Genasis

Famous OSS Claude Code Agent Collections — Field Survey

Reference doc for evaluating .claude/agents/*.md providers. Compiled 2026-05-05. Star counts and “last pushed” timestamps captured live from the GitHub REST API on this date — re-pull before quoting numbers in conversation.

1. Methodology

I weighted candidates by (a) raw popularity (GitHub stars + fork count via GET /repos/{o}/{r} on api.github.com), (b) commit recency (pushed_at within last 60 days = active), (c) presence in Anthropic’s official plugin marketplace (anthropics/claude-plugins-official marketplace.json), (d) inclusion in the de-facto curation lists (hesreallyhim/awesome-claude-code 42.6K★, hesreallyhim/a-list-of-claude-code-agents), and (e) actual file format — I fetched representative agent files from each repo and verified they ship valid .claude/agents/*.md with YAML frontmatter (name, description, optional tools, model). Repos that turned out to be skills frameworks, methodology guides, or CLAUDE.md-only patterns are flagged and demoted, since they are not drop-in replacements for ECC’s .claude/agents/. Discussion-quality signals (HN thread #44742081, Augment Code’s running ECC star-count series, daily.dev posts, dev.to writeups) were used as tiebreakers, not primary evidence.

2. Comparative table — top OSS Claude .claude/agents/*.md collections

Repo Stars Last commit Format Install method Strengths Weaknesses Best roles
affaan-m/everything-claude-code (ECC) 173,533 2026-05-03 Plain .md agents + plugin manifest + skills + hooks + rules git clone … && bash install.sh (also /plugin marketplace add affaan-m/everything-claude-code) 48 ships agents, strong language-reviewer matrix (typescript, python, rust, go, java, kotlin, c++, c#, dart, flutter), code-reviewer with 80%-confidence filtering, opensource-* utilities, harness-optimizer; cited as “community standard”; multilingual READMEs Big surface (997 tests, multi-runtime support); some Augment Code reviews call it overkill for solo devs; opinionated about workflow code-reviewer, planner, architect, language-specific reviewers, refactor, docs
obra/superpowers 178,678 2026-05-04 Skills, not agents (no .claude/agents/); ships methodology + skills like subagent-driven-development, using-git-worktrees, writing-plans Anthropic-official: /plugin install superpowers@claude-plugins-official Author Jesse Vincent; officially listed in Anthropic’s plugin marketplace; brainstorming → plan → TDD → subagent-driven-development pipeline; cross-harness (Codex, Gemini CLI, Cursor) Not an agent collection — if you’re looking to swap .claude/agents/*.md, this is the wrong shape. Pair it with one of the agent collections Workflow harness, TDD enforcement, plan→execute
bmad-code-org/BMAD-METHOD 46,362 2026-05-04 Skills-as-folders under src/bmm-skills/ (analysis → plan → solutioning → implementation); not raw .claude/agents/*.md npx bmad-method install Strong scrum/agile lifecycle: PRD → architecture → story → dev story → QA E2E; sprint-planning + retrospective skills Skill-graph format, not agent files; integration into existing .claude/agents/ requires translation Sprint planning, PRD writing, scrum/PM workflow
wshobson/agents 34,783 2026-05-02 Plugin marketplace that bundles .md agents under plugins/<plugin-name>/agents/ /plugin marketplace add wshobson/agents then /plugin install <plugin-name> 80 focused plugins, 185 agents, 153 skills, 100 commands; covers SEO, blockchain, fintech, quant, embedded; plugin-isolated context (only loads what you install) Some HN/issue feedback says agents are “too generic” (issue #24, discussion #42); install requires plugin per role frontend, backend, mobile (Flutter+iOS+RN), DB, K8s, security, full-stack-orchestrator
VoltAgent/awesome-claude-code-subagents 19,103 2026-04-20 Plain .md per category + plugin manifest claude plugin marketplace add VoltAgent/awesome-claude-code-subagents or clone + ./install-agents.sh 131+ agents, 10 clean categories (core dev / lang / infra / quality / data-AI / DX / specialized / business / meta-orch / research); explicit mobile-developer, flutter-expert, expo-react-native-expert, swift-expert, kotlin-specialist, ios-developer (under multi-platform) Less “production-tested” reputation than wshobson; some agents include heavy Communication Protocol JSON request blocks that pad context Language specialists, frontend, mobile, security-auditor, accessibility-tester
hesreallyhim/awesome-claude-code 42,554 2026-04-27 Curation list (meta), not a shipped agent set git clone to read The de-facto “where do I look first” index for Claude Code resources Reorg in progress per current README; doesn’t ship agents itself Discovery
hesreallyhim/a-list-of-claude-code-agents 1,239 2025-11-10 Community-submitted agent index (links + samples) git clone Pure agent curation (sister to awesome list) Slower-moving (last commit 2025-11) Discovery
vijaythecoder/awesome-claude-agents 4,234 2025-10-30 Plain .md agents organized as agents/{core,orchestrators,specialized,universal} git clone + cp Stack-specialized teams: Django / Laravel / Rails / React / Vue specialists; tech-lead-orchestrator, team-configurator, project-analyst Last commit 2025-10 — slowing; smaller surface Stack-specific full-team scaffolding
iannuttall/claude-agents 2,049 2025-07-25 Plain .md under agents/ git clone + cp Tight 7-agent set (prd-writer, frontend-designer, code-refactorer, security-auditor, content-writer, vibe-coding-coach, project-task-planner); often quoted in tutorials Stale (last push 2025-07); low coverage PRD writing, refactor — historic reference only
lst97/claude-code-sub-agents 1,558 2025-08-15 Plain .md git clone + cp Personal full-stack collection, well-loved on dev.to Stale (2025-08); single-maintainer Solo full-stack developer template
0xfurai/claude-code-subagents 884 2025-10-15 Plain .md cd ~/.claude && git clone … (auto user-global) Best dedicated mobile coverage of any pure-agent repo: ios-expert, swiftui-expert, android-expert, react-native-expert, flutter-expert, expo-expert, swift-expert, kotlin-expert, dart-expert Stale-ish (Oct 2025); low star count vs. surface area Mobile dev, language specialists
dl-ezo/claude-code-sub-agents 184 2025-07-30 Plain .md (35 agents) git clone + cp Strongest PM / requirements lifecycle: requirements-analyst, user-story-generator, business-process-analyst, requirements-validator, risk-manager, progress-tracker, stakeholder-communicator, uat-coordinator, training-change-manager Stale (July 2025); small audience Project planning, requirements, PM coordination
senaiverse/claude-code-reactnative-expo-agent-system 114 2025-10-05 Plain .md tiered (S/1/2/3) git clone RN-specific: Design Token Guardian, A11y Compliance Enforcer, Performance Budget Enforcer, Performance Prophet, Security Penetration Specialist Niche; small React Native + Expo only
keskinonur/claude-code-ios-dev-guide 655 2026-01 Guide doc, not agents (templates for you to author) n/a PRD-driven workflow + Swift 6 / SwiftUI / @Observable patterns; XcodeBuildMCP integration You build the agents; doesn’t ship them iOS Swift reference
anthropics/claude-plugins-official 18,566 2026-05-04 Marketplace registry (178 plugins) — Anthropic-curated Auto-loaded by Claude Code Vendor integrations (Atlassian, Auth0, AWS, Cloudflare, Cloudinary, Asana, Box…) + workflow plugins; only mobile-related entries: expo, kotlin-lsp, swift-lsp Mostly integrations, not generic dev-role agents; sparse coverage of frontend/backend/QA/architect roles Vendor SDK glue, LSP plumbing

Repos considered and dropped as either dead or low-signal: rahulvrane/awesome-claude-agents (334★, low fork ratio, supplanted by VoltAgent), rshah515/claude-code-subagents (49★, last push 2025-08), chusri/claude-code-agents (mirror of wshobson), mvandermeulen/wshobson-agents and destenson/wshobson--agents (forks).

3. De-facto standard per role — recommendation table

Rule used: if a role has a well-known dedicated agent file in a repo with >5K stars and commit activity in 2026, it is named as the standard. If no clear standard exists, the best available is named with the proviso.

