Inference Gateway Skills Adl

ADL (Agent Definition Language) Expert

Use this skill when working in any project whose root contains an agent.yaml manifest with apiVersion: adl.inference-gateway.com/v1 (or v2+ when it ships) together with an .adl-ignore file. That pair is the fingerprint of a project generated by adl-cli from a schema in inference-gateway/adl. Also use it when the user is authoring a fresh manifest (adl init) or planning the domain model for a new agent.

Mental model

ADL is "OpenAPI for AI agents". A single declarative manifest names everything the agent needs - capabilities, the AI provider, services, function-call tools, markdown skills, server/auth, language runtime, sandbox, deployment - and adl-cli turns it into an enterprise-ready, A2A-compatible project. You own the manifest and the bodies of TODO placeholders; the CLI owns everything else and will regenerate it on demand.

The contract has two halves you must keep straight:

If you forget which half a file belongs to, cat .adl-ignore - anything matched there survives adl generate --overwrite.

Spec at a glance. Every spec.* top-level field the v1 schema defines:

The sections below cover each in turn. Anything not in this table is not in v1 - if you see it in an existing manifest, treat it as a CLI extension and verify against adl-cli's changelog.

Schema-first / domain-first workflow

ADL is schema-driven by design: the manifest is the domain model. Sequence work so the manifest leads and the code follows.

1. Model the domain in YAML before touching any source file.
   • Name the bounded context in metadata.name (lowercase, hyphenated) and pin metadata.version (semver - ^\d+\.\d+\.\d+$).
   • Declare the required spec frame first: spec.capabilities (all three booleans - streaming, pushNotifications, stateTransitionHistory), spec.server.port, and at least one spec.language.{go|typescript|rust} target. Almost always set spec.agent (provider + model) and spec.card (A2A discovery) too - they're optional in the schema but the agent is useless without them.
   • Enumerate services (spec.services.*) as ports: each gets a type (service/repository/client/middleware), an interface, a factory, and a description. One service per responsibility - don't conflate database and cache.
   • Enumerate config sections (spec.config.*) as value objects. Use dotted-name injection (config.database) to give a tool only the subsection it needs.
   • Enumerate tools (spec.tools[]) as commands - the verbs the model can call. Each one declares its JSON Schema (schema:) and the services it depends on (inject:).
   • Enumerate skills (spec.skills[]) as the procedural knowledge - the "how" and "when" for using the tools. Skills are markdown, not code.
2. Validate the manifest before generating. adl validate agent.yaml checks shape against the pinned schema and rejects typos in reserved namespaces (e.g. spec.config.tools.bash.tymeout_seconds).
3. Generate. adl generate --file agent.yaml --output . writes the scaffold. The first run creates .adl-ignore with every file containing a TODO marked for protection. Re-run with --overwrite to refresh non-protected scaffolding after a manifest edit.
4. Implement TODO bodies. Each generated tools/<name>.{go,rs,ts} and each internal/<service>/*.{go,rs,ts} ships with // TODO: markers spelling out what to replace. The factory signature and the handler signature are locked - don't rename them, or you'll desync from the generator's expectations on the next regeneration of main.go.
5. Test what you implemented (see "Testing custom tools" - this is the non-negotiable part).
6. Run. task build && task run (Go/Rust) or the equivalent in the chosen language. The generated Taskfile.yml is the entry point.

Capabilities, the LLM, and the agent card

Three top-of-spec blocks shape what the agent advertises, which model it talks to, and how clients discover it. Set them before anything domain- specific.

spec.capabilities (required). All three booleans must be present - the validator rejects the manifest if any is missing:

capabilities:
  streaming: true # SSE-based streaming responses
  pushNotifications: false # webhook callbacks on long-running tasks
  stateTransitionHistory: true # record task state transitions for replay

spec.agent (optional but near-universal). The LLM the generated agent defers to. Provider is a fixed enum: openai, anthropic, ollama, deepseek, google, mistral, groq, or "" for "configure at runtime via env vars only". Temperature is bounded to 0-2; maxTokens must be ≥1.

