Fix: Always write parsed JSONL response to output_file
Problem: Codex CLI's -o (--output-last-message) flag only captures the last agent message, which may be a brief acknowledgment instead of the full analysis. The wrapper's mtime-based check would then trust this brief output over the comprehensive JSONL-parsed stdout.
This particularly affected the planner role, where the model produces multiple messages interleaved with tool calls — the full plan would be in an intermediate message, but -o only captured the final summary.
Fix:
- Remove
-oflag from Codex CLI invocation entirely - Always write parsed JSONL stdout (which contains ALL agent messages) to the output file
- Apply same fix to Gemini CLI output handling
- Remove mtime-based "did CLI write?" check (no longer needed)
Files changed: codex-core.ts, gemini-core.ts + built artifacts