Verification record

September 18, 2026 · View on GitHub

Environment: Windows, Node.js 24.13.1, Codex CLI 0.153.4. Real TypeSafe API credentials were not available. No live Jev performance or correctness claims are made.

Test-first record

  • Core test batch first failed because the curator did not exist, then passed after implementation.
  • CLI/stdio entrypoint tests first failed because the entrypoint files did not exist, then passed.
  • Follow-up tests exposed explicit result verified: false being ignored and ambiguous source-content statistics; both failed before their fixes.
  • A real long, unbroken log fixture exposed quadratic path matching (about 10.9 seconds for 100,000 characters). The regression failed before changing to bounded, per-token matching; afterward the same case took about 2 ms locally and remained unchanged with evidence_too_large.

Acceptance

  • npm test: 37 passing tests covering API shape, full evidence, uncertainty, atomic batches, missing keys, HTTP 429, network errors, timeout, malformed/missing/out-of-range probabilities, source protection, tool pairs, oversized evidence, quoted filenames, safe CLI parse errors and real stdio JSON-RPC.
  • npm run check: 10 JavaScript syntax checks plus JSON parsing, version and marketplace/MCP launch consistency. This is not a TypeScript compilation claim.
  • Locally bundled official plugin-creator/scripts/validate_plugin.py: passed for plugins/codex-jev-compaction.
  • Locally bundled official skill-creator/scripts/quick_validate.py: passed for skills/jev-handoff.
  • npm run smoke:codex: passed in an isolated temporary CODEX_HOME; real marketplace install, plugin install, skill discovery, ephemeral thread, MCP inventory, MCP tool invocation, and removal. The expected output preserved a protected user instruction and reported zero API requests.
  • npm run benchmark: 50 measured offline iterations, injected fixed probabilities, all retained source blocks exact; content 7,081 → 848 bytes, input JSON 8,044 bytes, output JSON approximately 2,613 bytes and Markdown 1,940 bytes. One local run measured median 0.33 ms, p95 0.79 ms, excluding any real service latency.

The failed host launch experiment with a Claude path variable was corrected before release. The actual tested Codex configuration is command: node, args: ["./scripts/server.mjs"], cwd: ".", resolved relative to the installed plugin root. No path-variable compatibility layer is shipped.

Limits of the evidence

The offline probability fixture is predetermined and cannot evaluate classification quality. Host smoke does not run a model turn, perform real Jev selection, or establish a long-task success rate. Linux/macOS runtime behavior depends on their own published CI or acceptance runs; local verification above occurred on Windows. No real Codex sessions were read or rewritten, and no global plugin configuration was changed during acceptance.

GitHub publication, release assets and installation from the remote repository are separate final delivery steps. The source ZIP contains committed source plus manifests, not a bundled Node or Codex runtime.