rheinmir orca-onboard
🧭 Dispatch backend — chọn agent/model rẻ, chạy nhiều worker song song, cú pháp opencode/orchestration → xem orca-dispatch-reference (nguồn chân lý duy nhất, đừng nhân bản syntax ở đây).
Onboard codebase via distilled understand-anything pipeline (graph + git history, no plugin), then domain enrichment + wiki + HTML.
Triggers
- "onboard codebase", "analyze codebase", "knowledge graph", "guided tour"
Options
--full— Delete.overstack/graph/and rebuild--update— Incremental: re-analyze only changed files since last run--language <lang>— Output language (vi, en, zh, ...)--skip-wiki— Skip wiki generation<path>— Target directory (default:.)
Progress
[Phase 1/4] Graph generation (scan + git history + batch analyze + layers/tour)...
[Phase 2/4] Domain enrichment (Claude)...
[Phase 3/4] Wiki generation (opencode + DeepSeek Flash v4)...
[Phase 4/4] HTML docs (opencode + DeepSeek Flash v4)...
⚠️ HARD RULES
⛔ DO NOT run Phase 0.5 or Phases 1–4 until Phase 0 completes.
Strict order — no reversal:
- Phase 0 → complete, verify
llmwiki/exists ✅ - Phase 0.5 (Gate) → create draft file FIRST, then ask user
- Phases 1–4 → execute, update status after EACH phase immediately
llmwiki/missing → bootstrap in Phase 0, NEVER defer- Bootstrap fail → STOP completely, report error
- After EACH phase → update draft status NOW, never at end
Dispatch Rules
TASK TYPE → AGENT → MODEL
──────────────────────────────────────────────────────
Scan + git history → bash/python direct → (no LLM)
Batch file analyze → python static parse → (no LLM — PRIMARY; opencode chỉ enrich)
Merge + validate → python inline → (no LLM)
Layers / tour → Claude main thread → Sonnet (never dispatch out)
Domain reasoning → Claude main thread → Sonnet (never dispatch out)
Wiki / HTML render → opencode → opencode/deepseek-v4-flash-free
Reasoning tasks MUST stay in Claude main thread. NO reasoning to opencode/agy. opencode + DeepSeek: template fill, wiki render, HTML only.
Agent Binaries
| Agent | Binary | Default model | Check |
|---|---|---|---|
| Antigravity | agy |
Claude (Sonnet/Opus) | agy --version |
| OpenCode | opencode |
Configurable | opencode --version |
| Kiro | kiro-cli |
— | kiro-cli --version |
Không cần plugin Understand-Anything. Phase 1 dùng phương pháp đã distill từ Understand-Anything (scan → batch analyze → merge → layers → tour → validate) viết thẳng trong skill này — chạy bằng bash/python + opencode prompt thuần + Claude main thread. Không cài thêm gì vào agy/opencode.
OpenCode with DeepSeek Flash v4:
opencode run "$SPEC" --model opencode/deepseek-v4-flash-free < /dev/null \
# Fallback if flag unsupported:
echo "$SPEC" | opencode --model opencode/deepseek-v4-flash-free
# Last resort: Claude main thread
Status Update Helper
update_phase_status() {
local keyword="$1"
local status="$2"
[ -z "$DRAFT_FILE" ] && echo "[WARN] DRAFT_FILE unset" && return 0
local f=$(echo "$DRAFT_FILE" | tr '' '/')
[ ! -f "$f" ] && echo "[WARN] DRAFT_FILE not found: $f" && return 0
perl -i -pe "s/(\Q$keyword\E.*?)\| (?:pending|in-progress|done)/\$1| $status/g if /\Q$keyword\E/" "$f" \
|| sed -i "s/\(.*$(echo "$keyword" | sed 's/[^^]/[&]/g').*\)| [a-z-]*/\1| $status/" "$f" \
|| echo "[WARN] status update failed for: $keyword"
}
Phase 0 — Pre-flight & Setup
# --- 1. Dependency check (agy optional — Phase 1 dùng distilled pipeline, không cần agy) ---
MISSING=()
opencode --version >/dev/null 2>&1 || MISSING+=("opencode")
ls ~/.agents/skills/docs-site-macos/SKILL.md >/dev/null 2>&1 || MISSING+=("docs-site-macos skill")
if [ ${#MISSING[@]} -gt 0 ]; then
echo "❌ Missing dependencies:"
for dep in "${MISSING[@]}"; do echo " - $dep"; done
echo ""
echo "docs-site-macos: npx skills add rheinmir/setup@orca --skill docs-site-macos --global -y"
exit 1
fi
# --- 2. Resume Mode check & Project Root resolution ---
if [[ "$1" == @* ]]; then
DRAFT_FILE="${1#@}"
DRAFT_FILE=$(echo "$DRAFT_FILE" | tr '' '/')
RESUME_MODE=true
PROJECT_ROOT=$(grep "Project root:" "$DRAFT_FILE" | grep -oP '(?<=`)[^`]+(?=`)' | head -1)
PROJECT_ROOT=$(echo "$PROJECT_ROOT" | tr '' '/')
else
RESUME_MODE=false
PROJECT_ROOT=${1:-.}
PROJECT_ROOT=$(echo "$PROJECT_ROOT" | tr '' '/')
fi
test -d "$PROJECT_ROOT" || { echo "❌ Project root directory not found: $PROJECT_ROOT"; exit 1; }
# --- 3. Update Mode check ---
UPDATE_MODE=false
for arg in "$@"; do
[ "$arg" = "--update" ] && UPDATE_MODE=true
done
if [ "$UPDATE_MODE" = "true" ]; then
[ ! -f "$PROJECT_ROOT/.overstack/graph/knowledge-graph.json" ] \
&& echo "❌ --update requires existing graph. Run without --update first." && exit 1
[ ! -f "$PROJECT_ROOT/.overstack/graph/meta.json" ] \
&& echo "❌ --update requires meta.json. Run without --update first." && exit 1
echo "[update] Incremental mode — reading changed files from meta.json"
CHANGED_FILES=$(python3 -c "
import json, sys
meta = json.load(open('$PROJECT_ROOT/.overstack/graph/meta.json'))
changed = meta.get('changedFiles', meta.get('changed', []))
print('\n'.join(changed))
" 2>/dev/null || echo "")
if [ -z "$CHANGED_FILES" ]; then
echo "[update] no changedFiles in meta.json — falling back to full rebuild"
UPDATE_MODE=false
else
echo "[update] Changed files:"; echo "$CHANGED_FILES"
