@srijanaryachecks…fix/normalize-string-tool-input-providers → next6 files · +111 −2review: @jkomynoupdated 1w agoExecuteToolFn is typed (toolSlug, input: Record<string, unknown>) — it promises an object. Per #2406, models intermittently emit tool-call input as a JSON string. The llamaindex (input as Record<string, unknown>) and claude-agent-sdk (args passed straight through) providers forwarded model input unchecked, so a stringified payload reached execution as a raw string (Input should be a valid dictionary).
const normalized = typeof input === 'string' ? JSON.parse(input) : input; — mirroring the guard already shipped in vercel/cloudflare/openai-agents.openai-agents guard had no string-path test. Backfilled string→object regression tests across all three providers ('{"x":1}' → {x:1}; {x:1} → passthrough).langchain/anthropic/google to in-flight #3438; didn't touch mastra (#3400).Refs #2406.
Hey @CryogenicPlanet — extended the string→object tool-input normalization from vercel/cloudflare/openai-agents to the llamaindex and claude-agent-sdk providers (they forwarded stringified input unchecked → the #2406 path) and backfilled the regression tests across all three (openai-agents didn't have one). Would love your eyes when you have a sec — happy to extend it to the remaining providers too.
Good catch. Now it wraps the parse, so a malformed string falls through to executeTool unchanged and you get the typed "input must be a dictionary" error instead of a raw SyntaxError. Added a regression test for the malformed case in both providers.
Went graceful to match the outer try/catch that's already there. Can make it fail-fast with a descriptive throw if you'd rather.
@srijanarya is attempting to deploy a commit to the Composio Team on Vercel.
A member of the Team first needs to authorize it.