truefoundry-deploy（truefoundry 部署）

Routing note: For ambiguous user intents, use the shared clarification templates in references/intent-clarification.md.

Deploy to TrueFoundry

Route user intent to the right deployment workflow. Load only the references you need.

Intent Router

Load only the reference file matching the user's intent. Do not preload all references.

General Principle: Ask, Don't Assume

When in doubt, ask. If any deployment parameter is ambiguous or missing — branch, workspace, image, port, resources, environment — ask the user rather than picking a value and proceeding silently. A wrong assumption can deploy to the wrong environment, from the wrong branch, or with the wrong configuration. The cost of one extra question is always lower than the cost of a bad deploy.

Examples of things to ask rather than assume:

• Which workspace to deploy to (even if only one exists)
• Which branch to build from (especially if the manifest branch differs from the local branch)
• Whether to use the existing manifest as-is or update it
• Which Docker image tag or registry to use
• Whether the service should be public or internal

Do NOT silently default to the current value of anything that could have changed or that the user has not explicitly confirmed for this deployment.

Prerequisites (All Workflows)

# 1. Check credentials
grep '^TFY_' .env 2>/dev/null || true
env | grep '^TFY_' 2>/dev/null || true

# 2. Derive TFY_HOST for CLI (MUST run before any tfy command)
export TFY_HOST="${TFY_HOST:-${TFY_BASE_URL%/}}"

# 3. Check CLI
tfy --version 2>/dev/null || echo "Install: pip install 'truefoundry==0.5.0'"

# 4. Check for existing manifests
ls tfy-manifest.yaml truefoundry.yaml 2>/dev/null

• TFY_BASE_URL and TFY_API_KEY must be set (env or .env).
• TFY_HOST must be set before any tfy CLI command. The export above handles this automatically.
• TFY_WORKSPACE_FQN required. HARD RULE: Never auto-pick a workspace. Always ask the user to confirm, even if only one workspace exists or a preference is saved. See references/prerequisites.md for the full workspace confirmation flow.
• For full credential setup, see references/prerequisites.md.

WARNING: Never use source.env. The tfy-api.sh script handles .env parsing automatically. For shell access: grep KEY.env | cut -d= -f2-

CRITICAL: tfy apply vs tfy deploy

HARD RULE: tfy apply does NOT support build_source.type: local. If the manifest has a local build source, you MUST use tfy deploy -f <manifest>. Using tfy apply with a local build source will fail with: must match exactly one schema in oneOf.

Before running any deploy command, check the manifest:

1. If build_source.type: local → use tfy deploy -f
2. Otherwise → tfy apply -f is fine

Pre-Flight Manifest Validation (MANDATORY)

Before attempting any deploy/apply, run these checks. Fix issues before deploying — do not deploy a known-bad manifest.

1. Exposed port requires host

If any port has expose: true, it must have a host field. Deploying without it will fail with: Host must be provided to expose port.

Auto-generate the host if missing:

TFY_API_SH=~/.claude/skills/truefoundry-deploy/scripts/tfy-api.sh

# Get cluster ID from workspace FQN (format: cluster-id:workspace-name)
CLUSTER_ID=$(echo "$TFY_WORKSPACE_FQN" | cut -d: -f1)

# Discover base domain from cluster manifest
bash $TFY_API_SH GET "/api/svc/v1/clusters/$CLUSTER_ID"
# → Response is at data.manifest.base_domains[] (array of strings)
# → Look for wildcard entry (e.g., "*.ml.example.truefoundry.cloud")
# → Strip "*." to get base domain: "ml.example.truefoundry.cloud"
# → Construct host: "{service-name}-{workspace-name}.{base_domain}"

Pattern: {service-name}-{workspace-name}.{base_domain}

2. Local build source requires tfy deploy

If the manifest contains build_source.type: local, ensure the deploy command is tfy deploy -f, NOT tfy apply.

3. capacity_type compatibility

spot_fallback_on_demand is not supported on all clusters. If you're unsure, use on_demand or omit capacity_type entirely to let the platform decide. Valid safe values: on_demand, spot.

4. build_spec.type must be exact

Only dockerfile and tfy-python-buildpack are valid. Do NOT use docker, build, python, or any other value.

