feat(github_copilot): route /v1/messages to Copilot native Anthropic endpoint

[mateo-berri]

Relevant issues

This is a copy of #28054 by @ririnto, rebased onto the current litellm_internal_staging so it can run through our CI. All credit for the implementation goes to @ririnto; this branch only resolves the merge conflict against the moved base and normalizes formatting to the current ruff

• Closes [Feature]: github_copilot — route /v1/messages to Copilot native Anthropic endpoint (restores Claude thinking) #28053
• Addresses [Bug]: GitHub Copilot provider does not support reasoning_effort from Claude Code #25666; restores Claude thinking end-to-end by avoiding the chat-completions conversion that strips it
• Refs [Bug]: content_block_start dropped during /v1/messages → GitHub Copilot streaming, triggering non-streaming fallback + output truncation #24765; uses raw native passthrough so streaming content_block_* events are preserved
• Refs Anthropic → OpenAI conversion drops reasoning_content, breaks multi-turn with reasoning models #27946; bypasses the Anthropic -> OpenAI conversion that loses reasoning_content
• Alternative to Convert 'thinking' parameter to 'reasoning_effort' for Claude models #28031; that PR converts thinking -> reasoning_effort on the chat path, this PR keeps Anthropic semantics intact via Copilot's native /v1/messages

Linear ticket

n/a

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

Direct upstream verification from the original PR (independent of LiteLLM, to isolate Copilot's behaviour):

# Before: Copilot /chat/completions cannot emit thinking
POST .../chat/completions  reasoning_effort=high
  -> 400 invalid_reasoning_effort
POST .../chat/completions  thinking={type:enabled, budget_tokens:16000}
  -> 200 plain text only

# After: Copilot /v1/messages emits thinking blocks
POST .../v1/messages  thinking={type:enabled, budget_tokens:16000}
  -> 200 content=[{type:"thinking", thinking:"..."}, {type:"text", text:"..."}]
          usage.output_tokens=8413  (thinking_chars=9094)
# Streaming:
POST .../v1/messages  stream=true  thinking={type:enabled, budget_tokens:16000}
  -> SSE: message_start, content_block_start(type=thinking), content_block_delta(thinking_delta)...,
          content_block_stop, content_block_start(type=text), ...

This PR makes LiteLLM produce the "After" response when a client calls POST /v1/messages against a github_copilot Claude model. Unit tests cover config wiring, URL construction, header merging precedence, and auth error mapping

Type

New Feature

Changes

• New litellm/llms/github_copilot/messages/__init__.py
• New litellm/llms/github_copilot/messages/transformation.py
  • GithubCopilotAnthropicMessagesConfig(AnthropicMessagesConfig)
  • Overrides validate_anthropic_messages_environment (Copilot auth + integration headers + anthropic-version) and get_complete_url (/v1/messages)
  • Caller-supplied api_base is intentionally ignored so the Copilot bearer token can never be routed to a caller-controlled URL
• litellm/integrations/websearch_interception/handler.py; keep the web-search short-circuit for github_copilot even though it now has a BaseAnthropicMessagesConfig, since Copilot does not handle web_search tools natively
• litellm/utils.py; ProviderConfigManager._get_provider_anthropic_messages_config_cached() returns the new config when provider == GITHUB_COPILOT and "claude" in model_lower
• litellm/model_prices_and_context_window_backup.json; add /v1/messages to supported_endpoints for the 3 base Claude models this routing enables
• 5+ new unit tests under tests/test_litellm/llms/github_copilot/messages/

---

Generated by Claude Code

---

Note

Medium Risk
Changes request routing and auth for GitHub Copilot Claude on /v1/messages (including deliberate api_base pinning); behavior is well-tested but affects thinking, streaming, beta headers, and web-search interception for that provider.

Overview
Adds GithubCopilotAnthropicMessagesConfig so github_copilot Claude models on POST /v1/messages go to Copilot’s native /v1/messages path (auth, messages-proxy headers, URL building) instead of the chat-completions conversion that drops thinking and native streaming blocks.

