원클릭으로
agent-issues
Use in repositories that track work with the agent-issues local issue workflow.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
메뉴
Use in repositories that track work with the agent-issues local issue workflow.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Abandon an active issue or release a stale issue claim so another agent can pick it up.
Claim exactly one issue, fix it, and create a pull request starting from a clean branch.
Prepare and open a pull request for the current branch after validating the full diff and running checks.
Push the current branch, open or update its PR, and loop through CI failures and review feedback until the PR is clean.
Like solve-issue, but skips plan confirmation and goes straight to implementation.
| name | agent-issues |
| description | Use in repositories that track work with the agent-issues local issue workflow. |
Use this skill when a repository uses agent-issues for local issue tracking,
agent branch workflow, and PR submission.
Read the repository's AGENTS.md or equivalent local agent instructions first.
Those files should contain project-specific policy: validation commands,
generated-output rules, product constraints, and repo-specific priority rubrics.
The shared rules below describe the common agent-issues workflow.
Issues live as individual JSON5 files under issues/ in the repository root.
The filename is the issue ID and must start with p0-, p1-, p2-, p3-,
p4-, or blocked-.
Resolved issues are deleted in the PR that fixes them. Do not mark resolved issues as closed in place.
Use a stable sequencing token in related issue filenames so ls issues/ keeps
the series grouped and ordered, for example blocked-migration-step5.json5
and later p3-migration-step5.json5.
Issue files use this shape:
{
"title": "Short summary of the issue",
"description": "Full description with context...",
"status": "open",
"priority": 3,
"type": "task",
"labels": ["backend"],
"created_at": "2026-02-09T14:30:00.000000-08:00",
"updated_at": "2026-02-09T14:30:00.000000-08:00"
}
Use real timestamps, not placeholder times.
Supported fields:
| Field | Type | Description |
|---|---|---|
title | string | Short summary |
description | string | Full description with context |
status | string | Always "open" |
priority | int | 0 highest through 4 lowest, unless a repo-specific rubric narrows this range |
type | string | Usually "task" |
labels | string[] | Tags for categorization |
created_at | string | ISO 8601 timestamp |
updated_at | string | ISO 8601 timestamp |
blocked | bool | string? | If truthy, filename must start with blocked- and issue-autoclaim skips it |
Create blocked issues when work requires human input or approval before an
agent can safely act. Use a blocked string that states the specific decision,
credential, external dependency, cost/rate-limit concern, or scope approval
needed.
Document follow-up work you discover by adding local issue files. Do not create, close, or comment on GitHub Issues unless the user explicitly asks.
Use the shared CLI instead of repo-local scripts:
issue-query
issue-query --label backend
issue-query --max-priority 2
issue-query --search "streaming"
issue-autoclaim
issue-autoclaim <issue-name>
issue-claim <issue-name>
issue-claim --current
issue-claim --list
issue-fmt
issue-lint
issue-autoclaim is the normal claim path because it merges the default branch
first and respects the shared local claim store. issue-claim is for narrower
cases where the caller explicitly needs to claim without that merge.
Releasing a stale claim is always authorized and does not require user approval. A claim is stale when the prior work is demonstrably finished or no longer active, for example when its issue was removed by a merged PR, or when the worktree is clean at the default branch with no open PR after completing the prior task.
When a stale claim prevents new work, run issue-abandon and retry the intended
claim. Treat this as claim-store bookkeeping: do not reset, switch, or clean the
branch when preflight already shows that the worktree is ready. If the prior
claim might still have uncommitted work or an open PR, it is not clearly stale;
preserve it and ask before abandoning it.
After editing issue files, run:
issue-fmt
issue-lint
Agents usually work on a branch with two random English words in the name. Stay on the current branch unless the user explicitly asks you to change branches. Do not create replacement branches as part of normal issue or PR workflow.
Merge the default branch into the work branch when the workflow calls for updates. Do not rebase or force-push unless the user explicitly approves that operation.
If a working tree has unexpected changes, treat that as important context. Do not silently discard, revert, or exclude files from the PR. Inspect the changes and ask the user when ownership is unclear or the changes affect the task.
Use the shared PR skills when available:
solve-issue claims exactly one issue, fixes it, deletes the issue file, and
submits the PR.create-pr prepares and opens a PR for already-completed local changes.submit-pr pushes the branch, creates or updates the PR, and watches CI and
review feedback through agent-submit.For direct CLI use, create or update PRs with:
agent-submit --title "Updated title" --body "$(cat <<'EOF'
...updated description...