5. Git branch mismatch (existing manifest + git source)

If an existing manifest has build_source.type: git with a branch_name set, compare it to the current local branch before deploying:

# Use only the specific manifest file for this deployment (not both at once)
# Use -h to suppress the filename prefix so the bare value can be compared
grep -h 'branch_name:' "$MANIFEST_FILE" 2>/dev/null | head -1 | sed 's/.*branch_name:[[:space:]]*//'

# Get current local branch
git branch --show-current 2>/dev/null

If the branches differ, stop and ask the user:

The manifest specifies branch_name: {manifest_branch}, but your current local branch is {current_branch}. Which branch should be deployed? 1. Keep manifest branch: {manifest_branch} (deploy as-is, no manifest change) 2. Use current branch: {current_branch} (update branch_name in the manifest)

Never silently override the manifest's branch_name with the current local branch.

Quick Ops (Inline)

Apply a manifest (pre-built image or git source)

# tfy CLI expects TFY_HOST when TFY_API_KEY is set
export TFY_HOST="${TFY_HOST:-${TFY_BASE_URL%/}}"

# Preview changes
tfy apply -f tfy-manifest.yaml --dry-run --show-diff

# Apply
tfy apply -f tfy-manifest.yaml

Deploy from local source

# tfy CLI expects TFY_HOST when TFY_API_KEY is set
export TFY_HOST="${TFY_HOST:-${TFY_BASE_URL%/}}"

# MUST use tfy deploy (not tfy apply) for local builds
tfy deploy -f truefoundry.yaml --no-wait

Reminder: tfy apply does NOT support build_source.type: local. Use tfy deploy -f for local builds.

Minimal service manifest template

name: my-service
type: service
image:
  type: image
  image_uri: docker.io/myorg/my-api:v1.0
ports:
  - port: 8000
    expose: false  # Set true + add host for public access
    app_protocol: http
resources:
  cpu_request: 0.5
  cpu_limit: 1
  memory_request: 512
  memory_limit: 1024
  ephemeral_storage_request: 1000
  ephemeral_storage_limit: 2000
env:
  LOG_LEVEL: info
replicas: 1
workspace_fqn: "WORKSPACE_FQN_HERE"

Public access template (when expose: true)

ports:
  - port: 8000
    expose: true
    host: my-service-my-workspace.ml.your-org.truefoundry.cloud  # Auto-generate from cluster discovery
    app_protocol: http

Host is REQUIRED when expose: true. Auto-generate it: {service-name}-{workspace-name}.{base_domain}. Get base_domain from cluster discovery (see cluster-discovery.md).

Check deployment status

TFY_API_SH=~/.claude/skills/truefoundry-deploy/scripts/tfy-api.sh
bash $TFY_API_SH GET '/api/svc/v1/apps?workspaceFqn=WORKSPACE_FQN&applicationName=SERVICE_NAME'

Or use the applications skill.

Post-Deploy Monitoring (MANDATORY)

HARD RULE: After every successful tfy apply or tfy deploy command, you MUST monitor the deployment to completion. Do NOT stop after the apply/deploy command returns. Do NOT ask the user "should I monitor?" — just do it. Do NOT say "you can check the status" — YOU check the status. The deployment is not done until you confirm a terminal state.

Monitoring procedure

Immediately after deploy/apply succeeds, start polling. Do not wait for the user to ask.

Poll loop — execute this yourself, do not delegate to the user:

TFY_API_SH=~/.claude/skills/truefoundry-deploy/scripts/tfy-api.sh

# Run this in a loop with sleep between checks:
# Every 15s for first 2 min, every 30s for min 2-5, every 60s after that
# Timeout after 10 minutes
bash $TFY_API_SH GET '/api/svc/v1/apps?workspaceFqn=WORKSPACE_FQN&applicationName=SERVICE_NAME'

Or use MCP tool call if available:

tfy_applications_list(filters={"workspace_fqn": "WORKSPACE_FQN", "application_name": "SERVICE_NAME"})

How to check: The response is at data[0].deployment.currentStatus. Use state.isTerminalState as the authoritative check.

