Skills Inventory: Updates, Renames, and New Knowledge

Skills Updated

Skill	Version Change	Key Updates
hosted-model-ctl	2.3 → 4.0	Added rhel-ai host profile (4× L4); g30b-96k and g8b-128k models; InstructLab container support; port 8000→10000 for rhtevan-work; port 9000 for rhel-ai
skupper-model-provider	1.0 → 4.0	Complete rewrite: edge→interior mode; linux→podman platform; official container image; dual routing keys; inter-router links
goose-skupper-provider	1.0 → 3.1	Multi-port routing; defensive JSON template with warning; port 10000/9000
skill-index	2.3 → 2.4	Fixed signals extraction bug (regex→state machine); 44/48→48/48 skills with signals
agentfs-setup	3.8 → 3.9	“Never improvise” routing rule; “Backup untracked files” guardrail; log.md comment removal
skill-gen	1.5 → 1.6	“Defensive file templates” writing guidance; writes-files schema field

Skills Renamed

Old Name	New Name	Reason
local-model-ctl	hosted-model-ctl	Reflects “self-hosted” models, not “local” — models run on remote GPU hosts

SSH Profile Created

Profile	Host	User	Purpose
rhel-ai	bastion.g7cpg.sandbox600.opentlc.com	cloud-user	4× NVIDIA L4 cloud instance for large model serving

AgentFS Guardrail Changes (Template v3.9)

Guardrail	Change
Signal Routing Rules	Added: “Never improvise when a skill exists”
#9 Checkpoints	Added: “Backup untracked files” for non-git-tracked files
#5 Filesystem Integrity	Changed log.md insertion anchor from HTML comment to heading

Model Registry (Final State)

rhtevan-work (RTX A500, 4 GB VRAM)

Alias	Model	Engine	Port
g350m	Granite 4.0 350M	vLLM FP16	10000
g1b	Granite 4.0 1B	vLLM INT4	10000
g8b	Granite 4.1 8B	llama.cpp Q4_K_M	10000

rhel-ai (4× NVIDIA L4, 92 GB VRAM)

Alias	Model	Engine	TP	Port	Context
g30b-96k	Granite 4.1 30B	vLLM BF16	4	9000	96K
g8b-128k	Granite 4.1 8B	vLLM BF16	2	9000	128K

Final Skupper VAN Architecture

localhost (interior, outbound only, podman, skrouterd 3.5.2)
  ├── 3× inter-router → 192.168.1.177:55671 (rhtevan-work)
  │     Listener :10000 ← model-api-rhtevan-work
  │
  └── 3× inter-router → 3.23.208.217:8000 (rhel-ai)
        Listener :9000  ← model-api-rhel-ai

Port 8000 unoccupied on localhost

Candidate Knowledge for Future Extraction

Topic	Type	Scope
Skupper V2 podman platform operational guide	Knowledge bundle	Cross-project
vLLM tensor parallelism VRAM calculator	Skill or reference doc	Cross-project
InstructLab container as vLLM runtime	Knowledge concept	Cross-project
Goose custom provider JSON schema	Reference doc	Cross-project
AWS security group port workarounds	Knowledge concept	Cross-project