oh-my-colab

Chain PLAN + BUILD + REVIEW automatically. Pauses at plan for human OK.

BUILD workflow: load plan → dispatch subagents → TDD enforcement → verify

Enter persistent compressed communication mode

External document intake — normalize user-provided specs into .ohc/doc-sources.md and a traced plan (continues `/ohc-plan`).

EXPLORE workflow: read codebase, populate PROJECT.md with stack, conventions, gotchas

PLAN workflow: interview → brainstorm → write task plan to .ohc/plans/

Activate ralph persistence mode — keep iterating until the task is done or max iterations reached

Cross-session research cache. Inspect, search, prune, or verify cached library docs that the researcher agent fetched via Context7/Brave/GitHub.

RETRO workflow: session diff → extract patterns → update memory files → draft skills

REVIEW workflow: spec compliance (pass 1) then code quality (pass 2)

Run interactive oh-my-colab onboarding: platforms, MCP, HUD

SHIP workflow: pre-merge check → changelog → PR description → clean worktrees

Manage skills: /ohc-skill list | promote <name> | draft <name>

Spawn N parallel agent workers that run an internal five-stage pipeline (plan → prd → exec → verify → fix → merge). Usage: /ohc-team N:provider "task".

Run tests and static quality checks (typecheck + lint) to verify task completion. Use after executor finishes a task. Reports exact pass/fail counts, hands back to executor on failure or to reviewer on pass.

Transform ambiguous goals into precise executable task lists. Use when a task is >30min, has multiple tradeoffs, or has unclear requirements. Writes a plan to .ohc/plans/ and awaits user confirmation before handing to executor.

Pre-execution gap analysis and risk assessment agent. Reads plan files, PROJECT.md, and known gotchas to produce a risk surface report. Run before starting a major BUILD workflow.

Fetch accurate current documentation and API signatures. Use before writing any code that calls external libraries. Cache-first — check .ohc/research/ before hitting Context7/Brave/GitHub. Persists every fetch for cross-session reuse.

Design systems and write Architecture Decision Records. Use when the task needs structural decisions, ADRs, or component/data-flow diagrams. Hands to planner after user confirms design.

Enable seamless handoffs between developers and sessions. Use when handing off work, picking up someone else's work, or needing a session context summary. Writes updated notepad.md and commits it.

Find root cause using hypothesis-driven investigation. Use when something is broken, failing, or behaving unexpectedly. Never guesses — always hypothesizes then tests. Hands to executor with confirmed root cause.

Write code following the active plan. Use after the PLAN gate clears, for any implement/build/code task in .ohc/plans/. Hands off to verifier when done.

Read large files or directories into subagent context and return structured summaries with exact line references. Use to prevent blowing out the main session context window when searching for specific implementations.

Two-pass code review — spec compliance then code quality. Use after verifier passes. Blocks shipping on critical issues, hands to executor for fixes or to writer for changelog.

Write human-facing text artifacts — changelogs, PR descriptions, README updates. Use after reviewer passes. Terminal agent — does not hand off further.

Chain PLAN + BUILD + REVIEW automatically. Pauses at plan for human OK.

Structured ideation before committing to a design. Socratic questioning to surface hidden requirements and constraints. Produces 2-3 concrete approaches with explicit tradeoffs. Runs at the start of the PLAN workflow.

Ultra-compressed communication mode that removes filler while preserving technical accuracy. Use when the user explicitly asks for "caveman mode", "compressed mode", or to stay terse/compressed until told to stop.

Enforces use of Context7 MCP for all external library documentation. Prevents API hallucination by requiring real-time doc fetching before any external library call. Auto-triggered when external library usage is detected in code context.

Normalize user-provided specs in any format into .ohc/doc-sources.md plus a traced plan. Triggers: from my docs, existing spec, external spec, /ohc-document-intake.

Systematically understand an unfamiliar codebase or feature area. Outputs a populated .ohc/PROJECT.md with stack, architecture, conventions, entry points, and gotchas. Run at project start or when entering new feature territory.

Pre-merge checklist, changelog entry, PR description, and worktree cleanup. Ensures no debug code, all tests pass, documentation updated. Triggers on keyword "ship" or when SHIP workflow is active.

Anti-overengineering rules. Use when writing, reviewing, or refactoring code. Enforces surgical changes, explicit assumptions, minimal scope, and verifiable success criteria. Extended from Karpathy guidelines for team AI workflows.

PRD-loop persistence mode: iterate BUILD until verifier passes or max blocked attempts. Triggers ralph keep going keywords and /ohc-ralph. Deactivate via stop ralph.

Two-pass code review: spec compliance first, code quality second. Produces structured report with ✓/⚠/✗ items. Blocking issues must be fixed before ship. Triggers on keyword "review" or when REVIEW workflow is active.

Post-session learning loop. Compares plan vs actual, extracts novel patterns, updates USER.md and PROJECT.md, promotes draft skills. Run after any session >30 minutes. Triggers on keyword "retro" or auto-suggested by on-stop hook.

Dispatch multiple specialized subagents to work in parallel on independent tasks using git worktrees to prevent conflicts. Use when a plan has 3+ independent tasks or total work >2h. Triggers when BUILD workflow runs a multi-task plan.

Hypothesis-driven debugging loop. Forces explicit hypothesis formation before any fix attempt. Prevents guess-and-check debugging. Use when investigating or fixing an observed bug, failing test, runtime error, broken behavior, or performance regression.

RED-GREEN-REFACTOR discipline. Write failing test first, minimum code to pass, then refactor. Enforces verifier check before claiming any task done. Triggers on keyword "tdd" or when BUILD workflow is active.

Transform a goal into a precise executable task list saved to .ohc/plans/. Run before any task >30 minutes. Includes deep interview, approach selection, task decomposition with success criteria. Triggers on keyword "plan".

Meta-skill for creating new skills in oh-my-colab format. Use when RETRO identifies a pattern worth formalizing. Ensures correct folder structure, 200-line discipline, proper frontmatter, and token-efficient design.

12 hooks across 12 events

GitHub MCP — repos, PRs, issues, CI

Linear — project management

Context7 — live library docs (free; key optional)

Firecrawl — web scraping

Playwright — browser automation, e2e testing

Brave Search — web search