dev-sprint
暂无简介
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-15
- Stars
- ★ 4
- Forks
- 1
02 资源文件 4
- 参考文档references/plugin-augmentation.md1.3 KB
- 脚本scripts/create-composite.sh3.2 KB
- 脚本scripts/list-issues.sh4.6 KB
- 脚本scripts/resolve-base-branch.sh4.1 KB
03 SKILL.md 正文
<!-- markdownlint-disable MD003 MD013 MD022 MD041 -->
name: dev-sprint description: Use when user asks "start sprint", "pick up work", "what should I work on". Selects 1-5 related issues from active milestone, synthesizes into composite work item (or picks single issue), hands off to ship-issue. model: haiku resources:
- scripts/list-issues.sh
- scripts/create-composite.sh
- scripts/resolve-base-branch.sh
- references/plugin-augmentation.md
Dev Sprint
Pick 1-5 related issues, prioritizing lowest active milestone. Run a single issue by choice, or group related issues: same subsystem, shared code paths, blocking order.
Prioritize bug fixes over new features.
Pick one issue deliberately, or as many as five. Either way, skip composite creation only if it's a single issue (Phase 3) — but still run the full Phase 4 handoff, not a stripped-down one. "Directly" means "no composite," NOT "skip the augmentation."
Synthesize into composite.
Hand off to ship-issue. ship-issue owns branching, TDD, CI, pre-pr-review, merging. dev-sprint owns selection + synthesis.
Single-issue runs are not a fast path. Whether the sprint is one issue or five, Phase 4's handoff augmentation is mandatory: the required pre-pr-review scans (slop-scan, etc.), the ponytail:ponytail-review pass after pre-pr-review, the propagated directives (subagent/executing-plans, review autonomy), and Phases 5–8 (docs, semver, changelog, deferred-work filing, verification). The only thing a single issue skips is building a composite issue (Phase 3) and the multi-issue closing refs. Everything else runs exactly as it does for a multi-issue sprint.
Hard rule — no orphan sprints: Once a sprint composite is created (Phase 3), it owns the run. If scope must shrink, update the composite or close it with justification (see Phase 3.5). Never create a second composite while the first is open.
Setup
Create .tmp/ at project root before starting:
mkdir -p .tmp
All temp files in dev-sprint + sub-skills (ship-issue, pre-pr-review, etc.) use .tmp/ at project root, not /tmp. Keeps artifacts local, no system temp interference.
Phase 1: Triage Unmilestoned Issues
Before selecting work, every open issue must have a milestone. List the orphans directly:
"${CLAUDE_CONFIG_DIR}/skills/dev-sprint/scripts/list-issues.sh" --no-milestone
(Add --json for machine-readable output to pipe through jq/python — full bodies included, so no follow-up gh issue view needed.)
Assign each unmilestoned issue to the right milestone:
-
Survey existing milestones first. List with due dates + themes to place each orphan accurately:
gh api repos/:owner/:repo/milestones --jq '.[] | "\(.title) — \(.description // "no description") (due \(.due_on // "none"))"' -
Match each orphan to best-fitting milestone by topic/subsystem + ordering (which release it belongs to). Bugs in shipped behavior → earliest open milestone; new features → milestone owning that area.
gh issue edit <issue> --milestone "<milestone-title>" -
No existing milestone fits? OK to create one — or rename existing milestones for better ordering or topic coverage. Prefer reuse/rename over proliferating near-duplicate milestones.
# create gh api repos/:owner/:repo/milestones -f title="vX.Y" -f description="<theme>" -f due_on="<ISO8601 or omit>" # rename / re-describe (milestones are addressed by number, not title) gh api -X PATCH repos/:owner/:repo/milestones/<number> -f title="<new title>" -f description="<new theme>"
Goal: after this phase, No Milestone is empty (or holds only issues you've deliberately, with justification, decided are out of scope for any milestone). Re-run list-issues.sh to confirm before selection.
Phase 2: Select Related Issues
"${CLAUDE_CONFIG_DIR}/skills/dev-sprint/scripts/list-issues.sh"
Pick lowest milestone with open issues. Choose 1-5 related issues from that milestone: one issue alone (deliberately), or multiple related issues (same subsystem, shared code paths, blocking order).
