Menu
Session start orchestrator. Sequences retrospective, sprint setup, work selection, and housekeeping. Use at the start of every session.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Baseado na classificação ocupacional SOC
Review the completed sprint. Compares what shipped vs planned, updates execution_plan.json milestones with delivered status, and records sprint velocity. Use when all stories are done or deferred.
Create sprint.json from an execution plan milestone or product spec. Decomposes milestones into context-rich stories with file domains, interface contracts, and inlined design context. Deep codebase dive identifies story boundaries. Customer confirms scope before writing.
The universal per-story execution-shape decision. Reads the session default tier and the plan-reviewer's per-story recommendation, then applies a behavior table: it may exit in-agent (no spawn — continue in the existing checkout) or spawn exactly ONE teammate at the chosen tier (the story's executor_model forwarded to spawn_teammate's --model flag). Valid for solo and parallel frontiers alike. Runs once per story after that story's plan is reviewed.
Push the sprint branch, fork xp-close-reviewer for cross-cutting review, and merge into the target with cleanup. Auto-invoked at the end of /xp-sprint-review; degrades gracefully when gh is unavailable.
Verify acceptance criteria for in-progress stories. Present criteria, guide e2e test execution, mark stories done or deferred, update sprint.json.
This skill should be used when the user needs to select work for the session, review retro Try items, triage open questions, or pick sprint stories. Invoked as part of /xp-kickoff or standalone via /xp-work-selection.
| name | xp-kickoff |
| description | Session start orchestrator. Sequences retrospective, sprint setup, work selection, and housekeeping. Use at the start of every session. |
| allowed-tools | ["Bash(*/append.sh *)","Bash(*/init.sh)","Bash(*/skills/*/scripts/*)","Bash(python3 */smm/retro_cli.py *)","Bash(python3 */smm/smm_cli.py *)","Bash(python3 */scripts/branching.py *)","Bash(python3 */scripts/cadence_cli.py *)","Bash(python3 */scripts/teammate_config_cli.py *)","Read"] |
!CLAUDE_PLUGIN_DATA="${CLAUDE_PLUGIN_DATA}" ${CLAUDE_SKILL_DIR}/scripts/check_session_needs.sh
Sequential discipline. The harness batches independent tool calls in parallel; this skill is step-gated. Run Step 0 → 1 → 2 → 3 → 4 → 5 → 6 → 7 strictly, one step per turn — make the call, observe, then decide the next. Never put an
AskUserQuestionand the action it gates in one block (the Step 2 session-mode question vs the skill it launches); never spawn the same subagent twice (e.g. duplicate retrospectives). Independent read-only calls may still batch.
You MUST complete ALL steps below in order. Do NOT stop after any single step. Do NOT start working on the user's goal until ALL steps are done. Housekeeping (Step 6) MUST always run — it is not optional. Only begin session work after housekeeping completes.
Bootstrap system context. If the preload shows NEEDS_SYSTEM_CONTEXT, invoke /xp-system-context and wait. It sets the branching stage, so it MUST run before the stage gate below and before the mode fork.
Read the branching stage. Take STAGE from the preload's ### STAGE=N marker — that is the default source, no Python call needed. If no ### STAGE marker is present, read it from Python (covers a fresh project before /xp-system-context and the rare case where the preload could not compute the stage). Also re-read it from Python after any path that mutated the stage since the preload ran: after /xp-system-context or after /xp-stage-migration (below):
STAGE=$(python3 ${CLAUDE_PLUGIN_ROOT}/scripts/branching.py --smm-dir <SMM_DIR> stage)
If STAGE < 2, invoke /xp-stage-migration (it sets the Stage 2 floor directly, unless the user declines). Re-read STAGE after it returns.
Orphan free-branch triage. If the preload shows ORPHAN_FREE_BRANCHES, ask via AskUserQuestion for each: merge / keep / delete?
/xp-free-close against that branch (check it out first if needed).branching.py delete --cwd . --branch <name>. Do not auto-delete branches with commits ahead of primary.Orphan story-branch triage. If the preload shows ORPHAN_STORY_BRANCHES, ask via AskUserQuestion for each: merge / keep / delete?
branching.py merge-branch --cwd . --branch <name>, then branching.py delete --cwd . --branch <name>.branching.py delete --cwd . --branch <name>. Do not auto-delete branches with commits ahead of base.Invoke the xp-retrospective agent via the Agent tool (subagent_type=xp-agents:xp-retrospective). Invoke it synchronously — pass run_in_background:false (the harness backgrounds Agent-tool subagents by default; a backgrounded retrospective races the kickoff sequence and can bury its render before Step 2). The agent handles both populated and empty .retro-input.json — on a fresh project with no prior session_end, it emits a seed retrospective (Keep around adopting XP, Try as skill suggestions, no Fix). Do NOT gate this step on the existence of .retro-input.json; the agent owns that branch.
