{
  "markdown": "<p align=\"center\">\n  <img src=\"https://img.shields.io/badge/Claude_Code-Plugin-7C3AED?style=for-the-badge&logoColor=white\" alt=\"Claude Code\" />\n  <img src=\"https://img.shields.io/badge/Codex-Compatible-10B981?style=for-the-badge&logoColor=white\" alt=\"Codex\" />\n  <img src=\"https://img.shields.io/badge/Cursor-Compatible-1F2937?style=for-the-badge&logoColor=white\" alt=\"Cursor\" />\n  <img src=\"https://img.shields.io/badge/Gemini-Compatible-4285F4?style=for-the-badge&logoColor=white\" alt=\"Gemini\" />\n  <img src=\"https://img.shields.io/badge/version-5.5.0-blue?style=for-the-badge\" alt=\"Version\" />\n  <img src=\"https://img.shields.io/badge/license-MIT-green?style=for-the-badge\" alt=\"License\" />\n</p>\n\n<p align=\"center\">\n  <img src=\"https://img.shields.io/badge/agents-45-orange?style=for-the-badge\" alt=\"Agents\" />\n  <img src=\"https://img.shields.io/badge/commands-38-red?style=for-the-badge\" alt=\"Commands\" />\n  <img src=\"https://img.shields.io/badge/hooks-25-blueviolet?style=for-the-badge\" alt=\"Hooks\" />\n  <img src=\"https://img.shields.io/badge/skills-34-teal?style=for-the-badge\" alt=\"Skills\" />\n</p>\n\n<p align=\"center\">\n  <sub>45 agents = <strong>26 framework-agnostic</strong> (code, growth, debate, planning) + <strong>19 stack specialists</strong> (WordPress, Next.js, Supabase, mobile, n8n, WooCommerce)</sub>\n</p>\n\n<h1 align=\"center\">ForgeBee</h1>\n\n<p align=\"center\">\n  <strong>A colony of AI agents forging your product</strong><br/>\n  45 specialist agents. 34 skills. 38 slash commands. 25 lifecycle hooks.<br/>\n  Three execution modes: inline skills, context:fork, subagents.<br/>\n  Adaptive pipeline. Karpathy principles. Adversarial debate. Continuous learning.<br/>\n  <em>Works with Claude Code, Codex, Cursor, Gemini, and OpenClaw.</em>\n</p>\n\n<p align=\"center\">\n  <a href=\"#quick-start\">Quick Start</a> &nbsp;&bull;&nbsp;\n  <a href=\"#commands\">Commands</a> &nbsp;&bull;&nbsp;\n  <a href=\"#agents\">Agents</a> &nbsp;&bull;&nbsp;\n  <a href=\"#hooks\">Hooks</a> &nbsp;&bull;&nbsp;\n  <a href=\"#continuous-learning\">Learning</a> &nbsp;&bull;&nbsp;\n  <a href=\"#growth-os\">Growth OS</a> &nbsp;&bull;&nbsp;\n  <a href=\"#project-management\">PM System</a> &nbsp;&bull;&nbsp;\n  <a href=\"#openclaw\">OpenClaw</a>\n</p>\n\n---\n\n## Why ForgeBee?\n\nClaude Code and OpenClaw are powerful out of the box. ForgeBee makes them **opinionated and structured** — so your agent plans before it codes, debates before it ships, and verifies before it marks anything \"done.\"\n\n| Without ForgeBee | With ForgeBee |\n|:--|:--|\n| Agent jumps straight into coding | Agent plans, debates requirements, then codes |\n| \"It should work\" | Evidence-based verification with actual test output |\n| Single-agent, single-pass | 45 agents + 34 skills working in parallel with blind review |\n| Manual project tracking | Automated state.yaml + markdown dashboards |\n| No marketing workflow | Full 9-phase Growth OS with 11 marketing agents + 3 strategy debate agents |\n| Every session starts from scratch | Continuous learning — heuristic pattern detection + pending instinct approval |\n| Edits break silently | Auto-format, typecheck, and lint on every edit |\n| Permission prompts everywhere | Mode-aware permissions — fully yields to auto-mode's classifier; enforces blocklist in default and bypass |\n| Commands duplicate agent logic | Commands delegate to specialist agents with automatic fallback |\n| Review finds issues that should have been caught | Quality pipeline — specialists self-review, code-skeptic validates, review-all just confirms |\n| Agents take shortcuts under pressure | Every agent has explicit \"Never\" rules — hard boundaries that can't be rationalized away |\n| Skill descriptions drift silently | All skill descriptions follow \"Use when...