Features

Status Line (HUD)

OMH replaces Claude Code's default status line with a real-time dashboard:

[OMH] | 5h:14%(3h51m) | wk:7%(6d5h) | session:29m | ctx:39% | 🔧53 | agents:2 | opus-4-6

Segment	Meaning
`5h:14%(3h51m)`	5-hour rate limit usage 14%, resets in 3h 51m
`wk:7%(6d5h)`	Weekly rate limit usage 7%, resets in 6d 5h
`session:29m`	Current session duration
`ctx:39%`	Context window usage (green → yellow at 70% → red at 85%)
`🔧53`	Total tool calls this session
`agents:2`	Currently running subagents
`opus-4-6`	Active model

Rate limit data is fetched from the Anthropic OAuth API and cached for 90 seconds.

Smart Defaults — What OMH Does Automatically

OMH hooks into Claude Code's lifecycle and activates automatically. No manual intervention needed.

┌─────────────────────────────────────────────────────────────────┐
│  You type a prompt                                              │
│                                                                 │
│  ┌──────────────────────┐   ┌──────────────────────┐            │
│  │ 🔍 Ambiguity Guard   │   │ 📋 Auto-Plan Mode    │            │
│  │ Vague request?       │   │ 3+ tasks detected?   │            │
│  │ → Ask for scope      │   │ → Suggest plan first  │            │
│  └──────────────────────┘   └──────────────────────┘            │
│                                                                 │
│  Claude starts working                                          │
│                                                                 │
│  ┌──────────────────────┐   ┌──────────────────────┐            │
│  │ 🛡️ Dangerous Guard   │   │ 📁 Scope Guard       │            │
│  │ rm -rf / force push? │   │ Edit outside allowed  │            │
│  │ → Warn + confirm     │   │ paths? → Warn         │            │
│  └──────────────────────┘   └──────────────────────┘            │
│                                                                 │
│  ┌──────────────────────┐   ┌──────────────────────┐            │
│  │ 🤖 Model Routing     │   │ 📝 Commit Convention  │            │
│  │ Delegates to the     │   │ git commit detected?  │            │
│  │ right model tier:    │   │ → Remind format       │            │
│  │ haiku/sonnet/opus    │   │                       │            │
│  └──────────────────────┘   └──────────────────────┘            │
│                                                                 │
│  Task completes                                                 │
│                                                                 │
│  ┌──────────────────────┐   ┌──────────────────────┐            │
│  │ ✅ Test Enforcement   │   │ 💾 Context Snapshot   │            │
│  │ Code changed?        │   │ Context compaction?   │            │
│  │ → Verify tests exist │   │ → Save state first    │            │
│  └──────────────────────┘   └──────────────────────┘            │
└─────────────────────────────────────────────────────────────────┘

Model Routing in Action

When Claude delegates to subagents, OMH automatically selects the right model:

Agent Tier	Model	When Used	Example Tasks
`harness:quick`	Haiku	Simple lookups, exploration	"Find all TODO comments", "What's in this file?"
`harness:standard`	Sonnet	Implementation, fixes	"Fix this bug", "Add validation", "Write tests"
`harness:architect`	Opus	Architecture, design	"Design the auth system", "Security review", "Complex refactor"

The current model is always visible in the HUD status line.

Feature Tags — `[omh:*]`

Every OMH action is prefixed with a tag so you always know which feature fired:

[omh:ambiguity-guard]    → Asking for clarification on a vague request
[omh:auto-plan]          → Detected 3+ tasks, suggesting plan mode
[omh:dangerous-guard]    → Warning before destructive command
[omh:model-routing → sonnet] → Delegating to sonnet for implementation
[omh:test-enforcement]   → Reminding to verify tests after code change
[omh:commit-convention]  → Showing commit format after git commit
[omh:scope-guard]        → Warning about edit outside allowed paths
[omh:convention-detect]  → Detected project conventions on session start
[omh:context-snapshot]   → Saving state before context compaction

Example session output:

⏺ [omh:convention-detect] Project: node | test: vitest | lint: eslint
  ...
⏺ [omh:ambiguity-guard] 요청이 모호합니다. 구체적 범위를 확인합니다.
  ...
⏺ [omh:model-routing → haiku] Finding all TODO comments...
  ...
⏺ [omh:model-routing → sonnet] Implementing the auth middleware...
  ...
⏺ [omh:dangerous-guard] WARNING: rm -rf detected. Confirm with user.
  ...
⏺ [omh:test-enforcement] 코드 변경 감지. 테스트 존재 여부 확인.