Role Recommended source File path inside repo Why this one wins Runner-up
Frontend developer (web) wshobson/agents plugins/frontend-mobile-development/agents/frontend-developer.md React 19 / Next 15 / RSC explicit; Anthropic-marketplace-shaped; isolated plugin context VoltAgent categories/01-core-development/frontend-developer.md
Backend developer wshobson/agents plugins/backend-development/agents/backend-architect.md Splits backend-architect / event-sourcing / graphql / temporal-python-pro into one plugin; tdd-orchestrator co-located VoltAgent backend-developer.md + microservices-architect.md
Full-stack wshobson/agents plugins/full-stack-orchestration/agents/ (deployment-engineer, performance-engineer, security-auditor, test-automator) Multi-agent workflow orchestrator, designed for parallel execution VoltAgent categories/01-core-development/fullstack-developer.md
Mobile (cross-platform) VoltAgent categories/01-core-development/mobile-developer.md + categories/02-language-specialists/flutter-expert.md + expo-react-native-expert.md Largest single-repo mobile surface in a 19K★ active repo 0xfurai/claude-code-subagents agents/{ios,android,flutter,expo,react-native}-expert.md
Mobile iOS (Swift native) 0xfurai/claude-code-subagents agents/ios-expert.md, agents/swiftui-expert.md, agents/swift-expert.md Only repo shipping all three iOS-specific agents (UIKit, SwiftUI, Swift lang) as actual files VoltAgent swift-expert.md (no UIKit-specific)
Mobile Android (Kotlin/Compose) 0xfurai/claude-code-subagents agents/android-expert.md, agents/kotlin-expert.md Dedicated Android-Compose agent VoltAgent categories/02-language-specialists/kotlin-specialist.md
Mobile React Native / Expo senaiverse/claude-code-reactnative-expo-agent-system tiered agents under .claude/agents/ Only repo specifically targeting RN/Expo with accessibility + perf budget agents pre-wired VoltAgent expo-react-native-expert.md
QA / test automation VoltAgent categories/04-quality-security/test-automator.md, qa-expert.md, ui-ux-tester.md 6 quality/security agents in one folder; explicit accessibility-tester wshobson plugins/api-testing-observability + plugins/unit-testing
Security reviewer ECC agents/security-reviewer.md Cited as standout in awesome-claude-code; pairs with silent-failure-hunter, ad-security-reviewer VoltAgent security-auditor.md (also strong)
DevOps / infra VoltAgent categories/03-infrastructure/ — full set: cloud-architect, kubernetes-specialist, terraform-engineer, sre-engineer, devops-incident-responder, network-engineer, platform-engineer Most complete infra cluster (17 agents, separated cleanly) wshobson plugins/kubernetes-operations + plugins/cloud-infrastructure
Architect ECC agents/architect.md and agents/code-architect.md Two-layer architect (system + code); type-design-analyzer co-located VoltAgent microservices-architect.md + wshobson architect-review
Code reviewer ECC agents/code-reviewer.md Most-praised single agent in the ecosystem: 80%-confidence filtering + auto-consolidation + security-first ordering. Multiple writeups call it the gold standard. wshobson plugins/comprehensive-review/agents/code-reviewer.md (model: opus)
Planner / PM dl-ezo/claude-code-sub-agents project-planner.md, requirements-analyst.md, user-story-generator.md, progress-tracker.md, risk-manager.md, stakeholder-communicator.md Only repo with a complete requirements→PM lifecycle as .md agents. ECC’s planner.md is good but standalone BMAD-METHOD (skill-shaped, needs translation)
Performance wshobson/agents plugins/application-performance/agents/ + performance-engineer Application-tier focus + tied into orchestration ECC performance-optimizer.md
Refactor ECC agents/refactor-cleaner.md + code-simplifier.md Two complementary agents; pairs naturally with silent-failure-hunter VoltAgent refactoring-specialist.md
Docs ECC agents/doc-updater.md + docs-lookup.md doc-updater specifically wired for sync after code changes wshobson plugins/code-documentation
Accessibility ECC agents/a11y-architect.md Sole “a11y-architect” framing in any major repo VoltAgent accessibility-tester.md
Workflow / methodology obra/superpowers skills (not agents) — install via /plugin install superpowers@claude-plugins-official Anthropic-blessed; layers cleanly over any agent collection above BMAD-METHOD

4. Installation recipes

All targets assume Claude Code v2.0+ in repository /work/secusy. .claude/agents/ is the project-scoped location (committed); ~/.claude/agents/ is user-global (not committed). Pick deliberately.

# In Claude Code, register the marketplace once
/plugin marketplace add wshobson/agents

# Then install per-role plugins as needed
/plugin install backend-development@claude-code-workflows
/plugin install frontend-mobile-development@claude-code-workflows
/plugin install comprehensive-review@claude-code-workflows
/plugin install kubernetes-operations@claude-code-workflows
/plugin install full-stack-orchestration@claude-code-workflows

Equivalent raw clone (when .md files need to live in your repo for editing):

git clone --depth=1 https://github.com/wshobson/agents.git /tmp/wshobson
mkdir -p .claude/agents
cp /tmp/wshobson/plugins/comprehensive-review/agents/code-reviewer.md .claude/agents/
cp /tmp/wshobson/plugins/frontend-mobile-development/agents/*.md .claude/agents/
cp /tmp/wshobson/plugins/backend-development/agents/*.md .claude/agents/

License: MIT (Seth Hobson). Free to vendor.

B. VoltAgent/awesome-claude-code-subagents

# As a Claude Code plugin
claude plugin marketplace add VoltAgent/awesome-claude-code-subagents
claude plugin install voltagent-core-dev
claude plugin install voltagent-lang
claude plugin install voltagent-infra
claude plugin install voltagent-meta

# Or interactive installer (clone form)
git clone https://github.com/VoltAgent/awesome-claude-code-subagents.git
cd awesome-claude-code-subagents
./install-agents.sh   # installs to ~/.claude/agents/

# Or per-file copy into project
cp categories/01-core-development/frontend-developer.md /work/secusy/.claude/agents/
cp categories/02-language-specialists/flutter-expert.md /work/secusy/.claude/agents/
cp categories/02-language-specialists/swift-expert.md /work/secusy/.claude/agents/
cp categories/04-quality-security/{code-reviewer,security-auditor,test-automator,qa-expert,accessibility-tester}.md /work/secusy/.claude/agents/
cp categories/03-infrastructure/{kubernetes-specialist,terraform-engineer,sre-engineer}.md /work/secusy/.claude/agents/

License: MIT (VoltAgent, 2025). Free to vendor.

C. ECC (current incumbent)

# Plugin form (Anthropic-style)
/plugin marketplace add affaan-m/everything-claude-code

# Or direct install script (clones & wires up agents/skills/hooks/rules)
git clone https://github.com/affaan-m/everything-claude-code.git /tmp/ecc
cd /tmp/ecc && bash install.sh

# Cherry-pick the standout files into your repo
cp /tmp/ecc/agents/code-reviewer.md       /work/secusy/.claude/agents/
cp /tmp/ecc/agents/architect.md           /work/secusy/.claude/agents/
cp /tmp/ecc/agents/security-reviewer.md   /work/secusy/.claude/agents/
cp /tmp/ecc/agents/silent-failure-hunter.md /work/secusy/.claude/agents/
cp /tmp/ecc/agents/typescript-reviewer.md /work/secusy/.claude/agents/

License: MIT.

Note: obra/superpowers is installed as a skill plugin (/plugin install superpowers@claude-plugins-official) and does not populate .claude/agents/. It complements any of A/B/C above.

D. 0xfurai/claude-code-subagents (mobile gap-fill)

# User-global install (drops directly into ~/.claude/agents/)
cd ~/.claude && git clone https://github.com/0xfurai/claude-code-subagents.git

# Or selective copy into project
git clone --depth=1 https://github.com/0xfurai/claude-code-subagents.git /tmp/furai
cp /tmp/furai/agents/{ios-expert,swiftui-expert,swift-expert}.md /work/secusy/.claude/agents/
cp /tmp/furai/agents/{android-expert,kotlin-expert}.md /work/secusy/.claude/agents/
cp /tmp/furai/agents/{react-native-expert,expo-expert,flutter-expert,dart-expert}.md /work/secusy/.claude/agents/

E. dl-ezo/claude-code-sub-agents (PM lifecycle gap-fill)

git clone --depth=1 https://github.com/dl-ezo/claude-code-sub-agents.git /tmp/dlezo
cp /tmp/dlezo/{project-planner,requirements-analyst,user-story-generator,progress-tracker,risk-manager,stakeholder-communicator,business-process-analyst,requirements-validator,uat-coordinator}.md /work/secusy/.claude/agents/

5. Verdict — is ECC still the best?

Short answer: ECC is still the strongest single-repo foundation, but the optimal 2026 setup for secusy is a hybrid.

Reasoning:

  1. ECC wins on canonical quality for code-reviewer, security-reviewer, architect, planner, refactor-cleaner, silent-failure-hunter, a11y-architect, doc-updater, language reviewers (typescript-reviewer.md, python-reviewer.md, rust-reviewer.md, go-reviewer.md, flutter-reviewer.md, kotlin-reviewer.md, dart-reviewer.md). These are the most-cited individual agents in the ecosystem (Augment Code series, dev.to writeups, awesome-claude-code).
  2. ECC has gaps — by design it’s a harness, not a per-stack scaffold:
    • No frontend-developer / backend-developer implementer agents (only reviewers)
    • No mobile coverage at all (no iOS/Android/RN/Expo agent)
    • Thin DevOps/infra (only e2e-runner, no kubernetes / terraform / cloud-architect)
    • No GraphQL/WebSocket specialists
    • PM/requirements layer is one file (planner.md) vs. dl-ezo’s 8-file lifecycle
  3. wshobson/agents is the natural complement for implementers (frontend, backend, mobile, K8s, cloud, full-stack-orchestration). The plugin model means context isn’t bloated — install per role.
  4. VoltAgent fills the language-specialist + accessibility + DevOps holes when a raw .md cherry-pick (not a plugin) is wanted.
  5. 0xfurai is unique for native-iOS (ios-expert, swiftui-expert) and Android (android-expert) — no other major repo ships those as named files.
  6. dl-ezo is unique for full PM lifecycle (requirements-analyst, user-story-generator, risk-manager, progress-tracker, stakeholder-communicator, uat-coordinator).
  7. obra/superpowers is the workflow skin — install on top, regardless of which agents are picked. It’s the only .claude/agents/-adjacent thing officially endorsed by Anthropic via claude-plugins-official.
Layer Source What to copy
Reviewers + harness (keep) ECC code-reviewer.md, security-reviewer.md, architect.md, silent-failure-hunter.md, a11y-architect.md, doc-updater.md, tdd-guide.md, e2e-runner.md, refactor-cleaner.md, typescript-reviewer.md
Implementers (add) wshobson/agents (plugins) backend-development, frontend-mobile-development, full-stack-orchestration plugins via /plugin install
Mobile (add if relevant) 0xfurai ios-expert.md, swiftui-expert.md, android-expert.md, react-native-expert.md, expo-expert.md
Language specialists (gap-fill) VoltAgent react-specialist.md, nextjs-developer.md, swift-expert.md, kotlin-specialist.md, python-pro.md
PM lifecycle (replace ECC’s single planner.md) dl-ezo requirements-analyst.md, user-story-generator.md, progress-tracker.md, risk-manager.md
Workflow skin obra/superpowers /plugin install superpowers@claude-plugins-official

This is consistent with the project’s existing scrum + Plane workflow: ECC reviewers stay as the quality bar, wshobson plugins become the implementer pool agents called by pm during sprint dispatch, dl-ezo augments the planner, superpowers enforces the brainstorming → plan → TDD pipeline already encoded in the project’s CLAUDE.md.