fi
fi
# --- 4. Bootstrap llmwiki + harness if missing ---
# Harness có 2 mode: GLOBAL (~/.claude/harness + hooks trong ~/.claude/settings.json,
# bảo vệ mọi project llmwiki trên máy này) và PER-PROJECT (harness/ commit vào repo,
# bảo vệ teammate). Global đủ cho máy dev cá nhân — chỉ WARN per-project khi cả 2 thiếu.
GLOBAL_HARNESS=false
[ -f "$HOME/.claude/harness/hooks/pre_tool_use.py" ] \
&& grep -q 'harness/hooks/pre_tool_use.py' "$HOME/.claude/settings.json" 2>/dev/null \
&& GLOBAL_HARNESS=true
echo "[harness] global=$GLOBAL_HARNESS project=$([ -d "$PROJECT_ROOT/harness" ] && echo yes || echo no)"
if [ ! -d "$PROJECT_ROOT/llmwiki" ] || { [ "$GLOBAL_HARNESS" = "false" ] && [ ! -d "$PROJECT_ROOT/harness" ]; }; then
echo "[orca-onboard] bootstrapping llmwiki + harness..."
git clone https://github.com/rheinmir/setup.git /tmp/orca-llmwiki-bootstrap --depth 1 -b orca -q
[ ! -d "$PROJECT_ROOT/llmwiki" ] && cp -r /tmp/orca-llmwiki-bootstrap/llmwiki "$PROJECT_ROOT/llmwiki"
if [ "$GLOBAL_HARNESS" = "false" ] && [ ! -d "$PROJECT_ROOT/harness" ]; then
bash /tmp/orca-llmwiki-bootstrap/harness/scripts/install-harness.sh "$PROJECT_ROOT" \
&& echo "[orca-onboard] harness installed OK" \
|| echo "[WARN] harness install failed/denied — user chạy 1 trong 2: 'install-harness.sh .' (per-project, cho team) hoặc 'install-harness.sh --global' (cả máy, khuyên dùng cho máy dev)"
fi
rm -rf /tmp/orca-llmwiki-bootstrap
echo "[orca-onboard] bootstrap done"
fi
# --- 5. Create dirs ---
mkdir -p $PROJECT_ROOT/.overstack/onboard/{intermediate,tmp}
mkdir -p $PROJECT_ROOT/llmwiki/wiki/draft/{cave,uiux,orca}
# --- 6. File count & availability check ---
git rev-parse HEAD 2>/dev/null > $PROJECT_ROOT/.overstack/onboard/tmp/commit.txt
git ls-files > $PROJECT_ROOT/.overstack/onboard/tmp/files.txt 2>/dev/null \
|| find $PROJECT_ROOT -type f \
! -path "*/.overstack/onboard/*" ! -path "*/llmwiki/*" ! -path "*/.git/*" \
> $PROJECT_ROOT/.overstack/onboard/tmp/files.txt
FILE_COUNT=$(wc -l < $PROJECT_ROOT/.overstack/onboard/tmp/files.txt | tr -d ' ')
# Probe agent availability
AGY_OK=$(agy --version 2>/dev/null && echo "✅ usable" || echo "❌ not found")
OC_OK=$(opencode --version 2>/dev/null && echo "✅ usable" || echo "❌ not found")
KIRO_OK=$(kiro-cli --version 2>/dev/null && echo "✅ usable" || echo "❌ not found")
echo "[pre-flight] $FILE_COUNT files | agy=$AGY_OK | opencode=$OC_OK"
# Parse statuses if in resume mode
parse_status() {
local keyword="$1"
awk '/## Agent Task Assignment/{p=1} p && /^## [^A]/{p=0} p' "$DRAFT_FILE" \
| grep "$keyword" | grep -oE 'pending|in-progress|done' | head -1
}
if [ "$RESUME_MODE" = true ]; then
PHASE1_STATUS=$(parse_status "Phase 1 —")
PHASE2_STATUS=$(parse_status "Phase 2 —")
PHASE3_STATUS=$(parse_status "Phase 3 —")
PHASE4_STATUS=$(parse_status "Phase 4 —")
echo "[Resume] P1=$PHASE1_STATUS P2=$PHASE2_STATUS P3=$PHASE3_STATUS P4=$PHASE4_STATUS"
fi
Phase 0.5 — Gate
Required: create draft file FIRST, then ask user. Never reverse this order.
