50 of 51 criteria have a sourced finding; 50 were also cross-checked by a second agent. Every comparison that includes claude-code-action 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.
The product is a GitHub composite Action. Humans invoke it through GitHub issues and PRs; its code supplies no native desktop or phone client or separately distributed user-facing CLI.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Source pinned at 4fa93a88de49bbef68b10eb15aabedf28ca0e1ed; checked 2026-11-13. Default mentions and configured automation are distinguished; retained older citations identify their own revisions. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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 named documented or inspected parts are established; other atomic questions retain their recorded gaps.
9 atomic answers
Desktop os [unsupported] The pinned repository distributes a composite action, not a native macOS, Windows or Linux desktop client. Sources: claude-code-action:gap-action-client.
Mobile apps [unsupported] No native iOS or Android application is distributed in this action repository. Sources: claude-code-action:gap-action-client.
Browser [unsupported] No product-owned browser interface exists; users interact through GitHub's own issue and pull request pages by commenting, and the action's single tracking comment appears there. Sources: claude-code-action:v276-n-ui, claude-code-action:v276-dcap.
Cli [unsupported] The action orchestrator is a runner entrypoint, not a separate product-owned user CLI. The Claude CLI it installs is the underlying harness. Sources: claude-code-action:gap-action-client.
Editor extensions [unsupported] No editor extension ships in the repository; the quickstart relies on the separate Claude Code CLI's /install-github-app command. Sources: claude-code-action:v276-n-ui.
Task actions [supported] Humans invoke the action from GitHub issue/PR comments; its tracking comment displays progress and results. Sources: claude-code-action:readme.
Notifications [partial] The action posts and edits one tracking comment, so notifications come from GitHub for that comment; the action has no notification channel of its own. Sources: claude-code-action:v276-comment, claude-code-action:v276-dcap.
Cross device continuity [partial] Requests and results live in the GitHub issue or PR thread and can be followed from any device with GitHub access; a running Claude session cannot be re-attached, though session_id is emitted as an output. Sources: claude-code-action:v276-dcap, claude-code-action:v276-yml.
The action runs the Claude Code harness through its SDK run path. Provider configuration changes model authentication, and the custom executable input expects Claude Code; the pinned code does not implement a native route for other coding harnesses.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Source pinned at 4fa93a88de49bbef68b10eb15aabedf28ca0e1ed; checked 2026-11-13. Default mentions and configured automation are distinguished; retained older citations identify their own revisions. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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 named documented or inspected parts are established; other atomic questions retain their recorded gaps.
5 atomic answers
Harnesses [supported] The action runs Claude Code, with model/provider configuration separate. Sources: claude-code-action:readme.
Model providers [supported] Anthropic API, Bedrock, Vertex AI and Microsoft Foundry authentication are documented. Sources: claude-code-action:cloud.
Custom adapters [unsupported] The harness is fixed: the action installs the Claude Code CLI and path_to_claude_code_executable must point to Claude Code; there is no adapter interface for other harnesses. Sources: claude-code-action:v276-install, claude-code-action:v276-yml.
Session integration [supported] Workflow inputs configure the Claude Code execution on a GitHub runner. Sources: claude-code-action:usage.
Mixed harness operation [unsupported] The action and base-action run Claude Code only; there is no alternative coding-harness dispatcher in the pinned implementation. Sources: claude-code-action:gap-harness.
Claude Pro/Max users can generate a Claude Code OAuth token using claude setup-token for the action. Direct API key, workload identity federation and supported cloud-provider authentication are documented.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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 [supported] Claude Pro/Max users can generate a Claude Code OAuth token using claude setup-token for the action. Sources: claude-code-action:setup.
Vendor credits [unsupported] The action has no vendor-credit mechanism; it needs a provider credential (API key, OAuth token, federation or cloud account) and usage is billed by that provider. Sources: claude-code-action:v276-venv, claude-code-action:v276-provider.
Local models [unsupported] No local-model route exists; a credential for Anthropic, Bedrock, Vertex or Foundry is required, and ANTHROPIC_BASE_URL is only forwarded for gateway use. Sources: claude-code-action:v276-venv, claude-code-action:v276-provider.
Account switching [unsupported] One credential set is used per run and more than one cloud provider is rejected; selecting different accounts means different workflow secrets. Sources: claude-code-action:v276-venv.
Vendor account requirement [supported] A provider account or token is required; validation fails without an API key, OAuth token, workload identity or a complete Bedrock, Vertex or Foundry configuration. Sources: claude-code-action:v276-venv.
Provider restrictions [unknown] Provider terms on using subscription OAuth tokens in automation are vendor policy outside this repository; the checked docs, README and validation code state none. Reason: external vendor terms.
A repository admin installs the GitHub App, stores credentials and configures the workflow. The documented workflow operates on GitHub issues, PRs and a runner checkout. Setup documents repository-admin and secret permissions.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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] A repository admin installs the GitHub App, stores credentials and configures the workflow. Sources: claude-code-action:setup.
No git [unsupported] The orchestrator expects a Git checkout at the workspace root and runs git branch, fetch and push operations; a plain directory is not handled. Sources: claude-code-action:v276-branch, claude-code-action:v276-dsec.
Local git without remote [unsupported] Runs require a GitHub event payload, repository identity and GitHub API calls, so a local repository without a GitHub remote is not supported. Sources: claude-code-action:v276-run, claude-code-action:v276-ctx.
Hosted git [supported] The documented workflow operates on GitHub issues, PRs and a runner checkout. Sources: claude-code-action:readme.
Managed cloud setup [unsupported] No vendor-managed execution environment exists to set up; the only hosted piece is the GitHub App token exchange endpoint, and the job runs on the user's runner. Sources: claude-code-action:v276-steps, claude-code-action:v276-token.
Byo vps onboarding [unsupported] Self-hosted runners work as a GitHub feature, but the repository has no installer or onboarding flow for a user-owned server beyond custom executable paths. Sources: claude-code-action:v276-steps, claude-code-action:v276-dfaq.
The workflow accepts prompt/claude_args and emits structured JSON action outputs. GitHub App/token authorization and provider API or OAuth credentials are separate.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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 [unsupported] No product-owned CLI exists; the orchestrator is an internal entrypoint run by the composite action, and the CLI that installs the app is the separate Claude Code CLI. Sources: claude-code-action:v276-n-ui, claude-code-action:v276-run.
