INVARA

INVARA

Engineering invariants for AI-built software.

INVARA works with Claude Code, Codex, and any coding agent that can use a shell. It verifies declared constraints against observable evidence; it does not prove software correctness.

seal before work -> agent does the work -> judge

| PASS / BLOCK / UNVERIFIABLE / HUMAN_REVIEW

Start with the path that matches your environment:

• Claude Code: install the bundled editor/plugin experience with /plugin marketplace add Jujitae/invara, then /plugin install invara.
• Codex or another shell-capable agent: run uvx --from invara==0.3.0 invara list, or install persistently with python -m pip install invara==0.3.0.
• Tell your agent: use the copy-paste instructions in AI Agent Quickstart, then preserve the exact verdict it receives.

The first result is a verdict, not an agent self-report. PASS means every declared check returned what it promised and protected paths stayed unchanged. BLOCK, UNVERIFIABLE, and HUMAN_REVIEW stay exactly that; none is success by reinterpretation.

Current public release: 0.3.0. Status: Alpha. INVARA Free is distributed publicly; INVARA Pro is a separate paid product. INVARA is not autonomous authorization and not an enterprise compliance control.

Transformation assurance in 0.3.0, Gen 1

Two local CLI workflows can now check a declared behavior envelope around a refactor or migration:

• invara assure compares an existing before/after pair under an Equivalence Manifest and can search for a minimized counterexample.
• invara repair governs a repair one unit at a time, keeping a unit only after the declared checks accept it.

Gen 1 measures only the inputs, observations, policies, and claims declared in the manifest. A PASS is bounded to that envelope; it is not a general proof of correctness, security, or bug-free software. The workflows run user-named commands with the user's permissions. They make no outbound network call; an HTTP probe may connect only to a loopback service that the manifest starts for the session.

Start with the CLI guide and the manifest guide. Synthetic examples and display reports are under fixtures/ and docs/transformation-assurance/samples/.

---

What this is not

Read this part before the install line. It is short on purpose, and the word that carries the promise above is *verifiable* — not *correct*.

INVARA does not prove software correctness.

It verifies declared constraints against observable evidence.

Specifically:

• It does not inspect your codebase. It never asks whether the code is good, whether the architecture holds, or whether the tests are the right tests. Stack a new floor on junk and, if this floor was built to the declared spec, the verdict is PASS.
• It does not find bugs. A completion check that runs your suite is only as strong as your suite. INVARA reports the exit code; it does not have an opinion about coverage.
• It does not decide what should have been promised. You write the contract. A weak contract earns a weak PASS, and the contract is stored verbatim so anyone can see how weak it was.
• It is not a sandbox. Completion checks are commands and they run with your permissions. Do not seal a contract you have not read.
• It does not judge intent. intent is prose; nothing checks that the work matched it. The checks are what bind.
• It is not an AI code generator and not an LLM code-review bot. There is no model anywhere in the verdict path.

The narrow claim, which is what the machine actually does:

It decides, independently, whether this change kept the promises it made.

---

CLI first verdict, in more detail

Nothing to configure. No API key, no service, no account. Python 3.12+ and uv.

uvx --from invara==0.3.0 invara list

or, to keep it:

python -m pip install invara==0.3.0

The package declares zero runtime dependencies, so this pulls only the standard library. The pinned install reports the installed INVARA version, and pip list shows no additional runtime packages.

From a checkout it is the same program, but install it first — the source lives under src/, so a bare python -m invara in the repository root finds nothing to run:

pip install -e .

python -m invara list

Release-candidate verification before publish

The bundled plugin and the distributable package must expose the same six MCP tools. Before a maintainer publishes a new package version, build its wheel and run the repository's clean-environment proof. It creates a fresh virtual environment, installs only that wheel without using an index, starts the MCP server over stdio, verifies the exact tool list, then seals, judges, lists, logs, rebuilds the chain, and replays a verdict.

python -m pip wheel --no-deps . --wheel-dir .runtime/invara-dist

python scripts/verify_fresh_install.py --wheel-dir .runtime/invara-dist --plugin-root plugin

This is a release-candidate proof, not a publish command. The CI workflow runs it independently on Windows and Ubuntu for every candidate pull request.

1. Write the contract *before* the work

task.json, next to the repository you are about to change:

{

"task_id": "2026-08-17-tidy-the-parser", "intent": "Speed up the CSV parser without changing what it accepts", "constraints": [ { "kind": "paths_unchanged", "paths": ["tests/test_parser.py"], "reason": "a speedup that edits its own test is not a speedup" } ], "done_when": [ { "id": "suite", "command": ["python", "-m", "pytest", "-q"], "expect_exit": 0, "reason": "the whole suite" } ] }

invara seal task.json

Sealing takes the digests of the protected paths now, before anyone knows what the verdict will be. That ordering is the entire guarantee.

