English | 한국어 | 中文 | 日本語 | Español
Multi-agent orchestration for GitHub Copilot. Supercharged productivity.
#multi-agent #mcp #workflow-automation #context-preservation #agentic-ai
Get Started • Agents • Skills • MCP Server • Architecture
▶ Watch OMG in Action on YouTube
oh-my-githubcopilot (OMG) brings the multi-agent orchestration paradigm — pioneered by oh-my-claudecode (OMC) for Claude Code — to GitHub Copilot, now further supercharged with the best features of Everything Claude Code (ECC).
Where OMC supercharges Claude Code with specialized agents and workflow automation, OMG does the same for Copilot's agent mode in VS Code. And with the ECC integration (v1.1.0) plus the adapted OMC skill port, OMG now includes language-specialist reviewers, TDD enforcement, rapid security scanning, durable ultragoal checkpoints, research helpers, release guidance, visual QA, and more. Instead of a single assistant doing everything, OMG coordinates 28 specialized agents and 37 reusable skills through an MCP server, giving you structured workflows for planning, execution, review, and verification — all within your existing Copilot setup.
This is not a fork or copy of OMC or ECC. It is an independent implementation built from scratch to leverage GitHub Copilot's agent customization features (
.agent.md,.prompt.md,SKILL.md, MCP tools), drawing architectural inspiration from OMC's multi-agent approach and selectively integrating ECC's proven patterns.
- Works inside VS Code and Copilot CLI — No extra CLI tools, no external processes. VS Code agent mode or the standalone
copilotCLI. - Specialized agents — 28 purpose-built agents with scoped access (read-only analysts, full-access executors, 8 language reviewers)
- Workflow automation — From autonomous
omg-autopilotto persistentralphloops to parallelultrawork - Safety guardrails — Pre/post tool-use hooks prevent destructive operations automatically
- MCP-powered state — Persistent workflow state, PRD tracking, and project memory across sessions
- Natural language triggers — Say "omg-autopilot build me a REST API" and the system handles orchestration
- Verification-first — Separate authoring and review passes; claims require evidence
- VS Code with GitHub Copilot Chat enabled
- Agent mode / agent customization support available in your Copilot setup
- Node.js and npm installed so the MCP server can be built locally
- Workspace opened as a trusted folder so MCP, prompts, and customization files can load normally
- Install the extension (either method below):
- Method 1 — VSIX (CLI)
code --install-extension ./vscode-omg/oh-my-githubcopilot-1.4.4.vsix
If you downloaded the VSIX elsewhere, replace the path with your local file path.
- Method 2 — VS Code Extensions tab (UI)
Open VS Code → Extensions (
⇧⌘X/Ctrl+Shift+X) → searchoh-my-githubcopilot→ Install.
-
Open your project in VS Code.
-
⚡ Run
OMG: Initialize Workspace(REQUIRED)
Cmd+Shift+P (macOS) / Ctrl+Shift+P (Windows/Linux) → "OMG: Initialize Workspace"
Important
Installing the extension alone is NOT enough. You must run OMG: Initialize Workspace from the Command Palette after installation. This command scaffolds all .github/ convention files (agents, skills, hooks, prompts, copilot-instructions.md) and builds the MCP server in your workspace. Without this step, Copilot will not have access to any OMG agents or skills.
- When prompted, click "Reload Window" to activate all agents and skills.
- Start using OMG in Copilot Chat (agent mode).
- Clone the repository:
git clone https://github.com/jmstar85/oh-my-githubcopilot.git
cd oh-my-githubcopilot- Build the MCP server:
cd mcp-server
npm install
npm run build
cd ..-
Open the project in VS Code with GitHub Copilot Chat enabled.
-
In Copilot Chat (agent mode), say:
omg-autopilot: build a REST API for managing tasks
- OMG takes over — planning, implementing, reviewing, and verifying.
OMG works with the standalone Copilot CLI (copilot binary). The CLI reads the same .github/ convention files (agents, skills, hooks, prompts).
- Install the Copilot CLI binary.
- Clone OMG and build the MCP server:
git clone https://github.com/jmstar85/oh-my-githubcopilot.git cd oh-my-githubcopilot cd mcp-server && npm install && npm run build && cd ..
- Create a project-local
.copilot/mcp-config.json(or use--global-mcpto install globally at~/.copilot/):scripts/omg-adopt.sh --target . --mode template --target-env cli - Run
copilotfrom the project directory:copilot
- Try
/statusor@omg-coordinatorto verify OMG is loaded.