\" triggering format — Claude reads triggers, not workflow summaries |\n| Version numbers drift across manifests | `.version-bump.json` + `scripts/bump-version.sh` — single source of truth, drift detection, full-repo audit |\n| Debug loops forever on the same bug | 3-failed-fix Iron Law — debugger-detective escalates to architecture question after 3 failed attempts |\n| User forgets to run `/learn` | Auto-nudge on SessionStart when ≥5 pending instincts and no `/learn` in 24h |\n| One implementation = one harness | Single source ships to Claude Code, Codex, Cursor, Gemini via dedicated manifests |\n| Agents make silent assumptions | Karpathy P1–P4 baked into every code-producing agent: trace test, senior-engineer check, YAGNI timing, orphan rule |\n| Orchestrators idle after dispatch | P5 Anti-Stop Rule in `/workflow` + `/team`: continue with next-step work immediately, the harness wakes you when sub-agents return |\n| Review reviewer drift | P6 Severity Standard (`Critical/High/Medium/Low`) across all review skills — enables cross-skill aggregation |\n| Discovery searches hang on vendored subtrees | T1–T5 Tool Discipline in `CLAUDE.md`: prefer `rg` over `grep -r`, bound every search (scope/type/timeout), `Glob` before `grep`, exclude `vendor`/`node_modules` — kills 10-minute stalls on `wp-content/` and monorepo roots |\n| Untrusted code can hijack agents | 6-line Adversarial Input Hardening preamble in all 45 agents — homoglyphs, urgency, role-play overrides flagged not executed |\n| Sub-agent reports waste orchestrator context | `terse-report` skill cuts ~65% of report tokens while preserving code/citations exact |\n| Decisions get lost between sessions | `/workflow` and `/plan` emit `.decision-log.md` + `addendum.md` — re-read on next run |\n| Recursive debate fan-out runs away | Budget circuit breaker (`maxHops`, `maxTokens`, `maxUsd`) on every dispatch with constant-string errors |\n| Silent picks mid-task | `surface-ambiguity` micro-skill forces listing of interpretations + rationale before non-trivial choices |\n| Debugging without evidence | Failure-Capture template (7 fields) required before any recovery action in `debugger-detective` |\n| Diagnosis vs fix blur together | `/investigate` produces a forensic case file (Confirmed / Deduced / Hypothesized) — `debugger-detective` then fixes |\n| Reviews skim or lose-the-forest | Checkpoint Preview phase: diff-by-concern with `[auth]` / `[schema]` / `[billing]` risk tags before debate |\n| Plans drift without stress-test | `/elicit` applies 18 named methods (Pre-mortem, Red Team, Inversion, Stakeholder Round Table…) to your own output |\n| Quality decays as the kit grows | `/audit-self` re-runs the full scorecard on demand and surfaces regressions since last audit |\n| Claude scans every agent/skill/command frontmatter block per routing decision | Auto-generated `forgebee/INDEX.md` loaded once on SessionStart — replaces speculative scanning with one indexed read |\n\n---\n\n## Quick Start\n\n**Two commands. That's it.**\n\n```bash\n# Inside Claude Code:\n/plugin marketplace add forbee-dev/ForgeBee\n/plugin install forgebee@forbee-dev\n```\n\nThen edit `CLAUDE.md` with your project details and start working. Hooks activate automatically, commands are available immediately.\n\n<details>\n<summary><strong>Install from local directory</strong></summary>\n\n```bash\ngit clone git@github.com:forbee-dev/ClaudeDevKit.git\nclaude --plugin-dir ./ClaudeDevKit/forgebee\n```\n\n</details>\n\n<details>\n<summary><strong>Requirements</strong></summary>\n\n- Claude Code 1.0.33+\n- Node.js 18+ (all hooks are pure Node.js — no bash, no python, no jq)\n- Agent Teams: set `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`\n\n</details>\n\n---\n\n## Commands\n\nInvoke with a slash: `/review`, `/debug`, `/workflow`, etc.\n\n### Planning\n\n| Command | Description |\n|:--------|:------------|\n| `/plan` | Phased planning: Brief &rarr; Requirements &rarr; Architecture &rarr; Sprint Stories |\n\n### Development\n\n| Command | Description |\n|:--------|:------------|\n| `/review` | Multi-dimensional code review (structure, security, perf, correctness) |\n| `/debug` | Delegates to `debugger-detective` agent; systematic reproduce &rarr; isolate &rarr; fix |\n| `/architect` | Architecture decisions with trade-off matrices and ADR generation |\n| `/refactor` | Safe refactoring with test verification at each step |\n| `/test` | Delegates to `test-engineer` agent; unit &rarr; integration &rarr; e2e |\n| `/docs` | API docs, READMEs, ADRs, guides |\n| `/security` | Delegates to `security-auditor` agent; OWASP Top 10 + anti-rationalization gate |\n| `/perf` | Profile &rarr; optimize &rarr; measure |\n| `/migrate` | Version/framework migration with rollback plans + anti-rationalization