The agent that grows with you
Python aiai-agentai-agentsanthropicchatgptclaudeclaude-codecodexhermeshermes-agentllmnous-researchopenai
fix(bedrock): strip prior reasoning for the gpt-6 family on Converse
liuhao1024e49938daf7a6
diff --git a/agent/bedrock_adapter.py b/agent/bedrock_adapter.pyindex f4142cb48d1..2a266fd706c 100644--- a/agent/bedrock_adapter.py+++ b/agent/bedrock_adapter.py@@ -864,14 +864,16 @@ def _assistant_blocks(msg: dict, content, claude: bool = True, own: bool = False# the signature and all previous messages"; Bedrock ignores older turns server-side for free).# strip — DeepSeek-R1 (AWS's multi-turn sample removes reasoningContent) and Kimi K3 (Converse raises# InternalServerException "when reasoning content from earlier turns is included"): prior turns-# lose reasoning, the in-flight loop replays the model's own blocks as captured.+# lose reasoning, the in-flight loop replays the model's own blocks as captured. Also OpenAI's+# reasoning GPTs (gpt-6 family): they reject readable reasoningText in assistant history too+# ("This model doesn't support the reasoningContent.reasoningText.text field").# unsigned — every other model: prior turns carry readable reasoningText only, since their signatures and# redacted blobs may be another model's (a /model switch from Claude) and non-Claude models# reject them ("This model doesn't support the reasoningContent.reasoningText.signature field");# the in-flight loop was produced by the model now answering and replays as captured.# A turn stamped by another model (CONVERSE_ISSUER_TYPE) never gets the in-flight verbatim replay: a fallback or# switch mid-loop hands the loop to a model that cannot open the previous one's sealed blobs (_reasoning_filter).-_CONVERSE_REPLAY_STRIP_PATTERNS = ("deepseek.r1", "deepseek-r1", "moonshotai.kimi-k3")+_CONVERSE_REPLAY_STRIP_PATTERNS = ("deepseek.r1", "deepseek-r1", "moonshotai.kimi-k3", "openai.gpt-6")# reasoning_details record naming the model that minted a turn's Converse reasoning: its signatures are thatdiff --git a/tests/agent/test_bedrock_converse_reasoning_replay.py b/tests/agent/test_bedrock_converse_reasoning_replay.pyindex f67630ad51d..138d574e13b 100644--- a/tests/agent/test_bedrock_converse_reasoning_replay.py+++ b/tests/agent/test_bedrock_converse_reasoning_replay.py@@ -4,7 +4,8 @@ Converse is the bearer-token Claude path and the sticky fallback after a stream-error, so a Claude turn captured in Anthropic shape (``anthropic_content_blocks``) must keep its signedthinking ahead of its toolUse. AWS docs: Claude needs the signature and all previous messages; DeepSeek-R1'ssample removes prior reasoning; Kimi K3 raises InternalServerException on prior-turn reasoning; non-Claude-models reject the ``reasoningText.signature`` field.+models reject the ``reasoningText.signature`` field; OpenAI's reasoning GPTs (gpt-6) reject ``reasoningText``+in assistant history outright."""from __future__ import annotations@@ -65,6 +66,9 @@ def _reasoning(blocks):("global.moonshotai.kimi-k3", [], [{"text": "own sig_loop"}]),# gpt-oss rejects reasoningText.signature: prior turns go back as readable text only.("openai.gpt-oss-120b-1:0", [{"text": "own sig_prior"}], [{"text": "own sig_loop"}]),+ # Reasoning GPTs reject readable reasoningText in assistant history outright (#140177): prior+ # turns lose their reasoning; the in-flight loop still replays the model's own blocks.+ ("global.openai.gpt-6.1-sol", [], [{"text": "own sig_loop"}]),],)def test_converse_replays_each_models_reasoning_contract(model, prior, loop):