Okteto Onboarding

Okteto Onboarding Skill

1. Activation

This skill activates when all of the following are true:

• The project has no okteto.yaml or okteto.yml at the repo root
• The user is asking about Okteto, dev environments, or onboarding (e.g., "how do I get this onto Okteto", "set this repo up for Okteto", "create an Okteto manifest")

Do NOT activate if:

• An okteto.yaml or okteto.yml already exists at the repo root — that is the existing okteto skill's domain. Defer to it.
• The user is asking how to use Okteto with an existing manifest. Defer to the existing okteto skill.

Pre-flight check. Before starting Phase 1, run:

ls okteto.yaml okteto.yml 2>/dev/null

If anything is returned, stop and tell the user: > "I see okteto.yaml already exists. The okteto-onboarding skill is for repos without a manifest. For working with the existing manifest, use the okteto skill or run /dev-setup."

Two operating modes: collaborative (default — user is in the loop) and autonomous (no human; opens a PR). See Section 7. Most of the workflow is mode-agnostic; only Phase 6 and the resolution of "ask the user" branches differ.

2. Phase 1 — Discover

Build an internal model of the repo: services, their build contexts, ports, dev commands, and any existing deploy artifacts. Read signals in priority order. Do not write the manifest yet.

2.1 Signal priority

1. docker-compose.yml / compose.yaml — the richest signal for services and dev, and a valid deploy: source on its own. If present, it is the primary blueprint for build: and dev:, and Okteto can deploy it directly via deploy: compose: (Section 4.1) — no chart or k8s manifests required. See the mapping table below.
2. Existing Helm chart — any Chart.yaml under chart/, charts/, helm/, or deploy/. If found, deploy: will be a helm upgrade --install command. Do not generate or modify the chart.
3. Existing k8s manifests — .yaml files with kind: headers under k8s/, manifests/, deploy/ — but excluding any templates/ subdirectory (those belong to a Helm chart and are not directly applyable). If found, deploy: will be kubectl apply -f .... Do not author these manifests.
4. Per-service Dockerfiles — for repos without compose, each top-level Dockerfile is a candidate service. Service name comes from the parent directory.
5. Language manifests — package.json, go.mod, pom.xml, pyproject.toml, Gemfile, Cargo.toml. Used to pick the dev image and infer the dev command. When this is the only signal (no Dockerfile, no compose), name the service after the project's name field (e.g., pyproject.toml's [project] name, package.json's name) or, failing that, the repo directory.
6. Procfile / Makefile — secondary signals when language manifests are ambiguous.

These signals are complementary, not exclusive. A repo can carry more than one. Common combinations:

• Compose only — compose drives build:, dev:, and deploy: (via deploy: compose:). Enough to reach Level 2 with no chart or k8s manifests.
• Chart or k8s only — the artifact drives deploy:; derive dev: from the deployed workload names (Phase 3, bullet 4).
• **Compose and a chart/k8s — the deploy source is ambiguous. Compose files are often the dev path and charts the prod path, but not always. Do not assume the chart wins deploy:.** Ask the user which should drive it — compose, the chart/manifests, or both combined (Section 2.5).

Apply each signal to its own section of the manifest rather than treating priority order as a winner-takes-all.

Don't stop at the root compose file. Real projects routinely ship several: a base or production compose (often at the repo root, using published image:s) plus one or more development composes under docker/, dev/, or named *.dev.yml / compose.dev.yml. Glob for **/docker-compose*.y*ml and **/compose*.y*ml (skip node_modules, vendor, and build dirs) — not just ./docker-compose.yml. The split usually maps cleanly:

• The dev compose — the one that builds from source (build: with context:/target:, plus source bind mounts) — drives build: and dev:.
• A base / production / *.preview* compose can drive deploy:.

When more than one compose file exists, list them and ask which describes development and which is the deploy source (Section 2.5). Do not assume the root file is the dev blueprint — a root compose built entirely from published images is a run/deploy file, and developing against a published image is meaningless.

Follow include: and resolve anchors. A compose file may pull in others via include: (or the older extends:) — read those too, or you'll miss services. Parse the resolved YAML, not raw lines: dev composes commonly rely on anchors and x- extension fields (build: *some-anchor, ` | build. (when there's a build: block) and dev. | | build.context, build.dockerfile | build..context, build..dockerfile | | image (no build) | dev..image directly — **do not** add a build. entry for pre-built images | | ports (Level 1) | dev..forward — preserves the user's local-port intent | | ports (Level 2+) | exposed via the deploy: source (compose deploy:, Helm, or k8s); also keep dev..forward for active dev sessions | | command | starting point for dev..command | | volumes (host bind mounts only) | dev..sync candidates | | depends_on | ordering hint for the deploy step | | environment | dev..environment` |

Ignore compose-only concepts that don't translate (networks, restart, named volumes, extends).

