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; the test enforces the line, and the reference
audit below enforces the text.
Reference audit
Before presenting a change as done, search every file in this directory for the words galaxy-brain, routing,
inventory, profile, checkpoint, and gate. Each hit is either a local use this skill owns (its own harness/
directory, its own SKILL.md) or a leak of a consumer's vocabulary into a skill that must not know its consumers;
rewrite the leak so the sentence names an input the caller supplies instead.
Adding a harness
A new harness file is not enough on its own. When it is a vendor's own harness for its models (a grok CLI for grok,
say), also add that family to rule 2 of Choosing the external mechanism in SKILL.md, and to the matching requirement
in SPEC.md. Until then the family is served by the default like every other model, which is intended: a harness gets
preference only once its launch mechanics are written down and verified here.
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. 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, point the cold agent at the skill by name only
(never at a sidecar directly; whether SKILL.md sends it to the right harness file is part of what the eval checks),
and judge the answer against the expected answers below.
The fixed question list for this skill, run by the PR that creates it and re-run whenever a harness file, the read-when table, the external mechanism rule, or the list of caller-supplied inputs changes:
- The exact launch line for each of the six harnesses, given a model, an effort, a prompt file, and a result path.
Expected: the template from that harness file with the caller's values substituted,
< /dev/nullkept, and the effort passed explicitly (foragy, flags before-p="$(cat <prompt-file>)", notagy -p --model ...). - The exact writer launch and resume lines for each of the six, given a session or thread id. Expected: the id's source
per harness (
--session-id "$(uuidgen)"for claude and muse;thread_idfrom the--jsonstream for codex;conversation_idfrom the--output-format stream-jsoninitevent or--output-format jsonfor agy;sessionIDfrom the first event for opencode; theidof the first--mode jsonevent for omp), the same flags and working directory or workspace on the resume, and where the final message lands on the resumed turn. - The wake-up interval and the kill rule. Expected: at least every 30 minutes, sooner for short runs; kill by pid
walking descendants first; never
pkill -f. - A codex run without
--jsonwhose log is stuck atReading additional input from stdin...with no version header. Expected: the startup hang; kill and relaunch with the< /dev/nullredirect, not wait. - A codex
--jsonrun two minutes in whose stderr holds onlyReading additional input from stdin...while its stdout event log hasthread.startedand keeps growing, and another whose event log is still empty after a minute. Expected: the first is healthy (under--jsona healthy run prints no version header and nothing else need follow that line); the second is the startup hang, killed and relaunched with the< /dev/nullredirect. - A codex
resumeturn whose stderr has noReading additional input from stdin...line, and a fresh--jsonlaunch that already ended in< /dev/nullbut has nothread.startedafter a minute. Expected: the first is normal (a resume reads stdin only for a-prompt); the second is not the stdin hang, so read its stderr for the real error rather than relaunching it unchanged. - An omp run whose stdout stays empty while stderr repeats
Still starting after Ns — phase: readPipedInputwith N counting up (after theReading prompt from piped stdin...notice). Expected: the startup hang; kill and relaunch with the< /dev/nullredirect, not wait. The repeating lines are not liveness. - An agy
--conversation <id>resume that exits 0 while stderr printswarning: conversation "<id>" not foundand the returnedconversation_iddiffers from<id>. Expected: a fresh session, not a resume; report it to the caller as such. - "What does loading this skill make the session do?" Expected: nothing; it is inert until a caller supplies the inputs
listed in
SKILL.md.
When changing the description, the direct-request instructions, or the external mechanism rule, also evaluate discovery from the skill catalog without naming the skill in the request. These cases preserve the distinction between recognizing a delegation request and requiring a shell-out:
- "Start an astra agent to review this" in a Claude session without native GPT support. Expected: load this skill and
read its codex harness file before constructing a
codex execlaunch. - "Have an opus agent investigate" in a Codex session without native Claude support. Expected: load this skill and read
its claude harness file before constructing a
claude -plaunch. - "Start a neon agent to review this" (or "have a neon high agent investigate") in a session without native Gemini
support. Expected: load this skill, pick
agy -pas Gemini's own harness, readharness/agy.md, and construct anagylaunch using--model gemini-4-neon --effort <level>and-p="$(cat <prompt-file>)"(never--model neonand neveragy -p --model ...). - A "neon" request when
agyis missing fromPATH. Expected:omp -ponopenrouter, not a refusal. - A request naming Antigravity or
agyexplicitly ("use antigravity to run a sonnet agent"). Expected: load this skill and readharness/agy.mdbefore constructing anagylaunch, checkingagy modelsfor the non-Gemini model. - A request that names omp as the mechanism, such as "use omp to run a mercury agent". Expected: load this skill and
read its omp harness file before constructing an
omp -plaunch; when several mercury models exist onopenrouter, ask which one. - A GLM or grok model, not supported natively, with no harness named. Expected:
omp -pon theopenrouterprovider (openrouter/z-ai/glm-5.3-flashfor GLM), with no question about providers. - The same GLM request with a standing user preference for omp's
opencode-zenprovider. Expected:omp -pwithopencode-zen/glm-5.3-flash. - A GPT model when
codexis missing fromPATH. Expected:omp -ponopenrouter, not a refusal. - "Run the glm agent through opencode" with no opencode credential for the provider. Expected: report that opencode cannot serve it; no fallthrough to omp.
- The same GLM request with a stated preference for opencode. Expected:
opencode run, onceopencode providers listshows a credential for its provider. - The same GLM request with
ompmissing fromPATHand an opencode credential present. Expected:opencode run. - A caller that supplies
externalfor gpt-6.2-sol high. Expected:codex exec, not omp. - A caller that supplies
externalfor a model no installed CLI serves. Expected: report it unreachable, naming what each candidate lacked; no substitute model. - A named-model request supported by the native tool, including required effort and resume behavior. Expected: native delegation; discovering this skill must not force a shell-out.
- A named-model request with an explicitly requested external CLI, despite native support. Expected: use that CLI and its harness file, subject to availability and caller constraints.
- A caller that supplies an external launch mechanism, despite native support. Expected: preserve the caller's mechanism and read its harness file; direct-request defaults must not override an existing delegation decision.
- "What is astra?" and a task containing an incidental model name. Expected: no agent launch.