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, and SKILL.md
carries one marked stanza per dependency (between <!-- dependency: <name> --> and <!-- /dependency -->) whose
wording the same test checks against its canonical template. Change the stanza only by changing the template in the
test, and then in every skill that carries one.
The Codex path in the stanza
When Codex supplies no skill location, the stanza falls back to ${CODEX_HOME:-$HOME/.codex}/skills/<name>/SKILL.md,
the root verified for Codex 0.152 (the bundled skill-installer skill and codex-rs/skills both say so). A
harness-supplied location takes precedence. Codex's public docs already describe a .agents/skills root. When Codex
changes its skills root, or the installer starts writing somewhere else for Codex, re-verify the stanza's path with a
live codex exec run in an isolated CODEX_HOME and update the template in tests/skill_deps.rs and every stanza
together.
Evaluating changes
After changing any file in this directory that an agent reads (SKILL.md, feedback.md, the specs it points at), eval the change with a fresh-context sub agent before presenting the work as done. 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.
How to run an eval:
- Spawn a sub agent with no prior context (a general-purpose Agent tool task works).
- Tell it only where the relevant skill definitions live — this SKILL.md, plus any other skill involved in the scenario
— and give it a realistic user question whose answer the change should affect. For routing changes, "if I asked you
to use scode-galaxy-brain to do X, which model would you use?" works well. Do not point it at
feedback.mdor atscode-model-routing,scode-agent-delegation, orscode-harness-shelloutdirectly: SKILL.md is supposed to send it there when and only when the scenario calls for them, and whether it went is part of what the eval checks. For a change that touches a launch, ask for the exact launch command it would run, which it can only produce by having loaded the delegation skill, then the shell-out skill and its harness file; for a routing or delegation change, that skill's ownAGENTS.mdcarries the fixed question list, asked through this skill. The verdict-to-action table is this skill's own to eval: for each verdict, ask what the session does next and check it against the table. - Judge whether the answer reflects the intended behavior, not merely whether it quotes the new text. If it doesn't, revise the wording and re-run until it does.
Every change to a stanza or to what SKILL.md says about a dependency also runs the dependency checks: a temporary
consumer whose stanza names a dependency that exists nowhere, one whose name is wrong, and one whose needed sidecar is
missing, each of which must stop and name the skill and the path or tool. Positive cases cover a catalog-supplied file
path without a dedicated loader, a skill-scoped resolver without filesystem metadata, and recovery of truncated output
through the tool's continuation or artifact. Verify that unknown or denied access, invalid identity, and incomplete
content with no recovery path still stop; a complete skill needing no sidecars needs no base directory.
The fixed activation and composition questions, run whenever "Staying active", "Composing with other skills", or the
load-when text changes, each asked of a cold agent with the full situation stated (a Claude Code session on fable high,
no preference, no config file, every supported CLI on PATH, unless the situation says otherwise):
| Situation | Expected answer |
|---|---|
| the task galaxy-brain was invoked for is done; the next message does not mention galaxy-brain | routing continues; any delegation in the new task is announced with its model and effort |
a compaction summary omits the skill but mentions a running codex exec writer stopped at a checkpoint |
galaxy-brain is assumed active and the assumption is stated; the resume proceeds from the recorded session id and run directory, and no skill is re-read on that account alone |
| an invocation that limited scope up front ("for this one thing"); that thing is done | no activation for the next task |
| the swarm is invoked while galaxy-brain is active, no preference | correctness, security, spec-compliance, and test-quality reviewers on fable high native; the rest on sonnet high native; the choice announced; no checkpoint addendum or gate applied to them |
a clear-spec writer routed gpt-6-luna high external comes back execution-path failure because the shell-out skill found no installed CLI for any GPT model |
no relaunch; the gpt family recorded as unreachable on external and sent with every later routing request; the route re-requested |
the routing answer is external with a launch preference |
the preference handed to the delegation skill unchanged with the unit; no CLI chosen by galaxy-brain itself |
When a scenario depends on the model routing config file, ~/.scode-model-routing.md, run it in an isolated home; never
create or edit that file in the real home for an eval — it belongs to the user.