Work Session
Orchestrate AI-assisted development with configurable human control, using ohno for task management via MCP.
When NOT to Use
- Planning work — Use
planningto create epics/stories/tasks before starting sessions - Quick one-off tasks — Use
/quickfor simple tasks that don't need the full subagent pipeline - Revising a plan — Use
plan-revisionto modify scope, dependencies, or task hierarchy - Running a spike — Use
spikefor time-boxed investigation before committing to implementation
Key Principles
- Fresh context per task via subagent dispatch (no context degradation)
- Configurable checkpoint control: supervised, semi-auto, auto, or unattended
- Smart worktree isolation by task type (feature/bug → worktree, chore/docs → in-place)
- Hooks handle lifecycle automatically (sync, commit, tests)
- ohno MCP provides session continuity across conversations
- Evidence before completion claims (
verification-before-completion) - Root-cause-first bug handling (
systematic-debugging) - Explicit branch finish choice: merge, PR, keep, or discard (
finishing-branch) - Token-aware dispatch: stay inline for tiny work, use focused agents when isolation or parallelism is worth the cost
Quick Start Checklist
- Initialize ohno:
npx @stevestomp/ohno-cli init - Get session context: MCP
get_session_context() - Get next task: MCP
get_next_task() - Dispatch subagent for implementation
- Review results at checkpoints (based on mode)
Operating Modes
| Mode | Task Complete | Story Complete | Epic Complete |
|---|---|---|---|
| supervised | PAUSE | PAUSE | PAUSE |
| semi-auto | log | PAUSE | PAUSE |
| auto | skip | log | PAUSE |
| unattended | skip | skip | skip |
References
| Reference | Description |
|---|---|
| dispatch-preparation.md | Task extraction, brainstorm gate, design review gate, skill routing, template filling |
| review-pipeline.md | Two-stage review: spec compliance then code quality (+ design compliance) |
| dispatch-errors.md | Recovery when ohno, routing, or dispatch fails |
| checkpoint-types.md | PAUSE, REVIEW, NOTIFY checkpoint patterns |
| skill-routing.md | Task type to skill mapping |
| operating-modes.md | Supervised, semi-auto, auto, unattended details |
| worktree-management.md | Setup, completion, merge/PR workflows |
| parallel-execution.md | Parallel Execution: benefits, tradeoffs, dependency handling |
| hook-integration.md | Work loop with hooks, mode-specific behavior |
| ohno-integration.md | MCP tools and CLI commands reference |
| error-recovery.md | Build failures, blocked tasks |
| approval-policy.md | Runtime approval defaults for safe pokayokay automation |
| token-budgeting.md | Token/context budgeting rules for agents, skills, parallelism, and handoffs |
| anti-patterns.md | Common mistakes and fixes |
| bug-fix-pipeline.md | Agent pipeline for /fix --thorough and /hotfix commands |
| pre-flight-checks.md | Checks run before unattended/headless sessions |
Runtime Notes
- Claude Code: the coordinator dispatches yokay-brainstormer, yokay-design-reviewer, yokay-implementer, yokay-fixer, yokay-spec-reviewer, yokay-quality-reviewer, yokay-browser-verifier, and yokay-auditor via the Task tool with
subagent_type: "pokayokay:yokay-<name>", as described in dispatch-preparation.md. - Codex: there is no subagent dispatch. Run the pipeline inline in the current session — for each stage, read the corresponding
agents/yokay-<name>.mdand follow its Behavioral Defaults, Critical Rules, and Output Contract. Execute stages (design review (conditional) → implement → spec review → quality review) as separate, sequential passes; do not skip review stages because dispatch is unavailable. Parallel batch execution (parallel-execution.md) is Claude Code-only — process tasks sequentially on Codex.