Planning phase errors have blank detail messages after Vercel AI SDK migration

[thameema]

After the Python-to-TypeScript migration (Vercel AI SDK v6), the planning phase produces error entries in task_logs.json with completely empty detail fields.

Steps to Reproduce

1. Create a task (from any source - manual, GitHub, GitLab, or JIRA)
2. Start the task (moves to planning phase)
3. Planning phase completes with errors
4. Check Logs tab - error entries show red 'error' badges with no message

Expected

Error entries should contain the actual error message (e.g., rate limit, auth failure, tool error) so users can diagnose and fix the issue.

Actual

All error entries have empty detail fields:

{"type": "error", "phase": "planning", "detail": ""}

Environment

• Aperant 2.8.0-beta.1
• Claude Code subscription (OAuth, not API key)
• Claude CLI 2.1.81 works fine independently
• macOS

Analysis

The issue appears to be in the agent session runner or worker bridge - errors are caught but the message string is not passed through to the task log entry. The Claude CLI itself works (tested with claude -p 'say hello').

This affects all task types, not just a specific integration.

[thameema]

Additional investigation: The blank error details come from TaskLogWriter.processEvent() at task-log-writer.ts:150:

case 'error':
  this.addEntry(logPhase, 'error', event.error.message);

The StreamEvent error type wraps a SessionError object. When the Vercel AI SDK streamText() throws (e.g., auth failure, rate limit), the error gets classified by error-classifier.ts which produces a proper message. However, the error event being written to task_logs.json has an empty event.error.message.

Likely root cause: the error object from streamText() may have the message on a different property, or the onStepFinish error event from AI SDK v6 structures the error differently than expected.

The orchestrator also logs errors via postLog() (line 671 of worker.ts) rather than emitting proper error stream events, so orchestrator-level errors may not reach task_logs.json at all.

[thameema]

Root cause found: The planning errors are NOT auth or rate limit issues.

The actual error is:

Model completed session without making any tool calls — expected files not created: context.json

The AI model connects and responds, but doesn't make any tool calls during the planning phases (complexity_assessment, discovery). The session completes immediately without the model using Read/Write/Glob/Grep tools to create the required spec files.

This suggests the Vercel AI SDK streamText() configuration may not be properly passing tools to the model, or the system prompt isn't triggering tool use.

The blank error display was a separate issue - the error message 'Error' was being written to task_logs.json without the actual cause. Fixed in our branch by extracting the cause from the agent event system (PLANNING_FAILED event contains the real message).

[thameema]

@AndyMik90 This issue blocks all task execution in the app - not just JIRA tasks. The model connects but doesn't call any tools during planning phases.

Error: Model completed session without making any tool calls — expected files not created: context.json

This appears to be a Vercel AI SDK streamText() configuration issue. Could you take a look? Without this fix, no tasks can progress past the planning phase.