If your requirements are vague or you want structured clarification:
deep-interview "I want to build a task management app"
OMG uses Socratic questioning to surface hidden assumptions and clarify requirements before any code is written.
OMG is workspace-scoped, so the recommended way is to apply it per project.
This repository now includes an adoption script:
macOS / Linux (Bash):
scripts/omg-adopt.sh --target <your-project-path> --mode <template|submodule|subtree> [--target-env vscode|cli|both]Windows (PowerShell):
scripts/omg-adopt.ps1 -Target <your-project-path> -Mode <template|submodule|subtree> [-TargetEnv vscode|cli|both]The --target-env flag (default: both) controls what gets set up:
vscode— VS Code only (.vscode/mcp.json)cli— Copilot CLI only (.copilot/mcp-config.json+hooks.json)both— Both environments (default)
Use this for greenfield projects where you want OMG files copied directly into the project.
# macOS / Linux
scripts/omg-adopt.sh --target ~/work/my-new-app --mode template# Windows
scripts/omg-adopt.ps1 -Target ~/work/my-new-app -Mode templateUse one of these when you want a sync strategy for future OMG updates.
# macOS / Linux — Submodule strategy
scripts/omg-adopt.sh --target ~/work/my-app --mode submodule
# macOS / Linux — Subtree strategy
scripts/omg-adopt.sh --target ~/work/my-app --mode subtree# Windows — Submodule strategy
scripts/omg-adopt.ps1 -Target ~/work/my-app -Mode submodule
# Windows — Subtree strategy
scripts/omg-adopt.ps1 -Target ~/work/my-app -Mode subtreeWhat the script applies to the target project:
.github/copilot-instructions.md.github/agents/.github/skills/.github/hooks/.github/prompts/.vscode/mcp.jsonmcp-server/(withnpm install && npm run build, unless--skip-build)
After applying, open the target project as a trusted workspace and validate in Copilot Chat (agent mode):
/status
OMG includes 28 specialized agents, each with defined roles and access levels. Agents are declared as .agent.md files under .github/agents/.
| Agent | Role | Access |
|---|---|---|
| @omg-coordinator | Main orchestrator — coordinates workflows, omg-autopilot, ralph loops | Full |
| @executor | Focused task implementation — code changes, features, bug fixes | Full |
| @debugger | Root-cause analysis, stack traces, build error resolution (7-language guide) | Full |
| @architect | Architecture analysis, system design, debugging guidance | Read-only |
| @planner | Strategic planning with interview workflow | Plans only |
| @analyst | Requirements analysis, gap detection, scope risk | Read-only |
| @verifier | Evidence-based completion checks, test adequacy | Test runner |
| @code-reviewer | Severity-rated code review, SOLID checks, canonical coding standards | Read-only |
| @security-reviewer | OWASP Top 10, secrets detection (sk-/ghp_/AKIA), auth/authz audit | Read-only |
| @critic | Thorough plan/code review gate, pre-mortem analysis | Read-only |
| @test-engineer | Test strategy, TDD workflows, framework detection, flaky test hardening | Full |
| @designer | UI/UX design and frontend implementation | Full |
| @writer | Technical documentation — README, API docs, CODEMAP generation | Full |
| @tracer | Evidence-driven causal tracing with competing hypotheses | Full |
| @scientist | Data analysis, statistical analysis, visualization | Read-only |
| @qa-tester | Interactive CLI testing via VS Code terminal, Playwright POM, E2E | Full |
| @git-master | Atomic commits, rebasing, history management | Git only |
| @code-simplifier | Code clarity, complexity metrics, simplification patterns | Full |
| @explore | Codebase search, file finding, structure mapping | Read-only |
| @document-specialist | External documentation research, API reference lookup | Read-only |
Invoke with @mention for language-specific code review:
| Agent | Language | Key Rules |
|---|---|---|
| @typescript-reviewer | TypeScript | Strict mode, type safety, no-any, exhaustive checks |
| @python-reviewer | Python | PEP 8, type hints, idiomatic patterns |
| @rust-reviewer | Rust | Ownership, borrow checker, unsafe blocks, clippy |
| @go-reviewer | Go | Idiomatic Go, goroutine safety, error handling |
| @java-reviewer | Java | SOLID, Spring patterns, null safety |
| @csharp-reviewer | C# | Nullable analysis, async/await, C# idioms |
| @swift-reviewer | Swift | Memory safety, Swift concurrency, SwiftUI patterns |
| @database-reviewer | SQL/ORM | Query performance, parameterization, schema design |
Skills are reusable workflow routines triggered by slash commands or natural language keywords. Defined under .github/skills/.