gate |\n| `/deploy` | Pre-flight checks, rollout, post-deploy verification + anti-rationalization gate |\n| `/browser-debug` | Console, network, rendering, Core Web Vitals |\n| `/codemaps` | Token-lean architecture docs for AI context consumption |\n| `/design-system` | `audit` a Figma page against code, `onboard` a new component, or audit `tokens` — code-first, Figma is derived |\n\n### Growth & Marketing\n\n| Command | Description |\n|:--------|:------------|\n| `/growth` | **9-phase Growth OS** with adversarial debate (16 agents) |\n| `/content` | Quick content production without full pipeline |\n| `/gtm` | Go-to-market planning with launch checklists |\n| `/seo` | Technical SEO audit + keyword research |\n| `/social` | Platform playbooks, hook formulas, content calendars |\n| `/launch` | Product Hunt / HN / press launch execution |\n| `/competitive` | FIA battlecards + niche intelligence |\n| `/landing` | Conversion-optimized landing page (HTML + Tailwind) |\n| `/payments` | Stripe / LemonSqueezy / Paddle integration |\n| `/analytics` | Event tracking, dashboards, metrics framework |\n\n### Learning\n\n| Command | Description |\n|:--------|:------------|\n| `/learn` | Review pending instincts (auto-detected) + analyze observations for new patterns |\n| `/evolve` | Cluster related instincts into skills, commands, or agents |\n| `/instinct-status` | Show all learned instincts (project + global) with confidence scores |\n| `/instinct-export` | Export instincts to a shareable file |\n| `/instinct-import` | Import instincts from a file |\n\n### Meta\n\n| Command | Description |\n|:--------|:------------|\n| `/workflow` | **Full pipeline**: Plan &rarr; Debate &rarr; Architect &rarr; **Implementation Plan** (lightweight default, `--scrum` for full sprint stories) &rarr; Execute &rarr; **Spec Compliance** &rarr; **Checkpoint Preview** (default, `--skip-checkpoint` to opt out) &rarr; Code Debate &rarr; Deliver. Pass `--strict` to require a design spec via `brainstorming` before any code. Auto-offers `/elicit` methods at phase boundaries |\n| `/investigate` | Forensic diagnosis — produces a case file with Confirmed / Deduced / Hypothesized grading. Hand off to `debugger-detective` for the fix. Different from `/debug` (which fixes directly) |\n| `/elicit` | Stress-test the most recent plan or design with one of 18 named reasoning methods (`pre-mortem`, `red-team`, `inversion`, `stakeholder-round-table`, `tree-of-thoughts`, …). Applied to OUTPUT, not requirements |\n| `/audit-self` | Re-run the ForgeBee quality scorecard across all skills, agents, and commands. Detects regressions since last audit. Timestamped findings |\n| `/team` | Multi-agent orchestration with dependency graphs + checkpoints at 3+ agents |\n| `/idea` | Idea &rarr; validate &rarr; debate &rarr; MVP &rarr; roadmap |\n| `/pm` | Project dashboard from `state.yaml` |\n| `/audit` | Governance audit trail — permission decisions, debate rulings, verification results |\n\n---\n\n## Agents\n\n45 specialist agents + 34 skills for Claude Code's Agent Teams. Use them directly or let `/team` and `/workflow` orchestrate automatically.\n\n> **v5.4 highlights — routing actually fires.** Five defects were keeping the 117 surfaces unreachable: the SessionStart index was dropped for exceeding the `additionalContext` size limit (everything past ~2KB, including the whole WordPress roster, never arrived), `skill-activator.js` emitted a top-level `additionalContext` that the harness discards, and it scanned neither agents nor commands — so no agent or slash command could ever be recommended. Now all three surface types are scored together, stack-boosted from `project-triage.json`, capped at 5 candidates, and acronym-aware (ACF, SCF, SEO, API, RLS were silently below the word-length floor). New **Routing Discipline R1–R4** in `CLAUDE.md`: name the route before the first edit, prefer the stack-specific agent. New design-system trio: **`figma-code-sync`** skill (code-first Figma reconciliation, 12 defect priors, ~18 Plugin API traps), **`wp-design-system`** agent (`theme.json`-versus-SCSS layer ownership, token pipeline, patterns/variations), **`/design-system`** command. `wordpress-backend` gains full ACF/SCF field architecture — field-group registration, JSON sync, immutable field keys, repeater meta storage and why it cannot be `meta_query`'d, and ACF-PRO-to-SCF migration risk.