Instructions for agents changing this skill
SPEC.md
The skill must conform to SPEC.md next to this file. Read it before changing any file in this directory. If the change
you are making, or the text you find, disagrees with SPEC.md, treat that as a bug: fix the skill, or update SPEC.md
explicitly in the same change with the reason. Never leave the two apart, and never satisfy a spec requirement by
narrowing what the requirement says.
SPEC.md opens with a Dependencies: line that tests/skill_deps.rs at the repository root parses. This skill depends
on nothing, and nothing in its text may point a reader at another skill.
Reference audit
Before presenting a change as done, search every file in this directory for the words galaxy-brain, delegation,
harness/, gate, checkpoint, and run directory. Each hit is either a request input the text names as something
the caller supplies, or a leak of a consumer's vocabulary into a skill that must not know its consumers; rewrite the
leak.
Evaluating changes
Skill text is consumed by agents that have none of your conversation context, so your own reading of the new wording
proves nothing about how it lands cold; misrouting has repeatedly been caught only by asking a clean agent. After
changing any file here, eval the change with a fresh-context agent before presenting the work as done: install the skill
into an isolated home, and ask a cold agent each question below with the full situation stated, either directly ("load
scode-model-routing and answer") or through a temporary caller prompt that supplies the request the way a consumer
would. Judge the answer against the expected one, not against whether it quotes the new text. Do not point the agent at
a sidecar directly; whether SKILL.md sends it there is part of what the eval checks.
Unless a question says otherwise, the situation is: a Claude Code session on fable high (sota), no provider
preference, no config file, own decomposition, first attempt. The list is run by the PR that creates this skill and
re-run whenever the profile table, the precedence list, the request or answer shape, or the inventory changes.
| Situation | Expected answer |
|---|---|
| clear-spec writer, short, input not large | gpt-6-luna high external, cross-family yes (workhorse default) |
| clear-spec writer, tiny | orchestrator |
| read-only scan of one subsystem, input not large | haiku high native |
| read-only whole-repo scan, input large | haiku high native (read-only work never takes the luna route, so the long-context exception has nothing to redirect) |
| design decision, own decomposition | orchestrator |
design decision from a sonnet orchestrator (not sota) |
fable high native (delegate up) |
UI styling from a Codex session on gpt-6.2-sol high under prefer-gpt |
opus high external, diverged from preference yes (visual) |
routine authored under prefer-muse |
muse-spark-1.3-contributor medium external, escalation high |
critical review under prefer-glm |
fable high native, diverged from preference yes (no suitable model) |
critical review under prefer-glm from a Codex session on gpt-6.2-astra high |
gpt-6.2-astra high native, diverged from preference yes |
| config file removes fable; critical review | gpt-6.2-astra high external, cross-family yes (the only remaining sota model) |
design decision from a Codex session on gpt-6.2-sol high (not sota) |
gpt-6.2-astra high native (delegate up) |
config file removes sonnet medium; clear-spec writer under prefer-claude |
sonnet high native |
clear-spec writer, short; input 13 lists the gpt family as unreachable on external |
sonnet medium native, reason says the gpt path is unavailable |
clear-spec writer, large input; input 13 lists gpt-6.2-sol and gpt-6-luna as unreachable on external |
a Claude route native, not sol or luna again |
routine authored under prefer-gpt; input 13 lists the gpt family as unreachable on external |
sonnet low native, diverged from preference yes |
config file says "use GLM through the zai provider"; routine authored under prefer-glm |
glm-5.3-flash high external, launch preference carrying those words verbatim |
config file says "run sonnet through claude -p"; clear-spec writer under prefer-claude |
sonnet medium external, launch preference carrying those words |
user demands gpt-6-luna; input 13 lists it as unreachable on external |
no suitable route, reason names the demand |
a Codex session; input 13 lists gpt-6-luna as unreachable on native; clear-spec writer, short |
gpt-6-luna high external |
| a swarm-defined design role (process-defined spawn) | inherit |
a swarm-defined mechanical-review reviewer under prefer-gpt (process-defined, native only) |
sonnet high native, diverged from preference yes (mechanism fixed) |
| user demands opus for a mechanical task | opus high native, reason attributes the demand |
| user demands claude-fable-5, the older release, for critical review | claude-fable-5 high external, reason attributes the demand (native model parameter takes only aliases) |
| user writes "fable-5 high" for a mechanical task, copied from an older prompt | fable high native, reason attributes the demand (the old label means the alias) |
| critical review with independent perspective requested | gpt-6.2-astra high external, cross-family yes (independence) |
clear-spec writer after one substantive failure on gpt-6-luna high |
gpt-6.2-sol medium external (not sonnet) |
the same after execution-path failure |
gpt-6-luna high external again |
the same after substantive failure (lost context) |
gpt-6.2-sol medium external, reason says no attempt consumed |
Focused-review scenarios:
| Situation | Expected answer |
|---|---|
| Codex on sol high; each of idiomaticity, AI slop, and docs/comment correctness, input not large | gpt-6-luna high native; focused review |
| Codex on sol high; each of data-flow and edge-input correctness lenses, with strong general correctness coverage in the panel, input not large | gpt-6-luna high native; full charter retained |
| the preceding focused review with large input, or after lost-context failure on luna high | gpt-6.2-sol high native; no additional GPT rung |
| the preceding focused review after ordinary substantive failure on luna high | gpt-6.2-sol high native; route exhausted yes, endpoint trusted no |
| Codex on sol high; standalone correctness review focused on data flow | gpt-6.2-astra high native; critical review |
| Codex on sol high; simplification review | gpt-6.2-sol medium native; mechanical review |
| Claude Code on fable high; docs/comment correctness, process-defined native-only role | sonnet high native; focused review, Claude route |
| Claude Code on fable high; data-flow correctness lens in a panel, process-defined native-only role | fable high native; non-GPT correctness keeps critical review |
When a scenario depends on the model routing config file, the isolated home is where it goes; never create or edit
~/.scode-model-routing.md in the real home for an eval, it belongs to the user.