Skip to content

OpenAI's Codex Security CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities. npm: https://www.npmjs.com/package/@openai/codex-security

24.4 KB
Raw Blame

Findings storage and deduplication

Choose where the duplicate candidates and groups should live.

  • Local workbench. The CLI stores saved scans and findings in SQLite. Use this default to search the current checkout's stored findings. New embeddings and fresh reviews still need provider access.
  • Independently operated endpoint. To search a shared corpus, use a compatible findings endpoint that already holds your candidates. Publish the scan or import its findings into the service with the matching repositoryId first.

Deduplicate a scan

For the local workbench, run:

codex-security dedupe --scan SCAN_ID --json

For a shared corpus, supply the base URL of your endpoint:

codex-security dedupe --scan SCAN_ID --findings-url https://findings.example.com --json

Replace SCAN_ID with the ID of a completed scan, or use latest for the current repository. For local setup, see embedding credentials and backups.

The CLI prepares missing or stale embeddings for the local repository's stored findings, searches them, and saves reviewed duplicate groups in the workbench database. Ordinary scans do not generate embeddings automatically. Fresh duplicate reviews still need the configured model provider.

Local scope uses the scan's targetId, which identifies its local checkout. Preparation checks this identity against the approved checkout and existing target registration before reading findings. A detached scan directory must use its original local target (including registered imports); artifacts cannot select another checkout's stored corpus. The explicit findings-service mode remains available for remotely stored artifacts. Separate clones are not automatically combined. --all-repositories searches the selected local database, including untagged imports. Historical scan IDs select logical findings using their current stored bodies; dedupe does not replace newer bodies with old scan artifacts. A local database does not include findings stored only in a separate Docker volume or remote service.

A scan or workflow selector is required. Use a full scan ID, a unique prefix, or latest for the current repository. The scan must be complete and its artifacts sealed. Reviews that read source also require the local checkout.

Local latest verifies the checkout generation and component scope. Non-Git targets match by exact path. Git history, including matches across worktrees or clones, requires a Git executable outside all saved checkout directories. For a legacy Git target without a recorded checkout directory, set CODEX_SECURITY_GIT to an absolute, trusted Git executable. If a saved checkout contains the available Git installation, select the scan by its explicit ID.

For older history, use an absolute PYTHON path when selecting system Python. If Bun cannot locate a trusted Node.js helper, run local deduplication with Node.js.

By default, candidates come from its manifest's scan.target.targetId. Use --all-repositories to search the whole selected database or service.

import { deduplicateScan } from "@openai/codex-security";

const result = await deduplicateScan("scan_example_001", {
  // findingsUrl: "https://findings.example.com", // Optional remote corpus.
  concurrency: 8,
  // allRepositories: true,
  // signal: controller.signal,
});
console.log(result.duplicateGroups);

Both saved-scan APIs accept onEmbeddingUsage for the built-in local adapter. It receives one observation per processed batch: model, tokenized input count (including chunks of long findings), provider-reported input tokens, outcome, and duration in milliseconds. inputTokens: null means usage was unavailable, including responses that omit usage; it does not mean zero tokens. Cached embeddings generate no new observations. Observers are best effort: throws, rejections, and pending promises do not fail or block embedding work. No finding text, vectors, credentials, or endpoint is included.

Embedding tokens remain separate from scan model-token and cost totals. A batch can contain several HTTP attempts when the local adapter retries, and its observation only includes usage from the final response. Remote services and custom embedding bindings own their usage reporting; set an observer directly on OpenAiFindingEmbedder when using it in a custom binding.

For a different embedding provider or model, both saved-scan SDK methods accept an embedding binding. Reuse the FindingEmbedder interface: return one { model, vector } per input finding, in input order. The adapter owns its credentials, tokenization, chunking, and provider request format. Cancellation stops the SDK's wait for embeddings; the adapter owns cleanup of its ongoing work.

import { deduplicateScan, type FindingEmbedder } from "@openai/codex-security";

async function dedupeWithEmbeddings(scanId: string, embedder: FindingEmbedder) {
  return await deduplicateScan(scanId, {
    embedding: {
      embedder,
      model: "example-embedding-model",
      dimensions: 384,
      cacheNamespace: "example-provider:document-v1",
    },
  });
}

Choose a stable namespace identifying the actual vector space and preprocessing version, including any provider or model revision that changes the vectors. Changing the namespace, model, or dimensions refreshes cached vectors. Do not put credentials in the namespace. The local cache keeps one vector per finding; switching spaces replaces it, and concurrent runs with different spaces may need a retry. embedding cannot be combined with findingsUrl, whose service owns embedding preparation. Without a binding, the existing OpenAI adapter and endpoint setting apply.