ECC alone is no longer the best choice for a project that needs frontend/backend/mobile implementers in addition to reviewers — and secusy does, given the agent roster in .claude/agents/. Augmenting with wshobson + VoltAgent + dl-ezo + (optional) 0xfurai/superpowers closes every observable gap with no overlap.

6. Sources

GitHub repositories (verified live 2026-05-05)

Anthropic official docs

Industry / community writeups (used for reception, not raw fact)

Hacker News / Reddit threads

Single-agent file references used in §3

Discussion / criticism (used for “Weaknesses” column)

7. Curated .claude/agents/ plan — every dev role covered, single folder

Goal: a single .claude/agents/ folder that covers all common SW-development roles by cherry-picking from the repos surveyed above. No plugin marketplace dependency at runtime — just .md files committed to this repo. Sources stay traceable per-file (§9).

Naming convention: {role}.{origin-tag}.md where origin-tag ∈ {secusy, ecc, wshobson, voltagent, furai, dlezo, senaiverse, vijaythecoder}. This makes provenance grep-able (ls .claude/agents/ | cut -d. -f2 | sort -u) and prevents same-name collisions when two repos ship a code-reviewer.md. secusy is reserved for files genuinely written from scratch by this project’s team — files derived from ECC (even if heavily modified) are tagged ecc so attribution is correct.

Total: 47 cherry-picked files + 7 secusy-owned orchestrators = 54 agents in one folder.

7.1 Origin audit of files currently in .claude/agents/

Before designing the new layout, verify what each existing file’s true upstream is. This is the only honest basis for the §7.2 plan and §8 script. Origin determined by (a) presence in ECC’s published agents/ directory at the time of secusy’s genesis/create-agentic-team.sh STEP 2 clone, (b) git history of the file in this repo, (c) content inspection.

Current file True origin Status today Tag in new layout
architect.md ECC (modified by secusy 2026-04-13 to add Plane/Mattermost lifecycle) ECC-derived architect.ecc.md
backend.md secusy-custom (no ECC counterpart — ECC only ships reviewers, not implementers) secusy-original backend.md
build-error-resolver.md ECC unmodified ECC build-error-resolver.ecc.md
chief-of-staff.md ECC unmodified ECC; off-scope for team scrum dropped
code-reviewer.md ECC (modified 2026-04-13) ECC-derived code-reviewer.ecc.md
database-reviewer.md ECC unmodified ECC database-reviewer.ecc.md
designer.md secusy-custom (Gumroad design system orchestrator — no ECC counterpart) secusy-original designer.md
devops.md secusy-custom (Plane/CI orchestrator — no ECC counterpart) secusy-original devops.md
docs-lookup.md ECC unmodified ECC docs-lookup.ecc.md
doc-updater.md ECC unmodified ECC doc-updater.ecc.md
e2e-runner.md ECC unmodified ECC e2e-runner.ecc.md
frontend.md secusy-custom (Plane/Mattermost orchestrator — no ECC counterpart) secusy-original frontend.md
harness-optimizer.md ECC unmodified ECC harness-optimizer.ecc.md
loop-operator.md ECC unmodified ECC; off-scope dropped
performance-optimizer.md ECC unmodified ECC performance-optimizer.ecc.md
planner.md ECC (modified 2026-04-12 to reference Plane) ECC-derived planner.ecc.md
pm.md secusy-custom (Sprint/Plane Cycle orchestrator — no ECC counterpart) secusy-original pm.md
qa.md secusy-custom (E2E ownership matrix — no ECC counterpart, distinct from ECC’s e2e-runner) secusy-original qa.md
refactor-cleaner.md ECC unmodified ECC refactor-cleaner.ecc.md
security.md secusy-custom (security domain ticket owner, distinct from ECC’s security-reviewer) secusy-original security.md
security-reviewer.md ECC unmodified ECC security-reviewer.ecc.md
tdd-guide.md ECC unmodified ECC tdd-guide.ecc.md
typescript-reviewer.md ECC unmodified ECC typescript-reviewer.ecc.md

Summary: of 23 current files, 16 are ECC-origin (14 unmodified + 2 modified) and 7 are secusy-original (the orchestrators). The new §7.2 inventory tags them accordingly.

7.2 Full inventory — every file in the planned .claude/agents/

Each row carries a clickable upstream URL pointing to the exact upstream blob. License column is MIT for every row (audited live in §10).

# Category File in .claude/agents/ Origin Upstream URL (full path) Lic.
1 Orchestrator pm.md secusy (no upstream — written by secusy team) —
2 Orchestrator frontend.md secusy (no upstream — written by secusy team) —
3 Orchestrator backend.md secusy (no upstream — written by secusy team) —
4 Orchestrator qa.md secusy (no upstream — written by secusy team) —
5 Orchestrator security.md secusy (no upstream — written by secusy team) —
6 Orchestrator devops.md secusy (no upstream — written by secusy team) —
7 Orchestrator designer.md secusy (no upstream — written by secusy team) —
8 Architect architect.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/architect.md MIT
9 Architect code-architect.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/code-architect.md MIT
10 Architect architect-review.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/plugins/comprehensive-review/agents/architect-review.md MIT
11 Architect microservices-architect.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/01-core-development/microservices-architect.md MIT
12 Code review code-reviewer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/code-reviewer.md MIT
13 Code review code-reviewer.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/plugins/comprehensive-review/agents/code-reviewer.md MIT
14 Code review silent-failure-hunter.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/silent-failure-hunter.md MIT
15 Code review typescript-reviewer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/typescript-reviewer.md MIT
16 Code review python-reviewer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/python-reviewer.md MIT
17 Code review rust-reviewer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/rust-reviewer.md MIT
18 Code review go-reviewer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/go-reviewer.md MIT
19 Code review kotlin-reviewer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/kotlin-reviewer.md MIT
20 Code review flutter-reviewer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/flutter-reviewer.md MIT
21 Security security-reviewer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/security-reviewer.md MIT
22 Security security-auditor.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/04-quality-security/security-auditor.md MIT
23 Frontend (web) frontend-developer.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/plugins/frontend-mobile-development/agents/frontend-developer.md MIT
24 Frontend (web) react-specialist.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/02-language-specialists/react-specialist.md MIT
25 Frontend (web) nextjs-developer.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/02-language-specialists/nextjs-developer.md MIT
26 Backend backend-architect.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/plugins/backend-development/agents/backend-architect.md MIT
27 Backend graphql-architect.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/plugins/backend-development/agents/graphql-architect.md MIT
28 Backend python-pro.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/02-language-specialists/python-pro.md MIT
29 Backend nodejs-developer.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/02-language-specialists/nodejs-developer.md MIT
30 Full-stack fullstack-developer.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/01-core-development/fullstack-developer.md MIT
31 Full-stack deployment-engineer.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/plugins/full-stack-orchestration/agents/deployment-engineer.md MIT
32 Mobile X-plat mobile-developer.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/01-core-development/mobile-developer.md MIT
33 Mobile X-plat flutter-expert.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/02-language-specialists/flutter-expert.md MIT
34 Mobile X-plat expo-react-native-expert.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/02-language-specialists/expo-react-native-expert.md MIT
35 iOS native ios-expert.furai.md 0xfurai https://github.com/0xfurai/claude-code-subagents/blob/main/agents/ios-expert.md MIT
36 iOS native swiftui-expert.furai.md 0xfurai https://github.com/0xfurai/claude-code-subagents/blob/main/agents/swiftui-expert.md MIT
37 iOS native swift-expert.furai.md 0xfurai https://github.com/0xfurai/claude-code-subagents/blob/main/agents/swift-expert.md MIT
38 Android native android-expert.furai.md 0xfurai https://github.com/0xfurai/claude-code-subagents/blob/main/agents/android-expert.md MIT
39 Android native kotlin-expert.furai.md 0xfurai https://github.com/0xfurai/claude-code-subagents/blob/main/agents/kotlin-expert.md MIT
40 DB / data database-optimizer.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/plugins/comprehensive-review/agents/database-optimizer.md MIT
41 DB / data database-reviewer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/database-reviewer.md MIT
42 DB / data data-engineer.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/data-engineer.md MIT
43 DevOps / infra cloud-architect.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/03-infrastructure/cloud-architect.md MIT
44 DevOps / infra kubernetes-specialist.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/03-infrastructure/kubernetes-specialist.md MIT
45 DevOps / infra terraform-engineer.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/03-infrastructure/terraform-engineer.md MIT
46 DevOps / infra sre-engineer.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/03-infrastructure/sre-engineer.md MIT
47 DevOps / infra devops-incident-responder.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/03-infrastructure/devops-incident-responder.md MIT
48 QA / test test-automator.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/04-quality-security/test-automator.md MIT
49 QA / test qa-expert.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/04-quality-security/qa-expert.md MIT
50 QA / test ui-ux-tester.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/04-quality-security/ui-ux-tester.md MIT
51 QA / test accessibility-tester.voltagent.md VoltAgent https://github.com/VoltAgent/awesome-claude-code-subagents/blob/main/categories/04-quality-security/accessibility-tester.md MIT
52 QA / test e2e-runner.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/e2e-runner.md MIT
53 QA / test tdd-guide.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/tdd-guide.md MIT
54 Performance performance-engineer.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/plugins/full-stack-orchestration/agents/performance-engineer.md MIT
55 Performance performance-optimizer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/performance-optimizer.md MIT
56 Refactor refactor-cleaner.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/refactor-cleaner.md MIT
57 Refactor code-simplifier.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/code-simplifier.md MIT
58 Docs doc-updater.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/doc-updater.md MIT
59 Docs docs-lookup.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/docs-lookup.md MIT
60 Docs docs-architect.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/docs-architect.md MIT
61 Accessibility a11y-architect.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/a11y-architect.md MIT
62 Planner / PM planner.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/planner.md MIT
63 Planner / PM requirements-analyst.dlezo.md dl-ezo https://github.com/dl-ezo/claude-code-sub-agents/blob/main/requirements-analyst.md MIT⚠
64 Planner / PM user-story-generator.dlezo.md dl-ezo https://github.com/dl-ezo/claude-code-sub-agents/blob/main/user-story-generator.md MIT⚠
65 Planner / PM progress-tracker.dlezo.md dl-ezo https://github.com/dl-ezo/claude-code-sub-agents/blob/main/progress-tracker.md MIT⚠
66 Planner / PM risk-manager.dlezo.md dl-ezo https://github.com/dl-ezo/claude-code-sub-agents/blob/main/risk-manager.md MIT⚠
67 Planner / PM stakeholder-communicator.dlezo.md dl-ezo https://github.com/dl-ezo/claude-code-sub-agents/blob/main/stakeholder-communicator.md MIT⚠
68 Planner / PM uat-coordinator.dlezo.md dl-ezo https://github.com/dl-ezo/claude-code-sub-agents/blob/main/uat-coordinator.md MIT⚠
69 Build / utility build-error-resolver.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/build-error-resolver.md MIT
70 Build / utility harness-optimizer.ecc.md ECC https://github.com/affaan-m/everything-claude-code/blob/main/agents/harness-optimizer.md MIT
71 Build / utility error-detective.wshobson.md wshobson https://github.com/wshobson/agents/blob/main/error-detective.md MIT

