Skelm

skelm Skill

When to activate

Activate when the user is working in a skelm project, wants to author or modify a *.pipeline.mts / *.workflow.mts file, asks about AgentPermissions, skelm.config.ts, MCP server wiring, or wants to run, inspect, or operate a pipeline.

The unit of work

A pipeline is a TypeScript file that exports a pipeline() call. It has:

• id — stable string identifier
• description — optional human-readable summary
• input / output — Zod schemas (validated at run boundaries)
• steps — ordered array of Step values
• optional finalize — transforms accumulated step outputs into the declared output shape

Step kinds: code | llm | agent | parallel | forEach | branch | loop | wait | pipelineStep | idempotent

Import everything from 'skelm' (or '@skelm/core' in library contexts — 'skelm' re-exports '@skelm/core').

Authoring a pipeline

Minimal example — one deterministic step:

import { code, pipeline } from 'skelm'
import { z } from 'zod'

export default pipeline({
  id: 'greet',
  description: 'Greets a user by name.',
  input: z.object({ name: z.string().min(1) }),
  output: z.object({ greeting: z.string() }),
  steps: [
    code({
      id: 'build-greeting',
      run: (ctx) => {
        const { name } = ctx.input as { name: string }
        return { greeting: `hello, ${name}` }
      },
    }),
  ],
})

Access prior step outputs via ctx.steps['step-id']. Access run metadata via ctx.run.runId.

For multi-step composition, LLM inference, and control flow, see references/pipeline-authoring.md.

Adding an LLM step

llm() is for single-shot inference — chat-completion-style requests with optional structured output. The recommended backend is 'openai' (the OpenAI factory talks to anything that exposes the OpenAI Chat Completions shape: hosted OpenAI, vLLM, llama.cpp, sglang, ollama with /v1):

import { llm, pipeline } from 'skelm'
import { z } from 'zod'

export default pipeline({
  id: 'classify',
  input: z.object({ text: z.string() }),
  output: z.object({ label: z.string(), confidence: z.number() }),
  steps: [
    llm({
      id: 'classify-text',
      backend: 'openai',
      prompt: (ctx) =>
        `Classify the sentiment and return JSON { label, confidence }:\n${ctx.input.text}`,
      output: z.object({ label: z.string(), confidence: z.number() }),
      maxTokens: 1024,
    }),
  ],
})

When output is supplied, the runtime requests structured output from the backend and validates the parsed JSON against the schema before recording it.

Adding an agent step

Default-deny: every AgentPermissions field defaults to deny when omitted. An agent with no permissions block cannot call tools, read files, execute binaries, attach MCP servers, or make network requests.

The recommended agent backend is 'pi' (the pi coding-agent SDK, via @skelm/pi → createPiSdkBackend). The CLI's pi config key wires up the RPC backend; for the SDK backend, register an instance.

// skelm.config.ts
import { defineConfig } from 'skelm'
import { createPiSdkBackend } from '@skelm/pi'

export default defineConfig({
  backends: { agent: 'pi' },
  instances: [createPiSdkBackend({ id: 'pi' })],
})

Minimal agent step with explicit permissions:

import { agent, pipeline } from 'skelm'
import { z } from 'zod'

export default pipeline({
  id: 'ticket-to-pr',
  input: z.object({ ticketId: z.string() }),
  output: z.object({ prUrl: z.string() }),
  steps: [
    agent({
      id: 'implement',
      backend: 'pi',
      prompt: (ctx) =>
        `Implement ticket ${(ctx.input as { ticketId: string }).ticketId}. Return JSON {prUrl}.`,
      permissions: {
        allowedTools:       ['gh.*'],                      // GitHub CLI tools (prefix match)
        allowedExecutables: ['git'],
        allowedMcpServers:  ['github'],
        allowedSkills:      [],
        fsRead:             ['./'],
        fsWrite:            ['./src/', '/tmp/'],
        networkEgress:      { allowHosts: ['api.github.com'] },
      },
      output: z.object({ prUrl: z.string() }),
      workspace: { mode: 'ephemeral', cleanup: 'on-run-end' },
      maxTurns: 8,
    }),
  ],
})

For full agent() options, workspace modes, MCP wiring, and multi-turn control, see references/agent-step.md.