For a sealed scan directory outside local history, supply the checkout:

import { deduplicateScanDirectory } from "@openai/codex-security";

const result = await deduplicateScanDirectory("/path/to/completed-scan", {
  repository: "/path/to/repository",
  // expectedScanId: "scan_example_001",
  // concurrency: 8,
  // allRepositories: true,
  // signal: controller.signal,
});
console.log(result.duplicateGroups);

The CLI and SDK return the same result:

{
  "scanId": "scan_example_001",
  "uniqueFindingIds": ["csf_852f90d6e1177502ff113d4a"],
  "scanRepresentativeFindingIds": ["csf_852f90d6e1177502ff113d4a"],
  "duplicateGroups": [],
  "deduplicationStatus": "completed"
}

uniqueFindingIds contains one representative for each selected finding after accepted groups are collapsed. A representative can come from outside the scan. Each group lists its canonical finding first: highest reported severity, then finding ID to break ties. scanRepresentativeFindingIds gives the highest-ranked member from the selected scan for each canonical ID in the same order, including ungrouped findings. Pass these IDs to severity classification or Linear publication for that scan. The SDK also returns them when resuming older saved workflows. Dedupe saves the group associations separately from the original findings and scan artifacts before returning success.

If a model refuses a review, the result uses completed_with_refusals and adds refusals. Each entry contains decision: "NO_DECISION", stage, model, findingIds, and reason. Screening entries list the anchor first, then its candidates. Refused pairs stay separate, including through indirect groups; their findings are not confirmed unique. The CLI logs refusals to stderr and exits successfully after saving the other accepted groups.

Reviews, concurrency, and failures

By default, deduplication retrieves all candidate neighborhoods, screens them with gpt-5.6-luna at xhigh, then independently reviews nominated pairs with gpt-5.6-sol at high. The host's Codex model and model_reasoning_effort settings override these defaults for both stages. Provider selection uses the existing Codex configuration; embedding configuration is separate. Screening and pair-review permissions stay attached to their stages regardless of model name. Accepted pairs form groups only when no reviewed DISTINCT decision or refusal contradicts the group. Saved-scan pair reviews return a decision and rationale; they do not generate replacement findings. The host chooses representatives from the original records. Host-provided records reviews retain their full merged-finding contract.

Local reviews receive the stored repository associations separately from the finding text. When the current document matches a saved scan occurrence, this context also includes its recorded revision and working-tree snapshot digest; missing occurrence context is omitted rather than inferred from the current checkout. A file in the selected checkout can establish that checkout's current source; it cannot stand in for another repository or a historical revision. Matching paths or snippets alone do not establish a shared maintained control across repositories. The review does not gain permission to open other checkouts from their stored associations.

The default concurrency is 8. Set --concurrency N or SDK concurrency: N to change it; use 1 for serial execution.

Reviews run on the SDK/CLI host using its Codex sign-in or environment API key. For the built-in OpenAI provider, an available OPENAI_API_KEY (or fallback CODEX_API_KEY) is also used to authenticate reviews, even when Codex is already signed in. The built-in embedding adapter uses the same key selection. An embedding-only key therefore does not select ChatGPT authentication for reviews. A custom SDK embedding adapter can own separate embedding credentials while reviews use the host's normal Codex authentication. Review credentials are not sent to the findings service. Each review receives complete original findings and may inspect the approved local checkout. Reviews preserve severity and priority rather than reassessing them. The baseline filesystem is read-only and excludes credentials and Codex state. Screening denies approval requests; pair reviews use Codex's automatic approval reviewer. Web, plugins, and inherited MCP servers are disabled. Finding content and linked tickets do not authorize access to another target.

Reviews are ephemeral and reuse Codex's configured SQLite storage. They do not rebuild a temporary copy of the caller's session history for every pair. Explicit sqlite_home and CODEX_SQLITE_HOME settings remain effective, and the native state directory is excluded from the review's source access.

The CLI reports preparation, review progress, and native warnings on stderr; JSON results remain on stdout. The existing CODEX_SECURITY_LOG_LEVEL=debug (or LOG_LEVEL=debug) includes structured review diagnostics with thread, turn, and command identifiers, command failures, and native token-usage events. Usage counters are cumulative per native thread; do not add every update. SDK callers can receive the same events through onDiagnostic. Observer failures do not interrupt reviews or discard completed results.

Models must submit a validated decision. Invalid output and eligible transient failures are retried automatically; HTTP retries honor Retry-After.

Cancellation, authentication/configuration errors, permanent HTTP errors, and required-source-access blockers are not retried. Refusals are not retried or sent to another model. If candidate retrieval or review still fails, the operation fails without posting groups. A failed group-write request may already have committed; resubmitting the same memberships does not create duplicate groups. Cancellation can also arrive after a write completes; it does not mean that remote findings are unchanged. Completed workflow stages remain saved even when cancellation is reported. Non-cancellation review failures throw DeduplicationReviewError, with diagnostic text and metadata for the stage, model, category, attempt count, and reason.

