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, <<: *some-env), which only make sense after expansion.

An existing Okteto-aware compose is a strong signal. If a compose file already uses Okteto Compose extensions (endpoints:, public:) or is clearly an okteto deploy file, prefer it as the deploy: source and surface it to the user rather than authoring something new — the customer has already expressed how they want to deploy.

2.2 Compose → Okteto mapping

When docker-compose.yml is present, map fields like this:

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.

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

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

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.<svc>.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_<SERVICE>_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_<SERVICE>_IMAGE — full image reference (the one you almost always want)
• OKTETO_BUILD_<SERVICE>_REGISTRY, _REPOSITORY, _SHA — the parts, if you need them

<SERVICE> is the build: key uppercased, with - replaced by _ (so build.web-api → OKTETO_BUILD_WEB_API_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.<svc>.image: ${OKTETO_BUILD_<SVC>_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 upgrade --install myapp ./chart --set api.image=${OKTETO_BUILD_API_IMAGE}  # built api image wired in

dev:
  api:
    image: okteto/golang:1.23  # tag from api/go.mod (go 1.23)
    command: bash               # Go services usually want a shell to `go run` manually
    sync:
      - ./api:/usr/src/app      # per-service sync; tighten if you have large generated dirs

test:
  api:
    image: okteto/golang:1.23
    context: ./api
    commands:
      - go test ./...           # detected from the project's test layout

Example: Level 2 fragment (compose-driven deploy, no chart)

build:
  api:
    context: ./api
    dockerfile: Dockerfile

deploy:
  compose: docker-compose.yml   # deploy the stack straight from compose — no chart or k8s manifests needed

dev:
  api:
    image: okteto/golang:1.23   # tag from api/go.mod (go 1.23)
    command: bash
    sync:
      - ./api:/usr/src/app

Compose-as-deploy caveats. A compose file written for local development often won't deploy cleanly to a cluster as-is. Before committing to deploy: compose:, scan the compose file and warn the user about:

• Host-IP port bindings (e.g. 127.0.0.1:9229:9229) — Okteto's compose deploy rejects them (Host IP is not allowed). Drop the host-IP prefix (9229:9229).
• Source bind-mount volumes (e.g. ./api:/usr/local/app) — on a cluster these become empty volumes that shadow the image's application code, so the service crashes (can't open file 'app.py'). They belong in dev.<svc>.sync, not in a deployed compose.

If the compose file leans on these dev-only constructs, surface it up front — don't discover it at deploy time. In preference order:

1. Generate a deploy-ready compose — derive a compose.okteto.yaml from the original with the cluster-hostile parts removed, and point deploy: compose: compose.okteto.yaml at it. See "Generating a deploy-ready compose" below.
2. Reuse an existing deploy compose — if the repo already ships one (e.g. a *.preview* or production compose without the dev volumes/ports), point deploy: compose: at that.
3. Use the chart/k8s manifests for deploy: instead, if the repo has them. This is why a repo with both compose and a chart often wants the chart for deploy: even though compose drives dev: (Section 2.5).

Generating a deploy-ready compose (`compose.okteto.yaml`). When compose is the chosen deploy source but the file carries dev-only constructs, write a derived copy rather than editing the user's original. This is a mechanical transform of the user's own file — not a hand-authored deploy artifact — so show it to the user as a diff against the source. Reference it explicitly: deploy: { compose: compose.okteto.yaml }. Apply these transforms:

• Drop host-IP port prefixes — 127.0.0.1:9229:9229 → 9229:9229. Debug-only ports (inspectors) can be removed entirely; they belong in dev.<svc>.forward.
• Remove source bind-mount volumes — ./src:/app and the like. The image already contains the code; these belong in dev.<svc>.sync. Keep named/data volumes (e.g. db-data).
• Replace dev-container placeholders — a service whose entrypoint/command is sleep infinity won't run the app when deployed. Restore its real start command (from the Dockerfile CMD or a production compose), or drop the override.

Two things the transform cannot silently fix — flag these to the user rather than pretending the result is clean:

• Host-path config/script mounts (e.g. an init script mounted as the container's entrypoint) — removing the mount doesn't make the service work; the file has to be baked into an image. Leave it and warn.
• Anything else service-specific the scan can't reason about.

Add a header comment to compose.okteto.yaml: that it was derived from <original> by the skill, what changed, and that it's safe to edit or regenerate. The generated file is best-effort — still climb the Phase 5 ladder (validate → build → deploy) to confirm the stack actually comes up.