Terminal states (state.isTerminalState === true) — stop polling:

• DEPLOY_SUCCESS → report success, replicas, endpoint URL
• BUILD_FAILED, DEPLOY_FAILED, FAILED → fetch logs, diagnose, suggest fix (see below)
• PAUSED → report paused
• CANCELLED → report cancelled

Non-terminal states — keep polling, report progress each time:

• INITIALIZED → "Deployment initialized, waiting..."
• BUILDING (status) or transition BUILDING → "Build in progress..."
• BUILD_SUCCESS → "Build succeeded, deploying..."
• ROLLOUT_STARTED or transition DEPLOYING → "Deploying (X/Y replicas ready)..."
• DEPLOY_FAILED_WITH_RETRY → "Deploy failed, retrying..."

On success

1. Report final status and replicas (e.g., "2/2 ready")
2. Show endpoint URL if service has an exposed port
3. Run a quick HTTP health check if endpoint is available: curl -sf -o /dev/null -w '%{http_code}' "https://ENDPOINT_URL" || true

On failure

1. Fetch recent logs (last 5 minutes) using logs skill or direct API
2. Identify root cause from logs (OOMKilled, CrashLoopBackOff, ImagePullBackOff, port mismatch, probe failure, build error)
3. Follow deploy-debugging.md for diagnosis
4. Apply one fix and retry once; if still failed, report to user with summary and log excerpt and stop

On timeout (10 minutes)

Report current state and elapsed time. Do NOT silently give up — tell the user:

Monitoring timed out after 10 minutes. Current status: ROLLOUT_STARTED (transition: DEPLOYING).
The deployment is still in progress. You can re-run monitoring or check the TrueFoundry dashboard.

NEVER end your response after a deploy/apply command without reporting a terminal deployment status (state.isTerminalState === true). If you are about to end your response and you have not confirmed DEPLOY_SUCCESS, DEPLOY_FAILED, BUILD_FAILED, FAILED, PAUSED, or CANCELLED, you are violating this rule — go back and poll.

Post-Deploy Configuration (Ask After Success)

After deployment succeeds (DEPLOY_SUCCESS), ask the user about the following configuration options. Do not silently skip these — present them as a checklist and let the user decide.

1. Public vs Private URL

Ask the user:

Your service is deployed. How should it be accessed?
1. **Public URL** — Accessible from the internet (expose: true with a host)
2. **Private/Internal only** — Only accessible within the cluster (expose: false)

If the user picks public and the port doesn't already have expose: true + host, update the manifest and redeploy.

2. Authentication

Ask the user:

Do you want to add authentication to your service?
1. **No auth** — Anyone with the URL can access it
2. **TrueFoundry login** — Users must log in via TrueFoundry (truefoundry_oauth)
3. **JWT auth** — Verify JWT tokens from a custom identity provider
4. **Basic auth** — Username/password protection

If the user picks an auth option, add the appropriate auth block to the port configuration and redeploy.

3. Auto-shutdown vs Always Running

Ask the user:

Should the service auto-shutdown when idle?
1. **Always running** — Keep replicas up at all times (default)
2. **Auto-shutdown after idle** — Scale to zero after no requests for a period (saves cost)
   → Recommended wait_time: 900 seconds (15 min) for dev, longer for staging

If the user picks auto-shutdown, add the auto_shutdown block to the manifest:

auto_shutdown:
  wait_time: 900  # seconds of inactivity before scaling to zero

Skip these prompts if the user explicitly said they don't want changes, or if this is a redeploy of an existing service that already has these configured.

REST API fallback (when CLI unavailable)

See references/cli-fallback.md for converting YAML to JSON and deploying via tfy-api.sh.

Auto-Detection: Single vs Multi-Service

Before creating any manifest, scan the project:

1. Check for docker-compose.yml, docker-compose.yaml, or compose.yaml first. If present (or user mentions docker-compose), treat it as the primary source of truth: load deploy-multi.md and compose-translation.md, generate manifests from the compose file, wire services per service-wiring.md, then complete deployment. Do not ask the user to manually create manifests when a compose file exists.
2. Look for multiple Dockerfile files across the project
3. Check for service directories with their own dependency files in services/, apps/, frontend/, backend/

• Compose file present or user says "docker-compose" → Multi-service from compose: load deploy-multi.md + compose-translation.md
• Single service → Load references/deploy-service.md
• Multiple services (no compose) → Load references/deploy-multi.md

Multi-Service Deployment Order (MANDATORY)

HARD RULE: When deploying multiple services, you MUST deploy in dependency order, create secrets between tiers, and wire services before deploying dependents. Never deploy all services at once.