agent:
  provider: deepseek
  model: deepseek-v4-flash
  systemPrompt: |
    You are a helpful A2A agent. Use the AVAILABLE SKILLS playbooks for
    workflows; call tools for deterministic actions.
  maxTokens: 4096
  temperature: 0.3

spec.card (optional). Static fields for the A2A agent card served at /.well-known/agent-card.json. Used by other agents and the Inference Gateway registry for discovery. Fields are all free-form strings/arrays (no enum constraints in v1):

card:
  protocolVersion: "0.3.0"
  preferredTransport: JSONRPC
  defaultInputModes: [text, voice]
  defaultOutputModes: [text, audio]
  url: "https://my-agent.example.com:8443"
  documentationUrl: "https://github.com/company/my-agent/docs"
  iconUrl: "https://github.com/company/my-agent/icon.png"

Server, auth, and language targets

spec.server (required). Only port is mandatory. scheme (http/ https), debug, and auth.enabled are optional. v1 only models the auth on/off toggle - specific providers (OIDC, JWT, ...) live in the generated code and config sections, not the schema:

server:
  port: 8443
  scheme: https
  debug: false
  auth:
    enabled: true

spec.language (required). At least one of three children; the generator emits the corresponding scaffold. Required pairs per language:

language:
  go:
    module: github.com/example/my-agent
    version: "1.26.2"

vendor.{deps,devdeps} (CLI extension, not in v1 schema). adl-cli reads spec.language.<lang>.vendor.deps[] (runtime) and vendor.devdeps[] (code generators / dev tools) for Go and Rust to pre-populate go.mod / Cargo.toml. The v1 JSON Schema doesn't validate these keys, but it permits additional properties so they round-trip cleanly:

language:
  go:
    module: github.com/example/my-agent
    version: "1.26.2"
    vendor: # CLI extension
      deps:
        - github.com/stretchr/testify@v1.10.0
      devdeps:
        - golang.org/x/tools/cmd/stringer@v0.20.0

Verify against the latest adl-cli changelog before relying on it; treat schema-validated fields as load-bearing and extensions as best-effort.

Tools vs Skills (the often-confused distinction)

ADL distinguishes two complementary surfaces, and conflating them produces unmaintainable agents. Apply the rule first, then write the entry.

A skill that needs to read files (e.g. its own SKILL.md body, or bundled templates) requires - id: read in spec.tools and spec.config.tools.read.enabled: true. The validator enforces this; don't disable it.

Reserved built-in tools

spec.tools[] recognises five reserved ids that map to framework-supplied implementations. They ship with their own unit tests (see builtin/*_test.go.tmpl in the CLI templates), so you do not need to write tests for them. You activate them - that's all:

All five default to enabled: false. Opt in by listing the id alone (no name, description, or schema - the generator owns those) and setting enabled: true in the matching spec.config.tools.<id> block. The reserved config block accepts only the typed keys the generator knows; typos like tymeout_seconds fail validation by design.

Resolution precedence at runtime is env > compile-time literal > built-in default (disabled). The kill-switch envs (A2A_BASH_DISABLED=1, A2A_FETCH_DISABLED=1, etc.) override the compile-time enabled: true.

Each built-in accepts its own typed config keys under spec.config.tools.<id>. Anything not on this list fails validation:

A representative configuration that opts in the three tools with the richest config surfaces:

config:
  tools:
    read:
      enabled: true
      max_lines: 2000
    bash:
      enabled: true
      whitelist: [ls, cat, grep, find, rg, jq, wc, head, tail, git, go]
      timeout_seconds: 30
    fetch:
      enabled: true
      allowed_domains:
        - pkg.go.dev
        - .rust-lang.org # any subdomain of rust-lang.org
        - raw.githubusercontent.com
      max_bytes: 5242880 # 5 MiB
      timeout_seconds: 20
      allow_downloads: true
      download_dir: /tmp/adl-fetch-cache
tools:
  - id: read
  - id: bash
  - id: fetch

Custom tool implementation

A user tool is a full spec.tools[] entry - id, name, description, optional tags, inject, and a JSON Schema for schema. The generator produces tools/<name>.{go,rs,ts} with:

• a struct holding the injected dependencies (logger, services, optional config or config.<section> subsections),
• a constructor New<PascalName>Tool(...) server.Tool whose signature mirrors inject: in declaration order, and
• a handler method <PascalName>Handler(ctx, args) (string, error) whose body is a single // TODO: Implement <name> logic comment plus a placeholder return.

You replace the handler body. Do not change:

• the struct name or its field order (regeneration of main.go wires fields positionally),
• the constructor signature,
• the handler method name.

