coding-agent-orchestration-harness

Orchestrates work in a workspace: Workspace orchestrator. Clarifies requirements, gathers context via a Research subagent, creates and gets approval for plans, delegates execution to subagents, integrates results, invokes a review gate, routes git operations through git-workflow, routes skill governance through skills-maintenance, and updates repo rule files.

Research-only subagent. Gathers workspace context and returns plan-fill inputs to the parent Orchestrator. Does not edit files, does not write plan files, and does not interact with the user. May run bounded UI exploration using a selected browser automation provider such as playwright-cli (local URLs only) when it materially improves planning and validation design.

Review-only subagent. Reviews changes after an implementation wave/phase completes. Verifies objectives/acceptance criteria, best practices, and validation evidence. May run bounded UI evidence checks using a selected browser automation provider such as playwright-cli (local URLs only) when UI/user flows are impacted. Does not edit files.

Execution subagent for a single atomic Task_X within owns scope. Implements changes, runs required validation, returns a strict YAML report (subagent-report-contract), and does not perform shared-state git mutations unless explicitly instructed by the Orchestrator for that task.

Trigger for non-trivial implementation, PR/code review, and bug fix/refactor work; routes validation depth and evidence expectations, including required checks, through progressive disclosure across architecture, stack, language/tech, validation, and security.

Standardizes safe Git workflow decisions for branch safety, logical commit chunking, commit hygiene, and explicit non-interactive or non-destructive defaults. Use when planning or executing commit-affecting work, checking whether a commit should proceed, deciding how to split changes into coherent commits, or routing shared-state Git mutations through the Orchestrator.

Self-improvement loop for Orchestrator-driven workflows. Use immediately after any user correction or behavior change request (workflow, validation, scope, tone, output format), after missed required gates (e.g., validation), or after review misses. Requires completing a post-correction micro-checklist before ending the turn.

Standard execution plan format for decomposing non-trivial work into Task_X tasks with owns/depends_on, explicit validation ownership, and explicit Task Waves (parallel dispatch sets). Use when asked to make a plan, break down tasks, outline implementation steps, execute multi-step work, or when non-trivial work requires a plan.

Automates browser interactions for web testing, form filling, screenshots, and data extraction. Use when the user needs to navigate websites, interact with web pages, fill forms, take screenshots, test web applications, or extract information from web pages.