Model Prefs Test

Skill: /model-prefs-test

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

Reliability harness for the in-context model-preference layer. It answers the question the spec flags as untested: when I reference a group, does the right model actually get used? It has two modes — a cheap dry-run that shows how each group resolves, and a --live mode that spawns small agents by group and has them self-report which model ran.

Source of truth for the layer and member grammar: $HORIZON_ETC/horizon_aios_model_prefs.md (+ the gitignored horizon_aios_model_prefs.local.md). Read both before resolving anything.

Quick Reference

• Purpose: verify group → model resolution for the model-preference config, and optionally prove a spawn honors it.
• Triggers: "test my model prefs", "check which model #lowcost uses", "validate model groups resolve", /model-prefs-test.
• Modes: dry-run (default, no spawns, free) · --live (spawns test agents, costs tokens).
• Args: /model-prefs-test · /model-prefs-test --live · /model-prefs-test --live #lowcost #highcap (restrict to named groups).

When to invoke

After editing the extend file, after /model-catalog-refresh changes members, or any time you want to confirm a group resolves to what you expect in the current runtime.

What it tests

1. Resolution (both modes): for each group, which member is the first one runnable in the current runtime, and which members were skipped (wrong runtime / not available).
2. Routing (both modes): each task-class routing rule → its group → resolved model.
3. Spawn honoring (--live only): an agent actually spawned on the resolved model performs a tiny task and reports its own model identity, so you can see expected vs. spawned-as vs. self-reported.

Procedure

Step 1 — Load config and identify runtime

1.1 Read $HORIZON_ETC/horizon_aios_model_prefs.md and $HORIZON_ETC/horizon_aios_model_prefs.local.md (if the extend file is absent, say so — groups are empty and there is nothing to resolve). 1.2 Identify the current runtime — the harness you are executing in. For Claude Code: claude: members are runnable; ollama: members are skipped unless an Ollama bridge is configured. State which runtime you assumed.

Step 2 — Resolve (dry-run; always done)

2.1 For each group, walk members top-to-bottom; the resolved model is the first member runnable in the current runtime. Strip the runtime prefix for the model name (claude:haiku → haiku). List skipped members and why. 2.2 Resolve each Task-Class Routing rule the same way (class → group → model). 2.3 Print the resolution table:

| group | members (in order) | resolved (this runtime) | skipped |

plus a routing table: | class | → group | resolved |

2.4 If not --live, stop here. This proves nothing about spawning — say so.

Step 3 — Live test (--live only)

3.1 Choose the groups to exercise: the named ones if the user passed any, else a representative set — #lowcost, #investigate, #highcap (add #fast / #debug if the user wants broader coverage). Keep it to a few agents. 3.2 For each chosen group, spawn ONE agent with the model parameter set to the resolved alias from Step 2 (this is the mechanism under test — that group selection can drive spawn model). Give it the matching tiny task: - #lowcost, #fast → "Tell one short programming-related joke." - #midcost, #investigate, #highcap → "In 3 bullets, summarize the engineering values in $HORIZON_DOCS/dev_values.md." - #debug → "In 2 sentences, explain the bug in for i in range(len(a)): a.pop(i)." - Append to every task: "Then state, on its own line, MODEL: <your model family/id as best you can identify it>." 3.3 Collect each agent's final line and the model you spawned it with.

Step 4 — Report

4.1 Print the live results table:

| group | expected (config) | spawned-as (model param) | self-reported | match? |

4.2 Mark match ✓ (self-report agrees with spawned-as), ✗ (disagrees), or ? (agent couldn't identify itself — common; not a failure of the config). 4.3 Summarize: did every group resolve? did every spawn accept the model param? any surprises (a group falling through to default, a routing class with no runnable member)?

Interpreting results — caveats

• Self-report is corroboration, not ground truth. Models often can name their family (haiku/sonnet/opus/fable) but not the exact version, and may guess. A ? or a vague answer is not a config failure. The reliable signal is that the spawn accepted the resolved model and ran.
• The real pass/fail is resolution + spawn acceptance, not the prose answer. The joke/summary just gives the agent something cheap to do.
• Runtime matters. In Claude Code, ollama: members are expected to be skipped — that is correct behavior, not a miss. Flag it as skipped, not failed.
• A group whose only runnable member is missing should fall through to the harness default; note when that happens.
• Dry-run cannot detect access-suspension. It resolves by runtime prefix only, so a claude: member that the provider has access-restricted (e.g. a suspended model) still shows as "resolved" in dry-run. Only --live reveals it — the spawn fails and resolution must fall through to the next member. Never trust a green dry-run for a model you know may be access-restricted; confirm with --live.

Notes for the executing agent

• Dry-run is the default — never spawn agents unless --live is present.
• Spawn live test agents in parallel where the harness allows; keep tasks tiny to bound cost.
• This skill only reads config and the values doc and spawns short-lived agents; it writes nothing and touches no privileged paths — safe for brains.
• Resolve $HORIZON_ETC / $HORIZON_DOCS from the environment; never hardcode paths.