Private repository enclaves | GitHub Agentic Workflows
Skip to main content Private repository enclaves The top-level enclaves array enables finite-disclosure access to approved private repositories. The compiler registers enclave_run_script or enclave_run_agent from the keyed entries present on the awf-enclave MCP route. Omit the array to disable enclaves. Enclaves require AWF network isolation, which every supported sandbox.agent.runtime profile provides, so the compiler launches mcpg in bridge mode and AWF can attach it to the isolated topology. sandbox : agent : id : awf enclaves : - script : repos : - repo : octo-org/private-service sensitivity : confidential timeout : 45 - agent : model : gpt-5 repos : - repo : octo-org/private-service sensitivity : confidential timeout : 180 Each repos entry uses public , trusted , internal , confidential , or sealed sensitivity. trusted is unmetered and permits free-form string values inside an otherwise strict structured response schema, but is appropriate only when repository content is approved for unrestricted return to the primary agent without confidentiality accounting.
Do not select it merely to obtain string output. All other sensitivities are finite-schema-only. Each type can appear at most once. When the same repository appears in both entries, its sensitivity must match because its information budget is shared across executor types. AWF fixes the script enclave network and interpreter and the agent enclave network internally; workflows cannot override those security invariants. The generated gateway upstream uses a fresh masked capability for each workflow run.
That capability is passed only to mcpg and AWF and is excluded from the primary agent environment. The gateway allows 120 seconds for the AWF-owned HTTP upstream to become available. It enforces a 4,860-second tool timeout, covering AWF’s maximum 4,800-second finite-disclosure timing bucket plus a 60-second transport allowance. Executor timeouts are capped at 4,740 seconds because AWF reserves 60 seconds in the final bucket for processing and cleanup.
The gateway timeout is an enforcement bound, not an absolute AWF wall-clock guarantee under pathological host cleanup or scheduler stalls. Docker setup and host repository credentials Section titled “Docker setup and host repository credentials” The default AWF release is v0.29.7 . For Docker enclaves, compilation generates authenticated installation and probing of the same-release, runner-UID-scoped storage helper.
Only helper provisioning is privileged; AWF remains rootless with network isolation and writable invocation storage mounted noexec . Copilot native addons load from the sealed image runtime, not writable storage. Use sandbox.agent.host-github-token for host-side repository staging: sandbox: agent: runtime: docker host-github-token: ${{ secrets.GH_AW_GITHUB_TOKEN }} The execution step rejects a missing staging credential and binds it as host GH_TOKEN , automatically excluded from the primary agent environment.
This is separate from the enclave GitHub-tool credential and does not persist secrets through GITHUB_ENV . Grant only the repository read access needed for staging. Both agent and detection installers verify the release-bound checksum manifest against a trusted SHA256 pin and verify payload checksums, without querying release attestations. Public Docker diagnostics Section titled “Public Docker diagnostics” Set sandbox.agent.enclave-debug-public: true only for static, public-only Docker enclaves using official release images.
The default is false , explicitly serialized to disable AWF’s automatic public capture. Confidential, private, mixed, dynamic, custom-image, and unsupported runtime configurations cannot opt in. AWF additionally verifies repository visibility anonymously before capture. Public collection currently supports the reviewed AWF v0.29.5 release; later releases require aligned collector and interpreter pins before they can opt in.
Compilation installs the version-aligned upstream diagnostic collector and a checksum-verified Node executable under protected host paths. It generates prepare before execution, collect after execution and teardown (including execution failure), and a pinned artifact upload of only the collector’s validated runner-readable upload/ directory. Collection rejection can publish bounded status without partially validated streams; execution failure, collection failure, and absent evidence remain distinct.
Diagnostic artifacts have a fixed one-day retention ceiling, unaffected by repository retention variables. Credential redaction does not redact private repository content: never enable public diagnostics for confidential targets. No collector, interpreter, trusted-root, or retention override is exposed. Cloud Hypervisor enclave runtime Section titled “Cloud Hypervisor enclave runtime” Set runtime: cloud-hypervisor on a static script or agent entry to select AWF’s trusted host enclave microVM executor.
Omitting runtime , or setting it to docker , retains Docker execution. Cloud Hypervisor enclaves require AWF v0.28.47 or newer, GitHub-hosted Ubuntu x86_64 KVM runners, and release-attested enclave artifacts. The compiler provisions the primary and enclave artifacts from the same AWF release tag so their identities stay coordinated. on: workflow_dispatch runs-on: ubuntu-24.04 engine: copilot sandbox: agent: runtime: cloud-hypervisor version: v0.28.51 enclaves: - script: runtime: cloud-hypervisor repos: - repo: octo-org/private-service sensitivity: confidential The compiler enables the AWF preview and stages its runtime artifacts without changing the primary agent’s runtime, mounts, or TTY.
A Cloud Hypervisor primary may be combined with static script enclaves when both explicitly select cloud-hypervisor and AWF is v0.28.51 or newer; pin sandbox.agent.version explicitly when the compiler’s default is older. This primary-plus-enclave configuration is experimental and not evidence of Linux/KVM isolation acceptance or production readiness. Static agent enclaves with a Cloud Hypervisor primary remain unsupported except for the acceptance fixture described below.
All enclave entries must select Cloud Hypervisor together: mixed Docker/default and VM entries, dynamic , and enclave image overrides are unsupported. The primary agent cannot use sbx or nvx ; runner.topology: arc-dind and Docker host path prefixes are also unsupported. Agent enclaves additionally require agent.model , the API proxy with a configured provider route matching their fixed engine/profile, and no Docker-in-Docker.
Only the default agent.engine: copilot is currently implemented; its route requires the emitted COPILOT_GITHUB_TOKEN or COPILOT_PROVIDER_API_KEY credential, not just a target hostname or pricing overlay. Static agent.github and agent.tools.github access is unavailable in this VM preview because the scoped executor bearer handoff is not yet supported. Repository disclosure limits and the existing MCP tool contract remain unchanged.
Cloud Hypervisor agent acceptance fixture Section titled “Cloud Hypervisor agent acceptance fixture” The checked-in boolean features.cloud-hypervisor-agent-acceptance: true admits only smoke-enclave-cloud-hypervisor-copilot.md , with on: workflow_dispatch as its sole trigger and exactly if: github.repository == 'github/gh-aw-firewall' && github.ref_type == 'branch' . An ambient GH_AW_FEATURES flag cannot enable it.
Both runtimes must explicitly select cloud-hypervisor ; the primary uses Copilot, and the sole enclave uses explicit agent.engine: copilot , agent.model: claude-sonnet-5 , no profile override, and only github/gh-aw-firewall with sensitivity: public . This fixture requires sandbox.agent.version: v0.28.62 , timeout: 90 , max-output-bytes: 256 , max-invocations: 1 , agent.max-task-bytes: 4096 , agent.max-model-requests: 8 , and agent.max-model-tokens: 2048 .
The model request/token settings are advisory, not enforced inference ceilings in the pinned guest. Ex
The rest of this page is on the live site.