init
Step 1 of the five-step flow. Run before any change. Inventory existing tools, read the original design intent, recall memory, claim the issue, and create an isolated worktree.
SKILL.md.claude/skills/init/SKILL.md
GitHubinit: question the requirement, inventory, isolate
Run this before the first edit.
1. Inventory before you build
Follow the Key List Protocol:
- Check the Key List: Read the repository's SSOT index (
docs/ssot/MANIFEST.yaml,common/meta/data/MANIFEST.yaml, or the App's contracts index).- Scan the capability keys (e.g.,
deploy_v2,infra2-sdk,vault,signoz,pr_merge_gate).
- Scan the capability keys (e.g.,
- Reuse if present: If an existing component covers the need, reuse it directly. Do not build private alternatives.
- Append when new: If no entry fits and a new capability is necessary:
- Build the minimal implementation.
- Register the new key and summary into
MANIFEST.yamlin the same PR.
- First-turn contract: State in your opening response:
"Reusing existing component: <key>"OR"New capability: will register <key> in MANIFEST.yaml".
2. Read the original intent before you change a mechanism
- Find the issue, PR, or SSOT page that created the mechanism. Read it first.
- Write one sentence: "This exists because ...". If you cannot, you are not ready to change it.
- A fix that "does not take effect" needs the deep cause. Check that the host really loads the file. Do not retry the same edit.
3. Recall
Use the recall skill for the four-level search. Read hits as claims, not facts: check current code.
4. Scan issue and claim the lock
- Read the latest issue discussion (
gh issue view <N> --comments). - Search for an existing issue before you create one. Update it when it exists.
- Another worktree or running process for the same issue is an active lock. Do not take over. Match the full token:
git worktree list | grep -F "_issue<N>_"
A substring match is wrong: issue12 would match issue123.
Scan open production decisions: in a checkout of wangzitian0/infra2, run
uv run python -m tools.harness status --fetch --online. A release tag that the production marker has not reached waits for a production decision. Aholdin the handover notes (recallskill) answers it.Scan stale leftovers: Run these in bash or zsh, in the repository you work in. The three counts are: branches whose upstream is gone, stashes older than 30 days, and local branches whose name is the head of a merged PR. Put each count above zero in the "facts you recalled" line. A count above zero shows a cleanup that nothing triggered. Do not delete what another session owns. A
gherror is a failed scan, not a zero.
git fetch --prune --no-tags --no-write-fetch-head --no-auto-gc origin || echo "fetch failed: the counts can be stale"
LC_ALL=C git for-each-ref --format='%(upstream:track)' refs/heads | awk '/\[gone\]/ {n++} END {print n+0}'
git stash list --format=%ct | awk -v c=$(( $(date +%s) - 2592000 )) '$1 < c {n++} END {print n+0}'
gh pr list --state merged --limit 300 --json headRefName --jq '.[].headRefName' | sort -u | comm -12 - <(git for-each-ref --format='%(refname:short)' refs/heads | sort) | awk 'END {print NR}'
5. Init Swarm (M=4)
For complex tasks or large repositories, dispatch 1 Init Swarm (M=4 Interns via subagent_batch):
- Intern 1 (Key List): Scans
MANIFEST.yamlandtools/for reusable components. - Intern 2 (Intent & History): Runs git log search for past PRs touching the target subsystem.
- Intern 3 (Memory Recall): Executes global
search_notesacross all projects. - Intern 4 (Worktree Locks): Checks active worktrees and process locks across repositories.
6. Create the worktree
- Name:
<repo>_issue<N>_<slug>. Branch:feat/issue<N>-<slug>. Base: the latestorigin/main. - Never edit in a shared checkout (
git addshares.git/index), and never auto-remove IDE worktrees under~/.gemini/antigravity/worktrees/. - The worktree must work alone. Tools and tests must not use
../..to find files outside it.
7. Report the baseline
Reply in the owner's language with four lines: the goal, the tools you reuse, the facts you recalled, the worktree path. Do not execute test suites during baseline reporting. The baseline report is task metadata only.