| Skill | What It Does | Trigger Keywords |
|---|---|---|
/omg-autopilot |
Full autonomous execution from idea to working code | omg-autopilot, build me, create me |
/ralph |
PRD-driven persistence loop — won't stop until verified complete | ralph, don't stop, finish this |
/ultrawork |
Parallel execution engine for high-throughput tasks | ulw, ultrawork, parallel |
/ultragoal |
Durable MCP-backed goal tracking with checkpoints | ultragoal, goal loop, durable goal |
/team |
N coordinated agents on a shared task list with staged pipeline | team, multi-agent, swarm |
/plan |
Structured planning with optional interview workflow | plan this, let's plan |
/ralplan |
Consensus planning with Planner/Architect/Critic loop | ralplan, consensus plan |
/ccg |
Triple-model analysis (Claude + Codex + Gemini perspectives) | ccg, tri-model, cross-validate |
/autoresearch |
Stateful research loop with evaluator gates | autoresearch, research loop |
| Skill | What It Does | Trigger Keywords |
|---|---|---|
/deep-interview |
Socratic requirements clarification with ambiguity gating | deep interview, ask me everything |
/deep-dive |
Two-stage pipeline: trace → deep-interview | deep dive, investigate deeply |
/trace |
Evidence-driven causal tracing with competing hypotheses | trace this, root cause analysis |
/external-context |
Focused external documentation research | external context, docs research |
/sciomc |
Bounded scientist-style research analysis | sciomc, scientist research |
/ask |
Route multi-perspective questions through /ccg |
ask, multi-model ask |
/verify |
Evidence-based completion verification | verify this, prove it works |
/review |
Code review with severity ratings and spec compliance | review this, code review |
/ultraqa |
QA cycling — test, verify, fix, repeat until green | ultraqa, fix all tests |
/ai-slop-cleaner |
Detect and fix AI-generated code smells | deslop, anti-slop, cleanup slop |
/self-improve |
Autonomous evolutionary code improvement with tournament selection | self-improve, evolve code |
/visual-verdict |
Structured visual QA for screenshot comparisons | visual verdict, screenshot compare |
| Skill | What It Does | Trigger Keywords |
|---|---|---|
/remember |
Save information to project memory | remember this, store this |
/wiki |
Maintain lightweight project knowledge base entries | wiki, knowledge base |
/writer-memory |
Track writing project memory and story canon | writer memory, character memory |
/skillify |
Turn repeatable workflows into reusable OMG skills | skillify, make a skill |
/deepinit |
Generate or refresh repository agent guidance | deepinit, repo init docs |
/release |
Prepare safe release notes, checks, and approval gates | release, publish version |
/configure-notifications |
Guide notification integration setup | configure notifications, alert me |
/hud |
Explain OMG StatusBar and TreeView status surfaces | hud, status display |
/cancel |
Cancel active execution modes | cancel, stop, abort |
/status |
Show current workflow status and active agents | status, what's running |
OMG includes a TypeScript MCP (Model Context Protocol) server that provides persistent state management for workflows. Registered via .vscode/mcp.json, it exposes tools for:
| Tool Group | Tools | Purpose |
|---|---|---|
| State | omg_read_state, omg_write_state, omg_clear_state, omg_list_active |
Workflow state CRUD and active mode listing |
| PRD | omg_create_prd, omg_read_prd, omg_update_story, omg_verify_story |
PRD creation, story tracking, and verification |
| Workflow | omg_check_completion, omg_next_phase, omg_get_phase_info |
Phase transitions, completion checks, phase inspection |
| Memory | omg_read_memory, omg_write_memory, omg_delete_memory |
Project-scoped knowledge persistence |
| Model Router | omg_select_model |
Model recommendation based on task complexity |
| Ultragoal | omg_ultragoal_create, omg_ultragoal_status, omg_ultragoal_checkpoint, omg_ultragoal_complete |
Durable goal tracking with fail-closed checkpoint evidence |
State is stored under .omg/ in the workspace:
.omg/
├── state/ # Workflow state files per mode
├── ultragoal/ # Durable ultragoal ledger and checkpoints
├── plans/ # Work plans for execution
├── prd.json # Product Requirements Document
└── project-memory.json # Project-scoped knowledge store
OMG includes pre/post tool-use hooks (.github/hooks/) that act as safety nets:
Pre-tool-use guards:
- Blocks modifications to
node_modules/ - Prevents direct
.envfile edits - Stops deletion of critical config files (
package.json,tsconfig.json,.gitignore) - Prevents
git push --forceand destructive git operations
Post-tool-use tracking:
- Logs tool usage for debugging (
OMG_DEBUG=1) - Tracks file modifications for omg-autopilot phase awareness
- Monitors test results for ultraqa detection
All numbers are derived from the actual
oh-my-githubcopilotgit history, test suite, andnpm auditresults. No synthetic data.
