Cuff

Make claims checkable. Reject evidence when stale.

Cuff ties one completion claim to one exact subject, runs the verifier you choose, and checks whether the latest passing evidence still matches the current Git state.

It turns a completion statement into a durable, checkable record without deciding what should prove the work or what action should follow.

Requirements

• Python 3.11 or newer;
• uv on PATH (0.11.32 is the tested recommendation); and
• an existing Git worktree. Its root is the only valid Cuff workspace.

Git is mandatory. Cuff never initializes a repository, selects another worktree, or stages, commits, fetches, pushes, releases, or deploys anything.

Install

Install a released version as a standard uv-managed tool:

uv tool install cuff-cli==0.2.2
cuff --version

For local development, install the checkout explicitly:

uv tool install --editable .

Cuff has no runtime dependencies. It is distributed as a standard wheel and source distribution; it contains no bundled Python or native executable.

Five-command quickstart

Run initialization at the exact Git worktree root:

cuff init --json
git add .fab7/cuff/project.json
git commit -m "Initialize Cuff"

The marker is exactly {"schema":1} and records live under .fab7/cuff/records/. An incompatible marker is never rewritten or migrated.

The preferred path atomically appends a claim and its observed evidence:

cuff seal \
  --work-item task-1 \
  --summary "Implementation complete" \
  --subject-path src \
  --json \
  -- python -m pytest

cuff check --work-item task-1 --json

Recovery controllers that need the latest linked observation, including a recorded verifier failure, may opt into the closed seven-field projection:

cuff check --work-item task-1 --json --include-latest-record

Without that flag, check --json retains its exact six-field response.

The split path is available when the claim must exist before verification:

cuff claim \
  --work-item task-1 \
  --summary "Implementation complete" \
  --subject-path src \
  --json

cuff verify \
  --work-item task-1 \
  --claim rec_REPLACE_ME \
  --json \
  -- python -m pytest

The public surface is exactly:

cuff init
cuff claim
cuff verify
cuff seal
cuff check

Every claim, verification, seal, and check names its work item explicitly. Declared subjects use the complete {kind, ref, digest} identity; file and tree subjects use --subject-path and a Cuff-computed manifest digest.

Proof boundary

• Claims and evidence are closed generation-1 JSONL records.
• Every evidence record contains the HEAD commit observed before execution.
• Verifier argv is executed literally without a shell.
• Non-ledger dirtiness before or after verification records no evidence.
• seal appends its linked pair in one locked atomic replacement.
• check enforces subject freshness, commit ancestry, changed paths, non-ledger cleanliness, and append-only ledger changes.

Cuff treats verifier argv as opaque. It does not select the command, import an extension, interpret domain output, or grant merge, release, deployment, spend, or residual-risk authority.

Static host integrations

One native payload lives in plugins/cuff and contains both host manifests, Claude Code commands, and Codex skills. The shared fab7hq/fab7 marketplace owns registration; this repository owns the payload. The assets require the uv-managed cuff executable on PATH.

# Codex
codex plugin marketplace add fab7hq/fab7
codex plugin add cuff@fab7

# Claude Code
claude plugin marketplace add fab7hq/fab7
claude plugin install cuff@fab7 --scope user

Validate the built candidate and both host payloads without touching the normal host configuration:

uv build --out-dir ../sandbox/cuff-02/dist
uv run python tools/local_release_check.py --host all
uv run python tools/local_release_check.py --host all --prepare-auth \
  --candidate-commit COMMIT \
  --evidence-dir ../sandbox/cuff-02/e2e
# Log each CLI in using the isolated home paths printed by the prepare phase.
uv run python tools/local_release_check.py --host all --live --reuse-prepared \
  --candidate-commit COMMIT \
  --evidence-dir ../sandbox/cuff-02/e2e \
  --qualification-manifest ../sandbox/cuff-02/control/qualification.json \
  --preflight-evidence ../sandbox/cuff-02/control/preflight.json \
  --codex-model MODEL \
  --claude-model MODEL

Before --live, freeze the exact candidate, host order, current and stale cases, three-valid-sample rule, prompts, tool policies, and containment policies in the qualification manifest. A trusted non-LLM runner must then record a passing containment preflight for the same qualification and policy digests. The checker rejects missing or mismatched controls and runs six fresh workspaces and host sessions per host. Follow the root LLM_VERIFICATION.md; unsupported isolation is INCONCLUSIVE, not a release pass.

See RUNBOOK.md for operations, the architecture overview for ownership, and the ledger contract for the record and gate invariants.

Development

uv sync --locked
uv run --locked python -m pytest
uv run --locked python -m compileall -q core/cuff
uv build
git diff --check

Community and support

• Use Cuff Discussions for usage questions and design proposals.
• Report reproducible defects through GitHub Issues.
• Read CONTRIBUTING.md before proposing a change.
• Report vulnerabilities privately as described in SECURITY.md.

Cuff is licensed under the Apache License 2.0.