start-session

name	start-session
description	Dùng khi user nói 'start new session', 'bắt đầu session mới', '/start-session', 'what's the state?', 'tình trạng hiện tại', hoặc khi cần load context sau khi /clear. Load: CLAUDE.md digest, current wave, open PRs, failing CI, active blockers. Output: single summary block. Cũng check session-lock conflicts khi chạy parallel agents.
user-invocable	true
argument-hint	[--quick] [--no-lock] [--no-aws|--refresh-aws] [--audit-snapshot]

/start-session — Session Orchestration

Prepare fresh session bằng cách load đúng context + detect conflicts với concurrent sessions.

When to use

• Bắt đầu conversation mới trong repo này
• Sau khi /clear hoặc /compact
• Trước critical work (wave merge, audit, release)
• Khi unsure "làm gì tiếp?" → chạy skill này thay vì đọc mò

Process

Step 1 — Run state collector

./.claude/skills/workflow/start-session/scripts/collect-state.sh

Script output: active branch, open PRs count, failing CI runs, top 3 blocker gaps, current wave, last /compact timestamp (if logged).

Step 2 — Read digest sources

Nếu script không đủ context, đọc theo thứ tự:

1. CLAUDE.md (top 50 lines — rules, wave strategy, naming)
2. documents/03-planning/waves/ → file newest (current wave plan)
3. documents/04-quality/gaps/ROADMAP.md (top 30 lines — current status snapshot)
4. documents/03-planning/MASTER-GAPS-FIX-PLAN.md (if planning horizon needed)

KHÔNG đọc toàn bộ ROADMAP/plan — chỉ header + current section.

Step 3 — Session-lock check (optional, --no-lock skips)

1. ls .claude/session-locks/ — list active locks
2. Nếu có lock match branch/gap user định touch → WARN user: "Session X đang edit branch Y / GAP-Z, continue = risk conflict"
3. User confirm → create own lock: session-{timestamp}-{hostname}.lock
4. Lock content: YAML với branch, gaps, started, turn_estimate

Chi tiết: reference/context-template.md §Session Locks.

Step 4 — Output summary block

Format theo reference/context-template.md. Data sources (from collect-state.sh):

Field	Source
Wave	ROADMAP.md "Next recommended wave" line (per GAP-206 fix 2026-04-24)
Blocker gaps	ROADMAP.md "GA Blockers remaining" table (per GAP-206)
Repo level	scripts/repo-status.sh --json → .level (4 factors)
CVE / stale branches	Same --json → .security, .branches
Recent merges	git log main --since='3 days ago' --oneline (last 5)
Branch state	git diff --name-only; documents/action-2.md alone = "scratchpad only"
MCP servers	claude mcp list → connected/total + failed names. Added 2026-04-26 after Wave 6 anti-pattern: session defaulted gh CLI all wave because MCP availability never checked. Per .claude/rules/mcp-first-with-fallback.md §3 must verify at session start.
AWS Phase 1 BETA	Tier 1 read-only commands per .claude/rules/agent-aws-access.md §2.1: sts get-caller-identity / ec2 describe-instances / rds describe-db-instances / elbv2 describe-load-balancers / cloudtrail describe-trails + get-trail-status / cloudwatch describe-alarms --state-value ALARM. Cached 30 min in .claude/session-aws-cache/snapshot.json (gitignored). Added 2026-05-09 after user-flagged miss "session report thiếu trạng thái AWS" — Phase 1 BETA stack on account 906286017800 (ap-southeast-1) needs visibility at session start so drift / alarms / EC2 stop don't surprise mid-wave.

Ví dụ output (tiếng Việt per CLAUDE.md §CRITICAL — GAP-207):

## Ngữ cảnh session (2026-05-09 09:30)
**Wave:** Wave 48 SHIPPED — chưa có wave kế hoạch tiếp theo (theo ROADMAP §Next Action)
**Nhánh:** main (clean) / worktrees: 0
**Mức repo:** GREEN (CI green, 0 CVE, 0 branches cũ, 0 audit P0)
**PRs đang mở:** 0
**MCP servers:** 1/1 connected (github)
**AWS Phase 1 BETA (cache 5m):**
  · Account/Region: 906286017800 / ap-southeast-1
  · EC2: 2 running, 0 stopped (kh_backend=running, kc_backend=running)
  · RDS: 1 — kite-postgres-prod=available
  · ALB: 1 — kite-alb=active
  · CloudTrail: kitehub-main=IsLogging:True
  · Alarms ALARM: 0
**Gaps blocker (top 6):** GAP-453, GAP-223, GAP-222a, GAP-016, GAP-011, GAP-014
**Sức khỏe context:** fresh session
**Đề xuất tiếp theo:** GAP-453 (PR-time E2E gate, B.2 narrow-subset) hoặc /repo-status để triage CVE

Cờ collect-state.sh

Cờ	Tác dụng
--quick	One-line summary; AWS thu gọn còn EC2-running / alarms
--json	Structured JSON cho hooks/agents khác consume
--no-aws	Skip phần AWS hoàn toàn (offline / no-creds / không quan tâm)
--refresh-aws	Bỏ qua cache 30m, gọi AWS thật ngay
--audit-snapshot	Ghi documents/04-quality/audits/aws-verification/<date>-session-start-snapshot.md per agent-aws-access.md §5
--no-lock	Bỏ qua session-lock check / create

KHÔNG infer wave từ ls -t mtime hoặc blockers alphabetical grep (bugs cũ pre-GAP-206). Luôn parse ROADMAP.md.

