Skip to content

Skills ​

A skill provides reusable procedures such as review checklists, writing templates, or translation rules. Built-in skills and legacy README.md skills still enter the stable prompt prefix in full. The 1.1 development line adds SKILL.md metadata indexes and on-demand bodies. Validate their effect with scenario evaluations.

Select a skill ​

yaml
agents:
  reviewer:
    systemPrompt: Review the supplied contract using the project's rules.
    skills:
      - legal-review

Built-in IDs remain code-review, weekly-report, and translation. Project skills live under skills/<id>/ in the configuration directory, or the SDK's configDir. The directory name is the ID. CoreMind resolves only explicitly selected IDs; it does not discover ancestor directories or remote skills.

Built-ins take precedence over project skills with the same ID. A directory containing only README.md retains full prefix injection and the existing built-ins-first, project-skills-second order. If both files exist, SKILL.md wins with a warning. An invalid standard file fails rather than falling back to README. Truly missing IDs still warn and are ignored.

Standard skill format ​

Place this file at skills/legal-review/SKILL.md:

markdown
---
name: Contract review
description: |
  Check payment, delivery, and liability terms.
  Report risks and unresolved questions.
---

Read the supplied contract and project rules before checking each clause. Mark missing evidence explicitly.

YAML name and description must be non-empty strings; Unicode names and multiline descriptions are supported. Names are display metadata, not aliases. Duplicate display names warn; selecting the same standard ID repeatedly produces one index entry.

On-demand loading ​

The first model request contains only the authorized standard skills' IDs, names, descriptions, and body SHA-256 values. The current agent can call load_skill({"id":"legal-review"}) to receive the frozen body. The tool accepts no path or agent override and cannot load an unselected ID. Calls still pass through ToolPolicy, approvals, and budgets. Skill instructions grant no additional file, network, or process permissions. A conflicting custom tool name is rejected when standard skills are enabled.

After a successful load, a repeated call returns a short reference while the body remains in the final model working set. If compaction removes the body, another call reloads the same frozen bytes. A reference, user claim, or historical load marker does not prove that the body is present in the current request. Normal tool-result messages feed budgets, actual requests, and reconstruction; the body is not also appended to the stable prefix.

Limits, freezing, and recovery ​

  • Each project source file is limited to 256 KiB including frontmatter. Each resolver result and the complete snapshot across all agents are separately limited to 2 MiB of serialized UTF-8 JSON. Exceeding a limit fails explicitly.
  • Canonical directory and file boundaries reject escaping links. Invalid UTF-8, permission failures, or source identity changes during reading fail rather than becoming missing skills. IDs cannot specify arbitrary paths.
  • Body bytes and line endings after frontmatter are preserved. Body and snapshot SHA values are raw 64-character hexadecimal strings without a sha256: prefix.
  • When standard skills are enabled, Runtime creation freezes selected content for its Runs and Children. Later file edits do not change that Runtime; create a new one to load edits. Resuming a standard-skill Run uses its original start snapshot without rereading modified or deleted source files. Invalid snapshots or selection mismatches fail recovery, and existing configuration and resume gates still apply.

The raw RunStore is trusted host data containing frozen bodies and load facts. Public projections, Worker events, and telemetry exports strip those private attached fields. Loaded text can still appear in normal sessions, model requests, or business output. SHA values check consistency; they are not signatures or protection against a host rewriting its own Store.

Disabling Session does not provide full conversation recovery. Frozen snapshots and load facts support skill-result reconstruction; full request reconstruction still requires persisted session entries, compaction records, and the corresponding prefix. Store durability follows its existing contract.

CoreMind does not automatically load neighboring references, execute skill scripts, install dependencies, synchronize remote skills, or provide a skill marketplace.

Released under the MIT License.