Feature Details

1. Convention Auto-Detect

Scans project root on session start and injects detected conventions as context. Results are cached for 1 hour.

Project File	Language	Detected Tools
`package.json`	Node.js	jest / vitest / mocha, eslint / biome, prettier, typescript / vite / webpack
`pyproject.toml`	Python	pytest, ruff / flake8, black, mypy
`go.mod`	Go	go test, golangci-lint
`Cargo.toml`	Rust	cargo test, clippy, rustfmt
`build.gradle`	Java	junit, gradle
`pom.xml`	Java	junit, maven

Session start message example: [oh-my-harness] Project: node | test: vitest | lint: eslint | fmt: prettier

2. Test Enforcement

After code changes (Edit / Write / NotebookEdit), injects a reminder at session stop:

Verify test files exist for changed code
Each test file has at least N cases (configurable, default: 2)
Suggest adding tests if missing

Tests must cover happy path, edge case, and error case at minimum.

3. Model Routing

Three agent tiers for cost-efficient subagent delegation:

Agent	Model	Use For
`harness:quick`	haiku	File lookups, simple questions, exploration
`harness:standard`	sonnet	Implementation, bug fixes, debugging
`harness:architect`	opus	Architecture, complex analysis, security review

CLAUDE.md instructs Claude to delegate to the appropriate tier automatically based on task complexity.

4. Auto-Plan Mode

Detects 3+ independent tasks in a single message:

Numbered items (1. 2. 3.)
Bullet points (-, *)
Korean conjunctions (그리고, 또한, 추가로, 아울러, 더불어)

Calls EnterPlanMode tool to switch to real plan mode (Shift+Tab equivalent).

5. Ambiguity Guard

Detects vague requests using a scoring system (threshold: 2):

Signal	Score	Example
Vague references	+1	"fix this", "change that"
Scope-less verbs	+1	"refactor" (no file/function target)
Open-ended choices	+1	"or something", "whatever"
Very short message	+1	< 15 chars without specific identifiers
English scope-less	+1	"fix it", "clean up" without target

When score >= threshold, Claude must ask for clarification before starting work.

6. Dangerous Guard

Warns before potentially destructive operations:

Bash tool patterns:

Pattern	Warning
`rm -rf`, `rm --force`	File deletion
`git push --force`	Force push
`git reset --hard`	Hard reset
`git clean -f`	Git clean
`DROP TABLE / DATABASE`	Database destruction
`TRUNCATE TABLE`	Table truncation
`DELETE FROM` (no WHERE)	Mass deletion
`chmod 777`	Unsafe permissions
`curl \| sh`	Remote execution
`npm publish`	Package publish
`docker system prune`	Container cleanup

Write/Edit tool patterns:

Pattern	Warning
`.env` files	Environment secrets
`credentials`	Credential files
`secret`	Secret files
`id_rsa`, `.pem`, `.key`	Private keys

Warning only — does not block execution. Asks Claude to confirm with user.

7. Context Snapshot

Before context compaction (PreCompact), saves current state to .claude/.omh/context-snapshot.md:

Session summary
Active tasks
Reminder to review snapshot after compaction

8. Commit Convention

When git commit is detected, reminds the commit format.

Auto-detection priority:

commitlint config files -> Conventional Commits
gitmoji dependency in package.json -> Gitmoji
commitizen in package.json -> Conventional Commits
Default -> Conventional Commits

# Conventional Commits
<type>(<scope>): <description>
# Types: feat, fix, docs, style, refactor, perf, test, build, ci, chore

# Gitmoji
<emoji> <description>

9. Scope Guard

When enabled with allowedPaths, warns if Edit/Write targets files outside the allowed directories.