KHÔNG output bằng English — vi phạm CLAUDE.md §CRITICAL. Field labels + prose phải tiếng Việt. Giữ English cho: technical terms (CI, CVE, PR, gap, wave, branch — đã là loanwords trong context project), file paths, command output, code.

Step 4.5 — Parallel-eligibility hint (BẮT BUỘC sau khi xác định "Đề xuất tiếp theo")

Mỗi khi đề xuất next action, scan xem nó có wave-eligible không. Nếu yes → nhắc user dùng wave + parallel agents thay vì serial PRs.

Heuristic — gap/feature là wave-eligible khi:

1. Có ≥3 sub-tasks được phân chia rõ (Phase 1/2/3, Sub-PR a/b/c, Layer 1/2/3...)
2. Files mỗi sub-task touch là disjoint (không share application.yml, không share migration version, không share service class)
3. Mỗi sub-task self-contained TDD/build cycle (không cần wait sub-task khác xong)

Heuristic — KHÔNG wave-eligible khi:

• Chỉ 1-2 sub-tasks
• Sub-tasks share migration version slot (V_n và V_n+1 collide nếu chạy parallel)
• Sub-tasks share application.yml (rebases sẽ phải sequential)
• Foundation chưa ship (cần PR-0 ship interfaces/shared lib trước, sau đó parallel)

Nếu wave-eligible → output line:

**Wave-eligible:** YES — 5 sub-tasks disjoint files. Đề xuất: tạo wave plan + spawn 4-5 agents parallel (`isolation:worktree`, max 5).
Reference: `feedback_parallel_agent_strategy.md` + `feedback_wave_plan_before_serial_prs.md` + skill `quality-plan` / `priority-pr-planning` / `gap-to-pr-converter`.

Nếu KHÔNG wave-eligible → output line:

**Wave-eligible:** NO — {lý do, e.g. "shared application.yml", "1 sub-task only"}. Tiếp tục serial PR là OK.

Anti-pattern (phát hiện 2026-04-26 GAP-229): Pick gap có "Phase 1 / Phase 2 / Phase 3" rồi nhảy thẳng Phase 1 trên branch riêng, ship serial PRs liên tiếp. ~90min serial vs ~30min nếu parallel. Memory feedback_wave_plan_before_serial_prs.md chứa case study + checklist 3-câu.

Step 5 — Handoff to next skill

User decides:

• Wave plan (nếu Step 4.5 báo YES) — ưu tiên cao nhất; dùng quality-plan HOẶC viết wave plan ngắn (5-10 dòng) rồi single-message multi-Agent dispatch
• /continue — execute top priority from plan (single sub-task scope)
• /session-docs-check — Living Docs completeness gate (chạy trước commit/PR/merge để bắt status flip thiếu, ROADMAP entry thiếu, skill index chưa update)
• /repo-status — deeper health check
• /gap-triage — review gap backlog
• /pr-health 290-300 — audit recent merges

Context-degradation heuristic

Flag session as degraded if ANY:

• Turn count > 40 (manual count — harness doesn't expose; estimate from conversation length)
• Last /compact > 2 hours ago
• User reported output quality drop
• Multiple retries on simple tasks

Action khi degraded: nudge user to /clear + re-run /start-session. Critical work (wave merge, production deploy) → DON'T proceed in degraded session.

Session lock convention

Directory: .claude/session-locks/ (git-ignored, per-machine).

File naming: session-{YYYYMMDD-HHMMSS}-{hostname}.lock

Schema: see reference/context-template.md §Lock File Schema.

Lock lifecycle:

• Created on /start-session (unless --no-lock)
• Updated when switching branch/gap mid-session
• Removed on explicit /end-session OR stale after 4h (any new session cleans stale locks)

MVP: lock read + warn only, no enforcement hook. Phase 2 (post-pilot): add pre-commit hook to block commits on locked branch from different session.