2. Do the work.

3. Judge

invara judge 2026-08-17-tidy-the-parser            # dry run

invara judge 2026-08-17-tidy-the-parser --commit # record it

BLOCK: 1 protected path(s) changed: tests/test_parser.py: changed (a speedup that edits its own test is not a speedup) decided by: constraint_breaks

That is the first verdict. Everything below is detail.

---

The four verdicts

Constraint breaks outrank everything. A run that touched what it promised not to touch is not partially fine.

BLOCK is two rules wearing one word — a protected path that changed and a check that came back wrong are not the same accusation. So the verdict also records which rule decided it, named after the evidence it decided on (constraint_breaks, failed, unrunnable, needs_human, passed), and judge and log print it. Verdicts recorded before this existed do not have one, and do not get one fitted after the fact.

Sealing refuses more than it accepts

seal will not write a contract that cannot fail the work. It refuses a task with no completion condition, a condition with no command to check it, a contract with no protected paths ("a task allowed to change anything cannot be said to have respected anything"), a duplicate check id, a protected path that does not exist, and — the one that matters most — a contract where every condition defers to a person.

That last one is why this is not a rubber stamp: if the only evidence is somebody saying yes, there is no contract.

Other commands

invara init              print a task.json template for this repository

invara list sealed tasks and their latest verdict invara log every verdict this task has ever had invara show the contract, exactly as sealed invara replay recompute a recorded verdict and compare it invara chain rebuild both hash chains

invara replay is the one that answers "was this verdict reproducible?" — it recomputes from the observations stored at the time. A verdict recorded before replay support refuses with no_current_stored rather than guessing.

Verdicts live in .runtime/verify.db (--db to move it). A contract is sealed once and judged many times; the history is append-only and chained.

---

Inside the editor

The buyer this was built for does not open a terminal. So the same package ships an MCP server, self-contained as a Claude Code plugin:

/plugin marketplace add Jujitae/invara

/plugin install invara

The plugin bundles the INVARA source and runs via python -m, requiring only Python 3.12+ — no uv, no network, zero runtime dependencies. If the server will not start, /invara:doctor diagnoses the environment without assuming Python exists — the known silent case is Windows without Python, where the Microsoft Store's python alias spawns and dies with exit 9009 and seven bytes of stderr (Python ); that death happens before any INVARA code runs. The configuration the plugin writes looks like this:

{

"mcpServers": { "invara": { "type": "stdio", "command": "python", "env": { "PYTHONPATH": "${CLAUDE_PLUGIN_ROOT}/src" }, "args": ["-m", "invara.mcp"] } } }

Six tools — invara_seal, invara_judge, invara_list, invara_log, invara_chain, invara_replay. They are the commands above, and they change nothing about what a verdict is: it is still computed from file digests and exit codes, the checks that run are the ones a sealed contract already named, and there is still no field anywhere for an agent to assert that the work is done. An agent can ask for a verdict here. It cannot give one.

invara there is a local label — call it what you like. The server's actual identity is io.github.Jujitae/invara, which is how it is listed in the MCP registry and how a client that resolves through the registry will find it. It cannot be used as the key above: claude mcp add answers Names can only contain letters, numbers, hyphens, and underscores.

Not a sandbox, and this does not make it one. invara_seal takes a task file, that file names commands, and judging runs them. An agent that can write a task file can cause those commands to run — which is no more than the shell it already has, but better said here than discovered.

---

Determinism

Same repository state, same contract, same environment → same verdict. The verdict is a function of file digests and command exit codes, and nothing else; there is no model in the path and no clock in the decision.

Measured rather than asserted, 2026-08-17: contract 2026-08-17-route-discovery was judged twice, four minutes apart, against an unchanged tree — 7 checks each including a 2,021-test suite. Both runs returned the same status and the same reason string, and both rows are in the chain:

$ invara log 2026-08-17-route-discovery

2026-08-17 09:17 UTC PASS 7 check(s) passed and 6 protected path(s) are unchanged 2026-08-17 09:21 UTC PASS 7 check(s) passed and 6 protected path(s) are unchanged

Run that yourself on your own contract before you trust it on ours.

The word doing work there is environment. See the first failure story below: a contract that passed locally and failed in CI was not non-deterministic —