| Metric | Value |
|---|---|
| Total codebase | 25,964 lines |
| Development span | 12 days (Apr 6–17, 2026) |
| Total commits | 33 |
| Agents | 28 (20 core + 8 language reviewers) |
| Skills | 37 |
| MCP tools | 23 |
| Metric | v1.0 (initial) | v1.3.0 (after OMG pipeline) |
|---|---|---|
| Test pass rate | N/A | 46 / 46 (100%) |
| TypeScript errors | Not checked | 0 |
| Known CVEs | 7 (2 prod + 5 dev) | 0 |
| Pre-hook safety guards | 0 | 6 |
| Post-hook tracking features | 0 | 8 |
| Metric | Value |
|---|---|
| Files changed | 60 |
| Lines added | 5,844 |
| New agents | 8 language reviewer agents |
| New skills | 4 (/tdd, /security-scan, /coding-standards, /skill-stocktake) |
| Defects caught pre-merge | Shell injection in hooks + 7 CVEs |
| Decisions reviewed | Passed | Rejected by @critic |
Rejection rate |
|---|---|---|---|
| 9 | 7 | 2 | 22% |
22% of design decisions were revised at planning stage — before any code was written.
| Category | Found | Fixed |
|---|---|---|
| Hardcoded secrets | 0 | — |
| Production CVEs | 2 moderate | ✅ |
| Dev CVEs | 5 moderate | ✅ |
Shell injection (hook FILE_PATH) |
1 | ✅ |
.env missing from .gitignore |
1 | ✅ |
| Total vulnerabilities post-fix | 0 |
| Guard | Blocked Operation |
|---|---|
node_modules write protection |
Edit/create inside node_modules/ |
.env secret protection |
Direct .env file modification |
| Critical config deletion block | Delete package.json, tsconfig.json, .gitignore |
| Force push prevention | git push --force |
| Hard reset prevention | git reset --hard, git clean -fd |
| Path traversal sanitization | ../ and metacharacters in FILE_PATH |
oh-my-githubcopilot/
├── .github/
│ ├── copilot-instructions.md # Root orchestration instructions
│ ├── agents/ # 28 specialized agent definitions (20 core + 8 language reviewers)
│ ├── skills/ # 37 workflow skill routines
│ ├── hooks/ # Pre/post tool-use safety guards
│ └── prompts/ # Quick-fix, quick-plan, quick-review templates
├── mcp-server/ # TypeScript MCP server
│ └── src/
│ ├── index.ts # Server entry point
│ ├── state-tools.ts # Workflow state management
│ ├── prd-tools.ts # PRD and story tracking
│ ├── workflow-tools.ts # Phase transitions and completion checks
│ ├── memory-tools.ts # Project memory persistence
│ └── model-router.ts # Task complexity-based model routing
├── .vscode/mcp.json # MCP server registration for VS Code
└── .omg/ # Runtime state directory (gitignored)
- Instructions (
.github/copilot-instructions.md) define the orchestration rules, delegation logic, and agent catalog - Agents (
.github/agents/*.agent.md) are specialized personas with scoped tool access and model preferences - Skills (
.github/skills/*/SKILL.md) are workflow routines loaded on-demand via keyword triggers or slash commands - MCP Server provides persistent state, PRD tracking, and project memory via the Model Context Protocol
- Hooks guard against destructive operations and track execution for workflow awareness
OMG uses structured git trailers to preserve decision context:
fix(auth): prevent silent session drops during long-running ops
Auth service returns inconsistent status codes on token expiry,
so the interceptor catches all 4xx and triggers inline refresh.
