[billxc]

Summary

• Normalize equivalent max_tokens and max_completion_tokens only at the Copilot Chat target boundary, keeping the combined cap and removing the redundant legacy field.
• Leave single, missing, and distinct caller limits unchanged so upstream validation still owns invalid or unrepresentable combinations.
• Preserve the original payload object and upstream result.

Reproduced problem

Upstream Menci#673's input IR Chat writer emits both cap spellings for compatibility. Copilot rejects an equal pair such as max_tokens=32000 and max_completion_tokens=32000 with:

max_tokens and max_completion_tokens cannot both be set

All three Gemini Chat models currently advertised by the tested Copilot account reproduced the same HTTP 400 with equal dual caps and accepted a modern-only control. A real original Anthropic Messages request also reproduced the rejection.

When both bounds are equal, the combined reasoning-plus-visible cap already implies the same visible cap. Removing the legacy constraint therefore preserves the effective budget without guessing how to reconcile distinct values.

Why the fix is Copilot-scoped

The initial candidate removed the legacy spelling from generic translation and added an Ollama adapter. Adversarial review and a deterministic reproduction showed that this would silently drop budgets on legacy-only custom endpoints. This revision instead follows Copilot's established provider-owned boundary pattern.

There are no changes to generic translation, Ollama, custom or Azure runtime behavior, configuration, flags, or migrations. A translated legacy-custom regression verifies that its existing legacy cap still reaches the non-Copilot target.

Pre-existing dual-cap rejection on strict non-Copilot targets is outside this narrowly scoped fix; this PR does not claim a universal token-cap dialect migration.

Verification

• 176 tests passed across six explicitly named provider/gateway/protocol files.
• Unit cases cover equal numeric, zero and null caps; either single spelling; distinct numeric and mixed-null caps; absent caps; input non-mutation; and result passthrough.
• A real provider-chain test verifies the new interceptor is registered and the serialized Copilot request contains only the combined cap.
• A real gateway translation test preserves the legacy cap for a custom Chat target.
• ESLint passed for the five changed files. TypeScript passed for Copilot, gateway and the Node runtime. Diff whitespace checks passed.
• Original request replay through an independent Node process, fresh migrated SQLite database and loopback listener:
  • Source request unchanged, HTTP 200, real tool output and completed SSE without an error event.
  • max_completion_tokens=32000 on actual outgoing traffic; max_tokens absent.
  • All six screenshots retained as WebP image_url inputs.
  • Actual input usage 104,370 tokens, versus the preceding original request's 1,048,204.
  • No runtime payload override, feature-flag workaround, retry or token exchange; returned tool calls were not executed.

Scope

One commit, five files, no playground changes. This is a downstream preview PR against synchronized main; no upstream PR or merge has been performed.