accomplish-ai/accomplish @accomplish_ai/agent-core@0.4.0

on GitHub

Minor Changes

• f954f59: Add needs_planning classification to start_task tool. The agent still calls start_task for every message (preserving discipline), but now sets needs_planning: false for simple messages like greetings and knowledge questions. When false, the adapter skips plan card emission and todo creation, and the completion enforcer treats it as a conversational turn — no continuation prompts needed.

Patch Changes

• 68094c9: refactor(agent-core): remove stale duplicate files from opencode/ directory

  Deleted 6 dead files left behind after the encapsulation refactor: opencode/task-manager.ts,
  opencode/adapter.ts, opencode/stream-parser.ts, opencode/log-watcher.ts, opencode/index.ts
  (barrel), and internal/classes/CompletionEnforcer.ts. The canonical implementations live in
  internal/classes/ and are used via the factory pattern. Re-pointed test imports to the active
  internal/classes/ sources. No behavior changes; public API unchanged.

• 0c555bf: fix(agent-core): serialize non-string error values in TaskResult to prevent downstream TypeError

  The stream parser casts JSON.parse output to OpenCodeMessage without runtime validation. When the
  OpenCode CLI emits an error message with a non-string error field (e.g. an object like
  {name: "APIError", data: {...}}), the raw object was passed through to TaskResult.error which is
  typed as string | undefined. Downstream consumers calling .toLowerCase() on this value would
  crash with TypeError: errorName.toLowerCase is not a function.

  Added defensive coercion in the adapter's error handling path: string values pass through unchanged,
  non-string values are serialized via JSON.stringify to preserve error details.