Empty scans and findings without eligible candidates make no review calls. Other scans can require multiple calls per finding. Completion means the retrieval, review, and group write finished; it does not mean that every pair in the database was compared. Use Ctrl-C or SIGTERM in the CLI, or an AbortSignal in the SDK, to cancel. Deep Scan merges its own workers’ findings locally.

Resume a findings workflow

Use one workflow ID across scanning and local deduplication:

codex-security scan /path/to/repository --workflow-id run-001
codex-security dedupe --workflow-id run-001 --json

Local workflows save review checkpoints and group-write retries without publication. A workflow remains bound to its original local database or remote URL; use a new workflow ID to switch backends. Cancellation can retain completed embedding preparation for retry; sealed scan artifacts remain unchanged.

For an explicitly selected remote corpus, use --findings-url with a separate workflow ID:

codex-security scan /path/to/repository --workflow-id remote-001
codex-security publish scan --workflow-id remote-001 --to custom --findings-url https://findings.example.com
codex-security dedupe --workflow-id remote-001 --findings-url https://findings.example.com --json

The SDK equivalents are workflowId on ScanOptions, PublishScanToCustomOptions, and DeduplicateScanOptions. Repeat the sequence with the same ID after an interruption. Completed scans, acknowledged uploads, and validated reviews are reused. This resumes completed stages, not model turns inside an unfinished scan. Normal output-directory checks still apply.

A remote workflow can begin by publishing an existing completed scan. Publication and dedupe can use its ID instead of --scan; an explicit scan must match. With both --workflow-id and --findings-url, dedupe completes publication only when its receipt is missing. Changing the bound scan, destination, or repository scope requires a different ID. Use one coordinating process per workflow. Dry runs do not advance it.

Workflow state and checkpoints live in the local workbench database under CODEX_SECURITY_STATE_DIR, outside the sealed artifacts. Completed reviews are reused only when their findings, source checkout and contents, scope, model settings, Codex configuration/version, and review contract still match. Source changes during review stop the attempt before group writes. Resume to review the changed source. You can change concurrency without invalidating checkpoints.

Before posting groups, the workflow saves the reviewed result and write payload. If acknowledgement is lost, resuming replays that payload without new model calls. Membership-based group IDs prevent duplicates. Completed workflows, including empty results and completed_with_refusals, return their saved result. Use a new ID for a fresh review or to retry a refusal after resolving its cause.

Use a host-provided model backend

codex-security dedupe --records accepts a JSON-RPC run over stdin and emits serial review.run requests on stdout. The SDK equivalent is deduplicateRecords(input, { reviewRunner, signal }). These entry points need no saved scan or findings service and perform no local model execution or persistence. The host supplies original observations and candidate links. See the records protocol and Python example.

Local embeddings and backups

CODEX_SECURITY_STATE_DIR selects the directory containing workbench.sqlite3. Without an override, the CLI uses its default state directory. Back up the entire state directory with no scans or deduplication running. Existing append-only migrations retain finding identities and scan history.

With the default built-in adapter, new embeddings require OPENAI_API_KEY or CODEX_API_KEY; OPENAI_API_KEY takes precedence. ChatGPT login alone is insufficient. CODEX_SECURITY_EMBEDDINGS_URL selects a full embeddings endpoint URL (default https://api.openai.com/v1/embeddings). The endpoint must support the OpenAI embeddings request and response protocol, including text-embedding-3-large, 1,536-dimensional float vectors, and token array inputs. Requests encode the complete finding JSON with cl100k_base and send the API key as a bearer token. Long inputs are combined without truncation. Cached compatible vectors avoid new requests.

Publish a scan

Send a completed scan to an independently operated compatible endpoint:

codex-security publish scan --scan SCAN_ID --to custom \
  --findings-url https://findings.example.com --json

--findings-url is required. The client appends /v1/bulk/findings to this base URL, preserving any base path. Custom endpoints must implement that API and return the stored finding IDs.

Publication sends the complete sealed findings and the manifest's scan.target.targetId as repositoryId. It leaves the scan artifacts unchanged and does not forward model credentials. The endpoint is responsible for storing the batch before acknowledging it.