DATE=$(date +%d%m%y)
PROJECT_SLUG=$(basename "$PROJECT_ROOT" | tr '[:upper:]' '[:lower:]' | tr ' ' '-')
DRAFT_FILE="$PROJECT_ROOT/llmwiki/wiki/draft/orca/${DATE}-onboard-${PROJECT_SLUG}.md"
# Graph = distilled pipeline (no plugin): scan/history bash → batch opencode → layers/tour Claude
OC_AVAIL=$(opencode --version 2>/dev/null && echo "yes" || echo "no")
if [ "$OC_AVAIL" = "yes" ]; then
AGENT_GRAPH="distilled pipeline: bash + opencode (batch) + Claude (layers/tour)"
else
AGENT_GRAPH="distilled pipeline: bash + Claude main thread (opencode unavailable)"
fi
cat > "$DRAFT_FILE" << EOF
---
type: draft
title: ${DATE}-onboard-${PROJECT_SLUG}
status: proposed
tags: [orca-onboard, output-report]
timestamp: $(date +%Y-%m-%d)
---
# ${DATE}-onboard-${PROJECT_SLUG}
**Status:** proposed
## Agent CLI Availability
| Agent | Binary | Status |
|-------|--------|--------|
| Antigravity | \`agy\` | $AGY_OK |
| OpenCode | \`opencode\` | $OC_OK |
| Kiro | \`kiro-cli\` | $KIRO_OK |
## Agent Task Assignment
| Task | Agent | Model | Status |
|------|-------|-------|--------|
| Phase 1 — Graph generation ($FILE_COUNT files) | $AGENT_GRAPH | DeepSeek Flash v4 + Sonnet | pending |
| Phase 2 — Domain enrichment | Claude main thread | Sonnet | pending |
| Phase 3 — Wiki generation | opencode | DeepSeek Flash v4 | pending |
| Phase 4 — HTML docs | opencode | DeepSeek Flash v4 | pending |
## What
Onboard \`$PROJECT_ROOT\` — understand-anything graph, domain enrichment, wiki, HTML.
## Output
- \`.overstack/graph/knowledge-graph.json\` (tree-sitter + Louvain)
- \`.overstack/graph/ONBOARDING.md\` (~20k tokens distilled)
- \`.overstack/onboard/intermediate/domain-graph.json\`
- \`llmwiki/wiki/\` (index, concepts, entities, architecture, tours)
- \`llmwiki/html/onboarding-${PROJECT_SLUG}.html\`
- \`llmwiki/html/wiki-graph.html\` (vector quan hệ concept↔code — STEP C)
## Files
| File | Action |
|------|--------|
| \`.overstack/graph/knowledge-graph.json\` | created by Phase 1 pipeline |
| \`.overstack/graph/ONBOARDING.md\` | created by Phase 1 pipeline |
| \`.overstack/onboard/intermediate/domain-graph.json\` | created by Claude |
| \`llmwiki/wiki/index.md\` | created/modified |
| \`llmwiki/wiki/concepts/architecture.md\` | created |
| \`llmwiki/wiki/concepts/*.md\` | created |
| \`llmwiki/wiki/entities/*.md\` | created |
| \`llmwiki/wiki/concepts/onboarding-tour.md\` | created |
| \`llmwiki/html/onboarding-${PROJECT_SLUG}.html\` | created |
| \`llmwiki/html/wiki-graph.html\` | created by STEP C (vector) |
## Notes
- Invoked via: \`/orca-onboard\` skill
- Project root: \`$PROJECT_ROOT\`
- Files tracked: $FILE_COUNT
- Reasoning phases in Claude main thread — NOT dispatched to cheap models
- Mechanical phases: opencode + DeepSeek Flash v4
## Cost Estimate
| Phase | Agent | Est. tokens | Est. cost |
|-------|-------|-------------|-----------|
| Phase 1 (graph) | bash + opencode batches + Claude layers/tour | ~1.5M (DeepSeek) + ~50k (Sonnet) | ~\$0.50 |
| Phase 2 (domain) | Claude Sonnet | ~50k | ~\$0.50 |
| Phase 3 (wiki) | DeepSeek Flash | ~100k | ~\$0.02 |
| Phase 4 (HTML) | DeepSeek Flash | ~50k | ~\$0.01 |
## Origin
- **Draft:** \`wiki/draft/orca/${DATE}-onboard-${PROJECT_SLUG}.md\`
- **Commit:** _(filled by verify-before-commit)_
- **Date promoted:** _(filled by verify-before-commit)_
EOF
# Update wiki index + log
echo "| [${DATE}-onboard-${PROJECT_SLUG}](draft/orca/${DATE}-onboard-${PROJECT_SLUG}.md) | draft | $(date +%Y-%m-%d) |" >> "$PROJECT_ROOT/llmwiki/wiki/index.md"
echo "## $(date +%Y-%m-%d) — orca-onboard — onboard-${PROJECT_SLUG}" >> "$PROJECT_ROOT/llmwiki/wiki/log.md"
echo "DRAFT_FILE=$DRAFT_FILE"
# --- Terminal dispatch board (REQUIRED — must print in terminal BEFORE the gate question) ---
P1_NOTE="(distilled pipeline: 1.1-1.2 bash, 1.3 opencode batches, 1.4/1.6 python, 1.5 Claude)"
echo ""
echo "==================== ORCA-ONBOARD DISPATCH BOARD ===================="
echo " Project : $PROJECT_SLUG"
echo " Files : $FILE_COUNT"
echo " Draft : $DRAFT_FILE"
echo "----------------------------------------------------------------------"
printf " %-7s %-22s %-24s %-20s %s\n" "PHASE" "TASK" "AGENT" "MODEL" "EST.COST"
printf " %-7s %-22s %-24s %-20s %s\n" "1" "Graph generation" "bash+opencode+Claude" "DeepSeek+Sonnet" "~\$0.50"
printf " %-7s %-22s %-24s %-20s %s\n" "2" "Domain enrichment" "Claude main thread" "Sonnet" "~\$0.50"
printf " %-7s %-22s %-24s %-20s %s\n" "3" "Wiki generation" "opencode" "DeepSeek Flash v4" "~\$0.02"
printf " %-7s %-22s %-24s %-20s %s\n" "4" "HTML docs" "opencode" "DeepSeek Flash v4" "~\$0.01"
[ -n "$P1_NOTE" ] && echo " Phase 1 $P1_NOTE"
echo "======================================================================"
# --- Print propose file path + content (REQUIRED — user reviews this before approving) ---
echo ""
echo "PROPOSE FILE: $DRAFT_FILE"
echo "----------------------------------------------------------------------"
cat "$DRAFT_FILE"
⛔ The dispatch board AND the propose file content above MUST be printed via bash (visible in terminal) — restating them only in chat or only in the draft file is NOT sufficient. The user gates orchestration based on this board + propose. The draft's Agent Task Assignment rows must match the board (e.g. if Phase 1 is rerouted to opencode for >100 files, the draft must say opencode, not agy).