Injection patterns you can use in inject::

inject:
  - logger # always available, *zap.Logger (Go)
  - config # the whole *config.Config
  - config.database # only *config.DatabaseConfig - principle of least privilege
  - cache # any name declared in spec.services

Always prefer config.<section> over config - it keeps the tool's blast radius small, makes unit tests trivial to set up, and the validator catches mismatches between inject: and spec.config.* at generate time.

Testing custom tools (mandatory)

The built-in reserved tools (read, bash, write, edit, fetch) ship with generated unit tests. Custom tools do NOT. This is the single most common gap in ADL projects.

Verification step for every PR that adds or modifies spec.tools[]:

1. For each entry whose id is not in the reserved set above, confirm a sibling test file exists - tools/<name>_test.go (Go), tools/<name>.rs tests block (Rust), tools/<name>.test.ts (TypeScript) - and the file is listed in .adl-ignore so regeneration won't clobber it.
2. The test file MUST exercise the handler against a service mock for every dependency in inject:. Use the service's interface (declared in spec.services.*.interface) - that's why it exists.
3. Use table-driven tests (Go) / parameterised tests (Rust) / describe+it blocks (TypeScript) covering at least: happy path, invalid args, service error.
4. Run the language-native test command from the generated Taskfile: task test (Go and Rust) or the equivalent target. CI must pass.

A minimal Go template for a custom tool test (adapt to the actual struct name and interface):

package tools

import (
    "context"
    "errors"
    "testing"

    "go.uber.org/zap"
    "github.com/stretchr/testify/require"
)

type stubDatabase struct{ err error; out string }

func (s *stubDatabase) Query(ctx context.Context, sql string) (string, error) {
    return s.out, s.err
}

func TestQueryDatabaseTool_Handler(t *testing.T) {
    cases := []struct {
        name    string
        args    map[string]any
        db      *stubDatabase
        wantErr bool
    }{
        {"happy path", map[string]any{"query": "SELECT 1", "table": "t"}, &stubDatabase{out: `{"rows":1}`}, false},
        {"missing query", map[string]any{"table": "t"}, &stubDatabase{}, true},
        {"backend error", map[string]any{"query": "SELECT 1", "table": "t"}, &stubDatabase{err: errors.New("boom")}, true},
    }
    for _, tc := range cases {
        t.Run(tc.name, func(t *testing.T) {
            tool := &QueryDatabaseTool{logger: zap.NewNop(), database: tc.db}
            _, err := tool.QueryDatabaseHandler(context.Background(), tc.args)
            if tc.wantErr {
                require.Error(t, err)
            } else {
                require.NoError(t, err)
            }
        })
    }
}

Skip this and the agent ships untested business logic - the schema can't catch a misread args["query"].(string) or a swallowed service error.

The .adl-ignore contract

adl-cli writes this file on first generation. It works like .gitignore but applies to the generator - anything matched is preserved across adl generate --overwrite. The generator automatically adds:

• every custom tool file (tools/<name>.<ext>),
• every custom service implementation (internal/<service>/*),
• every bare: true skill directory (skills/<id>/).

You can extend it to protect anything else you've hand-edited:

# .adl-ignore
Dockerfile          # custom container build
k8s/                # using ArgoCD overlays elsewhere
Taskfile.yml        # extended with project-specific targets
README.md           # written by hand

Patterns: # for comments, trailing / for directories, * wildcards, exact paths.

When a regeneration goes wrong: check .adl-ignore first. A missing entry there is the usual cause of "my custom code disappeared after I edited the manifest."

Services and configuration

Services are first-class - declare them in spec.services with type, interface, factory, and description, and the generator creates internal/<service>/<service>.{go,rs} with an interface stub and a factory that takes (*zap.Logger, *config.Config).

type is a closed enum - pick the one that best names the role so the generator can emit idiomatic scaffolding:

interface and factory must match ^[a-zA-Z][a-zA-Z0-9_]*$ - they become real identifiers in the generated code. Three rules keep this maintainable:

1. One interface per responsibility. Don't ship a UtilService. Split database and cache even if they share a backend.
2. Inject the interface, never the implementation. Tools depend on database.DatabaseService, not on *database.databaseService. This is what makes the unit-test stubs above trivial.
3. Configuration mirrors services. A service named googleCalendar reads its settings from spec.config.googleCalendar -> generated as GoogleCalendarConfig with GOOGLE_CALENDAR_* env prefix. Don't hand-wire env vars; let the prefix mapping do it.

