The layers

muster checks seven layers of the agent-file stack, plus how those layers compose. Every layer has a static mode that runs offline and produces the same result every time. Most layers also have a behavioral mode that grades a live model, and that mode runs only when you give it an endpoint. When no endpoint is set, the behavioral cases are skipped and recorded as skipped, not counted as failures.

Every command shares two global flags, --mode <strict|permissive> (default strict) and --json, and the same exit codes: 0 conforming or all passed, 1 violations or at least one failure, 2 execution error such as an unreadable or invalid manifest.

The examples below assume a source checkout. Replace node dist/cli/index.js with muster if you installed the package globally.

Persona (Soul.md)

The original layer. A Soul.md file defines an agent’s voice, values, interaction style, safety posture, composition from reusable mixins, and reactive state. This is the Soul.md RFC-1 format, and muster is its reference CTS-1 harness.

Static checking parses the front matter, enforces the Soul-YAML subset (anchors, aliases, merge keys, and tags are refused rather than expanded), validates against the Appendix E JSON Schema and the §25 keyspace rules, and resolves composition to a byte-stable canonical form. Behavioral grading scores a model on three axes: verbosity, brief refusals, and dynamic state shifts.

# Validate one document and print a §25.1 report
node dist/cli/index.js check examples/soul/Soul.md --json

# Print the effective config after full §7.5 resolution
node dist/cli/index.js resolve examples/soul/Soul.md --output-format canonical-json

# Run the CTS-1 static fixture suite
node dist/cli/index.js cts run examples/cts/manifest.yaml

# Grade a live model against the soul's axes
node dist/cli/index.js behave run examples/behave/manifest.yaml --base-url https://api.openai.com/v1 --model gpt-4o

check also accepts --adapter rfc1|heartbeat|a2a for a static lint of those file types, plus --profile, --state, and --restrict-refs [dir] to confine reference loading. behave run reads the API key from MUSTER_API_KEY, falling back to OPENAI_API_KEY.

Skills (SKILL.md)

A skill is a directory with a SKILL.md and optional bundled files, following the agentskills.io spec. Static checking validates the front matter, the directory layout, and bundled-file safety, so a skill cannot reference a file outside its own directory or collide with a reserved name. Behavioral checking sends trigger queries to a model and confirms it routes to the skill when it should and stays away when it should not.

node dist/cli/index.js skills run examples/skills/manifest.yaml

The manifest lists skill directories and the cases to run against each. Trigger conformance runs when MUSTER_ENDPOINT is set.

SOP (AGENTS.md)

An AGENTS.md file is a standard operating procedure: the rules an agent must follow and the order they take precedence in. Static checking confirms each rule in the manifest has matching text in the document, flags precedence that is left undefined, and detects drift between the tools the SOP names and the tools the environment actually offers. Behavioral checking runs compliance probes (does the model follow the rule) and adversarial probes (can a crafted message get it to break the rule).

node dist/cli/index.js sop run examples/sop/manifest.yaml

The probe cases run when MUSTER_ENDPOINT is set. The CLI command is sop; the adapter is named openclaw-sop internally.

Tools (TOOLS.md)

A TOOLS.md manifest declares the functions an agent may call and their parameters. Static checking lints the manifest for missing sections, duplicate tool names, and parameter problems, and can compare it against an environment descriptor to catch drift. Behavioral checking presents the model with a scenario and confirms it selects the right tool.

node dist/cli/index.js tools run examples/tools/manifest.json

A case may set expect: "fail" for a negative test, where the case passes only when the lint, drift, or selection check produces at least one failure.

Memory (MEMORY.md / USER.md)

Memory files hold what an agent should remember: MEMORY.md for facts it has learned and USER.md for facts about the person it serves. Static checking finds stale facts and contradictions between the two files. It reads no clock; the reference date comes from the manifest, so the same input always produces the same findings. Behavioral checking runs recall probes (does the model use a fact it should know) and privacy probes (does it leak a fact it should keep private).

# Static lint
node dist/cli/index.js memory run examples/memory/manifest.json

# Add behavioral recall and privacy probes
node dist/cli/index.js memory run examples/memory/manifest.json --behavioral --base-url http://localhost:11434/v1 --model llama3.2

The behavioral endpoint defaults to a local Ollama instance (http://localhost:11434/v1, model llama3.2) and is overridable with --base-url and --model.

Heartbeat (HEARTBEAT.md)

A HEARTBEAT.md is the checklist an agent works through on a schedule. Static checking lints the document and its interval configuration. Behavioral checking covers three cases: action-diff (does the model take the right action for the current tick), idempotency (does a repeated tick avoid repeating work already done), and quiet-ack (does it stay silent when there is nothing to do).

# Static lint and interval checks
node dist/cli/index.js heartbeat run examples/heartbeat/manifest.json

# Add behavioral probes
MUSTER_ENDPOINT=https://api.openai.com/v1 node dist/cli/index.js heartbeat run examples/heartbeat/manifest.json

Behavioral cases run when MUSTER_ENDPOINT is set, with MUSTER_MODEL defaulting to gpt-4o-mini.

A2A (Agent Card)

An A2A agent card is the JSON document an agent publishes to describe itself to other agents: its skills, its authentication, and an optional signature. Static checking validates the card schema and verifies signatures offline when the JWKS is a local file. Live checking calls a deployed agent to confirm a skill behaves as advertised, that unauthorized requests are rejected, and that a signed card verifies against its live key set.

# Static card lint
node dist/cli/index.js a2a run examples/a2a/manifest.json

# Live conformance against a deployed agent
MUSTER_A2A_ENDPOINT=https://my-agent.example.com node dist/cli/index.js a2a run examples/a2a/manifest.json

A2A uses its own environment namespace and never touches the variables the other layers use. Set MUSTER_A2A_ENDPOINT to enable the live cases, MUSTER_A2A_TOKEN for the authorized leg of the auth probe, and MUSTER_A2A_TIMEOUT_MS to change the request timeout (default 10000).

Cross-layer composition

Layers do not run in isolation. A persona, an SOP, and a skill stack on top of each other, and the combination can behave differently than any single layer predicts. The cross-layer checks assemble a layer stack, lint it for composition problems such as contradictions and undefined precedence, and run rule-survival cases that confirm a rule declared in one layer still holds once the others are applied.

# Static composition checks only
node dist/cli/index.js crosslayer run examples/crosslayer/manifest.yaml --static-only

# Add behavioral rule-survival cases
MUSTER_ENDPOINT=https://api.openai.com/v1 node dist/cli/index.js crosslayer run examples/crosslayer/manifest.yaml

Behavioral cases use MUSTER_ENDPOINT or an endpoint block in the manifest. Use --static-only to skip them without setting up an endpoint.

A note on credentials

No command takes an API key as a flag, reads one from a manifest, or writes one to disk. A manifest names the environment variable to read; muster resolves the value from the environment at request time. A repository invariant test fails the build if a secret-shaped string is ever committed.