Ask for confirmation before continuing.
Phase 1 — Graph Generation (distilled understand-anything — KHÔNG cần plugin)
Pipeline: scan (bash) → git history (bash) → batch analyze (opencode + DeepSeek) → merge (python) → layers + tour (Claude main thread) → validate + save (python + Claude). Phương pháp distill từ Understand-Anything; mọi prompt/schema nằm ngay dưới đây — không dispatch /understand, không cài plugin.
Required output:
.overstack/graph/knowledge-graph.json(nodes, edges, layers, tour).overstack/graph/ONBOARDING.md(~20k tokens distilled).overstack/graph/meta.json
Skip check: RESUME_MODE=true + PHASE1_STATUS=done → skip to Phase 2.
Decision table:
| Graph exists | Flag | Action |
|---|---|---|
| No | (none) | full pipeline |
| Yes | (none) | skip Phase 1, use existing graph |
| Yes | --update |
re-analyze only git diff <meta.gitCommitHash>..HEAD --name-only files, merge into existing graph |
| Yes | --full |
delete .overstack/graph/, full pipeline |
1.1 SCAN (bash — no LLM)
mkdir -p "$PROJECT_ROOT/.overstack/graph" "$PROJECT_ROOT/.overstack/onboard/tmp"
git -C "$PROJECT_ROOT" ls-files > "$PROJECT_ROOT/.overstack/onboard/tmp/files.txt"
# Loại file nhị phân/lock/vendor; phân loại theo extension: code/config/docs/infra/data/script
# Đọc context: README (3000 chars đầu), manifest (package.json/pyproject.toml/Cargo.toml/go.mod), dir tree 2 cấp
# Detect entry point: src/index.ts, main.py, app.py, main.go, src/main.rs, cmd/*/main.go, manage.py, __main__.py...
1.2 GIT HISTORY (bash — no LLM)
Tín hiệu lịch sử git làm giàu graph — thứ phân tích tĩnh không có:
# Churn — file nóng (hay sửa nhất, 365 ngày)
git -C "$PROJECT_ROOT" log --since="365 days ago" --name-only --pretty=format: \
| grep -v '^$' | sort | uniq -c | sort -rn | head -40 > "$PROJECT_ROOT/.overstack/onboard/tmp/churn.txt"
# Recent — file vừa đổi (30 ngày)
git -C "$PROJECT_ROOT" log --since="30 days ago" --name-only --pretty=format: | grep -v '^$' | sort -u \
> "$PROJECT_ROOT/.overstack/onboard/tmp/recent.txt"
# Commit subjects gần nhất — chủ đề đang phát triển
git -C "$PROJECT_ROOT" log --since="90 days ago" --pretty=format:'%s' | head -50 \
> "$PROJECT_ROOT/.overstack/onboard/tmp/recent-subjects.txt"
# Co-change — python đọc `git log --name-only` theo commit, đếm cặp file đổi cùng nhau, ngưỡng ≥3
Áp dụng: top churn → tag node hot; trong recent.txt → tag recent; cặp co-change ≥3 → edge related (weight 0.5); tour + ONBOARDING ưu tiên file hot; recent-subjects cho mục "What's being worked on".
1.3 ANALYZE — PRIMARY: static parse (0 token) · opencode chỉ để enrich
Mặc định dùng python static parse (bài học run 120626-zca-bridge — vừa rẻ vừa deterministic):
regex import/export ... from + require() → resolve relative path → edges imports;
test/**.test.ts → tested_by; summary = comment đầu file / danh sách exports / heading md;
type theo extension (file/config/document). Imports là việc mechanical — LLM không chính xác hơn regex.
Chỉ dispatch opencode khi cần summary giàu ngữ nghĩa hơn cho file phức tạp (optional enrichment):
Chia file list thành batch ~25 file (file liên quan cùng batch: Dockerfile+compose, migrations, CI configs, docs). Mỗi batch dispatch:
opencode run "$BATCH_SPEC" --model opencode/deepseek-v4-flash-free < /dev/null
# ⚠ KHÔNG thêm --dangerously-skip-permissions: Claude Code auto-mode classifier sẽ DENY lệnh dispatch.
# opencode treo chờ permission / bị chặn → dùng STATIC PARSE fallback bên dưới (ưu tiên mặc định).
BATCH_SPEC (inject đủ context — DeepSeek không tự đọc gì ngoài danh sách được giao):
Analyze these files in project (, languages: ). For EACH file produce GraphNode
{id, type, name, filePath, summary (1-2 câu), tags[]}and GraphEdges{source, target, type, weight}for imports/calls/configures/documents visible in file content. Read each listed file. Write ONLY valid JSON{"nodes":[...], "edges":[...]}to.overstack/onboard/tmp/batch-<i>.json. ID convention:file:<relpath>,config:<relpath>,document:<relpath>. Files: <list kèm line counts>
Tối đa 5 batch song song. Batch fail → retry 1 lần → fail nữa thì bỏ qua, ghi PHASE_WARNINGS (partial graph > no graph). opencode unavailable → Claude main thread tự analyze batch.
1.4 MERGE + NORMALIZE (python inline — no LLM)
Claude viết script .overstack/onboard/tmp/merge.py (logic distill từ merge-batch-graphs.py của Understand-Anything): đọc mọi batch-*.json → gộp nodes/edges → chuẩn hoá ID prefix (bỏ double-prefix, thêm prefix thiếu) → dedupe node theo id (giữ bản cuối), edge theo (source,target,type) → drop edge dangling (log stderr) → thêm edges related từ co-change (1.2) → ghi assembled-graph.json.
