| name | start-session |
| description | Use when the user says 'start new session', 'bắt đầu session mới', '/start-session', 'what's the state?', 'tình trạng hiện tại', or when context must be reloaded after /clear. Loads: branch, open PRs, failing CI, recent commits, plus optional gap/wave/cloud sections when those artifacts exist. Output: single summary block. Also checks session-lock conflicts when running parallel agents. |
| user-invocable | true |
| argument-hint | [--quick] [--no-lock] [--cloud|--refresh-cloud] |
/start-session — Session Orchestration
Prepare a fresh session by loading the right context + detecting conflicts with concurrent sessions. Works in ANY project; project-specific sections (gap pipeline, wave history, cloud stack) degrade gracefully when absent.
When to use
- Starting a new conversation in this repo
- After
/clear or /compact
- Before critical work (wave merge, audit, release)
- When unsure "what next?" → run this instead of reading blindly
Process
Step 1 — Run the state collector
./.claude/skills/workflow/start-session/scripts/collect-state.sh
Output: active branch, open PR count + per-PR CI/merge state, repo health level, current wave (if any), top blocker gaps (if a gap pipeline exists), MCP server status, recent commits. Cloud stack is opt-in via --cloud.
Step 2 — Read digest sources (context-aware — minimize auto-load)
Default to script/CSV queries over narrative file reads. Reading files under documents/** may trigger path-scoped rule auto-loads and inflate context (per context-budget-mandate.md).
Default flow (skip file reads):
- Branch / PRs / CI / blockers — already in
collect-state.sh output.
- Gap detail (if project uses a gap pipeline) — query
gap-status.csv instead of reading each gap file.
- Recent waves (if project uses wave-pack methodology) —
tail -3 .claude/skills/quality/wave-pack-planner/data/wave-history.jsonl.
Only Read files when the user explicitly asks OR the script output is insufficient for the task at hand. CLAUDE.md auto-loads every session — no need to re-read.
Step 3 — Session-lock check (optional, --no-lock skips)
ls .claude/session-locks/ — list active locks.- If a lock matches a branch/gap the user intends to touch → WARN: "Session X is editing branch Y / GAP-Z, continuing risks conflict."
- User confirms → create own lock:
session-{timestamp}-{hostname}.lock.
- Lock content: YAML with
branch, gaps, started, intent.
Schema + overlap rules: reference/context-template.md §Lock File Schema.
Step 4 — Output summary block
Format per reference/context-template.md. Data sources (from collect-state.sh):
| Field | Source |
|---|
| Branch state | git diff --name-only |
| Open PRs | gh pr list (per-PR CI + mergeStateStatus) |
| Repo level | scripts/repo-status.sh --json .level if present, else CI-only |
| CVE / stale branches | same --json → .security, .branches (if repo-status.sh present) |
| Recent commits | git log <default-branch> --since='3 days ago' --oneline |
| MCP servers | claude mcp list → connected/total + failed names. Verify at session start per mcp-first-with-fallback.md §3 (avoid defaulting to CLI all session because MCP availability was never checked). |
| Wave (optional) | wave-history.jsonl tail — only when wave-pack methodology in use |
| Blocker gaps (optional) | gap-status.csv — only when a gap pipeline exists |
| Cloud stack (opt-in) | --cloud → read-only aws sts/ec2/rds/cloudwatch describe; cached 30m |
Example output (bilingual — labels in Vietnamese per kit convention; technical terms kept English):
## Ngữ cảnh session (2026-01-01 09:30)
**Nhánh:** main (clean) / worktrees: 0
**Mức repo:** GREEN (CI green, 0 CVE, 0 stale branches, 0 audit P0)
**PRs đang mở:** 0
**MCP servers:** 1/1 connected (github)
**Gaps blocker (top):** GAP-101, GAP-102 ← only if gap pipeline exists
**Wave hiện tại:** Wave example — shipped ← only if wave methodology in use
**Sức khỏe context:** fresh session
**Đề xuất tiếp theo:** pick a blocker gap to fix OR /repo-status for a deeper triage
Step 4 flags (collect-state.sh)
| Flag | Effect |
|---|
--quick | One-line summary |
--json | Structured JSON for hooks / other agents |
--no-lock | Skip session-lock check / create |
--cloud | Opt-in: probe cloud stack (AWS read-only) |
--refresh-cloud | Bypass 30m cloud cache, fetch now |
Step 4.5 — Parallel-eligibility hint (mandatory after determining "next action")
Whenever proposing a next action, scan whether it is wave-eligible. If yes → nudge the user toward a wave + parallel agents instead of serial PRs.
