plugins-claude/session/skills/session-issue/SKILL.md
Browse open issues, pick one, and start work on it
npx skillsauth add st0nefish/claude-toolkit session-issueInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
The discovery door: rank the open issues, pick one, then run the shared
begin-work spine (explore → plan). To start from your own description instead, use
/session:session-start.
CRITICAL: You MUST drive this to a plan. NEVER print "suggested first steps" or ask "ready to start?" — the flow does not end until you have explored the code with research agents and called
EnterPlanMode. And once the approved work is implemented you STOP and hand back to the user (spine Phase 5) — never commit, push, open/merge a PR, or finalize on your own. Plan approval ≠ permission to publish.
Fetch and rank ALL open issues using a subagent. Launch an Agent
(subagent_type: general-purpose) with this prompt:
Fetch open issues, rank them, and return ALL of them (not a top-N subset).
Run this command:
bash ${CLAUDE_PLUGIN_ROOT}/scripts/git-cli issue list --limit 50 --state openFrom the returned JSON array, rank by priority using these criteria:
- Labels indicating urgency:
critical,blocker,high-priority,bugrank higher- Issues with a milestone set rank higher than those without
- More comments → higher priority (community signal)
- Older issues rank higher than newer (age as proxy for neglect)
Return EVERY open issue, highest priority first. Start your reply with a single line
COUNT: <n>giving the total number of open issues. Then, for each issue, emit a two-line block:#N — Title [label1, label2] <one-sentence summary of the issue body, ≤25 words>Derive the summary from each issue's
bodyfield in the JSON (it is already included — do not fetch issues individually). If a body is empty, write(no description).
Pick the issue based on the total COUNT of open issues:
/session:session-start to begin from your own description. Stop here.#N — Title) plus its one-line summary, then
ask the user to confirm before starting (e.g. "This is the only open issue —
work on it now?"). Do not auto-proceed: the user may want to defer it, do it from
a specific machine/node, or start from their own description instead. Only enter
Phase 2 once they confirm.AskUserQuestion (the picker caps at 4 options, so the
whole set fits). Each option label is #N — Title; the description carries the
labels and the one-line summary.AskUserQuestion. Print the
full ranked list as plain text — every issue, each as its #N — Title [labels]
line followed by its one-line summary — then ask the user to type the number of
the issue they want to work on. Wait for their text reply and use that number.
(Forcing dozens of issues through a picker, or pre-truncating to a "top N", buries
real choices behind the agent's guess at what "top" means — let the user scan the
whole list and pick.)Fetch the full issue (save the body and labels — the spine needs them as context):
bash ${CLAUDE_PLUGIN_ROOT}/scripts/git-cli issue show <N>
Determine the branch type from the issue labels:
bug, fix → bugenhancement, feature, improvement → enhancementdocs, chore, refactor, maintenance → chorefeatureBuild the base name as <type>-<slug>, where <slug> is a kebab-case
3-5 word slug from the issue title. Example: issue #42 "Fix login crash on empty
password" → bug-fix-login-crash. The branch name does not encode the issue
number — the linkage lives in the PR's Closes #N, which is what auto-closes
the issue on merge.
Read the shared begin-work spine and execute it (use the Read tool):
${CLAUDE_PLUGIN_ROOT}/reference/spine.md
Hand it the issue body + labels as context and the base branch name from Phase 2.
The spine drives Isolate (worktree) → Escalate-to-orchestrate? → Explore (parallel
research agents) → Plan (EnterPlanMode) → Hand-off. Complete every MANDATORY
phase — the flow ends only once you have presented a plan.
development
Start work from your description — explore the codebase and plan
data-ai
Multi-phase, multi-agent feature workflow: spec → plan → refine → divide → execute → review. Invoke when the user escalates a session-start/session-issue flow to orchestration, or asks to run a non-trivial feature (multiple files, design ambiguity, cross-cutting concerns, correctness-critical paths) through the full multi-agent workflow. For small fixes, prefer session-start.
tools
Interact with GitHub and Gitea issue trackers and CI systems. List and show issues, file bugs, comment on issues or PRs, list and show pull requests, and fetch CI run logs — all from any repo context without leaving the session.
tools
MUST be used for ALL GitHub/Gitea CLI operations. Never invoke `gh`, `tea`, or their subcommands directly — always go through git-cli. Use for issues, PRs, CI runs, repo, and api calls; auto-detects the platform from the git remote.