LLM Mart Basic

The terminal step of /feature and /sprint. Routed to once commit-gate has cleared, to decide the branch's fate — merge via PR, open a PR, or discard. Direct merge to the default branch is forbidden; every change lands through a PR. Under /sprint the skill auto-selects "open PR" a

Finish an already-merged branch. Proves the branch is contained in the fetched default, classifies leftover artifacts as unique / redundant / superseded, returns to a clean --ff-only default checkout, and deletes the merged local branch — every discard confirmed per item. Routed

The behavior-preserving restructure gate. Routed to by /refactor for a rename, extract, inline, move, dedup, or internal-implementation swap. Six gated phases prove behavioral parity through unmodified pre-existing tests; any diff that classifies as `feat` is not a refactor and i

The single permitted path to a version tag, for any target declared in the project's .codearbiter/release-targets.md. Routed to when the user invokes /release on a non-default branch with a green suite. Takes the declared target as its one argument, derives the SemVer bump from C

The secret-source gate. Routed to when changed code reads, writes, or passes a secret — API key, token, password, connection string, signing key, certificate, or any value that grants access. Validates that every secret comes from the approved store and never lands in source, log

Optional, opt-in STRIDE threat pass for a sensitive feature — invoked deliberately via /threat-model, never forced on ordinary changes. Walks the change's attack surface and security boundaries (governed by ${CLAUDE_PROJECT_DIR}/.codearbiter/security-controls.md), surfaces threat

The test-first gate. Routed to by /feature (after the spec is approved), /fix, and /refactor before any implementation code is written. Six gated phases — obligation scan, red, green, obligation verify, coverage, lint. No feature code exists before Phase 1 clears; nothing reaches

The deep, rarely-convened whole-codebase audit lane. Routed to when the user invokes /ca:tribunal. Seven gated phases — cost/model, map, roster dispatch, triage, report, approval+filing, telemetry. Costs on the order of millions of tokens; proceeds only after the user acknowledge

OPTIONAL per-task isolation for autonomous parallel work. Routed to only on explicit opt-in by subagent-driven-development or dispatching-parallel-agents, so parallel units mutate files without colliding. Stands up one worktree per unit, works it in isolation, then integrates eac

The spec-to-plan bridge. Routed to by /feature once the brainstormed spec is approved, and by /sprint before execution. Decomposes the spec into 2–5 minute tasks, each carrying its exact file path(s) and a concrete verification step that maps to a tdd obligation. Writes the plan

Run Claude Code INSIDE a ca-sandbox box (`--with-claude`). Routed to when the user wants an agent loop running against an isolated, ephemeral sandbox rather than the host. Authenticates via an env-injected CLAUDE_CODE_OAUTH_TOKEN with no host bind of ~/.claude; the image pins the

The lifecycle gate for a local Codespace-equivalent sandbox. Routed to when the user invokes /ca-sandbox:sandbox to pull an untrusted repo into an ephemeral, host-FS-isolated Docker container, or any of the interaction commands (/ca-sandbox:sandbox-shell, /ca-sandbox:sandbox-exec

The Socratic spec-refinement front of /feature, and the planning front of /sprint. Routed to BEFORE any code — it takes a one-line idea and drives it to an approved, concrete spec with testable acceptance criteria. Five gated phases — frame, shape, refine, write, review-and-appro

The only path to a commit. Routed to when the user invokes /commit or otherwise instructs codeArbiter to persist staged changes. Nine gated phases — permission, branch, classification, verification (test/lint/secrets), behavioral proof, diff review, selective stage, message, comm

Optional manual drift audit — report stale provenance-tracked docs (via _provenancelib drift detection across .codearbiter/.provenance/), then per stale doc offer re-scout / re-baseline / defer. Not the daily loop; commit-gate auto-heal owns routine maintenance.

The brownfield back-fill. Routed to by /create-context, and by startup when .codearbiter/CONTEXT.md lacks the <!--INITIALIZED--> body marker but source code exists. Six gated phases — pre-flight, scout dispatch, synthesis, gap interview, write, lock. Reads the existing codebase t

The banned-primitive gate. Routed to when changed code hashes, signs, encrypts, derives keys, generates security-relevant randomness, configures TLS, or imports a crypto library. Rejects broken primitives, disabled TLS verification, and home-rolled crypto; the approved-primitive

Investigate-then-decide root-cause analysis for a defect whose cause is unknown (distinct from /fix, which assumes a known bug). Five gated phases: capture, hypothesize, gather, decide, hand off. Investigation only, no code changes; exits to /fix, /adr, or a no-action close.

Author and track Architecture Decision Records. Routed to when the user invokes /adr to record a new decision or /adr-status to list ADR health. Authors numbered, dated, user-attributed ADRs under .codearbiter/decisions/, maintains supersede chains, and reports status read-only.

Reconcile the project's architectural artifacts against the scaffold and prior decisions, then present each variance as a SMARTS analysis for the user to decide. Routed to when the user asks to arbitrate, reconcile, or consolidate architectural context, requests a variance report

Fourteen posts of being wrong in production, compressed to checkboxes

Healthy nodes, a quiet network, 300 restarts in three days, and a latency budget measured in milliseconds

Discovery worked. Ping worked. Every TCP connection timed out, and later the tunnel only worked when someone had a terminal open.

Every VM came back. The cluster did not. Declarative systems converge on config, and the datapath isn't config.

A surprising share of AI-in-the-terminal failures aren't the AI. They're zsh, and a version of bash from 2006.

A Claude Code plugin turns standalone project configuration into a namespaced, installable extension that teams and communities can update as one unit.

None of the safety came from the model. It came from six boring habits.

