Command Schema

Version: 0

Template

• Shape: file_with_prompt_or_body
• Filename: ``

Fields

Name	Type	Required	Notes
prompt	string	Yes	The instruction text the LLM executes. In .md files, this is the markdown body after frontmatter. In .toml files, this is an explicit prompt = "..." field. dotpack treats them as the same canonical field; adapters re-emit native form.
description	string	No	Brief one-line summary. Present in the anthropics example and all three Gemini TOML examples; absent in wshobson's model-only commands. Used by host /help menus.
model	string	No	Per-command model selection. Values seen: full model ID ("claude-sonnet-4-0", "claude-opus-4-1") — wshobson uses full IDs. Other corpora may use aliases (sonnet/opus/haiku). Adapter resolves per host; Gemini commands inherit session model and have no native slot for this.
allowed-tools	string | array<string>	No	KEPT despite below-floor (1/7) because absence of this field changes runtime semantics in Claude Code (without it, every Bash invocation prompts for approval; with it, listed tools are auto-approved). The anthropics official example uses "Bash(git add:), Bash(git commit:)" filter syntax. Per ADR-0003, this is technically adapter-layer (runtime permissions), but is retained here because Claude's documented spec includes it and it has corpus presence. Adapters that don't support it warn at install time.

Ecosystem Notes

• Two formats, one concept. Claude commands are markdown + optional YAML frontmatter; Gemini commands are TOML files containing a prompt field. The translator unifies on the dotpack representation; adapters re-emit the host-native shape on install.
• Argument placeholders diverge: Claude uses $ARGUMENTS (and $1, $2 positional); Gemini uses {{arg0}} double-curly templating. The translator handles substitution syntax during cross-format port; the schema does not enforce one.
• Filename → command-name mapping: both hosts derive the command name from the filename (minus extension). Subdirectories become namespaces: Claude uses subdir prefixes; Gemini uses colons (agents/start.toml becomes /agents:start).

Deliberately Excluded Concepts

Concept: claude_command_runtime_overrides

Documented in the Claude Code command frontmatter reference but absent from corpus. These encode Claude-native command ergonomics: CLI argument hinting (argument-hint) and invocation gating (disable-model-invocation, which prevents the model from auto-triggering the command). Grouped into one canonical_concept because corpus presence is 0/7 across the group. Adapter behaviour: - claude-code adapter: emit natively when present. - non-claude-code adapters: surface as lossy. disable-model-invocation is particularly load-bearing — silently dropping it on a host that lacks the concept would change runtime invocation semantics. Install proceeds only with --allow-lossy. Note: the universal allowed-tools field is PROMOTED to the schema (not deliberately_excluded) under the Deviation 2 rule of ADR-0011 — see the field's notes block above. It is NOT a lossy-detection concern; adapters that don't support it warn at install time rather than block.

Aliases:

Host	Field Name
claude-code	argument-hint
claude-code	disable-model-invocation

Body Section Normalisation

Survey clusters.md noted recurring body sections in wshobson Markdown commands (## Context, ## Requirements, ## Instructions). Not adopted — these are author conventions for prompt structure, not parseable schema. Same posture as ADR-0006 for agent body. NOT a field; no canonical_concept; not a lossy-detection participant.