Permissions are part of the API

The permission model has these dimensions. Every field is optional and defaults to deny.

Composition is intersection-only. Project defaults → permission profile → step-level permissions. Each layer can only narrow, never widen. If a project default denies network and a step sets networkEgress: 'allow', the resolved policy is still deny.

Named profiles in skelm.config.ts:

defaults: {
  permissionProfiles: {
    'read-only':  { fsRead: ['./'], networkEgress: 'deny' },
    'full-write': { fsRead: ['./'], fsWrite: ['./'], networkEgress: 'allow' },
  },
}

Apply a profile in a step: permissions: { profile: 'read-only', allowedTools: ['rg'] } — the step's allowedTools are still intersected with the profile.

See references/permissions.md for the full enforcement model.

Project layout

my-project/
├── skelm.config.ts          # Required for gateway and agent steps
├── workflows/
│   └── hello.workflow.mts    # One pipeline per file
├── package.json             # { "dependencies": { "skelm": "^0.3.7", "zod": "^4" } }
└── tsconfig.json

Pipeline files may be named *.workflow.{mts,ts} or *.pipeline.{mts,ts} — the discovery glob in skelm.config.ts decides what gets registered.

For skelm.config.ts shape and all config options, see references/config.md.

Running and inspecting

# Run once, pass input as JSON
skelm run ./workflows/ticket-to-pr.workflow.mts --input '{"ticketId":"PROJ-42"}'

# Run with input from file
skelm run ./workflows/ticket-to-pr.workflow.mts --input-file input.json

# Run a workflow project by directory (resolves skelm.config `entrypoint`,
# else index.workflow.mts, else the single workflow file in the dir)
skelm run ./builder --input '{"spec":"..."}'

# List discovered pipelines
skelm list

# Describe a pipeline (human-readable or mermaid diagram)
skelm describe <workflow-id-or-path>
skelm describe <workflow-id-or-path> --format mermaid

# View run history
skelm history --last 10 --json

Full CLI reference at references/cli.md. Exit codes: 0 ok, 1 CLI error, 2 schema validation, 3 run failed, 4 cancelled, 5 wait timeout, 6 permission denied.

The gateway is the trust boundary

The gateway owns all security-critical infrastructure: permission resolution, secret resolution, audit log, approval gating. Never write permission enforcement in pipeline or step code — call gateway helpers or use TrustEnforcer from @skelm/core only in tests and demos.

skelm gateway start            # foreground; SIGTERM/Ctrl-C drains and exits
skelm gateway status           # shows pid / url / state
skelm gateway stop             # stop a running gateway
skelm gateway install --systemd  # install ~/.config/systemd/user/skelm-gateway.service

Agent steps are enforced by the gateway at dispatch time. A step that declares allowedMcpServers: ['github'] cannot attach any other server, regardless of what the backend requests. See references/gateway.md.

Tests are not optional

Every behavior change ships with tests. For permission paths:

• Default-deny fixture — prove the action is denied when the permission field is omitted.
• Explicit-deny fixture — prove the action is denied when the field is present but excludes the target.
• Both fixtures are required, not one or the other.

pnpm check   # build → typecheck → lint → unit → guards → adversarial → contract → doc-links

Never claim a task done until pnpm check is green.

Common pitfalls

• Widening at step level — setting networkEgress: 'allow' in a step when the project default is deny has no effect. Intersection always wins.
• Missing Zod schema — input / output are validated at run boundaries; omitting them skips validation silently.
• agent() with an unregistered backend — the step fails at runtime if backend references an id with no matching entry under backends: and no matching instance in instances:. The pi SDK backend in particular must be added to instances: (the CLI's pi shorthand wires the RPC variant only).
• Editing dist/ — generated files; never edit. Run pnpm build to regenerate.
• Step id collisions inside parallel() — sibling ids must be unique; the runtime tracks uniqueness globally per run.

Scaffold a new pipeline

bash skill/skelm/scripts/new-pipeline.sh my-pipeline "What this pipeline does"

This copies assets/pipeline.template.ts to ./my-pipeline.pipeline.mts and substitutes the id and description. Use assets/agent-pipeline.template.ts for pipelines that include an agent step.