EQTY Skills

Three agent skills for EQTY lineage manifests — signed, content-addressed records of what a computation did: its inputs, tool calls, model calls and outputs, checkable by a third party without trusting anyone's word for it.

Skill	Fires when	What it does
eqty-manifest	a manifest is in hand	parses it, verifies its signatures, content addresses, statement integrity and hardware attestation evidence, and writes a report
eqty-instrument	you want your code to emit one	makes a repo emit a manifest, delivered as a patch for a human to review
eqty-explorer	you want to see one	shows a manifest's lineage graph in your browser and drives it while the agent answers: shows only the nodes an answer is about, filters, and selects

eqty-instrument produces a manifest and points at it; eqty-manifest reads it back; eqty-explorer draws it. Answers always come from eqty-manifest, and the picture only illustrates them.

Install

All three skills install together, in Claude Code or Codex.

Claude Code plugin:

/plugin marketplace add eqtylab/eqty-skills
/plugin install eqty-skills@eqty-lab

Codex plugin:

codex plugin marketplace add eqtylab/eqty-skills
codex plugin add eqty-skills@eqty-lab

Manually — copy the three skill directories from plugins/eqty-skills/ into a skills folder, side by side (eqty-instrument and eqty-explorer call eqty-manifest as their sibling):

Agent	For you, in every project	For one project
Claude Code	~/.claude/skills/	.claude/skills/
Codex	~/.agents/skills/	.agents/skills/

cp -R plugins/eqty-skills/eqty-manifest plugins/eqty-skills/eqty-instrument plugins/eqty-skills/eqty-explorer ~/.claude/skills/    # or ~/.agents/skills/ for Codex

Update

New versions are listed in CHANGELOG.md. After updating, restart the agent so it loads the new skills.

Claude Code — refresh the marketplace, then update the plugin:

claude plugin marketplace update eqty-lab
claude plugin update eqty-skills@eqty-lab

Codex — refresh the marketplace, which updates the plugin from it:

codex plugin marketplace upgrade eqty-lab

claude plugin list and codex plugin list show the installed version. Manually, copy the three skill directories again, over the old ones.

Each skill is self-contained: its directory carries everything its SKILL.md tells an agent to read or run, and refers to its own files relative to that directory. An install copies plugins/eqty-skills/ and nothing else: the three skills, the plugin manifests, LICENSE and NOTICE. The rest of this repository (examples/, docs/, this README) is for people and is never installed.

Requirements

• Python 3.10+.
• eqty-manifest needs eqty-sdk (2.5.0 or later), cryptography and base58 to verify. With uv, uv run installs them per script; otherwise pip install -r plugins/eqty-skills/eqty-manifest/requirements.txt. Without them the scripts still parse and report, and every check they cannot run says not checked — never a pass.
• eqty-instrument needs nothing installed to run: its scripts (detect.py, check_graph.py) use the standard library only. What it produces is a patch to your code, and that patch adds eqty-sdk to your project's dependencies, because the instrumented code imports it. In Mode 1 the patch also adds EQTY's LangChain or DeepAgents handler package (eqty-lineage-langchain or eqty-lineage-deepagents), also on public PyPI.
• eqty-explorer needs a browser, and nothing installed: explorer.py uses the standard library only. It loads the lineage explorer's code from jsDelivr (@eqtylab/explorer, pinned by version and hash); the manifest stays on your machine.

The skills build on the open-source EQTY SDK (eqty_sdk): eqtylab/integrity-py, installed with pip install eqty-sdk.

Try it

A real manifest comes with the examples. Verify it:

uv run plugins/eqty-skills/eqty-manifest/summary.py \
  examples/mode1/deepagent/research_agent/research_agent_manifest.json

Summary
=======
173 / 173 MetadataRegistration verified
120 / 120 DataRegistration verified
52 / 52 ComputationRegistration verified

Signers:
- did:key:z6MkqWiZdvahw5qQMBvrJfgTh7sLafRq18VpubrhnGnkRsrj signed 366 / 366 valid credentials
- 366 / 366 credentials: issuer is the registration's signer

Hashes:
- Total of 163 hashes
- 163 have pre-images attached
- 163 / 163 verified
- No tampering detected
- No missing pre-image blobs

Or ask Claude about it — "what does this manifest show?" — and eqty-manifest takes over. Then "show me the graph", and eqty-explorer opens it in your browser.

eqty-instrument, in two modes

• Mode 1 — repos built on LangChain, LangGraph or DeepAgents. The framework already fixed the node boundaries, so the edit is the same every time: initialise the SDK once, attach one callback handler per run, fingerprint the repo's own tools, export. detect.py plans it.
• Mode 2 — arbitrary Python. No framework has decided which operations deserve nodes, so the skill derives a control-flow graph first, you choose auto or human-in-the-loop node selection, and only then does it touch code. check_graph.py checks the result.

Examples

examples/ holds runs of both modes, for people to read; the agent does not use them. Start with examples/README.md. The Mode 1 examples show the current edit. The Mode 2 examples, lora_banking77 and pygit, show the current flow: blind CFG, data flow read with pyright, the two joined, the node selection, the patch and the prediction, with nothing run. More on the skill in docs/eqty-instrument.md.

Known limitations

See LIMITATIONS.md.

License

Apache License 2.0 — see LICENSE. Third-party material under plugins/eqty-skills/eqty-instrument/references/upstream/ and examples/ carries its own license beside it; see NOTICE.