Copilot NES: add ExP toggle to omit max_tokens

[Shengjie Ma (bluerose73)]

Summary

Adds the ExP-capable boolean setting github.copilot.chat.advanced.inlineEdits.xtabProvider.useMaxTokens (default: true).

When usePrediction=false, setting useMaxTokens=false omits max_tokens, allowing reasoning-model deployments that reject the legacy parameter to choose their server-side output budget. The default preserves the endpoint limit (currently 4096). Prediction-enabled requests are unchanged.

ExP treatment variable: copilotchat.config.chat.advanced.inlineEdits.xtabProvider.useMaxTokens.

Tests cover the toggle combinations, unset/default behavior, ExP assignment and user override precedence, and actual serialized request omission. The new NES cases use an isolated endpoint provider to avoid unrelated model-metadata cache access. No nullable ExP plumbing or numeric-limit setting is included.

Current NES model requests also omit max_tokens

In current NES model requests, max_tokens is not set, because usePrediction=true

Therefore, I do not expect github.copilot.chat.advanced.inlineEdits.xtabProvider.useMaxTokens = false to worsen latency or oversized suggestions.

Validation

• 37 focused tests passed (configuration, shared fetcher/retry suite, and all five new NES toggle cases).
• Copilot extension TypeScript check passed.
• Changed-file ESLint and staged whitespace checks passed.
• A prior broader NES test run reported three failures in existing cases and LFS-backed SQLite cache errors. Baseline reproduction has not been established; this is not a claim that the full suite passes.

[vs-code-engineering]

📬 CODENOTIFY

The following users are being notified based on files changed in this PR:

Ulugbek Abdullaev (@ulugbekna)

Matched files:

• extensions/copilot/src/extension/xtab/node/xtabProvider.ts
• extensions/copilot/src/extension/xtab/test/node/xtabProvider.spec.ts

[Copilot]

🟢 Approval recommended

The focused implementation preserves existing defaults and is covered across configuration, request construction, and serialization.

0 open findings

What changed in this PR

Adds an experiment-controlled NES option to omit max_tokens when prediction is disabled, enabling reasoning-model deployments to use server-side output budgets.

Changes:

• Adds the useMaxTokens configuration with user-over-ExP precedence.
• Conditionally suppresses max_tokens while preserving prediction-enabled behavior.
• Adds configuration, request-construction, and serialization tests.

File	Description
configurationService.ts	Defines the new experiment-based setting.
configurationServiceImpl.spec.ts	Tests defaults, ExP assignment, and user overrides.
xtabProvider.ts	Applies omission to prediction-disabled requests.
xtabProvider.spec.ts	Covers toggle and prediction combinations.
chatMLFetcherRetry.spec.ts	Verifies serialized request behavior.

🧠 Review effort: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.