\n>\n> **v5.1 highlights:** Karpathy P1–P6 principles baked into every code-producing agent (trace test, senior-engineer check, YAGNI timing, orphan rule, anti-stop rule, severity standard). Seven new skills: `brainstorming` (opt-in via `--strict`), `surface-ambiguity`, `terse-report`, `checkpoint-preview`, `investigate`, `elicitation` (18 methods), `audit-self`. Three new commands: `/investigate`, `/elicit`, `/audit-self`. Adversarial Input Hardening preamble on all 44 agents. Budget circuit breaker on dispatch (maxHops/maxTokens/maxUsd). Decision logs, failure-capture templates, learnings compression. Bloat trim: 6 over-budget agents trimmed to ≤200 lines via references/. **Auto-generated `forgebee/INDEX.md`** — Claude reads one indexed routing map on SessionStart instead of scanning every skill description. See [`CHANGELOG.md`](./CHANGELOG.md) for full release notes.\n>\n> **v5.0 highlights:** Multi-platform manifests (Codex/Cursor/Gemini), brainstorming hard-gate skill (opt-in via `/workflow --strict`), two-stage review (spec compliance before code debate), `debugger-detective` Iron Law (3 failed fixes → escalate), auto-learn SessionStart nudge, \"Use when...\" skill descriptions, and `.version-bump.json` drift detection.\n>\n> **v4.1 migration:** 21 agents moved to skills — 9 debate agents and 12 review agents now use `context:fork` for efficient isolated execution. `review-all` is an inline skill that runs in session context.\n\n<details>\n<summary><strong>Development</strong> (8 agents)</summary>\n\n| Agent | Use when... |\n|:------|:------------|\n| `frontend-specialist` | React, Vue, Svelte, Angular, CSS, UI work |\n| `backend-engineer` | APIs, Express, FastAPI, Django, Go, Rust |\n| `database-specialist` | Schema, migrations, queries, ORMs |\n| `security-auditor` | Auth changes, data handling, OWASP reviews |\n| `test-engineer` | Test generation, coverage improvement |\n| `devops-engineer` | Docker, CI/CD, deployment, infrastructure |\n| `performance-optimizer` | Profiling, bottlenecks, bundle analysis |\n| `debugger-detective` | Errors, test failures, root cause analysis |\n\n</details>\n\n<details>\n<summary><strong>Quality Gates</strong> (5 agents) &mdash; NEW in v2.3</summary>\n\n| Agent | Use when... |\n|:------|:------------|\n| `verification-enforcer` | Task completion &mdash; demands test output, build results, evidence |\n| `tdd-enforcer` | TDD discipline &mdash; blocks code written before tests |\n| `delivery-agent` | Final packaging after code debate + verification |\n| `dashboard-generator` | Regenerating PM dashboards from state.yaml |\n| `contract-validator` | Agent handoffs — validates output contracts between pipeline phases |\n\n</details>\n\n<details>\n<summary><strong>Dev Debate</strong> (6 context:fork skills)</summary>\n\n| Skill | Role |\n|:------|:-----|\n| `requirements-advocate` | Defends planning artifacts (blind, isolated) |\n| `requirements-skeptic` | Challenges planning artifacts (blind, isolated) |\n| `requirements-judge` | Rules: approve / block / flag |\n| `code-advocate` | Defends implementation (blind, isolated) |\n| `code-skeptic` | Challenges implementation (blind, isolated) |\n| `code-judge` | Rules: approve / block / flag |\n\n</details>\n\n<details>\n<summary><strong>Strategy Debate</strong> (3 context:fork skills)</summary>\n\n| Skill | Role |\n|:------|:-----|\n| `strategy-advocate` | Defends marketing strategy (blind, isolated) |\n| `strategy-skeptic` | Challenges marketing strategy (blind, isolated) |\n| `strategy-judge` | Rules: approve / block / flag |\n\n</details>\n\n<details>\n<summary><strong>Growth OS</strong> (11 agents)</summary>\n\n| Agent | Use when... |\n|:------|:------------|\n| `brand-strategist` | Brand positioning, archetypes, voice |\n| `market-intel` | Competitive research, battlecards |\n| `audience-architect` | ICPs, personas, buyer journeys |\n| `content-strategist` | Pillars, topic clusters, hub-and-spoke, content ideas, angle mining, editorial calendar |\n| `hook-engineer` | Scroll-stopping hooks, pattern interrupts |\n| `engagement-strategist` | Community, reciprocity loops, DM flows |\n| `content-creator` | Platform-native posts, threads, scripts, long-form blogs, landing copy |\n| `growth-engineer` | Growth