--- sidebar_label: fullsend inference --- # fullsend inference Manage the inference credentials agent runs use. `provision`, `deprovision` and `status` create, inspect, and remove the GCP Workload Identity Federation (WIF) pool, OIDC provider, and IAM bindings that let GitHub Actions workflows authenticate with GCP for Agent Platform (Vertex) access. [`openai`](#inference-openai) enrols repositories with OpenAI WIF for GPT models on the pi runtime. [`gateway status`](#inference-gateway-status) is not part of WIF: it checks a repository's `inference.gateway` block and whether the gateway accepts the job's OIDC assertion through direct OIDC trust (no token exchange). ## Commands | Command | Description | Who runs it / permission | |---------|-------------|--------------------------| | `fullsend inference provision ` | Create WIF pool/provider and grant Agent Platform access | GCP project admin, with `roles/iam.workloadIdentityPoolAdmin` and `roles/resourcemanager.projectIamAdmin` | | `fullsend inference deprovision ` | Remove org or repo from WIF | GCP project admin, with `roles/iam.workloadIdentityPoolAdmin` | | `fullsend inference status ` | Check WIF health and print config values | Read access on the GCP project: `roles/iam.workloadIdentityPoolViewer` and `roles/browser` | | `fullsend inference openai request [,…]` | Generate WIF provider/mapping request for OpenAI admin | Anyone; needs no credentials | | `fullsend inference openai import [reply.json]` | Import OpenAI WIF identifiers into config | Repo admin. Writes `.fullsend/config.yaml` locally, to commit afterwards; `--variables` needs a GitHub token that can set repository variables | | `fullsend inference openai status ` | Check OpenAI WIF configuration and exchange status | Config checks: anyone, from a checkout of the repository. Live exchange: only inside a GitHub Actions job in `` with `id-token: write`, so a repo writer adds the workflow | | `fullsend inference gateway status ` | Check inference gateway configuration and authentication | Config checks: anyone, from a checkout of the repository. Live check, `oidc` mode only: inside a GitHub Actions job in `` with `id-token: write`, so a repo writer adds the workflow | ## `inference provision` Creates a WIF pool (`fullsend-inference`), an OIDC provider (`github-oidc`), and grants `roles/aiplatform.user` to the WIF principal. Idempotent and safe to re-run. ```bash fullsend inference provision \ --project "" ``` Per-repo mode scopes the WIF provider to a single repository: ```bash fullsend inference provision \ --project "" ``` ### Flags | Flag | Default | Description | |------|---------|-------------| | `--project` | | GCP project ID | | `--region` | `global` | GCP region | ### Required IAM roles | Role | Description | |------|-------------| | `roles/iam.workloadIdentityPoolAdmin` | Create WIF pool and provider | | `roles/resourcemanager.projectIamAdmin` | Grant `roles/aiplatform.user` to WIF principals | ### Required GCP APIs ```bash gcloud services enable \ iam.googleapis.com \ cloudresourcemanager.googleapis.com \ aiplatform.googleapis.com \ --project="$GCP_PROJECT" ``` ## `inference deprovision` Removes an org or repo from WIF by deleting the IAM binding and (optionally) the WIF provider. ```bash fullsend inference deprovision \ --project "" ``` ### Required IAM roles | Role | Description | |------|-------------| | `roles/iam.workloadIdentityPoolAdmin` | Modify WIF pool and provider | ## `inference status` Checks WIF health and prints the configuration values needed for `github setup`. ```bash fullsend inference status \ --project "" ``` Read-only — makes no changes. ## `inference openai` Commands for enrolling repositories with OpenAI Workload Identity Federation (see the [operator guide](../guides/infrastructure/openai-workload-identity.md)). They need neither GCP access nor an OpenAI key. What reaches the network, and what does not: | Command | Network | |---|---| | `request` | None. The document is computed from the repository names. | | `import` | None, unless `--variables` is passed: that calls the GitHub API through the forge client to set the three repository variables. | | `status` | Reads configuration only, except inside a GitHub Actions job with `id-token: write`, where it performs one token exchange with OpenAI to prove the mapping accepts *that* job's repository — reporting the granted scope and expiry, never the token. | No OpenAI API key is used or created by any of them. ### `inference openai request` Generates the request document an administrator needs to enable OpenAI WIF for one or more repositories. Every value is computed from the repository names. Nothing is sent anywhere; the command needs no credentials. ```bash fullsend inference openai request [,…] \ [--audience ""] \ [--project ""] \ [--service-account ""] \ [--ref "refs/heads/"] \ [--format json|md] \ [--out ] ``` | Flag | Default | Description | |------|---------|-------------| | `--audience` | `fullsend://` | OpenAI Workload Identity audience | | `--project` | *(empty)* | OpenAI project name or ID for the service accounts | | `--service-account` | *(none)* | Map this existing service account instead of asking for `fullsend--ci` to be created in the mapping | | `--ref` | *(none)* | Optional tightening. Must be a full ref (`refs/…`). When set, emits **two mappings** per repository — one asserting the given ref (e.g. `refs/heads/main`) and one asserting `refs/pull/*` — so both branch and PR-review-triggered runs work. Without `--ref`, mappings assert `iss`, `aud`, and `repository` only (no ref), matching the Vertex path's repository-scoped trust. The two-mapping cost halves the 50-mapping-per-provider budget to 25 repositories | | `--format` | `json` | Output format: `json` (versioned schema) or `md` (copy-paste ticket) | | `--out` | *(stdout)* | Write output to a file | ### `inference openai import` Takes the