Rules

• TUYỆT ĐỐI giao tiếp bằng tiếng Việt per CLAUDE.md §CRITICAL Communication Language. Field labels, prose, recommendations trong output — tất cả tiếng Việt. Chỉ giữ English cho: technical terms (CI, CVE, PR, gap, wave, branch, main, merge — đã là loanwords trong project context), file paths, command output, code.
• LUÔN chạy collect-state.sh trước — không tự suy diễn status
• Nếu script fail (gh unauthed, hook missing) → báo rõ, không đoán
• Wave + blockers LUÔN parse từ ROADMAP.md, không dùng ls -t mtime hoặc alphabetical grep (pre-GAP-206 bugs)
• Output format tuân reference/context-template.md (có ví dụ VN trong Step 4)
• MCP status PHẢI nêu trong summary — nếu line "MCP servers" báo failed servers hoặc 0/N, suggest user fix trước critical work (GitHub MCP cần cho merge/PR ops); per .claude/rules/mcp-first-with-fallback.md §3 default sang CLI fallback nhưng phải biết để swap khi fix xong. Anti-pattern bắt nguồn 2026-04-26 (Wave 6 session default gh CLI suốt 4 sub-PR vì không check MCP đầu session).
• Step 4.5 BẮT BUỘC — luôn output line "Wave-eligible: YES/NO" sau "Đề xuất tiếp theo". Anti-pattern bắt nguồn 2026-04-26 (GAP-229 3 phases serial = ~90min vs ~30min nếu parallel). Per memory feedback_wave_plan_before_serial_prs.md + feedback_parallel_agent_strategy.md.
• AWS section trong summary: nếu AWS Phase 1 BETA line báo no-auth, no-cli, hoặc bất kỳ alarm ALARM > 0, PHẢI surface trong "Đề xuất tiếp theo" (vd: "Triage alarm X trước infra work" hoặc "Fix AWS auth trước deploy work"). Anti-pattern: silently swallow AWS state, ship infra change rồi mới phát hiện EC2 đã stopped sáng nay. Per .claude/rules/agent-aws-access.md §1 + aws-observability-first.md.

Gotchas

• collect-state.sh needs gh CLI authenticated — fall back to manual gh pr list nếu script fails
• .claude/session-locks/ KHÔNG commit (add to .gitignore nếu chưa có)
• Stale locks (>4h) auto-purged — nhưng crash mid-session có thể leave orphan; --no-lock hoặc manual rm clear
• Hostname trên WSL = wsl-name; multiple Claude Code windows cùng host vẫn tạo unique lock (timestamp differs)
• KHÔNG rely solely trên lock để prevent conflict — lock là hint, git branch isolation (worktree) mới là enforcement thật
• MCP transient connect failure: claude mcp list có thể báo ✗ Failed rồi retry trong cùng session báo ✓ Connected — lý do thường là Docker daemon vừa khởi động hoặc image đang pull. Khi gặp failed: docker ps (verify daemon) → docker pull ghcr.io/github/github-mcp-server → claude mcp list lại. Nếu vẫn failed, báo user, fallback CLI per mcp-first-with-fallback.md §3.
• AWS auth states: no-cli (CLI chưa cài), no-auth (sts get-caller-identity fail — kiểm aws configure list + key valid), cached (đọc snapshot 30m), fresh (vừa fetch lại), skipped (--no-aws). Khi no-auth xuất hiện trên Phase 1 BETA branch, cảnh báo user trước khi làm bất kỳ infra/deploy work — drift visibility = 0.
• AWS cost: Tier 1 describe/list calls đều free-tier within reason (vài chục ngàn requests/month per service). Cache 30m + opt-out --no-aws cover heavy session pattern. KHÔNG add aws cloudtrail get-trail-status data-events query — đó là Tier 2 + tốn $0.10/100k (per agent-aws-access.md §3).
• AWS section bypasses agent-aws-access.md §5 logging by default: theo §5.3 single ad-hoc no-log; multi-command verification logs only when --audit-snapshot. Cache file is gitignored, contains no secret material (Tier 1 commands return metadata only, never get-secret-value).
• AWS region: lấy từ aws configure get region (default profile) → fallback $AWS_REGION env → ?. Phase 1 BETA dùng ap-southeast-1; nếu profile shifted region, snapshot sẽ phản ánh region đó (KHÔNG cross-region multi-call để tiết kiệm time).

Skill contents

• SKILL.md — this file
• reference/context-template.md — output format + lock schema + example
• scripts/collect-state.sh — state collector (full / --quick / --json modes; AWS flags: --no-aws / --refresh-aws / --audit-snapshot)
• data/ — session history log (optional, append-only)
• .claude/session-aws-cache/snapshot.json — AWS Tier 1 snapshot, 30m TTL, gitignored (created lazily on first AWS-enabled run)

Related

• .claude/skills/workflow/continue/SKILL.md — use after /start-session to act on top priority
• .claude/skills/workflow/repo-status/SKILL.md — deeper remote health
• .claude/skills/quality/rework-audit/SKILL.md — consumes degradation signal for retrospective audit
• Rule: .claude/rules/output-review-mandate.md (sessions produce outputs)
• Rule: .claude/rules/agent-aws-access.md (Tier 1 read-only commands powering AWS section; --audit-snapshot artifact format)
• Rule: .claude/rules/aws-observability-first.md (CloudTrail IsLogging line in AWS section closes the audit-baseline gap surfaced by this rule)
• Memory: feedback_parallel_agent_strategy.md (multi-agent safety)