Muscle
Run the CSM delegation loop: hand grunt/bulk work (building, drafting, coding, doc generation, deep debugging) to GitHub Copilot CLI GPT-5.6 Sol as a background muscle job, then gate-check the FULL result hands-on with real inputs, fix via --resume rounds until it stands, and add the Fable-only layer last. USE THIS SKILL when Kody says: 'delegate to sol', 'muscle this', 'have sol/copilot build it', 'spawn the muscle', 'send it to the muscle', 'copilot the grunt work', 'ration fable', or whenever a task is bulk artifact generation (code, tests, docs, ports, sweeps) that does not need Fable-level judgment to produce — the output is an artifact, not a decision. Also use when a hands-on verification loop exceeds ~2 probes on the same failure (deep debugging is muscle work).From its SKILL.md
npx -y skills add kody-w/rapp-skills --skill muscleAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 24 days oldThe repository was created 24 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
17.7 KB, ~10.0k tokens by cl100k_base, as published. Nobody here has run it
muscle — the CSM delegation loop, as one verb
Canon: memory [[csm-doctrine]] + [[fable-copilot-loop-lessons]] (read them if judgment
calls arise). This skill is the runnable mechanics. Audit evidence (2026-07-16): the
loop was performed BY HAND 129 times across 8 of the last 50 sessions, with 39
--resume correction calls — this skill replaces that ritual.
Division of labor (decision rule): output is a decision -> Fable inline. Output is an artifact -> muscle. Output is a handoff -> spine (you, orchestrating).
1. Write the order (never prompt ad hoc)
Write the order to a file in the target workdir (ORDER-<slug>.md), then point the
muscle at it. Brief anatomy that field-tested well:
- Intent — one paragraph, what and why.
- Inputs to study — exact paths.
- Rules / do-not-touch — includes file-ownership lines when jobs run in parallel: "only create the files named below; a concurrent job may be writing X — leave it."
- Numbered acceptance checks — runnable commands, verbatim-output requirement.
- Done-when + report format — require it to RUN its own verification, end with a report, and include a "flags / surprises" section ("flag uncertainty, don't paper over") — that section yields real gold.
2. Spawn the muscle (background, parallel-safe)
cd <workdir> && copilot -p "Read ORDER-<slug>.md in the current directory and execute it completely. You are the muscle in a cortex/muscle pattern: do the full grunt build, run your own verification per the Done-when section, and end with the report." \
--model gpt-5.6-sol --allow-all-tools --log-level none 2>&1 | tail -40
- Big orders:
run_in_background: trueon the Bash call; several jobs in parallel are fine only with file-ownership lines in each brief. - Never poll with foreground
sleep(audit: 419 sleep-then-check calls wasted hours of wall clock). Background the job and use the Monitor tool / background-task notification; check output withtailonly when woken. - The output ends with
Resume: copilot --resume=<session-id>— capture that id; it is the correction channel. - Sanity ping if the CLI has not been used this session:
copilot -p "reply with exactly: MUSCLE-ONLINE" --model gpt-5.6-sol --allow-all-tools --log-level none 2>&1 | tail -3
3. Gate-check: touch the FULL E2E yourself (never skip, strictly ordered)
The muscle's green suites lie — every defect that ever mattered was invisible to its own tests and caught only by hands on REAL artifacts. Sequence (Kody-mandated):
a. Touch the full end-to-end — wait until the muscle is completely done, then exercise the entire finished artifact yourself: run it, render it, curl it, drive the whole flow. Re-run the acceptance commands verbatim PLUS one probe against a real file / live system the brief never mentioned (the live-system probe is the highest-yield test). Never trust the muscle's report. b. Criticize the whole — one written adversarial critique of the complete shape: gaps, wrong turns, integration seams, taste failures. c. Grunt-level defects go back via:
copilot --resume=<session-id> -p "<repro command + observed error + root cause + exact fix + acceptance check>"
Corrections written that way converge in one shot. Repeat a->b until it stands. d. THEN the Fable-only layer, last — naming, architecture judgment, taste, partner-facing prose (Fable writes every customer/partner-facing word). Never interleave touches with review.
4. Guards
- The slip: verification is a bounded final touch, not an iterative debug loop.
More than ~2 hands-on probes on the same failure -> STOP, package the loop as a
--resumeorder with the evidence gathered so far. - Report to Kody as one unified result with evidence — what was delegated, what the gate-check observed, what was fixed, what Fable added.
Parameters
The typed contract this capability answers to (JSON Schema — the deterministic layer):
{
"properties": {
"slug": {
"description": "Derived from `<slug>` used in the documented command at line 13.",
"type": "string"
},
"workdir": {
"description": "Derived from `<workdir>` used in the documented command at line 28.",
"type": "string"
}
},
"required": [],
"type": "object"
}
Deterministic steps
Lifted verbatim from the procedure above by toaster.py toast. Run them in order, substituting the typed parameters; do not paraphrase:
cd <workdir> && copilot -p "Read ORDER-<slug>.md in the current directory and execute it completely. You are the muscle in a cortex/muscle pattern: do the full grunt build, run your own verification per the Done-when section, and end with the report." \
What ships with it: 1 file
21.8 KB alongside SKILL.md, 1 of them executable
- muscle_agent.pyruns21.8 KB