Tokentoll

Patch release driven by dogfooding v0.8.0 against five popular TS LLM repos: vercel/ai-chatbot, langchain-ai/langchainjs, anthropics/anthropic-sdk-typescript, openai/openai-node, mckaywrigley/chatbot-ui. Three real bugs surfaced, each with a minimal regression test pinning the fix.

Fixes

TypeScript wrapper expressions are now unwrapped

Code like:

client.chat.completions.create({
  model: chatSettings.model as ChatCompletionCreateParamsBase[\"model\"],
  max_tokens: 1024 as number,
});

previously failed to resolve because the resolver bailed on as_expression. Now as, satisfies, !, parens, and the legacy <T>value syntax are all transparent.

Member chains rooted at this and super

Class-based SDKs use this.client.x.y.create(...) everywhere. The detector required an identifier base segment, so every such call slipped through. Detection counts after the fix:

| Repo | Before | After |
|---|---|---|
| anthropic-sdk-typescript | 2 | 7 |
| langchainjs | 12 | 24 |
| openai-node | 23 | 25 |

Vercel AI SDK detector no longer matches method calls

generateText, streamText, embed, embedMany are imported from @ai-sdk/* and called bare. Accepting member-expression callees (this.client.embed(req)) produced false positives for libraries like langchain-cohere that expose methods with the same names. Removed 7 false positives from the langchainjs scan.

Tests

141 passing (was 138). Three new fixture files plus three regression tests, each derived from the failing real-world pattern:

• ts_wrappers.ts covers as/satisfies/!/parens/<T>
• this_chain.ts covers this.client.x.y.create
• cohere_embed_not_vercel.ts covers the Vercel false-positive case

Upgrade

pip install --upgrade tokentoll

- uses: Jwrede/[email protected]

No schema or CLI changes. Existing .tokentoll.yml files keep working.