Example: Level 1 fragment (dev-only, single service)

dev:
  myapp:                        # service name from pyproject.toml [project] name
    image: okteto/python:3.12   # pyproject.toml declares requires-python >=3.12
    command: bash               # no FastAPI/Flask signal — shell is the safe default
    sync:
      - .:/usr/src/app          # full-repo sync; exclude .venv, __pycache__ if they grow

4.2 Refine (Phase 4)

Show the file to the user. In collaborative mode, ask:

"Here's the draft. Want me to change anything before we validate? Common edits: adjust sync paths to exclude node_modules/vendor/target, change the dev image version, add an env var, or swap the deploy command."

Common edit patterns:

Iteration is cheap — keep editing until the user is satisfied. Do not move to Phase 5 until they say "looks good" or equivalent.

In autonomous mode, skip the ask and move directly to Phase 5. Edits will be requested via the PR review.

5. Phase 5 — Validate (tiered)

Climb the validation ladder as far as the environment supports. Tier 1 is mandatory; Tiers 2 and 3 are opt-in.

5.1 Tier 1 — okteto validate (always)

Run:

okteto validate

What it catches: YAML syntax, schema violations, missing required fields. What it does NOT catch: wrong sync paths, missing services in deploy:, broken Helm refs, images that fail to build.

If it fails, treat as a Phase 4 issue and fix before continuing.

The skill does not finish without Tier 1 passing.

5.2 Tier 2 — okteto build (offered if Dockerfiles or build: exist)

Pre-check: run okteto context show. If it errors or returns no context, skip Tier 2 and inform the user:

"Skipping the build check — no Okteto context. The manifest is syntactically valid but I haven't proven the Dockerfiles build."

Otherwise, ask:

"I can run okteto build to prove every Dockerfile resolves and pushes. This takes a few minutes per service. Skip / build one service / build all?"

Default to build all. If the user has many services (≥ 4) and is in a hurry, offer narrowing.

Run:

okteto build

or

okteto build <service>

5.3 Tier 3 — okteto deploy --wait (offered if Tier 2 passed and user has a context)

Note: there is no `okteto deploy --dry-run` flag. Full deploy is the only Tier 3 option.

Ask:

"I can do a full deploy to verify the manifest works end-to-end. This will create resources in your namespace <ns>. After it succeeds, I'll show you the endpoints. You can okteto destroy after if you want. Proceed?"

Run:

okteto deploy --wait
okteto endpoints

On success: print endpoints. Do not run okteto destroy automatically — that's the user's call.

5.4 On failure at any tier

1. Surface the raw CLI error verbatim. Do not paraphrase.
2. Diagnose the likely cause based on the manifest section involved (e.g., a Helm error → deploy:; a build error → build: or the Dockerfile).
3. Propose a concrete edit to the manifest. Show the diff, not "you should change X."
4. After the user approves the fix:

• If the edit changed manifest structure (added/removed/renamed sections, changed YAML shape), re-run Tier 1 first to catch new schema problems, then re-run the failing tier.
• If the edit was purely a value change (image version, sync path, env var), re-run only the failing tier.

5.5 Final summary

Once the ladder is climbed (or stopped), summarize:

"✅ okteto validate passed ✅ okteto build passed for all services ⏭️ okteto deploy skipped (you opted out) Next: run /dev-setup or invoke the okteto skill to deploy and start developing."

If tiers were skipped due to environment (no context), say so:

"⚠️ Skipped Tiers 2 and 3 (no Okteto context). The manifest is syntactically valid but not deploy-tested."

6. Phase 6 — Handoff or PR

6.1 Collaborative mode: handoff

Point the user at the next step:

"The manifest is in place. To bring up your environment, you can: - Run /dev-setup for a guided deploy + dev mode - Or invoke the okteto skill in any future session — it'll pick up the manifest you just created If you change the manifest, re-run okteto validate (or come back to me)."

State which validation tiers were run and which were skipped (see Section 5.5).

Do NOT delete the manifest, run `okteto destroy`, or push to a remote. The work product is the file on disk.

6.2 Autonomous mode: PR

Create a branch, commit the manifest, push, and open a PR. The PR is the human review gate.

Before running the commands below, substitute every `<placeholder>` with a real value from your Phase 1 / Phase 5 results. Do not send literal <list>, <N>, <bulleted list of services...>, etc. to a real PR. For the validation checklist, use [x] if the tier passed and [ ] if it was skipped or failed.

git checkout -b okteto/onboarding
git add okteto.yaml
git commit -m "Add Okteto manifest

Generated by okteto-onboarding skill.
Discovered services: <list>
Scope level: <N>
Validation: <tiers passed>"
git push -u origin okteto/onboarding
gh pr create --title "Add Okteto manifest" --body "$(cat <<'EOF'
## Summary
This PR adds an Okteto manifest generated by the `okteto-onboarding` skill.