loops, flywheels, funnels, CRO audits, conversion optimization |\n| `marketing-analyst` | Marketing KPIs, attribution, A/B testing |\n| `seo-specialist` | Keyword strategy, pillar/cluster SEO |\n| `email-strategist` | Automation flows, segmentation, deliverability |\n\n</details>\n\n<details>\n<summary><strong>Design, Research & Platform</strong> (8 agents)</summary>\n\n| Agent | Use when... |\n|:------|:------------|\n| `ux-designer` | User flows, wireframes, accessibility |\n| `scrum-master` | Sprint planning, story decomposition |\n| `deep-researcher` | Documentation, APIs, technical questions |\n| `session-librarian` | Session history, context management |\n| `supabase-specialist` | Supabase auth, RLS, edge functions |\n| `ios-expert` | Swift, SwiftUI, Xcode, App Store |\n| `flutter-expert` | Dart, Flutter, cross-platform |\n| `n8n-builder` | n8n workflows, API integrations |\n\n</details>\n\n<details>\n<summary><strong>WordPress Specialists</strong> (8 agents) &mdash; NEW in v3.0</summary>\n\n| Agent | Use when... |\n|:------|:------------|\n| `wordpress-backend` | WordPress PHP backend, REST endpoints, ACF/SCF field architecture, hooks |\n| `wordpress-frontend` | WordPress theme dev, block themes, template hierarchy |\n| `wordpress-content` | WordPress Gutenberg content, block patterns, ACF content |\n| `wordpress-security` | WordPress security audit, sanitization, WPCS |\n| `wordpress-seo` | WordPress SEO, Yoast/RankMath, XML sitemaps |\n| `phpunit-engineer` | WordPress PHPUnit testing, WP_UnitTestCase |\n| `woocommerce-cro` | WooCommerce checkout/product page CRO |\n| `wp-design-system` | `theme.json` token architecture, Figma-to-presets pipeline, patterns/variations as design-system components |\n\n</details>\n\n<details>\n<summary><strong>Next.js Specialists</strong> (3 agents) &mdash; NEW in v3.0</summary>\n\n| Agent | Use when... |\n|:------|:------------|\n| `nextjs-frontend` | Next.js App Router, Server/Client Components, Supabase SSR |\n| `nextjs-content` | Next.js MDX content, Contentlayer, static generation |\n| `nextjs-seo` | Next.js Metadata API, sitemap.ts, OG image generation |\n\n</details>\n\n<details>\n<summary><strong>CRO Specialists</strong> (1 agent) &mdash; NEW in v3.0</summary>\n\n| Agent | Use when... |\n|:------|:------------|\n| `saas-cro` | SaaS landing page/pricing/signup CRO |\n\n</details>\n\n<details>\n<summary><strong>Review Skills</strong> (12 skills) &mdash; migrated to skills in v4.1</summary>\n\n| Skill | Type | Focus |\n|:------|:-----|:------|\n| `review-all` | **Inline** (session context) | Full pre-push quality gate with severity calibration |\n| `review-code` | context:fork | Logic errors, DRY, error handling, dead code |\n| `review-code-style` | context:fork | Convention adherence, imports, naming, file org |\n| `review-security` | context:fork | OWASP Top 10, injection, auth, secrets |\n| `review-performance` | context:fork | N+1 queries, memory leaks, missing caching |\n| `review-accessibility` | context:fork | WCAG 2.1 AA compliance |\n| `review-api` | context:fork | API design, validation, rate limiting, REST consistency |\n| `review-database` | context:fork | Migrations, RLS, schema, query patterns |\n| `review-tests` | context:fork | Coverage, test quality, mocking, structure |\n| `review-docs` | context:fork | Docblocks, comments, parameter docs |\n| `review-best-practices` | context:fork | SOLID, design patterns, architecture health |\n| `review-wordpress` | context:fork | WP coding standards, security, plugin architecture |\n\n`review-all` runs inline for maximum efficiency. For large diffs (>500 lines), it delegates to specialized review skills via `context:fork`.\n\n</details>\n\n---\n\n## Hooks\n\n25 hooks run automatically on Claude Code lifecycle events across 10 event types. No invocation needed.