agentic

My personal AI coding agent configuration, with any quirks, oddities, opinionated rules, and hallucination-inducing prompt fragments I live with on a daily basis.

One repo, one base rule set, rendered into per-agent global instruction files. Supports Claude Code, Codex, and any tool that reads AGENTS.md.

git clone https://github.com/jimeh/agentic.git ~/.config/agentic
cd ~/.config/agentic
mise run setup                  # install dependencies and local git hooks
mise run agent-config:install   # create symlinks (skips existing)
# mise run agent-config:force   # replace existing (backs up to .bak)

This creates symlinks from the repo into ~/.claude/, ~/.agents/, and ~/.codex/, and registers configured plugin marketplaces via the Claude CLI. Use the mise run agent-config:* tasks for installation.

Install only project dependencies with:

mise run deps:install

Git hooks are managed by Lefthook and installed by mise run setup. To install or refresh hooks directly:

mise run hooks:install

• rules/ — Source Markdown for global behavior rules. Each render target declares itself with type: agentic-rules frontmatter and composes its content with <!-- include: path --> directives. See rules/README.md.
• generated/ — Rendered global rule files (CLAUDE.md, CODEX.md, OPENCODE.md). These are symlinked into agent config directories; edit rules/ and run mise run rules:build instead of editing generated files directly.
• agent-config.toml — Installer config for fixed symlinks, skill symlink roots, stale-link cleanup paths, Claude plugin marketplaces, and auto-installed Claude plugins. It points editors at schemas/agent-config.schema.json; JSON and YAML config files are also supported.
• claude/ — Claude Code settings and statusline script.
• codex/ — OpenAI Codex config.
• skills/ — Custom skills (auto-discovered by the agent config installer).
• thirdparty/ — Vendored third-party skills plus manifest and lock metadata.
• plugins/ — Source retained for deprecated Claude Code plugins.
• packages/agent-config/ — Config installer, schema generator, and the compatibility CLI used by Mise tasks.
• packages/agent-rules/ — Global rule renderer and drift checker.
• packages/agent-harness/ — Agent metadata checks and executable skill and plugin test runner.
• packages/agent-headless/ — Shared core for the streamed headless CLI runners: artifacts, stream handling, progress, and exit policy.
• packages/claude-headless/ — Streamed Claude CLI runner with private run artifacts and package-owned integration tests.
• packages/codex-headless/ — Streamed Codex CLI runner with the same artifact layout, sandbox and session handling, and integration tests.
• packages/agent-pr-monitor/ — Read-only GitHub PR snapshots and quiet waiting, with persisted change detection and compact results for agents.
• packages/agent-judge/ — Typed TypeSafe evaluations through mise run judge, with private optional artifacts and a bounded history reranking experiment.
• packages/vendor-skills/ — Reviewed third-party skill intake and update tooling.
• docs/references/ — External articles and guides.

Skills are auto-discovered — drop a directory in the right place, re-run the agent config installer, done. Vendored third-party skills under thirdparty/skills/ are installed the same way, but only the explicit update task fetches from upstream. No plugins are currently published through the local Claude marketplace; the remaining plugin sources are retained but deprecated.

Bounded codex-* and claude-* analysis, implementation, and review skills are available to both engines. They start workers through the headless runners, or as T3 Code child tasks through t3-delegation when the session runs inside T3 Code. implement-feature owns implementation orchestration for both engines. react-high-performance and the vendored unslop and web-design-guidelines skills are excluded from installation; shared writing guidance lives in rules/base.md. See agent instructions for selection and worker boundaries.

Use mise run setup to bootstrap a delivery worktree, but do not run agent-config install there. Deliver configuration changes through a PR; installer tests use synthetic roots and homes.

Update vendored third-party skills with:

mise run thirdparty:add-skills -- vercel-labs/agent-skills
mise run thirdparty:add-skills -- vercel-labs/agent-skills --skill skill-name
mise run thirdparty:update-skills
mise run thirdparty:update-skills:dry-run
mise run thirdparty:update-skills:check

Those tasks call the repo-local vendor-skills CLI through Bun. Its implementation and colocated tests live under packages/vendor-skills/src/. The add command accepts full git URLs or GitHub owner/repo shorthand and opens a multi-select prompt when --skill is not provided, then vendors selected skills and updates the lockfile. --skill matches a skill's frontmatter name or its upstream directory name. Individual manifest skill entries can set ref to pin or test a skill separately from the source default, and "overrides": { "explicitOnly": true } to vendor a skill that loads only when the user invokes it: /name in Claude, or an explicit mention in Codex.

