42 of 51 criteria have a sourced finding; 40 were also cross-checked by a second agent. Every comparison that includes gh-aw links here for scope, limitations, atomic answers and sources.
Humans approve the rubric, templates and prototype; they do not review every finding. An agent source cross-check is not a hands-on product test. A missing sourced finding is not evidence that a capability is absent.
gh aw ships a native GitHub CLI extension for macOS, Windows and Linux; this is CLI/TUI access, not a separate desktop GUI. It supplies no released product browser or native phone app.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
9 atomic answers
Desktop os [supported] Native release extension binaries cover macOS, Windows and Linux; Windows installer uses Git Bash/MSYS/Cygwin. Sources: gh-aw:gap37-client-boundary.
Mobile apps [unsupported] No native iOS/Android application; Android Termux CLI installation is not a native phone app. Sources: gh-aw:gap37-client-boundary.
Browser [unsupported] GitHub workflow pages are the external platform UI; the extension does not supply an ADE browser client. Sources: gh-aw:gap37-client-boundary.
Mobile browser [unsupported] No product-owned phone-browser ADE interface. Sources: gh-aw:gap37-client-boundary.
Cli [supported] gh aw is a GitHub CLI extension for compiling and operating agentic workflows. Sources: gh-aw:readme.
Editor extensions [partial] No official IDE extension is documented; gh aw init configures VS Code Copilot Chat to use the gh-aw MCP server and an experimental browser compiler playground exists. Sources: gh-aw:d80-editors, gh-aw:d80-mcpsrv.
Task actions [supported] The CLI adds, compiles, runs, audits and updates workflows, and runs start from GitHub triggers or the Actions tab. Sources: gh-aw:d80-cli.
Notifications [partial] Workflows report through issues, comments and run footers; there is no dedicated push channel in gh-aw. Sources: gh-aw:d80-safeout.
Cross device continuity [supported] Runs, issues and pull requests live in GitHub and are reachable from any device with access. Sources: gh-aw:d80-safeout.
Built-in engines and custom imported definitions are separate. Public experimental samples launch Cursor Agent, OpenCode and Aider without compatibility guarantees; these are runnable integrations, not supported built-in engines.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
The released Agy engine is experimental, Linux x64 only, and supports Gemini API keys rather than Google WIF, ADC or Vertex AI. It cannot disable native shell, editing or web tools.
5 atomic answers
Harnesses [supported] Published experimental custom-engine samples launch Cursor Agent, OpenCode 2 and Aider; existing built-in engines retain their own scope. Sources: gh-aw:gap37-experimental-engines.
Model providers [supported] Engines use Copilot inference, Anthropic API keys or workload identity, OpenAI, Gemini and Azure OpenAI BYOK, with model selection and optional routing. Sources: gh-aw:d80-engines, gh-aw:d80-claude, gh-aw:d80-routing.
Custom adapters [supported] Compiler accepts imported custom engine definitions, including the shipped samples. Sources: gh-aw:gap37-experimental-engines.
Session integration [supported] Compiled workflows launch the configured native engine with declared tools and sandbox policy. Sources: gh-aw:engines.
Mixed harness operation [partial] Each workflow picks one engine; different workflows in a repository can use different engines. Sources: gh-aw:d80-engines.
A personal Copilot subscription can use COPILOT_GITHUB_TOKEN. Claude subscription OAuth and ChatGPT subscription login are unsupported. Claude and Codex support API credentials and provider-specific inference routes.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
7 atomic answers
Agent subscriptions [partial] A personal Copilot subscription can use COPILOT_GITHUB_TOKEN. Claude subscription OAuth and ChatGPT subscription login are unsupported. Sources: gh-aw:copilot, gh-aw:claude, gh-aw:codex.
Api accounts [supported] Claude and Codex support API credentials and provider-specific inference routes. Sources: gh-aw:claude, gh-aw:codex.
Vendor credits [partial] Copilot inference is organisation-billed and reported as AI Credits; no gh-aw credits are sold. Sources: gh-aw:d80-cost, gh-aw:d80-copilot.
Local models [partial] engine.api-target and custom engines can point at another endpoint; no local model runtime is provided. Sources: gh-aw:d80-engines.
Account switching [partial] Credentials are repository secrets per workflow, so different workflows can use different accounts. Sources: gh-aw:d80-claude, gh-aw:d80-copilot.
Vendor account requirement [supported] A GitHub repository with Actions and an engine account or Copilot access is required. Sources: gh-aw:d80-quick, gh-aw:d80-billing.
Provider restrictions [supported] A Claude subscription OAuth login is explicitly not supported for the Claude engine. Sources: gh-aw:d80-claude.
Install the extension, select/authenticate an engine, add Markdown workflow source and compile its lock file. The generated workflow runs on GitHub Actions against the configured repository.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
7 atomic answers
First task setup [supported] Install the extension, select/authenticate an engine, add Markdown workflow source and compile its lock file. Sources: gh-aw:readme.
No git [unsupported] A GitHub repository is required: workflows are GitHub Actions workflows. Sources: gh-aw:d80-quick.
Local git without remote [unsupported] The workflow runs only through GitHub Actions on a GitHub remote. Sources: gh-aw:d80-quick.
Hosted git [supported] The generated workflow runs on GitHub Actions against the configured repository. Sources: gh-aw:readme.
Prerequisite guidance [supported] The quick start lists maintainer access and the GitHub CLI. Sources: gh-aw:d80-quick.
Managed cloud setup [supported] GitHub-hosted runners run the agent job by default with no extra setup. Sources: gh-aw:d80-ent.
Byo vps onboarding [partial] A self-hosted runner page documents runs-on, host isolation and enforcement but there is no one-step installer. Sources: gh-aw:d80-shr.
gh aw compile, logs and audit provide workflow authoring and inspection operations. gh aw mcp-server exposes workflow operations, read-only unless write tools are enabled.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
8 atomic answers
Cli [supported] gh aw compile, logs and audit provide workflow authoring and inspection operations. Sources: gh-aw:readme, gh-aw:cost.
Api [partial] Automation goes through the GitHub Actions and API surfaces and the gh-aw MCP server; gh-aw defines no separate service API. Sources: gh-aw:d80-mcpsrv.
Sdk [partial] A WebAssembly build of the compiler can be embedded in a web page. Sources: gh-aw:d80-editors.
Mcp server [supported] gh aw mcp-server exposes workflow operations, read-only unless write tools are enabled. Sources: gh-aw:mcp.
Webhooks [partial] Experimental on.gateway-dispatch subscriptions need a hosted receiver that gh-aw does not deploy. Sources: gh-aw:d80-webhook.
Operations [supported] The CLI covers add, compile, run, logs, audit, update and upgrade for workflows and enterprise scopes. Sources: gh-aw:d80-cli.
Authentication [supported] Access uses gh authentication and GitHub tokens with configurable permissions. Sources: gh-aw:d80-perms.
