K-stack

A Claude-only engineering stack for Claude Code: a rigorous working mode, per-role Sonnet/Opus effort routing, focused skills, and daily-driver hooks.

K-stack is my personal Claude Code stack, and the next version of what used to be Claude Agent System (CAS). I merged the parts of CAS I still use every day (Spectre research, the safety hooks, the Codex and Cursor delegation skills) with my global setup (e2e QA through AgentController, notifications, commit guards). I removed the routers and swarm wrappers that current models no longer need. I also brought in the working mode and playbooks from pstack by Lauren Tan (MIT), and rebuilt them around Claude-only Sonnet and Opus routing.

Install

/plugin marketplace add Kasempiternal/K-stack
/plugin install kstack@k-stack
/kstack:k-setup
/kstack:k-mode

k-setup writes a managed kstack-routing block into ~/.claude/CLAUDE.md and asks for a preset. k-mode starts the mode.

How it works

The lead is Opus 5.5 at High. It plans, briefs, reviews diffs, and talks to you. It edits directly only for one-or-two-turn changes.
Work routes by role, never by model name. A playbook step names a role (build, explore, review, ...), and the role resolves to a tier agent: one kstack:k-* agent per model/effort pair, since the Agent tool cannot carry effort.
Burn is a deliberate spend-more switch. Say "burn" or name a Max tier and the current task escalates to Sonnet Max or Opus Max. It never turns on by itself.
The loop is design → build → review → fix. Nontrivial changes go through k-design first, every worker diff goes through k-review before landing, and verification happens on the real surface (k-e2e-qa for UI and devices).
Nothing is done without evidence. Workers paste the command and its outcome; unverifiable claims get labeled, not laundered.

Skills

Skill	What it does
`/kstack:k-mode`	The mode: playbooks, role routing, evidence discipline
`/kstack:k-setup`	Writes the role→tier routing block into `~/.claude/CLAUDE.md`
`/kstack:k-explore`	Read-only subsystem explanations and git-history archaeology
`/kstack:k-design`	Design before code: data shape, interface, file plan, failure cases
`/kstack:k-review`	Diff review on separate Opus lenses, judged by the lead
`/kstack:k-write`	Prose rules for docs, commits, PRs, and replies
`/kstack:k-spectre`	Parallel web/codebase research with independent claim validation
`/kstack:k-e2e-qa`	UI and device verification: macOS, iOS, Android, web
`/kstack:k-gpt-architect`	Optional: delegate to the Codex MCP to spend ChatGPT quota
`/kstack:k-cursor-architect`	Optional: delegate to `cursor-agent` (Grok) to spend Cursor quota

Balanced routing (default)

Role	Covers	Tier
`mechanical`	renames, formatting, boilerplate, small CSS, changelogs	sonnet-low
`docs`	README, developer docs, docstrings, API docs	sonnet-medium
`build`	normal features and UI, simple bugs, tests	sonnet-high
`build-hard`	large features, complex state, refactors, hard bugs	sonnet-xhigh
`explore`	unfamiliar code, repo-wide search, regression archaeology	sonnet-xhigh
`research`	web and document research seats (k-spectre)	sonnet-high
`design`	architecture, API design, plans, RFCs	opus-high
`review`	diff review, simplification, security, claim validation	opus-high
`sweep`	final pre-release sweep, stubborn cross-system root cause	opus-xhigh

Max never appears in a preset; Burn reaches it per task. The full table with Lean and Quality presets lives in kstack/skills/k-mode/references/routing.md.

The routing shape follows Artificial Analysis measurements quoted 2026-09-29: Sonnet 5.5 High→XHigh is the useful coding step (Coding Agent Index 55→63), XHigh→Max gains little at ~4x cost, and Opus 5.5 High beats Sonnet XHigh on intelligence per dollar, so judgment roles go to Opus and code volume goes to Sonnet.

Hooks

Hook	Event	What it does
`protect-secrets.js`	PreToolUse `Read\|Edit\|Write\|Bash`	Asks before touching `.env`, private keys, credential files, or commands that expose/exfiltrate secrets
`sensitive-command-approval.cjs`	PreToolUse `Bash`	Requires approval for `rm`, `git commit`, and `git push`, including through `sudo`, `env`, `xargs`, `bash -c`, `eval`, and `find -exec`
`no-claude-attribution.js`	PreToolUse `Bash`	Denies `git commit` commands containing `Co-Authored-By` lines
`kill-before-open.js`	PreToolUse `Bash`	macOS only. When a command opens a `.app`, kills the running instance first so you never test a stale build
`notify-session-start.sh`	SessionStart	Records the terminal tab and project so the notifiers below can route
`notify-attention.sh`	Notification	macOS only. Posts a notification on permission and idle prompts
`notify-stop.sh`	Stop	macOS only. Posts a notification when a task finishes

Requirements

Claude Code with plugin support, plus node and bash for the hooks (python3 for the notification scripts).
macOS for the notification and kill-before-open hooks; they no-op silently elsewhere.
Optional, only if you use the matching skill: the codex MCP for k-gpt-architect, cursor-agent for k-cursor-architect, AgentController and Maestro for k-e2e-qa.

Migrating from CAS

The old cas plugin is preserved at tag cas-v7.37.0-final. Renamed: spectre → k-spectre, review → k-review, gpt-architect → k-gpt-architect. Removed: cccontrol, cyberconan, faster, gonk-test, hydra, l30, legion, orchestrate, pcc, pcc-opus, phoenix, setup-hooks, setup-swarm, shared, siege, systemcc, zk, the cccontrol-mcp bridge, and the spectra-mcp-server. They were older-model scaffolding now covered natively by Claude Code subagents, effort control, and worktree isolation.

License

MIT. See LICENSE and NOTICE.