chore(release): 0.5.0

.github/release-notes.md

@@ -1,84 +1,181 @@
-<!-- release-version: 0.4.0 -->
+<!-- release-version: 0.5.0 -->
 
 <!-- release-section: highlights:start -->
 
 ## Highlights
 
-Codex Security 0.4.0 adds hosted Standard scan execution and improves token
-accounting, bulk-scan recovery, and report formatting. The minor version marks
-the removal of obsolete ranking helpers; direct helper users should follow the
-migration guidance below.
-
-- Hosts can run Standard repository or path scans through the TypeScript
-  `runHostedScan(input, { executor, signal, onEvent })` API or the
-  `codex-security scan --host` JSON-RPC adapter. Hosted execution is opt-in and
-  requires a host-provided executor, frozen checkout, and explicit model and
-  storage settings. Hosts own authorization, cancellation, and durable execution
-  state. See the [hosted scan guide](https://github.com/openai/codex-security/blob/npm-v0.4.0/sdk/typescript/docs/hosted-scans.md).
-  ([#1559](https://github.com/openai/codex-security/pull/1559))
-- Scan token totals include recorded usage from earlier Deep Scan worker
-  attempts and context compaction without counting resumed history twice.
-  The SDK and compatible hosts can finish accounting for the final response
-  after report generation. Missing records remain marked incomplete or
-  unavailable rather than being presented as an exact total.
-  ([#1657](https://github.com/openai/codex-security/pull/1657))
-- Bulk scans preserve successful results and original scan failures when
-  temporary-checkout cleanup fails. Cleanup errors are reported separately, and
-  reusing a completed scan can retry its recorded cleanup without another scan.
-  ([#1433](https://github.com/openai/codex-security/pull/1433))
-- Reports render malformed evidence ranges as a valid single line or just the
-  path. Finding text beginning with tilde fences or parenthesized list markers
-  no longer changes the formatting of surrounding report sections. These fixes
-  preserve the underlying findings.
-  ([#1656](https://github.com/openai/codex-security/pull/1656),
-  [#1604](https://github.com/openai/codex-security/pull/1604))
-- The package includes Codex CLI/SDK 0.163.0 and restores the **Codex Security**
-  display name in plugin listings, prompts, and tool labels. The npm package
-  remains `@openai/codex-security`, with the `cs` and `codex-security` commands.
-  ([#1670](https://github.com/openai/codex-security/pull/1670),
-  [#1681](https://github.com/openai/codex-security/pull/1681))
+Codex Security 0.5.0 adds investigation time budgets, host-provided finding
+reviews, and more inference provider options. It also improves saved finding
+evidence, usage reporting, and recovery after interrupted scans and publication.
+The minor version marks the retirement of advisory capability preflight;
+custom integrations should follow the migration guidance below.
+
+- Set an investigation allowance with `codex-security scan . --max-time 15`,
+  SDK `maxTimeMinutes`, or project setting `scan.max_time_minutes`. Values are
+  positive minutes, including fractions. Standard and diff scans treat the
+  allowance as advisory; Deep Scan enforces it for discovery, then combines
+  completed findings. Reporting can continue beyond the deadline. See the
+  [time-budget guide](https://github.com/openai/codex-security/blob/npm-v0.5.0/sdk/typescript/docs/cli.md#set-investigation-time).
+  ([#1674](https://github.com/openai/codex-security/pull/1674))
+- Hosts can validate findings against source with `codex-security validate --records`
+  and `validateRecords()`, or classify canonical reports against a supplied
+  severity rubric with `codex-security classify-severity --records` and
+  `classifySeverityRecords()`. Both use host-provided model execution;
+  classification requires no repository or saved scan. See the
+  [validation guide](https://github.com/openai/codex-security/blob/npm-v0.5.0/sdk/typescript/docs/validate-records.md)
+  and [classification guide](https://github.com/openai/codex-security/blob/npm-v0.5.0/sdk/typescript/docs/classify-severity-records.md)
+  for input contracts and host responsibilities.
+  ([#1682](https://github.com/openai/codex-security/pull/1682),