1.5 LAYERS + TOUR (Claude main thread — REASONING, không dispatch)
Từ assembled-graph + dir tree + README:
- Layers:
[{id: "layer:<kebab>", name, description, nodeIds[]}]— mỗi file-level node thuộc ĐÚNG MỘT layer; cấu trúc thư mục là bằng chứng mạnh nhất. - Tour: 5-15 bước
[{order, title, description, nodeIds[]}]— bước 1 luôn là project overview (README), theo entry point, ưu tiên filehot; kể cùng câu chuyện README kể nhưng qua code thật.
1.6 VALIDATE + SAVE (python + Claude)
Validate deterministic (script): không duplicate node ID, không dangling edge, mọi file node trong đúng 1 layer, tour nodeIds tồn tại, đếm orphan. Sửa tự động được thì sửa (drop dangling, fill default), còn lại ghi warning — partial graph vẫn save.
Ghi:
knowledge-graph.json:{version, project: {name, languages, frameworks, description, analyzedAt, gitCommitHash}, nodes, edges, layers, tour}meta.json:{lastAnalyzedAt, gitCommitHash, analyzedFiles}ONBOARDING.md(Claude distill, ~20k tokens): overview, layers + mô tả, key flows, hot files + lý do (churn), "What's being worked on" (recent-subjects), tour narrative, stats.
Cuối: xoá batch-*.json, báo summary (files analyzed, nodes/edges by type, layers, tour steps, warnings).
Schema (distill từ Understand-Anything)
- Node types:
file, function, class, module, concept, config, document, service, table, endpoint, pipeline, schema, resource— ID:<type>:<relpath>[:<name>] - Edge types chính: imports, exports, contains, inherits, implements, calls, reads_from, writes_to, depends_on, tested_by, configures, related, documents, deploys, triggers, routes, defines_schema
- Weights: contains 1.0 · inherits/implements 0.9 · calls/exports 0.8 · imports 0.7 · depends_on/configures 0.6 · còn lại 0.5
# Gate cuối Phase 1
[ ! -f "$PROJECT_ROOT/.overstack/graph/knowledge-graph.json" ] \
&& echo "❌ Phase 1 FAIL: knowledge-graph.json missing" && exit 1
[ ! -f "$PROJECT_ROOT/.overstack/graph/ONBOARDING.md" ] \
&& echo "❌ Phase 1 FAIL: ONBOARDING.md missing" && exit 1
echo "✅ Phase 1 done"
# update_phase_status "Phase 1 —" "done"
1.7 CODE-GRAPH MCP INDEX (optional — chỉ khi server code-graph có trong mcpServers)
Tầng truy vấn quan hệ code chính xác theo line bổ sung cho graph distill ở trên: search_symbols, get_callers/callees cho impact-check tức thì. Server code-graph (~/.claude.json → mcpServers) ghi 1 db SQLite mỗi repo tại <repo>/.graph-agent/index.db.
Khi nào chạy: project có submodule/repo code thật (Go/TS/Py…) và code-graph THẬT SỰ dùng được — kiểm bằng python3 harness/scripts/dep-health.py --json và đòi status == "ok", ĐỪNG chỉ xem nó có được khai báo trong ~/.claude.json hay không. Khai báo trong config ≠ server sống: DB 0 byte, DB thiếu bảng symbols, hay server đã chết đều vẫn "đang khai báo" (bài học 2026-07-20 — code-graph hỏng nhiều tuần mà mọi phiên vẫn bị lùa vào nó). degraded/absent thì bỏ qua bước này.
# 1. Index từng repo code (auto-detect root từ .git/go.mod/package.json).
# Gọi MCP tool reindex_repo(<abs path repo>, force=true) cho MỖI repo code.
# Lưu ý: code có thể nằm ở SUBDIR của submodule (vd <sub>/<svc>/go.mod chứ không phải <sub>/).
# → trỏ vào thư mục chứa go.mod/package.json, không phải submodule root rỗng.
# 2. reindex_repo tự thêm repo vào registry ~/.graph-agent/repos.txt
# → server --watch re-watch sau restart (durability).
# 3. Verify: get_stats() phải thấy symbols thật (>0), không phải stub.
echo "[1.7] code-graph: gọi reindex_repo cho mỗi repo code có manifest"
Durability harness: hook SessionStart code_graph_keeper.py (cài kèm install-harness.sh) tự re-register repo nếu rớt registry + cảnh báo repo chưa index. Fail-open, no-op nếu project không dùng code-graph — điều kiện là có .graph-agent/index.db dùng được (mở được + có bảng symbols/files), không phải chỉ có file: DB rỗng từng khiến keeper ghi nhầm registry và dập tắt luôn cảnh báo "chưa index". KHÔNG reindex per-edit (watcher đã lo, debounce 2s). Sau đó ghi 1 wiki concept page concepts/code-graph-nav.md (cách dùng tool + ví dụ caller trace) để team biết dùng.
Freshness sau đó tự động: edit file → watcher reindex ~2s. Không cần reindex tay trừ khi clone mới hoặc đổi hàng loạt lúc server tắt.
Phase 2 — Domain Enrichment
Agent: Claude main thread (no dispatch) | Model: Sonnet
READ FIRST (in order, before writing anything):
.overstack/graph/ONBOARDING.md— full read (~20k tokens).overstack/graph/knowledge-graph.json— readlayersarray +entry-pointtagged nodes ONLY (⛔ NOT full file — context overflow)- If
UPDATE_MODE=true: read existing.overstack/onboard/intermediate/domain-graph.json(merge, don't overwrite)
DO:
- From entry points → identify HTTP endpoints / CLI commands / events / cron jobs
- Reverse-engineer: entry point → flow (process) → steps (actions @ file:line)
- Build
domain → flow → stephierarchy - Write
.overstack/onboard/intermediate/domain-graph.json
Domain graph schema:
{
"domains": [
{
"id": "domain:order-management",
"name": "Order Management",
"flows": [
{
"id": "flow:create-order",
"name": "Create Order",
"steps": [
{ "order": 0.1, "id": "step:validate-input", "file": "src/orders/validator.ts", "line": 42 },
{ "order": 0.2, "id": "step:check-inventory", "file": "src/inventory/checker.ts", "line": 15 }
]
}
]
}
]
}
Output: .overstack/onboard/intermediate/domain-graph.json
Skip check: RESUME_MODE=true + PHASE2_STATUS=done → skip to Phase 3.
