Description

Implement a two-pipe system for delegate_task to allow the main model to discover authenticated providers and reasoning effort levels (Discovery Pipe) and then spawn subagents with these specific parameters (Action Pipe).

Implementation Plan

Reference: .hermes/plans/per-call-provider-model-delegation.md

Key Deliverables:

1. Discovery Pipe: Implement build_delegation_capabilities_prompt() and inject it into stable_parts of the system prompt in run_agent.py.
2. Action Pipe:
   • Update DELEGATE_TASK_SCHEMA to support provider, model, and reasoning_effort.
   • Implement resolution chain: per-call $\rightarrow$ config $\rightarrow$ parent.
   • Move credential resolution inside the batch loop to support heterogeneous tasks.
   • Fix 'Provider-only override' trap (resolve default model for target provider).
3. Verification: Implement 24 test cases covering priority chain and edge cases.

Technical Constraints

• Priority: Per-call argument > delegation.* config > laurent inheritance.
• Performance: Discovery guidance must be in stable_parts to maintain prompt cache efficiency.
• Safety: Maintain fix for bug [Bug]: Delegation subagents fail to resolve credentials and hit 404s on mixed-model providers (OpenCode) #20558 (api_mode not inherited when provider changes).

Labels

• in-progress, delegation, staff-sde