Skills inside an ADL agent

spec.skills[] accepts three entry shapes, resolved by adl-cli:

source: shorthand:

- id: skill-creator
  source: skill-creator # inference-gateway/skills, main
- id: skill-creator
  source: skill-creator@v1.0 # pinned tag
- id: pdf
  source: anthropics/skills/pdf # different repo
- id: pdf
  source: anthropics/skills/pdf@abc1234 # pinned commit SHA
- id: custom
  source: https://github.com/my-org/my-repo/tree/release/path/to/skill

At runtime, the generated agent walks first-level subdirectories under skills/, parses each <id>/SKILL.md's frontmatter, and appends an AVAILABLE SKILLS: block to the system prompt - the bodies are not inlined. The model loads them on demand via the read tool, so a skills-using agent must opt read in (see "Reserved built-in tools").

license: on a skill entry must be one of the SPDX identifiers the schema accepts (MIT, Apache-2.0, BSD-2-Clause, BSD-3-Clause, GPL-2.0, GPL-3.0, LGPL-2.1, LGPL-3.0, MPL-2.0, ISC, CC0-1.0, CC-BY-4.0, CC-BY-SA-4.0, Unlicense) or the literal Proprietary. SPDX expressions like MIT OR Apache-2.0 are not currently accepted.

Local development: sandbox and AI assistants

spec.development configures the experience of working on the agent locally - reproducible dev environments and AI-assistant onboarding files. Two independent subsections, both optional.

spec.development.sandbox. Three alternative packagings - pick any combination; each declares its own enabled boolean. Generated artefacts:

spec.development.ai. Per-agent toggles for the coding assistants the project is meant to be edited with. Each is independent; all default off. The CLI generates each agent's onboarding file the first time it's enabled, and keeps it in sync on subsequent adl generate --overwrite runs unless you list it in .adl-ignore:

A small example combining sandbox + AI toggles:

development:
  sandbox:
    flox:
      enabled: true
    devcontainer:
      enabled: false
    dockerCompose:
      enabled: true
  ai:
    claudecode:
      enabled: true
    codex:
      enabled: true

spec.development.deps[] (CLI extension, not in v1 schema). adl-cli v0.38.0+ reads a cross-cutting list of sandbox-level tool dependencies to install across every enabled sandbox flavour. Treat it as a CLI convention - useful, but not schema-validated.

SCM, CI/CD, and deployment

Two related blocks that together drive everything outside main.go: where the code lives, how it ships, and where it runs.

spec.scm. All fields optional; the provider enum is closed:

spec.deployment. Choose type: kubernetes or type: cloudrun; the matching sub-block carries the detail. Both share an image shape with registry, repository, tag, optional useCloudBuild.

type: kubernetes generates k8s/deployment.yaml:

deployment:
  type: kubernetes
  kubernetes:
    image:
      registry: ghcr.io
      repository: example/my-agent
      tag: v1.0.0

type: cloudrun generates cloudrun/ helpers plus a deploy target in the Taskfile.yml. Fields map 1:1 to Cloud Run concepts:

deployment:
  type: cloudrun
  cloudrun:
    image:
      registry: gcr.io
      repository: my-agent
      tag: v1.0.0
      useCloudBuild: true
    resources:
      cpu: "1"
      memory: 512Mi
    scaling:
      minInstances: 0
      maxInstances: 100
      concurrency: 1000
    service:
      timeout: 3600
      allowUnauthenticated: true
      serviceAccount: my-agent@PROJECT_ID.iam.gserviceaccount.com
      executionEnvironment: gen2
    environment:
      LOG_LEVEL: info
      ENVIRONMENT: production

adl generate exposes equivalent CLI flags (--ci, --cd, --deployment kubernetes|cloudrun, --flox, --devcontainer) that OR with the manifest values. Prefer the manifest for anything that needs to be reproducible across machines and CI runs - treat the flags as one-off escape hatches.

