Support typed agentConfig model and permissions across all agents

[harshitsinghbhandari]

This was generated by AI during triage.

Summary

PR #154 introduces typed per-project agentConfig.model and agentConfig.permissions, but consumption is currently Claude Code-specific. We should support these fields consistently across all agent adapters with a provider-qualified model naming scheme.

Desired behavior

• agentConfig.model should be accepted and resolved for every supported agent adapter that can select a model.
• Model values should use a qualified format:
  • claude.claude-opus-4-8
  • codex.gpt-5-5
  • <agent-or-provider>.<model-id>
• agentConfig.permissions should keep the shared AO vocabulary:
  • default
  • accept-edits
  • auto
  • bypass-permissions
• Each adapter maps the shared permission value to its own native implementation/flags/config. If an adapter cannot support a mode, it should fail clearly or document the fallback.

Scope

• Define the canonical model string format and validation rules.
• Add adapter-owned model/permission mapping for all relevant agents.
• Keep API/CLI values stable and agent-agnostic; implementation differences stay inside adapters.
• Add tests per adapter for:
  • model flag/config application
  • each supported permission mode
  • unsupported model/provider/mode errors

Acceptance criteria

• A project can set a model such as claude.claude-opus-4-8 or codex.gpt-5-5 and have the selected adapter launch with that model.
• A project can set the same permissions value regardless of adapter, and each adapter translates it appropriately.
• Unsupported values are rejected with actionable errors.
• Documentation explains the model naming format and permission compatibility per adapter.