Stability [partial] Weekly or bi-weekly minor releases with experimental flags; several features are labelled experimental. Sources: gh-aw:d80-release, gh-aw:d80-staged.
Declared MCP servers are filtered through the configured gateway tool allowlist. Engine-specific plugins are supported where listed; agent plugins remain experimental. Gateway allowlists constrain MCP calls independently of Claude permission mode.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
7 atomic answers
Mcp client [supported] Declared MCP servers are filtered through the configured gateway tool allowlist. Sources: gh-aw:claude.
Plugins [supported] Engine-specific plugins are supported where listed; agent plugins remain experimental. Sources: gh-aw:engines.
Custom tools [supported] Custom tools arrive through MCP servers, MCP scripts, custom safe-output jobs and workflow steps. Sources: gh-aw:d80-safeout.
Instructions import [supported] Workflows import shared instruction files and the add command imports workflows from other repositories. Sources: gh-aw:d80-cli.
Authentication [supported] Integration credentials are Actions secrets and the MCP gateway routes authenticated calls. Sources: gh-aw:d80-sandbox.
Permission scope [supported] Gateway allowlists constrain MCP calls independently of Claude permission mode. Sources: gh-aw:claude.
Integration owner [supported] gh-aw owns the MCP gateway and tool configuration while engines own their native tools. Sources: gh-aw:d80-sandbox.
gh-aw is a Go command-line tool distributed as a GitHub CLI extension. A WebAssembly build of its compiler powers an experimental browser playground; there is no desktop or mobile client.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: gh-aw v1.2.2 (399a445, latest stable release; v1.3.0 is a pre-release) compiler, CLI and documentation; source inspection 2026-11-16. Execution depends on GitHub Actions and the chosen engine. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
Partly answered: hybrids (partial).
4 atomic answers
Desktop implementation [supported] Go CLI (gh extension) compiling Markdown workflows to Actions YAML. Sources: gh-aw:d80-cli.
Mobile implementation [not-applicable] No mobile client exists; GitHub mobile shows the resulting issues and pull requests. Sources: gh-aw:d80-editors.
Hybrids [partial] The same compiler runs as a WebAssembly build in the browser playground. Sources: gh-aw:d80-editors.
Implementation evidence [supported] The pinned tree has Go packages under pkg and cmd and a gh-aw-wasm build. Sources: gh-aw:d80-editors.
Workflow Markdown holds natural-language task instructions; YAML frontmatter sets triggers and policy. Issues, PRs, comments and other configured GitHub events can trigger workflows.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
9 atomic answers
Local creation [partial] Workflows are authored as Markdown files and created with gh aw add or create; there is no work item creation outside GitHub. Sources: gh-aw:d80-cli.
External import [supported] Issues, discussions, Linear, Jira and Azure DevOps items are read and written through safe outputs and tools. Sources: gh-aw:d80-safeout.
Chat git triggers [supported] Issues, PRs, comments and other configured GitHub events can trigger workflows. Sources: gh-aw:triggers.
Authoritative tracker [supported] GitHub issues and projects remain authoritative for work items. Sources: gh-aw:d80-safeout.
Sync [partial] Safe outputs write to the tracker; the work queue can mirror tasks to issues but the queue log is authoritative. Sources: gh-aw:d80-workq.
Supported edits [supported] Safe outputs create, update, close, label, assign and link issues. Sources: gh-aw:d80-safeout.
External roundtrips [partial] Operations without a safe-output type require the tracker UI. Sources: gh-aw:d80-safeout.
gh-aw has no epic object: GitHub sub-issue links and the native work queue, which supports task dependencies and fair scheduling, give limited hierarchy and dependency handling.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: gh-aw v1.2.2 (399a445, latest stable release; v1.3.0 is a pre-release) compiler, CLI and documentation; source inspection 2026-11-16. Execution depends on GitHub Actions and the chosen engine. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Epics [unknown] Hierarchy, dependencies and evolving scope, epics: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Hierarchy [partial] link-sub-issue builds parent and child issue links; depth is GitHub own. Sources: gh-aw:d80-safeout.
Blockers [partial] The work queue supports task dependencies; issues use GitHub sub-issue links. Sources: gh-aw:d80-workq, gh-aw:d80-safeout.
Readiness [partial] The queue dispatches tasks whose dependencies are satisfied. Sources: gh-aw:d80-workq.
Enforcement [unknown] Hierarchy, dependencies and evolving scope, enforcement: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
State propagation [unknown] Hierarchy, dependencies and evolving scope, state_propagation: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Task to epic [unknown] Hierarchy, dependencies and evolving scope, task_to_epic: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Identity retention [unknown] Hierarchy, dependencies and evolving scope, identity_retention: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Workflow constraints [unknown] Hierarchy, dependencies and evolving scope, workflow_constraints: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Skills [unknown] Shared context and memory, skills: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Project docs [unknown] Shared context and memory, project_docs: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Memory [partial] Cache-memory persists declared files between runs with workflow scope/retention. Sources: gh-aw:memory.
Scope [supported] Cache memory is scoped per workflow and repo memory per branch and id. Sources: gh-aw:d80-cache, gh-aw:d80-repomem.
Provenance [unknown] Shared context and memory, provenance: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Correction [unknown] Shared context and memory, correction: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Deletion [unknown] Shared context and memory, deletion: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Freshness [partial] Cache memory is evicted after 7 days of disuse; repo memory has unlimited retention. Sources: gh-aw:d80-cache, gh-aw:d80-repomem.
Takeover continuity [supported] Repo memory on a Git branch and repository files persist for any maintainer. Sources: gh-aw:d80-repomem.
The native work queue (tools.work-queue) keeps Work items with dependencies, priorities and attempt state in a Git-backed ledger, separate from individual workflow runs. Operators submit, list, filter, search, dispatch, cancel and reprioritize Work with gh aw work-queue; GitHub Issues can mirror items but are not the store.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-10 (12 days old)
Agent cross-check
Not separately cross-checked
Discussion, labels, assignment to humans, archive and reopen were not established for queue Work. The queue is agent-worker infrastructure rather than a human backlog UI. No installed-product test was run.
17 atomic answers
Create [supported] submit-work and submit-graph admit Work items, optionally as a dependency graph. Sources: gh-aw:src1-workqueue-ops.
Edit [partial] Queue Work is changed only through administrator reprioritize and cancel operations; update-issue edits GitHub Issues, which are separate from queue Work. Sources: gh-aw:d80-workq, gh-aw:d80-safeout.
Discuss [partial] The queue has no comment thread; backing Issues mirror queue Work and carry ordinary GitHub comments. Sources: gh-aw:d80-workq, gh-aw:d80-safeout.
Prioritize [supported] Work has a priority (default 3); administrators can reprioritize Work from 1 to 5. Sources: gh-aw:src1-workqueue-ops.
Label [partial] Backing Issues get a tracking label and work:<status> labels projected from queue state; add-labels edits GitHub Issue labels separately. Sources: gh-aw:d80-workq, gh-aw:d80-safeout.