Prefer: shared files, unblock each other. Avoid: mixed domains.
Phase 3: Create Composite Issue
Composite body structure:
## Issues
- #N — <title>
- #N — <title>
## Scope
<Shared subsystem/data flow/behavior. Why together.>
## Acceptance Criteria
<Merged & deduplicated criteria from all issues>
## Implementation Notes
<Ordering constraints. Known shared files.>
CRITICAL: List all sub-issues in composite body. Source of truth for what PR must close.
"${CLAUDE_CONFIG_DIR}/skills/dev-sprint/scripts/create-composite.sh" "Sprint: <theme>" "<body>" "<milestone>" "<labels>"
Milestone required; labels optional (upstream ship-issue/pr-review no longer enforce labels, label sets vary per repo). Labels = 4th arg, comma-separated for multiple — e.g. "bug,area-core". Use labels that exist in repo (gh label list); script validates each, attaches the existing ones, skips missing ones with a warning rather than failing. Script prints new issue number on its final line—capture it, pass to ship-issue.
Phase 3.5: Composite is Immutable (READ BEFORE PROCEEDING)
Once the composite issue is created, it IS the sprint. You may not silently abandon it and create a new one.
This is the orphan-sprint failure mode: composite #A is created with sub-issues #X, #Y, #Z; mid-run you decide #Y and #Z are too much; you create composite #B covering only #X and ship that; #A is left OPEN with #Y and #Z dangling, milestone view corrupted, planned work silently dropped.
Forbidden: creating a second composite during the same dev-sprint run while the first is still OPEN. Tempted to? Stop, apply one resolution below instead.
If scope must shrink after Phase 3 (e.g. one bug harder than expected, another turns out blocked):
- Preferred — keep the composite, shrink the PR: Pick the subset you'll actually ship. Update composite body to mark deferred sub-issues:
## Issues
- #X — <title> ✅ (this PR)
- #Y — <title> ⏸ deferred → #<followup>
- #Z — <title> ⏸ deferred → #<followup>
For each deferred sub-issue, post a comment explaining the defer + linking the follow-up plan (or leave it open with no comment if follow-up is just "next sprint"). Composite stays open until its PR merges; on merge, Closes #<composite> + Closes #X only. #Y and #Z stay open for a future sprint.
gh issue edit <composite> --body "$(updated body)"
gh issue comment <Y> --body "Deferred from sprint #<composite>. Will be picked up in a follow-up sprint."
-
Alternative — abandon the composite cleanly: Original framing wrong and you want a fresh composite? You must close the old one first with a justification comment:
gh issue comment <composite> --body "Closing as superseded. Original scope (#X #Y #Z) was too broad for one sprint. Replacing with #<new-composite> covering #X only. #Y and #Z return to milestone backlog." gh issue close <composite>Only after the old composite is CLOSED may you create a new one.
Self-check before running create-composite.sh a second time in one run: Has a composite already been created in this run? If yes, you are in the orphan path. STOP and apply resolution 1 or 2 above. Never create a second composite without closing or updating the first.
Phase 4: Hand Off to ship-issue
Applies to single-issue runs too. Where this phase says "composite-id," use the single issue's number instead. The base-branch resolution, scans, ponytail pass, propagated directives, and PR verification all run regardless of issue count.
Phase 4a: Select the base branch (milestone → release branch)
The composite's milestone decides the base branch — not main by default. A milestone titled X.Y — <theme> (or vX.Y) is a patch line: if a release/X.Y.x branch exists on the remote, branch from it so the work doesn't drag in unreleased feature work that lives only on main (and doesn't duplicate commits already on the release branch). Only fall back to the default branch when no matching release/X.Y.x exists.
Use the script — don't eyeball it. resolve-base-branch.sh performs this evaluation deterministically (same result every run) and prints the base branch on stdout:
# By composite issue number (looks up its milestone via gh):
BASE_BRANCH="$("${CLAUDE_CONFIG_DIR}/skills/dev-sprint/scripts/resolve-base-branch.sh" --issue <composite-id>)"
# …or by milestone title directly:
BASE_BRANCH="$("${CLAUDE_CONFIG_DIR}/skills/dev-sprint/scripts/resolve-base-branch.sh" --milestone "<milestone>")"
echo "Base branch for this sprint: $BASE_BRANCH"
The script falls back to the repo default branch when the milestone has no matching release/X.Y.x (a milestone with no release line is not an error). Diagnostics go to stderr; only the branch name is on stdout, safe for $(...) capture.