+  [#1662](https://github.com/openai/codex-security/pull/1662))
+- Scans support Cheaper Inference through `--provider cheaperinference`, an
+  explicit model, and `CHEAPER_INFERENCE_API_KEY`. The GitHub Action adds a
+  `provider` input for Amazon Bedrock, OpenRouter, and Fireworks alongside the
+  default OpenAI provider; non-OpenAI providers require an explicit model and
+  the selected provider's credentials. See the
+  [CLI provider documentation](https://github.com/openai/codex-security/blob/npm-v0.5.0/sdk/typescript/docs/cli.md)
+  and [Action setup guide](https://github.com/openai/codex-security/blob/npm-v0.5.0/github-action/README.md#inference-providers).
+  ([#1623](https://github.com/openai/codex-security/pull/1623),
+  [#1677](https://github.com/openai/codex-security/pull/1677))
+- Hosted scans pass `reasoningEffort` strings through to the executor, including
+  `max` and future values supported by the selected model or provider. The
+  GitHub Action's `effort` input also accepts provider-supported strings and
+  retains its `xhigh` default. Deep Scan forwards resolved native settings to
+  fresh and resumed workers while retaining worker-specific permissions.
+  ([#1712](https://github.com/openai/codex-security/pull/1712),
+  [#1717](https://github.com/openai/codex-security/pull/1717))
+- Standard and diff scans use available workers and continue review in the
+  parent when delegation is unavailable. Optional worker or goal-tool
+  availability no longer depends on the removed capability registry.
+  ([#1609](https://github.com/openai/codex-security/pull/1609))
+- Patching and fix verification use complete sealed findings from saved scans,
+  preserving current triage and remediation state. Workbench excerpts stay
+  tied to the source and scope of the saved scan, and publishing older findings
+  preserves their historical root-cause details and code evidence.
+  ([#1608](https://github.com/openai/codex-security/pull/1608),
+  [#488](https://github.com/openai/codex-security/pull/488),
+  [#1756](https://github.com/openai/codex-security/pull/1756))
+- Usage reporting distinguishes recorded tokens from missing worker receipts
+  or unfinished turns. Partial totals display as "at least X tokens; some usage
+  unavailable." Saved scan history includes parent/worker totals and
+  measurement times alongside cost estimates; historical account-limit readings
+  are not a live allowance check. See
+  [progress and cost](https://github.com/openai/codex-security/blob/npm-v0.5.0/sdk/typescript/docs/cli.md#progress-and-cost).
+  ([#1640](https://github.com/openai/codex-security/pull/1640))
+- Bulk-scan recovery validates completed artifacts before reusing them and
+  preserves interrupted receipt bytes. Parallel workers serialize receipt
+  writes, and publication replaces its latest receipt atomically. Completed
+  Linear publication results remain available after late cancellation during
+  a failed final receipt write.
+  ([#1720](https://github.com/openai/codex-security/pull/1720),
+  [#1748](https://github.com/openai/codex-security/pull/1748),
+  [#1735](https://github.com/openai/codex-security/pull/1735),
+  [#1738](https://github.com/openai/codex-security/pull/1738),
+  [#1725](https://github.com/openai/codex-security/pull/1725))
+- Matching GitHub Action distributions publish their manifest, SBOM, and full
+  distribution commit SHA in a companion `action-vX.Y.Z` release. Artifact
+  upload failures propagate the original stream error and cancel active streams.
+  ([#1711](https://github.com/openai/codex-security/pull/1711),
+  [#1671](https://github.com/openai/codex-security/pull/1671))
+- Failed or canceled `dedupe --json --full-output` commands no longer print a
+  false success envelope. Diff scans handle root commits in SHA-256
+  repositories and read blobs with Git versions older than 2.38.
+  ([#1071](https://github.com/openai/codex-security/pull/1071),
+  [#1716](https://github.com/openai/codex-security/pull/1716),
+  [#1715](https://github.com/openai/codex-security/pull/1715))
+- DOCX knowledge-base extraction excludes deleted revisions and field
+  instructions while preserving visible field results.
+  ([#1714](https://github.com/openai/codex-security/pull/1714))
 
