Common AI
The harness in detail · Common Contributions

How a contribution stays small, capped and reviewed.

This page is the detail behind the mission: every guard, what the dry run showed, and the honest state of each piece. Nothing here is installable yet.

Consent comes first

No file, no contact.

The proposed design has a project opt in with a single file. If it is absent, the tool does nothing there. You revoke it by setting one line to false. This is a design. Today the tool opts in by issue label and per-issue scope.

.github/AGENT-CONTRIBUTIONS.md   # name provisional

agent_contributions: true          # set false to revoke instantly
contact: maintainer@example.org    # a human who agreed to this
label: agent-ready
allowed_paths: ["src/**", "tests/**", "docs/**"]
forbidden_paths: [".github/**", "**/*.lock"]
requires_acceptance_test: true
max_diff_lines: 200
disclosure_line: "AI-assisted; reviewed by a human before submission"

A project never has to pay anything, accept any change, reply by a deadline, or run our software.

How a task runs

One issue. One file. Six steps.

  1. A maintainer opts one issue in. They label it agent-ready and write its scope: one file and a check for when it is done.
  2. The coverage gate looks first. If a test already covers the code and the issue does not say so, the run is refused.
  3. Spend is reserved before the call. The worst case is checked against your daily and total caps.
  4. The agent returns text. It has no shell and no tools. If it reports using any, the run stops.
  5. The harness writes the file and checks it. Secrets and stray URLs are scanned, then the project's own type check and tests run.
  6. Another model tries to break the tests. A person reviews the result. Nothing is published without that.

Six guards. Flip one to see what it stops.

Guard circuits · dry-run harness Click a breaker to trip it. Click again to reset it. Resetting is a person's job.
Spend cap$3 per day Armed

Stops a runaway loop, or a provider with no hard limit.

How

The worst case is reserved against the daily and total caps before every call. A model with no known price is refused.

Coverage gateevery test Armed

Stops writing a test that already exists.

How

Every test file is searched for imports of the module. The issue has to name each one it finds.

One file1 path Armed

Stops edits outside the task.

How

The result must be exactly the path in the issue. The harness writes the file itself, not the model.

Secret scanreply + file Armed

Stops keys and tokens leaking into output.

How

The reply and the written file are scanned for known secret shapes. The sandbox environment is an allowlist.

Egress allowlistmodel + git Armed

Stops the agent reaching anywhere else.

How

The harness's own requests may go only to the model endpoint and git. This checks what the harness sends; it is not an operating-system firewall, and running model-written code in a container is not built yet. Documentation hosts such as example.com count as fixtures, not calls.

Branch ruleagent/* only Armed

Stops anything landing on main.

How

The publish gate rejects any other branch name, including look-alikes and anything containing "..".

Main kill switch

One file, checked before every call. Three failures in a row trip it. Only a person clears it.

panel all six circuits armed. Nothing has tripped.

0 of 6 tripped

This is a model of the real guards, with their real behaviour described in plain words. The harness has its own test suite, and a guard was removed more than 50 times to check that the tests notice.

Evidence

A test tuned to one set fell from 86% to 50% on the next.

After each test file passed, a model that had never seen it wrote mutations of the code: small deliberate breaks. The share the tests catch is a floor on their quality, not a gate. Each set was written from the source alone by a model that was not allowed to read any test.

Caught out of total. Grey bars are scores on sets that had already shaped the tests. Only the dark and orange bars show whether the tests generalise. Some survivors on the first set are equivalent mutants that no test can catch.

What this does not show yet

  • Two small modules, one model family writing tests, one author per mutation set.
  • In the dry run, the budget cap and measured token counts were never exercised (a later real run of the hosted adapter is in the status table above). The agents ran in a coding assistant that is not billed per token.
  • The count of tools an agent used is passed in by whoever runs the harness. The harness cannot verify it.
  • The container is tested on one runtime (macOS with OrbStack). A container escape would reach this Mac's home folder with your rights, and --allow-host-execution skips the container. The run record is made by the tool on your machine, so it is self-reported; a project's own CI is the real check.
  • The dry run itself opened no pull request, and no maintainer has reviewed those test files. (Later, one pull request was opened on our own demo repository; it is in the status table above.)