Pull latest of $BASE_BRANCH before invoking. Pass $BASE_BRANCH into the ship-issue handoff (see example below) so it branches from the right place instead of assuming main. Check references/plugin-augmentation.md for plugin, include in message.
IMPORTANT: All temp files in dev-sprint + sub-skills use .tmp/ at project root. Ensure .tmp/ exists before invoking sub-skills. Sub-skills must not use /tmp.
CRITICAL: Include sub-issue list in ship-issue invocation. Propagates to pre-pr-review so PR has all closing refs.
Example message:
Sprint: Validation & Error Handling
Base branch: release/1.0.x ← branch from this, NOT main (from Phase 4a)
Sub-issues to close on PR merge:
- #315 — Code quality: socket validation
- #325 — P0: Add error handling and validation
- #330 — Sprint composite issue
[Continue with ship-issue instructions...]
For agentic workers: REQUIRED: Use superpowers:subagent-driven-development (if subagents available) or superpowers:executing-plans to implement this plan. Steps use checkbox (- [ ]) syntax for tracking.
Applies to any sub-skill that launches agents to do the work (ship-issue, pre-pr-review). Include this directive in the handoff message so it propagates.
Review phase autonomy: ship-issue → pre-pr-review fixes apply automatically, no user prompts. Fix or defer, then create PR.
Required scans during pre-pr-review: Include in ship-issue handoff so it propagates to pre-pr-review:
-
/Users/phaedrus/git/my-llmenv/bundles/base/scripts/slop-scan.sh scan . --lintAlways invoke slop-scan through this pinned wrapper (it runs
slop-scan@0.3.0and is on the permission allowlist). Do not callnpx slop-scandirectly — the unpinned, agent-chosen form is blocked by the supply-chain auto-policy.
After pre-pr-review completes: Run ponytail:ponytail-review and apply all suggestions automatically before creating the PR. No user prompts — apply and commit.
PR closing references (MANDATORY):
Closes #<composite-id>
Closes #<issue-1>
Closes #<issue-2>
Closes #<issue-3>
Composite first, then sub-issues. GitHub auto-closes on merge. Every sprint PR closes all sub-issues, not just composite.
Sprint meta-issue: Sprint tracked via issue? Add closing ref:
Closes #<composite-id>
Closes #<issue-1>
...
Closes #<sprint-tracking-issue>
Branch (created by ship-issue from $BASE_BRANCH selected in Phase 4a):
feat/<composite-id>-<theme-slug> # forked from $BASE_BRANCH (release/X.Y.x or default)
Follow ship-issue exactly—no skips.
Phase 5: Documentation & Semver
After ship-issue, update docs + version.
Docs: Update README.md, docs/ for:
- Features/APIs → examples + usage
- Behavior changes → update sections
- Types/functions → API docs
- Config changes → options/env vars
Semver:
- MAJOR: Breaking API/behavior/config
- MINOR: New features, APIs, config options (backward-compatible)
- PATCH: Bug fixes, internal improvements
Update package.json / Cargo.toml / pyproject.toml + any version refs. Commit with code.
CHANGELOG (always run the keepachangelog skill at the end of the sprint): Invoke the keepachangelog skill to update CHANGELOG.md if the sprint introduced any user-visible changes (features, behavior changes, config changes, bug fixes). If keepachangelog modifies CHANGELOG.md, commit it. If nothing warrants a changelog entry, leave it unchanged. This runs on every sprint — never skip it.
Phase 6: Deferred Work & Pre-Existing Issues
Defer only if:
- Pre-existing bugs outside sprint scope
- Features needing brainstorm/design
- Work needing
feature-devor architecture review - External dependency blocks
Do NOT defer: Fix non-complex issues immediately, mid-sprint OK.
gh issue create \
--title "Title" \
--body "Found during sprint X, [details]" \
--milestone "vX.X" \
--label "bug" # optional; use a label that exists in this repo (gh label list)
REQUIRED: All issues from any skill need:
--milestone(no exceptions)--labelis optional — if used, must be a label that exists in repo (gh label list). Don't invent labels.
