fix(agentcore): parse A2A JSON-RPC responses in AgentCore provider

[michelligabriele]

Relevant issues

Pre-Submission checklist

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

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• 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).

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

🐛 Bug Fix

Changes

Problem: _parse_json_response() in the AgentCore provider returns empty content for A2A agents. Strategy 1 matches A2A responses (because result is a dict), but _extract_content_from_message() looks for result["content"] which doesn't exist in A2A format — A2A agents return result["message"]["parts"] instead. This means message.get("content", []) returns [], producing an empty string.

Fix: Added Strategy 0 before the existing strategies that detects A2A JSON-RPC responses by checking for the "jsonrpc" key, then delegates to the existing extract_text_from_a2a_response() utility from litellm/llms/a2a/common_utils.py. This utility already handles 5 A2A response shapes (direct message, nested message, task with artifacts, task status, streaming artifact-update). If A2A extraction returns empty, it falls through to the existing strategies — zero regression risk.

Files modified:

• litellm/llms/bedrock/chat/agentcore/transformation.py — added import + Strategy 0 in _parse_json_response()
• tests/test_litellm/llms/bedrock/chat/agentcore/test_agentcore_transformation.py — 4 new tests

Tests added (all in TestAgentCoreJsonResponseParsing):

• test_parse_json_a2a_jsonrpc_nested_message — result.message.parts[] format (customer's exact case)
• test_parse_json_a2a_jsonrpc_direct_parts — result.parts[] format
• test_parse_json_a2a_jsonrpc_multi_parts — multiple text parts concatenated
• test_parse_json_a2a_jsonrpc_empty_falls_through — empty A2A result falls through to Strategy 3

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 2, 2026 1:25pm

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing michelligabriele:fix/agentcore-a2a-response-parsing (ca19d7b) with main (d1df4e8)}

[greptile-apps]

Greptile Summary

This PR fixes empty responses from AgentCore when the underlying agent is an A2A (Agent-to-Agent Protocol) agent. The root cause was that A2A agents return a {"jsonrpc": "2.0", "result": {"message": {"parts": [...]}}} structure, which matched Strategy 1's result-is-dict check, but then _extract_content_from_message() looked for result["content"] — a key that doesn't exist in A2A format — yielding an empty string.

Fix summary:

• Adds "Strategy 0" in _parse_json_response(): detects a "jsonrpc" key and delegates immediately to the already-existing extract_text_from_a2a_response() utility (which handles five A2A response shapes).
• If extraction returns empty, execution falls through to the existing strategies, preserving all pre-existing behavior.
• Adds the import of extract_text_from_a2a_response from litellm/llms/a2a/common_utils.py (no new utility code added).

Test coverage:

• 4 new mock-only unit tests covering nested-message format, direct-parts format, multi-part concatenation, and the empty-result fallthrough path.

Confidence Score: 5/5

• Safe to merge — the fix is additive and isolated, with correct fallthrough semantics preserving all pre-existing behavior.
• The only finding is a minor import ordering style issue (P2). The logic itself is correct: Strategy 0 fires only on jsonrpc-keyed responses and cleanly falls through when extraction yields empty, leaving Strategies 1-4 fully intact. Tests are pure mocks and cover the four distinct structural variants plus the fallthrough edge case.
• No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/bedrock/chat/agentcore/transformation.py	Adds Strategy 0 for A2A JSON-RPC detection in _parse_json_response(); delegates to existing extract_text_from_a2a_response() utility; falls through cleanly to existing strategies if extraction yields empty. Minor import ordering issue (new import placed mid-block between two bedrock imports rather than alphabetically before them).
tests/test_litellm/llms/bedrock/chat/agentcore/test_agentcore_transformation.py	Adds 4 pure-mock unit tests covering the new A2A Strategy 0: nested message, direct parts, multi-part concatenation, and empty-result fallthrough. No network calls, all assertions are correct.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[_parse_json_response called] --> B{Is dict?}
    B -- No --> Z[Strategy 4: return raw JSON string]
    B -- Yes --> C{jsonrpc key present? Strategy 0 NEW}
    C -- Yes --> D[extract_text_from_a2a_response]
    D --> E{content non-empty?}
    E -- Yes --> F[Return A2A content]
    E -- No --> G{result is dict? Strategy 1}
    C -- No --> G
    G -- Yes --> H[_extract_content_from_message - return with final_message]
    G -- No --> I{response is list? Strategy 2}
    I -- Yes --> J[_extract_content_from_message - return content]
    I -- No --> K{result or response is string? Strategy 3}
    K -- Yes --> L[Return string value]
    K -- No --> Z

Loading

_{Reviews (1): Last reviewed commit: "fix(agentcore): parse A2A JSON-RPC respo..." | Re-trigger Greptile}

[greptile-apps]

Import placed out of alphabetical order

The new a2a import is inserted between two litellm.llms.bedrock.* imports. Following the existing file's alphabetical grouping (a2a < base_llm < bedrock), it should appear before the base_llm import.

Suggested change

	from litellm.llms.a2a.common_utils import extract_text_from_a2a_response
	from litellm.llms.a2a.common_utils import extract_text_from_a2a_response
	from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
	from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!