harvis-vibecoding

name	harvis-vibecoding
description	Coding agent skill. Activate for: writing code, editing files, fixing bugs, adding features, refactoring, running tests, opening PRs, reading the repo, vibe coding, any task that touches source files or the terminal.
metadata	{"openclaw":{"emoji":"💻"}}

Harvis Vibe Coding Skill

You are a senior engineer. Read this whole skill before touching a single file.

---

0. Before You Do Anything — Memory + Scan First

memory_search: "harvis repo patterns"
memory_search: "python_back_end main.py"

If memory has relevant context, use it. Then scan the repo:

exec: find /home/node/.openclaw/workspace/Harvis -maxdepth 3 -type f -name "*.py" -o -name "*.ts" -o -name "*.tsx" | grep -v node_modules | grep -v __pycache__ | head -50
exec: grep -rn "target_function_or_keyword" /home/node/.openclaw/workspace/Harvis/python_back_end/ | head -20

Your workspace root: /home/node/.openclaw/workspace/Harvis/

---

1. read — Read files (always before editing)

read python_back_end/main.py
read python_back_end/main.py offset=140 limit=25    ← lines 140–165 only

• Use offset+limit for large files — don't read 500 lines when you need 20
• Read returns line numbers — copy the exact text INCLUDING indentation for edit

---

2. write — Create new files or full rewrites only

write python_back_end/new_module.py
<complete file content here>

Never use write to modify an existing file. Use edit instead.

---

3. edit — Modify existing files (THE MOST IMPORTANT TOOL)

Does a literal exact string replace. Whitespace, tabs, newlines must match exactly.

Workflow — never skip step 1:

Step 1: Read the exact lines
  exec: grep -n "def target_function" python_back_end/main.py
  read: python_back_end/main.py offset=<line-from-grep> limit=15

Step 2: Copy text verbatim from read output → paste as old_string
  edit python_back_end/main.py
  old_string: <exact text from read, character for character>
  new_string: <your replacement>

When edit fails ("Could not find the exact text"):

1. The file may have changed since you last read it — re-read
2. Your old_string has wrong whitespace — copy again from read output
3. Try a shorter, more unique old_string (just the key line + 1–2 lines of context)
4. Fall back to apply_patch (see below)

Never write old_string from memory. Always copy from read output.

---

4. apply_patch — Unified diff edits (fallback when edit keeps failing)

Use when edit fails repeatedly OR when making changes in multiple places at once.

apply_patch python_back_end/main.py
--- a/python_back_end/main.py
+++ b/python_back_end/main.py
@@ -142,6 +142,7 @@
 def my_function(x):
-    return x
+    return x * 2
+    # doubled for local kimi

Rules:

• Context lines (no +/-) must also match exactly
• The @@ -old_line,count +new_line,count @@ numbers tell it where to look
• Include 2–3 context lines around each change for reliable matching

---

5. exec — Shell commands

# Find things
exec: grep -n "keyword" python_back_end/main.py
exec: grep -rn "function_name" python_back_end/ | head -20
exec: find . -name "*.py" -newer python_back_end/main.py | head -10

# Git
exec: cd /home/node/.openclaw/workspace/Harvis && git status
exec: cd /home/node/.openclaw/workspace/Harvis && git diff python_back_end/main.py
exec: cd /home/node/.openclaw/workspace/Harvis && git log --oneline -5

# Run/test
exec: cd /home/node/.openclaw/workspace/Harvis && python3 -m pytest python_back_end/tests/ -x -q
exec: cd /home/node/.openclaw/workspace/Harvis && python3 -c "import main; print('ok')"
exec: pip show python-docx

# GitHub
exec: cd /home/node/.openclaw/workspace/Harvis && gh pr list --state open
exec: cd /home/node/.openclaw/workspace/Harvis && gh pr create --title "fix: desc" --body "## What\n- change"

---

6. process — Background processes (long-running tasks)

Use when you need to run something and keep doing other work simultaneously:

# Start a background process
process action=start name="test-watcher" command="cd /home/node/.openclaw/workspace/Harvis && python3 -m pytest -x -q --tb=short -f"

# Check its output
process action=output name="test-watcher"

# Kill it when done
process action=stop name="test-watcher"

# List running processes
process action=list

Use cases: running a dev server while coding, watching test output, long installs.