Phase 6a: Issue Creation Standards (All Skills)
MANDATORY for all issues:
- Milestone: Active milestone or current sprint (required)
- Body: Where found, why deferred, repro steps (if bug), blockers
- Label: optional — if used, must already exist in the repo (
gh label list)
Enforce in skill scripts: require the milestone, but treat the label as optional and validate it against the repo (warn, don't reject) rather than against a fixed allowlist:
if [[ -z "$MILESTONE" ]]; then
echo "Error: Milestone is required for all issues" >&2
exit 1
fi
declare -a GH_ARGS=(issue create --title "$TITLE" --body "$BODY" --milestone "$MILESTONE")
if [[ -n "${LABEL:-}" ]]; then
if gh label list --limit 200 --json name --jq '.[].name' | grep -Fxq "$LABEL"; then
GH_ARGS+=(--label "$LABEL")
else
echo "Warning: label '$LABEL' not found in repo; creating without it." >&2
fi
fi
gh "${GH_ARGS[@]}"
Phase 7: Verify PR (base branch + closing references)
After the PR is created, verify two invariants before allowing the merge.
7a: Base branch matches the milestone
The PR must target the branch resolve-base-branch.sh selected in Phase 4a — not whatever the agent defaulted to. Recompute the expected base and compare it to the PR's actual base:
EXPECTED_BASE="$("{${CLAUDE_CONFIG_DIR}/skills/dev-sprint/scripts/resolve-base-branch.sh" --issue <composite-id>)"
ACTUAL_BASE="$(gh pr view <pr-number> --json baseRefName --jq .baseRefName)"
if [[ "$EXPECTED_BASE" != "$ACTUAL_BASE" ]]; then
echo "BASE MISMATCH: PR targets '$ACTUAL_BASE' but milestone wants '$EXPECTED_BASE'" >&2
# Fix in place — retarget the PR (no reopen needed):
gh pr edit <pr-number> --base "$EXPECTED_BASE"
# Re-verify, and confirm the diff still makes sense against the new base
# (a wrong base can hide or invent changes). If the branch was forked from
# the wrong base, rebase onto $EXPECTED_BASE before retargeting.
fi
Failure condition: PR merges into the wrong base (e.g. main for a 1.0 milestone). This is the exact misfire Phase 4a exists to prevent — 7a is the catch-net if the branch was still created from the wrong place.
7b: Closing references
Verify a closing ref for every sub-issue:
gh pr view <pr-number> --json body --jq .body | grep -c "^Closes #"
Count must equal sub-issue count. If not, fix immediately:
gh pr edit <pr-number> --body "$(gh pr view <pr-number> --json body --jq .body)
Closes #<missing-issue-id>"
Failure condition: PR merges without closing all sub-issues. Sprint not complete until all closed.
Phase 8: Task Cleanup
TaskList. Mark stale tasks deleted via TaskUpdate. Keep composite task if planning new sprint.
References
Plugin/agent guidance: references/plugin-augmentation.md
04 同仓库更多 Skills
clippy
Rust clippy linter skill
hello
Hello world skill
setup-llmenv
Interactive setup wizard for new llmenv users. Evaluates existing Claude Code configuration, helps sort into bundles, and writes llmenv-native equivalents.
llmenv
How to use llmenv's built-in features (task tracker, memory, context-mode, codebase-memory) effectively. Load this first; it points to a reference file per enabled feature rather than dumping all of them into context.
fyi
Use when the user asks for an "fyi", "morning todo", "morning briefing", "what's on my plate", "what do I have open", "what's in flight", "where did I leave off", "what needs my attention today", or "my open work across projects". Surveys the user's currently-open, in-flight, and pending work across GitHub, Linear, Slack, Pylon, and Notion, then produces a ranked, GitHub-markdown briefing for the day. Forward-looking sibling of daily-update (which reports what already happened).
daily-update
Use when the user asks for a "standup", "scrum update", "daily update", "status update", "what did I work on", or "what did I do yesterday". Scans llmenv session logs, Linear, Slack, and GitHub for the user's own activity on a given day and produces a short scrum-style update, copied to the clipboard as Slack markdown.