菜单
Orchestrates a complete feature lifecycle: receive goals, synthesize specs, propose via OpenSpec, commit specs to PR, apply tasks, commit implementation, archive, update PR, and post audit results.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/avoidwork/skills --skill create-feature命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | create-feature |
| description | Orchestrates a complete feature lifecycle: receive goals, synthesize specs, propose via OpenSpec, commit specs to PR, apply tasks, commit implementation, archive, update PR, and post audit results. |
⚠️ EXECUTION RULE: Every code block in this skill is a shell command to execute. Do not print them as text, explain them, or treat them as examples — run them directly.
Orchestrate a complete feature lifecycle from raw goals to shipped code. This is the full pipeline — synthesis, specification, implementation, verification, delivery, and archive.
Before starting, ensure:
package.json, openspec/)gh CLI is authenticated and configuredopenspec CLI is availablemain (or will be checked out at step 0)The input may include a BRANCH_TYPE prefix (e.g., BRANCH_TYPE=fix create-feature ...). Extract it if present in the conversation context.
Read BRANCH_TYPE from the conversation history if provided by the invoking skill. If not present, default to feat.
The user provides a list of goals/features in any format (natural language, JSON, markdown list, etc.). Parse and normalize them into a clean array of goal strings.
Example input:
Add a new tool that can summarize web pages, and improve the TUI memory panel to show retention stats.
Normalized goals:
Determine the target repository dynamically from the git remote (needed for later gh commands):
GIT_REMOTE=$(git remote get-url origin 2>/dev/null)
if echo "$GIT_REMOTE" | grep -q '^git@'; then
GH_REPO=$(echo "$GIT_REMOTE" | sed 's/.*@[^:]*:\(.*\).git$/\1/')
elif echo "$GIT_REMOTE" | grep -q 'github\.com'; then
GH_REPO=$(echo "$GIT_REMOTE" | sed 's/.*github\.com[/:]\(.*\).git$/\1/')
fi
# Only checkout main if not already on it
if [ "$(git branch --show-current)" != "main" ]; then
# Verify main exists on remote before pulling
REMOTE_MAIN=$(git ls-remote --heads origin main 2>/dev/null)
if [ -n "$REMOTE_MAIN" ]; then
git fetch origin main
git checkout main
git pull origin main
else
echo "WARNING: No remote branch 'main' found. Current branch '$(git branch --show-current)' will be used as-is."
fi
fi
Verify the working tree is clean:
git status --porcelain
If there are uncommitted changes, report them and stop. Do not proceed with a dirty tree.
Capture a unique session identifier so that multiple instances (e.g., subagents) can run in parallel without file collisions.
# Portable: works in Alpine, minimal images, and standard Linux
# Generates 8 random hex chars from /dev/urandom, falls back to $RANDOM if unavailable
SESSION_ID=$(head -c 8 /dev/urandom 2>/dev/null | od -An -tx1 | tr -d ' \n' || echo $RANDOM$RANDOM$RANDOM)
echo "SESSION_ID=$SESSION_ID"
Where to store temp files: The agent running this skill decides where to place temp files. Use example names like ${SESSION_ID}-feature-goals.md, ${SESSION_ID}-feature-prompt.md, ${SESSION_ID}-audit-results.md, and ${SESSION_ID}-pr-number.txt. The agent may place them in tmp/, state/, or any other directory that fits the project's conventions.
Set up guaranteed cleanup: Register a trap so temp files are removed regardless of how the skill exits (success, failure, interruption):
trap 'rm -f "${SESSION_ID}-feature-goals.md" "${SESSION_ID}-feature-prompt.md" "${SESSION_ID}-audit-results.md" "${SESSION_ID}-pr-number.txt" "${SESSION_ID}-test-patterns.md"' EXIT
Take the raw goals from the input and expand each into a more detailed, actionable specification. For each goal, produce:
Write the detailed goals to a file named ${SESSION_ID}-feature-goals.md (use an example name — the agent decides where to place it). This file serves as the source of truth for all subsequent audits.
Using the detailed goals from Step 1, synthesize a multi-paragraph proposal prompt suitable for feeding into openspec-propose. The prompt should:
CHANGE_NAME: <kebab-case-name> (e.g., CHANGE_NAME: web-page-summarizer-tui-memory-stats)openspec new change + artifact generation will produce useful specsWrite the proposal prompt to a file named ${SESSION_ID}-feature-prompt.md (use an example name — the agent decides where to place it).
Parse the change name from the proposal prompt in Step 2. This is the kebab-case identifier that will be used for the OpenSpec change directory.