if [ "$RESUME_MODE" != "true" ] || [ "$PHASE2_STATUS" != "done" ]; then
update_phase_status "Phase 2 —" "in-progress"
# Claude main thread — no dispatch
if [ "$UPDATE_MODE" = "true" ] && [ -n "$CHANGED_FILES" ]; then
echo "[Phase 2] Update mode: re-enrich domain steps for changed files only"
# Read existing domain-graph.json
# Filter steps with filePath in CHANGED_FILES
# Re-analyze those steps from ONBOARDING.md
# Merge back into domain-graph.json (keep unchanged steps)
else
echo "[Phase 2] Full domain enrichment from ONBOARDING.md"
# Read ONBOARDING.md + layers from knowledge-graph.json, write new domain-graph.json
fi
update_phase_status "Phase 2 —" "done"
fi
Phase 3 — Wiki Generation
Agent: opencode | Model: opencode/deepseek-v4-flash-free
READ FIRST (inject into SPEC before dispatch — DeepSeek won't read files unless injected):
.overstack/graph/ONBOARDING.md— full read.overstack/onboard/intermediate/domain-graph.json— full read- If
UPDATE_MODE=true: grepllmwiki/wiki/for refs toCHANGED_FILES→ identify stale pages
- ⛔ DO NOT read
.overstack/graph/knowledge-graph.json— too large, context overflow
DO: Fill wiki templates from distilled content. No reasoning — mechanical template fill.
Output structure:
llmwiki/wiki/
├── index.md
├── concepts/ ← architecture.md, message-sync/flows, onboarding-tour.md (R5: KHÔNG tạo folder architecture/ hay tours/ — validator folder_structure chặn)
└── entities/ ← domain entities + project-structure.md
# Mọi trang PHẢI có '## Origin' (R2) và row trong index.md (R3)
Skip check: RESUME_MODE=true + PHASE3_STATUS=done → skip to Phase 4.
if [ "$RESUME_MODE" != "true" ] || [ "$PHASE3_STATUS" != "done" ]; then
update_phase_status "Phase 3 —" "in-progress"
if [ "$UPDATE_MODE" = "true" ] && [ -n "$CHANGED_FILES" ]; then
STALE_PAGES=$(grep -rl "$CHANGED_FILES" "$PROJECT_ROOT/llmwiki/wiki/" 2>/dev/null | tr '\n' ' ')
if [ -n "$STALE_PAGES" ]; then
echo "[Phase 3] Stale wiki pages: $STALE_PAGES"
SPEC="Update these wiki pages based on changes in ONBOARDING.md: $STALE_PAGES
Changed source files: $CHANGED_FILES
Do NOT regenerate unchanged pages. Preserve existing content in unchanged pages."
else
echo "[Phase 3] No stale wiki pages — skipping"
update_phase_status "Phase 3 —" "done"
fi
else
SPEC="Generate wiki pages from .overstack/graph/ONBOARDING.md and .overstack/onboard/intermediate/domain-graph.json.
Create: llmwiki/wiki/index.md, llmwiki/wiki/concepts/*.md (architecture.md + flow pages + onboarding-tour.md),
llmwiki/wiki/entities/*.md (domain entities + project-structure.md).
R5: chỉ được ghi vào concepts/ hoặc entities/. R2: mọi trang có '## Origin'. R3: cập nhật index.md.
Use wikilink format [[page-name]]. Do NOT read knowledge-graph.json directly."
fi
opencode run "$SPEC" --model opencode/deepseek-v4-flash-free < /dev/null \
|| echo "[WARN] opencode unavailable — Claude main thread fallback for wiki"
update_phase_status "Phase 3 —" "done"
fi
Phase 4 — HTML Docs (skeleton v2 + JSON island)
Agent: assemble JSON = opencode (mechanical) / Claude fallback · fill skeleton = python (no LLM)
Cách hoạt động (KHÁC bản cũ — chống sơ sài + đứt UX): KHÔNG để model sinh HTML thô
(DeepSeek tự chế CSS/JS → mất "gương", ấn không ăn, output 173–816 dòng không ổn định).
Thay vào đó: model CHỈ phát một object JSON theo schema; UI do skeleton v2 frozen
(assets/docs-site-skeleton.html, đã áp đúng design-system /docs-site-macos) render.
Skeleton = nav 5 tab CỐ ĐỊNH (Overview/Architecture/Guided Tour/Modules/Docker) + sidebar
fixed + collapse + scroll-spy + tour master-detail + draggable diagram; Modules/Docker tự
ẩn khi mono. Nội dung con (layer/tour/module/docker) DATA-DRIVEN từ JSON.
READ FIRST (nguồn GIÀU — inject vào SPEC; DeepSeek không tự đọc file):
.overstack/graph/ONBOARDING.md— full (overview, hot files, flows, tour narrative).overstack/onboard/intermediate/domain-graph.json— domain→flow→step (tour + lifecycle)knowledge-graph.jsonchỉlayers+tour+ node entry-point (⛔ KHÔNG full — overflow) ⛔ KHÔNG dùng wiki md mỏng làm nguồn chính (đó là lý do bản cũ sơ sài).
