One plugin to bridge and delegate across Claude Code, Codex CLI, and Antigravity — single-source AGENTS.md, shared skills, mirrored hooks and MCP servers, full Claude↔Codex bidirectional delegation (including Codex reading Claude's session history), a Claude→Grok Build delegation lane over ACP, a bounded read-only Claude→Qwen Code review lane (Safe Mode + Plan mode + sandbox, isolated target copies, verified tool boundary), project-scoped advisor agents, opt-in bridging to more coding agents (Grok Build, opencode, Qwen Code, Kimi CLI) via a tool-profile registry, never-stale model selection (the config records the 'latest' policy), a structured diagnostics engine (enabled-tools-aware health checks with verified auto-repair), and a cross-repo sweep that finds every project on the machine cc-suite has bridged and repairs bridge leftovers in one pass — dead MCP registrations, and skills symlinks left pointing at a pruned plugin version, which SessionStart now also re-points on its own. Supersedes cc-bridge and codex-toolkit.
13 skills · 46 commands · hooks
Enforces strict TDD discipline across unit/integration/e2e/contract lanes — specification-strength levels S1-S6 with property tooling, adversarial spec review before implementation, red receipts proving spec/implementation separation, per-path critical thresholds, 9-format coverage merging, and mutation testing
32 skills · 16 commands · 7 agents · hooks
Documentation freshness, accuracy and coverage checks, plus commit and push guards that resolve the target repository and map outgoing source changes to documentation.
35 skills · 5 commands · 5 agents · hooks
Natural-Language Programming Manager — score, check, fix, and test NL artifacts across Claude Code, Codex CLI, and Antigravity. Tier-aware scoring with per-tool overlays. v1.0.0 marks the multi-tool branded release: nlpm:conventions splits into universal floor + per-tool overlays (conventions-claude, conventions-codex, conventions-antigravity); the scoring rubric splits Hooks per tool (the three event vocabularies are not 1:1 mappable); new artifact rows for .codex-plugin/plugin.json, .agents/plugins/marketplace.json, agents/openai.yaml sidecars, gemini-extension.json, LSP, monitors. Standalone Python validator (bin/nlpm-check) remains tool-agnostic and Claude-Code-free for pre-commit hooks and CI.
17 skills · 15 commands · 8 agents · hooks
English coaching with configurable on-demand and sampled modes, bounded provider calls, literal-preservation checks and progress reports; the original user prompt remains authoritative.
23 skills · 10 commands · 4 agents · hooks
Personal IT guy for non-technical macOS users — 12 plain-language commands, 4 agents, observation-first onboarding, a machine profile that retires its own stale beliefs by retesting them, a locked and transactional state writer, a PreToolUse guard that blocks destructive shell commands, a toolbox that turns repeated chores into reusable micro-tools and offers automations you did not know to ask for, learning maps built from your own measurements, and answers in your chosen language while everything on disk stays English. Ships a Codex layout with the same hooks. macOS only. 129 assertions.
8 skills · 12 commands · 4 agents · hooks
Turn AI sessions into a maintained, human-reviewed knowledge base: capture sessions to append-only logbooks, compile into consistency-checked cabinet dossiers gated by trust tiers (proposed → verified → canonical), and render an offline gazette (built by the bundled press). A typed review queue (bureau:review) orders the backlog upstream-first and names the exact action per page (approve / reapprove / confirm / resolve / repair); a human applies decisions one at a time or as a reviewed, content-bound, commit-gated batch (approve --from a manifest, or --all with a warning). Record architecture decisions as MADR ADR pages (bureau:adr) with typed supersedes edges — a superseded decision leaves current fact only once the superseding ADR is content-bound approved (ADR-0006). Delegate the review queue to Codex as your representative (bureau:codex-review): it recommends approve/hold, and where a workspace opts the codex authority into trust_policy it commits under --by codex — an honest machine authority, never a human one (ADR-0007). Bulk-approve the whole backlog in one confirmation (bureau:approve-all): the AI shows the batch and confirms, then hands you the single --by human line to fire yourself — it never signs for you. A deterministic recursion engine tracks claim-level dependencies (rests_on + author-anchored spans) via an append-only decision log, per-span revisions, an early-cutoff gate, and an fsck byte-fixpoint — so changing an upstream claim flags every downstream page for review. A live Engine view (bureau:serve) paints three signals onto the board as you edit — dependency freshness (needs-review/stale), artifact currency (a verified file that drifted out from under a claim), and a convergence trend (is the canon settling or thrashing) — and git-backed versioning renders any past board (build --at), diffs two versions, and pins named snapshots. Enable a crew of specialized agents or author your own. Karpathy LLM-wiki pattern + session provenance + a human rev
4 skills · 18 commands · hooks
Mine past Claude Code conversations for values, decisions, mistakes, patterns, and wisdom. Manage memory lifecycle: audit staleness, extract knowledge, prune waste.
