Open Spec · v1.0.0

Describe the whole AI harness.

Cromus STACK.md inventories an AI system as one governed unit: the files and runtime components around the model, not only a single skill or workflow.

Eight components, one manifest

Each section is optional and can reference a repository file, embed text inline, or combine both. Missing sections are reported as Structural Gaps; references alone remain declared but are not fetched or scored.

system_prompt

The system prompt and operating role.

guardrails

Behavioral policy and constraints.

memory

Persistent and session context.

tool_guidance

Tool instructions, MCP dependencies, and side-effect labels.

subagents

GRAPH.md topology and AGENTS.md instructions.

skills

The SKILL.md capabilities the harness uses.

scripts

Runnable code, purpose, and side-effect classification.

plugins

Installed plugins and MCP servers.

Required frontmatter: cromus_stack: "1.0.0", stack_id, spec_version, owner

---
cromus_stack: "1.0.0"
stack_id: support-agent
spec_version: "1.0.0"
owner: Support Engineering
system_prompt:
  ref: prompts/support-system.md
guardrails:
  ref: ETHOS.md
tool_guidance:
  tools:
    - name: search_support_docs
      usage: Search approved support documentation before answering.
      side_effect_class: read_only
scripts:
  - path: scripts/build-report.py
    language: python
    purpose: Build a local support report.
    side_effect_class: writes_local
---
Deterministic by design

Harness scoring makes zero LLM calls and does not retain submitted content.

Only frontmatter marked cromus_stack: "1.0.0" identifies a Cromus manifest. Ordinary tech-stack documents receive a neutral notice and no score.

Inline SKILL.md, ETHOS.md, MEMORY.md, and GRAPH.md content uses the existing canonical Cromus engines.

Each missing top-level component adds 25 Croms to Structural Gaps only. The composite score is capped at 1000.

Each component uses a repository ref, an inline value, or both. References are never fetched; supplied bundle files can resolve local references.

Harness formats already exist

Cromus did not invent the harness manifest. Our validator also scores read-only inventories of htekdev/ai-harness harness.md and reachjalil/harness-config .harness/ source bundles. Results identify the source format and exclude unsupported or unsubmitted components from missing-component penalties.

Scripts and tools carry side-effect labels

Use read_only, writes_local, network, or destructive. Connect a declaration to a GRAPH.md tool node with graph_node_id so judge-placement analysis sees the same risk. Unclassified scripts are treated as destructive until labeled.

Canonical component scoring

Available inline component Croms are averaged, then missing-component penalties are added. SKILL.md uses Operational Croms, ETHOS.md uses Ethics Croms, MEMORY.md quality is inverted onto the Croms scale, and GRAPH.md uses the Graph Layer score and signals. Adding or removing a missing component changes Structural Gaps only—not the other pillars.

Check a harness manifest

Paste a marked STACK.md, harness.md, or .harness/ files bundle for validation and canonical scoring.

Open validator