AHO — Agent Harness Orchestrator

Use case

AHO picks up an issue from the tracker and runs a configured plan: a single coder (single), an escalating cascade from cheap to premium backend (cascade), parallel specialists in their own git worktrees (lead-specialists), or coder + cross-provider reviewer (single-with-review). Successful runs go through AI review, which either merges directly or spawns a fix issue as a follow-up task.

What's done so far

• MVP tag mvp-hello-aho-v0.1.0, schema v0.4.0
• Four plan patterns in production: single, cascade, lead-specialists, single-with-review
• Pattern-agnostic post-success finalize hook: auto-commit, ai_review merge into default branch, auto-spawn of fix issues on reviewer reject
• Fix-issue auto-dispatch with parent_issue_id, aho/<parent> branch reuse, capped via max_review_iterations
• Provider-agnostic backend abstraction (Claude, OpenAI, local models)
• Postgres audit trail per plan run, SSE stream + minimal operator dashboard

In progress

Cleaner cross-provider reviewer flow, more robust worktree lifecycle handling, telemetry on subagent run time and token cost, plus more anti-patterns as executable hooks.