ComposioHQ/composio @composio/mastra@0.9.3

on GitHub

Patch Changes

• b69cef1: Tolerate dangling $ref pointers in tool schemas the Composio API ships without a matching $defs entry. Some toolkits (e.g. GMAIL_FETCH_EMAILS) emit outputParameters with "$ref": "#/$defs/FetchEmailsResponse" while never declaring a top-level $defs block. After the strict resolver shipped with the previous Mastra fix, this caused composio.tools.get(...) to throw JsonSchemaRefResolutionError upfront, making every Gmail / Slack / Google-Calendar tool unusable through MastraProvider. The SDK now degrades the unresolvable branch to a permissive object schema and surfaces a single observability warning per (toolSlug, ref) pair instead of crashing.

  • dereferenceJsonSchema accepts a new optional second argument { onUnresolved?: 'throw' | 'sentinel'; onReplace?: (ref, reason) => void }. Default behavior is unchanged ('throw') — first-party / custom-tool schemas with a typo'd $ref still surface as a hard error. Pass 'sentinel' to replace unresolved branches with the cycle-break sentinel ({ type: 'object', additionalProperties: true }) that the resolver already uses for $ref cycles. The replaced sentinel carries a default description hint so LLMs consuming the wrapped tool's schema get an in-band signal that the branch is opaque; a caller-provided description sibling overrides the default (Draft 2020-12 sibling-keyword merge). Safety caps (MAX_REF_CHAIN_DEPTH, MAX_NODE_DEPTH) keep throwing in both modes. New UnresolvedRefStrategy, UnresolvedRefReason, and DereferenceJsonSchemaOptions type exports.
  • MastraProvider.wrapTool opts both inputParameters and outputParameters into 'sentinel' mode and emits one logger.warn per (toolSlug, ref) pair via the provider-scoped dedup Set. User-controlled segments in the warning (tool.slug, toolkit.slug, ref) are JSON.stringifyd to neutralize embedded newlines / ANSI escapes / control bytes that could otherwise forge log lines (CWE-117). A matching one-shot telemetry event (composio.mastra.wrapTool.danglingRef) fires next to the warn so the Composio team has aggregate visibility into which toolkits are affected; the event respects COMPOSIO_DISABLE_TELEMETRY=true.
  • The telemetry instance from @composio/core is now publicly re-exported alongside logger, so providers can emit aggregate signals without reaching into the package's internals.
  • Resolvable $defs / definitions continue to be inlined exactly as before — no regression in the type-info preservation contract introduced by the previous Mastra fix.

• ce4b213: fix(providers): normalize string tool-call arguments across all providers

  Models occasionally emit tool-call arguments as a JSON string instead of an
  object (most visibly with COMPOSIO_MULTI_EXECUTE_TOOL on the Vercel AI SDK),
  which broke downstream validation with errors like
  tool_use.input: Input should be a valid dictionary.

  @composio/core now exposes a single normalizeToolArguments helper, and every
  provider routes model-supplied arguments through it. Object payloads pass
  through unchanged, JSON strings are parsed, empty/null payloads become {},
  and anything that cannot resolve to an object throws a typed
  ComposioInvalidToolArgumentsError instead of a raw SyntaxError or a silently
  forwarded malformed string. This replaces the inconsistent per-provider guards
  that previously existed only in vercel, cloudflare and openai-agents.