Current version: 0.2.0

Stop AI agents from shipping plausible wrong work.

Krypton is the goal-based planning and proof gate for running Claude Code, Codex, and other AI coding agents against production-grade codebases.

The failure mode Krypton is built for is not bad syntax. It is the right-looking feature that compiles while living on the wrong layer, inventing a second source of truth, skipping cutover, or claiming success without proof from the real path.

Krypton forces the plan before the code:

• truth owner
• contract boundary
• displaced path
• cutover decision
• acceptance evidence
• review gates

Built for /goals, equivalent goal-based agent workflows, Claude Code skills/plugins, Codex skills, and operators who need agent speed without letting the codebase rot.

Modern agents can create weeks of architectural debt in one enthusiastic session. Krypton turns "build this" into an operational contract before the agent touches code:

• what product or engineering outcome the work serves
• what current behavior is replaced, redirected, deleted, or demoted
• who owns the truth
• what contract crosses the boundary
• what evidence proves the result from the target person's perspective
• what kill criteria prevent duplicate paths from living forever

That contract is what lets an operator keep using agents on production systems without letting the codebase become a pile of current-looking alternatives.

Krypton is best used with /goals or any equivalent goal-based agent workflow. It is not meant to be a one-shot "please implement this" prompt.

The planning session creates the durable handoff:

docs/goals/<goal-slug>/PLAN.md
docs/goals/<goal-slug>/GOAL.md
docs/goals/<goal-slug>/EVIDENCE.md

PLAN.md is the source plan. GOAL.md is the compact /goal prompt. EVIDENCE.md is where execution records the actual proof from the real command, artifact, payload, browser state, trace, or operator-visible result. Start the execution session with the GOAL prompt, then load krypton-execution so the main agent preserves the plan's ownership, cutover, review, and evidence gates.

If your harness does not have /goals, paste the contents of GOAL.md into a fresh Codex or Claude session. The shape still works.

Use the open skills CLI when you want Krypton installed into supported agent skill directories:

Codex only:

npx --yes skills add jturntdev/krypton --skill '*' --agent codex --copy -y

Claude Code only:

npx --yes skills add jturntdev/krypton --skill '*' --agent claude-code --copy -y

That installs every Krypton skill for the named agent only. If you want Codex, Claude Code, OpenCode, or another agent together, run the command without --agent ... and use the interactive selector:

npx --yes skills add jturntdev/krypton

See the package page:

https://skills.sh/jturntdev/krypton

This repo ships a Codex marketplace file at .agents/plugins/marketplace.json. For Codex clients that support repo marketplaces:

codex plugin marketplace add \
  "https://github.com/jturntdev/krypton.git" \
  --ref "main" \
  --sparse ".agents/plugins" \
  --sparse ".codex-plugin" \
  --sparse "assets" \
  --sparse "skills"

codex plugin install krypton --source krypton

Use the plugin route when you want Claude Code to manage Krypton as a plugin:

/plugin marketplace add jturntdev/krypton
/plugin install krypton@krypton-dev
/reload-plugins

Claude Code namespaces plugin skills, so invoke them as:

/krypton:krypton-planning
/krypton:krypton-execution

This repo includes the Claude plugin files Claude Code expects:

.claude-plugin/plugin.json
.claude-plugin/marketplace.json
skills/krypton-planning/SKILL.md
skills/krypton-execution/SKILL.md

Use manual install when you want the skills available as personal Claude Code skills without the plugin marketplace:

git clone https://github.com/jturntdev/krypton.git
mkdir -p ~/.claude/skills
cp -R krypton/skills/krypton-planning ~/.claude/skills/
cp -R krypton/skills/krypton-execution ~/.claude/skills/

Manual personal skills invoke without a plugin namespace:

/krypton-planning
/krypton-execution

For Codex skills:

git clone https://github.com/jturntdev/krypton.git
mkdir -p ~/.codex/skills
cp -R krypton/skills/* ~/.codex/skills/

• krypton-planning: turn a request into an outcome contract, architecture slice, task plan, evidence gate, and /goal handoff prompt.
• krypton-execution: execute an approved plan without drifting from ownership, cutover, or proof requirements.

