Repository navigation
SDK with BYOK providers: short model names aren't matched to the provider's models (custom agents fail; a short session model fails with a misleading "No GitHub OAuth token" error) #5127
Description
Describe the bug
When an SDK session configures BYOK providers and models, its models are only known as
provider-qualified ids such as copilot/gpt-5.4-mini. Any model given by its short name is not
matched to them, even when exactly one configured provider serves that model:
-
Custom agents. An agent whose
modelis a short name fails when the main agent starts it.
The tool call returns
Model 'gpt-5.4-mini' is not available. Available models: copilot/gpt-5.6-sol, copilot/gpt-5.6-luna, copilot/gpt-5.4-mini, copilot/gpt-5.6-sol, copilot/gpt-5.6-luna, copilot/gpt-5.4-mini.
The list itself names the model, and also lists every model twice. This happens for agents
passed ascustomAgentsand for.github/agents/*.agent.mdfiles loaded with
enableConfigDiscovery: true. -
Session model.
createSessionaccepts a shortmodel(session.startreports
selectedModel: "gpt-5.6-sol"), but the first turn fails with
Execution failed: InvalidArg, No GitHub OAuth token or Copilot HMAC key provided. The debug log
shows the turn quietly switched toclaude-sonnet-4, a model that isn't configured, on GitHub's
own inference:[WARNING] model resolution before the turn did not commit a new selection {"error":"No model available. Check policy enablement under GitHub Settings > Copilot"} [WARNING] turn model is in neither the picker nor the unfiltered catalog, ...; falling back to default capabilities and the chat-completions wire {"model":"claude-sonnet-4",...}The error points at authentication, not at the model name. When the user is also signed in to
GitHub, the request would presumably go to GitHub's own inference instead of the configured
provider, without any error.
Why it matters. Agent files are shared between Copilot CLI and SDK sessions. In CLI BYOK mode
the same .agent.md must use the short name, and a file can't know which provider name an SDK host
chose. With discovery on, the main agent sometimes recovers by retrying with the qualified name. That
costs a failed tool call, depends on the model noticing, and in our runs a retry sometimes picked a
different model instead. GitHub Agentic Workflows currently rewrites every agent's model to a
qualified name before creating the session; we'd rather not keep that workaround.
This short-name matching is probably also the SDK-mode cause of #5126 (built-in agents' default
models never match).
Affected version
@github/copilot-sdk 1.0.19 against copilot --headless from Copilot CLI 1.0.96. Also seen with SDK 1.0.16 and CLI 1.0.90.
Steps to reproduce the behavior
Start copilot --headless --port <port> with no GitHub sign-in, and use one BYOK provider:
{
"providers": [{ "name": "copilot", "type": "openai", "wireApi": "responses",
"baseUrl": "https://api.githubcopilot.com", "bearerToken": "<token>",
"headers": { "Copilot-Integration-Id": "copilot-developer-cli" } }],
"models": [{ "id": "gpt-5.6-sol", "provider": "copilot" },
{ "id": "gpt-5.6-luna", "provider": "copilot" },
{ "id": "gpt-5.4-mini", "provider": "copilot" }]
}All models are on the same wire API, so #5103 doesn't apply.
- Custom agent with a short model: session
model: "copilot/gpt-5.6-sol", plus
customAgents: [{ "name": "quick-checker", "description": "Finds where a function is defined", "prompt": "...", "model": "gpt-5.4-mini" }].
Prompt:Use the task tool with agent_type "quick-checker" to find which file defines add_numbers.
Observed: the first task call fails withModel 'gpt-5.4-mini' is not available. Available models: ...copilot/gpt-5.4-mini.... - Same agent from a file: no
customAgents,enableConfigDiscovery: true, and
.github/agents/quick-checker.agent.mdwithmodel: gpt-5.4-miniin its front matter. Observed:
the same error. - Control: the same agent with
"model": "copilot/gpt-5.4-mini"runs on mini
(modelSelectionSource: "agent_definition_default"). - Short session model:
model: "gpt-5.6-sol"and promptReply with the single word OK.
Observed:session.errorExecution failed: InvalidArg, No GitHub OAuth token or Copilot HMAC key provided,
with theclaude-sonnet-4warnings above in the debug log.
Expected behavior
- A short model name resolves to the configured provider model with that id when exactly one
provider serves it:gpt-5.4-mini→copilot/gpt-5.4-mini. This applies to the session model,
custom agents fromcustomAgentsor discovered files, and built-in agent defaults (BYOK: built-in agents (explore, task, research) never use their default models and silently run on the session model #5126). - If several providers serve the id, pick deterministically (for example, the provider whose wire
API the model supports), or report the ambiguity and list the qualified choices. - A session model that can't be resolved fails at
createSessionwith an error that names the
model and the available ids. It isn't silently replaced by another model, and the error doesn't
point at authentication.
Additional context
- Related: BYOK: built-in agents (explore, task, research) never use their default models and silently run on the session model #5126 (built-in agents ignore their default models under BYOK) and BYOK: sub-agents always use the session's wire API, so a sub-agent on a model from the other family fails with a 400 #5103 (one wire
API per BYOK session in CLI mode). - Environment: Linux ARM64 (Ubuntu 24.04), Node 22. Also seen on GitHub Actions Linux x64 runners
through GitHub Agentic Workflows' AWF API proxy as the BYOK provider.
Activity
rishidevcli54-mcp commentedon Oct 13, 2026