Skip to content

Codex Security 0.3.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 15 Oct 04:06
· 57 commits to main since this release
Immutable release. Only release title and notes can be modified.
npm-v0.3.0
5c5239d
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Highlights

Codex Security Local 0.3.0 reduces unnecessary work in Deep Scan, includes more source files, and makes scan exit codes easier to use in scripts. It also improves recovery when scans are interrupted.

  • Deep Scan waits for earlier results to be checked before starting more reviews, avoiding extra reviews when it already has enough information to stop. Fresh API-key and Amazon Bedrock scans also skip a model step during startup when using the bundled plugin and default Node runtime. (#1630, #1555)
  • Source selection now includes text files without recognized extensions, build and test files, and JSP and ASP.NET templates. Fixes also preserve unusual filenames and the original contents of saved Git-diff and directory scans. (#1347, #1414, #1345, #1280)
  • Resumed scans preserve their original knowledge context, model/provider settings, and selected paths. Interrupted scans retain usable results, and Deep Scan cancellation waits for worker cleanup and state updates before returning. Exhausted usage stops workers instead of triggering more retries. (#1556, #1501, #1602, #1260, #1638)
  • The terminal dashboard batches updates during busy scan replays and avoids rereading unchanged logs. It shows token usage even when pricing is unavailable for the selected model. Cost reporting adds Cyber pricing, restores GPT-6 Sol estimates, and includes Codex context-compaction usage. (#1276, #1336, #1585, #1590)

Scan exit codes

Coverage warnings no longer turn a completed scan into a failure. Deep Scans can also finish when an optional detailed writeup is missing; the findings remain in the report.

Exit code Meaning
0 The scan completed without a configured severity-threshold failure.
1 Findings met the configured --fail-on-severity threshold.
2 Execution, result integrity, or required output failed.

A successful exit does not mean every file was reviewed. Check coverage details when your workflow requires that information. (#1633, #1624)

Other CLI and SDK additions

  • Use cs as a shorter alias for codex-security. Policy drafting now works when Codex is installed under your home directory, and app clients can rename and search saved scans. (#1415, #1416, #1180)
  • Generate and verify fixes from TypeScript with security.patch(), using the SDK client's authentication, model settings, and cancellation support. Your application controls review, commits, and publication. When you request a pull request for a supplied issue, CLI results now include its branch and URL. (#1635, #1645)
  • Owner suggestions prefer committed CODEOWNERS before falling back to Git history. In Codex, the new $cloud-findings skill reads findings, reports, and recorded validation from existing Codex Security Cloud scans. (#1519, #1422)

Other fixes cover GitLab merge-request publication, Windows process launch and network-share database paths, configuration errors, YAML error messages, and plugin integrity checks after Python creates cache files. The package includes Codex CLI/SDK 0.162.1. Display names now say Codex Security Local; the package remains @openai/codex-security.

GitHub Action

A new GitHub Action runs repository and pull-request scans in CI. Set the source scope and time and cost budgets, review job summaries, and download saved results. You can upload findings to GitHub code scanning as SARIF and fail a job at a chosen severity.

The initial Action supports GitHub.com, Linux x64 runners, and API-key authentication. Action releases follow the verified npm release. See the Action guide. (#1015, #1418)

Upgrade notes

Read the sections that apply to your scripts, saved scans, or integrations before updating from 0.2.0.

Scripts and integrations that read scan results

If you consume scan JSON or SDK types, replace coverage.completeness, CoverageCompleteness, bulk/component incomplete counters, comparison comparable, and Action scan-status: incomplete handling with manifest.scan.status and exit codes. Action scan-status now uses completed, failed, or skipped. If a later scan does not find a previously reported finding in the code it scanned, comparison returns not_observed. Excluded paths and uncertain matches remain unknown; resolved requires explicit closure as already fixed.

New drafts, imports, worker output, and custom-validation results must use the current codeEvidence, rootCause, attackPath.dataflow, and evidenceRefs fields. Migrate unfinished drafts or finish them with their original version. Completed historical reports remain readable and exportable. (#1633)

Standalone validation

validate now assesses source in a read-only sandbox. It does not build or execute the target, run tests or proofs of concept, or contact services. Save the assessment from CLI stdout or SDK result.report; the validator no longer writes report files:

codex-security validate findings.json > assessment.md

Run a separate reproduction when you need runtime proof. Validation within Standard, Deep, and Diff scans keeps its existing behavior. (#1423)

Resuming existing scans

Keep the private .scan-knowledge.json file with scan artifacts. Resume uses this captured knowledge-base context even if the original documents change or disappear. Older individual scans that used knowledge bases but lack this snapshot require a new scan. Scans without knowledge bases and bulk scans with captured context remain resumable. Reruns use current inputs. (#1556)

Reading compressed .jsonl.zst sessions requires Node 22.15+ within 22.x, or Node 24/26. Node 22.13–22.14 still supports plain logs but cannot resume compressed sessions. (#1428)

Owner suggestion consumers

The owner-suggestion JSON report now uses schemaVersion: 2. Owners can be GitHub users, teams, or email-only identities. Check for a non-null owner, then use "handle" in owner before reading owner.handle or owner.email. Do not assume every owner has a Git name/email pair. (#1519)

Installation and Windows settings

Global installs now create the cs alias. If npm reports EEXIST because another tool owns that name, use npx @openai/codex-security. Windows defaults to the elevated sandbox backend and honors an explicit backend selection. (#1415, #1281)

Custom plugins and direct helper users

Upgrade custom plugins to the current archive/comparison protocols. New and resumed Deep scans also require directDeepScanEngine: 1, workerProviderSnapshot: 4, and isolated workers. Alternatively, omit pluginPath/--plugin-path to use the bundled plugin. Replace direct node ./mcp/server.mjs launches with the current plugin launcher.

For retained findings-storage and severity operations, replace direct calls to retired Python helpers with scripts/launch_codex_security_mcp --helper COMMAND (the .cmd launcher on Windows). Pass JSON on stdin with an absolute stateDirectory and the helper payload. Normal CLI/SDK callers need no helper migration. The selected state directory must be writable; the temporary-state fallback is removed. See the SDK compatibility notes and helper reference. (#1548, #1408, #1543)

Local deduplication and findings service removal

Saved-scan deduplication now uses local SQLite storage by default, without a running service. It caches embeddings and review progress so interrupted work can resume.

The serve command, its --port option, start:server, startFindingsServer, browser dashboard, and findings-service Compose deployment are removed. Use codex-security dedupe --scan SCAN_ID locally, or supply --findings-url for an independently hosted compatible service.

Keep existing state directories and volumes. Select the local database through CODEX_SECURITY_STATE_DIR; a separate service volume is not selected automatically. Generating or refreshing embeddings with the built-in embedder requires an API key. The @openai/codex-security/server entrypoint retains its storage and embedding exports. See the findings guide. (#1340, #1338)

What's Changed

Breaking changes

Features

Fixes

Documentation

Other changes

New Contributors

Full Changelog: npm-v0.2.0...npm-v0.3.0

Contributors

kylos101, mldangelo, and 19 other contributors
Assets 4
Loading