CHANGE_NAME=$(grep -o 'CHANGE_NAME: .*' "${SESSION_ID}-feature-prompt.md" | sed 's/^CHANGE_NAME: //' | sed 's/ *$//' | head -1)
if [ -z "$CHANGE_NAME" ]; then
CHANGE_NAME=$(basename "$(pwd)" | tr '[:upper:]' '[:lower:]' | tr ' ' '-')
fi
echo "CHANGE_NAME=$CHANGE_NAME"
Determine the branch type from the input context. If the input contains a reference to an issue with a known type (e.g., from fix-issue chain), extract it. Otherwise, default to feat.
Valid types: feat, fix, chore, docs, test.
Strip any existing type prefix from $CHANGE_NAME to avoid double-prefixing:
# Determine branch type — default to feat if not specified
BRANCH_TYPE="${BRANCH_TYPE:-feat}"
# Validate branch type
case "$BRANCH_TYPE" in
feat|fix|chore|docs|test) ;;
*) BRANCH_TYPE="feat" ;;
esac
# Strip any existing type prefix from CHANGE_NAME to avoid double-prefixing
BRANCH_NAME=$(echo "$CHANGE_NAME" | sed 's/^\(feat\|fix\|chore\|docs\|test\)\///')
git checkout -b "${BRANCH_TYPE}/${BRANCH_NAME}"
Verify the branch was created:
git branch --show-current
Run the OpenSpec propose workflow for the synthesized change:
openspec new change "$CHANGE_NAME"
Then generate all artifacts using this procedure (optimized for the automated pipeline):
Get the artifact build order:
openspec status --change "$CHANGE_NAME" --json
For each artifact that is ready (dependencies satisfied):
openspec instructions <artifact-id> --change "$CHANGE_NAME" --jsonopenspec status --change "$CHANGE_NAME" --json after each artifactContinue until all applyRequires artifacts are complete.
After completion, verify all expected files exist:
ls -la openspec/changes/$CHANGE_NAME/
Expected files: .openspec.yaml, proposal.md, design.md, tasks.md, and any spec deltas in specs/.
If any expected file is missing, stop and report the error.
Read the detailed goals from the file named ${SESSION_ID}-feature-goals.md (use an example name — the agent decides where to place it) and the generated spec documents (proposal.md, design.md, tasks.md, and any spec deltas).
Perform a thorough audit:
Write audit findings to a file named ${SESSION_ID}-audit-results.md (use an example name — the agent decides where to place it).
If errors are found:
If no errors are found:
Purpose: Lock the spec documents into a pull request before implementation begins.
Do not perform git operations inline. Invoke the commit-push skill to handle staging, committing, pushing, and PR creation.
The commit-push skill will:
PR_NUMBER=<number> and PR_URL=<url>Read the PR_NUMBER from the structured output printed by commit-push. Do not attempt to grep a shell variable — the value is in the conversation history.
Add a reusable capture pattern to extract PR_NUMBER from the conversation history after commit-push completes:
# Capture PR_NUMBER from the conversation history (last occurrence wins)
# The chained skill prints lines like: PR_NUMBER=42 or PR_URL=https://...
# Use grep to find lines matching the pattern, take the last one
PR_NUMBER=$(echo "$CONVERSATION_HISTORY" | grep -oE '^PR_NUMBER=[0-9]+' | tail -1 | cut -d= -f2)
if [ -z "$PR_NUMBER" ]; then
echo "ERROR: Could not capture PR_NUMBER from commit-push output. Conversation history:"
echo "$CONVERSATION_HISTORY" | tail -20
exit 1
fi
echo "PR_NUMBER=$PR_NUMBER"
Write it to a file for later use in Step 12:
echo "<PR_NUMBER_VALUE>" > "${SESSION_ID}-pr-number.txt"
Replace <PR_NUMBER_VALUE> with the actual number read from the conversation history.
If commit-push fails, report the error and stop. Do not attempt to recover with manual git commands.
After reading the PR number, continue to Step 7. Do not stop or wait for further input — the pipeline proceeds automatically.
Before implementing, read existing tests to understand the project's testing conventions, patterns, and structure. This ensures new tests (if any are required by the tasks) follow the established style.
Identify source files from the spec: Look at the tasks in tasks.md, the design document, and spec deltas to determine which source files will be created or modified.