Footnote on MIT⚠ (rows 63-68): dl-ezo declares MIT in its README but has no LICENSE file in the repo. Treated as MIT for §10.2 compatibility, mitigated per §10.4.

Origin breakdown (54 imported + 7 secusy-original = 61 unique* files / 71 rows):

*Numbering went 1–71 to keep distinct rows; some upstream sources contribute multiple files.

7.3 What changes vs. today

7.4 Why this composition

  1. ECC stays the review/quality bar (every reviewer agent is ECC’s). Empirically the most-praised single-repo reviewer matrix in the ecosystem (§5).
  2. wshobson supplies implementer + cross-cutting orchestration (frontend-developer, backend-architect, deployment-engineer, performance-engineer, error-detective, docs-architect, data-engineer). Plugin source structure preserved as bare files.
  3. VoltAgent fills language specialists, infra cluster, and quality testers that ECC and wshobson are thin on (kotlin/swift specialists exist in voltagent and furai both — voltagent picked for cross-platform, furai for native-only).
  4. 0xfurai is the only repo with swiftui-expert + ios-expert + android-expert as named files — used solely for mobile-native gap.
  5. dl-ezo’s PM lifecycle agents replace the gap left by ECC’s single planner.md for projects with real requirements/UAT phases.
  6. senaiverse and vijaythecoder are excluded from default cherry-pick — niche enough that they should be added per-project, not in baseline.
  7. obra/superpowers is NOT in .claude/agents/ — it’s installed as a skill plugin. Listed in §8 as a separate post-install step.
  8. secusy-original tag is reserved for the 7 orchestrators (pm, frontend, backend, qa, security, devops, designer). Any other file currently sitting under bare names in .claude/agents/ originated from ECC and is correctly retagged in §7.1.

8. Cherry-pick install shell script

Save as genesis/install-curated-agents.sh. Idempotent — re-running cleans .claude/agents/_imported/ and re-pulls.

#!/usr/bin/env bash
# install-curated-agents.sh — populate .claude/agents/ from the curated §7 inventory.
# License: see §10. Each imported file retains its upstream `Source-*:` header (§9).
# Re-run: safe. Removes only files under .claude/agents/_imported/ marker dir.

set -euo pipefail

# ─── 0. Guards ──────────────────────────────────────────────────────────────
[[ -d ".claude/agents" ]] || { echo "Run from repo root with .claude/agents/"; exit 1; }
command -v git >/dev/null || { echo "git required"; exit 1; }
command -v rsync >/dev/null || { echo "rsync required"; exit 1; }

TMP="$(mktemp -d)"
trap 'rm -rf "$TMP"' EXIT
TARGET=".claude/agents"
MARK="$TARGET/.imported-manifest"
mkdir -p "$TARGET"

# ─── 1. Clean prior import (orchestrators kept) ─────────────────────────────
# Anything tagged with one of our origin tags = previously imported, safe to delete
echo "[1/6] Cleaning prior import…"
find "$TARGET" -maxdepth 1 -type f \
  \( -name '*.ecc.md' -o -name '*.wshobson.md' -o -name '*.voltagent.md' \
     -o -name '*.furai.md' -o -name '*.dlezo.md' -o -name '*.senaiverse.md' \
     -o -name '*.vijaythecoder.md' \) -delete
rm -f "$MARK"

# ─── 2. Clone upstreams (shallow) ───────────────────────────────────────────
echo "[2/6] Cloning upstream repos (shallow)…"
git clone --depth=1 --quiet https://github.com/affaan-m/everything-claude-code.git    "$TMP/ecc"        &
git clone --depth=1 --quiet https://github.com/wshobson/agents.git                    "$TMP/wshobson"   &
git clone --depth=1 --quiet https://github.com/VoltAgent/awesome-claude-code-subagents.git "$TMP/voltagent" &
git clone --depth=1 --quiet https://github.com/0xfurai/claude-code-subagents.git      "$TMP/furai"      &
git clone --depth=1 --quiet https://github.com/dl-ezo/claude-code-sub-agents.git      "$TMP/dlezo"      &
wait

# ─── 3. Helper: copy + inject provenance header ─────────────────────────────
# Adds an HTML-comment provenance block at the very top of each .md.
# Claude Code's frontmatter parser ignores HTML comments before the YAML block,
# so this is a runtime no-op — but on GitHub it makes origin clickable.
import_file () {
  local src_repo="$1" src_path="$2" src_tag="$3" dst_name="$4" src_license="${5:-MIT}"
  local src_full="$TMP/$src_tag/$src_path"
  local dst_full="$TARGET/$dst_name"
  [[ -f "$src_full" ]] || { echo "  MISSING: $src_repo : $src_path"; return 1; }

  local commit_sha
  commit_sha="$(git -C "$TMP/$src_tag" rev-parse HEAD)"
  local source_url="https://github.com/$src_repo/blob/$commit_sha/$src_path"
  local source_url_main="https://github.com/$src_repo/blob/main/$src_path"
  local imported_at
  imported_at="$(date -u +%Y-%m-%dT%H:%M:%SZ)"

  {
    cat <<EOF
<!--
Source-Repo: $src_repo
Source-Path: $src_path
Source-URL: $source_url_main
Source-Pinned-URL: $source_url
Source-License: $src_license
Source-Commit: $commit_sha
Imported-At: $imported_at
-->
EOF
    cat "$src_full"
  } > "$dst_full"

  printf '%s\t%s\t%s\t%s\t%s\t%s\n' \
    "$dst_name" "$src_repo" "$src_path" "$commit_sha" "$src_license" "$source_url_main" \
    >> "$MARK"
}

echo "[3/6] Importing curated files…"
echo "# .claude/agents import manifest — generated $(date -u +%Y-%m-%dT%H:%M:%SZ)" > "$MARK"
printf '# columns: imported-name\tsource-repo\tsource-path\tcommit-sha\tlicense\tsource-url\n' >> "$MARK"

# ── ECC ────────────────────────────────────────────────────────────────────
import_file affaan-m/everything-claude-code agents/architect.md              ecc architect.ecc.md
import_file affaan-m/everything-claude-code agents/code-architect.md         ecc code-architect.ecc.md
import_file affaan-m/everything-claude-code agents/code-reviewer.md          ecc code-reviewer.ecc.md
import_file affaan-m/everything-claude-code agents/silent-failure-hunter.md  ecc silent-failure-hunter.ecc.md
import_file affaan-m/everything-claude-code agents/security-reviewer.md      ecc security-reviewer.ecc.md
import_file affaan-m/everything-claude-code agents/typescript-reviewer.md    ecc typescript-reviewer.ecc.md
import_file affaan-m/everything-claude-code agents/python-reviewer.md        ecc python-reviewer.ecc.md
import_file affaan-m/everything-claude-code agents/rust-reviewer.md          ecc rust-reviewer.ecc.md
import_file affaan-m/everything-claude-code agents/go-reviewer.md            ecc go-reviewer.ecc.md
import_file affaan-m/everything-claude-code agents/kotlin-reviewer.md        ecc kotlin-reviewer.ecc.md
import_file affaan-m/everything-claude-code agents/flutter-reviewer.md       ecc flutter-reviewer.ecc.md
import_file affaan-m/everything-claude-code agents/database-reviewer.md      ecc database-reviewer.ecc.md
import_file affaan-m/everything-claude-code agents/refactor-cleaner.md       ecc refactor-cleaner.ecc.md
import_file affaan-m/everything-claude-code agents/code-simplifier.md        ecc code-simplifier.ecc.md
import_file affaan-m/everything-claude-code agents/performance-optimizer.md  ecc performance-optimizer.ecc.md
import_file affaan-m/everything-claude-code agents/doc-updater.md            ecc doc-updater.ecc.md
import_file affaan-m/everything-claude-code agents/docs-lookup.md            ecc docs-lookup.ecc.md
import_file affaan-m/everything-claude-code agents/a11y-architect.md         ecc a11y-architect.ecc.md
import_file affaan-m/everything-claude-code agents/planner.md                ecc planner.ecc.md
import_file affaan-m/everything-claude-code agents/tdd-guide.md              ecc tdd-guide.ecc.md
import_file affaan-m/everything-claude-code agents/e2e-runner.md             ecc e2e-runner.ecc.md
import_file affaan-m/everything-claude-code agents/build-error-resolver.md   ecc build-error-resolver.ecc.md
import_file affaan-m/everything-claude-code agents/harness-optimizer.md      ecc harness-optimizer.ecc.md