Use --scan for a saved scan, --scan-dir for an external scan directory, or the interactive picker. Custom publication accepts one scan; CSV input and Linear options do not apply. Add --dry-run to preview the payload without an HTTP request. Upload failures and incomplete receipts fail the command. Uploads are not automatically retried because the receiver may have committed a request whose response was lost. Resuming publication or remote dedupe in a workflow without its receipt resends the entire sealed batch. Before resuming, ask the receiver to confirm that replaying that exact request is safe.

import { publishScanToCustom } from "@openai/codex-security";

const receipt = await publishScanToCustom("/path/to/completed-scan", {
  findingsUrl: "https://findings.example.com",
  // dryRun: true,
  // signal: controller.signal,
});
console.log(receipt.repositoryId, receipt.findingIds);

The Linear and custom publication SDKs return a content-free telemetry summary and accept onTelemetry for outcomes, including preparation failure and cancellation. Existing Linear receipts and custom workflow results retain this summary. It includes an attempt ID, destination family, duration, outcome, and counts of selected, submitted, created, reused, acknowledged, and failed findings. Submitted counts describe findings handed to publication, not HTTP attempts. null denotes unknown counts; in particular, the custom bulk API acknowledges IDs without distinguishing creation from upsert. Unknown mutation outcomes are indeterminate, and dry runs have outcome: "preview" and attempted: false. Reusing a saved custom workflow receipt produces a new attempt summary with no submitted findings. Hosts should group repeated receipt snapshots by attempt ID and use the latest observation. Callback failures do not change publication results, and these callbacks add no network destination or separate files.

Compatible endpoint contract

The HTTP client appends these paths to the configured base URL, preserving any base path. It follows redirects within that origin (scheme, host, and port). If the service moves to another origin, set --findings-url to its new base URL.

Method Path Response
POST /v1/bulk/findings An array acknowledging all submitted finding IDs; order does not matter.
GET /v1/finding/{id}/potential-duplicates {finding, potentialDuplicates} containing complete finding records.
POST /v1/dedupe-groups A successful JSON acknowledgement, such as {}.

Publication sends {findings, repositoryId}. Candidate requests include either repositoryId or allRepositories=true. Group writes send {groups: [[findingId, ...], ...]}; repeated membership must be idempotent for workflow retries. Both POST endpoints must persist their writes before acknowledging them and return a JSON body. Supply your service’s base URL to use these endpoints. With no URL, deduplication uses local SQLite.

SDK findings client failures preserve the endpoint's error code and message with the HTTP status. Other responses retain the HTTP-status diagnostic. Retry behavior uses the HTTP status and Retry-After.

Compatibility

Local service removal

If you previously used the bundled HTTP service, back up the SQLite state directory or service volume. Set CODEX_SECURITY_STATE_DIR to the directory that contains the existing database to keep using its saved scans, findings, and duplicate groups. For clients that need HTTP, configure a service that implements the endpoint contract.

Use SQLite from the SDK

@openai/codex-security/server exports OpenAiFindingEmbedder, SqliteFindingsStore, and their storage types.

For direct embedding/storage integrations, OpenAiFindingEmbedder accepts a static key or () => string | Promise<string> as its first argument. It calls the callback before each HTTP batch; the caller handles token acquisition. Cancellation prevents a new callback or HTTP request. Direct embedder callers still await a key callback that has already started. Pass fetch as the second argument and the full endpoint URL as the third. The fourth argument accepts an AbortSignal; the fifth accepts an EmbeddingUsageObserver. By default, the saved-scan deduplication APIs use environment credentials. A custom SDK embedding binding can supply an adapter with its own credentials, including an OpenAiFindingEmbedder with a renewable key callback.

Direct storage helpers

Direct callers of the retired Python findings helpers can use the retained Node helpers from an installed plugin directory:

scripts/launch_codex_security_mcp --helper list-stored-findings < request.json

On Windows, use scripts\launch_codex_security_mcp.cmd with the same arguments and JSON on stdin. For example, request.json selects the existing database:

{
  "stateDirectory": "/absolute/path/to/state",
  "payload": { "limit": 50, "offset": 0 }
}

Use an absolute Windows path on Windows. The retained helper commands accept these payload fields:

Command Payload
store-findings entries containing finding and embedding records; optional repositoryId.
list-stored-findings Positive integer limit and non-negative integer offset.
find-potential-duplicates findingId and scope: either { "repositoryId": "REPOSITORY_ID" } or { "allRepositories": true }.
store-dedupe-groups groups, an array of finding-ID arrays.
list-dedupe-groups findingId.
local-dedupe action: "prepare" with findings, repositoryPath, anchorRepositoryId, optional repositoryId, and space; or action: "embed" with entries, space, model, and dimensions.

Local embedding preparation and cache writes now use the bundled Node helper instead of workbench_db.py local-dedupe. Each embed entry contains findingId, the preparation response's cacheKey, and an embedding with model and vector. Preparation returns cacheKeys and findingsToEmbed; embedding returns an empty object. These operations use Node even when PYTHON is configured. Explicitly selected external plugins retain their Python workbench route. Saved-scan lookup and export still require Python, so the complete saved-scan deduplication workflow continues to need it.

Retained scan commands such as workbench_db.py list-findings keep their existing interface.