Assign [partial] dispatch-next commits claim assignments to workers; human assignees exist only on GitHub Issues through assign-to-user. Sources: gh-aw:d80-workq, gh-aw:d80-safeout.
Search [supported] state --search TEXT searches the queue tree. Sources: gh-aw:src1-workqueue-ops.
Archive [partial] cancel-work and clear remove unfinished Work and compact checkpoints the ledger; close-issue closes GitHub Issues separately. Sources: gh-aw:d80-workq, gh-aw:d80-safeout.
Reopen [unknown] Native issue management and idea backlog, reopen: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Project epic views [unknown] Native issue management and idea backlog, project_epic_views: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Idea parking [unknown] Native issue management and idea backlog, idea_parking: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Explicit dispatch [supported] dispatch-next commits assignments and dispatch activates a verified dispatcher workflow for queued Work. Sources: gh-aw:src1-workqueue-ops.
Task session separation [supported] Work is an immutable task definition; each attempt is a separate Claim bound to a workflow run. Sources: gh-aw:src1-workqueue.
Metadata [unknown] Native issue management and idea backlog, metadata: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Durable states [supported] The version-3 transaction log on the queue branch is authoritative and records Claims, Completions and Results. Sources: gh-aw:src1-workqueue.
External tracker requirement [supported] No external tracker is required: Issues can mirror Work but cannot store the queue. Sources: gh-aw:src1-workqueue.
Agents operate on the same GitHub issue and pull request numbers humans use, through safe outputs that run as separate permission-limited jobs with bot attribution and footers, and optional staged preview.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: gh-aw v1.2.2 (399a445, latest stable release; v1.3.0 is a pre-release) compiler, CLI and documentation; source inspection 2026-11-16. Execution depends on GitHub Actions and the chosen engine. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Workflows on distinct issues/PRs/branches may run concurrently; default per-engine groups serialize each engine.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
5 atomic answers
Concurrent sessions [supported] Workflows on distinct issues/PRs/branches may run concurrently; default per-engine groups serialize each engine. Sources: gh-aw:concurrency.
Concurrent tasks [supported] Workflows on different issues, PRs or branches run concurrently under per-workflow groups. Sources: gh-aw:d80-conc.
Quotas [supported] engine.concurrency limits agent jobs per engine across workflows and max-turns and max-ai-credits bound each run. Sources: gh-aw:d80-conc, gh-aw:d80-engines.
Observed concurrency [not-applicable] The catalog runs no products. Sources: gh-aw:d80-conc.
Status visibility [supported] Run status is shown in the GitHub Actions UI and gh aw logs. Sources: gh-aw:d80-audit.
Claims [partial] Work-queue claims scope writes to a task. Sources: gh-aw:d80-workq.
Queues [supported] queue:max and cancel-in-progress settings determine pending-run retention and obsolete-PR cancellation. Sources: gh-aw:concurrency.
Routing [partial] Triggers and label commands route work to workflows. Sources: gh-aw:d80-conc.
Duplicate prevention [supported] Concurrency groups and queue claims prevent duplicate runs on the same entity. Sources: gh-aw:d80-conc, gh-aw:d80-workq.
Ownership enforcement [partial] Queue claims and concurrency groups enforce exclusivity per run. Sources: gh-aw:d80-workq, gh-aw:d80-conc.
The product supplies producer/dispatcher agent roles and native queue tools for submitting tasks, dispatching approved worker workflows and inspecting task/claim/result snapshots.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
unknown
Delivery
unknown
Checked
2026-11-13 (9 days old)
Agent cross-check
2026-11-17
Source review and pinned implementation inspection only; no installed-product, account or multi-user scenario was executed. Unanswered atomic questions remain open.
12 atomic answers
Roles [supported] Product-owned producer/dispatcher and worker roles have distinct tools and authority. Sources: gh-aw:gap37-native-delegation.
Model harness effort [supported] Each workflow sets engine, model and reasoning effort, with optional routing. Sources: gh-aw:d80-engines, gh-aw:d80-routing.
Messages [partial] Workflows exchange work through dispatch-workflow, call-workflow and repository dispatch. Sources: gh-aw:d80-safeout.
Results return [supported] Verified Results and Claims are retained in the authoritative Git-backed queue and visible in activation snapshots. Sources: gh-aw:gap37-native-delegation.
Orchestration owner [supported] Native work-queue submission/dispatch and verified worker routes are product orchestration rather than provider-native subagents. Sources: gh-aw:gap37-native-delegation.
Activation [supported] Delegation needs no flag beyond configured safe outputs. Sources: gh-aw:d80-safeout.
Authentication [supported] Each dispatched workflow uses its own secrets and permissions. Sources: gh-aw:d80-perms.
Live progress [supported] Run logs stream in the Actions UI. Sources: gh-aw:d80-audit.
Schedule frontmatter can run a recurring workflow such as a daily report. GitHub events can start automation without a human mentioning the agent on every run. Schedules execute through GitHub Actions rather than a local browser or desktop scheduler.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
8 atomic answers
Schedules [supported] Schedule frontmatter can run a recurring workflow such as a daily report. Sources: gh-aw:triggers.
Events [supported] GitHub events can start automation without a human mentioning the agent on every run. Sources: gh-aw:triggers.
Recurring tasks [supported] Schedule triggers use a documented cron syntax. Sources: gh-aw:d80-eph.
Completion conditions [partial] stop-after disables a workflow at a deadline; runs end at timeout or turns. Sources: gh-aw:d80-eph, gh-aw:d80-frontmatter.
Scheduler location [supported] Schedules execute through GitHub Actions rather than a local browser or desktop scheduler. Sources: gh-aw:readme, gh-aw:triggers.
Overlaps [supported] Concurrency groups serialize or cancel overlapping runs per trigger context. Sources: gh-aw:d80-conc.
Retries [unknown] Unattended triggers, retries: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Client independence [supported] Schedules run on GitHub and need no client. Sources: gh-aw:d80-rate.
gh-aw runs unattended on GitHub: runs are stopped from the Actions UI or by concurrency groups, approvals come from environment protection, and failures show in logs and audit reports. There is no mid-run steering.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: gh-aw v1.2.2 (399a445, latest stable release; v1.3.0 is a pre-release) compiler, CLI and documentation; source inspection 2026-11-16. Execution depends on GitHub Actions and the chosen engine. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Pending decisions [supported] Environment protection can require approvals before safe outputs apply. Sources: gh-aw:d80-safeout.
Escalation [partial] noop and missing-tool reports surface problems; no paging. Sources: gh-aw:d80-safeout.
Visible errors [supported] Actions logs and audit reports show failures. Sources: gh-aw:d80-audit.
Chat terminal parity [unknown] Intervention and attention, chat_terminal_parity: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Retained history [supported] Run logs and artifacts remain per Actions retention. Sources: gh-aw:d80-audit.