# ── wshobson ───────────────────────────────────────────────────────────────
import_file wshobson/agents plugins/comprehensive-review/agents/code-reviewer.md       wshobson code-reviewer.wshobson.md
import_file wshobson/agents plugins/comprehensive-review/agents/architect-review.md    wshobson architect-review.wshobson.md
import_file wshobson/agents plugins/comprehensive-review/agents/database-optimizer.md  wshobson database-optimizer.wshobson.md
import_file wshobson/agents plugins/frontend-mobile-development/agents/frontend-developer.md wshobson frontend-developer.wshobson.md
import_file wshobson/agents plugins/backend-development/agents/backend-architect.md    wshobson backend-architect.wshobson.md
import_file wshobson/agents plugins/backend-development/agents/graphql-architect.md    wshobson graphql-architect.wshobson.md
import_file wshobson/agents plugins/full-stack-orchestration/agents/deployment-engineer.md  wshobson deployment-engineer.wshobson.md
import_file wshobson/agents plugins/full-stack-orchestration/agents/performance-engineer.md wshobson performance-engineer.wshobson.md
import_file wshobson/agents data-engineer.md         wshobson data-engineer.wshobson.md
import_file wshobson/agents docs-architect.md        wshobson docs-architect.wshobson.md
import_file wshobson/agents error-detective.md       wshobson error-detective.wshobson.md

# ── VoltAgent ──────────────────────────────────────────────────────────────
import_file VoltAgent/awesome-claude-code-subagents categories/01-core-development/microservices-architect.md voltagent microservices-architect.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/01-core-development/fullstack-developer.md     voltagent fullstack-developer.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/01-core-development/mobile-developer.md        voltagent mobile-developer.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/02-language-specialists/react-specialist.md    voltagent react-specialist.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/02-language-specialists/nextjs-developer.md    voltagent nextjs-developer.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/02-language-specialists/python-pro.md          voltagent python-pro.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/02-language-specialists/nodejs-developer.md    voltagent nodejs-developer.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/02-language-specialists/flutter-expert.md      voltagent flutter-expert.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/02-language-specialists/expo-react-native-expert.md voltagent expo-react-native-expert.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/03-infrastructure/cloud-architect.md           voltagent cloud-architect.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/03-infrastructure/kubernetes-specialist.md     voltagent kubernetes-specialist.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/03-infrastructure/terraform-engineer.md        voltagent terraform-engineer.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/03-infrastructure/sre-engineer.md              voltagent sre-engineer.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/03-infrastructure/devops-incident-responder.md voltagent devops-incident-responder.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/04-quality-security/test-automator.md          voltagent test-automator.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/04-quality-security/qa-expert.md               voltagent qa-expert.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/04-quality-security/ui-ux-tester.md            voltagent ui-ux-tester.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/04-quality-security/accessibility-tester.md    voltagent accessibility-tester.voltagent.md
import_file VoltAgent/awesome-claude-code-subagents categories/04-quality-security/security-auditor.md        voltagent security-auditor.voltagent.md

# ── 0xfurai (mobile native) ────────────────────────────────────────────────
import_file 0xfurai/claude-code-subagents agents/ios-expert.md         furai ios-expert.furai.md
import_file 0xfurai/claude-code-subagents agents/swiftui-expert.md     furai swiftui-expert.furai.md
import_file 0xfurai/claude-code-subagents agents/swift-expert.md       furai swift-expert.furai.md
import_file 0xfurai/claude-code-subagents agents/android-expert.md     furai android-expert.furai.md
import_file 0xfurai/claude-code-subagents agents/kotlin-expert.md      furai kotlin-expert.furai.md

# ── dl-ezo (PM lifecycle) — see §10.4: MIT declared in README only, no LICENSE file
DLEZO_LICENSE="MIT (declared in upstream README; no LICENSE file — see _licenses/LICENSE.dl-ezo-claude-code-sub-agents.NOTE)"
import_file dl-ezo/claude-code-sub-agents requirements-analyst.md       dlezo requirements-analyst.dlezo.md     "$DLEZO_LICENSE"
import_file dl-ezo/claude-code-sub-agents user-story-generator.md       dlezo user-story-generator.dlezo.md     "$DLEZO_LICENSE"
import_file dl-ezo/claude-code-sub-agents progress-tracker.md           dlezo progress-tracker.dlezo.md         "$DLEZO_LICENSE"
import_file dl-ezo/claude-code-sub-agents risk-manager.md               dlezo risk-manager.dlezo.md             "$DLEZO_LICENSE"
import_file dl-ezo/claude-code-sub-agents stakeholder-communicator.md   dlezo stakeholder-communicator.dlezo.md "$DLEZO_LICENSE"
import_file dl-ezo/claude-code-sub-agents uat-coordinator.md            dlezo uat-coordinator.dlezo.md          "$DLEZO_LICENSE"

# ─── 4. Vendor LICENSE files ────────────────────────────────────────────────
echo "[4/6] Vendoring upstream LICENSE files into .claude/agents/_licenses/"
mkdir -p "$TARGET/_licenses"
cp "$TMP/ecc/LICENSE"        "$TARGET/_licenses/LICENSE.affaan-m-everything-claude-code"
cp "$TMP/wshobson/LICENSE"   "$TARGET/_licenses/LICENSE.wshobson-agents"
cp "$TMP/voltagent/LICENSE"  "$TARGET/_licenses/LICENSE.VoltAgent-awesome-claude-code-subagents"
cp "$TMP/furai/LICENSE"      "$TARGET/_licenses/LICENSE.0xfurai-claude-code-subagents"
# dl-ezo: README declares MIT but no LICENSE file — vendor the README excerpt instead
{
  echo "# License notice for dl-ezo/claude-code-sub-agents"
  echo
  echo "Upstream repository does not ship a LICENSE file."
  echo "README.md declares (verbatim):"
  echo
  echo "    ## 📄 License"
  echo "    MIT License - feel free to use, modify, and distribute these agents for any purpose."
  echo
  echo "Source: https://github.com/dl-ezo/claude-code-sub-agents/blob/main/README.md"
  echo "Captured: $(date -u +%Y-%m-%dT%H:%M:%SZ)"
} > "$TARGET/_licenses/LICENSE.dl-ezo-claude-code-sub-agents.NOTE"

# ─── 5. Print manifest summary ─────────────────────────────────────────────
echo "[5/6] Manifest:"
column -t -s $'\t' "$MARK" | head -10
echo "  …(see $MARK for full list)"

# ─── 6. Optional: install superpowers as skill plugin ──────────────────────
echo "[6/6] To install obra/superpowers as a workflow skill plugin, in Claude Code run:"
echo "      /plugin marketplace add anthropics/claude-plugins-official"
echo "      /plugin install superpowers@claude-plugins-official"
echo
echo "Done. $(/bin/ls "$TARGET" | wc -l) files in $TARGET/"

Make executable and run:

chmod +x genesis/install-curated-agents.sh
./genesis/install-curated-agents.sh

Re-running pulls the latest commits and rewrites the imported files. Files that begin with the Source-Repo: HTML comment are treated as “imported” and removable on next run; orchestrators (no header) are preserved.

9. Attribution rules — how to cite each imported file

Goal: when this .claude/agents/ is published as a public GitHub repo, each imported file’s origin is unambiguous and auditable, and Anthropic’s runtime parser is undisturbed.

9.1 Per-file provenance header (required)

Every imported file gets a 7-line HTML comment above any YAML frontmatter — Claude Code’s parser ignores HTML comments before the frontmatter, so this is invisible at runtime but visible on GitHub. Source-URL always points at the upstream main branch (so the link keeps working as the upstream evolves); Source-Pinned-URL is the immutable commit-pinned permalink (so the exact version that was imported can always be retrieved):

<!--
Source-Repo: wshobson/agents
Source-Path: plugins/frontend-mobile-development/agents/frontend-developer.md
Source-URL: https://github.com/wshobson/agents/blob/main/plugins/frontend-mobile-development/agents/frontend-developer.md
Source-Pinned-URL: https://github.com/wshobson/agents/blob/7f3a4b2c.../plugins/frontend-mobile-development/agents/frontend-developer.md
Source-License: MIT
Source-Commit: 7f3a4b2c…
Imported-At: 2026-05-05T09:50:00Z
-->
---
name: frontend-developer
description: …
---
…rest of agent body…

This is what import_file in §8 injects automatically.