Skip check: RESUME_MODE=true + PHASE4_STATUS=done → skip.
if [ "$RESUME_MODE" != "true" ] || [ "$PHASE4_STATUS" != "done" ]; then
update_phase_status "Phase 4 —" "in-progress"
# --- locate skeleton v2 ---
SKELETON=$(ls ~/.agents/skills/orca-onboard/assets/docs-site-skeleton.html \
~/.claude/skills/orca-onboard/assets/docs-site-skeleton.html \
"$PROJECT_ROOT/skills/orca-onboard/assets/docs-site-skeleton.html" 2>/dev/null | head -1)
[ -z "$SKELETON" ] && echo "❌ Phase 4: skeleton v2 not found" && exit 1
# --- detect Docker (multi-service → tab Docker; mono → docker:null, Modules:[]) ---
DKFILE=$(ls "$PROJECT_ROOT"/docker-compose.y*ml "$PROJECT_ROOT"/compose.y*ml 2>/dev/null | head -1)
N_SVC=0; [ -n "$DKFILE" ] && N_SVC=$(grep -cE '^ [a-zA-Z0-9_-]+:' "$DKFILE" 2>/dev/null || echo 0)
echo "[Phase 4] skeleton=$SKELETON | compose=$DKFILE | services=$N_SVC"
# --- STEP A: assemble ONBOARD_JSON (model emit JSON ONLY → file) ---
JSON_OUT="$PROJECT_ROOT/.overstack/onboard/tmp/onboard.json"
SPEC="Đọc .overstack/graph/ONBOARDING.md + .overstack/onboard/intermediate/domain-graph.json.
Phát ra MỘT object JSON (KHÔNG markdown, KHÔNG giải thích) theo ĐÚNG schema trong header của
$SKELETON, ghi vào $JSON_OUT. Yêu cầu chất lượng:
- project: name/subtitle/about thật; stack[] + versions[] từ manifest; stats[] (files, services, layers, tour steps).
- architecture.layers[] = layers thật; architecture.diagram.nodes[] (toạ độ trong ~760×rộng, rect≥70×30) + edges[] theo luồng.
- tour[] 5–15 bước, MỖI bước GIÀU: role(1-2 câu), file+line THẬT, in[]/out[], hot(churn), narr — KHÔNG 1–2 dòng.
- modules[] = mỗi container/image (đọc $DKFILE): name,img,meta(port/env/volume),life[[read|proc|write,desc]],arch. Repo mono ($N_SVC≤1) → modules:[].
- docker = {strategy, cmd[[text,c|p|]], table[[svc,img,cmd/port,deps]], order[]} nếu có compose; mono → docker:null."
opencode run "$SPEC" --model opencode/deepseek-v4-flash-free < /dev/null 2>/dev/null \
|| echo "[Phase 4] opencode unavailable — Claude main thread tự assemble $JSON_OUT theo schema skeleton"
# Nếu opencode treo/trống → Claude main thread đọc 2 nguồn trên, tự viết $JSON_OUT (reasoning OK ở đây).
# --- STEP B: fill skeleton (python, no LLM) — thay {{TITLE}} + {{ONBOARD_JSON}} ---
OUT="$PROJECT_ROOT/llmwiki/html/onboarding-${PROJECT_SLUG}.html"
mkdir -p "$PROJECT_ROOT/llmwiki/html"
python3 - "$SKELETON" "$JSON_OUT" "$OUT" <<'PY'
import json,sys,html
sk=open(sys.argv[1]).read(); data=json.load(open(sys.argv[2]))
name=(data.get("project") or {}).get("name","Project")
out=sk.replace("{{TITLE}}",html.escape(name)).replace("{{ONBOARD_JSON}}",json.dumps(data,ensure_ascii=False))
assert "{{ONBOARD_JSON}}" not in out and "{{TITLE}}" not in out, "token chưa thay hết"
open(sys.argv[3],"w").write(out); print("✅ filled →",sys.argv[3],len(out),"bytes")
PY
ls "$OUT" && echo "✅ Phase 4 done" || echo "❌ Phase 4 FAIL: HTML not found"
update_phase_status "Phase 4 —" "done"
fi
Fallback: opencode chết/treo → Claude main thread tự đọc
ONBOARDING.md+domain-graph.json, viết$JSON_OUTtheo schema (đây là assemble dữ liệu, OK ở Claude), rồi chạy STEP B. KHÔNG bao giờ tự viết HTML thô — luôn qua skeleton v2.
STEP C — vẽ vector wiki-graph.html (wiki vừa sinh ở Phase 3 → giờ CÓ concept để vẽ):
Onboard đẻ wiki nhưng KHÔNG tự vẽ vector quan hệ — vẽ NGAY tại đây để dự án code-only kết thúc
onboard là có luôn vector (đúng nút thắt: build-wiki-graph.py đọc llmwiki/wiki/*.md + code, không
đọc thẳng code để đẻ concept). Fail-open, không chặn onboard.
# engine: repo-local → global (cùng thứ tự hook Stop resolve)
WG=$(ls "$PROJECT_ROOT/fdk/tools/build-wiki-graph.py" \
"$HOME/.claude/harness/fdk/tools/build-wiki-graph.py" 2>/dev/null | head -1)
if [ -n "$WG" ] && [ -d "$PROJECT_ROOT/llmwiki/wiki" ]; then
ALSO=""; [ -d "$PROJECT_ROOT/fdk/wiki" ] && ALSO="--also fdk/wiki"
( cd "$PROJECT_ROOT" && python3 "$WG" llmwiki/wiki $ALSO --code-root . ) \
&& echo "✅ vector: llmwiki/html/wiki-graph.html (concept↔code)" \
|| echo "⚠️ wiki-graph vẽ lỗi — không chặn onboard (Stop hook sẽ vẽ khi wiki/code đổi)"
else
echo "⚠️ bỏ vẽ vector — thiếu engine build-wiki-graph.py (cài harness global) hoặc chưa có llmwiki/wiki"
fi
After completion:
lsof -ti :8765 >/dev/null 2>&1 || nohup npx serve -p 8765 > /tmp/serve.log 2>&1 &
echo "→ http://localhost:8765/llmwiki/html/onboarding-${PROJECT_SLUG}.html"
Rules
- READ BEFORE ACT — each phase reads all inputs in
READ FIRSTblock before any action. Never ask user about something already in a file. - NO full
knowledge-graph.jsonreads after Phase 1 — useONBOARDING.mdinstead - Real file paths only — never fabricate
- Domain graph: only reference real file:line, verify against code
- Wiki: wikilink format
[[page-name]] - Tour: 5-15 steps, start with project overview
- Reasoning tasks (domain, architecture): Claude main thread only
Errors
- Phase 1 batch fail (opencode) → retry 1 lần, fail nữa → Claude main thread tự analyze batch đó; ghi PHASE_WARNINGS
- Phase 1 fail (no graph) → kiểm tra batch-*.json có tồn tại không; merge script lỗi → đọc stderr, sửa, chạy lại
- Phase 2 domain empty → no HTTP/CLI/event entry points found; write empty domains array
- Phase 3 fail → check opencode model config; fallback Claude main thread
- Phase 4 fail → fallback: invoke docs-site-macos skill directly in Claude
Output Report
After all phases complete, write propose draft to wiki.
