Przejdź do treści
Biblioteka skilli
AI i agenci · Planowanie

human-runsheet

/human-runsheet

Zapisuje w trwałym pliku czynności, które może wykonać tylko osoba prowadząca, jako dokładne polecenia do uruchomienia, gdy sesja natrafia na blokadę, której agent nie może sam obejść.

Repozytorium
assay ↗
Licencja
Apache-2.0
Ostatnia zmiana repozytorium
27.10.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
1
Skrypty
0
Komentarze HTML
1
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 27.10.2026, bez odwołań do wycofanych modeli i API.

Licencja

Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.

Co sprawdzić przed instalacją
  • SKILL.md zawiera komentarze HTML (1). Nie widać ich w podglądzie Markdown, a model je czyta.

Co robi i kiedy

Zapisuje w trwałym pliku czynności, które może wykonać tylko osoba prowadząca, jako dokładne polecenia do uruchomienia, gdy sesja natrafia na blokadę, której agent nie może sam obejść.

Oryginalny opis (EN)

Write the acts owed to the driver — not decisions, ACTS only the driver can perform — as exact runnable commands in a durable file. Use when a session hits a change a client-side guard refuses, a staged-to-live promote whose scope its token does not hold, a workflow dispatch its App must not be granted, or a gate override reserved to a human, and would otherwise describe the blocker as prose in a console nobody is re-reading. Not for presenting a DECISION (that is `ask-decision`) and not for a multi-session operation's record (that is `author-drive-plan`).

Jak działa

  1. When it runs
  2. The entry — four parts, every time
  3. Never a credential in a runsheet
  4. Three boundaries
  5. Four example entries
  6. The closing rule
Sekcje z nagłówków SKILL.md, w oryginale · 128 linii, 1 316 słów
Podgląd SKILL.md
---
name: human-runsheet
description: >-
  Write the acts owed to the driver — not decisions, ACTS only the driver can perform — as exact
  runnable commands in a durable file. Use when a session hits a change a client-side guard
  refuses, a staged-to-live promote whose scope its token does not hold, a workflow dispatch its
  App must not be granted, or a gate override reserved to a human, and would otherwise describe
  the blocker as prose in a console nobody is re-reading. Not for presenting a DECISION (that is
  `ask-decision`) and not for a multi-session operation's record (that is `author-drive-plan`).
---
<!-- GENERATED from plugins/assay/skills/human-runsheet/SKILL.md by deskinstall --harness cursor. Do not edit. -->

# Human runsheet — the acts owed to the driver, written as exact commands

A session that hits a structural block — a guard it cannot clear, a scope its own token does not
hold, a permission its App must never be granted, a gate only a human may override — has exactly
one honest thing to do with that fact: hand it to the driver as something they can *run*, not as
a paragraph they have to reconstruct into a command. "I am blocked, here is what I think you need
to do" is lossy the moment the session that wrote it ends. A file of exact `! <command>` lines is
not.

## When it runs

A session invokes this the moment it recognises the blocker is not "I don't know how" but
"I structurally cannot" — the four classes below cover the shapes that recur:

- a change a client-side classifier or guard refuses to let the session write or commit;
- a staged-to-live promote whose scope the session's own token does not carry;
- a workflow dispatch that the session's App is deliberately not granted the permission to run;
- a gate override reserved to a human (a required check, a merge, a tag) that no App holds
  authority to waive.

It does not run for a DECISION — a fork where the driver picks among options and the session can
proceed once told which. That is `ask-decision`'s job: it presents options with a recommended
default and relays the ruling. This skill is for the opposite shape: the driver has already,
implicitly, made the call by being the only party who *can* act — there is nothing to decide,
only something to run.

## The entry — four parts, every time

`RUNSHEET.md`, kept in the session's own scratch home — **never the repo tree**. It is session
state describing what this session could not do, not a repo artifact anyone else's tooling reads.
One entry per owed act, in this shape:

1. **The command.** A line starting `! `, followed by the exact command, runnable as written,
   every placeholder already resolved to a real value. A driver who copies the line and presses
   enter should not have to fill anything in first.
2. **Why it is owed to the driver.** The specific structural reason — quote the guard's refusal,
   name the scope the token lacks, name the permission the App does not hold, name the gate.
   "I couldn't do it" is not a reason; the refusal text is.
3. **What the session did instead.** Parked the change at a stated path, filed the escalation
   issue, worked around it within its own authority where one existed — whatever actually
   happened, stated plainly.
4. **The resume step.** What the session will do once the act lands, and how it will notice —
   which file changed, which run turned green, which label came off. A runsheet entry with no
   resume step is a dead end dressed as a hand-off.