9.2 Top-level NOTICE.md

Add NOTICE.md to the repo root listing every upstream + commit SHA + license + clickable file-level URL. Generated from .claude/agents/.imported-manifest (6 columns, written by §8 step 3):

# regenerate-notice.sh
{
  echo "# Third-Party Notices"
  echo
  echo "This repository's \`.claude/agents/\` directory contains files imported from"
  echo "the OSS projects below. Each imported file retains its origin in an HTML"
  echo "comment header (\`Source-Repo\`, \`Source-URL\`, \`Source-Pinned-URL\`,"
  echo "\`Source-License\`, \`Source-Commit\`, \`Imported-At\`), and the upstream"
  echo "LICENSE files are vendored under \`.claude/agents/_licenses/\`."
  echo
  echo "| # | Imported file | Upstream repo | Upstream file (link) | Commit | License |"
  echo "|---|---|---|---|---|---|"
  i=0
  # Skip the 2 header lines, then format. Column order in manifest:
  #   1=imported-name 2=source-repo 3=source-path 4=commit-sha 5=license 6=source-url
  tail -n +3 .claude/agents/.imported-manifest | while IFS=$'\t' read -r name repo path sha license url; do
    i=$((i+1))
    printf '| %s | `%s` | [%s](https://github.com/%s) | [`%s`](%s) | `%s` | %s |\n' \
      "$i" "$name" "$repo" "$repo" "$path" "$url" "${sha:0:8}" "$license"
  done
} > NOTICE.md

9.3 LICENSE vendoring

Step 4 of the install script copies each upstream LICENSE verbatim into .claude/agents/_licenses/. MIT requires preserving the copyright notice and the permission-notice text (clauses 1–2). Vendoring the file itself is the minimal-friction way.

9.4 README disclosure

Add to repo root README.md:

## Attribution

Many `.claude/agents/*.md` files in this repo were imported from upstream OSS projects. See [`NOTICE.md`](./NOTICE.md) for the full per-file attribution table and [`/.claude/agents/_licenses/`](./.claude/agents/_licenses/) for the vendored upstream LICENSE files. Imported files carry an HTML comment header with their `Source-Repo`, `Source-Path`, `Source-Commit`, and `Source-License`.

9.5 Modifications to imported files

If you modify an imported file (e.g., adapt frontend-developer.wshobson.md to your project), append a Modified-By: line under Imported-At: in the header. Do NOT remove the Source-* lines — that would breach MIT’s attribution requirement. Example:

<!--
Source-Repo: wshobson/agents
Source-Path: plugins/frontend-mobile-development/agents/frontend-developer.md
Source-URL: https://github.com/wshobson/agents/blob/main/plugins/frontend-mobile-development/agents/frontend-developer.md
Source-Pinned-URL: https://github.com/wshobson/agents/blob/7f3a4b2c.../plugins/frontend-mobile-development/agents/frontend-developer.md
Source-License: MIT
Source-Commit: 7f3a4b2c…
Imported-At: 2026-05-05T09:50:00Z
Modified-By: secusy team — adapted to Plane/Mattermost workflow on 2026-05-12
-->

Better practice for secusy: don’t modify imported files at all. If a wshobson frontend-developer.wshobson.md needs project-specific tweaks (Plane lifecycle, Mattermost mentions), those tweaks belong in the secusy-original orchestrator (frontend.md) which then delegates to the unmodified wshobson specialist. That preserves the ability to re-run §8’s install script and pick up upstream improvements without merge conflicts.

10. License compatibility — MIT inbound, MIT outbound

10.1 Live audit — every candidate repo

Pulled directly from each repo’s LICENSE file via raw.githubusercontent.com on 2026-05-05:

Repo License file License Copyright holder Notes
affaan-m/everything-claude-code LICENSE MIT © 2026 Affaan Mustafa Standard MIT
wshobson/agents LICENSE MIT © 2024 Seth Hobson Standard MIT
VoltAgent/awesome-claude-code-subagents LICENSE MIT © 2025 VoltAgent Standard MIT
0xfurai/claude-code-subagents LICENSE MIT © 2025 0xfurai Standard MIT
dl-ezo/claude-code-sub-agents (no LICENSE file) MIT (declared in README) (no holder named) ⚠ Less defensible — see §10.4
vijaythecoder/awesome-claude-agents LICENSE MIT © 2024 Awesome Claude Agents Contributors Standard MIT
iannuttall/claude-agents LICENSE MIT © 2025 ian nuttall Standard MIT
lst97/claude-code-sub-agents LICENSE MIT © 2025 SIO TOU LAI [Nelson] Standard MIT
senaiverse/claude-code-reactnative-expo-agent-system LICENSE MIT © 2025 SenaiVerse Standard MIT
bmad-code-org/BMAD-METHOD LICENSE MIT © 2025 BMad Code, LLC Not used in §7 (skill-shaped) but listed for completeness
obra/superpowers LICENSE MIT © 2025 Jesse Vincent Used as plugin, not vendored

10.2 Compatibility verdict

No conflicts. Every source is MIT (one is MIT-by-README — see §10.4 mitigations). MIT is permissive and explicitly compatible with redistributing under MIT, including modified versions, provided the original copyright notice and permission notice are preserved (the two-paragraph LICENSE body). Outbound MIT for your repo is the natural choice — combining MIT-licensed inputs into an MIT-licensed output requires only:

  1. Preserve original copyright notices → §9.1 per-file headers + §9.3 vendored LICENSE files
  2. Preserve the MIT permission notice → §9.3 vendored LICENSE files
  3. Disclaim warranty / liability → satisfied by your MIT LICENSE in the repo root (the standard MIT body covers this)

There is no copyleft contamination, no GPL ambiguity, no Apache NOTICE-file requirement, no patent-grant divergence. MIT-into-MIT is the trivial case.

10.3 What you would need to change for any of these scenarios

If you pulled future agents from a repo with a different license, the picture changes:

Upstream license Compatible with MIT outbound? What you’d need
MIT / BSD-2 / BSD-3 / ISC ✅ Yes Just preserve copyright + license notice
Apache-2.0 ✅ Yes (MIT can redistribute Apache code) Vendor LICENSE and any NOTICE file; record any patent grants
MPL-2.0 ⚠ File-level copyleft Modified files stay MPL; can keep original-MPL files alongside MIT files but cannot relicense them to MIT
LGPL ⚠ Copyleft on modifications + linking quirks Avoid for simple text agents; legal review needed
GPL-2.0 / GPL-3.0 ❌ Incompatible with MIT outbound Cannot relicense; whole repo would have to be GPL
AGPL-3.0 ❌ As GPL plus network use Same: whole repo would have to be AGPL
CC-BY-4.0 ✅ Compatible (similar attribution) Preserve attribution; not ideal mixed with code MIT but works for prose/docs
CC-BY-SA-4.0 ❌ ShareAlike clause Cannot relicense to MIT
CC-BY-NC ❌ Non-commercial restriction Cannot use in any commercial context
Custom / unclear ❌ Treat as All Rights Reserved Skip until clarified by upstream

Translation: as long as you only cherry-pick from MIT/BSD/ISC/Apache-2.0 upstreams, the MIT outbound strategy holds. Any addition outside that set should trigger a review against this table.

10.4 dl-ezo edge case — MIT declared in README only

dl-ezo/claude-code-sub-agents has no LICENSE file in the repo. README.md says “MIT License - feel free to use, modify, and distribute these agents for any purpose.” This is the maintainer’s clear intent but is legally weaker than a LICENSE file because:

Mitigation strategy (already implemented in §8 step 4):

  1. Vendor .claude/agents/_licenses/LICENSE.dl-ezo-claude-code-sub-agents.NOTE quoting the README verbatim with timestamp + permalink
  2. In §9.1 file headers for those 6 imports, set Source-License: MIT (declared in upstream README)
  3. (Optional, recommended) Open a PR upstream adding a proper LICENSE file. Keeps everyone safer
  4. (Optional) If 3 doesn’t land in 30 days, replace the dl-ezo agents with hand-written equivalents derived from §3’s runner-up (BMAD-METHOD skills translated to agents) — that removes the legally weaker dependency entirely

For now the risk is low (clear public intent + small specialized files + §10.5 final defense), but worth tightening if the resulting repo is used in commercial contexts.

your-repo/
├── LICENSE                              ← MIT, your copyright, repo-wide
├── NOTICE.md                            ← §9.2 generated table
├── README.md                            ← §9.4 attribution section
└── .claude/
    └── agents/
        ├── architect.ecc.md             ← imported, with §9.1 header
        ├── …                            (53 more)
        ├── _licenses/                   ← §9.3 vendored upstream LICENSEs
        │   ├── LICENSE.affaan-m-everything-claude-code
        │   ├── LICENSE.wshobson-agents
        │   ├── LICENSE.VoltAgent-awesome-claude-code-subagents
        │   ├── LICENSE.0xfurai-claude-code-subagents
        │   └── LICENSE.dl-ezo-claude-code-sub-agents.NOTE
        └── .imported-manifest           ← §8 SHA-pinned manifest

Root LICENSE text (standard MIT, with one extra paragraph at the end making the layered nature explicit):

MIT License

Copyright (c) 2026 <YOUR NAME>

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

---

This repository includes third-party content imported from MIT-licensed
upstreams. Each such file carries a `Source-Repo:` header naming its origin,
and the upstream LICENSE files are vendored under
`.claude/agents/_licenses/`. See `NOTICE.md` for the full attribution table.

That’s the full discharge of MIT inbound + MIT outbound obligations. Compatible. Auditable. No surprises.