Backing services vs. dev targets. A compose service that uses a published image: with no build: and no source in the repo (databases, caches, object stores, mail catchers, message brokers — postgres, redis, minio, nats, mailpit, and the like) is infrastructure: include it in the deploy: stack, but it rarely belongs in dev:. Reserve dev: for the services you actually edit — the ones that build from source in the repo. Don't create a dev: entry for every compose service by default; a repo with ten compose services often has only one or two real dev targets.

2.3 Dev-image picks by language manifest

Pick the image family from the language manifest, and the tag from the version the repo declares — do not hard-code a version. Reading the declared version keeps the pick current instead of freezing a minor that goes stale.

| Language manifest | Image family | Where to read the version | |---|---|---| | package.json | okteto/node | engines.node | | go.mod | okteto/golang | the go 1.x line | | pom.xml / build.gradle* | okteto/maven or okteto/gradle | maven.compiler.release / sourceCompatibility, or the wrapper version | | pyproject.toml / requirements.txt | okteto/python | requires-python or .python-version | | Gemfile | okteto/ruby | .ruby-version or the ruby directive | | Cargo.toml | okteto/rust | rust-version (MSRV), if set |

Tag the image with the declared version, e.g. okteto/golang:1.23 when go.mod says go 1.23. State the source in an inline comment (# go.mod declares Go 1.23).

When the repo declares no version, do not invent a specific minor — it will be wrong as often as right. In collaborative mode, ask the user which version they target. In autonomous mode, pin to the family's current stable major and flag it in the PR as a decision to confirm (# go.mod has no version line — defaulting to a recent stable; confirm).

2.4 Dev-command picks

| Signal | Dev command default | |---|---| | package.json with scripts.dev | npm run dev | | package.json with scripts.start (no dev) | npm start | | go.mod | bash (Go projects usually want a shell to go run manually) | | pom.xml with spring-boot-maven-plugin | mvn spring-boot:run | | pyproject.toml with FastAPI/Flask | bash (varies too much to default) | | Procfile with a web: line | the value of web: | | compose entrypoint/command is sleep infinity or tail -f /dev/null | bash — this is a dev-container placeholder (the container stays alive so you exec in), not a real command. Don't copy it into dev..command; it's exactly the pattern Okteto's dev: replaces. | | None of the above | bash |

2.5 When discovery is ambiguous

If discovery leaves real ambiguity, ask one targeted question at a time rather than guessing. Examples:

• "I see Dockerfiles in api/ and web/ but no compose file. Should both be services?"
• "I found a Helm chart at chart/ and another at infra/helm/. Which one is the canonical deploy?"
• "This repo has both a docker-compose.yml and a Helm chart. Which should drive deploy: — the compose file, the chart, or both? (Compose is often the dev path and the chart the prod path, but I don't want to assume.)"
• "I see several compose files — docker-compose.yml at the root and docker/docker-compose.dev.yml. Which one describes how you develop (builds from source), and which should I use to deploy the stack?"
• "Your package.json doesn't have a scripts.dev — what command starts your dev server?"

Do not ask the user about everything. Only ask when a guess would be likely wrong. Trust the signals.

In autonomous mode, do not ask. Pick the most conservative interpretation, proceed, and note the ambiguity in the PR description (Section 7.2 covers this).

2.6 Output of Phase 1

Internally, you should now have a model like:

Services: [api, web, worker]
api:    Dockerfile @ ./api,    port 8080, dev cmd `bash`,        dev image `okteto/golang:1.22`
web:    Dockerfile @ ./web,    port 3000, dev cmd `npm run dev`, dev image `okteto/node:20`
worker: Dockerfile @ ./worker, no port,   dev cmd `bash`,        dev image `okteto/python:3.12`
Deploy: helm chart at ./chart
Tests: detected `go test ./...` in api, `npm test` in web

Show this summary to the user before moving to Phase 2.

3. Phase 2 — Negotiate scope

Before drafting, frame the choice and pick a level on the adaptive ladder.

3.1 The framing block (always show this to the user)

> In Okteto, deploy: describes how to provision the environment — from a Docker Compose file, a Helm chart, or k8s manifests. dev: describes how to live-edit a running service — file sync, the dev image, the startup command. You can use Okteto with just dev: if you already have a way to deploy your stack, or have Okteto handle both.

This same framing goes at the top of the generated okteto.yaml as a header comment.