A workflow is natural-language Markdown with engine, model and routing in frontmatter; changes take effect on the next run after recompilation. There is no mid-run reconfiguration.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: gh-aw v1.2.2 (399a445, latest stable release; v1.3.0 is a pre-release) compiler, CLI and documentation; source inspection 2026-11-16. Execution depends on GitHub Actions and the chosen engine. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Quick task [supported] A single Markdown workflow is enough. Sources: gh-aw:d80-quick.
Growing epic [unknown] Workflow flexibility and orchestration visibility, growing_epic: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Presets [partial] Imports and gallery workflows act as presets. Sources: gh-aw:d80-cli.
Natural language [supported] Workflows are written in natural-language Markdown and the compiler turns them into Actions YAML. Sources: gh-aw:d80-quick.
Agent configuration [supported] Engine, model, effort and routing are set in frontmatter. Sources: gh-aw:d80-engines, gh-aw:d80-routing.
In run changes [unknown] Workflow flexibility and orchestration visibility, in_run_changes: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Live roles [unknown] Workflow flexibility and orchestration visibility, live_roles: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Retained relationships [unknown] Workflow flexibility and orchestration visibility, retained_relationships: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Discoverability [supported] Documentation, gh aw help and the gallery explain options. Sources: gh-aw:d80-cli.
Compiled workflows execute on GitHub Actions runners, including user-controlled Linux runners. A runner on the user computer executes there; other self-hosted runners and GitHub-hosted runners remain separate options.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
12 atomic answers
Local [supported] A user-managed Linux runner on the user computer executes the compiled agent job there; compilation by the CLI alone is not agent execution. Sources: gh-aw:gap37-client-boundary.
Byo remote [supported] Self-hosted Linux container-capable runners can execute the compiled workflow. Sources: gh-aw:faq.
Self hosted [partial] Self-hosted Linux runners work; a GHES control plane is compatibility-tested only and its end-to-end availability is stated as unconfirmed. Sources: gh-aw:d80-shr, gh-aw:d80-ent.
Managed cloud [supported] On GitHub-hosted Actions runners the agent runs in an AWF Docker container by default. Sources: gh-aw:arch-runtime, gh-aw:sandbox.
Hybrid [supported] Workflows choose GitHub-hosted or self-hosted runners per workflow. Sources: gh-aw:d80-shr.
Provisioning [unsupported] gh-aw provisions no machines; GitHub or the user provides runners. Sources: gh-aw:d80-shr.
Payer [supported] Actions minutes bill the repository owner and inference bills the engine provider. Sources: gh-aw:d80-billing.
Resources [partial] Runner size is chosen by runs-on. Sources: gh-aw:d80-frontmatter.
Region [unknown] Laptop-independent execution, region: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Expiry [supported] The agent step times out after 20 minutes by default and jobs after 360 minutes. Sources: gh-aw:d80-rate.
Client disconnect [supported] Event and scheduled workflows execute on the Actions runner after the initiating client disconnects. Sources: gh-aw:triggers, gh-aw:readme.
Other device continuation [supported] Runs are visible and continue on GitHub regardless of device. Sources: gh-aw:d80-rate.
AWF Docker isolation is the default; opt-in Cloud Hypervisor/NVX preview runtimes use KVM microVMs on eligible runners. The compiler does not provision Git worktrees.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
MicroVM runtimes are previews limited to eligible KVM runners; the default is a Docker container.
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
12 atomic answers
Checkout [supported] Runs use a repository checkout on the runner. Sources: gh-aw:d80-sandbox.
Branch [unknown] Workspace and process isolation, branch: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Worktree [unsupported] Checkout/branch steps use the workflow workspace, not a product worktree isolation option. Sources: gh-aw:gap37-client-boundary.
Container [supported] The default docker runtime runs the agent in a rootless AWF container. Sources: gh-aw:sandbox.
Vm [supported] Preview cloud-hypervisor (GitHub-hosted Ubuntu x86_64 KVM runners) and nvx runtimes place the agent in a KVM microVM. Sources: gh-aw:sandbox, gh-aw:arch-runtime.
Os sandbox [supported] The AWF firewall sandbox is the default agent sandbox. Sources: gh-aw:d80-sandbox.
Filesystem [supported] The agent runs in an AWF container with a restricted environment. Sources: gh-aw:d80-sandbox.
Process [supported] The agent process is separate from safe-output jobs. Sources: gh-aw:d80-safeout.
Network [supported] AWF network isolation sends egress only through the proxy to configured domains. Sources: gh-aw:sandbox, gh-aw:arch-runtime.
Secrets [partial] Secrets stay in Actions and write permissions only in safe-output jobs. Sources: gh-aw:d80-perms.
Code separation [supported] Each run uses its own runner workspace. Sources: gh-aw:d80-sandbox.
Security boundary [partial] The docs state the sandbox does not make a self-hosted host clean. Sources: gh-aw:d80-shr.
Experimental engine.durable saves/restores native session checkpoints in artifacts, default seven days subject to repository retention. Checkpoint format and resume behavior differ by engine; durable mode does not automatically dispatch a continuation.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
12 atomic answers
Client disconnect [supported] Runs continue on the runner after any client disconnects. Sources: gh-aw:d80-rate.
Worker crash [unknown] Recovery, expiry and continuity, worker_crash: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Service restart [unknown] Recovery, expiry and continuity, service_restart: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Expiry [supported] Timeouts and cache eviction after 7 days define expiry. Sources: gh-aw:d80-cache, gh-aw:d80-rate.
Transcript reload [supported] Run logs and artifacts reload through gh aw logs and audit. Sources: gh-aw:d80-audit.
Process reconnect [unknown] Recovery, expiry and continuity, process_reconnect: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Task resume [supported] Experimental engine.durable saves/restores native session checkpoints in artifacts, default seven days subject to repository retention. Sources: gh-aw:engines.
State reconstruction [partial] Artifact-backed durable mode is experimental. Sources: gh-aw:d80-engines.
Context loss [partial] Checkpoint format and resume behavior differ by engine; durable mode does not automatically dispatch a continuation. Sources: gh-aw:engines.
Duplicate actions [unknown] Recovery, expiry and continuity, duplicate_actions: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Manual repair [unknown] Recovery, expiry and continuity, manual_repair: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
gh-aw leaves dependency setup to workflow steps; it provides Actions caches, cache and repo memory and secrets through Actions.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: gh-aw v1.2.2 (399a445, latest stable release; v1.3.0 is a pre-release) compiler, CLI and documentation; source inspection 2026-11-16. Execution depends on GitHub Actions and the chosen engine. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Dependencies [unknown] Development environment reliability, dependencies: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Snapshots [unknown] Development environment reliability, snapshots: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Services [unknown] Development environment reliability, services: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Ports [unknown] Development environment reliability, ports: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Duplicate installs [unknown] Development environment reliability, duplicate_installs: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Collisions [unknown] Development environment reliability, collisions: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Remote setup [supported] Setup happens in workflow steps on the runner. Sources: gh-aw:d80-sandbox.