 <!-- release-section: highlights:end -->
 
 <!-- release-section: upgrades:start -->
 
 ## Upgrade notes
 
-### Scripts and custom workflows that invoke ranking helpers
-
-The following plugin helper commands are removed:
-
-- `make-repo-rank-input` and `select-deep-review-input`
-- `make-rank-shards`, `validate-rank-shard`, and `merge-rank-outputs`
-- `make-rank-pool-plan`, `validate-rank-worker`, and `validate-rank-pool`
-
-Remove the ranking, shard/pool scheduling, and ranked-selection stages from
-custom workflows. Generate the scoped inventory and review every selected file:
-
-1. For explicit repository scopes, use
-   `scripts/generate_rank_input.py make-repo-scope-input --repo <repo_root> --scopes-file <scopes_json> --out <output_jsonl>`
-   with your Python interpreter and the script from the installed plugin.
-   The scopes file is a JSON array, such as `["src", "tests"]`; use `["."]`
-   for the whole repository instead of the old default `--scope .`.
-   Output rows contain only `path`, for example `{"path":"src/example.ts"}`.
-   Update your consumer to resolve each path against the repository root and
-   read that file directly. If it requires a `path`/`area` review worklist,
-   convert every row to `{"path":"src/example.ts","area":"repository"}`.
-   This inventory contains no `area` or `preview` fields and cannot be passed
-   directly to `copy-deep-review-input`.
-2. For Git diffs, retain `make-diff-rank-input` with its existing revision or
-   local-patch arguments. Copy all candidate rows into the review worklist with
-   `scripts/launch_codex_security_mcp --helper copy-deep-review-input --rank-input <input_jsonl> --out <review_jsonl>`.
-   On Windows, use the `.cmd` launcher.
-3. Review the entire resulting worklist instead of selecting a ranked subset.
-   `bind-repo-scopes` and the retained diff helpers keep their existing
-   arguments and worklist formats.
-
-See the [scoped review reference](https://github.com/openai/codex-security/blob/npm-v0.4.0/plugins/codex-security/skills/security-scan/references/scan-artifacts-and-ledger.md#scoped-deep-review)
-for commands, scope inputs, and Windows invocation details.
-([#1607](https://github.com/openai/codex-security/pull/1607))
-
-Normal CLI and SDK scans do not need this helper migration. Hosted execution is
-opt-in; existing local scans continue to use their current execution path.
+### Investigation time settings
+
+Prefer the new minute-based settings for Standard, diff, and Deep scans. For
+existing Deep Scan configurations, convert hours to minutes:
+
+| Previous setting                      | Preferred setting                    |
+| ------------------------------------- | ------------------------------------ |
+| `--max-time-hours 2`                  | `--max-time 120`                     |
+| SDK `maxTimeHours: 2`                 | SDK `maxTimeMinutes: 120`            |
+| Project `scan.deep.max_time_hours: 2` | Project `scan.max_time_minutes: 120` |
+
+The hour-based settings, including legacy TOML `deep_scan.max_time_hours`, remain
+accepted as deprecated Deep-only aliases. Explicit settings override
+lower-priority configuration across units. At the winning configuration source,
+equivalent durations are accepted; conflicting minute/hour values are errors.
+
+Standard scans still have no time allowance by default. An allowance is neither
+a hard timeout nor a minimum runtime. Deep Scan's default and maximum discovery
+duration remain 5760 minutes (96 hours). Resuming Deep Scan retains its original
+deadline; rerunning starts a new allowance. For batch and component scans,
+project time settings apply separately to each scan attempt, not the campaign
+as a whole. See the
+[time-budget guide](https://github.com/openai/codex-security/blob/npm-v0.5.0/sdk/typescript/docs/cli.md#set-investigation-time).
+([#1674](https://github.com/openai/codex-security/pull/1674))
+
+### Saved scan evidence and recovery
+
+Keep the complete sealed artifacts with saved scans used for patching or fix
+verification. Missing or mismatched evidence now stops those tasks instead of
+passing a shortened preview to the model. Bulk-scan resume validates saved
+contracts and seals before reusing a completed result; invalid or unsuccessful
+results start a new attempt while preserving the old artifacts.
+([#1608](https://github.com/openai/codex-security/pull/1608),
+[#1720](https://github.com/openai/codex-security/pull/1720))
+
+### GitHub Action release assets
+
+If automation downloads the Action manifest or SBOM from an `npm-vX.Y.Z`
+release, switch those downloads to the matching `action-vX.Y.Z` release. The npm
+release remains the canonical changelog and Latest release; Action version tags
+and workflow usage syntax are unchanged. See the
+[Action release documentation](https://github.com/openai/codex-security/blob/npm-v0.5.0/github-action/README.md#development).
+([#1711](https://github.com/openai/codex-security/pull/1711))
+
+### Custom capability-preflight integrations
+
+The plugin's `scripts/config_preflight.py` helper and
+`preflight/capability-profiles.toml` registry are removed. Custom capability
+registries and helper-derived progress inputs and events are no longer supported.
+Update integrations as follows:
+
+1. Remove helper invocations, custom-registry arguments, and gates that wait for
+   a helper `ready` result. Replace references to `config-preflight.md` and
+   `desktop-config-preflight.md` with the
+   [scan runtime guidance](https://github.com/openai/codex-security/blob/npm-v0.5.0/plugins/codex-security/references/scan-runtime.md).
+   Use available delegation tools and successful spawn results; continue in the
+   parent when workers cannot start. Do not edit persistent configuration to
+   enable optional workers or goal tools.
+2. Remove `kind: "preflight"` handling from `onWorkerStatus` consumers,
+   including reads of `delegation` and `configuredSlots`, and stop waiting for
+   the CLI's `worker.preflight` diagnostic. `onWorkerStatus` still emits
+   `kind: "dispatch"` with `phase`, `planned`, and `started`. These counts are
+   best-effort model reports; a missing event does not prove no worker ran.
+3. Stop sending `preflightChecks` to
+   `update_codex_security_scan_progress` and remove the workbench arguments
+   `--preflight-issues-json` and `--preflight-issues-json-stdin`. For host-backed
+   Standard and diff scans, use the authoritative scan context and advance to
+   `threat_model` before architecture mapping. Deep Scan's coordinator continues
+   to own its setup and progress.
+
+Saved preflight history remains readable, and the scan lifecycle still includes
+a `preflight` phase. Local SDK `preflight()` input checks, authentication,
+sandbox enforcement, and Deep worker configuration remain in effect. Normal
+CLI and SDK scans need no capability-preflight configuration migration.
+([#1609](https://github.com/openai/codex-security/pull/1609))
+
+### Selected-profile approval policy
+
+An explicitly selected Codex profile's `approval_policy` now takes precedence
+over the root policy. If you previously relied on a root `never` policy despite
+a selected profile specifying `on-request`, set the selected profile's policy
+to `never` to continue denying permission requests. An absent or null profile
+policy still inherits the root value; Deep Scan workers retain their explicit
+`never` policy.
+([#530](https://github.com/openai/codex-security/pull/530))
 