## Discovered services
<bulleted list of services with their Dockerfiles and dev commands>

## Scope level
Level <N> — <one-sentence rationale>

## Validation
- [<x or space>] `okteto validate` passed
- [<x or space>] `okteto build` ran for all services
- [<x or space>] `okteto deploy --wait` succeeded with endpoints

## Decisions to confirm
- [ ] Dev image picks (e.g., `okteto/golang:1.22`) match your toolchain
- [ ] Sync paths exclude appropriate generated dirs (node_modules, vendor, target)
- [ ] Helm/kubectl deploy command matches your usual workflow

🤖 Generated by the okteto-onboarding skill
EOF
)"

The skill never merges the PR. A human reviews and merges.

7. Operating modes

7.1 Collaborative (default)

A user is in the loop. Each phase that needs a decision asks a question. Defaults are presented but not auto-selected.

7.2 Autonomous (opt-in)

No human is expected to intervene. Inferred from context the same way the existing okteto skill does — for example, when invoked from a CI pipeline or a ticket-driven session.

In autonomous mode:

• Scope level → highest level the discovery supports. Level 2 requires a deploy source (a chart, k8s manifests, or a deployable compose file); Level 3 additionally requires detected tests. With no deploy source at all, stay at Level 1 even if tests are present. When both compose and a chart exist, pick the most conservative deploy source and note the choice in the PR.
• Validation tier → Tier 1 always; Tier 2 if okteto context show succeeds; Tier 3 only if the trigger explicitly authorizes a deploy (label, env var, or explicit instruction).
• Discovery ambiguities → pick the most conservative interpretation and note it in the PR description.
• Phase 6 → always opens a PR (Section 6.2), never hands off.

8. CLI quick reference

Never run okteto up from this skill — it's interactive and belongs to the existing okteto skill / the user. Never run okteto destroy — leave that to the user.

9. Common mistakes to avoid

• Triggering when an `okteto.yaml` already exists. Always do the pre-flight check from Section 1.
• Generating a Helm chart or k8s manifests. This skill does not author deploy artifacts. If the user has no chart, no k8s manifests, and no deployable compose file, recommend Level 1 and stop. A compose file alone is enough for Level 2 via deploy: compose: — don't drop to Level 1 just because there's no chart. (The one exception is a derived compose.okteto.yaml — a mechanical transform of the user's own compose, Section 4.1 — never a chart or k8s manifest written from scratch.)
• Assuming the chart drives `deploy:` when a compose file is also present. Ask which the user wants (Section 2.5); don't silently pick the chart.
• Recommending `deploy: compose:` without scanning for dev-only constructs. Host-IP port bindings and source bind-mount volumes break a cluster deploy (Section 4.1 caveats). Warn the user up front, don't surface it at deploy time.
• Reading only the root `docker-compose.yml`. Projects often keep the dev compose under docker/ or as *.dev.yml, separate from a published-images root compose. Glob for all compose files and follow include: (Section 2.1). Building dev: from a published-images compose produces a useless manifest.
• Putting every compose service in `dev:`. Backing services (databases, caches, object stores, brokers) are deploy-only infrastructure. dev: is for the services you edit (Section 2.2).
• Copying a `sleep infinity` entrypoint into `dev.<svc>.command`. That's a dev-container placeholder, not a real command — use bash (Section 2.4).
• Hard-coding a dev-image version. Read the version the repo declares (Section 2.3); only fall back to a default when none is declared, and flag the fallback.
• Forgetting to wire built images downstream. A build: entry isn't enough on its own for Helm/k8s deploys — pass ${OKTETO_BUILD_<SERVICE>_IMAGE} into the deploy command so the workload runs the image you just built.
• Skipping the framing block. The dev: vs deploy: framing in Section 3.1 must be shown to the user and written into the manifest as a header comment.
• Climbing the validation ladder without checking `okteto context show` first. Tiers 2 and 3 require a working context.
• Paraphrasing CLI errors on validation failure. Show the raw output; the user (or the next agent) needs to see exactly what Okteto said.
• Asking the user about everything. Trust the signals from Phase 1. Only ask when discovery is genuinely ambiguous.
• Merging the PR in autonomous mode. The PR is the human gate. Never merge.
• Recommending `okteto/samples/` templates. Those are demos. Build the manifest from discovered facts, not templates.