Automatic repair [unknown] Development environment reliability, automatic_repair: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Top-level max-turns defaults to 500 for supported engines; Agy lacks this limit. max-ai-credits bounds catalog-estimated inference consumption through the proxy. Claude recovery resumes the exact session; silent fresh restart after progress requires an explicit override.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
8 atomic answers
Time limits [supported] timeout-minutes defaults to 20 for the agent step. Sources: gh-aw:d80-frontmatter.
Turn limits [partial] Top-level max-turns defaults to 500 for supported engines; Agy lacks this limit. Sources: gh-aw:engines.
Token limits [partial] max-ai-credits bounds inference spend per run; no raw token cap. Sources: gh-aw:d80-engines.
Spend limits [supported] max-ai-credits bounds catalog-estimated inference consumption through the proxy. Sources: gh-aw:cost.
Timeouts [supported] Agent and job timeouts are enforced by Actions. Sources: gh-aw:d80-rate.
Retries [supported] Claude recovery resumes the exact session; silent fresh restart after progress requires an explicit override. Sources: gh-aw:claude.
Provider failures [partial] Retries are configured per engine. Sources: gh-aw:d80-engines.
Usage export [unknown] Truthful status and operational statistics, usage_export: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Task stats [unknown] Truthful status and operational statistics, task_stats: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Agent stats [partial] gh aw logs summarise runs by workflow and AI credits. Sources: gh-aw:d80-cost.
Project stats [unknown] Truthful status and operational statistics, project_stats: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Time stats [partial] Run durations appear in logs. Sources: gh-aw:d80-audit.
Failure status [supported] Actions run conclusions distinguish failed and cancelled runs. Sources: gh-aw:d80-audit.
Approval status [unknown] Truthful status and operational statistics, approval_status: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Reconnect consistency [unknown] Truthful status and operational statistics, reconnect_consistency: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Performance method [not-applicable] This catalog measures no performance. Sources: gh-aw:d80-audit.
Responsiveness [not-applicable] No hands-on responsiveness is measured by this catalog. Sources: gh-aw:d80-audit.
AI Credits estimate inference per workflow run; Actions compute is a separate cost component. Credit estimates use catalog prices and do not reconcile the providers invoice. Structured logs JSON includes per-run aic and episode total_aic fields.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
14 atomic answers
Task cost [supported] AI Credits are reported per run. Sources: gh-aw:d80-cost.
Session cost [supported] AI Credits estimate inference per workflow run; Actions compute is a separate cost component. Sources: gh-aw:cost.
Child cost [unknown] Per-task API and compute cost, child_cost: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Epic cost [unknown] Per-task API and compute cost, epic_cost: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Harness aggregation [unknown] Per-task API and compute cost, harness_aggregation: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Billed vs estimated [partial] Credit estimates use catalog prices and do not reconcile the providers invoice. Sources: gh-aw:cost.
Tokens [supported] Audit reports show raw token counts. Sources: gh-aw:d80-cost.
Subscription allowance [partial] Copilot inference is metered in AI Credits. Sources: gh-aw:d80-cost.
Provider reconciliation [unknown] Per-task API and compute cost, provider_reconciliation: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Compute cost [partial] Actions minutes are billed by GitHub and not itemised by gh-aw. Sources: gh-aw:d80-billing.
Refresh delay [unknown] Per-task API and compute cost, refresh_delay: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Alerts [unsupported] max-ai-credits is a limit, not an alert. Sources: gh-aw:d80-engines.
Export [supported] Structured logs JSON includes per-run aic and episode total_aic fields. Sources: gh-aw:cost.
Drilldown [unknown] Per-task API and compute cost, drilldown: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Workflow and engine concurrency groups bound overlap and preserve queued work where configured. PR updates can cancel outdated runs; mixed trigger queue behavior is documented separately.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
12 atomic answers
Concurrency [supported] Workflow and engine concurrency groups bound overlap and preserve queued work where configured. Sources: gh-aw:concurrency.
Queues [supported] The work queue provides durable queued tasks. Sources: gh-aw:d80-workq.
Backpressure [unknown] Compute-aware scheduling and verification, backpressure: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Host awareness [unknown] Compute-aware scheduling and verification, host_awareness: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Check reuse [unknown] Compute-aware scheduling and verification, check_reuse: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Check deduplication [unknown] Compute-aware scheduling and verification, check_deduplication: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Bounded tests [unknown] Compute-aware scheduling and verification, bounded_tests: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Cancellation [supported] PR updates can cancel outdated runs; mixed trigger queue behavior is documented separately. Sources: gh-aw:concurrency.
Fairness [supported] The work queue has fair scheduling. Sources: gh-aw:d80-workq.
Recovery [unknown] Compute-aware scheduling and verification, recovery: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Shared limits [supported] engine.concurrency applies across workflows. Sources: gh-aw:d80-conc.
The scoped workflow runtime accepts its invocation inputs and does not supply multiple humans with a shared live prompting channel. GitHub discussion and follow-on workflows are external platform interactions.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
unknown
Delivery
unknown
Checked
2026-11-13 (9 days old)
Agent cross-check
2026-11-17
Source review and pinned implementation inspection only; no installed-product, account or multi-user scenario was executed. Unanswered atomic questions remain open.
11 atomic answers
Visibility [supported] Issues, comments and PRs are visible to repository members. Sources: gh-aw:d80-safeout.
Discussion [supported] Comments and discussions carry conversation. Sources: gh-aw:d80-safeout.
Comments [supported] Issue and PR review comments are written and read. Sources: gh-aw:d80-safeoutpr.
Presence [unknown] Live human collaboration, presence: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Coprompting [unsupported] No product shared live-prompt channel to an existing agent run in the scoped extension/runtime. Sources: gh-aw:gap37-client-boundary.
Ordering [unknown] Live human collaboration, ordering: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Conflicts [unknown] Live human collaboration, conflicts: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Attribution [supported] GitHub attributes each actor. Sources: gh-aw:d80-safeout.
Human authentication and repository authorization belong to GitHub. Native queue observer/producer/worker roles are agent authority, not human workspace roles or product SSO.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
unknown
Delivery
unknown
Checked
2026-11-13 (9 days old)
Agent cross-check
2026-11-17
Source review and pinned implementation inspection only; no installed-product, account or multi-user scenario was executed. Unanswered atomic questions remain open.
9 atomic answers
Organization roles [unsupported] No separate gh-aw human workspace/team-role model. Sources: gh-aw:gap37-client-boundary.
Project roles [unknown] Roles and access boundaries, project_roles: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Task roles [unknown] Roles and access boundaries, task_roles: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Membership [unknown] Roles and access boundaries, membership: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Sso [unsupported] No product SSO implementation; GitHub login and enterprise authentication are external platform dependencies. Sources: gh-aw:gap37-client-boundary.