Render global instruction files after changing rules/:

mise run rules:build
mise run rules:check

Update pinned GitHub Actions with Pinact:

mise run actions:update

New worktrees can be bootstrapped with Treeboot:

mise run treeboot
mise run treeboot:check

Run local tests with:

mise run test
mise run test:plugins
mise run test:skills
mise run test:claude-headless
mise run test:codex-headless
mise run test:pr-monitor

Format Markdown and TypeScript with:

mise run format
mise run format:check

Mise-managed tools are locked in mise.lock. Bun stays on the 1.3 release line; other tools stay on their current major release lines with Mise's three-day minimum_release_age. Refresh tool lock metadata after changing mise.toml with:

mise lock --minimum-release-age 3d

No plugins are currently supported or published by this repository. Sources under plugins/ are deprecated and retained for reference; they are not available from the marketplace.

A PreToolUse hook that strips redundant git -C <cwd> flags from Bash commands when the path matches the current working directory. Claude Code tends to add these unnecessarily, and the -C flag changes the command string enough that pre-approved git commands no longer match the allowlist — causing repeated permission prompts.

Handles all -C syntax variants (space, =, bare, quoted) and compound commands (&&, ;).

Slash commands for common git workflows:

• /commit — Stage changes, create a well-formed commit.
• /commit-push-pr — Commit, push, and open a PR.
• /rebase — Rebase onto upstream main/master.
• /clean-gone-branches — Clean up branches deleted on remote.

Derived from the official commit-commands plugin, heavily modified.

Slash commands for managing AGENTS.md files:

• /claude-md-to-agents-md — Convert CLAUDE.md to AGENTS.md.
• /generate-agents-md — Generate AGENTS.md from codebase analysis.
• /refactor-agents-md — Refactor AGENTS.md for progressive disclosure.

A disciplined research-plan-implement workflow. Instead of jumping straight to code, you move through distinct phases so every decision is reviewed before implementation begins.

• /research — Deep-read a codebase area, write findings to research.md.
• /plan — Create a detailed implementation plan in plan.md.
• /refine — Address inline notes you've added to the plan.
• /todo — Add a granular task breakdown to the plan.
• /implement — Execute the plan mechanically, marking progress.
• /review — Optional post-implementation sanity check.

Also includes agent-agnostic prompt snippets for use with any AI coding assistant.

Initializes a lightweight Feature Design (FD) tracking system in any project. Scaffolds directory structure, templates, index, project-local slash commands, and CLAUDE.md conventions for the full FD lifecycle.

• /fd-init — Set up the FD system in the current project.

Based on the Feature Design system by manuelschipper.

agent-pr-monitor reads GitHub through Octokit and waits without model calls. The installer links it into ~/.local/bin/; from this checkout, use the Mise task directly:

mise run pr-monitor -- snapshot https://github.com/OWNER/REPO/pull/123
mise run pr-monitor -- wait https://github.com/OWNER/REPO/pull/123

The first snapshot establishes a durable baseline. Subsequent waits stay quiet until the head changes, a check fails, checks complete, another account leaves or edits feedback, a thread's state changes, the branch starts to conflict, or the PR closes, merges, or reopens. The authenticated account's own comments and reviews do not wake a wait. Each poll is a single GraphQL request, once a minute by default, and --initial-delay postpones the first poll when nothing can happen yet. Each invocation prints one JSON result and saves private artifacts. It never writes to GitHub or decides whether a PR is ready to merge. Run agent-pr-monitor --help for authentication, timeouts, and state paths. The babysit-pr monitoring reference describes agent execution and result handling.

Goal-aware probes evaluate selected conditions once or wait for all of them:

mise run pr-monitor -- evaluate "$PR_URL" --until checks-pass --checks required
mise run pr-monitor -- wait "$PR_URL" --until checks-pass \
  --until review-approved --reviewer coderabbitai
mise run pr-monitor -- evaluate "$PR_URL" --until review-finished \
  --reviewer coderabbitai

evaluate returns satisfied: true, false, or null (unknown), with evidence for each condition. Goal calls leave the change cursor untouched. Review conditions use submitted GitHub review states and current-head metadata. See the goal evaluator reference for condition semantics, exit codes, freshness, and evidence limits.

packages/agent-config exposes the agent-config CLI used by the Mise tasks. The CLI keeps the existing command contract while delegating rule and harness commands to their workspace packages. Its install command reads agent-config.toml to create fixed symlinks, discover configured skill roots, clean stale managed symlinks, register Claude plugin marketplaces, and install configured Claude plugins. To add or remove auto-installed plugins, edit agent-config.toml. The loader checks config files in this order: agent-config.toml, agent-config.yaml, agent-config.yml, then agent-config.json.

Config source paths are repo-relative. Home-side target paths must be explicit and start with ~/, including symlinks[].target, skillSymlinks[].targetRoots[], and staleSymlinkCleanup[].targetDir.

The JSON schema at schemas/agent-config.schema.json is generated from the package source. Run mise run agent-config:schema:build after changing the schema, or mise run agent-config:schema:check to verify it is current.

Requires the claude CLI. Skipped gracefully if it is missing.

• Bun 1.3+

This is a personal configuration repository. Feel free to use it as inspiration for your own setup, but there are no guarantees it won't teach your agents some questionable habits.