fix(a2a): populate response usage in a2a chat transformation

[michelligabriele]

Relevant issues

Linear ticket

Resolves LIT-3291

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• 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

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Called the a2a/ provider through /chat/completions on a local proxy configured with per-token pricing for the model:

curl -s -X POST http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model":"echo-a2a-chat","messages":[{"role":"user","content":"Say hello in one word"}]}'

Before fix:

{"choices": [...], "usage": {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}}

After fix:

{"choices": [...], "usage": {"prompt_tokens": 12, "completion_tokens": 9, "total_tokens": 21}}

usage is no longer 0/0/0, and spend logs now show real cost computed from the configured input_cost_per_token / output_cost_per_token against this usage (previously spend was $0.00 because cost is priced off a zero token count).

Type

🐛 Bug Fix

Changes

The a2a/ chat-completions provider (litellm/llms/a2a/chat/transformation.py) never set usage on the transformed response, so callers of model=a2a/<agent> always got usage: {prompt_tokens: 0, completion_tokens: 0, total_tokens: 0} directly in the API response, and per-token cost always computed to $0.00 since it prices off that zero usage.

A2A agents don't return token usage in their protocol responses, so this estimates it with litellm.utils.token_counter off the request messages and response text, following the same pattern already used by the langgraph and langflow agent-bridge providers for the same "agent doesn't report usage" situation.

[greptile-apps]

Greptile Summary

This PR fixes the A2A chat provider always returning usage: {prompt_tokens: 0, completion_tokens: 0, total_tokens: 0}, which caused per-token cost calculations to always produce $0.00. Since A2A agents don't include token counts in their protocol responses, the fix estimates usage via litellm.utils.token_counter (using the gpt-3.5-turbo tokenizer as a proxy), the same approach already used by the langgraph, vertex_ai/agent_engine, and azure_ai/agents providers.

• litellm/llms/a2a/chat/transformation.py: Adds a best-effort try/except block after the response is built that estimates prompt_tokens from the input messages and completion_tokens from the response text, then sets a Usage object on model_response.
• tests/test_litellm/llms/a2a/chat/test_a2a_chat_transformation.py: New mock-only regression test that asserts all three token counts are non-zero and that total_tokens equals the sum of the other two.

Confidence Score: 5/5

Safe to merge — the change is additive, wrapped in a best-effort try/except so a tokenizer failure cannot break the response path, and it exactly mirrors the pattern already in production for three other agent bridge providers.

The fix is a small, well-scoped addition that follows a pattern proven across three other providers in the same codebase. The estimation runs inside a try/except so any edge-case failure degrades gracefully to the existing 0/0/0 behaviour rather than raising. The new test is mock-only and validates the core regression directly.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/a2a/chat/transformation.py	Adds best-effort token usage estimation via token_counter, following the identical pattern used by langgraph, vertex_ai/agent_engine, and azure_ai/agents providers. Usage and import are correct.
tests/test_litellm/llms/a2a/chat/test_a2a_chat_transformation.py	New regression test using mocks only; correctly verifies that prompt_tokens, completion_tokens, and total_tokens are all non-zero and self-consistent after the fix.

_{Reviews (2): Last reviewed commit: "test(a2a): rename to avoid module basena..." | Re-trigger Greptile}

[greptile-apps]

The bare except Exception: pass silently swallows any token-counting failure with no log output, making it impossible to diagnose why usage stays at zero for a user. Both langgraph/chat/transformation.py and vertex_ai/agent_engine/transformation.py use the same pattern but emit a verbose_logger.warning so the failure is at least visible in debug logs.

Suggested change

	except Exception: # noqa: BLE001 - best-effort estimate; a tokenizer hiccup must not break the response
	pass
	except Exception as e: # noqa: BLE001 - best-effort estimate; a tokenizer hiccup must not break the response
	verbose_logger.warning(f"A2A: failed to estimate token usage: {e}")

[codecov]

Codecov Report

❌ Patch coverage is 75.00000% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/llms/a2a/chat/transformation.py	75.00%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[Sameerlite]

@greptileai

[Sameerlite]

LGTM