resources

name

resources

description

Use this skill to discover and configure compute resources — SSH keys, AWS credentials, GPU servers, model files, data directories, and Python environment managers. Triggers when the user asks about available resources, credentials, servers, or when another skill needs non-local access. Use for: '看看有什么资源', 'scan for GPUs', 'find credentials', 'set up server access', 'what envs do I have'. Also called automatically by run skills when credentials are missing.

/resources — Resource Discovery

Search the local machine for credentials, models, and data in common default locations. Can be invoked standalone or triggered by other skills when a resource is needed.

Ask one question at a time — multiple questions at once overwhelms users.

Follow the Workflow State Protocol from CLAUDE.md: push on entry, update step as you progress, pop on completion.

Prerequisites

Ensure {WORKSPACE}/resources.json exists (workspace root, shared across projects). If not, copy from lifecycle/resources.json. Resolve {WORKSPACE} from project.json -> workspace or parent directory of project root.

What to search

Credentials

Type	Default locations
SSH keys	~/.ssh/id_rsa, ~/.ssh/id_ed25519, ~/.ssh/*.pem, ~/.ssh/config
AWS	~/.aws/credentials, ~/.aws/config, env vars AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / AWS_PROFILE
Docker/Registry	~/.docker/config.json
Kubernetes	~/.kube/config
GCP	~/.config/gcloud/application_default_credentials.json
Azure	~/.azure/
Custom tokens	env vars matching *_TOKEN, *_API_KEY, *_SECRET

Servers

Discover from ~/.ssh/config (Host, HostName, User, IdentityFile) and ~/.ssh/known_hosts.

Run python lifecycle/scripts/resources/parse_ssh_config.py. Fallback: manually read ~/.ssh/config.

For each server found, create an entry with host, username, ssh_key_path, alias (SSH config Host name). Ask user for mlclaw_root (remote workspace root for path mapping) and python_path (try ssh <host> "which python3" to auto-detect). Leave gpu/description empty for user to fill.

Optional GPU probe: with user permission, try ssh <host> "nvidia-smi --query-gpu=name,count --format=csv,noheader".

Python Environment Manager

Check in preference order: mamba -> conda -> uv. Record first found in resources.json -> local.env_manager. Also run conda env list to record existing environments.

If none found, warn that /refactor-init needs mamba or conda.

Models & Artifacts

Type	Locations
HuggingFace cache	~/.cache/huggingface/hub/
Torch hub	~/.cache/torch/hub/
ONNX/TensorRT	~/models/, ~/weights/, D:\models\, stage artifacts/

Data

Common dirs: ~/data/, ~/datasets/, D:\data\, D:\datasets\, {PROJECT}/stages/*/data/.

Flow

Step 1: Check resources.json first

If already has non-empty values for the requested type, show what's cached and ask whether to re-search. If user says no, use cached values.

Step 2: Search

Run search for requested category. Report findings with types and sizes.

Step 3: Auto-save

Show proposed writes to {WORKSPACE}/resources.json. User confirms (y/n/edit). Existing non-empty values are shown side-by-side with new values so the user can choose — overwriting silently would lose manual configuration.

Step 4: If nothing found

Report what's missing, ask if user has credentials to provide manually (one field at a time).

Usage modes

Standalone (/resources): ask what to search for (credentials / models / data / all), then follow the flow.

Called by another skill: check cache first -> if valid, return immediately -> if not, search relevant category only -> save -> return.

AWS Credential Troubleshooting

When AWS SSO (aws sso login) fails:

Problem	Symptom	Fix
SSO config expired	InvalidRequestException on RegisterClient	Confirm sso_start_url with admin
Region mismatch	InvalidRequestException	Check sso_region matches Identity Center region
Network/proxy blocking	SSO hangs	Check corporate firewall/proxy
SSO cache corrupted	Various errors	Clear ~/.aws/sso/cache/ and retry

Fallback: SSO fails -> try default profile -> try static credentials in ~/.aws/credentials -> ask user. Report SSO error as warning so user can fix later, but continue with whatever works.

Safety

Secret values (keys, passwords, tokens) are never displayed — only paths, profiles, and metadata. This protects against accidental exposure in logs or shared screens.