Skip to content

cllama v0.7.0

Choose a tag to compare

@mostlydev mostlydev released this 10 Jun 01:41
· 8 commits to master since this release

Managed tool mediation

  • Pre-dispatch schema validation: model-emitted tool arguments are validated against the manifest inputSchema before the providing-service dispatch. Invalid calls are rejected in-round with a structured schema_validation tool result naming exact violations and JSON paths, including a wrong-nesting hint when a required property exists elsewhere in the payload. Validation fails open on schema constructs it does not understand and never blocks a call the provider would accept. New intervention type: managed_tool_schema_rejected. Kill switch: CLLAMA_TOOL_SCHEMA_VALIDATION=off.
  • Hash-free presented tool names: tools are presented to models without the hex hash suffix when the sanitized name is unique within the agent's manifest. Eliminates the suffix-dropping recovery class (and its per-call managed_tool_hashless_alias intervention noise) for typical pods. Resolution still accepts canonical, current presented, and legacy hashed forms; ambiguity remains a hard miss.

Feeds

  • Configurable fetch timeout: CLLAMA_FEED_FETCH_TIMEOUT_MS (default 3000, sanity range 100–120000) joins the existing feed budget knobs. Slow feed providers no longer silently drop context blocks at a hardcoded 3s.