Api [unsupported] The action exposes no HTTP API; its only input surfaces are workflow inputs and GitHub events, and its outputs are workflow step outputs. Sources: claude-code-action:v276-n-ui, claude-code-action:v276-yml.
Mcp server [unsupported] The action launches GitHub MCP servers only for the Claude session it starts; it does not expose the product to other agents over MCP. Sources: claude-code-action:v276-tag, claude-code-action:v276-mcp-ci.
Webhooks [partial] GitHub webhook events, including repository_dispatch, start runs through workflow triggers; the action sends no outbound webhooks of its own. Sources: claude-code-action:v276-ctx, claude-code-action:v276-dauto.
Operations [supported] The workflow accepts prompt/claude_args and emits structured JSON action outputs. Sources: claude-code-action:usage.
Authentication [supported] GitHub App/token authorization and provider API or OAuth credentials are separate. Sources: claude-code-action:setup.
Custom --mcp-config servers merge with built-in GitHub MCP servers; allowedTools gates enabled tools. Reading CI workflow/job logs requires actions:read and additional_permissions. The action wires GitHub tools into Claude Code and passes custom MCP configuration through Claude CLI flags.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Custom tools [supported] Custom MCP servers and tools are added with claude_args (--mcp-config, --allowedTools) and merged with the built-in GitHub servers. Sources: claude-code-action:v276-dcfg.
Instructions import [supported] Claude is told to follow the repository's CLAUDE.md, and .claude settings, commands and CLAUDE.md are restored from the PR base branch so repository instructions are loaded. Sources: claude-code-action:v276-prompt, claude-code-action:v276-restore.
Authentication [supported] Docs show passing MCP server secrets through workflow env and settings; provider and GitHub tokens are separate inputs. Sources: claude-code-action:v276-dcfg.
Permission scope [supported] Reading CI workflow/job logs requires actions:read and additional_permissions. Sources: claude-code-action:config.
Integration owner [supported] The action wires GitHub tools into Claude Code and passes custom MCP configuration through Claude CLI flags. Sources: claude-code-action:config.
The product is a composite GitHub Action with a TypeScript orchestrator run by Bun; it ships no desktop or mobile client to classify.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
No desktop, mobile or editor client exists, so client-implementation questions are answered as absent; only the runner-side implementation is described.
Mobile implementation [unsupported] There is no mobile client; no React Native, Expo or native mobile code exists in the repository. Sources: claude-code-action:v276-n-ui.
Issue and PR comments or an explicit workflow prompt supply task input. Mentions, issue events and configured PR events initiate action work. GitHub issues and PRs remain the external records; the action updates a tracking comment.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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 [unsupported] The action cannot create tasks of its own; a task is a GitHub comment, issue event or workflow prompt created outside it. Sources: claude-code-action:v276-n-store.
External import [supported] Issue or PR body, comments, reviews and changed files are fetched through the GitHub API into the run's prompt. Sources: claude-code-action:v276-fetch.
Authoritative tracker [partial] GitHub issues and PRs remain the external records; the action updates a tracking comment. Sources: claude-code-action:readme.
Plans are a flat checklist in one GitHub tracking comment that Claude maintains; there are no acceptance criteria, revisions or separate plan objects.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
The checklist lives in a comment overwritten by the action; no acceptance criteria, revision store or editor exists.
7 atomic answers
Plans [partial] Claude is instructed to keep a task checklist inside the tracking comment; it is a comment, not a separate plan object. Sources: claude-code-action:v276-prompt.
Acceptance criteria [unsupported] No acceptance-criteria field or check exists in the action. Sources: claude-code-action:v276-n-store.
Editors [partial] Humans can edit the GitHub comment, but the action overwrites the tracking comment on later updates; there is no plan editor. Sources: claude-code-action:v276-comment.
Revisions [unsupported] The action keeps no plan revision history; any edit history is GitHub's own comment history. Sources: claude-code-action:v276-n-store.
Amendment [partial] The prompt tells Claude to add or remove checklist items as requirements change during its run; a running job takes no new human input, so humans amend by commenting again. Sources: claude-code-action:v276-prompt.
Restart persistence [partial] The checklist stays in the GitHub comment after the job ends, and a new @claude comment starts a fresh run that re-reads thread history; session state is cleared and not resumed. Sources: claude-code-action:v276-comment, claude-code-action:v276-guard.
The action has no epics, hierarchy or dependencies; it handles one GitHub issue or PR per run with a flat checklist.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Only a flat checklist and unrestricted workflow choice exist; hierarchy, blockers and epics are absent.
Task to epic [unsupported] A task cannot be promoted to an epic; the only growth is additional checklist items. Sources: claude-code-action:v276-n-store.
Workflow constraints [supported] No rigid preselected workflow is required: mode is auto-detected from event and prompt, and a free-form prompt or mention is enough. Sources: claude-code-action:v276-dauto, claude-code-action:v276-dusage.
Repository instructions, skills and plugins are loaded through Claude Code from the checkout and base-branch config; the action keeps no memory store.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
No product-managed memory, provenance, correction or deletion exists; the harness features belong to the closed Claude CLI.
Skills [supported] Claude Code skills and commands in .claude are loaded from the base branch and plugins can be installed through the plugins input; the skill mechanism belongs to the Claude Code CLI. Sources: claude-code-action:v276-restore, claude-code-action:v276-plugins.
Freshness [partial] Comments edited after the trigger time are excluded to avoid stale or injected context; there is no other stale-context control. Sources: claude-code-action:v276-fetch.
Takeover continuity [partial] A later job by another collaborator sees the whole issue or PR thread and the branch; Claude's own transcript is not carried over. Sources: claude-code-action:v276-fetch, claude-code-action:v276-dcap.
Claude gets Glob, Grep, LS and Read and multimodal image input, plus GitHub CI retrieval; the action has no index, citation or conflict detection.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
No indexing, citation, conflict detection or review enforcement exists.
10 atomic answers
Code search [partial] Claude gets Glob, Grep, LS and Read over the checkout; the action provides no search service or index of its own. Sources: claude-code-action:v276-tag.
Indexing [unsupported] No repository indexing, embeddings or vector store exists. Sources: claude-code-action:v276-n-mem.
External knowledge [partial] External knowledge comes from Claude Code tools and user-configured MCP servers; the action itself ships no knowledge connectors. Sources: claude-code-action:v276-dcfg.
