase 0: Setup & Version Detection

1. Detect installed CC version:
   $ claude --version
 
2. Determine target version:
   - If args provided: use specified versions
   - If no args: search for latest available version
 
3. Create Task structure:
   TaskCreate: "[CC-Version-Analysis] CC v{from} → v{to}"
     ├── Task: "Phase 1: CC 변경사항 조사"
     ├── Task: "Phase 2: bkit 영향 분석"
     ├── Task: "Phase 3: Plan Plus 브레인스토밍"
     └── Task: "Phase 4: 보고서 작성"
 
4. Load previous analysis context:
   - Read memory/cc_version_history_*.md
   - Read last ENH number from MEMORY.md
   - Read existing PDCA status from .bkit/state/pdca-status.json

Phase 1: Research (Agent: cc-version-researcher)

Input: from_version, to_version Output: Structured CC Change Report

Launch the cc-version-researcher agent with:

Research CC CLI changes from v{from} to v{to}.
Sources: official docs, GitHub (issues/PRs/releases), npm, blogs.
Categorize by: Breaking/Feature/Fix/Performance/SystemPrompt/Hook/Config.
Rate impact: HIGH/MEDIUM/LOW.
Flag bkit-relevant changes.
Output structured markdown tables.

Parallel research tasks (when using Agent Team):

• Task 1: GitHub releases + changelog
• Task 2: GitHub issues (open + recently closed)
• Task 3: Official docs changes
• Task 4: System prompt diff analysis

Phase 1.5: Raw Source Verification Gate (MANDATORY — added 2026-05-20)

Why: cc-version-researcher's output may paraphrase, summarize, or miscount. The v2.1.145 cycle revealed that a single WebFetch (via the model-processed release tag page) under-counted by 1 bullet (reported 6 Added, actual 7). A second fetch against raw CHANGELOG.md is required to catch this.

Protocol (main session, not delegated):

1. Fetch raw GitHub release tag page:

`` WebFetch https://github.com/anthropics/claude-code/releases/tag/v{to_version} ``

1. Fetch raw CHANGELOG.md (authoritative):

`` WebFetch https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md `` prompt: "Show the FULL v{to_version} section verbatim. List every bullet exactly as written, in order, under the original headings (Added/Fixed/ Improved/Breaking/etc). Do NOT summarize."

1. Compare to cc-version-researcher's reported counts:

• Bullet count (total / per-heading)
• Breaking count
• HIGH/MEDIUM/LOW split
• Key item presence (security fixes, OTEL, hook input changes)

1. If raw and agent disagree → raw wins. Record discrepancy in report

under §3.0 "Verification Notes" with both numbers.

1. Spot-check ≥3 random Fixed bullets — verify verbatim text in agent report

matches raw.

1. Block Phase 2 progression if any mismatch is unresolved.

Output: §3.0 Verification table appended to Phase 1 output:

Phase 1.6: Opaque Release Protocol (MANDATORY when triggered — ENH-420)

Trigger: the range's CHANGELOG carries ≤ 1 bullet, OR its bullets are non-specific ("Bug fixes and reliability improvements"). CC v2.1.226 was the first instance and will not be the last.

Why this exists: an opaque release makes CHANGELOG-based analysis worth exactly nothing, and the cycle either invents an answer or measures one. Cycle

#35 measured, and the measurement carried the whole report — but it took 15+ manual tool calls to re-derive a procedure nobody had written down, and it produced a false positive along the way (ERRATA-35-1) before catching it.

Procedure (main session, not delegated):

1. Run node scripts/cc-binary-equivalence.js <from> <to> (ENH-421). It reports

file size, sha256, Mach-O segment sizes, and exact grep -a -o -F counts for every hook-contract marker, in one pass per build.

1. Read the segment table before the marker table. Identical __TEXT /

__DATA_CONST / __DATA with a moving __BUN means a payload swap, not a native rebuild — a distinction the file size alone hides.

1. If no marker moved, the range is equivalent on the measured surfaces, and

that is the finding. State the surfaces; do not generalize past them.

1. Transfer rule: when two builds are equivalent, every judgement made about

the earlier one carries forward unchanged — compatibility, risk, tracked-issue state. Nothing needs re-evaluating, and saying so explicitly is more useful than re-deriving it.

1. "No change" is itself a reason RECOMMENDED_VERSION cannot rise. An

equivalent release resolves nothing, so a hold stays a hold.

Method rules, each of which cost a cycle:

• ERRATA-35-1: a strings set diff is NOT evidence. Trailing-byte artifacts

alone produced 1,315 added / 1,226 removed on two builds that turned out to be equivalent. Only exact counts on the raw file count.

• ERRATA-37-2: a needle beginning with - is parsed as an option by ugrep and

silently returns 0, which reads as absence. Use grep -a -o -F -e '<needle>'.

• ERRATA-37-6: a misspelled needle also returns a silent 0. Measure spelling

variants together — runInBackground was 0 in three builds while run_in_background was 49.

• ERRATA-33-6: the CHANGELOG paraphrases the implementation. A count of 0 for

a term lifted from the release notes does not mean the fix is absent.

• ERRATA-36-6: cmp -l over two ~300 MB binaries exceeds a two-minute budget,

and cmp -n / -i return wrong results on this artifact. Do not use them.

Binary provenance (ENH-422) — record these in the report and in the cycle memory, for every build examined. They scope the claim honestly and make cross-platform gaps trackable rather than invisible:

The last row is the one that matters. GIT_SHA moving while the bundle does not means upstream commits exist; the honest claim is "those commits did not change the shipped macOS x86_64 bundle", never "there were no commits".

Phase 2: Analyze (Agent: bkit-impact-analyst)

Input: Phase 1 CC Change Report PLUS Phase 1.5 verification table Output: bkit Impact Analysis

Pre-flight check before launching the analyst agent: the analyst will measure bkit architecture stats (agents/skills/hooks counts) via Bash. If the analyst proposes a numeric correction to memory (e.g. "agents 34 → 36"), the main session MUST re-run the measurement command independently before accepting it. See "Numeric Correction Protocol" in bkit-impact-analyst.md.