4 skills · 8 commands · 5 agents
Recursive governance for Executable Operating Units (EOUs). Faceted classification, generating-EOU constraints, ECP-governed change, no-self-approval. v0.7.0: Stage 0 (reference-grounded captured_workflow + per-app domain_values constitutional layer + Rule 96). v0.8.0: agentic judgment (judgment_authorized flag, value_invocations trace, F14–F17 taxonomy, judgment_maturity J0–J4 axis, $audit-judgment skill, Rule 97, counterfactual-swap audit).
14 skills · 1 command
Deep codebase interrogation: a bounded three-finding quick review by default, with explicit full multi-agent reviews, five review styles and eight pressure tests.
8 skills · 1 command · 6 agents
Deep argument interrogation: a bounded quick review by default, with an explicit full adversarial panel and evidence-grounded verdicts.
8 skills · 1 command · 6 agents
Enforce per-file pure LOC limits — deterministic tokei-based counting, per-file ceilings and exemptions for files that are one thing by design, opus-powered extraction plans measured in pure LOC, and a --check mode that exits non-zero for CI and pre-commit hooks
8 skills · 2 commands · 2 agents
Block hardcoded UI values from LLM-written code — rewrite-first PreToolUse hook silently corrects literals to design-token references on the way to disk; uncertain matches deny with structured suggestions (strict mode, default) or surface as PostToolUse advisories (advisory mode); per-project surfaces allowlist narrows which file kinds the hook scans, plus an ignore glob list excludes paths (test fixtures, generated or vendored code) from hooks and audit alike; /ui-tokenize:review dispatches the token-reviewer agent to flag semantic mis-picks (right token vs. just any token); tokenize__* MCP tools for autonomous recovery
7 commands · 1 agent · 1 MCP server · hooks
Compute routing with guarded effort requests, a measured cheap-model worker and reports of observed main-agent and subagent effort; runtime escalation must be verified.
3 skills · 1 agent · hooks
Advisory responsive-design coach — flags off-catalog breakpoints, bare 100vh, fixed widths without max-width, and stylesheets with no @media; surfaces findings as PostToolUse additionalContext, never blocks tool calls
3 commands · 1 MCP server · hooks
A lean, self-measuring dev office for Claude Code and Codex CLI: three short agents (planner, implementer, verifier), four skills (adr, release, repo-init, tidy), perception hooks, test tiers, git routine and tidy, all host-neutral behind a single dispatcher with a four-layer config resolver (defaults, user, repo, repo-local).
4 skills
XROS — an eXecutable Research Operating System. Four commands: compile a long-horizon investigation into a verifiable methodology spec; sharpen an un-checkable question (orient a newcomer in a field's vocabulary → falsifiable claim → best-available check, with its ceiling); run the Tier-A/B multi-agent engine (diverse independent routes, adversarial refutation) gated on a real oracle's exit code; or reason through a no-oracle Tier-C claim with premises, pre-mortem, and dated tripwires. Founding rule: no verification, no claim. Ships a bundled dependency-free validator and conformance suite.
4 skills
A 260-token system prompt that overrides three structural presumptions every RLHF-trained LLM inherits from training: that you want confirmation, that old scarcity still applies, that best practices are ceilings.
1 command · 1 agent
Mermaid diagrams drawn inside Claude Code replies (optional Ghostty/kitty images, offline browser view), plus a truecolor status line under the prompt: git, model, context with a Token Weather forecast, cost and rate limits.