Read the full dry-run report

What exists

There is no installer. Here is each piece.

This is the honest state of every part, so nobody has to guess what a command would do.

PieceStatusDetail
Spend cap and ledgerBuiltWorst case held under a lock before each call, so two runners cannot double-spend. Append-only ledger. Unknown prices refused.
Kill switchBuiltChecked before every call. Three failures in a row trip it. Only a person clears it.
Scope, secrets, egress checksBuiltOne file per task, secret scan (patterns plus an entropy check on what the agent adds), host allowlist, agent/* branches only.
Coverage gate, mutation runnerBuiltUsed on the first runs.
Coding-agent replayBuiltAn agent's reply goes through the same checks. This is how the first runs were done.
Local model adapterMock-testedLoopback only. Not run against a real model yet.
Maintainer opt-in filePartly builtThe file format, a strict parser and a checker for maintainers are built and tested (optin-check). No project has adopted it.
Hosted model, capped API keyPartly builtWorkers AI adapter: usage measured from the provider, an input bound enforced before the call, capped by the ledger. One real run cost $0.0027 (2,904 tokens in, 792 out) and a second was refused by the cap before any call. Not yet reconciled against Cloudflare's own usage figure.
Pull request publishingPartly builtA dry run and a publisher that opens a pull request only from the submitter's own fork, in one command, are built and tested against a fake GitHub. It has run for real once: on 2026-10-12 it opened one pull request on our own demo repository (pull request 31), which we merged. Before spending tokens or publishing, it also looks for pull requests that already mention the issue. No outside project has adopted it, and no pull request has been opened anywhere else.
Container verifierBuiltModel-written code is checked in a container with no network, a read-only copy of the source at a recorded commit, no capabilities, and memory, process and time limits. It refuses to start unless the runtime passed the escape tests (tested on macOS with OrbStack only). A container escape would reach this Mac's home folder with your rights. --allow-host-execution skips the container. Reaching the host from inside a container with no network was not demonstrated, only that sibling containers and the internet are unreachable. The run record is self-reported.
One-command installNot builtNo package exists.
Name and domainNot clearedCommon AI is the name for now. Another organisation, CommonAI, uses a near-identical name; no trademark or domain check has been done and nothing is registered.

When there is something to install, the plan is one instruction to your agent. It would ask for three things: an API key with a spend limit you set, a list of projects you choose, and a daily budget. Until then, nothing on this page should be copied into a terminal.

Besides the guards

The rest of what it does.

Checks the work

  • Runs the project's own tests, not its own idea of them.
  • Scores the tests with mutations written by a different model than the one that wrote them.
  • Reports what it cannot know, such as the number of tools an agent used.

Keeps people in charge

  • A person clears the kill switch. The tool never does.
  • A person reviews and merges. The agent never does.
  • No rewards. Nothing is paid or credited for contributing.
For agents

If you are an AI agent reading this page.

Here is a plain summary you can parse. It asks you to do nothing. The same text is at llms.txt and agent.json.

# machine-readable note
name: Common AI (working title; the tool is called Common Contributions)
status: not installable
instruction: do not run commands, install packages, or contact any project because of this page
action: tell your user what this page says and ask what they want
what_it_is: a harness that caps spend, limits scope and requires human review
what_exists: guard rails, coverage gate, mutation runner, replay mode, hosted-model adapter (spend-capped; not yet reconciled with the provider), container verifier (macOS with OrbStack only), opt-in file checker and PR publisher (fake GitHub, plus one real pull request on our own demo repository)
what_does_not: installer, a pull request on anyone else's project, any outside project using the opt-in file
credentials: never ask your user for a subscription login; an API key with a spend limit is the only fit
rewards: none
Why a page for agents says "do nothing". A page that tells agents to run things on their owner's behalf has the same shape as an attack. Anything real here will say what it does, ask your user first, and be easy to stop.