After it completes, render the latest retrospective:
python3 ${CLAUDE_PLUGIN_ROOT}/smm/retro_cli.py --smm-dir <SMM_DIR> render
Output the render CLI's stdout to the user before continuing.
Unconditional gate — do not skip. Ask the user via AskUserQuestion: "Free session or Sprint session?"
Record the session-mode goal event:
${CLAUDE_PLUGIN_ROOT}/smm/append.sh --smm-dir <SMM_DIR> \
--type "goal" --agent "xp-kickoff" \
--content "Free session" # or "Sprint session: <sprint-id>"
Ask via AskUserQuestion: "What's the focus for this session?" Use the answer for both the branch slug and the session-focus goal event:
${CLAUDE_PLUGIN_ROOT}/smm/append.sh --smm-dir <SMM_DIR> \
--type "goal" --agent "xp-kickoff" \
--content "<user's session focus>"
If STAGE >= 2, create the free branch (slug = first 3-4 slugified words of the focus):
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/branching.py --smm-dir <SMM_DIR> \
create-free --cwd . --slug "<session-focus-slug>"
Tell the user which branch is in use. Then jump to Step 5.
Review cadence (opt-in). Ask the user via AskUserQuestion: "Review cadence: per-commit or per-story?"
/xp-quality-review runs before every commit (its xp-code-reviewer self-finds correctness)./xp-story-close (the merge); intermediate commits proceed with a one-line deferral advisory. Nothing reaches the sprint base unreviewed.Only when the user chooses story, write the cadence marker (the careful commit default needs no write — session-start already reset it):
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/cadence_cli.py --smm-dir <SMM_DIR> write story
Teammate support (two-step opt-in). Ask the user via AskUserQuestion: "Teammate support for this session?"
off and skip Q2:python3 ${CLAUDE_PLUGIN_ROOT}/scripts/teammate_config_cli.py --smm-dir <SMM_DIR> write off
Q2 — Ask via AskUserQuestion: "Default model for teammates?"
Four options; fable rides the auto "Other" escape:
On "Other", accept a valid token (e.g. fable), else re-ask.
Write the chosen token:
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/teammate_config_cli.py --smm-dir <SMM_DIR> write <token>
Then proceed to Step 3.
If the preload shows NEEDS_EXECUTION_PLAN, first Read the SMM file at <SMM_DIR>/shared_mental_model.json for context. Then invoke /xp-plan. Wait for it to complete.
If not shown, skip to Step 4.
If the preload shows NEEDS_SPRINT and you haven't read the SMM in Step 3, Read <SMM_DIR>/shared_mental_model.json first. Then invoke /xp-sprint-start. Wait for it to complete (it creates the sprint branch).
If not shown, skip to Step 5.
Run /xp-work-selection. In free mode, stop after Step 4 (skip sprint story selection).
Wait for it to complete.
Invoke the xp-housekeeper agent via the Agent tool (subagent_type=xp-agents:xp-housekeeper). This is mandatory — it curates the four-pillar SMM (Intent, Constraints, Risks, Wisdom). Kickoff is not complete until housekeeping finishes.
Do NOT run housekeeping in the background — pass run_in_background:false to the Agent tool. The harness backgrounds Agent-tool subagents by default; without the explicit flag the Stop housekeeping gate races. Wait for the subagent to complete before proceeding to Step 7.
If the user said "skip" at any earlier step, still run housekeeping.
After housekeeping completes, render:
python3 ${CLAUDE_PLUGIN_ROOT}/smm/smm_cli.py --smm-dir <SMM_DIR> render
Output the render CLI's stdout to the user, followed by the change summary the housekeeper agent returned (items added/removed/promoted/resolved, health warnings).
Kickoff is complete. Do NOT stop.
If stories were selected in Step 5, run /xp-schedule first — it promotes the next frontier (scheduled → in-progress), sets each story's execution_mode, and (solo) creates the branch. The schedule gate enforces this before planning. Then follow its handoff: enter plan mode, run /xp-review-plan after exiting, then (teammate mode only) run /xp-assign. If a story is already in-progress from a prior session, continue it directly into the plan cycle.
If no sprint is active or the user chose free session, begin working on the session goals.