跳转至

Historical version

This page records design, operation, or validation facts from an earlier release. It is not current V7.17.3 guidance. Start from the documentation hub for current instructions.

Codex Cross-Project Long-Term Engineering Assistant V6.4 Validation Report

Subject

  • Version: 6.4.0
  • Target host: native Windows Codex CLI 0.150.1
  • Recommended form: account-level Plugin
  • Upgrade baselines: 6.1.0, 6.2.0, and 6.3.0
  • Event contract: TaskOutcomeEvent 2.0
  • Automatic subagent ceiling: gpt-5.6-terra + high
  • Execution authorization: NONE

Candidate-Package Validation

Check Result Evidence Entry
Package unit and regression tests PASS, 67/67 python -m unittest discover -s tests -v
Shared runtime regressions PASS, 6/6 python -m unittest discover -s runtime/tests -v
Neutral language and structural semantics PASS scripts/semantic-lint.py
Nested Reparse protection in managed trees PASS test_nested_reparse_inside_managed_plugin_tree_is_rejected
Hard-crash Plugin recovery PASS tests/test_package_manager_security.py
Partial-event recovery and cross-segment continuity PASS tests/test_v64_resilience.py
Fail-closed invalid terminal outcome and schema PASS tests/test_v64_resilience.py
Exact actual_model allowlist PASS tests/test_v64_resilience.py
Unified release-validator contract PASS tests/test_v64_release.py
Deterministic ZIP and attestation contract PASS tests/test_v64_release_delivery.py
Correlated host-session acceptance without Hook model fields PASS test_lifecycle_uses_correlated_host_session_model_without_rewriting_hook_facts

The complete package suite ran on 2026-08-28 with 67 tests; the runtime suite had six. Formal machine evidence records actual command output and duration.

Independent Review

Two preimplementation Reviewers covered state/concurrency and compatibility/regression. In postimplementation round one, security and test/delivery Reviewers found three blockers and one nonblocking issue:

  • Explicit invalid terminal outcomes must not degrade to UNKNOWN.
  • Existing events with valid hashes but invalid schemas must be rejected.
  • actual_model must use an exact allowlist.
  • Link-like descendants inside managed trees must be rejected before recursive operations.

All four were repaired with regressions. Round two reused a newly frozen packet for targeted rereview; every original finding closed with no residual finding. Reviewer isolation was logical-readonly. Actual model and reasoning effort for round two were not confirmed by host evidence and remained unverified.

Formal Artifact and Host State

Before formal ZIP construction and real-account upgrade, these remained unverified:

Check Result
Byte-identical double clean build NOT_EXECUTED
Formal ZIP SHA-256 NOT_EXECUTED
Real-account V6.3 -> V6.4 upgrade NOT_EXECUTED
Plugin installed/enabled/version=6.4.0 NOT_EXECUTED
ZIP/Marketplace/cache payload digests agree NOT_EXECUTED
Ten Skills, seven Reviewers, six Hooks discovered NOT_EXECUTED
Five-event lifecycle in a new session NOT_EXECUTED
SessionEnd three-second compatibility NOT_EXECUTED
External unified validation and attestation NOT_EXECUTED

Update this report only after those evidence sources exist. Plans, source presence, or copied files do not replace real host readback.

Security Boundaries

  • Unknown account assets are not deleted.
  • config.toml and main-agent model configuration are not rewritten.
  • Reviewer models are not fixed in TOML.
  • Skills, Reviewers, routing, and business repositories are not modified automatically.
  • Evolution Proposals are not accepted or implemented automatically.
  • Commit, push, deploy, restart, and production operations are not automatic.
  • Project aggregation validates both project_id + repo_fingerprint.
  • Historical Events, Snapshots, Assessments, Proposals, and upgrade backups remain preserved.