\n\n**Session & state management:**\n\n| Hook | Event | What it does |\n|:-----|:------|:------------|\n| `session-load` | `SessionStart` | Restores previous session context |\n| `session-save` | `Stop` | Persists session state to JSON snapshot |\n| `project-triage` | `SessionStart` | Auto-detects project type, stack, and conventions |\n| `load-context-rules` | `SessionStart` | Loads contexts and language-specific rules |\n| `task-sync` | `SessionStart` + `Stop` | Bidirectional sync with TASKS.md |\n| `pm-sync` | `SessionStart` + `Stop` | Loads PM state, reports blockers |\n| `context-guard` | `PreCompact` | Backs up critical context before compaction |\n\n**Quality automation:**\n\n| Hook | Event | What it does |\n|:-----|:------|:------------|\n| `post-edit-format` | `PostToolUse` (Edit) | Auto-formats JS/TS after every edit (Biome or Prettier) |\n| `post-edit-typecheck` | `PostToolUse` (Edit) | Runs `tsc --noEmit` after editing .ts/.tsx files |\n| `post-edit-console-warn` | `PostToolUse` (Edit) | Warns about `console.log` in edited files |\n| `console-log-audit` | `Stop` | Audits all modified files for console.log at session end |\n| `permission-guard` | `PreToolUse` (Bash) | Mode-aware command safety: yields to classifier in auto-mode; blocklist + ask-cascade in default; blocklist-only in bypass |\n| `dev-server-blocker` | `PreToolUse` (Bash) | Blocks `npm run dev` outside tmux |\n| `git-push-reminder` | `PreToolUse` (Bash) | Warns before pushing to main/master |\n| `secret-scan` | `PreToolUse` (Bash) | Blocks commits/pushes that introduce hardcoded secrets (override: `FORGEBEE_ALLOW_SECRET=1`) |\n| `suggest-compact` | `PreToolUse` (Edit\\|Write) | Suggests `/compact` at logical breakpoints |\n\n**Continuous learning:**\n\n| Hook | Event | What it does |\n|:-----|:------|:------------|\n| `observe` | `PreToolUse` + `PostToolUse` (*) | Captures every tool call lifecycle for pattern extraction |\n\n**Intelligence & governance:**\n\n| Hook | Event | What it does |\n|:-----|:------|:------------|\n| `skill-activator` | `UserPromptSubmit` | Intent detection + skill recommendations |\n| `self-improve` | `Stop` | Captures patterns + runs heuristic engine to flag pending instincts |\n| `checkpoint` | `PreCompact` | Saves the active pipeline phase (derived from `state.yaml`) before context compaction, for crash recovery |\n| `audit-trail` | All governance events | Append-only JSONL log of permissions, debates, verifications |\n| `permission-denied-logger` | `PermissionDenied` | Logs auto-mode classifier denials to audit trail |\n\n**Quality gate hooks** (for Agent Teams):\n\n| Hook | Event | What it does |\n|:-----|:------|:------------|\n| `TaskCompleted` | Task marked done | Verifies completion — accepts config/markdown tasks, demands evidence for code tasks |\n| `TeammateIdle` | Agent going idle | Checks for unclaimed tasks to pick up |\n\n### Platform parity\n\nThe **skills, agents, and commands** are portable — they ship to Claude Code, Codex, Cursor, and Gemini via the per-IDE manifests. The **hooks above are Claude Code-only**: they key on Claude Code's lifecycle events, so the automatic enforcement they provide does not run on the other harnesses.\n\n| Capability | Claude Code | Codex / Cursor / Gemini |\n|:-----------|:-----------:|:-----------------------:|\n| Skills, agents, commands | ✅ | ✅ (via manifest) |\n| 25 lifecycle hooks (permission-guard, secret-scan, audit-trail, …) | ✅ | ❌ |\n| Governance / audit trail | ✅ | ❌ |\n| Continuous learning (observe → instincts) | ✅ | ❌ |\n| Auto project-triage routing | ✅ | ❌ |\n\nOn non-Claude harnesses you get the full prompt library; the guardrails, governance, and learning loop are Claude Code-only today.\n\n---\n\n## Quality Pipeline\n\nForgeBee embeds quality checks throughout the development flow so that `/review-all` is a **validation gate**, not a discovery phase.\n\n```\nSpecialist agents implement + self-review → report DONE/DONE_WITH_CONCERNS/BLOCKED/NEEDS_CONTEXT\n      │\n      ▼\nCode debate (context:fork skills — blind advocate/skeptic/judge)\n      │\n      ▼\nWorkflow/Team quality gate (tests + lint + build must pass)\n      │\n      ▼\nreview-all (inline skill — runs in session context, only Critical/High block push)\n```\n\n**How it works:**\n\n- **Every command** has an Objective (what success looks like) and Never rules (hard boundaries)\n- **Every code-producing agent** has a Self-Review section matching review-all's criteria\n- **Agent status protocol** — agents report `DONE`, `DONE_WITH_CONCERNS`, `BLOCKED`, or `NEEDS_CONTEXT`; orchestrators handle each appropriately\n- **Review calibration** — only Critical/High issues block the push; Medium/Low are informational\n- **`/workflow` Work Breakdown** defaults to a lightweight Implementation Plan (ordered workstreams + file scope + agent + acceptance criteria) — `scrum-master` full sprint stories are opt-in via `/workflow --scrum`\n- **`/workflow` code debate** uses `context:fork` skills for blind isolation with less token overhead\n- **`/team` quality gate** invokes review-all as an inline skill (session context, ~93% fewer tokens)\n- **Instruction priority** — CLAUDE.md > Inline skills > Forked skills > Subagents > Defaults\n\nIf the pipeline works correctly, `review-all` finds nothing. Issues caught in review-all mean the pipeline leaked.