agentkit

Reusable agents, skills, scripts, and configuration that turn an LLM-assisted IDE into a multi-repo, multi-agent dev workflow — with durable session memory, daily standup briefs, parallel-agent isolation, and PR review/author voices distilled from senior maintainers.

This is not a framework or an extension. It's a small, hostable set of markdown prompts and POSIX shell scripts that you drop into your existing setup (Claude Code, Cursor, Continue, Copilot Chat — anything that reads system prompts).

If you've ever ended a day with nine open chat sessions across five worktrees and three repos, then come back the next morning unsure what you were doing — this toolkit is the answer to "fix that."

What's inside

Pillar	Where	What it solves
1. Worktree isolation	scripts/wt + agents/worktree-setup.md + memories/worktree-workflow.md	Two agents on one repo never write to the same working tree. The agent gates editing tasks behind a worktree decision; wt is the helper.
2. Daily session digest	scripts/session-log-compile.sh + systemd/	A nightly compile of every chat transcript and git activity into a single Markdown file under ~/.local/state/session-logs/. Per-session: UID, workspace hash, primary repo, msg count, first/last prompt, last edit.
3. Daily Brief agent	agents/daily-brief.md	Reads the latest digest + live gh state and produces a prioritized standup: CHANGES → CI-FAIL → CONFLICT → REVIEW → READY → WIP → DRAFT → STALE. Read-only.
4. PR Reviewer voice	agents/pr-reviewer.md + skills/distributed-systems-pr-review/ + memories/pr-review-techniques.md	Seven concrete review moves (trace the flow, future fragility, lock scope, test-as-no-op, silent-fallback hazard, error formatting, platform limits) distilled from senior maintainer history.
5. PR Author voice	agents/pr-author.md + skills/distributed-systems-author-style/	Upstream-ready titles, commit messages, and PR bodies in a distilled-maintainer style; pre-open self-review. Suggests git/gh commands; doesn't run them.

Quick start

git clone https://github.com/jaydipgabani/agentkit.git
cd agentkit
# Read the install steps for your host:
$EDITOR INSTALL.md

The TL;DR for Claude Code on Linux:

# Agents and skills
mkdir -p ~/.claude/agents ~/.claude/skills
cp agents/*.md          ~/.claude/agents/
cp -r skills/*/         ~/.claude/skills/

# Scripts
install -Dm755 scripts/wt                    ~/.local/bin/wt
install -Dm755 scripts/session-log-compile.sh ~/.local/bin/session-log-compile.sh

# Repo list (edit before first run!)
mkdir -p ~/.config/agentkit
cp examples/repos.txt ~/.config/agentkit/repos.txt
$EDITOR ~/.config/agentkit/repos.txt

# Daily compile via systemd
mkdir -p ~/.config/systemd/user
install -m644 systemd/*.{service,timer} ~/.config/systemd/user/
systemctl --user daemon-reload
systemctl --user enable --now session-log-compile.timer

# Smoke test
~/.local/bin/wt list
~/.local/bin/session-log-compile.sh
ls ~/.local/state/session-logs/

Full host-specific instructions (Claude Code, Cursor, Continue, Codex, non-systemd setups) in INSTALL.md.

Repo layout

agentkit/
├─ agents/                     # general-purpose sub-agent prompts
│  ├─ daily-brief.md
│  ├─ pr-author.md
│  ├─ pr-reviewer.md
│  ├─ session-log.md
│  └─ worktree-setup.md
├─ skills/                     # general-purpose on-demand skill bundles
│  ├─ distributed-systems-pr-review/
│  ├─ distributed-systems-author-style/
│  └─ distributed-systems-security-hardening/
├─ scripts/
│  ├─ wt                       # git-worktree helper
│  └─ session-log-compile.sh   # daily digest compiler
├─ systemd/
│  ├─ session-log-compile.service
│  └─ session-log-compile.timer
├─ memories/                   # auto-loaded context (user-tier)
│  ├─ code-review-methodology.md
│  ├─ pr-review-techniques.md
│  ├─ sig-auth-rigor.md
│  ├─ worktree-workflow.md
│  └─ tool-use.md
├─ docs/
│  └─ talk-walkthrough.md      # 5-pillar live demo doc
├─ examples/                   # domain-specific templates
│  ├─ agents/
│  │  ├─ dalec-spec-bump.md
│  │  └─ gatekeeper-policy-author.md
│  ├─ skills/
│  │  └─ kubernetes-sig-auth-rigor/
│  └─ repos.txt
├─ INSTALL.md
├─ LICENSE
└─ README.md

Reuse model — what's portable, what isn't

Layer	Portable?	Notes
Agent prompts (agents/*.md)	Yes — pure Markdown + YAML frontmatter.	Path conventions are noted in the prompts; rewrite them for your host.
Skills (skills/*/SKILL.md)	Yes — pure Markdown.	Some hosts call these "rules" or "instructions"; the content is the same.
Memories (memories/*.md)	Yes, but the install path varies by host.	They're meant to be auto-loaded into every turn, not lazily fetched.
scripts/wt	Linux/macOS Bash.	Pure git/bash; no extra deps.
scripts/session-log-compile.sh	Linux/macOS Bash. Requires jq, find, git. Reads VS Code chat transcripts under ~/.vscode-server/data/User/workspaceStorage/.	The transcript path is VS Code-specific. For other editors, point the script at the equivalent transcript directory.
systemd/	Linux with systemd user instance.	Provided as a starting point; macOS users want launchd, others want cron. The systemd/README.md covers conversions.

Compatibility

• Editors: VS Code (with Copilot Chat or any chat extension that writes JSONL transcripts under workspaceStorage/). The session-log compiler is the only piece coupled to VS Code's storage layout; agents/skills are editor-agnostic.
• LLM hosts: Claude Code is the primary target (sub-agents + skills + memory tool match cleanly). The same Markdown works as Cursor "rules", Continue "config", or Copilot Chat "custom instructions" with light adaptation.
• OS: Linux (primary) and macOS for the scripts. The Markdown is OS-agnostic.

Provenance and credit

The PR review and author skills under skills/ and the corresponding notes under memories/ are abstracted patterns distilled from the public PR-review and commenting history of senior open-source maintainers in the Kubernetes / OPA / Azure ecosystems. The raw mining corpora — which contain personally-attributable phrasing and timing data — are not redistributed in this repo. If you do similar mining yourself, treat the corpus as private and publish only the abstracted ruleset.

Talk / walkthrough

docs/talk-walkthrough.md is a 5-pillar live demo doc — the same one used to walk teammates through the toolkit. Open the linked artifacts as you go.

Contributing

This is a personal toolkit shared for reuse, not a maintained product. PRs that improve portability (more hosts, better fallbacks, fewer hard-coded paths) are welcome. Domain-specific agents are best kept in your own fork.

License

Apache-2.0. See NOTICE for provenance and attribution requirements when redistributing.