ProviderConfigManager returns this config when the provider is GITHUB_COPILOT and the model name contains claude; non-Claude Copilot models stay off this path. Caller-supplied api_base is ignored in validation so the Copilot bearer token is only sent to the authenticated Copilot host.

Copilot is treated as an Anthropic passthrough for beta headers: should_filter_anthropic_beta_headers() is overridden to False so injected anthropic-beta values (e.g. context management, structured outputs) are not stripped. Model metadata now lists /v1/messages on the affected Claude Copilot entries.

Web-search interception no longer skips short-circuit merely because an Anthropic Messages config exists; it checks new handles_web_search_natively() (default True on the base config, False for Copilot) so Copilot still gets the synthetic web-search short-circuit even with the new messages config.

^{Reviewed by Cursor Bugbot for commit 54bff77. Bugbot is set up for automated code reviews on this repo. Configure here.}

[mateo-berri]

@greptileai

---

Generated by Claude Code

[greptile-apps]

Greptile Summary

Adds GithubCopilotAnthropicMessagesConfig so GitHub Copilot Claude models on POST /v1/messages are routed to Copilot's native /v1/messages path instead of the chat-completions conversion that strips thinking blocks and native streaming events. The implementation correctly pins the request host to the authenticated Copilot endpoint, normalises trailing slashes, disables the provider-scoped beta-header filter, and keeps the web-search interception short-circuit active via a new handles_web_search_natively() hook.

• New GithubCopilotAnthropicMessagesConfig (litellm/llms/github_copilot/messages/transformation.py): overrides URL construction, auth header injection, and beta-header filtering; caller-supplied api_base is intentionally discarded to prevent bearer-token leakage.
• ProviderConfigManager (litellm/utils.py): returns the new config for any github_copilot model whose name contains "claude"; non-Claude Copilot models are unaffected.
• BaseAnthropicMessagesConfig gains handles_web_search_natively() -> bool (default True), letting the web-search handler remain provider-agnostic while Copilot overrides it to False.

Confidence Score: 5/5

Safe to merge; the new routing path is isolated to GitHub Copilot Claude models on /v1/messages, existing providers are untouched, and the auth pinning prevents token leakage.

The implementation is well-contained and thoroughly tested with mock-only unit tests. The routing guard in utils.py is slightly broader than the three models whose metadata was updated, but this results in Copilot returning a model-not-supported error rather than any data corruption or security issue.

litellm/utils.py — the "claude" in model_lower routing guard routes all six Copilot Claude models through the new config, while only three received the /v1/messages metadata update.

Important Files Changed

Filename	Overview
litellm/llms/github_copilot/messages/transformation.py	New GithubCopilotAnthropicMessagesConfig class; correctly pins api_base to the authenticated Copilot host, normalises trailing slashes, and overrides beta-header filtering and web-search dispatch.
litellm/utils.py	Adds GITHUB_COPILOT branch to _get_provider_anthropic_messages_config_cached; routing guard ("claude" in model_lower) is broader than the three models that received /v1/messages in supported_endpoints metadata.
litellm/integrations/websearch_interception/handler.py	Short-circuit logic now checks handles_web_search_natively() instead of testing for config presence alone; correctly keeps Copilot on the short-circuit path even with its new messages config.
litellm/llms/base_llm/anthropic_messages/transformation.py	Adds handles_web_search_natively() hook to BaseAnthropicMessagesConfig with a safe default of True; no behaviour change for existing providers.
litellm/model_prices_and_context_window_backup.json	Adds /v1/messages to supported_endpoints for three Claude models (haiku-4.5, opus-4.5, sonnet-4.5); three other Copilot Claude models (opus-4.6-fast, opus-41, sonnet-4) are not updated but will still be routed through the new config.
tests/test_litellm/llms/github_copilot/messages/test_github_copilot_messages_transformation.py	Comprehensive mock-only unit tests covering URL construction, header injection, beta-header filtering, auth errors, and provider config dispatch; no real network calls.

