eve-deploy-debugging（前夕部署调试）

Eve Deploy and Debug

Use these steps to deploy and diagnose app issues quickly.

Environment Setup

• Get the staging API URL from your admin.
• Create and use a profile:

eve profile create staging --api-url https://api.eh1.incept5.dev
eve profile use staging

Infrastructure Change Policy

Never run kubectl apply, helm install, or any direct Kubernetes resource creation against shared infrastructure. All infrastructure changes go through Terraform. Use the Eve CLI (eve env, eve env deploy) to manage application deployments — the platform handles the underlying k8s resources.

Deploy Flow (Staging)

# Create env if needed
eve env create staging --project proj_xxx --type persistent

# Deploy (requires --ref with 40-char SHA or a ref resolved against --repo-dir)
eve env deploy staging --ref main --repo-dir .

# When environment has a pipeline configured, the above triggers the pipeline.
# Use --direct to bypass pipeline and deploy directly:
eve env deploy staging --ref main --repo-dir . --direct

# Pass inputs to pipeline:
eve env deploy staging --ref main --repo-dir . --inputs '{"key":"value"}'

Deploy Polling Flow

When eve env deploy is called:

1. Direct deploy (no pipeline): Returns deployment_status directly. Poll health endpoint until ready === true.
2. Pipeline deploy: Returns pipeline_run_id. Poll GET /pipelines/{name}/runs/{id} until all steps complete, then check health.

Deploy is complete when: ready === true AND active_pipeline_run === null.

Observe the Deploy

eve job list --phase active
eve job follow <job-id>              # Real-time SSE streaming
eve job watch <job-id>               # Poll-based status updates
eve job diagnose <job-id>            # Full diagnostic
eve job result <job-id>              # Final result
eve job runner-logs <job-id>         # Raw worker logs

Real-Time Debugging (3-Terminal Approach)

# Terminal 1: Pipeline/job progress
eve job follow <job-id>

# Terminal 2: Environment health
eve env diagnose <project> <env>

# Terminal 3: System-level logs
eve system logs

Debugging Workflows

Job Won't Start

1. Check dependencies: eve job dep list <job-id>
2. Check if blocked: eve job show <job-id> → look at blocked_by
3. Verify environment readiness: eve env show <project> <env>
4. Check orchestrator: eve system orchestrator status

Job Failed

1. Get the error: eve job diagnose <job-id>
2. Check logs: eve job follow <job-id> or eve job runner-logs <job-id>
3. If build failure: eve build diagnose <build-id>
4. If secret failure: eve secrets list --project <project_id>

Job Stuck Active

1. Check if waiting for input: eve job show <job-id> → effective_phase
2. Check thread messages: eve thread messages <thread-id>
3. Check runner pod: eve system pods

System Issues

1. API health: eve system health
2. Orchestrator: eve system orchestrator status
3. Recent events: eve system events

Common Error Messages

Build Failures

If a deploy pipeline fails at the build step:

eve build list --project <project_id>
eve build diagnose <build_id>
eve build logs <build_id>
eve secrets list --project <project_id>     # Required for BYO/custom registry: REGISTRY_USERNAME, REGISTRY_PASSWORD

Common build failures:

• Registry auth: For BYO/custom registry, verify REGISTRY_USERNAME and REGISTRY_PASSWORD secrets
• Dockerfile not found: Check build.context path in manifest
• Multi-stage build failure: BuildKit handles these correctly; Kaniko may have issues
• Workspace errors: Build context not available — check eve build diagnose

Worker Image Registry

Eve publishes worker images to the configured private registry with these variants:

Version pinning: Use semver tags (e.g., v1.2.3) in production. Use SHA tags or :latest in development.

Platform Environment Variables

Eve automatically injects these into every deployed service container:

Use EVE_API_URL for backend calls. Use EVE_PUBLIC_API_URL for browser/client-side code. Services can override any of these by defining them explicitly in their manifest environment section.

Access URLs

• URL pattern: {service}.{orgSlug}-{projectSlug}-{env}.{domain}
• Local dev default domain: lvh.me
• Ask the admin for the correct domain (staging vs production).

Environment-Specific Debugging

Private Endpoints (Tailscale)

Connect services on private networks (home lab GPUs, internal APIs, dev machines) to the Eve cluster. The platform creates K8s ExternalName services backed by Tailscale egress proxies.

# Register a private endpoint
eve endpoint add \
  --name lmstudio \
  --provider tailscale \
  --tailscale-hostname mac-mini.tail12345.ts.net \
  --port 1234 \
  --org org_xxx

# List and inspect
eve endpoint list --org org_xxx
eve endpoint show lmstudio --org org_xxx

# Diagnose connectivity
eve endpoint diagnose lmstudio

# Remove
eve endpoint remove lmstudio --org org_xxx

Each endpoint gets a stable in-cluster DNS name: http://{orgSlug}-{name}.eve-tunnels.svc.cluster.local:{port}. Wire it into apps/agents via secrets:

eve secrets set LLM_BASE_URL \
  "http://myorg-lmstudio.eve-tunnels.svc.cluster.local:1234/v1" \
  --scope project

Diagnostics check: operator status, K8s service existence, DNS resolution, TCP connectivity, and HTTP health.

Worker Toolchain-on-Demand

The default worker image is base (~800MB with Node.js, git, and all harnesses). Toolchains (Python, Rust, Java, Kotlin, media) are injected on-demand via init containers rather than bundled in a fat image.

Deployment impact: If an agent job needs toolchains, the runner pod starts init containers that copy toolchain binaries from small pre-built images. First pull adds ~5-10s; subsequent jobs on the same node use cached images.

Debugging toolchain issues:

# Check if toolchains are declared in agent config
# agents.yaml: toolchains: [python]

# If a toolchain binary is missing at runtime:
# 1. Verify agent config has the toolchain declared
# 2. Check init container logs on the runner pod
# 3. Verify toolchain images are available in the registry

To use the full image (all toolchains bundled): set EVE_WORKER_VARIANT=full or use --variant full locally.

App Undeploy/Delete Lifecycle

Remove environments and clean up resources:

# Undeploy services from an environment (stops pods, keeps env record)
eve env undeploy <project> <env>

# Delete the environment entirely (removes env record, managed DB, secrets)
eve env delete <project> <env>

When a managed DB is attached, eve env delete deprovisions it. Secrets scoped to the environment are cleaned up. The environment's pipeline history remains in the audit log.

For app-level cleanup, remove the project:

eve project delete <project-id>

This cascades: environments, secrets, pipeline history, and build artifacts are removed.

Workspace Janitor

Production disk management for agent workspaces:

• EVE_WORKSPACE_MAX_GB — total workspace budget
• EVE_WORKSPACE_MIN_FREE_GB — trigger cleanup threshold
• EVE_SESSION_TTL_HOURS — auto-evict stale sessions
• LRU eviction when approaching budget; TTL cleanup for idle sessions
• K8s: per-attempt PVCs deleted on completion

Related Skills

• Local dev loop: eve-local-dev-loop
• Secrets: eve-auth-and-secrets
• Manifest changes: eve-manifest-authoring