Find related tests using these rules:
src/foo/bar.js, look for tests/unit/foo/bar.test.js or tests/unit/bar.test.test.js. Mirror the directory structure.tests/ or tests/unit/ directory with no nesting, find tests by filename matching — e.g., src/foo/bar.js → tests/bar.test.js or tests/bar.test.test.js.src/foo/bar.js with src/foo/bar.test.js or src/foo/bar.spec.js.tests/unit/foo.test.js covering everything in src/foo/.For each source file being created or modified:
# Example: find related tests for src/scheduler/runner.js
# 1. Check nested test directory
ls tests/unit/scheduler/runner.test.js 2>/dev/null && echo "found: tests/unit/scheduler/runner.test.js"
# 2. Check flat test directory by name
ls tests/runner.test.js 2>/dev/null && echo "found: tests/runner.test.js"
# 3. Check sibling test file
ls src/scheduler/runner.test.js 2>/dev/null && echo "found: src/scheduler/runner.test.js"
# 4. Check parent module test
ls tests/unit/scheduler.test.js 2>/dev/null && echo "found: tests/unit/scheduler.test.js"
Read every test file you discover. Pay attention to:
assert, chai, node --test, custom helpersdescribe/it nesting, fixture setup, teardownWrite a brief summary of the patterns you found to ${SESSION_ID}-test-patterns.md (use an example name — the agent decides where to place it). This becomes context for Step 7 when implementing tests.
If no related tests exist, note "No existing tests found — follow project conventions from AGENTS.md §3.5" and proceed.
Invoke the openspec-apply-change skill to implement all tasks from tasks.md.
The openspec-apply-change skill will:
tasks.md[x] in tasks.mdWait for the skill to complete. Do not proceed to Step 7.5 until all tasks are marked complete.
If the skill reports that tasks are blocked (missing artifacts), report the error and stop. Do not proceed with incomplete work.
If interrupted (timeout, context limit): Re-invoke openspec-apply-change "$CHANGE_NAME" to resume from where it left off. The skill checks task status and only implements remaining tasks.
After completion, verify:
tasks.md are marked [x]After verification, continue to Step 7.5. Do not stop or wait for further input — the pipeline proceeds automatically.
Check which npm scripts are available before running them — not all projects define the same scripts:
AVAILABLE_SCRIPTS=$(node -e "console.log(Object.keys(require('./package.json').scripts || {}).join('\n'))" 2>/dev/null)
test script exists, run: npm run testlint script exists, run: npm run lintcoverage script exists, run: npm run coverageIf any verification fails, fix the issues and re-verify.
Purpose: Push the code produced by openspec-apply-change to the open PR.
Invoke the commit-push skill to stage, commit, push, and update the existing PR.
This will:
commit-push detects a pre-existing PRPR_NUMBER=<number> and PR_URL=<url>Read the PR_NUMBER from the structured output. If commit-push fails, report the error and stop. Do not attempt to recover with manual git commands.
Add a reusable capture pattern to extract PR_NUMBER from the conversation history after commit-push completes:
# Capture PR_NUMBER from the conversation history (last occurrence wins)
PR_NUMBER=$(echo "$CONVERSATION_HISTORY" | grep -oE '^PR_NUMBER=[0-9]+' | tail -1 | cut -d= -f2)
if [ -z "$PR_NUMBER" ]; then
echo "ERROR: Could not capture PR_NUMBER from commit-push output."
exit 1
fi
echo "PR_NUMBER=$PR_NUMBER"
After /commit-push completes, continue to Step 8. Do not stop or wait for further input — the pipeline proceeds automatically.
npm start)After tasks are applied, verify the application actually starts without crashing.
Run with a timeout — this is the preferred method:
timeout 10 npm start 2>&1 || true
This ensures the process terminates after 10 seconds even if it hangs.
If the application fails to start, fix the issue before proceeding. Do not skip this step — a crashing application means the implementation is flawed.
Read the original goals from the file named ${SESSION_ID}-feature-goals.md (use an example name — the agent decides where to place it) and the spec documents. Audit the implemented results:
Write audit findings to the file named ${SESSION_ID}-audit-results.md (use an example name — the agent decides where to place it) (overwrite previous results).
If errors are found:
If no errors are found:
Purpose: Archive the OpenSpec change and push the archive (and any remaining implementation fixes from Step 9) to the open PR.
Archive the OpenSpec change with automatic sync:
Archive with retry on validation failure (up to 3 attempts):
ARCHIVE_ATTEMPTS=0
ARCHIVE_MAX=3
while [ $ARCHIVE_ATTEMPTS -lt $ARCHIVE_MAX ]; do
ARCHIVE_ATTEMPTS=$((ARCHIVE_ATTEMPTS + 1))
echo "Archive attempt $ARCHIVE_ATTEMPTS/$ARCHIVE_MAX..."