Skills package instructions and references. Subagents run work in a separate context and return results. They solve different problems and can be composed deliberately.

Six hours in, one step left, everything green, and the incident that didn't happen

CLAUDE.md carries persistent project context. Skills load reusable procedures when relevant. Separating stable facts from task-specific workflows keeps both easier to maintain.

Twenty minutes recovering secrets that never existed, and the one sentence from a human that ended it

An API request routing a model's tool call through an approval gate to a remote MCP server

31 config keys, two audits, and why the first one was wrong in both directions

The official MCP Registry stores standardized server metadata rather than package code. Publishers verify a namespace, describe installation or remote access, and submit immutable versions.

Everyone looks at the Dockerfile. The file that actually leaked the key was the project file.

Remote MCP authorization uses established OAuth standards, but secure integration still requires issuer validation, least-privilege scopes, protected token handling, and server-side enforcement.

"Copy it over and switch the reference" is two steps, and the outage lives in the one nobody checks

stdio fits local processes and prototypes. Streamable HTTP fits hosted services and shared integrations. The right choice follows where the capability runs and who must reach it.

The most important rule wasn't about what I could change. It was about what I was allowed to display.

Tools perform operations, resources expose readable context, and prompts provide reusable templates. Choosing the correct primitive makes an MCP server easier to understand and govern.

`crabbox attach` follows the recorded events of an active coordinator run and

`crabbox azure` groups Azure provider setup commands. It currently has a single

`crabbox bench` records and reports local benchmark timing observations. It is a

`crabbox cache` inspects, purges, or warms package and build caches on a

`crabbox capsule` captures, replays, and tracks lightweight failure capsules.

Save the state of a lease, then restore it onto another box or fork it into a

`crabbox claims list` prints the lease claims stored on the current machine. It

`crabbox cleanup` sweeps direct-provider machines and local provider state that

`crabbox code` bridges a Linux lease's `code-server` workspace into the

`crabbox config` inspects and updates user configuration. It has three

`crabbox connect` resolves a lease and opens an interactive SSH session to it.

`crabbox cp` copies files or directories between the host and a Crabbox-owned

`crabbox desktop` drives a visible desktop session on a lease that was warmed

`crabbox doctor` runs a preflight before you commit to a long workflow. It is

`crabbox egress` gives a lease mediated outbound network: a lease-local browser

`crabbox events` prints the broker's event log for a recorded run.

`crabbox heartbeat` refreshes the idle deadline for one owned lease and prints

`crabbox history` lists recorded remote command runs from the broker. Each run is

`crabbox image` holds the trusted-operator controls for provider base images:

`crabbox init` onboards the current repository: it writes the minimal config

Your car as a chat-room agent: Raspberry Pi 5 + dashcam + local AI. CodeWatch's sibling for the garage.

The enterprise gateway for autonomous agents. Identity management, per-channel isolation, credential vault, per-session tamper-evident audit log.

AI agent for Home Assistant — talk to your home, create automations in plain language, analyze cameras with face recognition, and get proactive anomaly alerts.…

A desktop app to prototype agent ideas, inspect every harness step, replay failures, and evaluate performance, all in one place. Local-first, cloud-ready for ma…

Local-first AI memory you can see, edit, and override — portable across Claude Code, Codex, Cursor, Windsurf, and other MCP coding tools.

Web & desktop (Electron) client for the OCTO open workplace — one React + TypeScript codebase shipping browser and PC surfaces, with first-class AI agent UX.

同花顺官方 A股金融数据服务，提供股票实时行情、历史行情、财务报表、指数、板块、涨停等数据，适用于 AI Agent、量化研究和应用开发，支持 API、MCP、CLI 和 Python。Official Tonghuashun (HiThink) A-share financial data service provi…

Open-source desktop SQL workspace for PostgreSQL, MySQL/MariaDB, SQLite and 15+ more databases like DuckDB, ClickHouse, Redis and Firestore. Built-in MCP server…

🤖 Local-assist BOSS Zhipin CLI for AI agents — search, welfare filtering, shortlist, JSON-envelope output; low-risk & compliant by default.

The local-first sidebar AI agent for ComfyUI — runs on your own Claude OR ChatGPT subscription (no API keys, no extra LLM costs). Drives your live graph: edits,…

Persistent memory for Claude Code — identity, context, and continuity across sessions

🔬 A curated collection of 23,000+ agent skills for empirical research across 8 social science disciplines. | 精选 23,000+ AI Agent 技能库，覆盖8大社会科学学科的实证研究。CoPaper.AI…

A curated list of plugins, skills, MCP servers, patch/profile layers, orchestrators & UIs for DeepSeek Harness (DSH). Visualization · PPT · Coding · Agents · Lo…

World Memory Protocol.

Run your own organization of agents.

The most RAM efficient harness

AI-native, free, open-source alternative to Jira, Trello, ClickUp & Monday. Built for Scrum teams where humans and AI agents collaborate as equals — on the same…

Free open-source desktop app that scrapes JAV metadata and generates NFO + cover art for Jellyfin, Emby & Kodi. No Docker, no CLI — one-click install on Windows…

A cross-platform desktop app to manage Agent Skills in one place and sync them to multiple AI coding tools’ global skills directories — “Install once, sync ever…

天枢 (Tianshu) 是一个基于harness工程的终端编程智能体运行时（Tui X Gui），针对DeepSeek V4 做了前缀缓存工程优化（长会话实测稳态命中率 97–99%）和深度适配。它跳出了传统 AI 编程助手把大模型仅当成“工具”的局限，基于认知虚拟机 (CVM)、自感知层和信息素（Stigmergy…