Start-Task

Turn a task reference into a synced feature branch named {type}/{number}/{slug} plus a context file at specs/context/<id>.md. Zero confirmations on the happy path: detect → fetch → derive → execute → report.

1. Detect the source (no bash)

Match the args in this order:

1. Jira — [A-Z]+-\d+ (e.g. PROJ-42)
2. GitHub issue — a numeric token, stripping any of github, gh, issue, issues, #
3. Local spec — a path ending in .md, .txt, .yaml, .yml, or .json; reject any path containing ..
4. Ad-hoc — nothing matched: ask once (a single question) for the task type, title, and identifier, then skip to step 3

2. Fetch (1 bash call)

GitHub:

gh issue view <ISSUE_NUM> --json title,labels --jq '{title: .title, labels: [.labels[].name]}'

Jira: which acli && acli jira issue view <KEY> — extract summary, type, key. If acli is missing, hard-stop: the developer chose Jira; do not degrade to ad-hoc.

Local spec: read the file; extract title and scope. The identifier is the developer-provided one or the filename stem.

On any fetch failure, hard-stop and print the error verbatim — no retries, no paraphrasing, no follow-up questions.

3. Derive TYPE and SLUG (no bash)

TYPE — from GitHub labels (first match wins) or from the task's nature:

Signal	TYPE
bug, fix, hotfix	fix
enhancement, feature	feat
refactor	refactor
docs, documentation	docs
test, tests	test
ci, github-actions	ci
chore, deps, dependencies	chore
(no match)	feat

SLUG — from the title: lower-case; drop articles, prepositions, and helper verbs; keep the core action verb + object nouns; 2–4 words; hyphenated; only [a-z0-9-].

Examples: "Add user authentication to the login form" → add-user-auth; "Fix null pointer in payment processing flow" → fix-null-pointer-payment.

4. Execute

GitHub (1 bash call) — the script syncs the default branch, creates the branch, writes and commits the context file on it, then pushes:

bash {base_directory}/gh-start-task.sh <ISSUE_NUM> <TYPE> <SLUG>

Non-zero exit → hard-stop, print stderr verbatim. It fails on a dirty working tree by design — never stash or discard for the developer.

Jira / local / ad-hoc — do the same steps manually:

git status --porcelain            # any output → hard-stop: uncommitted changes
git fetch origin <default> && git checkout <default> && git pull origin <default>
git checkout -b <TYPE>/<ID>/<SLUG> && git push -u origin <TYPE>/<ID>/<SLUG>

If the branch already exists locally, switch to it instead of creating it. Then write specs/context/<ID>.md with: identifier, title, type, source, relevant details (description, acceptance criteria), and today's date. If the file exists, update it in place — don't overwrite wholesale.

5. Report (no bash)

Branch: <TYPE>/<ID>/<SLUG>
Task:   <ID> — <TITLE>
Context: specs/context/<ID>.md

Next:
  /plan-requirements   — capture WHAT and WHY
  /plan-architecture   — design the solution
  /implement           — start coding

You Must NOT

• Ask for confirmation on the happy path — ad-hoc's single intake question is the only interactive moment
• Stash, discard, or commit the developer's pre-existing uncommitted changes — hard-stop instead (the GitHub script committing the context file it just generated is the one sanctioned exception)
• Push to main/master, or use --force on any push
• Paraphrase errors — print them verbatim so the developer gets the exact actionable message
• Fall back to ad-hoc when a chosen CLI tool is missing — fail clearly
• Trigger automatically — this skill is opt-in