Constraint: Auth service does not support token introspection
Rejected: Extend token TTL to 24h | security policy violation
Confidence: high
Scope-risk: narrow
Available trailers: Constraint, Rejected, Directive, Confidence, Scope-risk, Not-tested
| Feature | OMC (oh-my-claudecode) | OMG (oh-my-githubcopilot) |
|---|---|---|
| Target Platform | Claude Code CLI | GitHub Copilot (VS Code) |
| Installation | npm package / plugin marketplace | Clone + build MCP server |
| Agent Count | 19+ (with tier variants) | 28 agents (20 core + 8 language reviewers) |
| Skills | 10+ workflow skills | 37 skills with keyword triggers |
| State Management | .omc/ directory |
.omg/ via MCP server |
| Multi-model | Codex/Gemini via tmux CLI | ccg skill (advisory) |
| Configuration | ~/.claude/settings.json |
.github/ + .vscode/mcp.json |
| Tool Safety | Plugin-level hooks | Pre/post shell hooks |
| HUD / Statusline | Built-in HUD | VS Code native |
- VS Code with GitHub Copilot extension
- GitHub Copilot Chat with agent mode enabled
- Node.js 18+ (for the MCP server)
Ports high-value OMC workflows into OMG with VS Code/Copilot-native behavior. This release expands OMG from 24 to 37 skills and adds MCP-backed durable goal tracking.
- New MCP-backed
ultragoalworkflow:omg_ultragoal_create,omg_ultragoal_status,omg_ultragoal_checkpoint, andomg_ultragoal_completestore fail-closed durable goal artifacts under.omg/ultragoal/. - 13 adapted OMC-inspired skills:
visual-verdict,release,deepinit,skillify,writer-memory,configure-notifications,ultragoal,wiki,autoresearch,hud,external-context,sciomc, andask. - Updated routing and health checks: Copilot instructions now include new slash commands, keyword triggers, ultragoal completion rules, hook gates, and skill thresholds (
CORE=24,EXTENDED=37). - Template sync: Extension templates now include the new skills,
ultragoalMCP server source, tests, and updated README content. - Verification: MCP build/test passed (23 tests), VS Code extension unit tests passed (29 tests), VSIX package verified under 20MB.
- VSIX:
oh-my-githubcopilot-1.4.4.vsix.
Closes the v1.4.0 bridge into a full OMG ↔ OMC round-trip. v1.4.0 shipped one-way omc → omg import; v1.4.3 adds the reverse omg → omc export so work done in GitHub Copilot can continue in Claude Code.
- New module
mcp-server/src/bridge/omg-exporter.ts: composes (does not blindly copy) —prd.jsonandproject-memory.jsonare file-copies;active_modes[]from the OMG checkpoint is decomposed into per-mode.omc/state/{mode}-state.jsonfiles (the inverse of v1.4.0's omc-importer composition);.omc/state/session-checkpoint.jsonis composed with translated provenance. - New shared helper
mcp-server/src/bridge/conflict-utils.ts:shouldSkipForConflict()uses embeddedtimestampfield as primary conflict primitive forsession-checkpoint.json(mtime fallback elsewhere — addresses git-checkout-touches-mtime brittleness);rotateBackup()keeps last N=3 timestamped.previous.{ISO}.jsonsnapshots, applied symmetrically on both import and export sides. - New MCP tool
omg_export_external_session({target: "omc", force?})with atomicity contract:.omg/state/last-export-token.jsonis the LAST side-effect — partial-write failures leave no token, so subsequent operations correctly treat the transaction as uncommitted. - Round-trip safety via dual-write loop guard: importer reads either
.omg/state/last-export-token.jsonOR the destination checkpoint'ssource_origin === "bridged-from-omg"— guard survives wipe of either side. No time-based TTL; token invalidates only when destination'ssource_originflips to"native"orsource_session_idchanges. - Provenance schema extended (backward-compatible): new
source_originfield ("native" | "bridged-from-omc" | "bridged-from-claude-code" | "bridged-from-omg") is the authoritative provenance signal; legacysource_tool: "copilot"is retained. Newworkspace_rootfield powers a project-identity guard that aborts before any mutation when source checkpoint's workspace path doesn't match the current workspace. /push-omcskill (6 steps): detect → compare → confirm → export → summarize → continue. Trigger keywords:/push-omc,"push omc","omc 푸시","omc로 보내기","sync to omc","export to omc". Mirror atvscode-omg/resources/templates/skills/push-omc/SKILL.md.- VS Code command
omg.pushExternal("OMG: Push to External Session (OMC)"): user-initiated only — no activation-time push notification (push remains a deliberate user action). omg_compare_checkpointsreports both directions: now returnsomg_newer_than_externalalongsidenewer_than_omgso/push-omcand/resume-claudeshare the same comparison surface.- Tests grow 8 → 18+ across 6 zones (Unit, Provenance, Round-trip guard, Composition, Conflict-path, Project-identity).