\n\n---\n\n## Continuous Learning\n\nForgeBee learns from your sessions automatically. Every tool call is observed (PreToolUse + PostToolUse), patterns are detected by a heuristic engine, and candidates are surfaced for your approval.\n\n```\nSession Activity (every tool call)\n      │\n      ▼\n  observations.jsonl (automatic — Pre + Post)\n      │\n      ▼  Stop hook heuristics (no API calls)\n  pending-instincts.jsonl (auto-detected candidates)\n      │\n      ▼  /learn (review + approve/reject)\n  instincts/personal/ (project-scoped, active)\n      │\n      ▼  /evolve\n  evolved/ → skills, commands, agents\n```\n\n**Key concepts:**\n\n- **Two-stage pipeline** — heuristics flag candidates automatically; `/learn` lets you review and approve\n- **Pending instincts** — auto-detected patterns start in `pending` state, never activate without your approval\n- **Instincts** are atomic patterns: one trigger, one action, confidence-scored (0.3–0.9)\n- **Project-scoped** by default — React patterns stay in your React project, Python conventions in your Python project\n- **Global promotion** — when the same instinct appears in 2+ projects with ≥0.8 confidence, it's promoted to global\n- **Zero-config observation** — the `observe.js` hook captures every tool call silently (3s timeout, never blocks)\n- **Portable** — export/import instincts across machines and teammates\n\nStorage lives at `~/.claude/forgebee-learning/` with per-project isolation via git remote URL hashing.\n\n---\n\n## Growth OS\n\nA 9-phase marketing pipeline with 18 specialist agents, invoked with `/growth`:\n\n```\nBrand Foundation ──► Market Intelligence ──► Content Architecture\n        │                    │                       │\n        ▼                    ▼                       ▼\n   Hook & Idea Engine ──► Strategy Debate ──► Execution Plan\n        │                    │                       │\n        ▼                    ▼                       ▼\n  Content Production ──► Distribution & CRO ──► Measure & Optimize\n```\n\nThe strategy debate phase uses blind Advocate/Skeptic/Judge review — the same pattern as `/workflow`'s code debate, applied to marketing strategy. Catches weak positioning before you invest in execution.\n\nIncludes dedicated agents for CRO (conversion rate optimization), email marketing automation, and competitive intelligence.\n\n---\n\n## Project Management\n\nForgeBee tracks features across sessions with a hybrid YAML + Markdown system:\n\n```\ndocs/pm/\n├── state.yaml        ← Machine-readable project state\n├── index.md          ← Auto-generated dashboard\n├── decisions.md      ← Decision log from all debates\n└── features/         ← Per-feature detail pages\n```\n\nEvery `/workflow`, `/growth`, `/idea`, and `/plan` run reads and writes `state.yaml` at each phase transition. The `pm-sync` hook loads state on session start. Run `/pm` anytime for a full status report.\n\n---\n\n## How It Works\n\nForgeBee is **markdown files and Node.js scripts**. No runtime dependencies, no build step, no bash, no python.\n\n- **Commands** are `.md` files — structured prompts that activate workflows\n- **Agents** are `.md` files — specialist definitions with trigger conditions\n- **Skills** are directories with `SKILL.md` + scripts — auto-triggered capabilities\n- **Hooks** are `.js` scripts wired to Claude Code lifecycle events via `hooks.json`\n- **Audit trail** is an append-only JSONL log — every permission, debate, and verification is recorded\n\nClaude Code reads them natively. The plugin system handles distribution.