Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Safely upgrade this Sutando checkout to the latest upstream code **without
.claude/skills/sonichi-self-upgrade/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | 144% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 0% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 7% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -11% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 6% | 0% |
Safely upgrade this Sutando checkout to the latest upstream code without bricking the running core session — the "success path" distilled from a real 2026-07-20 upgrade that would otherwise hang (and did, the first time).
Usage: /self-upgrade
A naive "pull + restart" self-upgrade gets stuck, because:
src/restart.sh ends with exec bash src/startup.sh.src/startup.sh runs foreground work and foreground-parents thecredential-proxy (a tsx process that never exits). The open-source core is headless: the optional Swift helpers (ax-read, Sutando.app) are built separately by the app packaging/setup workflow, not by core startup.
restart.sh inline from the core session never returns —the Bash call hangs forever, the task never gets a result, and from the owner's side you've "gone stuck."
The fix is simple once you know it: hand the restart to the same durable tmux server that owns the core. A plain nohup … & is not enough: the supported Codex executor tears down that process tree when the tool call ends. A detached tmux service pane survives that boundary, remains the parent of restarted services, and lets startup recreate the managed task notifier.
Run the helper. It aborts safely on a dirty tree or a non-fast-forward, pulls --ff-only, and launches src/restart.sh in the persistent sutando-services tmux session:
bashbash skills/self-upgrade/scripts/upgrade.sh # origin/main # bash skills/self-upgrade/scripts/upgrade.sh --no-restart # pull only
Exit 0 = upgraded (or already latest); exit 2 = aborted (dirty tree / not a fast-forward) — surface the reason and stop.
If the diff touched package*.json / tsconfig / *.swift / requirements (the script prints this), a rebuild may be needed. For *.swift, rebuild the optional menu-bar app and ax-read through the app setup workflow; core startup.sh intentionally does not build them. For npm deps run npm ci before relying on the TS services.
bashpython3 src/health-check.py
Expect "All systems operational." Confirm the core survived (the restart log contains sutando-core already running — restart.sh never touches the core CLI), the managed sutando-core-watcher tmux session exists, and bridges came back on new PIDs. telegram-bridge / slack-bridge warnings are fine if they were already optional/unconfigured.
For live-path evidence, submit one task through POST /task, write its result, and confirm GET /result/<id> returns that exact body after the restart.
Report to the owner: old → new commit, how many commits, whether a rebuild was needed, and that the core stayed up.
restart.sh / startup.sh inline from the core session, anddo not rely on plain nohup … &. Inline = stuck; an executor may reap the nohup child. Use the helper's durable tmux handoff.
sutando-services session to "tidy up."It deliberately parks after startup completes so background bridges keep their durable parent. The helper marks the session done and only replaces that completed session when a later upgrade actually needs another restart.
watch-tasks-stream also matches *other* installs (e.g. a /tmp/… checkout); match the full repo path, not a bare pattern.
clobber them; commit or stash before upgrading.
nohup with a durablesutando-services tmux handoff that outlives task executors and keeps the bridge parent alive after startup; completion markers are isolated per tmux socket.
behind → 0) where the naive inline restart hung on startup.sh's foreground Swift build + credential-proxy hold.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 1,901 | 4,560 | +140% | 1 | 1 | 0% | 279 | 1,301 | +366% | 0 | 0 | — |
case-02 | fail→fail | 10,028 | 4,992 | -50% | 1 | 1 | 0% | 1,370 | 1,303 | -5% | 0 | 0 | — |
case-18 | pass→pass | 8,574 | 2,307 | -73% | 1 | 1 | 0% | 1,281 | 1,530 | +19% | 0 | 0 | — |
case-03 | fail→fail | 5,832 | 5,485 | -6% | 1 | 1 | 0% | 325 | 1,370 | +322% | 0 | 0 | — |
case-04 | fail→fail | 9,231 | 5,460 | -41% | 1 | 1 | 0% | 1,529 | 2,018 | +32% | 0 | 0 | — |
case-05 | fail→pass | 4,141 | 2,678 | -35% | 1 | 1 | 0% | 627 | 1,533 | +144% | 0 | 0 | — |
case-06 | pass→pass | 9,534 | 3,474 | -64% | 1 | 1 | 0% | 1,513 | 1,675 | +11% | 0 | 0 | — |
case-07 | fail→pass | 11,506 | 4,055 | -65% | 1 | 1 | 0% | 1,740 | 1,744 | +0% | 0 | 0 | — |
case-08 | fail→pass | 8,870 | 2,274 | -74% | 1 | 1 | 0% | 1,353 | 1,452 | +7% | 0 | 0 | — |
case-09 | pass→pass | 9,697 | 5,631 | -42% | 1 | 1 | 0% | 1,625 | 2,022 | +24% | 0 | 0 | — |
case-10 | fail→pass | 12,706 | 6,316 | -50% | 1 | 1 | 0% | 2,072 | 1,837 | -11% | 0 | 0 | — |
case-11 | fail→pass | 12,795 | 5,481 | -57% | 1 | 1 | 0% | 1,901 | 2,009 | +6% | 0 | 0 | — |
case-17 | fail→pass | 11,948 | 2,512 | -79% | 1 | 1 | 0% | 1,871 | 1,491 | -20% | 0 | 0 | — |
case-12 | pass→pass | 9,812 | 5,095 | -48% | 1 | 1 | 0% | 1,560 | 1,922 | +23% | 0 | 0 | — |
case-13 | pass→pass | 3,806 | 2,905 | -24% | 1 | 1 | 0% | 568 | 1,644 | +189% | 0 | 0 | — |
case-14 | fail→pass | 4,761 | 1,630 | -66% | 1 | 1 | 0% | 683 | 1,305 | +91% | 0 | 0 | — |
case-15 | fail→pass | 6,642 | 2,154 | -68% | 1 | 1 | 0% | 988 | 1,442 | +46% | 0 | 0 | — |
case-16 | fail→pass | 14,188 | 6,497 | -54% | 1 | 1 | 0% | 2,246 | 2,074 | -8% | 0 | 0 | — |
case-19 | pass→pass | 18,209 | 6,107 | -66% | 1 | 1 | 0% | 2,720 | 1,770 | -35% | 0 | 0 | — |
case-20 | pass→pass | 4,841 | 4,146 | -14% | 1 | 1 | 0% | 817 | 1,725 | +111% | 0 | 0 | — |
case-21 | fail→pass | 9,649 | 5,179 | -46% | 1 | 1 | 0% | 1,608 | 1,423 | -12% | 0 | 0 | — |
case-22 | pass→pass | 4,814 | 3,583 | -26% | 1 | 1 | 0% | 756 | 1,686 | +123% | 0 | 0 | — |
case-23 | fail→pass | 9,811 | 2,895 | -70% | 1 | 1 | 0% | 1,624 | 1,564 | -4% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 20 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +48 percentage points is the difference between those two pass rates over the 20 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.