3.2 The adaptive ladder

| Level | What it produces | Requires | |---|---|---| | 1 — dev-only | dev: section only. User runs their own deploy externally. | Nothing extra. | | 2 — deploy + dev | build: + deploy: + dev:. okteto deploy brings up the stack. | A Helm chart, k8s manifests, or a Docker Compose file in the repo. | | 3 — full lifecycle | Adds test: containers wired to existing test commands. | Level 2 prereqs PLUS detected tests. |

3.3 Recommendation logic

Recommend a level based on what Phase 1 found:

A deploy source is a Helm chart, k8s manifests, or a deployable Docker Compose file.

• A deploy source found AND tests detected → recommend Level 3
• A deploy source found, no tests → recommend Level 2
• No deploy source at all (no chart, no k8s manifests, no compose) → recommend Level 1 (and explain why higher levels are unavailable)

When more than one deploy source exists (e.g. compose and a chart), confirm which one drives deploy: (Section 2.5) before drafting — don't default to the chart.

Then ask the user (collaborative mode) or accept the recommendation (autonomous mode):

> "Based on what I found, I'd recommend Level [N]. Want to go with that, pick a different level, or have me explain the trade-offs?"

3.4 Locking the level

Once chosen, the level is locked for the rest of the session. Do not negotiate level mid-flight. If the draft turns out to need a different level (e.g., the chart is broken), surface that as a Phase 5 failure and re-enter Phase 2 cleanly.

4. Phases 3–4 — Draft and refine

4.1 Draft (Phase 3)

Write okteto.yaml to the repo root. The user must see the actual file, not just a summary.

Required content of every draft:

1. Header comment block — the framing from Section 3.1, plus a note that the file was generated by okteto-onboarding and the user is expected to edit it. Include the chosen scope level and a one-line rationale.
2. build: section (Level 2+) — one entry per service with context and dockerfile. Skip pre-built images here (they go directly under dev..image).
3. deploy: section (Level 2+) — points at the chosen deploy source:

• Compose: deploy: { compose: docker-compose.yml } (optionally a services: subset). The simplest path when the repo has no chart or manifests.
• Helm: a helm upgrade --install command, passing built images via --set ...=${OKTETO_BUILD__IMAGE} (see "Wiring built images" below).
• k8s manifests: a kubectl apply -f ... command.

See the schema examples below.

1. dev: section (always) — one entry per service the user wants in dev mode, with image, command, and sync. The key name must match the workload Okteto will deploy:

• Compose repo: the compose service name (mapping table, Section 2.2).
• Helm/k8s repo (no compose): the name of the Kubernetes workload the chart/manifests create — the Deployment/StatefulSet metadata.name (render the chart or read the manifest to find it). A dev: key matching no deployed workload won't attach.

The image, command, and sync come from the Dockerfile + language manifest (Sections 2.3–2.4) regardless of repo shape.

1. test: section (Level 3) — one entry per detected test command. See the schema example below.

Sync-path defaults:

• Single-service repo (one Dockerfile or one language manifest at the root): .:/usr/src/app is a reasonable default. If the Dockerfile sets WORKDIR, use that path on the right side instead of /usr/src/app.
• Multi-service repo (per-service Dockerfiles or compose with multiple services): scope each service's sync to its own subdirectory, e.g., ./api:/usr/src/app. Syncing the whole repo into every container is almost always wrong.

Inline comments are required on every non-obvious choice (image picks, command picks, sync paths, forward: ports). Image and command picks are non-obvious by default — comment them.

Wiring built images into deploy: and dev:. Images defined under build: are pushed to the Okteto Registry and exposed to the rest of the manifest as environment variables:

• OKTETO_BUILD__IMAGE — full image reference (the one you almost always want)
• OKTETO_BUILD__REGISTRY, _REPOSITORY, _SHA — the parts, if you need them

` is the build: key uppercased, with - replaced by (so build.web-api → OKTETOBUILDWEBAPI_IMAGE`). Use these so the deployed workload runs the freshly built image instead of a stale or hard-coded one:

• Helm: --set api.image=${OKTETO_BUILD_API_IMAGE}
• k8s manifests: substitute the value (e.g. envsubst) before kubectl apply, or template the image field.
• Compose: the build: section already overrides the matching compose image: by service name — no --set needed.
• dev:: set dev..image: ${OKTETO_BUILD__IMAGE} when the user wants to develop against the built application image rather than a generic toolchain image (Section 2.3).

Example: Level 2/3 fragment (build + deploy + dev + test)

build:
  api:
    context: ./api
    dockerfile: Dockerfile  # using the existing Dockerfile

deploy:
  - name: Deploy chart
    command: helm upg

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [okteto](https://github.com/okteto)
- **Source:** [okteto/okteto-agent-skills](https://github.com/okteto/okteto-agent-skills)
- **License:** Apache-2.0

Install and usage instructions live in the source repository linked above.