Provisioning [unknown] Roles and access boundaries, provisioning: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Unauthorized access [unknown] Roles and access boundaries, unauthorized_access: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Revocation [unknown] Roles and access boundaries, revocation: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Plan scope [unknown] Roles and access boundaries, plan_scope: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
The gateway applies explicit tool allowlists; engine-level tool enforcement differs by engine. Firewall allowlists and scoped safe-output jobs provide configured isolation/write controls. Agents default to read-only access; GitHub mutations normally occur in separately scoped safe-output jobs.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
14 atomic answers
Command [unknown] Approvals and temporary access, command: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Tool [supported] The gateway applies explicit tool allowlists; engine-level tool enforcement differs by engine. Sources: gh-aw:claude, gh-aw:codex.
Network [supported] Firewall allowlists and scoped safe-output jobs provide configured isolation/write controls. Sources: gh-aw:architecture, gh-aw:sandbox.
Secret [unknown] Approvals and temporary access, secret: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Delivery [supported] Environment protection applies to safe-output jobs. Sources: gh-aw:d80-safeout.
Approver identity [supported] GitHub environment reviewers are identified by GitHub. Sources: gh-aw:d80-safeout.
Lifetime [unknown] Approvals and temporary access, lifetime: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Deny [supported] Unlisted domains and outputs are blocked. Sources: gh-aw:d80-network.
Revoke [unknown] Approvals and temporary access, revoke: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Tunnels [unknown] Approvals and temporary access, tunnels: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Defaults [supported] Agents default to read-only access; GitHub mutations normally occur in separately scoped safe-output jobs. Sources: gh-aw:readme.
Safer modes [supported] Staged mode previews writes and read-only permissions are the default. Sources: gh-aw:d80-staged, gh-aw:d80-perms.
Enforcement [supported] The compiler and safe-output jobs enforce permissions, network and output limits. Sources: gh-aw:d80-perms, gh-aw:d80-network.
History lives in GitHub: issues, PRs, comments, Actions logs, repo memory commits and audit artifacts; there is no separate gh-aw audit store.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: gh-aw v1.2.2 (399a445, latest stable release; v1.3.0 is a pre-release) compiler, CLI and documentation; source inspection 2026-11-16. Execution depends on GitHub Actions and the chosen engine. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Actions [supported] Actions logs, footers and audit reports record actions. Sources: gh-aw:d80-audit.
Approvals [supported] Environment approvals are recorded by GitHub. Sources: gh-aw:d80-safeout.
Edits [supported] Git history records edits. Sources: gh-aw:d80-repomem.
Discussion [unknown] Decision and activity history, discussion: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Decisions [unknown] Decision and activity history, decisions: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Cross user persistence [unknown] Decision and activity history, cross_user_persistence: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Cross session persistence [supported] Repo memory and artifacts persist across runs. Sources: gh-aw:d80-repomem.
Export [partial] Audit artifacts can be downloaded. Sources: gh-aw:d80-audit.
Mutability [unknown] Decision and activity history, mutability: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Retention [partial] Run retention follows Actions settings and cache eviction is 7 days. Sources: gh-aw:d80-cache.
Integrity [unknown] Decision and activity history, integrity: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
The extension exposes workflow logs, audit reports and output artifacts, not a product UI for reviewing proposed code diffs. GitHub pull-request review is separate.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
13 atomic answers
Diffs [unsupported] No product-owned proposed-diff review UI in the scoped extension/runtime; reports and GitHub PR pages are separate. Sources: gh-aw:gap37-client-boundary.
Documents [supported] upload-artifact and upload-asset attach outputs. Sources: gh-aw:d80-safeout.
Screenshots [unknown] Retained artifacts and outputs, screenshots: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Recordings [unknown] Retained artifacts and outputs, recordings: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Previews [unknown] Retained artifacts and outputs, previews: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Test outputs [unknown] Retained artifacts and outputs, test_outputs: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Reviewer outputs [supported] PR review and review comment outputs retain reviewer findings on the PR. Sources: gh-aw:d80-safeoutpr.
Gallery [unknown] Retained artifacts and outputs, gallery: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Task links [supported] Outputs link to the issue or PR. Sources: gh-aw:d80-safeout.
Persistence [supported] Workflow logs and artifacts retain run/session evidence with configured retention. Sources: gh-aw:engines, gh-aw:cost.
Child findings [unknown] Retained artifacts and outputs, child_findings: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Prompt leakage [unknown] Retained artifacts and outputs, prompt_leakage: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Workflow configuration separates deterministic CI jobs from agent reasoning and permitted outputs.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
11 atomic answers
Tests [unknown] Verification and spec gates, tests: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Lint [unknown] Verification and spec gates, lint: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Types [unknown] Verification and spec gates, types: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Ci [unknown] Verification and spec gates, ci: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Independent review [partial] Threat detection reviews output and agent review workflows can be authored. Sources: gh-aw:d80-threat.
Spec checks [unknown] Verification and spec gates, spec_checks: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Required checks [partial] Checks are repository settings; gh-aw triggers CI through a documented route. Sources: gh-aw:d80-safeout.
Failure blocking [supported] Threat detection runs before safe outputs apply and can block them. Sources: gh-aw:d80-threat.
Requirement conflicts [unknown] Verification and spec gates, requirement_conflicts: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Configured policy [supported] Workflow configuration separates deterministic CI jobs from agent reasoning and permitted outputs. Sources: gh-aw:readme.
Enforced policy [supported] Compiler and safe-output jobs enforce permissions and limits. Sources: gh-aw:d80-perms.
create-pull-request safe outputs propose code in a separate permission-scoped job. Workflow examples include CI-failure investigation; automatically triggering CI from generated PRs may require an additional token.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
9 atomic answers
Branches [supported] create-pull-request creates a branch and PR. Sources: gh-aw:d80-safeoutpr.
Worktrees [unknown] Git and pull request workflow, worktrees: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Pull requests [supported] create-pull-request safe outputs propose code in a separate permission-scoped job. Sources: gh-aw:pr, gh-aw:faq.
Comments [supported] Review comments and replies are safe outputs. Sources: gh-aw:d80-safeoutpr.
Ci response [partial] Workflow examples include CI-failure investigation; automatically triggering CI from generated PRs may require an additional token. Sources: gh-aw:faq.
Rebase [unknown] Git and pull request workflow, rebase: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Conflicts [partial] The docs describe conflicts only where a safe output reports them. Sources: gh-aw:d80-safeoutpr.
Product state [unknown] Git and pull request workflow, product_state: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Merge operations are separately configured safe outputs with scoped credentials; PR creation alone does not grant merging.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
6 atomic answers
Approve [supported] submit-pull-request-review and approve-workflow-run submit approvals. Sources: gh-aw:d80-safeoutpr.
Merge [partial] Merge operations are separately configured safe outputs with scoped credentials; PR creation alone does not grant merging. Sources: gh-aw:pr.
Deploy [unsupported] No deploy safe output exists in the list; deployment is a user workflow concern. Sources: gh-aw:d80-safeout.