---

7. memory_search + memory_get — Persistent knowledge

# Before coding anything, search for relevant context
memory_search: "python_back_end model proxy"
memory_search: "how to add a route to main.py"
memory_search: "harvis frontend api pattern"

# Pull specific memory file content
memory_get: "project-notes"
memory_get: "MEMORY.md"

After solving a tricky problem or discovering a pattern, save it:

memory_get: "MEMORY.md"   ← read first
# then write an updated version with your new knowledge

---

8. sessions_spawn — Spawn a sub-agent for a parallel task

sessions_spawn:
  agentId: "main"
  sessionKey: "harvis-task-scan-backend"
  message: "Read python_back_end/main.py and list all API routes defined with @app.get or @app.post"

• Each spawned agent has 600s (10 min)
• Don't spawn for tasks under 30s — do them inline
• After spawning, continue your own work; check results with subagents

---

9. subagents — Manage spawned sub-agents

# See all running sub-agents and their status
subagents action=list

# Read a sub-agent's output / steer it
subagents action=steer sessionKey="harvis-task-scan-backend" message="focus only on POST routes"

# Kill a sub-agent that's stuck
subagents action=kill sessionKey="harvis-task-scan-backend"

---

10. image — Vision (read screenshots, diagrams, UI mockups)

image path="/home/node/.openclaw/workspace/Harvis/screenshot.png"

Use when the user shares a screenshot of a UI bug, error, or design mockup. Describe what you see, then write code to match it.

---

Repo Structure

/home/node/.openclaw/workspace/Harvis/
├── python_back_end/
│   ├── main.py                    ← ALL routes registered here (@app.get/post)
│   ├── workspace/
│   │   ├── model_proxy.py         ← Kimi/NVIDIA NIM proxy (httpx, streaming)
│   │   └── openclaw_client.py     ← OpenClaw WebSocket client
│   ├── requirements.txt
│   └── research/                  ← Research agent code
├── front_end/jfrontend/
│   ├── app/                       ← Next.js 14 app router pages + API routes
│   ├── components/                ← React components (UI, chat, voice)
│   └── lib/                       ← Auth, DB, utilities
├── k8s-manifests/overlays/prod/
│   ├── openclaw.yaml              ← OpenClaw pod config (ConfigMaps, timeouts, skills)
│   └── skills/                    ← Skill SKILL.md source files
└── scripts/                       ← CI/CD: build-push.sh, deploy.sh

---

Git Workflow

Commit (always specific files, never git add .):

exec: cd /home/node/.openclaw/workspace/Harvis && \
  git add python_back_end/specific_file.py front_end/jfrontend/components/Thing.tsx && \
  git commit -m "fix: what you fixed and why

Co-authored-by: brandoz2255 <[e-mail]>"

Conventional commit prefixes: fix: feat: chore: docs: refactor:

Open a PR:

exec: cd /home/node/.openclaw/workspace/Harvis && \
  git push origin HEAD && \
  gh pr create \
    --title "fix: short description under 70 chars" \
    --body "## What changed
- bullet one
- bullet two

## Why
- reason"

K8s manifest changes: open PR only. Never kubectl apply directly.

---

The Rules (never break these)

Rule	Why
Always read before edit	edit needs exact text — memory is wrong
Copy old_string from read output	Whitespace errors kill the match
Re-read and retry on edit failure	Don't give up after one miss
process for background tasks	Don't block the main agent on long waits
memory_search before starting	Prior patterns save time
Stage specific files in git	Never accidentally commit .env or secrets
PR only for K8s changes	User applies after review
Co-author trailer on every commit	brandoz2255 gets credit