- Deferred to v1.5.x:
omg → claude-codedirection. Would require synthesizing validtool_use/tool_resultUUID chains compatible with Claude Code's resume-by-id contract; tracking: TBD. - VSIX:
oh-my-githubcopilot-1.4.3.vsix.
Made Opus 4.7 routing tier-safe by adding Claude Opus 4.6 (copilot) as an inline fallback.
- The 7 deep-reasoning agents (
analyst,architect,code-reviewer,critic,omg-coordinator,planner,security-reviewer) now use a YAML model array:["Claude Opus 4.7 (copilot)", "Claude Opus 4.6 (copilot)"]. - Why: VS Code Copilot blocks subagent calls that exceed the active chat model's cost-tier ceiling. With a 7.5x base model (e.g.
GPT-5.5), anOpus 4.7-only request was hard-failing instead of degrading. - New behavior: when the session ceiling allows it, Opus 4.7 is used. Otherwise the agent transparently falls back to Opus 4.6 instead of failing.
- No change to GPT-5.5 (15 agents) or Sonnet 4.6 (6 agents) routing.
- VSIX:
oh-my-githubcopilot-1.4.2.vsix.
Mapped all 28 agents to task-fit Copilot model preferences using qualified model names
- Updated agent
model:frontmatter to use official qualified model-name strings such asGPT-5.5 (copilot)andClaude Sonnet 4.6 (copilot). - Kept high-stakes reasoning and approval agents on Claude Opus 4.7; routed debugging, verification, research, testing, design, and specialist review agents to GPT-5.5; routed execution, exploration, documentation, QA driving, git, and simplification workflows to Claude Sonnet 4.6.
- Updated
omg_select_modelrecommendations to use the same allowed model set and role-based agent overrides, removing stalegpt-4.1,gpt-4.1-mini, and dot-format Claude model recommendations. - Model smoke tests:
GPT-5.5 (copilot)invoked successfully;Claude Opus 4.7 (copilot)is recognized by Copilot but was blocked in this environment by the current cost-tier limit, not by a stale model string.
| Agent | Model |
|---|---|
analyst |
Claude Opus 4.7 (copilot) |
architect |
Claude Opus 4.7 (copilot) |
code-reviewer |
Claude Opus 4.7 (copilot) |
critic |
Claude Opus 4.7 (copilot) |
omg-coordinator |
Claude Opus 4.7 (copilot) |
planner |
Claude Opus 4.7 (copilot) |
security-reviewer |
Claude Opus 4.7 (copilot) |
csharp-reviewer |
GPT-5.5 (copilot) |
database-reviewer |
GPT-5.5 (copilot) |
debugger |
GPT-5.5 (copilot) |
designer |
GPT-5.5 (copilot) |
document-specialist |
GPT-5.5 (copilot) |
go-reviewer |
GPT-5.5 (copilot) |
java-reviewer |
GPT-5.5 (copilot) |
python-reviewer |
GPT-5.5 (copilot) |
rust-reviewer |
GPT-5.5 (copilot) |
scientist |
GPT-5.5 (copilot) |
swift-reviewer |
GPT-5.5 (copilot) |
test-engineer |
GPT-5.5 (copilot) |
tracer |
GPT-5.5 (copilot) |
typescript-reviewer |
GPT-5.5 (copilot) |
verifier |
GPT-5.5 (copilot) |
code-simplifier |
Claude Sonnet 4.6 (copilot) |
executor |
Claude Sonnet 4.6 (copilot) |
explore |
Claude Sonnet 4.6 (copilot) |
git-master |
Claude Sonnet 4.6 (copilot) |
qa-tester |
Claude Sonnet 4.6 (copilot) |
writer |
Claude Sonnet 4.6 (copilot) |
Opus 4.6 remains an explicit fallback alternative in the model router rather than the default for any agent.