Exact revision [unknown] Merge and deployment authority, exact_revision: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Authority enforcement [supported] Write permissions exist only in safe-output jobs with limits. Sources: gh-aw:d80-perms.
Safe outputs link work to issues, PRs, commits and check runs, and outcomes record the observed repository state; there is no deployment record or rollback feature.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: gh-aw v1.2.2 (399a445, latest stable release; v1.3.0 is a pre-release) compiler, CLI and documentation; source inspection 2026-11-16. Execution depends on GitHub Actions and the chosen engine. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Deployment links [unknown] Completion evidence and rollback, deployment_links: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Delivery status [supported] Outcomes record merged, closed or ignored state. Sources: gh-aw:d80-outcomes.
Rollback [unknown] Completion evidence and rollback, rollback: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Revert [unknown] Completion evidence and rollback, revert: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Stale approval invalidation [unknown] Completion evidence and rollback, stale_approval_invalidation: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
The gh-aw compiler and runtime repository use MIT.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
6 atomic answers
Client license [supported] The gh-aw compiler and runtime repository use MIT. Sources: gh-aw:license.
Server license [not-applicable] There is no gh-aw server component; the runtime is GitHub Actions. Sources: gh-aw:d80-ent.
Hosted extensions [partial] GitHub Actions and Copilot services are GitHub-operated and closed. Sources: gh-aw:d80-ent.
Source visibility [supported] Source is public on GitHub. Sources: gh-aw:d80-license.
Source revision [supported] Inspected at v1.2.2, commit 399a445. Sources: gh-aw:d80-release.
Hosted parity [unknown] Source and license, hosted_parity: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
The product code and execution worker can run on user-controlled runners, but GitHub Actions supplies workflow control, triggers, secrets and logs. A self-managed execution worker alone does not establish a fully self-hosted core ADE; a complete user-managed control-plane deployment is unestablished by this source pass.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
8 atomic answers
Self hosting [partial] Self-managed runner execution is established. A full user-managed workflow/control-plane deployment is not established; runner support alone does not resolve core ADE self-hosting. Sources: gh-aw:gap37-client-boundary.
Vendor dependencies [supported] GitHub Actions and the selected inference provider remain dependencies. Sources: gh-aw:faq.
Export [unknown] Hosting and exit, export: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Import [supported] gh aw add imports workflows from repositories. Sources: gh-aw:d80-cli.
Formats [supported] Workflows are Markdown with YAML frontmatter and compiled lock files. Sources: gh-aw:d80-quick.
Cancellation [unknown] Hosting and exit, cancellation: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Account independence [unsupported] A GitHub account and Actions are required. Sources: gh-aw:d80-quick.
History portability [partial] Everything is in Git and Actions artifacts. Sources: gh-aw:d80-repomem.
gh-aw workflow OTLP telemetry export is configurable and can be disabled by removing workflow/import endpoints and clearing default endpoint secrets/variables. GitHub platform and coding-engine data policies remain separate.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
unknown
Delivery
unknown
Checked
2026-11-13 (9 days old)
Agent cross-check
2026-11-17
Source review and pinned implementation inspection only; no installed-product, account or multi-user scenario was executed. Unanswered atomic questions remain open.
15 atomic answers
Code data [partial] The engine, inference provider and tools determine where repository context is processed. Sources: gh-aw:d80-engines.
Prompts [unknown] Data handling and telemetry, prompts: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Credentials [supported] Credentials are Actions secrets and threat detection scans for leaks. Sources: gh-aw:d80-threat.
Logs [unknown] Data handling and telemetry, logs: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Artifacts [unknown] Data handling and telemetry, artifacts: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Retention [unknown] Data handling and telemetry, retention: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Deletion [unknown] Data handling and telemetry, deletion: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Training [unknown] Data handling and telemetry, training: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Residency [partial] GHEC with data residency endpoints are compatibility-tested only. Sources: gh-aw:d80-ent.
Subprocessors [unknown] Data handling and telemetry, subprocessors: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Telemetry disclosure [supported] OpenTelemetry export is user-configured and documented. Sources: gh-aw:d80-otel.
Telemetry defaults [unknown] Data handling and telemetry, telemetry_defaults: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Telemetry controls [supported] Clear both default endpoint values and remove workflow/import observability.otlp; export becomes a no-op. Sources: gh-aw:gap37-telemetry-control.
Consent prompt [unknown] Data handling and telemetry, consent_prompt: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Provider separation [supported] The control disables gh-aw OTLP export, not GitHub operational logs or coding-agent telemetry. Sources: gh-aw:gap37-telemetry-control.
The gh-aw extension/compiler is free MIT software without a paid extension plan. Inference is billed by the chosen provider/Copilot plan and Actions compute minutes remain separate.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: 399a445fe2d8926c616b54b1b82a16b5b2be4445 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
No provider or Actions compute price is inferred from the software license.
12 atomic answers
Currency [unknown] Price and included usage, currency: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Interval [unknown] Price and included usage, interval: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Seats [unknown] Price and included usage, seats: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Seat minimum [unknown] Price and included usage, seat_minimum: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Workspace fees [unsupported] No fee for the scoped MIT extension/compiler software; platform runner and inference charges are separate. Sources: gh-aw:gap37-extension-license.
Included usage [unknown] Price and included usage, included_usage: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Credits [unknown] Price and included usage, credits: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Trial [unknown] Price and included usage, trial: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Overages [unknown] Price and included usage, overages: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Taxes [unknown] Price and included usage, taxes: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Model cost [supported] Inference is charged by the selected provider or Copilot plan, separate from Actions minutes. Sources: gh-aw:faq.
Compute cost [supported] GitHub Actions compute minutes are separate from model inference. Sources: gh-aw:cost.
Release v1.2.2 (commit 399a445) was the latest stable release at the 2026-11-16 source pass, with v1.3.0 a pre-release; v1.2.1 dated 2026-11-06 had added experimental artifact-backed continuation and workflow fixes.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
Not separately cross-checked
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
8 atomic answers
Release stage [unknown] Availability and support, release_stage: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Access restrictions [unknown] Availability and support, access_restrictions: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Support [unknown] Availability and support, support: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Response commitments [unknown] Availability and support, response_commitments: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Export options [unknown] Availability and support, export_options: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
gh-aw itself is free; model and runner charges remain separate.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-17
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
10 atomic answers
Ongoing free [supported] gh-aw itself is free; model and runner charges remain separate. Sources: gh-aw:faq.
Trial [unknown] Useful free tier and evaluation access, trial: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Credits trial [unknown] Useful free tier and evaluation access, credits_trial: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Self host cost [unknown] Useful free tier and evaluation access, self_host_cost: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Paid only [unknown] Useful free tier and evaluation access, paid_only: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Early access [partial] Several features are labelled experimental; pre-releases are public. Sources: gh-aw:d80-staged.
