feat(cognitive): add local provider registry and discovery

This commit is contained in:
m
2026-08-17 09:28:33 +03:00
parent 9e99a04c76
commit a25a57ab69
38 changed files with 2344 additions and 116 deletions
+11 -2
View File
@@ -1,6 +1,15 @@
# Assistant project status
Audit date: 2026-08-16. Active implementation: DEV-000053 revision 1.
Audit date: 2026-08-17. Active implementation: DEV-000056 revision 1.
## Current milestone
**Provider-neutral Cognitive Architecture + Local AI integration.**
- Docker production baseline at `9e99a04c…` is completed and healthy.
- The OpenAI integration is a bootstrap adapter only. Real OpenAI credentials and production traffic are intentionally deferred.
- The current milestone integrates one normally available local CPU provider and one opportunistic local Ollama/GPU provider through typed discovery, health and routing boundaries.
- Discovery never activates a provider, model or binding. Until a later benchmark and explicit promotion task, all discovered bindings remain `INACTIVE` and production remains fully LLM-free.
## Delivered
@@ -20,7 +29,7 @@ Audit date: 2026-08-16. Active implementation: DEV-000053 revision 1.
## Remaining work requiring a new task decision
1. **Cognitive provider:** the cognitive worker intentionally remains the deterministic fake. Selecting a real model/provider, data policy and allowed proposal contract requires a separate capability/architecture decision; it must not gain canonical mutation authority.
1. **Cognitive qualification and promotion:** DEV-000056 discovers and registers local providers/models without activation. A later benchmark decision must establish demonstrated capabilities and explicitly promote exact binding revisions through `EVALUATING → VALIDATED → CANARY → ACTIVE`.
2. **Product breadth:** notification delivery is the only accepted real external operation. Any additional operation needs its own proposal, authority model, idempotency/reconciliation semantics and Capability binding.
3. **Artifact and Epistemics product behavior:** their ownership boundaries exist, but no product task has yet specified durable user-facing workflows for them.
@@ -0,0 +1,39 @@
{
"artifact_id": "gpu-sustained-benchmark-2026-08-17-v1",
"artifact_version": 1,
"source": "user-provided sustained benchmark",
"observed_at": "2026-08-17T00:00:00+03:00",
"hardware_profile": "NVIDIA RTX 4090 24 GB VRAM",
"runtime": "Ollama",
"settings": {
"temperature": 0,
"seed": 42,
"context_units": 4096,
"thinking_enabled": false
},
"results": [
{
"provider_model_id": "qwen3:30b",
"parameter_size": "30.5B",
"quantization": "Q4_K_M",
"memory_bytes": 18870000000,
"generation_tps": [
{"output_units": 128, "median": 224.99},
{"output_units": 512, "median": 219.40},
{"output_units": 1024, "median": 208.43}
]
},
{
"provider_model_id": "gemma4:26b",
"parameter_size": "25.8B",
"quantization": "Q4_K_M",
"memory_bytes": 17350000000,
"generation_tps": [
{"output_units": 128, "median": 168.10},
{"output_units": 512, "median": 165.54},
{"output_units": 1024, "median": 164.12}
]
}
],
"interpretation": "performance evidence only; no quality or activation conclusion"
}