needle

The failure occurs because cactus-needle 3.0.5 requests engine wheel 3.0.2, but the Hugging Face repository only has 3.0.0 and 3.0.1. Immediate workaround: downgrade to 3.0.4, which pins engine 3.0.1:

pip install --force-reinstall cactus-needle==3.0.4
needle fetch

The code fix changes ENGINE_VERSIONS[3] in needle/agent/fetch.py from 3.0.2 back to the published 3.0.1 wheel, or installs a release containing that fix. Alternatively, the 3.0.2 wheels must be uploaded to the Cactus-Compute/needle3 Hugging Face repository.

Do not pass --layers to needle finetune; the option is only advertised by stale help text and is not supported by the finetune parser. Run fine-tuning without it, for example needle finetune data.jsonl --checkpoint <path>. Rung selection belongs to the build step: use needle build --layers N. Local fine-tuning runs at the full 20 layers. Update to a release containing the CLI help fix if the outdated hint still appears.

This happens because complete()/_complete() treats each call as a turn in the same conversation. Unrelated prompts accumulate in the history, causing confidence to drop and tool calls to alternate with function_calls: [].

For independent prompts, reset the conversation between calls:

agent = needle.Needle(tools=[get_weather, send_message], generation=2)

for prompt in prompts:
    resp = agent._complete(prompt)
    print(resp["function_calls"], resp["confidence"])
    agent.reset()

agent.close()

Alternatively, use the stateless mode if your installed version includes it:

agent = needle.Needle(
    tools=[get_weather, send_message],
    generation=2,
    stateless=True,
)

stateless=True automatically resets before each query. Creating and closing a fresh Needle instance for each independent request is also a valid workaround. Only reuse the instance without resetting when the prompts are actually part of one conversation.

The failure occurs because needle requests engine wheel cactus_needle-3.0.2-py3-none-manylinux2014_x86_64.whl, but the Cactus-Compute/needle3 Hugging Face repo only publishes 3.0.0 and 3.0.1. Upgrade cactus-needle to a release containing the fetch fix. If no fixed release is installed yet, pin the engine to 3.0.1 locally:

import pathlib, needle
p = pathlib.Path(needle.__file__).parent / "agent" / "fetch.py"
s = p.read_text()
p.write_text(s.replace("3.0.2", "3.0.1"))

Then restart the Python process/Colab runtime and initialize needle.Needle() again. Alternatively, use a package version that still references the available 3.0.1 wheel.

This is a regex bug in needle/__init__.py in _source_years. The day/month pattern uses (\d{1,4})(?![0-9A-Za-z]), so 19 in 19:30, 14 in 14:00, and 7 in 7 pm are incorrectly treated as years. Upgrade to a release containing the issue #153 fix. If you need to patch it immediately, change the trailing year lookahead to reject time suffixes, for example: (\d{1,4})(?!\s*:\d{2}|\s*[ap]\.?m\.?(?![A-Za-z])|[0-9A-Za-z]) with case-insensitive matching. This still allows short years such as "5th September 42". As a temporary workaround, include an explicit four-digit year, e.g. "5 June 2026 19:30". Verify with needle._source_years("Book a table on 5 June 19:30"), which should return set(), then run pytest -q -m "not slow".