Repository navigation
plimsoll v0.24.0
A coding agent's "the tests passed" is its own account of itself. This release adds an example that makes it evidence instead: a Nemotron model on Nebius Token Factory writes code, runs it in a plimsoll sandbox behind gVisor, and every run comes back as a signed record that you can recheck offline, now in your browser too. The rest of the release is hardening from a whole-tree review.
Server and clients stay on protocol 4 and run-record version 4. Clients 0.23.0 keep working against a 0.24.0 daemon. Upgrade them anyway: 0.24.0 clients refuse an answer whose result is not the kind the request asked for, and a record that ends before it starts, before either moves a session's chain.
For agent-sandbox operators running hardened mode: set SANDBOX_AGENTSANDBOX_ENCRYPTED_POD_NETWORK=1 once your pod network encrypts traffic (a network plugin with WireGuard or IPsec, or a service mesh with mutual TLS), or the daemon refuses to start. plimsolld talks to each pod's sandboxd over plain HTTP/2, so the sandboxd token, the code, each run's report key and the output are only as private as that network; the daemon cannot check it, so it asks you to say so.
Also included:
- Three fixes from a whole-tree security scan of the published code: the TypeScript client refuses a bytes field that is not base64 text, where a 30-byte answer could make it allocate 2 GB; the daemon's open development mode (
PLIMSOLL_INSECURE=1) answers only requests addressed tolocalhostor a loopback address, so a web page that rebinds its own name to 127.0.0.1 cannot run code through it; and the agent-sandbox control channel above, now stated in the docs. examples/nemotron-agent: a Go program that gives a Nemotron model one tool, run code in a plimsoll sandbox, and records every call with the attest harness; its verifier is the same Go verification compiled to WebAssembly, so a page can recheck a bundle with the public key and no server. The recorded session is on its page: Nemotron 3 Super's first run failed three of eleven tests while exiting 0, it fixed its code, and its second run passed all eleven; the page rechecks the signed bundle in your browser (the page).- Docker provider: a guest that could print to the runner's stdout could make the daemon hash about 1.3 TiB from a 16 MiB stream; parsing now hashes at most twice the stdout cap across every candidate frame. A runner report already in stdout when the time backstop fires is the result, not a timeout with no steps. Orphan reconciliation reaps by label first and gives each quarantined removal an equal share of the pass, so one hung removal no longer starves the rest. The broker socket mount quotes its path, so a temp path with a colon mounts.
- WASM: a canceled run returns the context's error, as docker and E2B do, not a normal result with exit 4294967295. The broker charges a route's call cap last, once nothing else can refuse the call.
- RPC and daemon: a failed run's audit line carries the error text cut to 4 KiB, never a vendor's error page whole;
OpenSessionandCloseSessionrun under the same five-minute ceiling asRun; the orphan reconciler and the isolation re-check end after the drain, not at the signal. - E2B: a session's open now proves, as the guest, that the VM's agent refuses a token-less process start and a token reset; proven on the live service. A step the deadline ends keeps what it wrote.
- Docker Cloud: the in-VM wrapper always writes its trailer, so a guest without
/proccannot forge a line that marks its own run timed out or flooded. - OpenShell and agent-sandbox: the agent-sandbox startup check connects to the pod's control port as the guest and requires both a token-less and a wrong-token start to be refused, or the daemon serves nothing; every result on both providers carries the pinned image digest as its environment identity; a run whose grant relay failed to start is a marked refusal (the guest's code never ran); a sandbox recreated under a session's name ends the session as changed; an ingress policy source must be narrower than everything.
- Records and signing: the check refuses a response whose result kind differs from the request's and an answered record with no result;
plimsoll-attest verify-recordverifies a v0.10.0 to v0.17.x call line with no link; a harness whose write failed appends nothing more; a replay sends a request recorded with a grant profile only with-with-grants;open-statementreads at most one bundle line. A placement backend on another protocol number is unfit, so a rolling upgrade no longer blocks the pool behind one old daemon. - MCP server: a call that waited for its sandbox's turn used to run inside what was left of the call's deadline, and a cut-off lost the sandbox with everything in it; such a call is now refused before it is sent, and runs go under the server's own context. A
setup_failedproject reportsran: "no". Errors shown to the model never name the daemon's address or the receipts directory. An authenticated GET or DELETE on/mcpis 405 before the call meter. - Python and TypeScript clients: the result-kind and time-order checks above; Python catches a recursion error in an error body and hands IPv6 literals to its HTTP client in brackets; the TypeScript decoder of an unanswered record skips unknown fields by wire type.
- Conformance suites: a suspend is proven through a live interpreter; the disk budget case writes past the 64 MiB every session is opened with; one-shot runs add an artifact at a fifo.
Correction to v0.23.0's notes: they said sessions work on E2B and Docker Cloud. Sessions work on docker, E2B and OpenShell; Docker Cloud does not have them yet (sessions).
A signed record protects recorded statements from unnoticed changes. It does not prove that code is correct or safe. Isolation tiers describe configuration and provider evidence plus startup checks, not hardware attestation.