Multimodal inputs [supported] Screenshots attached to issues or comments are downloaded and shown to Claude, and the usage guide documents fixing a bug from a screenshot. Sources: claude-code-action:v276-images, claude-code-action:v276-dusage.
Freshness [partial] Context is the checkout at trigger time with a shallow fetch depth chosen from PR size; nothing refreshes it during the run. Sources: claude-code-action:v276-branch.
Access scope [supported] The token is scoped to the triggering repository and docs state there is no cross-repository access. Sources: claude-code-action:v276-dsec.
The action consumes GitHub issues and pull requests; it does not own a persistent issue/task store in the pinned src and base-action implementation.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Source pinned at 4fa93a88de49bbef68b10eb15aabedf28ca0e1ed; checked 2026-11-13. Default mentions and configured automation are distinguished; retained older citations identify their own revisions. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Only the named documented or inspected parts are established; other atomic questions retain their recorded gaps.
17 atomic answers
Create [unsupported] Action-owned issue creation/storage is absent; issue data is fetched through the GitHub API. Sources: claude-code-action:gap-session-boundary.
Edit [unsupported] The action owns no issues to edit; GitHub issues are edited in GitHub. Sources: claude-code-action:v276-n-store.
Discuss [unsupported] Discussion happens in GitHub comments, not in an action-owned issue object. Sources: claude-code-action:v276-n-store.
Label [unsupported] The action has no labeling feature of its own; label_trigger only reads a GitHub label as a trigger. Sources: claude-code-action:v276-n-store.
Assign [unsupported] The action does not assign issues; assignee_trigger only reads a GitHub assignment as a trigger. Sources: claude-code-action:v276-n-store.
Explicit dispatch [supported] A GitHub issue can be dispatched later by applying label_trigger or assigning assignee_trigger, or by an @claude mention; the issue itself remains a GitHub object. Sources: claude-code-action:v276-trig, claude-code-action:v276-yml.
Task session separation [unsupported] There is no separate task entity; each run is bound to one GitHub issue or PR event. Sources: claude-code-action:v276-n-store.
Metadata [unsupported] The action stores no task metadata beyond branch name and tracking comment. Sources: claude-code-action:v276-n-store.
Durable states [unsupported] Durable issue states belong to GitHub, rather than a product-owned action task store. Sources: claude-code-action:gap-session-boundary.
Agents operate on GitHub-owned issue and PR numbers through comments, commits and allowed tools; the action owns no issue store.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Agent operations are limited to GitHub-owned records; claim, assign, dependency and child operations are absent.
Human visibility [supported] Agent writes appear as GitHub comments, commits and branches visible to everyone with repository access. Sources: claude-code-action:v276-comment.
Attribution [supported] Commits default to the claude[bot] identity with configurable bot_id and bot_name, and the prompt asks for a Co-authored-by trailer for the triggering user. Sources: claude-code-action:v276-coauthor, claude-code-action:v276-yml.
The action keeps no product specifications; CLAUDE.md is only an instruction file.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
unknown
Delivery
unknown
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
9 atomic answers
Persistent requirements [unsupported] No persistent product requirements store exists; CLAUDE.md is an instruction file, not a spec. Sources: claude-code-action:v276-n-store.
Independent GitHub workflow jobs can execute this action concurrently. The action prepares a Claude run and stores execution output in each job's RUNNER_TEMP; GitHub runner and workflow concurrency limits remain external.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Source pinned at 4fa93a88de49bbef68b10eb15aabedf28ca0e1ed; checked 2026-11-13. Default mentions and configured automation are distinguished; retained older citations identify their own revisions. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Only the named documented or inspected parts are established; other atomic questions retain their recorded gaps.
5 atomic answers
Concurrent sessions [supported] The per-job orchestrator has no action-wide lock or shared session store and uses job-scoped execution files. Sources: claude-code-action:gap-session-boundary.
Concurrent tasks [supported] Each workflow job runs its own orchestrator with job-scoped files and no shared lock, so jobs triggered on different issues and PRs can run side by side. Sources: claude-code-action:v276-run, claude-code-action:v276-guard.
Quotas [unsupported] The action defines no concurrency quota; runner capacity and GitHub workflow concurrency groups set the limits. Sources: claude-code-action:v276-n-queue.
Dispatch is by GitHub events and an assignee or label trigger; there is no claim, queue or ownership enforcement.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
No claims, queues, routing between agents, duplicate prevention or enforced ownership exist.
7 atomic answers
Human assignment [partial] Assigning a GitHub issue to the configured assignee triggers work; this is GitHub assignment used as a trigger, not product-managed ownership. Sources: claude-code-action:v276-trig, claude-code-action:v276-yml.
Agent assignment [unsupported] No agent-to-agent or agent-to-human assignment exists. Sources: claude-code-action:v276-n-store.
Duplicate prevention [unsupported] The action does not prevent two runs on the same issue or PR; users can add GitHub concurrency groups. Sources: claude-code-action:v276-n-queue.
Ownership enforcement [unsupported] No ownership is enforced; the push wrapper only limits which branch a session may update. Sources: claude-code-action:v276-n-store.
The action runs a single Claude session per job; it supplies no coordinator, children or messaging.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Source pinned at 4fa93a88de49bbef68b10eb15aabedf28ca0e1ed; checked 2026-11-13. Default mentions and configured automation are distinguished; retained older citations identify their own revisions. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Only the named documented or inspected parts are established; other atomic questions retain their recorded gaps.
Live progress [unsupported] No child progress view exists; the tracking comment shows a single agent's checklist. Sources: claude-code-action:v276-n-store.
README links scheduled maintenance examples using GitHub Actions. Configured GitHub events run explicit-prompt automation without a mention. The schedule and event workflow runs on GitHub Actions, outside the initiating browser.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Events [supported] Configured GitHub events run explicit-prompt automation without a mention. Sources: claude-code-action:automation.
Recurring tasks [supported] A schedule cron workflow with a prompt runs the action without a human trigger. Sources: claude-code-action:v276-dsol.
Completion conditions [partial] A run ends on Claude finishing, max-turns or job timeout, and can emit structured_output; the action has no goal-completion loop. Sources: claude-code-action:v276-sdk, claude-code-action:v276-dcfg.
Overlaps [unsupported] The action has no overlap handling; overlapping scheduled runs are governed by GitHub concurrency settings the user chooses. Sources: claude-code-action:v276-n-queue.
Retries [unsupported] A failed run is not retried; only token exchange and install steps retry. Sources: claude-code-action:v276-n-queue.