 <!-- release-section: upgrades:end -->

.github/release-pr-state.json

@@ -1,13 +1,13 @@
 {
-  "baseVersion": "0.3.0",
-  "baseCommit": "5c5239dd8062a7d00916c4bc0d5aeaa3ed0c5477",
+  "baseVersion": "0.4.0",
+  "baseCommit": "559b492b22bb457aea8ff88c1bd41aacc2cec8bf",
   "sections": {
     "highlights": {
-      "generatedHash": "ed7fb1027c60901424f054c5b4b1e7ee19485193f00d37e2ae74f3fb4ba6f2bb",
+      "generatedHash": "07952b04e16a367470314af7e5b21770c4eba8b5b18088876541aaca543f61ec",
       "humanOwned": true
     },
     "upgrades": {
-      "generatedHash": "1b370c365d6e8307d660a71b96785ce8cd8d046a893dc4c3dacc37da4d85e27e",
+      "generatedHash": "afb441fc0f198f1c8bca468973660604d67d352a9a95c307afe16367cbb9279a",
       "humanOwned": true
     }
   }

examples/github-actions/codex-security.yml

@@ -32,7 +32,7 @@ jobs:
       security-events: write
       actions: read
     env:
-      CODEX_SECURITY_VERSION: "0.3.0"
+      CODEX_SECURITY_VERSION: "0.5.0"
       BEDROCK_MODEL_ID: ${{ vars.BEDROCK_MODEL_ID }}
       # Report-only by default. Set to "high" to fail on high/critical findings.
       FAIL_ON_SEVERITY: ""

github-action/package.json

@@ -1,6 +1,6 @@
 {
   "name": "codex-security-action",
-  "version": "0.4.0",
+  "version": "0.5.0",
   "private": true,
   "type": "module",
   "description": "Run Codex Security in GitHub Actions with explicit scope, budgets, and scan health reporting",

sdk/typescript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@openai/codex-security",
-  "version": "0.4.0",
+  "version": "0.5.0",
   "description": "TypeScript SDK and CLI for Codex Security",
   "license": "Apache-2.0",
   "author": "OpenAI",