autopilot-github-execution
/autopilot-github-execution
Wdraża zmianę ze zgłoszenia GitHub do gotowego pull requesta: wyjaśnia niejasne wymagania i nie zgaduje kryteriów.
- Autor
- anas-zenedge ↗
- Repozytorium
- agent-skills ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 07.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 07.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Wdraża zmianę ze zgłoszenia GitHub do gotowego pull requesta: wyjaśnia niejasne wymagania i nie zgaduje kryteriów.
Oryginalny opis (EN)
Use when implementing a GitHub issue or other GitHub change request, when the user says autopilot or ready for PR, or when an issue is unclear, contradictory, or missing acceptance criteria. Also use when tempted to guess the requirement, skip review, ask the user to approve a settled issue, or implement it yourself because the change looks small.
Jak działa
- Overview
- When to Use
- Understand gate
- Recipe
- Overrides
- Final message
- Quick reference
- Rationalizations
- Red flags
- Common mistakes
Podgląd SKILL.md
---
name: autopilot-github-execution
description: Use when implementing a GitHub issue or other GitHub change request, when the user says autopilot or ready for PR, or when an issue is unclear, contradictory, or missing acceptance criteria. Also use when tempted to guess the requirement, skip review, ask the user to approve a settled issue, or implement it yourself because the change looks small.
---
# Autopilot GitHub Execution
## Overview
Implement a GitHub change request without checking in, unless you cannot tell what done means.
**Core principle:** The issue, read in full, is the only spec. If you cannot state done without inventing a product rule, stop. If you can, run the recipe. A subagent approves the spec and plan. You do not.
**Violating the letter of the gates is violating the spirit of the gates.**
## When to Use
- User gives a GitHub issue URL and asks to implement, fix, or deliver it
- User says autopilot, or ready for PR, on a change request
- Issue is empty, contradictory, or missing acceptance criteria
- About to guess, skip a gate, or code it yourself because it looks small
**When NOT to use:** explaining an issue, reviewing an existing PR, or the user said not to implement.
```dot
digraph understand_gate {
"Issue and comments read?" [shape=diamond];
"Read them with GitHub MCP" [shape=box];
"Can you state done without inventing a product rule?" [shape=diamond];
"STOP" [shape=octagon, style=filled, fillcolor=red, fontcolor=white];
"Run the recipe" [shape=doublecircle];
"Issue and comments read?" -> "Read them with GitHub MCP" [label="no"];
"Read them with GitHub MCP" -> "Can you state done without inventing a product rule?";
"Issue and comments read?" -> "Can you state done without inventing a product rule?" [label="yes"];
"Can you state done without inventing a product rule?" -> "STOP" [label="no"];
"Can you state done without inventing a product rule?" -> "Run the recipe" [label="yes"];
}
```
## Understand gate
Read the issue body, labels, and every comment with GitHub MCP before scoping. Title-only is not read. A fetch failure is a failed gate. The user's paraphrase is not the spec.
You understand it only if you can write, from that text, all of:
- requested behavior
- acceptance criteria
- out-of-scope items the issue states
You do not understand it if a user-visible rule is missing, contradicted, or left for someone to make a reasonable call. Design choices the issue leaves open (file layout, internal names) are not this gate.
**You apply this gate. Do not delegate it.** If a subagent would have to decide what the issue is asking, you do not understand it.
### Stop message
The message IS these parts, in order, and nothing else:
1. Blocked. The change request is not understood.
2. Issue URL
3. The contradiction or missing fact
4. The one question that would unblock
No design, no branch, no partial fix, no safe subset.
Example: Product says a credit never changes the tax line. Finance says it must reduce tax proportionally. Stop. Ask which rule is the requirement. Do not pick.
## Recipe
Run every step, including for a one-line change. Documents may be short. Gates may not be skipped.
1. **Isolate.** **REQUIRED SUB-SKILL:** superpowers:using-git-worktrees. Do not implement on the default branch.
2. **Design.** **REQUIRED SUB-SKILL:** superpowers:brainstorming for the design and spec file only. Do not ask the user. Do not offer a visual companion. Do not wait for user approval.
3. **Spec decision and review.** Dispatch a fresh general-purpose subagent on the most capable model (see Model). It decides open design choices and returns only APPROVE or BLOCK. BLOCK for a missing product rule → stop message. BLOCK for a design defect → revise and re-review. No self-approval.
4. **Plan.** **REQUIRED SUB-SKILL:** superpowers:writing-plans. Do not offer an execution choice.
5. **Plan decision and review.** Fresh highest-reasoning subagent. Same APPROVE or BLOCK rule. Then implement. Do not ask which executor to use.
6. **Implement.** **REQUIRED SUB-SKILL:** superpowers:subagent-driven-development. The controller does not write the product code. Implementers use superpowers:test-driven-development.
7. **Review the branch.** **REQUIRED SUB-SKILL:** superpowers:requesting-code-review, highest-reasoning subagent. **REQUIRED SUB-SKILL:** superpowers:verification-before-completion before any ready claim.
8. **Report.** Do not invoke superpowers:finishing-a-development-branch. Do not push, open a PR, or merge.
### Model
Name the model on every subagent dispatch. A dispatch without one inherits the controller's model, so every subagent runs on the most expensive tier.