One-directional bridge: resume interrupted Claude Code or OMC sessions in GitHub Copilot
- Claude Code JSONL importer: Parses
~/.claude/projects/session logs to extract modified files, last user prompt, and last assistant response. Supports Write/Edit/MultiEdit tool_use block detection. - OMC importer: Maps
.omc/state directory (PRD, workflow state, checkpoint, project memory) to.omg/equivalents with mtime-based conflict resolution. - 3 new MCP tools:
omg_detect_external_session(read-only detection),omg_import_external_session(performs import with backup),omg_compare_checkpoints(timestamp comparison). - VS Code auto-detection: On activation, detects external sessions and shows notification ("Resume / Ignore / Always ignore"). Skips if OMG checkpoint is fresh (< 30 min).
/resume-claudeskill: 6-step workflow — detect → compare → confirm → import → summarize → continue. Keyword triggers: "resume claude", "resume work", "continue session".- Security: Imported checkpoint files set to
chmod 0600. Existing checkpoints backed up to.previous.jsonbefore import. - Checkpoint schema extension: 4 new optional fields —
source_tool,source_session_id,imported_at,imported_summary(backward-compatible).
Dual-compatibility for VS Code agent mode and the standalone Copilot CLI (Issue #4)
- Agent frontmatter normalization:
model:field changed from array to string format across all 28 agents. CLI tool equivalents (read,edit,shell,create,delete) added totools:lists. - Hook dual-mode input: Pre/post tool-use hooks now accept both VS Code environment variables and CLI stdin JSON. Tool name normalization maps CLI names to VS Code equivalents before guard logic.
hooks.jsonregistration: New CLI hook wrapper forpreToolUse/postToolUsediscovery.- Adopt script
--target-env: New--target-env vscode|cli|bothflag (default:both). CLI mode generates.copilot/mcp-config.jsonand skips.vscode/mcp.json.--global-mcpinstalls MCP config at~/.copilot/. - Skill CLI fallback: All 5 interactive skills (
deep-interview,omg-autopilot,ralplan,plan,self-improve) updated with CLI fallback — numbered markdown options whenvscode_askQuestionsis unavailable. - Documentation: CLI badge, "Option C: Copilot CLI" quick start,
--target-envusage added to all READMEs.
Three community-requested improvements (Issues #5, #6, #7)
- Windows PowerShell support (Fixes #5): Added
.ps1equivalents for all shell scripts —pre-tool-use.ps1,post-tool-use.ps1,omg-adopt.ps1. Hooks bundled in extension templates. READMEs updated with PowerShell examples. - Non-destructive copilot-instructions.md (Fixes #6):
initWorkspacenow appends OMG instructions to existing files instead of overwriting. Marker-based section detection supports re-init updates. - MIT License clarification (Fixes #7): Added root
LICENSEfile for GitHub API detection. Removed "All rights reserved" contradictions from all READMEs.
Upgraded all agent model references from Claude Opus 4.6 to Claude Opus 4.7
- Updated
model: [claude-opus-4-6]→model: [claude-opus-4-7]across all 8 Opus-routed agents. - Applied to both active agents (
.github/agents/) and extension templates (vscode-omg/resources/templates/agents/). - Affected agents: @architect, @code-reviewer, @planner, @security-reviewer, @analyst, @omg-coordinator, @code-simplifier, @critic.
- Verified: MCP server build + tests (18/18), vscode-omg build + tests (28/28), TypeScript type check — all passing.
State path consistency update across source + templates
- Renamed remaining state path references from
.omc/to.omg/across skills, agents, MCP templates, and extension templates. - Updated hook state variable naming from
OMC_STATE_DIRtoOMG_STATE_DIRwhere applicable. - Verified MCP server build and tests after migration (18/18 passing).
- Kept intentional
.omc/mentions only in OMC-vs-OMG comparison table rows.
Major reliability upgrade for long-running workflows
- Added Tier-2 memory enhancements to OMG MCP:
omg_search_memoryfor keyword lookup across keys, values, categories, and tagstagssupport inomg_write_memory- Backward-safe reads for existing memory entries
- Added Tier-3 context-pressure protocol:
omg_checkpoint,omg_restore_checkpoint,omg_context_status- Auto checkpoint advisory from pre/post hooks based on accumulated tool I/O bytes
- Configurable threshold via
OMG_CONTEXT_THRESHOLD(default: 400KB)
- Hardened hook safety and lifecycle behavior:
- Fixed checkpoint loop by resetting context counter after checkpoint
- Improved force-push detection (
--force,-f) while allowing--force-with-lease - Expanded destructive git guard coverage
- Added startup/session recovery guidance in
copilot-instructions.md
OMC-parity: mid-workflow user decisions via vscode_askQuestions
Five core skills now fire structured multiple-choice prompts at decision gates, mirroring OMC's gateway-level interrupt hooks — natively inside VS Code Copilot.