ARCHIVE_OUTPUT=$(openspec archive "$CHANGE_NAME" --yes 2>&1)
ARCHIVE_EXIT=$?
if [ $ARCHIVE_EXIT -eq 0 ]; then
echo "$ARCHIVE_OUTPUT" > "${SESSION_ID}-archive-output.md"
echo "Archive succeeded on attempt $ARCHIVE_ATTEMPTS."
break
fi
echo "Archive failed (attempt $ARCHIVE_ATTEMPTS): $ARCHIVE_OUTPUT"
echo "$ARCHIVE_OUTPUT" > "${SESSION_ID}-archive-error.md"
if [ $ARCHIVE_ATTEMPTS -ge $ARCHIVE_MAX ]; then
echo "ERROR: Archive failed after $ARCHIVE_MAX attempts. Stopping."
exit 1
fi
done
When presented with a sync prompt, automatically choose sync (the recommended option). Do not require user input — just select sync:
Resolve spec issues if archive validation failed (up to 3 fix iterations):
If the archive step above exited with an error (i.e., openspec archive returned non-zero and the loop terminated via exit 1), the file ${SESSION_ID}-archive-error.md contains the error details. Use those details to identify and fix the underlying spec issues:
${SESSION_ID}-archive-error.md to understand what validation failed (missing requirements, inconsistent specs, task-to-spec mismatches, etc.)specs/)If archive still fails after 3 fix iterations, report the remaining validation errors and stop. Do not proceed with an unarchived change.
Verify the archive was created:
ls openspec/changes/archive/
The change directory should now be under openspec/changes/archive/YYYY-MM-DD-<name>/.
Push the archive (and all remaining changes) to the open PR by invoking commit-push.
This will:
commit-push detects a pre-existing PRPR_NUMBER=<number> and PR_URL=<url>Read the PR_NUMBER from the structured output. If commit-push fails, report the error and stop. Do not attempt to recover with manual git commands.
Add a reusable capture pattern to extract PR_NUMBER from the conversation history after commit-push completes:
# Capture PR_NUMBER from the conversation history (last occurrence wins)
PR_NUMBER=$(echo "$CONVERSATION_HISTORY" | grep -oE '^PR_NUMBER=[0-9]+' | tail -1 | cut -d= -f2)
if [ -z "$PR_NUMBER" ]; then
echo "ERROR: Could not capture PR_NUMBER from commit-push output."
exit 1
fi
echo "PR_NUMBER=$PR_NUMBER"
After the archive and push complete, continue to Step 11. Do not stop or wait for further input — the pipeline proceeds automatically.
Purpose: Set the final, accurate PR title and description reflecting what was actually implemented.
Invoke the update-pr skill to update the PR. This skill scans project rules for conventions and the PR template, synthesizes both from the delta between the branch and target, and applies changes via gh api.
The update-pr skill will:
git diff main...HEAD --stat, git diff main...HEAD --name-only)gh apiAfter update-pr completes, verify the PR was updated correctly by checking the PR on GitHub.
Add a reusable capture pattern to confirm PR_NUMBER is still available from the conversation history after update-pr completes:
# PR_NUMBER was already captured from earlier commit-push steps.
# Re-capture to confirm it's still available after update-pr.
PR_NUMBER=$(echo "$CONVERSATION_HISTORY" | grep -oE '^PR_NUMBER=[0-9]+' | tail -1 | cut -d= -f2)
if [ -z "$PR_NUMBER" ]; then
echo "ERROR: PR_NUMBER not found in conversation history after update-pr."
exit 1
fi
echo "PR_NUMBER=$PR_NUMBER (verified after update-pr)"
After verification, continue to Step 12. Do not stop or wait for further input — the pipeline proceeds automatically.
Post the final audit results from Step 9 as a comment on the PR:
PR_NUMBER=$(cat "${SESSION_ID}-pr-number.txt")
gh pr comment "$PR_NUMBER" --body "$(cat "${SESSION_ID}-audit-results.md")" --repo "$GH_REPO"
Print a summary:
Feature lifecycle complete.
Change: $CHANGE_NAME
PR: <URL>
Goals addressed: <count>/<total>
Iterations: <spec-audit-x>, <impl-audit-y>
Tests: passing
Lint: passing
Coverage: maintained
Remove the intermediate memory files — they served their purpose and won't be needed again:
rm -f "${SESSION_ID}-feature-goals.md" "${SESSION_ID}-feature-prompt.md" "${SESSION_ID}-audit-results.md" "${SESSION_ID}-pr-number.txt" "${SESSION_ID}-test-patterns.md"
Verify cleanup:
ls "${SESSION_ID}-feature-*.md" "${SESSION_ID}-audit-results.md" "${SESSION_ID}-pr-number.txt" 2>&1
If any files remain, report them and remove manually.
gh API calls fail, report the error and provide the manual commands needed.SESSION_ID to avoid temp file collisions.commit-push is delegated, not inline. Steps 6, 7.5, and 10 invoke commit-push as a chain instruction — do not perform git operations inline at those steps.openspec-apply-change is resumable. If interrupted by timeout or context limit, re-invoke with the same CHANGE_NAME to resume from where it left off.
left off.