The FAQ describes using GitHub Actions logs and optional debug output for diagnostics.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Steer [unsupported] A running job accepts no further instructions; a new comment starts a new run. Sources: claude-code-action:v276-dcap.
Notifications [partial] GitHub notifies on comment updates; the action sends no other notification. Sources: claude-code-action:v276-comment.
Pending decisions [unsupported] There is no approval prompt in a workflow; commands that no rule covers are denied in non-auto modes. Sources: claude-code-action:v276-dsec.
Escalation [unsupported] There is no escalation channel; Claude is told to explain limits and link the FAQ in its comment. Sources: claude-code-action:v276-prompt-limits.
Visible errors [supported] The FAQ describes using GitHub Actions logs and optional debug output for diagnostics. Sources: claude-code-action:faq.
Chat terminal parity [unsupported] There is no separate terminal surface; the single comment is the only interface. Sources: claude-code-action:v276-n-ui.
The action ships no fleet or epic interface; its only surfaces are the GitHub tracking comment and workflow logs.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
unknown
Delivery
unknown
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
8 atomic answers
Fleet fixture [unsupported] No fleet or epic interface exists; GitHub's workflow run list is not part of the product. Sources: claude-code-action:v276-n-ui.
A mention or prompt starts one run; configuration is a free-form prompt and claude_args; no presets, growth or in-run changes exist.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
No presets, epics, in-run changes or orchestration visibility exist.
Natural language [supported] Behavior is configured in a free-form prompt and claude_args. Sources: claude-code-action:v276-dauto.
Agent configuration [partial] Model, tools, turns and settings are set per workflow in claude_args and settings; there is no multi-agent configuration. Sources: claude-code-action:v276-dcfg.
Retained relationships [unsupported] No relationships between runs are retained beyond the issue or PR thread. Sources: claude-code-action:v276-n-store.
Auth friction [partial] Setup needs repository-admin app install and provider secrets or federation before the first run. Sources: claude-code-action:v276-dsetup.
The action runs on the selected user-controlled GitHub runner. GitHub event/schedule workflows execute on their runner without the initiating browser remaining open.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Byo remote [supported] The action runs on the selected user-controlled GitHub runner. Sources: claude-code-action:readme.
Self hosted [supported] Runs on whichever runner the job selects, including a self-hosted runner; the action adapts install steps when sudo or apt are missing. Sources: claude-code-action:v276-steps.
Managed cloud [unsupported] The action provisions no vendor-run compute; GitHub-hosted runners are GitHub's, and the only Anthropic-hosted piece is the token exchange. Sources: claude-code-action:v276-steps, claude-code-action:v276-token.
Provisioning [unsupported] The action provisions nothing; the workflow chooses runs-on. Sources: claude-code-action:v276-steps.
Payer [unknown] Runner minutes are billed by GitHub and model usage by the chosen provider; billing is a hosted-service business fact not stated in the repository.
Resources [unknown] CPU and memory depend on the runner chosen in the workflow, a GitHub-hosted service setting not defined by the action.
Region [unknown] Region depends on the GitHub runner and model provider region (AWS_REGION, CLOUD_ML_REGION are forwarded); the action defines none.
Other device continuation [partial] Results can be followed and new runs triggered from any device through GitHub, but a running session cannot be attached to. Sources: claude-code-action:v276-dcap, claude-code-action:v276-yml.
The action expects a Git repository at the runner workspace root.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Checkout [supported] The action expects a Git repository at the runner workspace root. Sources: claude-code-action:security.
Branch [supported] Issue and closed-PR runs work on a new branch, open-PR runs on the PR branch, and the push wrapper may update only that branch. Sources: claude-code-action:v276-branch.
Worktree [unsupported] No git worktree is created; the run uses the single checkout. Sources: claude-code-action:v276-n-iso.
Container [unsupported] The action configures no container; the job runs on the runner directly. Sources: claude-code-action:v276-n-iso.
Vm [unsupported] The action provisions no VM; any VM isolation is GitHub's runner and not an option the action offers. Sources: claude-code-action:v276-n-iso.
Os sandbox [partial] Only with allowed_non_write_users on Linux does it install bubblewrap and run subprocesses with PID-namespace isolation, as a best-effort feature of the Claude CLI. Sources: claude-code-action:v276-steps, claude-code-action:v276-dsec.
Filesystem [partial] Config paths are restored from the base branch and session files are snapshotted, but Claude can still read and write the workspace and docs warn broad Write rules reach outside it. Sources: claude-code-action:v276-restore, claude-code-action:v276-guard.
Secrets [partial] With allowed_non_write_users, Anthropic, cloud and Actions secrets are scrubbed best-effort from subprocess environments; the app token stays in Claude's environment. Sources: claude-code-action:v276-dsec, claude-code-action:v276-otel.
Code separation [partial] On PRs by authors without write access, base-branch config is restored and code-running commands are denied, but untrusted PR code is checked out in the same workspace. Sources: claude-code-action:v276-untr, claude-code-action:v276-restore.
Security boundary [unsupported] Docs state the untrusted-PR command denial and session-file restore are not a sandbox; no hostile-code boundary is provided. Sources: claude-code-action:v276-dsec-notsandbox.
Work continues on the runner when the client disconnects; there is no crash, restart or session recovery, and each run starts a fresh Claude session from GitHub context.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
No worker-crash or service-restart recovery and no session reconnection; continuity relies on GitHub thread and branch.
Transcript reload [partial] The execution_file output holds the run's messages for that job; there is no transcript reload across jobs. Sources: claude-code-action:v276-sdk.
The action installs Bun and Claude Code itself with retries; project dependencies, caches, services and ports are left to the user workflow.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
No dependency cache, snapshots, services or port handling is provided.
claude_args supports --max-turns to bound a Claude invocation.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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 [partial] The timeout_minutes input was removed; the job-level timeout-minutes is the documented limit. Sources: claude-code-action:v276-dmigr.
Turn limits [supported] claude_args supports --max-turns to bound a Claude invocation. Sources: claude-code-action:usage.
Token limits [partial] The result log reports each model's context window and max output tokens and MAX_MCP_OUTPUT_TOKENS is forwarded; the action sets no token budget. Sources: claude-code-action:v276-sdk, claude-code-action:v276-otel.
Spend limits [unknown] The action has no budget input; a spend cap would be a Claude CLI flag in the closed binary, which this repository cannot answer. Reason: closed component.