Tier-by-tier flow:

TIER 0: Infrastructure (DB, Cache, Queue) → deploy → wait for pods ready → create TFY secrets
TIER 1: Backend (APIs, workers) → deploy with secrets + DNS wiring → verify connectivity
TIER 2: Frontend / gateway → deploy with backend URLs → verify end-to-end

Key rules:

• Create TFY secret groups with infra credentials between Tier 0 and Tier 1 — never put raw passwords in manifests
• SPA frontends (React, Vue) MUST use backend's public URL, not internal DNS
• DEPLOY_SUCCESS does NOT mean Helm pods are ready — poll actual readiness
• Present the dependency graph and deploy plan to the user before deploying

For step-by-step orchestration, examples, and common patterns, see deploy-ordering.md. For dependency graphs, DNS wiring, and compose translation, see deploy-multi.md, service-wiring.md, and dependency-graph.md.

Secrets Handling (MANDATORY: Always Use TFY Secrets)

HARD RULE: NEVER put sensitive values directly in the manifest env block. ALWAYS create a TrueFoundry secret group first, then reference the secrets using tfy-secret:// format. This is non-negotiable — even for "quick" or "test" deployments.

Workflow for any env var that looks sensitive (matches *PASSWORD*, *SECRET*, *TOKEN*, *KEY*, *API_KEY*, *DATABASE_URL*, *CONNECTION_STRING*, *CREDENTIALS*, or any value the user explicitly says is sensitive):

1. Ask the user for the secret values (or confirm they want to store them)
2. Create a secret group using the secrets skill: # Use the secrets skill to create a group with the sensitive keys # The skill will handle creating the group and individual secrets
3. Reference them in the manifest with tfy-secret:// format:

env:
  LOG_LEVEL: info                                              # plain text OK
  DB_PASSWORD: tfy-secret://my-org:my-service-secrets:DB_PASSWORD  # sensitive — ALWAYS use tfy-secret://
  API_KEY: tfy-secret://my-org:my-service-secrets:API_KEY          # sensitive — ALWAYS use tfy-secret://

Pattern: tfy-secret://<TENANT_NAME>:<SECRET_GROUP_NAME>:<SECRET_KEY> where TENANT_NAME is the subdomain of TFY_BASE_URL.

If the user provides a raw secret value in the manifest or asks you to put it directly in env:

1. Warn them: "Secrets should not be stored as plain text in manifests."
2. Offer to create a TFY secret group for them
3. Only proceed with raw values if the user explicitly insists after the warning

Use the secrets skill for guided secret group creation. For the full workflow, see references/deploy-service.md (Secrets Handling section).

File Mounts (Config, Secrets, Shared Data)

When users ask to mount files into a deployment, prefer manifest mounts over Dockerfile edits:

• type: secret for sensitive file content (keys, certs, credentials)
• type: config_map for non-sensitive config files
• type: volume for writable/shared runtime data

See references/deploy-service.md (File Mounts section) for the end-to-end workflow.

Shared References

These references are available for all workflows — load as needed:

Workflow-Specific References

Composability

• Find workspace: Use workspaces skill
• Monitor rollout: Use monitor skill to track deployment progress
• Check what's deployed: Use applications skill
• View logs: Use logs skill
• Manage secrets: Use secrets skill
• Deploy Helm charts: Use helm skill
• Deploy LLMs: Use llm-deploy skill
• Test after deploy: Use service-test skill

Success Criteria

• User confirmed service name, resources, port, and deployment source before deploying
• Deployment URL and status reported back to the user
• Deployment status verified automatically immediately after apply/deploy (no extra prompt)
• Health probes configured for production deployments
• Secrets stored securely (not hardcoded in manifests)
• For multi-service: all services wired together and working end-to-end