hotsheet
Read the Hot Sheet worklist and work through the current priority items
SKILL.md.claude/skills/hotsheet/SKILL.md
GitHub<!-- hotsheet-skill-version: 22 -->
Read .hotsheet/worklist.md and work through the tickets in priority order.
For each ticket:
- Read the ticket details carefully
- Implement the work described
- When complete, mark it done via the Hot Sheet UI
Work through them in order of priority, where reasonable.
If the worklist says "Auto-Prioritize", follow those instructions to choose and mark tickets as Up Next before working on them.
If API calls fail (connection refused or 403), re-read .hotsheet/settings.json for the current port and secret values — you may be connecting to the wrong Hot Sheet instance.
MCP tools (hotsheet_*) are preferred over curl when the channel is connected — see the worklist for per-operation guidance. The 14-tool surface covers ticket lifecycle (hotsheet_update_ticket, hotsheet_create_ticket, hotsheet_get_ticket, hotsheet_delete_ticket, hotsheet_restore_ticket, hotsheet_toggle_up_next, hotsheet_duplicate_tickets), bulk operations (hotsheet_batch), notes (hotsheet_edit_note, hotsheet_delete_note), attachments (hotsheet_add_attachment), channel signaling (hotsheet_signal_done), feedback sugar (hotsheet_request_feedback), and query (hotsheet_query_tickets). Curl stays supported as the universal fallback for non-Claude AI agents and human terminal callers.
Parallelizing work across the worker pool: hotsheet_get_worker_pool / hotsheet_set_worker_target / hotsheet_dispatch_tickets / hotsheet_drain_workers. When the project has "always preview agent plans" on (the worklist says so), call hotsheet_propose_partition with your whole proposed assignment INSTEAD of hotsheet_dispatch_tickets — it surfaces the plan in the owner's partition editor for review, and the UI dispatches on accept (you do not claim the tickets yourself).
Git: keep the target current + integrate worker branches
You run on the target branch (usually main) in the main worktree, so you are the single integrator for parallel worktree workers (docs/89). Distributed workers (/hotsheet-worker) commit their work on their own branches and rebase onto the target to stay current, but they never write the target — that's your job:
- Stay current — before integrating, bring the target up to date:
git fetchthengit pull --rebase(or rebase onto the upstream) when the repo has a remote, so you build on the latest. Commit or stash your own in-progress changes first so a merge doesn't tangle with them. - Integrate ready worker branches — periodically (e.g. when a batch of workers has finished, or the pool drains). Use the integration helpers (HS-9048) rather than hand-rolling the git:
GET /api/workers/integratablereturns the detected target branch + the ready worker branches (hotsheet/*ahead of the target, with ahead/behind counts); then for each, in ticket-priority order,POST /api/workers/integratewith{ "branch": "<name>" }does a guarded merge into the target. It returns astatus:merged(success),conflict(it captured the conflicted files + aborted cleanly — resolve them by hand or, if non-trivial, ask the maintainer),dirty-tree(commit/stash your own changes first),not-on-target/nothing-to-integrate. The helper never pushes — pushing still needs explicit permission. - A ready branch may carry a BATCH of several tickets. Workers batch small/related tickets onto one branch and refresh + gate once at the batch boundary (docs/98), so the "branch ready" signal fires once per batch, not per ticket — one
integrateBranchcall + one gate run covers the whole batch. When you integrate it, clearpending_integrationfor every ticket whose work it carried (next bullet), not just one. - Gates after a merge (and the optional in-helper gate, HS-9091). After a
mergedresult with nogatefield, run the project's gates yourself (type-check, lint, the relevant tests). If the project configured the opt-inintegrationGateshared setting, the helper ran that command inside the merge and you'll get either:mergedwithresult.gate.ran: true(the gate already passed — you do not need to re-run those gates for this branch);gate-failed(the gate command failed and the merge was already rolled back to the pre-merge target — do NOT re-merge blindly; readresult.gate.outputto decide whether to quickly fix-and-retry or, if non-trivial, ask the maintainer); orgate-timeout(same rolled-back state, but the gate exceeded its time limit). Ongate-failed/gate-timeoutthe branch is unmerged — leave its ticket'spending_integrationmarker set (don't clear it). - For each ticket whose work you just integrated, clear its "merge pending" marker:
hotsheet_update_ticketwith{ "id": <id>, "pending_integration": false }(the tickets markedpending_integrationare the ones awaiting integration). - Sensible conflict resolution, ask on the hard ones — auto-resolve trivial/mechanical conflicts; if a conflict is non-trivial or ambiguous, or the gates fail in a way you can't quickly and safely fix, stop and ask the maintainer rather than force it (leave the branch unmerged). Integrate only from committed branch state — never disturb a worker mid-ticket.
- NEVER
git pushwithout the maintainer's explicit permission — local integration only.