Provider failures [partial] A failed classification call posts comments unclassified; SDK errors write the execution file and fail the step; there is no provider failover. Sources: claude-code-action:v276-classify, claude-code-action:v276-sdk.
Enforcement [partial] A run that exceeds the configured max-turns is reported as failed; time limits are enforced by GitHub and spend limits are not enforced. Sources: claude-code-action:v276-sdk.
GitHub Actions logs expose execution diagnostics; full output may contain sensitive content. The FAQ documents permission, OIDC, tool and CI-access errors and their configuration remedies.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Failure reasons [supported] The FAQ documents permission, OIDC, tool and CI-access errors and their configuration remedies. Sources: claude-code-action:faq.
Time stats [supported] The tracking comment shows the run duration and the step-summary report shows duration and cost for that run; no cross-run time statistics exist. Sources: claude-code-action:v276-cost-ui, claude-code-action:v276-sdk.
Only a per-run total_cost_usd and model limits are logged; there is no aggregation, alert or budget.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
No per-epic, per-child or aggregated cost, no alerts and no reconciliation exist.
14 atomic answers
Task cost [partial] The run's total_cost_usd is logged and, with display_report on, printed as Cost in the step-summary report; it is not shown in the tracking comment and is not aggregated per issue. Sources: claude-code-action:v276-cost-ui, claude-code-action:v276-sdk.
Harness aggregation [unsupported] Only one harness runs, and nothing aggregates across runs. Sources: claude-code-action:v276-n-cost.
Billed vs estimated [unknown] total_cost_usd is produced by the closed Claude CLI and this repository does not say whether it is billed or estimated. Reason: closed component. Sources: claude-code-action:v276-sdk.
Tokens [partial] Only per-model contextWindow and maxOutputTokens appear in the sanitized log; token usage is withheld unless show_full_output is on. Sources: claude-code-action:v276-sdk.
Subscription allowance [unsupported] OAuth-token runs show no allowance information. Sources: claude-code-action:v276-n-cost.
Provider reconciliation [unsupported] No reconciliation with provider billing exists. Sources: claude-code-action:v276-n-cost.
The action has no resource-aware scheduling; concurrency and cancellation are GitHub workflow settings chosen by the user.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
unknown
Delivery
unknown
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
12 atomic answers
Concurrency [unsupported] The action sets no concurrency limit; GitHub concurrency groups are user workflow settings. Sources: claude-code-action:v276-n-queue.
Distinct GitHub users with repository write access can @claude on the same PR. Each instruction starts a new job that reads the shared PR history and operates on its head branch; an already-running job does not accept live input from other people.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Source pinned at 4fa93a88de49bbef68b10eb15aabedf28ca0e1ed; checked 2026-11-13. Default mentions and configured automation are distinguished; retained older citations identify their own revisions. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Shared work is the GitHub PR and its branch/history; each instruction starts a fresh action job, not a shared live terminal session.
Ordering [partial] Each trigger starts its own job and comments edited after the trigger are excluded; simultaneous instructions are not ordered by the action. Sources: claude-code-action:v276-fetch.
Conflicts [unsupported] No handling for two runs pushing to the same branch exists beyond git's own rejection. Sources: claude-code-action:v276-n-queue.
Permissions [supported] The action checks each triggering GitHub account for repository write/admin permission, with configurable explicit exceptions. Sources: claude-code-action:gap-shared-pr-users.
Repository write access normally gates supported actor-triggered events; documented allowed_non_write_users and bot overrides change this policy.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
allowedTools/disallowedTools configure Claude tool access, and GitHub token permissions bound API operations. The FAQ describes rebase/destructive Git restrictions as system-prompt instructions, not an independent enforcement boundary.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Defaults [supported] Docs define the default as auto mode, with acceptEdits as the earlier default for mentions, and a write-access gate. Sources: claude-code-action:v276-dusage.
Safer modes [supported] acceptEdits and default permission modes are documented for stricter control. Sources: claude-code-action:v276-dusage.
Enforcement [partial] The FAQ describes rebase/destructive Git restrictions as system-prompt instructions, not an independent enforcement boundary. Sources: claude-code-action:faq.
Interactive mode updates a single GitHub tracking comment with progress and final output.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Actions [supported] Interactive mode updates a single GitHub tracking comment with progress and final output. Sources: claude-code-action:readme.
Approvals [unsupported] The action records no approvals; the separate agent-approval-check action counts approvals as a commit status. Sources: claude-code-action:v276-n-store.
No review interface exists; Claude leaves inline PR comments and a PR link in GitHub.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
No product review UI, pending-review list, change requests, keyboard or narrow-screen support exists.
13 atomic answers
Pending reviews [unsupported] No review inbox exists; GitHub's PR list is outside the product. Sources: claude-code-action:v276-n-ui.
Change understanding [partial] The prompt asks Claude to summarize work in its comment; there is no guided change view. Sources: claude-code-action:v276-prompt.
Evidence navigation [supported] Inline comments are attached to files and lines of the PR. Sources: claude-code-action:v276-inline.
A colleague with write access can start a new run on the same issue or PR and branch, but the Claude transcript and runtime do not carry over.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Takeover is by new run; session history, runtime context and ownership transfer are not preserved.
Validated JSON results become workflow action outputs. With actions:read, the action can inspect CI runs, job logs and test results.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Screenshots [unsupported] The action does not capture or store screenshots; it only reads user-attached images. Sources: claude-code-action:v276-n-misc.
The action can retrieve configured CI evidence; enforcing branch protection remains a separate GitHub policy.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Lint [partial] Linters run only if Claude's rules allow them; the action does not require them. Sources: claude-code-action:v276-dcfg.
Types [partial] Type checks run only if Claude's rules allow them; the action does not require them. Sources: claude-code-action:v276-dcfg.
Ci [partial] The action can retrieve configured CI evidence; enforcing branch protection remains a separate GitHub policy. Sources: claude-code-action:config.
Required checks [partial] agent-approval-check posts a commit status that can be marked a required check; the main action's own runs are not required. Sources: claude-code-action:v276-appr-readme.
Failure blocking [partial] A pending agent-approval-check blocks merge only when GitHub branch protection requires it; the main action blocks nothing. Sources: claude-code-action:v276-appr-readme.
Default issue mentions create a work branch; existing open PR work pushes commits to that PR branch. The action creates and pushes commits; optional signing is configurable. Default mention mode returns a prefilled GitHub PR-creation link for the human to submit. The documented default workflow refuses rebase and directs the user to do it outside the action.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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] Default issue mentions create a work branch; existing open PR work pushes commits to that PR branch. Sources: claude-code-action:faq.
