ARCH-003: Remove all direct LLM calls from infrastructure code #121

New issue

Closed

opened 2026-04-11 01:22:14 +00:00 by Mark · 1 comment

Mark commented

2026-04-11 01:22:14 +00:00

Owner

Problem

LLM calls exist outside of worker execution context. Infrastructure components (mesh_node, evolution, validation) directly call LLMs instead of routing through workers.

Current Violations

Location	LLM Usage	Problem
`mesh_node.py`	None currently (after recent fixes)	Fixed
`evolution/championship.py`	`run_writer_capability()`	Evolution should use workers
`validation_agent.py`	`AgenticValidator` calls Ollama	Validation should be a worker
`ingress_controller/run.py`	`_call_llm()` method	Should be worker task

Why This Is Wrong

Infrastructure thinks - breaks agentic design
Cannot track resource usage - bypasses ledger
No skill/tool integration - direct LLM calls miss capabilities
Testing complexity - infrastructure shouldn't need LLM mocking

Fix Required

Evolution - Submit test intents to mesh, let workers execute, score results
Validation - Make ValidationAgent a proper worker that bids on validation intents
Ingress Controller - Route to workers instead of direct LLM calls
All LLM calls go through call_llm() in worker execution context only

Files Affected

evolution/championship.py - Remove direct capability calls
substrate/validation_agent.py - Convert to worker
capabilities/_muscle/ingress_controller/run.py - Route to workers

Acceptance Criteria

No LLM imports in infrastructure code
All LLM calls happen in worker context
Evolution uses mesh intent flow
Validation is a worker, not infrastructure

Notes

This ensures only agents (workers) think, infrastructure just routes.

See also: issues/04_remove_llm_from_infrastructure.md

## Problem LLM calls exist outside of worker execution context. Infrastructure components (mesh_node, evolution, validation) directly call LLMs instead of routing through workers. ### Current Violations | Location | LLM Usage | Problem | |----------|-----------|---------| | `mesh_node.py` | None currently (after recent fixes) | Fixed | | `evolution/championship.py` | `run_writer_capability()` | Evolution should use workers | | `validation_agent.py` | `AgenticValidator` calls Ollama | Validation should be a worker | | `ingress_controller/run.py` | `_call_llm()` method | Should be worker task | ### Why This Is Wrong - **Infrastructure thinks** - breaks agentic design - **Cannot track resource usage** - bypasses ledger - **No skill/tool integration** - direct LLM calls miss capabilities - **Testing complexity** - infrastructure shouldn't need LLM mocking ### Fix Required 1. **Evolution** - Submit test intents to mesh, let workers execute, score results 2. **Validation** - Make ValidationAgent a proper worker that bids on validation intents 3. **Ingress Controller** - Route to workers instead of direct LLM calls 4. All LLM calls go through `call_llm()` in worker execution context only ### Files Affected - `evolution/championship.py` - Remove direct capability calls - `substrate/validation_agent.py` - Convert to worker - `capabilities/_muscle/ingress_controller/run.py` - Route to workers ### Acceptance Criteria - [ ] No LLM imports in infrastructure code - [ ] All LLM calls happen in worker context - [ ] Evolution uses mesh intent flow - [ ] Validation is a worker, not infrastructure ## Notes This ensures only agents (workers) think, infrastructure just routes. See also: issues/04_remove_llm_from_infrastructure.md

Mark referenced this issue from a commit

2026-04-19 16:33:53 +00:00

Remove remaining local issue files now tracked in Forgejo

Mark commented

2026-05-05 01:14:12 +00:00

Author

Owner

Closing — RESOLVED per M008 validation and code inspection:

All openai/anthropic imports removed from docker/*.py (verified via grep — zero matches)
Infrastructure code (mesh_node.py, bid_manager.py, worker_spawner.py, etc.) has zero direct LLM calls
Workers at primitives/_muscle/* handle LLM calls through their own run.py execute functions
Verified by R121 and PROJECT.md

Closing — RESOLVED per M008 validation and code inspection: - All openai/anthropic imports removed from docker/*.py (verified via grep — zero matches) - Infrastructure code (mesh_node.py, bid_manager.py, worker_spawner.py, etc.) has zero direct LLM calls - Workers at primitives/_muscle/* handle LLM calls through their own run.py execute functions - Verified by R121 and PROJECT.md

Mark closed this issue

2026-05-05 01:14:15 +00:00

Mark referenced this issue

2026-06-07 05:57:18 +00:00

LLM Upgrade: intent_router semantic routing decisions #83

Mark referenced this issue

2026-06-07 06:35:01 +00:00

[Phase 2] Tailnet Federation BLOCKED - Key Management Infrastructure Required #56