## Never a credential in a runsheet

A line names the **identity** the command must run under — "as the driver's own account", "with
admin rights on the repo" — and never carries a token, a key, or a password, and never instructs
the driver to hand one over to the session. A runsheet is a set of commands, not a place secrets
pass through. If a command as actually run would need a secret filled in, the entry says so and
leaves it to the driver to supply it directly when they run it — the file itself never carries it.

## Three boundaries

- **Against `ask-decision`.** `ask-decision` presents a DECISION — the driver chooses among
  options and the session proceeds on the answer. This presents an ACT — there is no option set,
  only one command only the driver can run. A blocker that is genuinely a choice belongs there,
  not here.
- **Against `author-drive-plan`.** `author-drive-plan` is the durable record of a multi-session
  OPERATION — where a migration or a long sweep left off, across many sessions and many briefs.
  A runsheet is narrower and more perishable: the acts owed right now, by the session that wrote
  it, this run. A drive plan may reference that a runsheet exists; it does not replace one, and a
  runsheet is not a substitute for a drive plan when the work truly spans sessions.
- **Not a second escalation channel.** The filed issue (or PR comment) the session already opens
  for the blocker remains the escalation — the thing the driver's queue surfaces and the thing
  that ages if unanswered. The runsheet accompanies it as the exact command form; it is never
  written instead of filing, and it is never treated as having escalated anything on its own.

## Four example entries

Stripped of any house's own tool names — each names a class from `## When it runs` above, in the
four-part shape.

**Example 1 — a client-side guard refuses the write.** A content guard on the session's own write
path refuses a diff whose content matches a pattern it cannot clear on its own authority.

! git apply /path/to/parked.patch

Why: the guard binds this session's write path only, on the *text* of the change, and only a
human accepting the material outside that guard can land it. What the session did: left the
diff at the stated path and filed the blocking issue. Resume: on next boot, check whether the
patch's target file changed on the branch; if so, continue from the following step.

**Example 2 — a staged-to-live promote outside the token's scope.** The session's own token is
scoped to staging; promoting to the live environment needs a broader scope no App should hold
standing.

! deploy-promote --from staging --to production --release <release-id>

Why: `deploy-promote` is the driver's own promote tool, run under the driver's own account —
this skill names it only as an example command, never as a verb it defines or grants. What the
session did: staged the release and stopped short of promoting it. Resume: poll the release's
own status until it reads live, then continue the rollout checklist.

**Example 3 — a workflow dispatch the App must not hold.** Firing a given workflow needs a write
scope the session's App is deliberately kept without, because that scope can also cancel
unrelated runs or disable other workflows.

! gh workflow run <workflow-file> --ref <branch>

Why: the App's permission is deliberately read-only on workflow dispatch; broadening it to
dispatch one workflow also grants canceling and disabling every other one. What the session
did: prepared the branch the workflow should run against and stopped. Resume: watch for the
run to appear and pick up at the step gated on its result.

**Example 4 — a gate override reserved to a human.** A required check is red for a reason the
session has verified is unrelated to its own change, and only a human may override the merge
gate.

! gh pr merge <number> --admin --merge

Why: an admin override of a required check is a human-only gate by design — no App holds merge
authority over a protected branch. What the session did: posted the evidence that the red check
is unrelated and left the PR open. Resume: once merged, continue with the next dependent item.

Four illustrative entries, not a catalogue to keep current — a real runsheet holds only the acts
a given session actually owes, in whatever number that turns out to be.

## The closing rule

A runsheet accompanies the filed escalation; it never replaces it. If a driver reads only the
runsheet and the underlying issue was never filed, the escalation did not happen — the file is
the command form of something that must also exist as a durable, driver-visible record. Write
both, every time.

Wersje

RepozytoriumWynikLicencja
Ta strona · medici-finance/assay100Apache-2.0
medici-finance/assay100Apache-2.0

Migawka bazy z 13.11.2026 · wersja paczki 4da2c7b5cdaf (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 95ask-decision
    Przedstawia oczekujące decyzje człowieka pojedynczo: kontekst, opcje z domyślną na pierwszym miejscu i gotową akcję do wykonania.
  • 95ask-decision
    Przedstawia oczekujące decyzje człowieka pojedynczo: kontekst, opcje z domyślną na pierwszym miejscu i gotową akcję do wykonania.
  • 100parawatch
    Opisuje, jak agent pracuje z człowiekiem przez parawatch: pyta o decyzje, zatwierdzenia i czynności, których nie może wykonać sam.
  • 100dispatch-first
    Sprawdza przed planowaniem, zgłaszaniem issue, pytaniem człowieka lub startem pracy, czy ktoś już śledzi to zadanie albo ma w nim decyzję, używając polecenia dispatch.
  • 100human-gate
    Prowadzi ścieżkę weryfikacji przez człowieka w pętli agenta i dowodzi, że przegląd się odbył: buduje stronę HTML do przeglądu, zbiera uwagi jako artefakt i blokuje zamknięcie przy otwartym BLOCKER.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.