Wave-eligible when:
- ≥3 clearly-divided sub-tasks (Phase 1/2/3, Sub-PR a/b/c, Layer 1/2/3…)
- Files each sub-task touches are disjoint (no shared config, no shared migration version, no shared service class)
- Each sub-task is a self-contained TDD/build cycle (no waiting on another)
NOT wave-eligible when: only 1-2 sub-tasks; sub-tasks share a migration version slot; sub-tasks share a single config file; or foundation hasn't shipped yet (PR-0 needs to ship shared interfaces first).
If wave-eligible → output:
**Wave-eligible:** YES — N disjoint sub-tasks. Suggest: wave plan + spawn 4-5 parallel agents (`isolation:worktree`, max 5).
Reference: `.claude/skills/quality/wave-pack-planner/SKILL.md`.
Else:
**Wave-eligible:** NO — {reason, e.g. "shared config file" / "1 sub-task"}. Serial PR is fine.
Rationale: picking a multi-phase task and shipping serial PRs is ~3x slower than parallel agents on disjoint scope.
Step 5 — Handoff to next skill
User decides:
- Wave plan (if Step 4.5 says YES) —
.claude/skills/quality/wave-pack-planner/SKILL.md
/repo-status — deeper health check- Single-gap PR flow —
.claude/skills/workflow/start-pr/SKILL.md
/check-pr / /fix-pr — audit / repair an in-flight PR
Context-degradation heuristic
Flag the session degraded if ANY: turn count > ~40; last /compact > 2h ago; user reports output-quality drop; multiple retries on simple tasks.
Action when degraded: nudge user to /clear + re-run /start-session. Don't proceed with critical work (wave merge, production deploy) in a degraded session.
Session lock convention
Directory: .claude/session-locks/ (git-ignored, per-machine). File: session-{YYYYMMDD-HHMMSS}-{hostname}.lock. Schema: reference/context-template.md §Lock File Schema.
Lifecycle: created on /start-session (unless --no-lock); updated when switching branch/gap; removed on /end-session OR stale-purged after 4h. Lock is a HINT — git worktree isolation is the real conflict barrier.
Rules
- LUÔN chạy
collect-state.sh trước — không tự suy diễn status.
- If the script fails (gh unauthed, jq missing) → report it clearly, don't guess.
- MCP status MUST appear in summary — if MCP servers failed or 0/N connected, suggest fixing before critical work (GitHub MCP is needed for merge/PR ops); per
mcp-first-with-fallback.md §3 fall back to CLI but know to swap back when fixed.
- Step 4.5 is mandatory — always emit a "Wave-eligible: YES/NO" line after the next-action suggestion.
- Gap / Wave / Cloud sections are OPTIONAL — only surface them when the artifact exists.
Gotchas
collect-state.sh needs gh authenticated for PR/CI fields — falls back gracefully otherwise..claude/session-locks/ and .claude/session-cloud-cache/ MUST be git-ignored.- Stale locks (>4h) auto-purged; a crash mid-session can leave an orphan —
--no-lock or manual rm clears it.
- MCP transient connect failure:
claude mcp list may show ✗ Failed then ✓ Connected on retry (Docker daemon starting / image pulling). On failure: verify daemon, restart server, re-list; if still failed, fall back to CLI per mcp-first-with-fallback.md §3.
- Cloud section is opt-in (
--cloud) and read-only — never returns secret material (describe/list only). Cache is git-ignored.
Skill contents
SKILL.md — this filereference/context-template.md — output format + lock schema + examplesscripts/collect-state.sh — state collector (full / --quick / --json; --cloud / --refresh-cloud / --no-lock)
Related
.claude/skills/workflow/repo-status/SKILL.md — deeper remote health.claude/skills/workflow/end-session/SKILL.md — symmetric counterpart (docs-sync + handoff + lock release).claude/skills/quality/wave-pack-planner/SKILL.md — wave + parallel-agent methodology (Step 4.5 handoff)- Rule:
.claude/rules/output-review-mandate.md (sessions produce outputs)
- Rule:
.claude/rules/mcp-first-with-fallback.md (MCP status check at session start)
- Rule:
.claude/rules/context-budget-mandate.md (why Step 2 prefers queries over file reads)