10.6 Quick checklist before publishing

11. Ralph (snarktank/ralph) — fit assessment & positioning

Ad-hoc analysis added 2026-05-05 in response to a question: can Ralph’s functionality be expressed as .claude/agents/*.md and slotted into the §7 architecture? Conclusion below; verdict is partial fit — Ralph is mostly a runtime loop and a workflow skill, not an agent. It belongs alongside the agent set as a thin orchestrator layer, not inside it.

11.1 What Ralph actually is

Source: github.com/snarktank/ralph, MIT-licensed, based on Geoffrey Huntley’s Ralph pattern. Five components:

Ralph component Form Role
ralph.sh (~3.5KB bash) Outer driver for i in 1..N; spawn fresh AI instance with prompt.md/CLAUDE.md; check for COMPLETE; archive on branch change
prompt.md / CLAUDE.md Single-page agent prompt The instructions each fresh iteration receives — read PRD, pick highest-priority story where passes:false, implement, run checks, commit, append to progress.txt, set passes:true, repeat
prd.json Persistent state Backlog of user stories with {id, title, acceptanceCriteria, priority, passes:false}
progress.txt Persistent state Append-only log + a “Codebase Patterns” header section that future iterations read first
skills/prd/ + skills/ralph/ Two Claude Code skills (a) generate PRD from feature description, (b) convert PRD markdown → prd.json

Operating model: each iteration is a fresh Claude/Amp process — no persistent agent context. State carries through git history + progress.txt + prd.json. Stop condition is the LLM emitting <promise>COMPLETE</promise>. The “loop until all stories pass” loop is in bash, not in the agent.

11.2 Does Ralph slot into §7’s .claude/agents/*.md shape?

Mostly no — and that’s the right answer.

Ralph is structurally three things, and only one of them is an agent file:

  1. An outer driver script (ralph.sh) — this is not an agent at all. It belongs in genesis/ or scripts/, not .claude/agents/.
  2. A single agent prompt (prompt.md / CLAUDE.md) — this could be expressed as a .claude/agents/ralph-iterator.md, but it’s a one-shot self-driving prompt, not a Task-tool-callable specialist. Calling it through Claude Code’s subagent dispatch (Task(subagent_type="ralph-iterator")) doesn’t make sense — it expects to be the top-level conversation.
  3. Two skills (skills/prd, skills/ralph) — these are skills, not agents. They go in ~/.claude/skills/ or via /plugin install ralph-skills@ralph-marketplace. They have no native home in .claude/agents/.

The §7 architecture cannot host Ralph as a peer to frontend.md / backend.md. Ralph operates one level above the agent layer — it’s a session driver. Trying to force it into .claude/agents/ would either lose its essence (the fresh-context loop) or break the §7 contract that everything in that folder is a Task-callable subagent.

11.3 Where Ralph does fit — three positioning options

Option A — Sprint executor for a single feature branch (recommended)

Ralph becomes a fourth path beside the existing two paths in CLAUDE.md “Git 브랜치 전략” section:

Path Operator Use case
Path A Human maintainer Direct main commits
Path B Agent team (frontend/backend/qa…) Plane-ticket-driven feat branches
Path C (new) Ralph driver PRD-driven autonomous burn-down of a multi-story feature inside one sprint

Concretely:

Mapping to file system:

genesis/
├── ralph.sh                     ← copied from snarktank/ralph (MIT vendored)
├── ralph-prompt.md              ← extends Ralph's CLAUDE.md with secusy quality gates
└── install-curated-agents.sh    ← unchanged from §8

.claude/
├── agents/                      ← unchanged 54-file curated set from §7
└── skills/
    ├── prd/                     ← from snarktank/ralph (or via /plugin install)
    └── ralph/                   ← from snarktank/ralph

Ralph’s prompt then calls secusy’s existing agents per iteration:

Option B — Skills only (lightest touch, ship today)

Skip ralph.sh and the loop entirely. Just install the two skills (/prd, /ralph) globally:

# In Claude Code
/plugin marketplace add snarktank/ralph
/plugin install ralph-skills@ralph-marketplace

Now anywhere in the secusy project, a human or pm agent can run /prd "feature description" to draft a PRD, then /ralph tasks/prd-feature.md to turn it into the secusy backlog (adapt the schema converter to emit Plane intake tickets instead of prd.json).

This is a drop-in upgrade for pm.md’s requirement-intake step without taking on Ralph’s loop runtime.

Option C — Embed Ralph’s idea into pm.md (no vendoring)

Don’t import Ralph at all. Steal the four most valuable concepts and bake them into existing secusy agents:

Ralph concept Secusy equivalent today Action
Fresh-context iterations Each Plane-ticket-driven Task() spawn Already done
progress.txt Codebase Patterns header MEMORY.md + CLAUDE.md Already done
<promise>COMPLETE</promise> stop signal pm’s /sprint-end Already done
prd.json priority + passes flags Plane state machine (Backlog→Todo→Done) Already done
AGENTS.md auto-update by every iteration “When you discover reusable patterns, append to nearest CLAUDE.md / docs/” rule Missing — add to pm.md or create a memory-curator agent

The one genuinely novel idea Ralph brings is the enforce-pattern-extraction rule: “every iteration MUST update the nearest AGENTS.md / CLAUDE.md if it discovered a reusable pattern”. Adding this as a hook (PostToolUse on Edit) or as a step-9 in frontend.md / backend.md would capture Ralph’s distinctive signal without vendoring its loop.

Adopt Options B + C; defer Option A.

Reasoning:

11.5 If Option A is chosen later

Should secusy decide to add Ralph as an autonomous burn-down path (e.g., for off-hours or unattended weekend runs), the additions to §7 would be:

.claude/skills/
├── prd/                                   ← from /plugin install ralph-skills@ralph-marketplace
└── ralph/                                 ← same

genesis/
├── ralph.sh                               ← vendored from snarktank/ralph (MIT, attribution required)
├── ralph-prompt.secusy.md                 ← secusy's wrapper around Ralph's CLAUDE.md
│                                            (calls Plane API instead of editing prd.json,
│                                             posts to Mattermost instead of progress.txt)
└── install-curated-agents.sh              ← unchanged

§9 attribution rules would extend: any vendored Ralph file gets the same Source-Repo: snarktank/ralph header. License is MIT — fully compatible with the §10 strategy. No outbound license changes needed.

A new .claude/agents/ralph-driver.md would NOT be created — Ralph drives at the session level, above the subagent layer.

11.6 Verdict — one-line summary for the doc reader

Ralph is NOT an agent — it’s a 50-line bash loop + a prompt template + two skills. The pattern it embodies (PRD → fresh-context iterations → progress.txt learnings) is already implemented in secusy at a higher fidelity by the Plane-Cycle + agent-ownership architecture. Adopt Ralph’s two skills (/prd, /ralph) and its mandatory inline knowledge capture rule (§12 — Option B + C). Skip the loop driver (Option A) — it’s a regression versus secusy’s parallel agent model. License: MIT, no attribution drama if any of it is vendored.

12. Ideas worth borrowing from Ralph (reference cheat sheet)

Companion to §11. §11 is “should we adopt Ralph?” — answer was partially. This section is the concrete checklist of ideas to lift from Ralph if/when we choose to adopt them, with each idea boiled down to (a) what Ralph does, (b) the closest standard SE term, (c) how to express it in secusy’s existing structure. Source for every quote: snarktank/ralph (MIT) — prompt.md, CLAUDE.md, ralph.sh, README.md.

12.1 Mandatory inline knowledge capture (the one truly novel idea)

What Ralph does (CLAUDE.md “Update AGENTS.md Files” section):

Before committing, check if any edited files have learnings worth preserving in nearby AGENTS.md files:

  1. Identify directories with edited files
  2. Check for existing AGENTS.md
  3. Add valuable learnings — if you discovered something future developers/agents should know: API patterns or conventions specific to that module · Gotchas or non-obvious requirements · Dependencies between files · Testing approaches for that area · Configuration or environment requirements

And (CLAUDE.md “Consolidate Patterns” section):

If you discover a reusable pattern that future iterations should know, add it to the ## Codebase Patterns section at the TOP of progress.txt (create it if it doesn’t exist).

Closest standard SE term: living documentation (Cyrille Martraire) enforced as a commit gate. Adjacent academic term: knowledge externalization (Nonaka & Takeuchi SECI model — making tacit knowledge explicit).

Why it’s powerful: most “documentation drift” doesn’t happen because devs don’t want to document — it happens because there’s no forcing function. Ralph turns “update the docs if you discovered a pattern” from aspiration into commit-blocking step.

Filter rule (also from Ralph CLAUDE.md, equally important):

Do NOT add: Story-specific implementation details · Temporary debugging notes · Information already in progress.txt Only update CLAUDE.md if you have genuinely reusable knowledge that would help future work in that directory.

So the rule is bidirectional: forced to capture genuine patterns, forbidden from polluting with story-specific noise.

