Parent: TechDevGroup/devagentic#203.
Gap
Lane H (kind:reasoning-graft-candidate) is graph-resident — emitted by the confer-loop downstream — but workers have no hermes-native read surface. Per NousResearch#203 §3.2, a worker requesting a paid-tier reasoning lift should call requestReasoningGraft and read the result via the same tool surface.
What is needed
Add MCP tools:
requestReasoningGraft(prompt: str, lift_model?: str) — kicks off a Lane H lift via devagentic's confer/lift pipeline. Returns the resulting reasoning-graft-candidate doc id.
fetchReasoningGraft(graft_id: str) — fetch the lift result for the worker to use as preamble for its next moves.
Acceptance
- Both tools registered under
mcp_serve.py.
- Unit test verifying the round-trip: request → graft doc written → fetch returns content.
- Docs explain the lift-and-graft pattern (high-tier model = synthesis; cheap silos = breadth; lifts grafted DOWN as preamble for subsequent calls).
Notes
- Depends on G2 (fetchReasoningGraft is a specialized querySilo read). Could share an MCP module with G2.
Parent: TechDevGroup/devagentic#203.
Gap
Lane H (
kind:reasoning-graft-candidate) is graph-resident — emitted by the confer-loop downstream — but workers have no hermes-native read surface. Per NousResearch#203 §3.2, a worker requesting a paid-tier reasoning lift should callrequestReasoningGraftand read the result via the same tool surface.What is needed
Add MCP tools:
requestReasoningGraft(prompt: str, lift_model?: str)— kicks off a Lane H lift via devagentic's confer/lift pipeline. Returns the resultingreasoning-graft-candidatedoc id.fetchReasoningGraft(graft_id: str)— fetch the lift result for the worker to use as preamble for its next moves.Acceptance
mcp_serve.py.Notes