_{Reviews (6): Last reviewed commit: "test(github_copilot): remove dead branch..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

Thanks for the review. Addressed all three findings in ada5fcc

Provider-specific check outside llms/: replaced the hardcoded LlmProviders.GITHUB_COPILOT check in the web-search interception handler with a handles_web_search_natively() method on BaseAnthropicMessagesConfig (default True), overridden to False in GithubCopilotAnthropicMessagesConfig. The handler is now provider-agnostic and a future provider in the same spot needs no carve-out here; this mirrors the existing should_filter_anthropic_beta_headers() pattern on the same base class

get_complete_url re-reading from disk: it now reuses the api_base it is handed, which in the request flow is the value already resolved by validate_anthropic_messages_environment (llm_http_handler.py reassigns api_base from that return before calling get_complete_url), so the redundant authenticator read and the mid-request inconsistency window are gone. The caller-supplied base is still discarded in validate, which is the security boundary; get_complete_url only falls back to a fresh resolution if no base was passed

Double-slash URL: both methods now rstrip("/") the base before appending /v1/messages, so a tenant host with a trailing slash no longer yields //v1/messages

Added regression tests for the new method contract (Copilot False, base config True) and for trailing-slash normalization in both get_complete_url and validate_anthropic_messages_environment

@greptileai

---

Generated by Claude Code

[mateo-berri]

bugbot run

---

Generated by Claude Code

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

Autofix Details

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Beta headers stripped for Copilot
  • Overrode should_filter_anthropic_beta_headers() to return False on GithubCopilotAnthropicMessagesConfig so injected anthropic-beta values (context_management, structured outputs, etc.) reach Copilot's native /v1/messages verbatim instead of being dropped by the provider-scoped filter, and added a regression test.

_{You can send follow-ups to the cloud agent here.}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
2 out of 3 committers have signed the CLA.

✅ ririnto
✅ mateo-berri
❌ cursoragent
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[mateo-berri]

Bugbot's autofix in 210bc94 is correct; Copilot's /v1/messages is a native Anthropic passthrough, so overriding should_filter_anthropic_beta_headers() to False is the right call (the default filter would strip every anthropic-beta value since github_copilot has no entry in anthropic_beta_headers_config.json), and it comes with a regression test. Re-requesting review on the latest commit

@greptileai

---

Generated by Claude Code

[mateo-berri]

bugbot run

---

Generated by Claude Code

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 210bc94. Configure here.}

[mateo-berri]

@greptileai

[mateo-berri]

Addressed the test quality nit in 54bff77; the beta-header regression test had an unreachable if should_filter_anthropic_beta_headers() branch (always False). Replaced it with a direct assertion that running the provider-scoped filter for github_copilot strips every beta value, which proves why the override is load-bearing and fails if someone flips it back on

@greptileai

---

Generated by Claude Code

[mateo-berri]

bugbot run

---

Generated by Claude Code

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 54bff77. Configure here.}

[mateo-berri]

On the note about the utils.py routing guard being broader than the three models with /v1/messages metadata: this is intentional and benign, so leaving it as-is

The "claude" in model_lower dispatch is the original design; it routes every Copilot Claude model to the native endpoint. Worth clarifying that supported_endpoints does not gate this path: in anthropic/experimental_pass_through/messages/handler.py, ProviderConfigManager.get_provider_anthropic_messages_config(...) is resolved first, and only if it returns None does the _deployment_passes_through_anthropic_messages(model_info) (supported_endpoints) fallback run. Since the Copilot config is non-None for any Claude model, the metadata on the three base models is informational; it does not produce a model-not-supported rejection for the others. Those models still route natively, and the worst case for a Claude SKU that Copilot does not serve on /v1/messages is a clear upstream error, no data corruption or security impact, consistent with the 5/5 assessment

I am deliberately not widening the metadata here: declaring /v1/messages for the other Claude SKUs would assert upstream support that was only verified on the base models, and it changes nothing at runtime since the dispatch already covers them. If the team later wants dispatch and advertised metadata to be strictly 1:1, the cleaner move is the one the original PR already calls out, switching the guard to supported_endpoints membership, which is a follow-up design change rather than something to slip into this copy

---

Generated by Claude Code