Commits [supported] The action creates and pushes commits; optional signing is configurable. Sources: claude-code-action:security.
Pull requests [partial] Default mention mode returns a prefilled GitHub PR-creation link for the human to submit. Sources: claude-code-action:security.
The default action documentation says it does not merge branches; this is described as a prompt restriction.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Merge [unsupported] The default action documentation says it does not merge branches; this is described as a prompt restriction. Sources: claude-code-action:faq.
Exact revision [partial] agent-approval-check binds /approve to the current head SHA; the main action does not check revisions. Sources: claude-code-action:v276-appr-readme.
The default result includes a human PR-creation link, distinct from evidence that a PR was created.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Pr links [partial] The default result includes a human PR-creation link, distinct from evidence that a PR was created. Sources: claude-code-action:security.
Stale approval invalidation [partial] agent-approval-check marks /approve comments for an older head SHA stale after a push and recomputes the status. Sources: claude-code-action:v276-appr-readme.
No batch review or shipping queue exists in the main action. The separate agent-approval-check action counts human approvals per pull request against the head SHA.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Only per-PR approval counting against the head SHA exists; no batch queue, grouping, summaries or dependency-aware merge queue.
Exact revision approval [partial] agent-approval-check binds approval to the head SHA for one PR at a time; there is no batch flow. Sources: claude-code-action:v276-appr-readme.
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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 action repository code is MIT licensed. Sources: claude-code-action:license.
Server license [unknown] The only server-side piece is Anthropic's hosted token-exchange endpoint, which is not in this repository. Reason: closed hosted component.
Hosted extensions [unknown] The Claude Code CLI binary and the Anthropic API and token-exchange services are closed or hosted and have no license in the repository. Reason: closed components.
Source visibility [supported] The whole repository is public under MIT. Sources: claude-code-action:v276-lic.
Hosted parity [unknown] No hosted edition of the action exists, but the hosted token-exchange behavior cannot be compared to source. Reason: hosted service.
The core action executes on the users own GitHub runner, still using GitHub and the chosen model provider. GitHub workflow infrastructure and the selected model provider remain dependencies.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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 [supported] The core action executes on the users own GitHub runner, still using GitHub and the chosen model provider. Sources: claude-code-action:readme.
Cancellation [unknown] The action has no account; GitHub App, provider and GitHub cancellation are hosted-service business terms.
Account independence [partial] A custom GitHub App or github_token and any supported provider work; a provider credential is always required and the default flow depends on Anthropic's token exchange. Sources: claude-code-action:v276-yml, claude-code-action:v276-venv.
Full debug output may include sensitive context, tokens and tool content and requires access restrictions.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
15 atomic answers
Code data [supported] Code and context go to the selected model provider through the Claude CLI, and inline comment text goes to the Anthropic Messages API for classification when an API key is set. Sources: claude-code-action:v276-venv, claude-code-action:v276-classify.
Credentials [supported] Provider credentials come from workflow secrets, the app token is exchanged through Anthropic's endpoint and revoked after the job, and secrets are redacted from fallback summaries. Sources: claude-code-action:v276-revoke, claude-code-action:v276-token.
Logs [partial] Full debug output may include sensitive context, tokens and tool content and requires access restrictions. Sources: claude-code-action:security.
Retention [unknown] Retention of data sent to model providers and held by GitHub is set by those hosted services. Reason: hosted service policy.
Deletion [unknown] Deletion terms of the model providers and GitHub are hosted-service policy not stated in the repository.
Training [unknown] Training use is a provider policy outside this repository. Reason: hosted service policy.
Residency [unknown] Residency depends on the provider and region the user configures (AWS_REGION, CLOUD_ML_REGION are forwarded); no residency statement is in the repository. Reason: hosted service.
Subprocessors [unknown] Subprocessors of the model providers and GitHub are not listed in the repository. Reason: hosted service policy.
Telemetry disclosure [partial] The code forwards OTEL variables and sends a map of which inputs were set to the Claude CLI; no privacy statement about it is in the repository. Sources: claude-code-action:v276-otel, claude-code-action:v276-inenv.
Telemetry defaults [partial] The action enables no OpenTelemetry export unless the user sets variables; the input-presence map is always passed to the CLI, whose own defaults are in the closed binary. Sources: claude-code-action:v276-otel, claude-code-action:v276-inenv.
Telemetry controls [partial] No action input disables telemetry; only forwarding of CLAUDE_CODE_ENABLE_TELEMETRY and OTEL variables is exposed. Sources: claude-code-action:v276-otel.
Consent prompt [unsupported] There is no prompt or consent flow; runs are non-interactive. Sources: claude-code-action:v276-run.
Provider separation [supported] Docs separate Anthropic, Bedrock, Vertex and Foundry routes, and the orchestration vendor's token exchange is a separate step. Sources: claude-code-action:v276-dcloud.
The action itself is free (MIT) with no separate price; Claude usage is paid via the user's Anthropic API key, cloud provider (Bedrock/Vertex/Foundry) or Claude Pro/Max subscription, and it runs on the user's own GitHub runner.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: First-party pricing sources read 2026-11-10 Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
unknown
Delivery
unknown
Checked
2026-11-10 (12 days old)
Agent cross-check
2026-11-17
First-party pricing pages, terms and licences read 2026-11-10; no purchase, billing or authenticated account test.
Model cost [supported] Model calls are billed by the chosen provider: Anthropic API key, Bedrock, Vertex AI, Foundry, or a Claude Pro/Max OAuth token; no amounts in these sources. Sources: claude-code-action:readme-nov10, claude-code-action:setup-nov10.
Compute cost [partial] It executes on the user's own GitHub runner; runner costs are not stated in these sources. Sources: claude-code-action:readme-nov10.
The action adds no fee; scenario costs depend on provider token prices and runner minutes that are not in the repository.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Parts of this criterion are unsupported, partial or unknown; see the atomic answers.
11 atomic answers
One person [unknown] Pricing a workload needs provider token rates and runner minute prices, which are hosted-service business facts not in this repository.
Two people [unknown] Same gap: provider and runner prices are not in the repository.
Five people [unknown] Same gap: provider and runner prices are not in the repository.
Task assumptions [unsupported] The repository defines no standard task, token or runtime assumptions. Sources: claude-code-action:v276-lic.