How to graft into secusy:

Add a step-7.5 to frontend.md, backend.md, qa.md, security.md, devops.md, designer.md orchestrators (right before “PR 생성”):

## 7.5 커밋 직전 — 패턴 캡처 게이트 (Ralph-style)

자기 변경분을 살펴보고, 다음 중 하나라도 발견했으면 **커밋 전에** 가장 가까운 디렉터리의
`CLAUDE.md` 또는 `docs/<feature>.md` 에 한 줄 추가한다:

- 이 모듈만의 API/네이밍/설정 컨벤션
- 다른 파일과의 비명시적 의존 ("X 바꾸면 Y도 같이 바꿔야 함")
- 비자명한 gotcha (Plane state machine, hook 동작 순서, env var 종속 등)
- 이 영역 특유의 테스트 셋업 절차

다음은 절대 추가하지 않는다:
- 이번 티켓 한정의 구현 디테일
- 임시 디버깅 노트
- 이미 `progress.txt` / Mattermost 스레드에 있는 정보

판단 기준 한 줄: **"내가 다음 sprint에 이 디렉터리 처음 본 다른 agent라면 이 정보가 시간을
아껴줄까?"** 답이 yes 일 때만 커밋. (Ralph CLAUDE.md `Update AGENTS.md Files` §)

Optional 강제 hook (PostToolUse on Edit, 실험적):

// .claude/settings.json
{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "command": "node .claude/hooks/pattern-capture-reminder.js",
        "description": "If the edit touches a directory with no recent CLAUDE.md update, prompt the agent to consider updating it before commit"
      }
    ]
  }
}

훅이 차단(exit 2)하지는 않고, agent에게 stderr로 한 줄 알림만 띄우는 부드러운 강제. Ralph의 prompt 안 instruction → secusy의 hook으로 매체 변환.

12.2 Append-only progress log with consolidated patterns header

What Ralph does (prompt.md):

## Codebase Patterns                       ← 항상 최상단, 일반화된 학습만
- Use `sql<number>` template for aggregations
- Always use `IF NOT EXISTS` for migrations
- Export types from actions.ts for UI components

## [Date/Time] - [Story ID]                ← append-only timeline
- What was implemented
- Files changed
- **Learnings for future iterations:** ...
---

파일 구조의 핵심: 두 영역이 분리되어 있다.

Closest standard SE term: append-only event log + materialized projection — 이벤트소싱 패턴의 단순한 적용. 또는 engineering decision log (ADR의 가벼운 형태).

How to graft into secusy:

secusy는 이미 등가물을 갖고 있다 — Mattermost #scrum-secusy 스레드 + Plane 이슈 코멘트가 chronological log, MEMORY.md + 디렉터리별 CLAUDE.md 가 consolidated projection. 따라서 새로 만들 필요 없음. 다만 두 영역의 분리 규칙을 더 명시적으로 강제하는 게 가치가 있다:

Ralph location Secusy 등가물 어디에 올라가야 하는 내용
## Codebase Patterns (top of progress.txt) MEMORY.md + <dir>/CLAUDE.md + docs/design-system.md 재사용 가능한 패턴, 컨벤션, 결정
## [Date/Time] - [Story ID] (timeline) Mattermost 이슈 스레드 + Plane 이슈 코멘트 이번 작업 한정 진행 상황

위 표는 §12.1 step-7.5 보강을 위해 pm.md 또는 CLAUDE.md 본문에 그대로 넣을 수 있다.

12.3 Right-sized stories rule

What Ralph does (README.md “Small Tasks” section):

Each PRD item should be small enough to complete in one context window. If a task is too big, the LLM runs out of context before finishing and produces poor code.

Right-sized stories: Add a database column and migration · Add a UI component to an existing page · Update a server action with new logic · Add a filter dropdown to a list

Too big (split these): “Build the entire dashboard” · “Add authentication” · “Refactor the API”

Closest standard SE term: INVEST criteria (eXtreme Programming) — 특히 S = Small. 또는 vertical slice + story splitting (Mike Cohn).

Why it matters here: secusy의 estimate 옵션은 이미 3분 / 5분 / 10분 / 15분 / 30분 / 45분인데, 상한 45분이 LLM context 한계와 우연히 잘 맞는다 — Ralph의 “한 context window 안에 끝나야 함” 룰의 secusy 버전이다. 명시적으로 적어두면 estimate 결정이 더 일관돼진다.

How to graft into secusy:

pm.md 의 “Estimate” 섹션 또는 CLAUDE.md 의 “티켓 관리 원칙” 에 한 줄 추가:

### Estimate 가이드 (Ralph "right-sized stories" 원칙)
- 한 티켓 = 한 agent의 한 context window 안에 끝나야 함 (45분 상한 = ~한 세션)
- 끝나지 않을 가능성이 보이면 **분할**: 30분 + 30분 두 티켓이 60분 한 티켓보다 항상 낫다
- Ralph 가이드의 "too big" 예시: "전체 대시보드 구축", "인증 추가", "API 리팩토링" → 이런 사이즈는 Story가 아니라 Epic(=Module)

12.4 Stop signal as explicit completion contract

What Ralph does (prompt.md):

If ALL stories are complete and passing, reply with: <promise>COMPLETE</promise> If there are still stories with passes: false, end your response normally.

ralph.sh 가 stdout에서 그 토큰을 grep해서 루프 종료. machine-readable completion contract — LLM이 자유 텍스트로 “다 했어요!”라고 말해도 무시되고, 약속된 정확한 토큰만 신뢰.

Closest standard SE term: explicit termination protocol / sentinel value. 또는 분산 시스템의 commit token / ack message.

How to graft into secusy:

secusy는 이미 더 강한 등가물을 갖고 있다 — Plane 상태 머신 (In Review → Done) + qa 멘션 (@qa.{project} merge·Done 처리 가능). 텍스트 토큰보다 훨씬 안정적이라 그대로 가져올 가치는 없다. 다만 Mattermost 스레드 자동 파싱이 필요한 경우(예: pm 봇이 sprint-status를 자동 집계)에는 정확한 토큰 컨벤션이 도움이 된다. 후보:

[SCZ-N] ✅ Done — <한 줄 요약>      ← 이미 사용 중 (.claude/rules/agent-ownership.md §3)

이미 있으니 변경 없음. 단지 Ralph가 같은 패턴을 쓴다는 점이 검증.

12.5 Branch-scoped run archiving

What Ralph does (ralph.sh lines 38-58):

# Archive previous run if branch changed
if [ "$CURRENT_BRANCH" != "$LAST_BRANCH" ]; then
  ARCHIVE_FOLDER="$ARCHIVE_DIR/$DATE-$FOLDER_NAME"
  mkdir -p "$ARCHIVE_FOLDER"
  cp "$PRD_FILE" "$ARCHIVE_FOLDER/"
  cp "$PROGRESS_FILE" "$ARCHIVE_FOLDER/"
  # Reset progress file for new run
fi

작업 중인 feature 브랜치가 바뀌면 직전 run의 prd/progress를 자동으로 archive/YYYY-MM-DD-feature/ 로 옮기고 새 progress.txt를 시작.

Closest standard SE term: snapshot isolation between runs / generation-scoped state.

How to graft into secusy:

이미 등가물 존재 — Plane Cycle 종료 시 /sprint-end 가 Mattermost에 회고 요약 게시 + Cycle archive. secusy에는 추가할 게 없다. 단, sprint 단위가 아니라 개별 PR/feat 브랜치 단위로 학습을 보존하고 싶다면 Ralph 패턴을 흉내낼 수 있다:

docs/sprint-archives/
└── 2026-W18-SCZ-23-design-tokens/
    ├── plane-snapshot.json   ← 그 cycle 시작 시점 이슈 상태
    ├── mattermost-thread.md  ← 해당 이슈 스레드 export
    └── learnings.md          ← consolidated patterns from this run

/sprint-end skill 에 옵션 단계로 추가 가능. 우선순위는 낮음.

12.6 한눈에 보는 Borrow / Skip 표

Ralph 구성요소 secusy에 가져올까? 이유
/prd skill (PRD 생성) Borrow (§11 Option B) pm의 요구사항 수집 단계를 강화. 비용 0
/ralph skill (PRD→JSON 변환) Borrow (스키마 적응) Plane intake 자동 생성으로 변환 시 가치
§12.1 mandatory knowledge capture rule Borrow (필수) secusy에 빠진 유일한 가치, 도입 비용 작음
§12.2 append-only + consolidated 분리 규칙 Borrow as documentation 이미 구현되어 있지만 명문화 가치
§12.3 right-sized stories Borrow as guidance line estimate 가이드라인에 한 줄 추가
§12.4 stop signal token Skip secusy는 Plane state machine으로 이미 더 강하게 함
§12.5 branch-scoped archiving Skip (선택적) /sprint-end 로 충분, 추가 가치 한계적
ralph.sh 루프 driver Skip (§11 Option A 거부) 병렬 agent 모델 대비 회귀
prompt.md / CLAUDE.md 단일 prompt Skip secusy의 분산 agent 모델과 충돌
prd.json schema Skip Plane이 이미 단일 진실 소스
progress.txt 파일 Skip Mattermost #scrum-secusy + Plane comments로 대체

12.7 도입 시 첫 액션 3개 (구체적)

위 분석에 기반해 secusy에 실제로 적용하려면 다음 3개를 하면 된다 (모두 작은 PR 한 번씩):

  1. §12.1 step-7.5 추가: frontend.md, backend.md, qa.md, security.md, devops.md, designer.md, pm.md 의 본문에 “커밋 직전 — 패턴 캡처 게이트” 섹션을 같은 텍스트로 추가. 이 7개 orchestrator 가 모두 이 규칙을 따르게 만드는 게 핵심.
  2. §12.3 한 줄 추가: CLAUDE.md 의 “Estimate (분 단위 직접 값)” 섹션에 “한 티켓 = 한 agent context window 분량” 가이드라인 한 줄 추가.
  3. §11.4 Option B 실행: /plugin marketplace add snarktank/ralph + /plugin install ralph-skills@ralph-marketplace 를 genesis/setup-plugins.md 에 등록. pm agent가 intake 정제할 때 /prd 사용 가능해짐.

이 3개로 Ralph 가치의 ~95% 를 0.5일 분량 작업으로 흡수한다. 나머지 5%(loop driver, prd.json 스키마)는 secusy 구조와 충돌하므로 의도적으로 버린다.

This site is open source. Improve this page.