administrator's reply and writes `inference.openai` into `.fullsend/config.yaml` through the same setters as `fullsend github setup --openai-*`. All three identifiers must be present — a partial trio is refused, and the config is validated before it is written. The write is local: commit `.fullsend/config.yaml` afterwards, since fullsend reads the base branch for pull-request events. The reply file may be either the bare reply object or the whole document `request --format json` produced with its `reply` section filled in — an administrator can edit and return the same file. When it names service accounts for several repositories, pass `--repo ` to choose one, or `--service-account-id` to give the value outright. ```bash # From a reply JSON file: fullsend inference openai import reply.json # From flags: fullsend inference openai import \ --audience "fullsend://" \ --identity-provider-id "" \ --service-account-id "" # Set repository variables instead of config.yaml: fullsend inference openai import \ --variables --repo \ --audience "fullsend://" \ --identity-provider-id "" \ --service-account-id "" ``` | Flag | Default | Description | |------|---------|-------------| | `--audience` | | OpenAI Workload Identity audience | | `--identity-provider-id` | | OpenAI identity provider ID | | `--service-account-id` | | OpenAI service account ID | | `--fullsend-dir` | `.fullsend` | Path to the .fullsend configuration directory | | `--variables` | `false` | Set `FULLSEND_OPENAI_*` repository variables instead of config.yaml | | `--repo` | | Two roles: the target repository for `--variables`, and the repository to select from a reply that names several (`service_account_ids`). Matched case-insensitively | ### `inference openai status` Prints the resolved OpenAI WIF identifiers and their source (config layer or environment variables), and flags a partial trio. When nothing is configured, reports that a run will refuse the openai provider and names both remedies: the WIF trio, or the `FULLSEND_OPENAI_API_KEY` repository secret (exported as `OPENAI_API_KEY`). A static `OPENAI_API_KEY` already in the environment is reported as the source; in CI the runner warns and WIF remains preferred. When run inside a GitHub Actions job with `id-token: write`, performs one exchange and reports the returned scope and expiry without ever printing the token. ```bash fullsend inference openai status \ [--fullsend-dir ".fullsend"] ``` | Flag | Default | Description | |------|---------|-------------| | `--fullsend-dir` | `.fullsend` | Path to the .fullsend configuration directory | ## `inference gateway status` Prints the resolved `inference.gateway` block ([ADR 0137](../ADRs/0137-inference-gateway-credential-route.md)) — `url`, `audience`, the `auth` mode and the model list (`models` or `models_file`) — with the config layer each comes from (`config.yaml` or `config.base.yaml`), and flags a partial block (url missing, or audience missing in the `oidc` mode). The block has no environment-variable override. For a complete block it also reports the Claude Code route: the `ANTHROPIC_BASE_URL` a claude runtime `gateway/` run gets, and how the credential reaches it (an `apiKeyHelper` in the `oidc` mode, an `ANTHROPIC_AUTH_TOKEN` placeholder in the `api-key` mode, sent as `Authorization: Bearer` in both, and in the `oidc` mode Claude Code also copies the helper value into `x-api-key`; see [Claude Code § Inference gateway route](../runtimes/claude.md#inference-gateway-route)). In the `api-key` mode ([ADR 0138](../ADRs/0138-inference-gateway-api-key-credential-mode.md)) it warns that the mode relies on a long-lived secret and recommends `oidc` when the gateway supports it, reports whether `FULLSEND_INFERENCE_GATEWAY_API_KEY` is set (never its value), and stops at the config checks. In the `oidc` mode, when run inside a GitHub Actions job in `` with `id-token: write`, fetches one GitHub OIDC assertion for the configured audience, reports its expiry (`exp`, lifetime, time left), and sends one authenticated request, `GET /v1/models` with the assertion as a bearer token (under the url's path prefix when it has one, for example `GET https://gateway.example.com/team-a/v1/models`). It reports the HTTP status and reads the reply as an OpenAI-style model list (`{"data":[{"id":...}]}`, at most 1 MiB): a 2xx list with entries means the gateway accepted the assertion, and the command prints how many models the gateway authorises for `` and their ids (the first 20, with JWT-shaped values redacted, control characters replaced by `?` and every `::` printed as `: :` so an id cannot form a workflow command); an empty list means the assertion was authenticated but no models are authorised for the repository, which is an error; 401 or 403 means the gateway refused the assertion; 3xx is an error (redirects are not followed); any other status (for example 404, 405, 429 or 5xx), or a 2xx reply that is not a model list, means authentication could not be confirmed, which is also an error. The request goes out over `https` only, ignores proxy settings, and refuses to connect to loopback, private or other internal addresses. The assertion is never printed. Outside Actions, it says so and stops at the config checks. ```bash fullsend inference gateway status \ [--fullsend-dir ".fullsend"] ``` | Flag | Default | Description | |------|---------|-------------| | `--fullsend-dir` | `.fullsend` | Path to the .fullsend configuration directory | ## See also - [Getting inference for fullsend](../guides/getting-started/getting-inference.md) — getting started guide - [OpenAI Workload Identity](../guides/infrastructure/openai-workload-identity.md) — end-to-end OpenAI WIF setup guide - [Advanced setup](../guides/infrastructure/advanced-setup.md) — non-standard installation paths and WIF configuration - [CLI internals](../guides/dev/cli-internals.md) — command tree and implementation details