- **Spec verdict, plan verdict, and final branch review:** the most capable model. Also state in the dispatch prompt that the task needs the deepest reasoning. Always dispatch a general-purpose subagent, never a read-only explore or search agent, for these three. Do not downgrade these three to save time.
- **Every other dispatch** (implementation, fixes, re-reviews, searches, builds, test runs): follow subagent-driven-development's Model Selection. Pick the cheapest model that is reliable for the task.
Record the model that actually ran from the task result. Do not claim a higher model ran.
### Named skip
Skip a gate only when the user's latest message names that gate ("don't commit", "implement it yourself", "skip the plan", "open the PR"). Quote it in the decision log. Skip only that gate.
"Implement it yourself" names only the implement gate. Spec verdict, plan verdict, and branch review still run.
"Don't use subagents" or "no subagents" names every subagent gate: spec verdict, plan verdict, subagent-driven-development, and branch review. Skip those. Do not replace them with self-review and call them done. The understand gate still runs. Do not report ready for PR if branch review was skipped. Report the quote and what is unverified.
Not a named skip: a deadline, "don't stall", "skip the ceremony", "one-liner", "you know the codebase", "don't bother me", a senior's suggestion, a reporter saying "make a reasonable call."
### Follow-up issues
Filing any issue during the run (a plan task, a review finding, an out-of-scope bug): **REQUIRED SUB-SKILL:** filing-follow-up-issues, before the create command.
## Overrides
When this skill applies and the understand gate passes, these do not apply:
- brainstorming's user-approval hard gate and its questions to the user
- writing-plans' execution-choice prompt
- finishing-a-development-branch's option menu
A user instruction that names a gate still wins over this skill.
## Final message
In order:
1. Issue URL and understand-gate PASS
2. Branch and commit range
3. Decisions. Each is question, choice, who decided, model
4. Final branch review verdict and test evidence
5. Ready for PR: yes. PR not opened.
Confirm the change is committed before this message. Present every subagent decision here, not only in a file.
## Quick reference
| Situation | Action |
|---|---|
| Issue unread | GitHub MCP: body, labels, comments |
| Cannot state done | Stop message. End. |
| Understood | Recipe. No user check-in. |
| One-line change | Same recipe. Short docs. |
| Paraphrase conflicts with issue | Stop. Issue wins until the user amends it. |
| Ready for PR | Committed, reviewed, reported. Not pushed. |
## Rationalizations
| Excuse | Reality |
|---|---|
| Ceremony cannot change a one-liner | The recipe still runs. Docs can be short. |
| Don't stall, the meeting is in 15 minutes | Time pressure is not a named skip. |
| I'll implement it myself | The controller does not write the product code. |
| I already self-reviewed the plan | Self-review is not the spec or plan verdict. |
| The user didn't ask for subagents or a commit | This skill is that request. |
| Brainstorming says wait for the user | The subagent approves. See Overrides. |
| The subagent can pick the product rule | Failed understand gate. Stop. |
| Ready for PR means open it | Open or push only if the user named that gate. |
| The user's sentence is enough | Unread comments are an unread issue. |
| I'll note that I skipped SDD | Noting a skip is not permission. Run the step. |
| "Don't use subagents" doesn't name spec review | A subagent ban names every subagent gate. Do not dispatch one anyway. |
| Leave the model unset, the default is fine | Unset inherits the controller's model. Name it: most capable for the three verdicts, cheapest reliable for the rest. |
## Red flags
STOP and return to the gate you skipped:
- Coding the issue yourself
- Spec or plan with no subagent verdict
- Asking the user to approve a settled issue
- Asking a subagent what the requirement is
- A safe subset shipped as the issue
- Push, PR, or merge after ready for PR
- Ceremony, one-liner, don't stall, skip the ceremony
- "No subagents" followed by a spec, plan, or branch-review dispatch
- A subagent dispatch with no model named
- A spec, plan, or branch-review dispatch on less than the most capable model
## Common mistakes
- Treating a reporter's "don't block on me" as acceptance criteria
- Letting the user's paraphrase override unread issue comments
- Invoking finishing-a-development-branch and waiting on a menu
- Recording a skipped gate, then skipping it anyway
Wersje
W bazie nie ma innych skilli o nazwie autopilot-github-execution.
Migawka bazy z 13.11.2026 · wersja paczki 8fecbf3a6947 (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
- 100issue-run
Wyświetla ostatnie otwarte zgłoszenia GitHub lub Linear, pozwala wybrać jedno i uruchamia na nim autopilota przez skill run. - 100implement-issue
Czyta zgłoszenie GitHub, tworzy gałąź, implementuje rozwiązanie i otwiera pull request na podstawie numeru zgłoszenia. - 100github-fix-issue
Naprawia zgłoszenia GitHub od początku do końca: analiza, utworzenie gałęzi, implementacja, testy i wysłanie pull requesta, na podstawie numeru lub adresu zgłoszenia. - 100autopilot-ship
Scala gotowy pull request, śledzi wdrożenie na produkcji, sprawdza działanie, szuka regresji i zamyka zgłoszenie. - 86autopilot-implement
Wdraża zaplanowane zmiany od gałęzi po pull request, generując projekty Stitch, kod i walidację QA. - 100git-issue
Prowadzi od zgłoszenia GitHub do pull requesta: tworzy gałąź, implementuje przez TDD i otwiera PR, dla jednego zgłoszenia lub kilku równolegle.