Free scenarios [unknown] Useful free tier and evaluation access, free_scenarios: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Card required [not-applicable] No payment handling exists in gh-aw; Actions and engine billing are GitHub and provider concerns. Sources: gh-aw:d80-billing.
Usage caps [unknown] Useful free tier and evaluation access, usage_caps: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Paid evaluation [unknown] Useful free tier and evaluation access, paid_evaluation: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
The application adds no platform charge; inference and Actions minutes remain separate billed units.
Scope
Surfaces: gh aw CLI and compiled GitHub Actions workflows · Version: Repository 57111b2d03d5be9a30a998f518c9712700ebfa63; first-party source pass 2026-11-08. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-08 (14 days old)
Agent cross-check
2026-11-08
Only the cited atomic findings are established. Remaining atomic requirements and the full rubric scenario remain unverified.
9 atomic answers
Units [partial] The application adds no platform charge; inference and Actions minutes remain separate billed units. Sources: gh-aw:faq.
Included limits [unknown] Pricing transparency, included_limits: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Overages [unknown] Pricing transparency, overages: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Compute markup [unknown] Pricing transparency, compute_markup: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Model markup [unknown] Pricing transparency, model_markup: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Minimum commitment [unknown] Pricing transparency, minimum_commitment: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Cancellation [unknown] Pricing transparency, cancellation: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Sales only unknowns [unknown] Pricing transparency, sales_only_unknowns: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
Transparency classification [unknown] Pricing transparency, transparency_classification: GitHub Agentic Workflows's checked sources did not establish this requirement or its rubric scenario. No installed-product fixture was run.
The gh aw extension compiles Markdown/YAML repository automation into GitHub Actions workflows with configurable agents, read-only sandbox defaults and separate safe-output jobs.
The install script provides native extension binaries for Linux, macOS, Windows and other hosts. gh aw is a CLI/compiler plus GitHub Actions runtime, not a standalone browser/phone app or diff review interface. It executes agent work on configured Actions runners; a self-hosted runner can be on a user computer. Source checkout is managed by the workflow, with no worktree provisioning. GitHub supplies human accounts, repository permissions, SSO, workflow control and logs; no separate human-team or shared live-prompt service is implemented by the extension.
Kind
source-inspected
Locator
OS case 115-139; cmd/gh-aw and pkg/cli command registration; compiler/runtime; docs/reference/faq
Built-in engines include Copilot, Claude Code, Codex, deprecated Gemini CLI and Pi. Experimental durable mode checkpoints native sessions in artifacts; max-turns and AI-credit limits vary by engine.
The repository ships publicly importable experimental engine definitions that install/launch Cursor Agent, OpenCode 2 and Aider. They are explicitly unsupported examples for integration authors without a compatibility or maintenance commitment, rather than built-in supported engines. Custom engine definitions are accepted by the compiler.
Kind
source-inspected
Locator
Shared cursor.md, opencode.md and aider.md; docs/reference/third-party-agent.md Unsupported engine samples
Built-in engines are Copilot (default), Claude Code, Codex, Gemini CLI (deprecated), Pi and experimental Agy; max-turns defaults to 500 and max-ai-credits to 1000; imported custom engines are possible.
Kind
documented
Locator
docs/src/content/docs/reference/engines.md l12-l24; built-in engines and defaults; v1.2.2 (399a445)
Claude Code uses an Anthropic API key, Anthropic workload identity federation or Copilot inference; a Claude subscription OAuth login is not supported.
Kind
documented
Locator
docs/src/content/docs/engines/claude.md l6-l9; Connect your account; v1.2.2 (399a445)
Claude supports API/WIF or Copilot inference, not Claude subscription OAuth. dontAsk is default, tools.edit false denies edits, and MCP gateway permissions remain separate from agent approval modes.
Codex uses OpenAI API credentials or Copilot inference; ChatGPT subscription login is unsupported. Per-command Bash allowlists are unavailable for Codex.
GHES and GHEC data residency compatibility is implemented in the compiler but end-to-end availability is stated as unconfirmed; GitHub-hosted runners are not available for GHES.
Kind
documented
Locator
docs/src/content/docs/reference/enterprise-configuration.md l15-l27; deployment and runner applicability; v1.2.2 (399a445)
Self-hosted runners must be Linux with Docker; macOS and Windows are not supported; a self-hosted runner processes hostile content and the sandbox does not make the host clean.
Per-run logs/audit reports display tokens, duration and AI-credit estimates. One AIC represents $0.01 of catalog-estimated inference, distinct from Actions compute and provider invoices.
The agent runs in the AWF firewall sandbox by default and the MCP gateway routes tool calls; disabling the sandbox needs strict false and an explicit dangerous flag.
tools.work-queue is a native Git-backed queue whose transaction log is authoritative; it stores Work task definitions with dependencies, Claims for each attempt, Completions and Results. GitHub Issues can mirror tasks but cannot store the queue.
Kind
documented
Locator
docs/src/content/docs/reference/work-queue.md lines 6-31 (intro and record table), tag v1.2.2 commit 399a445
gh aw work-queue offers operator commands to submit Work, list state filtered by graph, pool, state or search text, inspect items, dispatch assignments, cancel Work or Claims and, for administrators, reprioritize Work; a keyboard TUI browses the queue.
Kind
documented
Locator
docs/src/content/docs/reference/work-queue.md lines 217-252 (Operator commands table), tag v1.2.2 commit 399a445
Separate workflows can run concurrently, with default one agent job per engine. Context-specific groups, cancellation and queue:max preserve or cancel pending runs according to trigger configuration.
The agent step has a default 20 minute timeout; other jobs default to the Actions 360 minutes; concurrency, safe output limits and delays bound runaway runs.
Kind
documented
Locator
docs/src/content/docs/reference/rate-limiting-controls.md l38-l42; timeouts and concurrency; v1.2.2 (399a445)
The default docker runtime runs the agent in a rootless AWF container with network isolation; preview cloud-hypervisor and nvx runtimes use KVM microVMs. Filesystem mounts and outbound network allowlists are configurable. The page does not itself describe the GitHub runner as a VM.
gh-aw is free with separate inference and Actions compute charges. It requires Linux container-capable runners, can use self-hosted runners, and retains GitHub workflow dependencies.
The default agent runtime uses Docker container isolation on GitHub-hosted or Linux self-hosted runners, with the Squid proxy as the only egress path. Eligible GitHub-hosted Ubuntu x86_64 KVM runners can instead place the agent in a Cloud Hypervisor microVM (preview).
Removing workflow/import observability.otlp and clearing both the secret and variable GH_AW_DEFAULT_OTLP_ENDPOINT disables gh-aw OTLP export. With both endpoint values unset export resolves to an empty endpoint and is a no-op. GitHub platform logs and coding-engine telemetry are separate.
Kind
documented
Locator
Organization-wide defaults at 101-118; workflow observability.otlp
The main extension/compiler software is MIT licensed and explicitly permits use free of charge; Actions runner compute and inference plans are separate dependencies.