Two GSD instances share model config — second overrides first

[originlabs-app]

When running two GSD CLI instances simultaneously on different projects, the model configuration bleeds
between them. After a few milestones, both instances align to whichever model was configured last.

Steps to reproduce:

1. Open terminal 1: cd ~/project-a && gsd (uses Claude Opus 4.6)
2. Open terminal 2: cd ~/project-b && gsd (uses Codex GPT-5.4)
3. Let both run for a few milestones
4. Observe: terminal 1 switches from Claude to Codex (or vice versa) without user action

Observed behavior:

• Terminal 1 (project ~/.openclaw/): started on Claude Opus 4.6, later switched to Codex GPT-5.4 with
  thinking off
• Terminal 2 (project ~/droid-mission-research/): started on Codex GPT-5.4
• After several milestones, both terminals show Codex GPT-5.4
• The status bar confirms the switch: (openai-codex) gpt-5.4 • thinking off on a project that was
  configured for Claude

Expected behavior:

Each GSD instance should maintain its own model configuration independently, regardless of other
running instances.

Environment:

• gsd-pi v2.14.4
• Two simultaneous instances on different repos
• macOS Darwin 25.0.0 (Apple Silicon, 16GB RAM)

[trek-e]

Fix in #652.

Root cause: setDefaultModelAndProvider() in settings-manager.ts always wrote to the global ~/.gsd/agent/settings.json. When two instances run concurrently and either changes models (interactive cycling, /model command, etc.), the global default gets overwritten. On the next context wipe, the other instance reads the contaminated global default.

Two-layer fix:

1. Project-scoped persistence — model setters now write to project-level settings (.pi/settings.json) when a project settings file exists, falling back to global only when no project context is available. This structurally isolates concurrent instances.

2. Auto-mode model capture — captures the session's model at auto-mode start and re-applies it at each unit dispatch (with persist: false) when no per-unit-type model preference is configured. Defense in depth for the case where project settings don't exist yet.

To avoid this in the meantime (before the fix lands): configure per-phase models in your project's .gsd/preferences.md:

models:
  research: claude-opus-4-6
  planning: claude-opus-4-6
  execution: claude-opus-4-6
  completion: claude-opus-4-6

This forces the model at every dispatch point, bypassing the global settings file entirely.

[originlabs-app]

Excellent, will try

[jeremymcs]

Thought I had already patched this issue. Good catch @trek-e