llamastash / llamastash / 26664094640

push

github

fix(init): reuse split_gguf shard helper; embed-shaped tool config

Two follow-ups from review on da624a4:

- Model-id derivation now routes through
  discovery::split_gguf::parse_shard_name instead of the inline
  strip_shard_suffix I'd duplicated. Same helper the discovery
  scanner uses to collapse shard sets — the id we hand to tools
  now matches the catalog row's name by construction.

- PatchContext gains is_embed; Continue.dev's roles flip from
  [chat, edit] to [embed] and pi.dev's api flips from
  openai-completions to openai-embeddings when the model id
  contains "embed" (nomic-embed-text, snowflake-arctic-embed,
  bge-embed, gte-embed). Setting chat/edit on an embedder gave
  confusing errors in Continue because the IDE tried to chat with
  an encoder-only model. Detection is a filename heuristic for v1;
  upgrading to gguf::metadata::ModeHint (BERT arch + pooling_type
  + name/tags scan) is a future swap.

61 of 73 new or added lines in 9 files covered. (83.56%)

41477 of 48125 relevant lines covered (86.19%)

1677.02 hits per line