Model Prefs Assign

Skill: /model-prefs-assign

Model preference: #lowcost (per horizon_aios_model_prefs.md; overridable by a prompt directive).

Maintenance tool for the per-skill model-preference callouts. Every AIOS skill should declare, as a one-line BODY callout, which model group (#lowcost, #midcost, #highcap, #investigate, #debug, #fast) its work should run on. The callout must live in the body — frontmatter is stripped before the model sees the skill, so a frontmatter field would never be read. This skill finds skills missing or mis-assigned and fixes them, then syncs the index columns.

Group semantics live in $HORIZON_ETC/horizon_aios_model_prefs.md — read it first.

Quick Reference

• Purpose: keep every AIOS skill labeled with an appropriate model group, and the skill indexes in sync.
• Triggers: "assign model groups to skills", "audit skills for model preference", "which skills are missing a model group", /model-prefs-assign.
• Scope: all SKILL.md under $HORIZON_SYSTEM/skills_bin/ and $HORIZON_SYSTEM/skills_sbin/ (owner-only; edits privileged skill files).

The callout (exact format)

One line, inserted immediately after the # Skill: heading (blank line before and after). Substitute the group:

**Model preference:** \#GROUP` (per `horizon_aios_model_prefs.md`; overridable by a prompt directive).`

Never put the group in frontmatter — it will not reach context.

Assigning a group — heuristics

Pick the group that matches the dominant nature of the skill's work:

1.1 #highcap — security-sensitive, privileged, or destructive OS changes; deep cross-referencing / judgment where a mistake is costly (e.g. brain provisioning, ACL hardening, consistency validation). 1.2 #investigate — research, live fetching, cross-source analysis, synthesis. 1.3 #midcost — structured authoring/summarization with moderate judgment (docs, handoffs, objectives, scaffolding). 1.4 #lowcost — mechanical, read-only, or report-only work with little reasoning. 1.5 #fast — trivial single actions where latency matters (e.g. launching a process). 1.6 #debug — step-by-step debugging / coding tasks.

When uncertain between two, prefer the cheaper group unless a mistake would be costly (then round up). Always propose the assignment to the user before applying.

Procedure

Step 1 — Enumerate

1.1 Glob all SKILL.md under skills_bin/ and skills_sbin/.

Step 2 — Audit

2.1 For each, grep for a line matching **Model preference:** #.... Classify: present (record the group), missing, or malformed (wrong format, unquoted, or in frontmatter). 2.2 Produce an audit table: skill | tier | current group | status.

Step 3 — Assign / refresh

3.1 For each missing/malformed skill, determine the group from the heuristics above based on the skill's own description and body. 3.2 If the user asked to MODIFY a specific skill's group, target that one. 3.3 Present the proposed changes (skill → group → reason) and get confirmation. 3.4 On confirm, Edit each SKILL.md: insert or correct the callout line after the # Skill: heading. Match each file's exact heading text. Do not touch frontmatter or other content.

Step 4 — Sync the indexes

4.1 Ensure both skills_bin/index.md and skills_sbin/index.md have a "Model group" column, and that every row's value matches the skill's callout. Add/repair rows as needed.

Step 5 — Report

5.1 List skills added, updated, and unchanged, and confirm index columns match. 5.2 Remind: changes to skills_sbin/ are owner-tracked — commit them (with the index updates) in one commit; the gitignored extend.md is unaffected.

Notes for the executing agent

• This edits privileged skill files — owner-only; never expose to brains.
• The callout is documentation the acting model reads when the skill runs; it does not enforce model selection, it directs it (in-context, per the BYO layer).
• New skills should ship the callout from creation — skill-creation is the place to enforce that; flag if a newly created skill lacks one.
• Keep each callout to exactly one line (token-aware). Do not restate group semantics in the skill — they live in the spec.
• Never hardcode paths; resolve $HORIZON_SYSTEM / $HORIZON_ETC from the env.