Verification

Verified on 2026-09-16. These checks cover local Windows Codex conversations.

Offline checks

• Windows PowerShell 5.1.19041.7663: 235 assertions passed, including the real 28-second supervisor timeout.
• PowerShell 7.6.5: 233 assertions passed with -SkipSlow.
• All PowerShell scripts/modules passed syntax parsing. No configured linter or installed PSScriptAnalyzer was found.

Tests cover exact command boundaries and CR/LF, ordinary prompt pass-through, removed :test pass-through, exact session selection and rewind links, call/result pairing, repeated actions, duplicate/self-audit exclusion, compaction, partial JSONL, truncation, credential redaction, deterministic budgets, response validation, worker failures/timeouts, compact/detail reports, and unchanged offline requests with -Details.

Installation checks cover repeat install/remove, preservation of unrelated settings, duplicate-hook suppression, and PowerShell launch paths containing spaces, apostrophes, dollar signs and backticks. Four synthetic traces cover clean success, recovery, unresolved failure and insufficient evidence. Expected labels remain outside outgoing state. Fake transport checks test integration behavior, not model accuracy.

Host integration

Interactive Codex CLI 0.153.4 and Codex desktop 0.154.0-alpha.6.2 were checked independently. Audit, ask, help and an invalid ask displayed reports/usage through the blocking response. Structured rollouts and per-turn logs showed zero Codex sampling/provider log entries and zero host-token delta for those commands.

Normal follow-ups produced the requested replies, sampling/provider logs, and host-token deltas of 20,647 in TUI and 198,864 in desktop. These totals include cached input. The user operated the desktop composer and confirmed visible results; GUI automation was unavailable. Native host checks preceded the report refinements and :test removal; the current formatting, flags and removal were verified offline.

Live audit/ask requests resolved to jev-1.13.0 and returned validated judgments and usage. Desktop audit usage was 18,689 input / 334 output tokens; ask usage was 18,245 input / 42 output tokens, each displaying 0.4 seconds request time. TypeSafe traffic was separate from inspected Codex sampling logs and was not packet-captured. These are smoke tests, not calibration or production-accuracy guarantees.

Limits

Other operating systems, host shells and remote/cloud tasks are unverified. Disabled/untrusted hooks, failed executable launches, or host termination before the supervisor responds can prevent interception. Partial traces limit judgments; absent evidence does not establish non-occurrence. Claude Code and OpenCode remain unsupported, with acceptance criteria in TODO.md.