second-claude-code

English | 한국어

---

Second Claude Code — Work OS for Knowledge Work

You type one prompt. Researchers crawl 20+ sources. An analyst finds patterns. A writer drafts 3,000 words — and before you even see it, five reviewers are already tearing the draft apart. One checks the logic. Another attacks the weak points. A third fact-checks every number.

One prompt. Full cycle. No duct tape between plugins.

This isn't a coding assistant. It's a work OS — it runs the full knowledge-work cycle autonomously: Plan → Do → Check → Act. Research, analysis, writing, and quality assurance in a single automated loop.

_{Image created with SceneSteller}

Docs · 한국어 문서 · User Manual · 사용 매뉴얼 · Skill Guides · GitHub Issues · 한국어 README

---

System at a Glance

flowchart TB
    U[One user prompt] --> R[Prompt router]
    R --> O{External plugin wins?}
    O -->|yes| E[Installed plugin capability]
    O -->|no| P[Second Claude PDCA]
    E --> P
    P --> PLAN[Plan: research and analyze]
    PLAN --> DO[Do: write or build]
    DO --> CHECK[Check: review and verify]
    CHECK --> ACT[Act: refine, commit, or route back]
    ACT --> OUT[Finished artifact plus cycle memory]

Second Claude Code is the control loop. The v1.4.0 orchestrator sits in front of that loop and gives installed plugins the first shot when they are a stronger fit.

---

What's New in v1.5.0

unblock skill — zero-key adaptive fetch chain. The 16th skill closes the gap where blocked URLs (4xx, captcha, WAF, JS-heavy SPAs) silently degraded research output. Eevee researcher and the auto-router now invoke a 9-phase escalation pipeline before giving up.

Phase 0a → 0b → 0c → 0d → 1 → 2 → 3 → 4 → 5 → 6
public APIs → Jina → yt-dlp → keyword → curl variants →
TLS rotation → LightPanda → Playwright → free archives → optional paid

• 11 public-API routes — Reddit, HN, arXiv, Bluesky, GitHub, NPM, Stack Exchange, Wikipedia, Mastodon (any host), Lemmy (any host), oEmbed fallback
• TLS multi-rotation Phase 2 — chrome131 → safari17_0 → firefox133 with cookie carry between attempts
• Hidden API discovery Phase 4 — Playwright network intercept surfaces internal JSON endpoints when the rendered HTML fails validation
• Free archive cluster Phase 5 — Wayback + archive.today + AMP raced in parallel; RSS/Atom discovery; OG-tag rescue
• Operational hardening — SSRF guard rejects RFC1918/loopback/cloud-metadata hosts; schema_version + idempotency_key envelope; stagnation detection short-circuits to archive after 3 same-reason failures
• 397 tests (394 pass, 0 fail, 3 skipped)

See docs/RELEASE-v1.5.0.md for full release notes and verification.

Previously in v1.4.0...

What's New in v1.4.0

Cross-Plugin Orchestrator — Second Claude Code now discovers and commands all your installed Claude Code plugins automatically.

You type "코드 리뷰해줘." The prompt-detect hook spots the intent. The orchestrator scans your plugin ecosystem in real-time and finds coderabbit installed. Instead of running its own review, it auto-dispatches: Skill: coderabbit-code-review. You type "커밋해줘" — it finds commit-commands and routes /commit-commands:commit. A direct plugin request works too: "posthog event analysis" routes to Skill: posthog-exploring-autocapture-events when that plugin is installed.

No hardcoded registry. No manual plugin wiring. No configuration files. The orchestrator discovers plugins at runtime, maps each one to the appropriate PDCA phase (Plan/Do/Check/Act), and generates the exact Skill tool invocation string. Install a plugin → it appears. Uninstall → it disappears. Zero maintenance.

graph LR
    U[User Prompt] --> PD[prompt-detect hook]
    PD --> P[PDCA Router]
    P --> |check phase| OC[Orchestrator]
    OC --> |scan| PL{Plugin Ecosystem}
    PL --> CR[coderabbit<br/>code-review]
    PL --> CC[commit-commands<br/>commit]
    PL --> FD[frontend-design<br/>design]
    PL --> CX[codex<br/>review]
    PL --> AT[agent-teams<br/>team-review]
    OC --> |dispatch| SK[Skill: plugin-skill]
    SK --> |execute| RS[Result]