Artifacts and post-generate hooks

Two small spec blocks that round out the manifest.

spec.artifacts.enabled. Set true to generate an artifacts server - a small HTTP service for storing task outputs - and to wire the matching create_artifact / read_artifact / list_artifacts helpers into the generated code. The backend (filesystem vs MinIO) is chosen at runtime via env vars on the generated binary, not via the manifest. When spec.development.sandbox.dockerCompose.enabled: true, the MinIO instance is wired into the generated docker-compose.yaml (adl-cli v0.39.0+).

artifacts:
  enabled: true

spec.hooks.post. A list of shell commands the CLI runs at the end of every adl generate. Use for the things a generator can't reasonably infer:

hooks:
  post:
    - go mod tidy
    - go generate ./...

Common commands

adl init is for bootstrapping a fresh agent.yaml; adl validate / adl generate are the steady-state loop. Run from the repo root:

adl init [name]                              # interactive wizard for a new agent.yaml
adl init [name] --defaults                   # non-interactive; accept the defaults
adl validate [file]                          # shape check; default file is agent.yaml
adl generate -f agent.yaml -o .              # initial scaffold
adl generate -f agent.yaml -o . --overwrite  # refresh after manifest edit (respects .adl-ignore)
adl generate ... --offline                   # skip skills registry; require local cache (~/.adl/skills-cache/)
adl generate ... --ci                        # also write .github/workflows/ci.yml
adl generate ... --cd                        # also write cd.yml + .releaserc.yaml
adl generate ... --deployment kubernetes     # also write k8s/deployment.yaml
adl generate ... --deployment cloudrun       # also write cloudrun/ helpers + Taskfile deploy target
adl generate ... --flox                      # enable Flox sandbox (OR with spec.development.sandbox.flox.enabled)
adl generate ... --devcontainer              # enable DevContainer sandbox (OR with spec.development.sandbox.devcontainer.enabled)
adl generate ... -t minimal                  # template selector (currently only "minimal" ships)

Once generated:

task build                                   # compile
task test                                    # run unit tests - INCLUDING custom tool tests
task run                                     # start the agent on spec.server.port
task deploy                                  # only present when spec.deployment is set

Validate the manifest in CI on every PR. Add a job step that runs adl validate agent.yaml before any code build - it's faster and surfaces schema drift early. Keep the manifest as the source of truth - prefer spec.scm.ci: true over --ci, spec.deployment.type: cloudrun over --deployment cloudrun, and so on. The flags exist for one-off enabling; they are not a substitute for declarative state.

Cross-repo awareness

ADL touches several repos in the inference-gateway org. When a change in one of them might ripple, surface it explicitly:

• inference-gateway/adl (schema) -> consumers must bump the pinned tag in their Taskfile.yml/internal/schema/. Breaking changes go to v2/, not on top of v1/.
• inference-gateway/adl-cli (generator) -> template edits land here, not in the schema. A generator change without a schema change means existing manifests keep working.
• inference-gateway/skills (this catalog) -> if a spec.skills[] entry uses source: <id> without an owner, it resolves here. Pin via @<tag> for reproducibility.

For ecosystem-wide concerns (release flow, conventional commits, cross-repo checklists), defer to the maintainer skill rather than restating it.

What NOT to do

• Don't hand-edit generated files that are not in .adl-ignore - they're overwritten on the next adl generate --overwrite. If you need them custom, add them to .adl-ignore (and accept that you've forked them from the generator).
• Don't rename a custom tool's struct, constructor, or handler. Rename via the manifest (spec.tools[].name) and regenerate.
• Don't duplicate a reserved built-in id (read, bash, write, edit, fetch) as a custom tool. The generator owns those.
• Don't inject config when you only need one section. Use config.<section> - smaller blast radius, simpler tests.
• Don't make schema-breaking changes in schema/v1/. They go to a new major (schema/v2/) per the additive contract.
• Don't ship a custom tool without a <name>_test.go (or language equivalent). The reserved built-ins are tested upstream; yours are not.
• Don't conflate ADL with ADK. ADL is the manifest format; the ADK (inference-gateway/adk) is the Go runtime the generated main.go imports.
• Don't omit spec.capabilities. All three booleans (streaming, pushNotifications, stateTransitionHistory) are required by the v1 schema; adl validate rejects the manifest if any are missing.
• Don't hand-edit generated CLAUDE.md, AGENTS.md, or GEMINI.md. They're re-emitted by the spec.development.ai.* generators on every adl generate --overwrite unless you add them to .adl-ignore. If you want the AI-onboarding docs to stay hand-written, ignore the file and accept that you've forked it from the generator.

Source: inference-gateway/skills — distributed by TomeVault.