Concurrency [unknown] Concurrency cost depends on runner pricing, a hosted-service fact.
Tokens [unknown] Token cost depends on the provider's price list, not stated in the repository.
Compute hours [unknown] Compute-hour cost depends on GitHub runner pricing or the user's own runner.
The latest retrieved version is v1.0.269 dated 2026-11-06.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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.
Support [partial] Issues are on GitHub and security reports go through HackerOne; no support channel or SLA is stated. Sources: claude-code-action:v276-secmd.
Response commitments [unsupported] No response-time commitment appears in SECURITY.md, README or docs. Sources: claude-code-action:v276-secmd.
Release activity [supported] The latest retrieved version is v1.0.269 dated 2026-11-06. Sources: claude-code-action:release.
Status history [partial] Release history is on GitHub; no status page is linked in the repository. Sources: claude-code-action:v276-release.
The action is MIT-licensed open source used from its public repository at no charge, with no paid plan shown. Model usage is paid to Anthropic or the chosen cloud provider, and runner costs are GitHub's.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
2026-11-17
Provider free tiers, GitHub Actions minute costs and card requirements were not checked. No workflow was run.
Self host cost [partial] Runs on the user's GitHub runner with no fee for the action; provider and runner costs remain separate. Sources: claude-code-action:src4-readme.
Free scenarios [partial] Every scenario runs for free except provider usage and runner minutes, which the user pays; no scenario works without a provider credential. Sources: claude-code-action:v276-lic, claude-code-action:v276-venv.
Paid evaluation [partial] Evaluation needs a paid provider account or token for model usage, but the action itself costs nothing. Sources: claude-code-action:v276-venv.
Pricing transparency: Checked Claude Code Action's pinned README, README.md, docs/setup.md, docs/usage.md, docs/configuration.md, docs/security.md, docs/faq.md, docs/cloud-providers.md, docs/custom-automations.md, ROADMAP.md, LICENSE, releases and recent public issues. This pass did not establish the remaining named requirement. No installed-product scenario or authenticated team fixture was run.
Scope
Surfaces: Claude Code GitHub Action and GitHub issue/PR comments · Version: Repository 115e0b2c49222c831de6748ccdf365bb8d3ea9a3; first-party source pass 2026-11-08. Default mention workflow and configurable automation are distinguished. Some older capability docs conflict with current v1 configuration. Source-depth answers added 2026-11-16 from release v1.0.276 (f3b070c88125); older citations identify their own revisions. · 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
Parts of this criterion are unsupported, partial or unknown; see the atomic answers.
The distribution is a composite GitHub Action that installs and runs Claude on the job runner. The repository contains action entrypoints, base-action code, GitHub MCP tools and docs, with no desktop/native-phone application or separately distributed user CLI.
Claude updates a single comment, creates a branch or pushes to the PR branch, cannot submit or approve PR reviews, cannot push to other branches, and does not rebase or merge.
Kind
documented
Locator
What Claude can and cannot do (v1.0.276, f3b070c88125)
Searched dashboard, webview, localhost, createServer, express, keyboard, aria, electron, vscode, vsix, ios and android across the repository: no web, desktop, mobile or editor client exists; docs/create-app.html is a one-off GitHub App manifest helper page.
Kind
source-inspected
Locator
bounded scan for own UI, dashboards and clients (v1.0.276, f3b070c88125)
The action installs Claude Code and invokes the Claude SDK run path. Its documented executable override expects a Claude Code executable; MCP tool customization does not replace the coding harness.
Installs the closed Claude Code CLI (pinned 2.1.325) with curl from claude.ai/install.sh, retrying up to three times with a 5 s pause, or uses path_to_claude_code_executable; no other harness can be selected.
Kind
source-inspected
Locator
buildInstallCommand and installClaudeCode (v1.0.276, f3b070c88125)
Repository admins install the app/workflow and set provider secrets; Claude Pro/Max users can generate a Claude Code OAuth token with claude setup-token.
Passes ANTHROPIC_API_KEY, CLAUDE_CODE_OAUTH_TOKEN, workload-identity IDs, ANTHROPIC_BASE_URL and Bedrock/Vertex/Foundry variables to the run; no local-model or credit mechanism.
The run fails unless one provider credential is present (API key, OAuth token, workload identity, or Bedrock/Vertex/Foundry variables) and rejects more than one cloud provider at once.
The action is a composite GitHub Action: it installs Bun, runs bun install and executes src/entrypoints/run.ts on whatever runner runs the job; optional bubblewrap/socat subprocess isolation installs only when allowed_non_write_users is set on Linux.
Single orchestrator: sets up the GitHub token, checks write permission, prepares tag or agent mode, installs Claude Code, restores config from the PR base, runs Claude and writes conclusion, execution_file, session_id and structured_output outputs.
Without an override github_token, the action exchanges the job OIDC token at https://api.anthropic.com/api/github/github-app-token-exchange for a GitHub App token, with retry; that endpoint is a hosted Anthropic service.
Kind
source-inspected
Locator
exchangeForAppToken and setupGitHubToken (v1.0.276, f3b070c88125)
Issues and closed PRs get a new branch named from branch_prefix and a template; open PRs are checked out and pushed in place; shallow fetch depth is chosen from the PR commit count. No worktree is created.
Write access gates triggers; allowed_bots and allowed_non_write_users override it; best-effort scrubbing and bubblewrap PID-namespace isolation apply with allowed_non_write_users; the token is repo-scoped and short-lived.
Kind
documented
Locator
Access Control and untrusted PR sections (v1.0.276, f3b070c88125)
Interactive mode gives Claude Glob, Grep, LS, Read, the tracking-comment tool, three CI read tools and either git wrapper scripts or API commit_files and delete_files; file edits come from the permission mode.
Kind
source-inspected
Locator
tagModeTools and prepareMcpConfig (v1.0.276, f3b070c88125)
The prompt tells Claude to keep a todo checklist in its single tracking comment, follow CLAUDE.md, add todos as scope grows, push to the working branch and give a pre-filled PR link; it also includes optional Fix-this links.
On pull requests, .claude, .mcp.json, .claude.json, .gitmodules, .ripgreprc, CLAUDE.md, CLAUDE.local.md and .husky are restored from the base branch; PR versions are kept under .claude-pr.
After the run, compares the branch to base, adds a branch link to the comment when commits exist and deletes the remote branch when Claude made no commits.
