Model profiles

ReviewPhin selects models through named profiles stored in the database. When no profile is active, it uses the bundled GitHub Copilot CLI.

A model profile changes how the Copilot CLI harness talks to a backend. Today ReviewPhin has one supported harness: the GitHub Copilot CLI wrapper. OpenAI-compatible, Azure OpenAI, and Anthropic profiles configure that harness with different provider settings; they do not switch to a different runtime.

Examples use reviewphin for readability — see running the CLI.

Resolution order

1. /reviewphin-profile <name> in the code review description.
2. The tenant profile set with tenant set-profile.
3. The database default profile.
4. Plain Copilot CLI fallback.

Copilot CLI profile

When no model profile is active, ReviewPhin uses native Copilot access. For non-interactive runs, set one GitHub token variable:

GH_TOKEN=github_pat_xxxxxxxxxxxxxxxxxxxxxxxxxxxx
# or
GITHUB_TOKEN=github_pat_xxxxxxxxxxxxxxxxxxxxxxxxxxxx
# or
COPILOT_GITHUB_TOKEN=github_pat_xxxxxxxxxxxxxxxxxxxxxxxxxxxx

The token owner needs GitHub Copilot access. If Copilot access comes through an organization or enterprise, Copilot CLI must also be enabled by policy. For local interactive runs, copilot auth login can be used instead of a PAT.

Create a native Copilot profile when you want to pin a model while keeping Copilot’s own backend:

reviewphin model-profile add \
  --name copilot-gpt5.4 \
  --review-model gpt-5.4 \
  --text-generation-model gpt-5.4-mini \
  --default

Before pinning a model, list the catalog available to the current Copilot identity:

reviewphin model-profile available-models

The catalog comes from GitHub Copilot at command time. It can vary by token, account, organization policy, and date. To inspect the credentials stored in a native profile, use --model-profile:

reviewphin model-profile available-models --model-profile copilot-team-a

To check a GitHub token before creating a profile, pass it directly:

reviewphin model-profile available-models --auth-token github_pat_xxxxxxxxxxxxxxxxxxxx

The direct token is used only for that catalog request. ReviewPhin does not store or display it. --model-profile and --auth-token cannot be used together.

The table includes the model ID, display name, supported reasoning efforts, default effort, and vision support. Use --output plain for tab-separated rows or --output json for structured output.

OpenAI-compatible endpoint

Any provider that exposes an OpenAI-compatible API can be used with --provider-type openai.

reviewphin model-profile add \
  --name byok-llama \
  --base-url http://vllm-host:8000/v1 \
  --provider-type openai \
  --review-model meta-llama/Llama-3.1-8B-Instruct \
  --ignore-missing-model

No --auth-token is needed when the endpoint runs without an API key. If the endpoint requires one, add --auth-token <key>.

For OpenAI’s hosted API, use https://api.openai.com/v1:

reviewphin model-profile add \
  --name openai-gpt5.4 \
  --base-url https://api.openai.com/v1 \
  --provider-type openai \
  --auth-token sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx \
  --review-model gpt-5.4 \
  --text-generation-model gpt-5.4-mini \
  --ignore-missing-model \
  --default

Azure OpenAI

Use the deployment name as --review-model.

reviewphin model-profile add \
  --name azure-gpt5.4 \
  --base-url https://my-resource.openai.azure.com \
  --provider-type azure \
  --auth-token your-azure-api-key \
  --review-model my-gpt5.4-deployment \
  --text-generation-model my-gpt5.4mini-deployment \
  --ignore-missing-model

Anthropic

reviewphin model-profile add \
  --name anthropic-claude \
  --base-url https://api.anthropic.com \
  --provider-type anthropic \
  --auth-token sk-ant-xxxxxxxxxxxx \
  --review-model claude-opus-4.8 \
  --text-generation-model claude-sonnet-4.6 \
  --ignore-missing-model

Use --wire-api completions only for compatibility endpoints that do not support the default Responses-style API.

Review and text models

For cost efficiency, configure a stronger review model and a lighter text-generation model:

reviewphin model-profile add \
  --name production \
  --base-url https://api.openai.com/v1 \
  --provider-type openai \
  --auth-token sk-xxx \
  --review-model gpt-5.4 \
  --text-generation-model gpt-5.4-mini \
  --ignore-missing-model \
  --default

When --text-generation-model is omitted, ReviewPhin uses the review model for all model-backed tasks. The text-generation model is used for lighter work such as memory coalescing and reply text.

Reasoning effort

Set reasoning effort independently for each role with --review-reasoning-effort and --text-generation-reasoning-effort. Accepted values are low, medium, high, and xhigh.

reviewphin model-profile add \
  --name gpt56-review \
  --review-model gpt-5.6 \
  --review-reasoning-effort high \
  --text-generation-model gpt-5.6-mini \
  --text-generation-reasoning-effort low \
  --default

The two roles are independent — the text-generation effort does not inherit the review effort, even when the text-generation model falls back to the review model.

When an effort is left unset (or cleared), ReviewPhin omits reasoningEffort from the model session entirely and the underlying harness keeps its own default. It does not substitute another value.

Clear a previously set effort with the matching clear flag:

reviewphin model-profile add --name gpt56-review --clear-review-reasoning-effort
reviewphin model-profile add --name gpt56-review --clear-text-generation-reasoning-effort

If a model and effort combination is not supported by the provider, ReviewPhin passes it through and surfaces the provider/CLI error unchanged rather than hiding it.

Model availability

GPT-5.6 models are shown as examples. Availability depends on the account or organization entitlement of the token or key backing the profile — not every account can use every model or effort.

Save-time model validation

When you create or update a profile, ReviewPhin resolves the complete result first. It then checks every explicit native Copilot model against the catalog for that profile’s credentials before changing storage. This includes stored model values retained during a partial update.

If an explicit model is missing, or the catalog cannot be loaded, the command fails and leaves both the profile and the current default unchanged. An unset native review model is valid: null tells Copilot CLI to choose its default and does not require a catalog lookup. An unset text-generation model inherits the review model, so the explicit review model is checked once.

Custom providers do not share one reliable model-discovery endpoint or metadata format. ReviewPhin therefore cannot verify their model IDs. Use --ignore-missing-model when you have verified a custom provider separately and want to save it anyway, as shown in the examples above. The flag still attempts validation where discovery is supported and prints a warning explaining whether a model was missing or availability could not be verified.

Validation happens only when model-profile add writes a profile. Selecting a stored profile with model-profile set-default, assigning it to a tenant, or starting a review does not query the catalog again. If availability changes later, the provider’s existing session error and retry behavior applies.

Updating profiles

Re-run model-profile add with the same --name to update fields. Nullable fields can be cleared:

reviewphin model-profile add --name my-profile --clear-base-url
reviewphin model-profile add --name my-profile --clear-auth-token

--clear-base-url also clears the stored provider type and wire API. You cannot set new values for provider type or wire API in the same command that clears the base URL.

Updates also validate the effective stored model values. If you deliberately need to retain an unavailable or unverifiable model, include --ignore-missing-model on that update.

Full flags for model-profile commands are in the CLI reference.