Windows: stale daemon can preserve missing embeddings env; search output can also crash on GBK consoles

Title: Windows: slim uvx cocoindex-code@latest install starts a stale daemon without embeddings; search output can also crash on GBK consoles

Repo: cocoindex-io/cocoindex-code
Version: cocoindex-code 0.2.33
OS: Windows

Summary

There are two separate problems:

1. A slim uvx cocoindex-code@latest install can start a daemon that lacks sentence_transformers, and later runs keep reconnecting to that stale daemon even after the dependency is added.
2. ccc search can crash on Windows consoles that use GBK when a result contains Unicode characters not encodable in that code page.

What I ran

First attempt:

uvx --prerelease=explicit --with cocoindex>=1.0.0a16 cocoindex-code@latest index

Observed:

Indexing failed: No module named 'sentence_transformers'

Then I retried with the missing dependency:

uvx --prerelease=explicit --with cocoindex>=1.0.0a16 --with sentence-transformers cocoindex-code@latest index

Observed:

Indexing failed: No module named 'sentence_transformers'

The dependency was actually present in the environment. These checks succeeded:

uv run --with sentence-transformers python -c "import sentence_transformers; print(sentence_transformers.__file__)"
uv run --with cocoindex-code --with sentence-transformers python -c "from cocoindex.ops.sentence_transformers import SentenceTransformerEmbedder; print(SentenceTransformerEmbedder)"

Root cause I found locally

• The first failed run had already started a daemon.
• Later runs kept reconnecting to that stale daemon.
• Stopping the daemon and rerunning in a correct environment fixed indexing.

Commands:

uv run --with cocoindex-code --with sentence-transformers python -m cocoindex_code.cli daemon stop
uv run --with cocoindex-code --with sentence-transformers python -m cocoindex_code.cli index

After stopping the stale daemon, indexing succeeded.

Recommended fix

• If embeddings are required, fail earlier with a direct install hint that points users to the supported cocoindex-code[full] install path.
• Consider including dependency/environment fingerprinting in the daemon handshake so a client does not reconnect to an incompatible stale daemon.
• If a required module is missing in the daemon, surface that as a daemon health error and suggest ccc daemon restart after install.

Second issue: Windows console encoding crash

After indexing succeeded, search returned results but then crashed while printing:

uv run --with cocoindex-code --with sentence-transformers python -m cocoindex_code.cli search obsidian --limit 3

Observed traceback tail:

UnicodeEncodeError: 'gbk' codec can't encode character '\u2194' in position 523: illegal multibyte sequence

Recommended fix

• Emit UTF-8 safely on Windows, or
• catch UnicodeEncodeError in result printing and fall back to replacement characters.

Notes

• Installing the supported full tool worked cleanly:

uv tool install --upgrade "cocoindex-code[full]"
ccc index
ccc mcp

• Once the full tool was installed, indexing worked on local projects and a Markdown notes directory.

Start with the daemon entry point used by python -m cocoindex_code.cli daemon stop and the index command, then inspect how clients reconnect after an environment change. Reproduce the stale-daemon case with and without sentence-transformers, and run search using a GBK console with Unicode results. Done means incompatible daemons report an actionable error or restart path, and search output no longer crashes.