\n\n<details>\n<summary><strong>Full installed structure</strong></summary>\n\n```\nyour-project/\n├── CLAUDE.md                          # Project memory\n├── TASKS.md                           # Auto-managed task tracking\n├── docs/\n│   ├── pm/                            # Project management\n│   │   ├── state.yaml\n│   │   ├── index.md\n│   │   ├── decisions.md\n│   │   └── features/\n│   ├── marketing/                     # Growth OS outputs\n│   └── planning/                      # Planning artifacts\n│       ├── briefs/\n│       ├── requirements/\n│       └── stories/\n└── .claude/\n    ├── settings.json\n    ├── sessions/                      # Session snapshots\n    ├── session-cache/                 # Permissions + skill manifest\n    ├── audit/                         # Governance audit trail (JSONL)\n    └── learnings/\n        ├── learnings.md              # Auto-captured patterns\n        └── pending-instincts.jsonl   # Heuristic-detected candidates\n\n~/.claude/forgebee-learning/           # Continuous learning (persists globally)\n├── projects.json                      # Project registry\n├── instincts/personal/                # Global learned instincts\n└── projects/<hash>/                   # Per-project isolation\n    ├── observations.jsonl             # Tool call observations\n    └── instincts/personal/            # Project-scoped instincts\n```\n\n</details>\n\n---\n\n## CLAUDE.md &mdash; Project Memory\n\nThe installed `CLAUDE.md` is a structured template Claude reads at the start of every session. Fill in your stack, conventions, key components, and team contacts.\n\nThe `self-improve` hook appends patterns to the **Learned Patterns** section automatically. The continuous learning system goes further — observing every tool call and building project-scoped instincts with confidence scoring that persist across sessions.\n\n---\n\n## OpenClaw\n\nForgeBee is fully compatible with [OpenClaw](https://github.com/openclaw/openclaw). All 45 agents and 38 commands convert to OpenClaw skills.\n\n```bash\n# Clone ForgeBee\ngit clone git@github.com:forbee-dev/ClaudeDevKit.git\n\n# Install for OpenClaw\nnode ClaudeDevKit/openclaw/install-openclaw.js\n```\n\nThis converts every agent and command into `SKILL.md` files in `~/.openclaw/workspace/skills/forgebee-*`. Skills auto-trigger based on their descriptions — the same ones optimized from the [Superpowers trigger pattern](#why-forgebee).\n\n<details>\n<summary><strong>How the conversion works</strong></summary>\n\n- Each agent becomes `forgebee-<agent-name>/SKILL.md`\n- Each command becomes `forgebee-cmd-<command-name>/SKILL.md` (user-invocable)\n- YAML frontmatter maps: `name` + `description` + `metadata.openclaw`\n- Agent body instructions transfer directly (markdown is markdown)\n- Model preferences are noted but OpenClaw handles model selection through its own config\n\n</details>\n\n---\n\n## Contributing\n\nContributions welcome! ForgeBee is markdown surfaces and Node.js scripts. See **[CONTRIBUTING.md](./CONTRIBUTING.md)** for the surface contracts (agent/skill/command structure, the load-bearing Status + adversarial-hardening contract, the model-tier policy, output modes) and **[ARCHITECTURE.md](./ARCHITECTURE.md)** for how the pieces fit together.\n\n1. Fork the repo\n2. Add your command (`forgebee/commands/your-command.md`), agent (`forgebee/agents/your-agent.md`), or skill (`forgebee/skills/your-skill/SKILL.md`) — follow the contract in CONTRIBUTING.md\n3. Run `node scripts/build-index.js` to regenerate the routing index\n4. Run `npm run check` — index sync + persona references + agent contract + version sync + eval must all pass (CI gates the PR on this)\n5. If you added/removed a surface, update the count strings in the manifests + READMEs (versions sync via `bump-version.sh`; counts are manual). Cut releases with `/release`.\n6. Open a PR\n\n**Release notes:** see [`CHANGELOG.md`](./CHANGELOG.md). The `v5.1.0` entry is the canonical reference for all the discipline patterns (Karpathy P1-P6, prompt defense, budget circuit breaker, terse-report, etc.).\n\n---\n\n## License\n\n[MIT](LICENSE) &mdash; use it however you want.\n\n<p align=\"center\">\n  <sub>Built with Claude Code by <a href=\"https://github.com/forbee-dev\">Forbee Dev</a> at <a href=\"https://raketech.com\">Raketech</a></sub>\n</p>\n",
  "bytes": 33698,
  "sha": "4cc9275aff307e0f2d4f89a173b4fc339164366444e4aa7a1e6d1da355c7edb0",
  "repo_slug": "forbee-dev/forgebee",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_forbee_dev_forgebee_f95c8b05/readme"
}