BYOK anthropic provider ignores model catalog for max_tokens, defaults to 8192

[TruaShamu]

Summary

When using type: "anthropic" with a BYOK provider config, the SDK sends max_tokens: 8192 (DEFAULT_MAX_OUTPUT_TOKENS) regardless of the model being used. The SDK's built-in model catalog (which correctly lists max_output_tokens: 32000 for claude-opus-4.5, 65536 for claude-sonnet-4.6, etc.) is not consulted for BYOK providers.

This causes 87% of agentic tool-use sessions to terminate prematurely because Claude's responses are truncated before the tool_use content block is emitted.

Note: This may be a regression or incomplete fix from #955 and #931, which were closed as completed on April 6 — but the underlying issue persists on SDK 0.2.0. The specific gap is that getEffectiveMaxTokens() does not look up the model catalog for BYOK providers.

Reproduction

const session = await client.createSession({
  model: "claude-opus-4.5",
  provider: {
    type: "anthropic",
    baseUrl: "https://my-proxy/v1",
    apiKey: "my-key",
  },
  tools: [myTool],
});

await session.sendAndWait({ prompt: "Use the tool to create a file with substantial content" });
// Session ends after 1 turn — tool_use block truncated at 8192 output tokens

Expected behavior

The SDK should resolve max_output_tokens from its built-in model catalog when the model name matches a known entry (e.g., claude-opus-4.5 → 32000). The DEFAULT_MAX_OUTPUT_TOKENS fallback should only apply when the model is genuinely unknown.

The catalog already exists in the SDK:

// From app.js
["claude-opus-4.5", { max_prompt_tokens: 168000, max_context_window_tokens: 200000, max_output_tokens: 32000 }]
["claude-opus-4.6", { max_prompt_tokens: 168000, max_context_window_tokens: 200000, max_output_tokens: 32000 }]
["claude-sonnet-4.6", { max_prompt_tokens: 168000, max_context_window_tokens: 200000, max_output_tokens: 65536 }]

Actual behavior

getEffectiveMaxTokens() returns this.options?.maxOutputTokens ?? DEFAULT_MAX_OUTPUT_TOKENS (8192). For BYOK providers, this.options.maxOutputTokens is never populated from the model catalog.

Impact

When Claude generates a response exceeding 8192 tokens (common for tool-calling agents that write substantial content):

1. Response is truncated — stop_reason becomes "max_tokens" instead of "tool_use"
2. The tool_use content block is incomplete or missing entirely
3. finish_reason maps to "length" → agent loop stops instead of continuing
4. Session emits session.idle prematurely — the agent never executes the tool

This is a silent failure — the session completes without error, but the agent did not accomplish its task.

Evidence

In a 54-trial evaluation run with claude-opus-4.5 via a BYOK proxy:

• 28 responses hit exactly 8192 output tokens with 0 tool calls (truncated)
• 19 responses were small enough to fit (tool calls for metadata lookups, not content creation)
• 7 responses happened to be concise enough to fit the create tool call within 8192 tokens

Suggested fix

In getEffectiveMaxTokens(), look up the model name in the built-in catalog before falling back to DEFAULT_MAX_OUTPUT_TOKENS:

getEffectiveMaxTokens() {
  const catalogMax = this.getModelCatalogMaxOutput(this.model); // new lookup
  const e = this.options?.maxOutputTokens ?? catalogMax ?? DEFAULT_MAX_OUTPUT_TOKENS;
  return this.isThinkingEnabled() ? Math.max(e, MIN_THINKING_BUDGET + 1) : e;
}

Workaround

Proxy servers can enforce a max_tokens floor by inspecting and overriding the value in the request body before forwarding to the upstream provider.

Environment

• @github/copilot-sdk: 0.2.0
• Models affected: Any Claude model via BYOK type: "anthropic" where the model's actual max output exceeds 8192