EOF
)"
agent-submit is the publishing path for both new PRs and updates. Do not use a
raw git push plus manual gh pr edit as a substitute for normal PR updates.
If CI fails or review feedback arrives, fix the root cause, commit, and submit
again.
Repos that need local automation on every PR submit can add
.agent-issues/submit-hooks.json5:
{
prepare: ["make agent-submit-prepare"],
after_publish: ["make agent-submit-after-publish"],
}
agent-submit runs prepare commands after its basic clean-tree preflight and
before pushing. Use this phase for repo-owned validation, generated artifact
refreshes, and commits of intentional generated diffs. A prepare hook may change
HEAD by creating commits, but it must finish on the same branch with a clean
working tree.
agent-submit runs after_publish commands after pushing and creating or
updating the PR, but before CI/review watching. Use this phase for signoff or
status creation that needs the final pushed SHA. The hook receives
AGENT_SUBMIT_SHA, AGENT_SUBMIT_BRANCH, AGENT_SUBMIT_BASE,
AGENT_SUBMIT_PR_NUMBER, AGENT_SUBMIT_PHASE, and AGENT_SUBMIT_REPO_ROOT.
An after-publish hook must not change branches, change HEAD, or leave a dirty
working tree.
Do not teach agents to bypass submit hooks with raw git push, gh pr edit,
or direct signoff commands. Put the repo-specific automation behind hook
commands and let agents use agent-submit.
Repo-local skills should contain only the parts that are specific to that
repository. Keep PR creation, PR updates, CI watching, review feedback loops,
timeout handling, local issue claiming, and GitHub Issue policy in the shared
create-pr, solve-issue, submit-pr, and agent-issues skills.
Use this shape for .claude/skills/create-pr-local/SKILL.md or
.agents/skills/create-pr-local/SKILL.md:
---
name: create-pr-local
description: Repo-specific instructions to use alongside the global create-pr skill.
---
# Create a Pull Request: Local Notes
Use these notes together with the global `create-pr` skill.
## Pre-Validation
- Read `Makefile` first and prefer existing project targets over direct tool
invocations.
- Before opening a PR, run `<repo test command>` and `<repo lint command>`.
- If you change generated-output inputs, run `<regeneration command>` and
commit intentional generated diffs.
- If `.agent-issues/submit-hooks.json5` owns validation or generated artifact
commits, say that agents should rely on `agent-submit` for those steps.
## Repo-Specific Constraints
- Do not hand-edit generated files under `<generated path>`.
- If validation dirties tracked files, inspect the diff and commit intentional
artifacts before creating the PR.
- If adding a dependency, document the repo-specific selection criteria and any
version policy that reviewers expect.
Use this shape for .claude/skills/solve-issue-local/SKILL.md or
.agents/skills/solve-issue-local/SKILL.md:
---
name: solve-issue-local
description: Repo-specific instructions to use alongside the global solve-issue skill.
---
# Solve an Issue: Local Notes
Use these notes together with the global `solve-issue` skill.
## Repo-Specific Commands
- Baseline verification is `<repo test command>` and `<repo lint command>`.
- If you touch `<generated input path>`, run `<regeneration command>` and
inspect the generated diff before committing it.
- If a targeted check exists for the touched area, run it before the full suite.
- If `.agent-issues/submit-hooks.json5` owns validation or generated artifact
commits, say that agents should rely on `agent-submit` for those steps.
## Repo-Specific Constraints
- Follow `<domain rule>` when changing `<area>`.
- Do not hand-edit generated files under `<generated path>`.
Good local skills may also list dependency approval requirements, generated
artifact commit hygiene, product/domain constraints, or project-specific review
steps. They should not describe how to call agent-submit, how to update PR
metadata, how to interpret watcher exit codes, how many CI/review loops to run,
or whether to create GitHub Issues unless the repo deliberately overrides the
shared policy. They also should not tell agents to run direct signoff/status
commands when submit hooks can own that automation.
Put repository-local temporary files under tmp/ in the worktree unless the
repo says otherwise. Avoid /tmp on hosts where tmpfs space is constrained.
Put repository-local logs under logs/ unless the repo defines a more specific
log location.