Searched sqlite, database, datastore, redis, persist, epic, blocked_by, blocker, subtask, backlog, priority, spec, requirement and drift: no action-owned issue, task, hierarchy, dependency or specification store exists; only GitHub issues, PRs and the single tracking comment are read and written.
Kind
source-inspected
Locator
bounded scan of src/, base-action/src/, action.yml and docs/ for issue, task and spec stores (v1.0.276, f3b070c88125)
Snapshots and restores GITHUB_ENV, GITHUB_PATH and git config after the session, records the only branch the push wrapper may update, disables hooks and clears session state; no session store persists between jobs.
Searched memory, remember, MEMORY.md, embedding, vector, index, semantic and code search in src/ and base-action/src/: no memory store, index or retrieval service exists; the only persistence is GitHub itself.
Kind
source-inspected
Locator
bounded scan for memory and context stores (v1.0.276, f3b070c88125)
Each workflow invocation prepares one Claude run and writes results to job-scoped RUNNER_TEMP. It reads GitHub issue/PR context, with no native issue database or live multi-human input channel. Separate workflow jobs have no action-wide concurrency lock.
Logs a sanitized result with duration_ms, num_turns, total_cost_usd, permission_denials_count and per-model contextWindow and maxOutputTokens; records session_id from the init message; reports failure when turns exceed the configured max-turns; writes an execution file on error.
Kind
source-inspected
Locator
sanitizeSdkOutput and runClaudeWithSdk (v1.0.276, f3b070c88125)
The only role model is GitHub repository permission: getCollaboratorPermissionLevel must report write or admin, unless allowed_non_write_users names the actor.
For PRs whose author lacks write access, Bash rules for package managers, test runners, interpreters, docker, make and shells are denied even if claude_args allows them.
Searched queue, mutex, lock, concurrency, quota, budget, max-budget, cpu, ram, throttle, retry-run and fairness in src/ and base-action/src/: none; concurrency groups are GitHub workflow features chosen by the user.
Kind
source-inspected
Locator
bounded scan for scheduling, queue and quota logic (v1.0.276, f3b070c88125)
Default workflow creates/pushes commits but directs humans to create PRs and rebase. Destructive Git restrictions are described as system-prompt rules, not a tool-enforced security boundary.
States that the untrusted-PR command list is not a sandbox, that the restore of session files is not a sandbox, and that a cancelled or timed-out run does not restore anything.
Kind
documented
Locator
Running code from pull requests and What a session can leave behind (v1.0.276, f3b070c88125)
Write-access checks gate supported actor-driven events, with explicit opt-in exceptions. Default mentions push a branch and return a prefilled PR link. Full debug output can expose private data.
Only forwards CLAUDE_CODE_ENABLE_TELEMETRY and OTEL_* variables, and MCP timeout variables, from the workflow env to the Claude CLI; there is no wrapper telemetry or opt-out input.
Kind
source-inspected
Locator
MCP and telemetry configuration env (v1.0.276, f3b070c88125)
Searched worktree, docker, Dockerfile, container, firecracker, vm, nsjail and network egress controls in src/, base-action/ and action.yml: the job runs on the runner directly with no container, VM, worktree or network policy configured by the action; only the optional bubblewrap install for allowed_non_write_users and the PR config restore exist.
Kind
source-inspected
Locator
bounded scan for isolation mechanisms (v1.0.276, f3b070c88125)
After the run, unconfirmed inline review comments are sent to the Anthropic Messages API (claude-haiku-4-5) using ANTHROPIC_API_KEY to be classified as real review versus probe; with no key they post unclassified.
Searched total_cost_usd, usage, cost, budget, billing and spend in src/ and base-action/src: total_cost_usd and duration are read from the Claude result for the sanitized log, the tracking comment (duration) and the step-summary report (cost and duration) of the single run; no aggregation, alert or budget logic exists.
Kind
source-inspected
Locator
bounded scan for cost and usage aggregation (v1.0.276, f3b070c88125)
The final tracking-comment update reads total_cost_usd, duration_ms and duration_api_ms from the last result element of the execution file; the comment body shows the run duration (comment-logic.ts), and format-turns.ts prints a Cost and Duration line in the step-summary report.
Kind
source-inspected
Locator
executionDetails read from the execution file (v1.0.276, f3b070c88125)
Each triggering GitHub actor is checked against repository write/admin access (or an explicitly configured user allowlist). Distinct authorized GitHub users can invoke the action on a shared PR.
Each mention job reads the shared PR body, comments, reviews and changed files. The current trigger comment supplies instructions; previous comments remain context. Multiple authorized people can give successive instructions on the same PR work, with each instruction starting a fresh job and no live input to an already-running job.
The action checks out the existing PR head branch, so follow-up jobs from distinct GitHub users operate on the same PR work rather than separate private copies.
Searched role, sso, saml, okta, scim, organization, membership and workspace across src/, base-action/src/ and action.yml: the only access model is the GitHub repository permission check; there are no roles, SSO or provisioning.
Kind
source-inspected
Locator
bounded scan for roles, SSO and membership (v1.0.276, f3b070c88125)
Searched export, import, rollback, revert, upload-artifact, gallery, screenshot, recording and preview in src/, base-action/src/, action.yml and docs/: no export or import tooling, artifact upload, gallery, recording, preview or rollback feature exists in the action.
Kind
source-inspected
Locator
bounded scan for export, import, rollback and artifact upload (v1.0.276, f3b070c88125)
A second action in the repository that gates pull requests containing agent-authored commits behind a configurable number of human approvals via an agent-approval-check commit status.
Kind
source-inspected
Locator
agent-approval-check composite action and inputs (v1.0.276, f3b070c88125)
Counts distinct human approvals of the current head SHA, flags earlier /approve comments stale after a push, posts a pending or success commit status, and relies on branch protection marking it required.
Kind
documented
Locator
What it does, Setup, Approving (v1.0.276, f3b070c88125)
The action runs on the user's own GitHub runner and authenticates to Anthropic API, Amazon Bedrock, Google Vertex AI or Microsoft Foundry, so model calls are billed by the chosen provider.
Kind
documented
Locator
Intro paragraph; Features > 'Runs on Your Infrastructure'
Setup installs the GitHub app and secrets from Claude Code with /install-github-app or manually; the README names provider credentials but no price or paid plan for the action, and states the MIT licence.
Kind
documented
Locator
README.md Quickstart (lines 24-33) and License (lines 69-71), commit 4fa93a8