Skip to content

Validated baselines

The versions below are validated host-contract baselines, not broad compatibility claims. A future support claim requires a new isolated verification cycle.

Component Validated baseline Evidence
OpenCode v1.18.18 tool.execute.before blocked before a stub effect.
Pi deterministic adapter v0.84.1 tool_call blocked before a stub effect.
Pi judge transport v0.84.2, commit 914cf1472e715297caa30db4b9535d534a9eb718 Temporary loopback probe passed ten transport and session cases.
Node.js test suite 24.9.0 Unit, runtime, integration, privacy, race, and conformance tests.

The Pi 0.84.2 probe used a temporary package, profile, and loopback OpenAI-compatible server. It did not use credentials, contact a real model, or modify the active Pi 0.84.1 installation.

The source-level adapter suite replays shared deterministic and judge corpora. Host doubles cover pre-effect blocking, exact trusted paths, judge allow/deny/unavailable/error/invalid output, session isolation, zero-tool requests, no recursion, model absence, cancellation, monotonic deadlines, synchronous provider blocking, late completion, shadow, off, and OpenCode degradation.

Runtime tests separately exercise host-owned configuration precedence, isolation, sanitized logs, malformed configuration, log-write failure, both host entry points, byte-preserving migration, destination precedence, symlink rejection, and non-directory failures. Isolated profile checks recorded in the compatibility evidence cover source loading and removal.

A supported release must repeat isolated integration tests that prove:

  • the hook blocks before a stub side effect;
  • timeout, cancellation, late completion, model absence, invalid output, and error paths fail closed;
  • configuration does not overwrite user files;
  • removal leaves unrelated configuration unchanged.

Source-profile and loopback results alone do not authorize a package release, real-model inference claim, or broader compatibility claim.