@basilica/openclaw 0.1.0

package/skills/basilica-operator/index.md

@@ -0,0 +1,56 @@
+# Basilica Operator Guide
+
+## Incident Handling
+
+### Deployment Unhealthy
+1. Check deployment health: `basilica_deployment_health`
+2. Check recent errors: `basilica_recent_errors`
+3. If errors indicate OOM or resource exhaustion, check GPU allocation
+4. Restart deployment: `basilica_restart_deployment`
+5. If restart fails, check model compatibility with allocated GPU type
+
+### Job Failures
+1. Check job status: `basilica_job_status`
+2. Check logs: `basilica_query_logs` with source filter matching the job
+3. Common causes:
+   - Insufficient GPU memory -- request larger GPU type
+   - Dataset format mismatch -- verify dataset metadata
+   - Configuration error -- review job config parameters
+4. After fixing, resubmit: `basilica_submit_job`
+
+### Authentication Issues
+1. Verify credentials: re-run `openclaw models auth login --provider basilica`
+2. If OAuth refresh fails, the token may have been revoked -- re-authenticate
+3. Check environment (production vs staging) matches intended target
+
+## Triage Patterns
+
+### Performance Degradation
+1. Query recent error logs for the affected project
+2. Check deployment health across all active deployments
+3. Inspect GPU allocation status -- look for "releasing" or "pending" states
+4. Review recent job submissions for resource contention
+
+### Resource Contention
+1. List GPU resources to identify availability
+2. Check for expired or idle allocations that can be released
+3. Consider region alternatives if primary region is constrained
+
+## Rollback Procedures
+
+### Deployment Rollback
+1. Stop the current deployment: use `basilica_restart_deployment` (will redeploy last known good state)
+2. If the model itself is the issue, deploy the previous model version
+
+### Job Recovery
+1. Cancel the failing job: `basilica_cancel_job`
+2. Review and correct the configuration
+3. Resubmit with corrected parameters
+
+## Operational Best Practices
+
+- Always specify explicit project scope for destructive operations
+- Check GPU availability before submitting large training jobs
+- Monitor deployment health after any configuration change
+- Use staging environment for testing new model deployments before production
+- Review error logs before escalating -- most issues are configuration-related

package/skills/basilica-platform/index.md

@@ -0,0 +1,67 @@
+# Basilica Platform
+
+## Overview
+
+Basilica is a GPU compute and AI infrastructure platform. It provides managed inference, training jobs, model deployments, and dataset management through a unified API.
+
+## Resource Model
+
+### Organizations
+- Top-level grouping for teams and billing
+- Each org has one or more projects
+
+### Projects
+- Primary scope for all resources (jobs, deployments, datasets, GPU allocations)
+- Tools require a project scope -- either from config or explicit parameter
+- Format: `org-name/project-name`
+
+### Models
+- Identified by `modelId` (e.g., `basilica/llama-3-70b`)
+- Available for inference and deployment
+- Some models require specific GPU types
+
+### Jobs
+- Types: training, inference, fine-tune, evaluation
+- Lifecycle: queued -> running -> completed/failed/cancelled
+- Each job is scoped to a project
+
+### GPU/Compute
+- Types: A100, H100, H200, L40S, T4
+- Allocations are time-bound and region-specific
+- Request compute before submitting large jobs
+
+### Deployments
+- Serve models as persistent endpoints
+- Lifecycle: deploying -> active/failed -> stopped
+- Support health checks, restart, and rolling updates
+
+### Datasets
+- Immutable versioned data artifacts
+- Metadata includes format, size, row count
+- Artifacts are the physical storage locations
+
+## Authentication
+
+Basilica supports two authentication methods:
+
+1. **OAuth (device code flow)** -- interactive login, produces refreshable credentials
+2. **API Token** -- non-interactive, exchanged for short-lived access tokens
+
+Credentials are stored in OpenClaw's standard auth-profile store.
+
+## Common Workflows
+
+### Run inference
+1. Verify auth: `openclaw models auth login --provider basilica`
+2. Use `basilica_start_inference` with model and input
+3. Poll with `basilica_query_inference` if async
+
+### Submit a training job
+1. Request compute: `basilica_request_compute` for required GPU type
+2. Submit job: `basilica_submit_job` with training config
+3. Monitor: `basilica_job_status`
+
+### Deploy a model
+1. Deploy: `basilica_deploy` with model and config
+2. Health check: `basilica_deployment_health`
+3. Restart if needed: `basilica_restart_deployment`