manifest-kit

Roham's toolkit for AI behavior — structural hooks that make agents behave correctly. Vision bridge, image-gen bridge, ForceLink, locate-yourself gate, search-before-assert, finish-not-flag, deliverable-quality, model-stack-persistence, agent-verify cognitive verification, and a reconciliation checker. Spec: SPEC-022.

Renamed from manifest-it / ai-manifest-it on 2026-06-26. The repo is now dapperlabs/manifest-kit.

What it is

A Claude Code plugin that bundles every structural fix for agent behavior. Two components:

Hooks (behavior guards)

The things the agent can't do without help and shouldn't do wrong:

1. Vision bridge — auto-describes images a text-only model can't see (PostToolUse on Read).
2. Image-gen bridge — routes image generation to OpenAI for text-only models.
3. ForceLink — blocks turns ending with bare file paths; forces clickable URLs.
4. Distilled ethos — portable operating principles (not essays).
5. Locate-yourself — SessionStart gate that tells the agent what room it's in (reads worldmodel.md).
6. Non-main-branch — blocks commits to non-main branches on non-boot repos.
7. Search-before-assert — blocks "doesn't exist" claims without search evidence.
8. Finish-not-flag — blocks "worth noting" when the task isn't actually done.
9. Deliverable-quality — asserts minimum quality for declared deliverables.
10. agent-verify input check — PreToolUse check that inputs contain real content, not metadata (from roham/agent-verify, wraps the council-verify ladder).
11. agent-verify output check — PostToolUse check that outputs contain actual value, not vacuous structure.

Checker (reconciliation)

Merged from SPEC-021 (was: RoboCop). A deterministic cron job that checks:

• Are the hooks installed and working?
• Are specs stale or dropped?
• Is data in the wrong org?
• Is worldmodel.md present and consistent with the README?
• Is the deployed VM code committed?

Not an agent — a deterministic script. Runs on cadence, surfaces divergences, never remediates.

Install

As a Claude Code plugin (recommended)

git clone https://github.com/dapperlabs/manifest-kit.git ~/.claude/plugins/manifest-kit

Then add to your ~/.claude/settings.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Read",
        "hooks": [
          { "type": "command", "command": "python3 ~/.claude/plugins/manifest-kit/hooks/vision_bridge_hook.py", "timeout": 130 }
        ]
      },
      {
        "matcher": "Bash|Task|Agent|Write",
        "hooks": [
          { "type": "command", "command": "python3 ~/.claude/plugins/manifest-kit/hooks/agent_verify_post_check.py" }
        ]
      }
    ],
    "PreToolUse": [
      {
        "matcher": "Read|Task|Agent",
        "hooks": [
          { "type": "command", "command": "python3 ~/.claude/plugins/manifest-kit/hooks/agent_verify_pre_check.py" }
        ]
      }
    ],
    "Stop": [
      {
        "matcher": "",
        "hooks": [
          { "type": "command", "command": "bash ~/.claude/plugins/manifest-kit/hooks/forcelink.sh", "timeout": 15 }
        ]
      }
    ]
  }
}

Restart Claude Code. Done.

Hooks only (no plugin system)

# ForceLink
curl -o ~/.claude/hooks/forcelink.sh https://raw.githubusercontent.com/dapperlabs/manifest-kit/main/hooks/forcelink.sh
chmod +x ~/.claude/hooks/forcelink.sh

# Vision bridge
curl -o ~/.claude/hooks/vision_bridge_hook.py https://raw.githubusercontent.com/dapperlabs/manifest-kit/main/hooks/vision_bridge_hook.py
curl -o ~/.claude/scripts/vision_describe.py https://raw.githubusercontent.com/dapperlabs/manifest-kit/main/scripts/vision_describe.py

# agent-verify hooks (requires agent-verify library)
pip install git+https://github.com/roham/agent-verify.git
curl -o ~/.claude/hooks/agent_verify_pre_check.py https://raw.githubusercontent.com/dapperlabs/manifest-kit/main/hooks/agent_verify_pre_check.py
curl -o ~/.claude/hooks/agent_verify_post_check.py https://raw.githubusercontent.com/dapperlabs/manifest-kit/main/hooks/agent_verify_post_check.py

Requirements

• python3 (pre-installed on macOS, most Linux distros)
• jq (optional — ForceLink falls back gracefully if not installed)
• For the vision bridge: at least one of GEMINI_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY
• For the image-gen bridge: OPENAI_API_KEY
• For agent-verify: the council plugin + llm CLI

Configuration

Env var	Purpose	Default
GEMINI_API_KEY	Vision bridge provider 1	—
OPENAI_API_KEY	Vision bridge provider 2 / image-gen	—
ANTHROPIC_API_KEY	Vision bridge provider 3	—
CAPABILITY_BRIDGE_IMAGE_MODEL	Image-gen model	gpt-image-2
SKIP_URL_GATE	1 to bypass ForceLink	—
SKIP_VISION_BRIDGE	1 to skip vision bridge	—
DAPPER_TOOLKIT_ADVISORY	1 = force all guards to advisory (never block)	—

Related

• agent-verify — cognitive verification library (council-verify ladder wrapper)
• SPEC-022 — the spec that governs this toolkit
• SPEC-024 — worldmodel.md spec
• SPEC-021 — reconciliation checker (merged into this repo)

License

MIT — see LICENSE. ForceLink is also MIT, bundled unmodified in logic.