1. Filename: DDMMYY-<ten>.md — today's date + 2-4 kebab-case summary words
2. Write llmwiki/wiki/draft/orca/DDMMYY-<ten>.md:
---
type: draft
title: DDMMYY-<ten>
status: proposed
tags: [orca-onboard, output-report]
timestamp: YYYY-MM-DD
---
# DDMMYY-<ten>
**Status:** proposed
## Agent Task Assignment
| Task | Agent | Model | Status |
|------|-------|-------|--------|
| Phase 1 — Graph generation | agy /understand | Claude (agy) | done |
| Phase 2 — Domain enrichment | Claude main | Sonnet | done |
| Phase 3 — Wiki generation | opencode | DeepSeek Flash v4 | done |
| Phase 4 — HTML docs | opencode | DeepSeek Flash v4 | done |
## What
<One sentence>
## Output
<Key artefacts>
## Files
| File | Action |
|------|--------|
| `.overstack/graph/knowledge-graph.json` | created |
| `llmwiki/wiki/index.md` | created/modified |
| `llmwiki/html/onboarding-<slug>.html` | created |
## Notes
- Invoked via: `/orca-onboard` skill
## Origin
- **Draft:** `wiki/draft/orca/DDMMYY-<ten>.md`
- **Commit:** _(filled by verify-before-commit)_
- **Date promoted:** _(filled by verify-before-commit)_
3. Update wiki index & log:
llmwiki/wiki/index.md— append:| [DDMMYY-<ten>](draft/orca/DDMMYY-<ten>.md) | draft | YYYY-MM-DD |llmwiki/wiki/log.md— append:## YYYY-MM-DD — orca-onboard — <ten>
4. Update statuses & sync push — REQUIRED:
- Update Status column in draft file to reflect actual run
- Clone
rheinmir/setupbranchorca, copy updated SKILL.md, push:git clone git@github.com:rheinmir/setup.git /tmp/rheinmir-setup-sync -b orca --depth 1 cp ~/.agents/skills/orca-onboard/SKILL.md /tmp/rheinmir-setup-sync/skills/orca-onboard/SKILL.md cd /tmp/rheinmir-setup-sync git add . git commit -m "skill: orca-onboard — wrap understand-anything, DeepSeek mechanical dispatch" git push origin orca rm -rf /tmp/rheinmir-setup-sync
Skip Output Report only if skill produced zero artefacts and zero decisions.
Output Report
After all phases complete, write propose draft to wiki.
1. Filename: DDMMYY-<ten>.md — today's date + 2-4 kebab-case summary words
2. Write llmwiki/wiki/draft/orca/DDMMYY-<ten>.md:
---
type: draft
status: proposed
tags: [orca-onboard, output-report]
proposed: YYYY-MM-DD
---
# DDMMYY-<ten>
## Agent Task Assignment
| Task | Agent | Model | Status |
|------|-------|-------|--------|
| Phase 1 — Graph generation | agy /understand | Claude (agy) | done |
| Phase 2 — Domain enrichment | Claude main | Sonnet | done |
| Phase 3 — Wiki generation | opencode | DeepSeek Flash v4 | done |
| Phase 4 — HTML docs | opencode | DeepSeek Flash v4 | done |
## What
<One sentence>
## Output
<Key artefacts>
## Files
| File | Action |
|------|--------|
| `.overstack/graph/knowledge-graph.json` | created |
| `llmwiki/wiki/index.md` | created/modified |
| `llmwiki/html/onboarding-<slug>.html` | created |
## Notes
- Invoked via: `/orca-onboard` skill
## Origin
- **Draft:** `wiki/draft/orca/DDMMYY-<ten>.md`
- **Commit:** _(filled by verify-before-commit)_
- **Date promoted:** _(filled by verify-before-commit)_
3. Update wiki index & log:
llmwiki/wiki/index.md— append:| [DDMMYY-<ten>](draft/orca/DDMMYY-<ten>.md) | draft | YYYY-MM-DD |llmwiki/wiki/log.md— append:## YYYY-MM-DD — orca-onboard — <ten>
4. Update statuses & sync push — REQUIRED:
- Update Status column in draft file to reflect actual run
- Clone
rheinmir/setupbranchorca, copy updated SKILL.md, push:git clone git@github.com:rheinmir/setup.git /tmp/rheinmir-setup-sync -b orca --depth 1 cp ~/.agents/skills/orca-onboard/SKILL.md /tmp/rheinmir-setup-sync/skills/orca-onboard/SKILL.md cd /tmp/rheinmir-setup-sync git add . git commit -m "skill: orca-onboard — wrap understand-anything, DeepSeek mechanical dispatch" git push origin orca rm -rf /tmp/rheinmir-setup-sync
Skip Output Report only if skill produced zero artefacts and zero decisions.