{
  "features": { "scopeGuard": true },
  "scopeGuard": { "allowedPaths": ["src/auth", "src/utils"] }
}

OFF by default. Enable when you want to restrict Claude's write scope.

10. Usage Tracking

Silently records every tool invocation to .claude/.omh/usage.json:

{
  "sessions": {
    "session-id": {
      "tool_counts": { "Edit": 5, "Bash": 3, "Read": 12 },
      "total_calls": 20,
      "started_at": "2026-03-23T10:00:00Z",
      "last_tool": "Edit"
    }
  }
}

11. Skill Scaffolding

Automatically generates project-specific skills in .claude/skills/ based on detected conventions.

Scaffolded skills:

Skill	What it does
`code-review`	Language-specific review checklist
`test-write`	Test writing conventions for detected framework
`lint-fix`	Lint check and auto-fix workflow

Supported languages:

Language	Test Framework	Linter	Notes
Node.js	vitest / jest / mocha	eslint / biome	TypeScript support auto-detected
Python	pytest	ruff / flake8	black / ruff formatting
Go	go test	golangci-lint	Table-driven test patterns
Rust	cargo test	clippy	rustfmt formatting
Java	junit	—	Gradle / Maven build
Kotlin	kotest / junit5	ktlint / detekt	Null safety checks

How it works:

Run /init-project or oh-my-harness init
OMH detects your project's language and tools
Skill templates are rendered with your specific tools (e.g., "vitest" not "test runner")
Skills are written to .claude/skills/ — Claude Code auto-discovers them
Customize freely — OMH never overwrites existing skills

Skills are user-owned files. oh-my-harness reset will NOT delete them.

Session hint:

If no project skills are detected, OMH shows a hint on session start:

[omh:skill-hint] No project skills found. Run /init-project to scaffold.

Configuration:

{
  "features": { "skillScaffolding": true }
}

Set to false to disable scaffold hints and skip skill generation during init.

12. Native Team

Orchestrate parallel work using Claude Code's built-in team system — no tmux or worktree dependencies.

Templates:

Template	Members	Model Routing
`fullstack`	frontend + backend + tester	All sonnet
`review`	reviewer + tester	opus + sonnet
`research`	researcher + implementer + architect	haiku + sonnet + opus

Commands:

Command	What it does
`/team-spawn [template\|N] [task]`	Create team, decompose tasks, spawn teammates
`/team-status`	Show teammate status and task progress
`/team-stop`	Shutdown team with incomplete task warnings

How it works:

Run /team-spawn fullstack build auth system
OMH creates a native team via TeamCreate
Tasks are decomposed and assigned to teammates
Teammates work in parallel, communicating via SendMessage
Check progress with /team-status
Shutdown with /team-stop when done

Configuration:

{
  "features": { "nativeTeam": true },
  "nativeTeam": {
    "maxTeammates": 4,
    "defaultTeamName": "omh-team"
  }
}

Custom templates can be added via nativeTeam.templates in the config.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Features

Status Line (HUD)

Smart Defaults — What OMH Does Automatically

Model Routing in Action

Feature Tags — `[omh:*]`

Feature Details

1. Convention Auto-Detect

2. Test Enforcement

3. Model Routing

4. Auto-Plan Mode

5. Ambiguity Guard

6. Dangerous Guard

7. Context Snapshot

8. Commit Convention

9. Scope Guard

10. Usage Tracking

11. Skill Scaffolding

12. Native Team

FilesExpand file tree

features.md

Latest commit

History

features.md

File metadata and controls

Features

Status Line (HUD)

Smart Defaults — What OMH Does Automatically

Model Routing in Action

Feature Tags — [omh:*]

Feature Details

1. Convention Auto-Detect

2. Test Enforcement

3. Model Routing

4. Auto-Plan Mode

5. Ambiguity Guard

6. Dangerous Guard

7. Context Snapshot

8. Commit Convention

9. Scope Guard

10. Usage Tracking

11. Skill Scaffolding

12. Native Team

Feature Tags — `[omh:*]`