| Skill | Hook Points |
|---|---|
/deep-interview |
Every interview round (with ambiguity %) · Spec approval · Execution bridge |
/plan |
Interview questions · Readiness gate · Trade-off selection · Critic rejection · Plan approval |
/ralplan |
Options selection · Architect concerns · Critic rejection · Final approval |
/self-improve |
Repo target · Trust confirmation · Goal interview · Harness rules (setup only — loop runs autonomously) |
/omg-autopilot |
Vague input redirect · Spec confirmation · QA stuck recovery · Validation rejection |
- At each decision gate the skill calls
vscode_askQuestionswith 3–5 labelled options, arecommendeddefault, andallowFreeformInput: true - User selects an option (or types a custom answer) — the workflow continues based on the response
- All hook calls use a unique
header(e.g."interview-round-3","ralplan-approval") for traceability - Global hook protocol documented in
copilot-instructions.md → Interactive Hook System
Major upgrade: best-of-ECC (Everything Claude Code) features merged into OMG
Eight new language-specialist review agents, each with 13–21 embedded style rules and a canonical coding standards reference:
- @typescript-reviewer — strict mode, no-any, exhaustive discriminated unions
- @python-reviewer — PEP 8, type hints, idiomatic patterns
- @rust-reviewer — ownership, borrow checker, unsafe justification
- @go-reviewer — idiomatic Go, goroutine safety, explicit error handling
- @java-reviewer — SOLID, Spring patterns, null safety
- @csharp-reviewer — nullable analysis, async/await, C# idioms
- @swift-reviewer — Swift concurrency, memory safety, SwiftUI patterns
- @database-reviewer — query performance, parameterization, schema design
/tdd— Full TDD enforcement: Red-Green-Refactor cycle, framework quick-ref (Jest/pytest/Cargo/Go test), gate table with fail conditions/security-scan— Rapid pre-commit security sweep: secrets patterns (sk-/ghp_/AKIA), CVE audit, input validation, auth checks/coding-standards— Canonical cross-language coding standards (naming, functions, errors, anti-patterns, SOLID). Cited by all reviewer agents./skill-stocktake— Audit skill inventory for frontmatter validity, template sync, stubs, and coverage gaps
| Agent | What Was Added |
|---|---|
| @debugger | 7-language build resolution guide (Node/TS, Python, Rust, Go, Java, C#, Swift) |
| @security-reviewer | Secrets detection regex patterns, JWT/session/crypto rules, OWASP annotations |
| @qa-tester | Playwright Page Object Model pattern, E2E classification table |
| @writer | CODEMAP generation template + auto-update workflow |
| @code-reviewer | Canonical D9 coding standards table embedded |
| @test-engineer | Framework detection table, coverage gap protocol, flaky test root causes |
| @code-simplifier | Complexity metrics table, simplification patterns, stability exceptions |
/remember— Quality gate added: 3-question filter (actionable? durable? unique?) before storing to project memory
- Tiered agent/skill thresholds in
convention.ts: warn <20/18, info <28/22, silent ≥28/22 (backward compatible) - Tree-view category grouping: agents auto-split into "Core Agents" + "Language Reviewers" when count >20
- Post-tool-use hook (
OMG_LINT_ON_EDIT=1): opt-in TypeScript typecheck + ESLint on every file edit withFILE_PATHsanitization guard - Security patches: hono/node-server CVEs fixed, vitest upgraded to 4.1.4,
.env*added to.gitignore
Bug Fix: Skill name collision with VS Code built-in Autopilot
- Renamed
/autopilotskill to/omg-autopilotto avoid collision with VS Code's built-in "Autopilot (Preview)" permission level - Fixed invalid YAML frontmatter in all SKILL.md files: removed unsupported
allowed-toolsfield, renamedhinttoargument-hintper VS Code spec - Root cause: The skill name
autopilottriggered VS Code's internal permission mode switch instead of loading the OMG skill instructions - Scope: Updated skill directories, MCP server code, agent definitions, cross-references, tests, and all documentation
MIT
Copyright (c) 2026 jmstar85. Licensed under the MIT License.
Inspired by: oh-my-claudecode by Yeachan Heo
The power of multi-agent orchestration, now for GitHub Copilot.