Inspect local Actions permission policies and job dependencies

Start with one workflow you can identify

A maintainer reported missing permission declarations in six jobs across three Viper workflows. T303 handles that review one copied workflow at a time. Select the UTF8 YAML from your own checkout; no repository credentials, token or network access is needed. Repeat separately for the other workflows and keep their source hashes with the exported reviews.

The report preserves the workflow-level declarations and every original job mapping, including steps, run text, conditions and matrices. These are parsed data. Shell commands, HTML strings, local action paths and remote uses are neither run nor fetched. This output is a review record, not an edited workflow to commit.

Read the selected policy before changing a file

A job declaration takes priority over the workflow declaration and supplied default context. An explicit mapping sets every omitted supported scope to none; an empty mapping sets them all to none. A job mapping therefore replaces the workflow mapping. The tool does not merge a workflow grant back into a job that omitted it.

In the checked six-job example, workflow contents: read and issues: write apply to build. Deploy explicitly requests id-token: write, so its contents and issues are none. Empty explicitly declares {}, leaving every scope none. The policy origin column distinguishes job, workflow, provided-default and unknown-default.

• The supported permission keys are fixed to the 2026-10-07 profile; an unknown future key fails rather than disappearing.
• id-token accepts write/none; vulnerability-alerts accepts read/none. Invalid access values fail the complete review.
• Global read-all/write-all stays a literal mode with scopes=null. It is not expanded into invented effective grants.
• The default JSON box is parsed only in mapping mode; its 4KiB UTF8 budget applies in every mode. Explicitly selected invalid context fails even when a workflow policy would take priority.

Separate real dependency edges from unresolved declarations

All declared needs remain in each job record. Only exact existing literal job IDs create dependency edges. Missing targets and expressions have distinct issues. Conditions, matrix declarations and job/step uses preserve their original text and positions; their runtime behavior is unknown.

The example has seven needs declarations and five existing edges. cycleA and cycleB form one exact strongly connected group. The downstream job after is not a cycle member; its missing target and expression remain two explicit issues. A self-edge forms a one-job cycle. This is a graph of declarations, not a simulation of which jobs execute.

Download the entire review and recover safely

The table contains at most 200 jobs and 2,000 UTF16 characters per cell. The summary explains the counts; complete JSON and CSV retain every job, original declaration, use reference and dependency. JSON also contains all graph edges, exact cycles, issues, default provenance and the source SHA256. A short table is not a partial export.

Local native tests deliver 1,000 jobs, 10,000 exact edges and 10,000 uses in one actual source, with all rows checked by a separate PyYAML reader and reachability oracle. A distinct source reaches exactly 10MiB of combined downloads; the next source byte exceeds the output budget. Native progress cancellation discards output, and the same supported input recovers. Browser acceptance is maintained separately by the controller.

• One input: 5MiB, 100,000 resolved value nodes, depth 64, 1,000 jobs, 10,000 needs and 10,000 uses; default JSON 4KiB; two downloads combined 10MiB. All limits apply together.
• Only YAML1.2 core, unique string keys and ordinary finite safe-number/string/boolean/null/mapping/array data are supported. YAML1.1, multiple documents, custom/Set/binary/date tags, unsafe integers, duplicate keys and cyclic aliases fail the whole task.
• Acyclic aliases use yaml2.9.1's weighted expansion guard of 50. It is not a literal promise of 50 references; the tested one-scalar anchor accepts 49 references and rejects the next.
• On error, correct the unsupported structure or actual limit and rerun. On cancellation, no partial success artifacts remain. Originals are never edited.
• Full schema validity, dependency vulnerabilities, permission necessity, live token access and automated repository remediation are outside this review.