1. Load krypton-planning with a feature request, bug, refactor, migration, or architecture goal.
2. The agent produces a goal package:

docs/goals/<goal-slug>/PLAN.md
docs/goals/<goal-slug>/GOAL.md
docs/goals/<goal-slug>/EVIDENCE.md

3. PLAN.md is the full implementation plan. GOAL.md is the short /goal prompt or handoff prompt for the next session.
4. Paste or run the GOAL.md prompt in Codex or Claude.
5. Pair it with krypton-execution when you want the main agent to use the same ownership, cutover, review, and evidence discipline.
6. Finish only when the acceptance evidence is captured from the real route, artifact, payload, trace, browser state, or operator-visible output and recorded in EVIDENCE.md.

The workflow is intentionally two-stage:

rough request
  -> krypton-planning
  -> PLAN.md + GOAL.md + EVIDENCE.md
  -> /goal handoff
  -> krypton-execution
  -> main-agent implementation + review gates + acceptance evidence

Krypton also ships a reusable PR gate, but treat it as beta until you have tested it in a throwaway repo that matches your workflow. The primary install path is still the skills.sh command above.

The gate fails non-trivial code changes that do not include a changed Krypton goal package with:

• PLAN.md
• GOAL.md
• EVIDENCE.md
• truth owner text
• deletion or cutover text
• evidence gate and acceptance evidence text

Beta workflow example:

name: Krypton Goal Gate

on:
  pull_request:

permissions:
  contents: read

jobs:
  krypton:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
        with:
          fetch-depth: 0
      - uses: jturntdev/krypton@v0.2.0

Run the same gate locally:

python3 scripts/check-krypton-goal.py --base-ref origin/main

Goal package templates live in:

templates/goal-package/

Before relying on the Action in a real repo, test three cases in a throwaway repository:

• code change with no goal package should fail
• docs-only change should pass
• code change with PLAN.md, GOAL.md, and EVIDENCE.md should pass

Krypton works best when the harness supports named agents:

• explorer: read-only source-of-truth and architecture mapping
• plan-reviewer: PRE and POST alignment checks
• reviewer: runtime correctness, security, and evidence check
• maintainer: codebase-shape, duplication, and cutover-debt check
• verifier: focused proof from the real path

See docs/required-roles.md for the role expectations. If your harness does not support named agents, the main agent can still follow the same gates, but independent exploration and review will be weaker.

Krypton keeps prompt templates individual instead of centralizing them into one large prompt file. Each skill owns the prompts it actually uses:

• skills/krypton-planning/plan-reviewer-prompt.md
• skills/krypton-execution/post-plan-reviewer-prompt.md
• skills/krypton-execution/reviewer-prompt.md
• skills/krypton-execution/maintainer-prompt.md

Bad agent path:

"Add sentiment to the dashboard."

The agent adds frontend-only sentiment logic and calls it done.

Krypton path:

Truth owner: market intelligence producer
Contract boundary: typed sentiment payload
Cutover: dashboard reads producer output only
Evidence: API response plus browser state showing the real payload

See examples/ and tests/pressure-scenarios/ for more.

This is the first public cut. It is intentionally small: two skills, individual prompt templates, agent role expectations, goal package templates, a local gate script, a beta reusable GitHub Action gate, examples, pressure scenarios, and validation scripts.

Krypton uses SemVer-style versions while the public package stabilizes.

• Version source of truth: VERSION
• Plugin metadata must match: .codex-plugin/plugin.json, .claude-plugin/plugin.json, and .claude-plugin/marketplace.json
• Release tags should use vX.Y.Z

Version policy:

• Patch: README, examples, prompt wording, metadata, validation improvements
• Minor: new skills, new gates, changed /goal workflow shape
• Major: breaking skill names, removed gates, incompatible plan or goal format

Run:

./scripts/validate.sh

The validator checks required files, skill metadata, JSON plugin metadata, and public-safety issues such as placeholders or private project references.

MIT