Build AI-agent collaboration infrastructure for codebases

ECL Harness Engineer designs and creates Harness Engineering infrastructure so AI agents can work reliably in a codebase, built on the core philosophy that "intelligence without infrastructure is just a demo" - the Agent Harness is the operating system, the LLM is just the CPU, and the repository becomes the single source of truth: if an agent can't see it in context, it doesn't exist. It follows one unified workflow regardless of project state (empty, existing code, or existing harness): detect the gap between current state and target state, then fill it. The default target is a core ECL harness including lightweight auto-evolve threshold checking (closed changes are counted, a pending-evolution note generated at threshold, and improvements applied only through evidence, validation, scoring, and rollback) - advanced capabilities like eval datasets, execution traces, durable state, checkpoints, and long-term memory stay optional, added only when explicitly requested.

Use it when a repository needs AI-agent collaboration infrastructure (AGENTS.md, docs/ECL.md, docs/STATUS.md, harness change tracking, mechanical validation gates), when auditing an existing Agent Harness for missing ECL lifecycle docs/change templates/lint checks/environment contracts/CI integration, or when converting repeated agent workflow failures into repository-local documentation, tests, lint rules, or auto-evolution checks. Do not use it for ordinary business feature implementation unless the request is specifically about the repository harness itself - it does not replace product requirements, implementation planning, code review, or release approval.

A five-phase process: (1) Quick Detection + Intent Confirmation - a sub-5-minute scan for file/code counts, existing harness components (AGENTS.md, docs/ECL.md, linters, change scripts), and tech stack, classifying project state as Empty/Code Only/Partial Harness/Harness Present and ECL readiness as Missing/Partial/Ready; (2) Analysis of architecture, harness state, environment, and project identity; (3) Intake Review + Delta Synthesis, classifying small vs structured work and computing exactly what to create or update; (4) Creation/Update of docs, status handoff, linters, ECL/change scripts, environment config, and CI; (5) Verification + Handoff, running checks, attributing failures, updating STATUS.md, triggering auto-evolve checks. A baseline verification snapshot runs only pre-existing project commands (lint/typecheck/test/build per ecosystem) before harness work, recording pass/fail/missing so pre-existing project debt is never misattributed to the harness change itself. Scope tiers: Core harness (default - AGENTS.md, docs/ECL.md, docs/STATUS.md, ECL changes, lightweight auto-evolve, linters, environment contract, CI), Advanced harness (core plus explicitly requested eval/trace/state/checkpoints/memory/metrics), and Documentation only (AGENTS.md and docs with no linters/scripts/CI).

Uses a user-confirmation tool where available (request_user_input in Codex, or the platform equivalent) to confirm scope before making changes, falling back to recorded assumptions if none exists. Auto-evolve recommendations are guidance only - harness changes go through normal review, validation, and rollback discipline rather than being applied as autonomous policy changes.

Teams building or auditing the infrastructure that lets AI coding agents work reliably and consistently in a repository - not implementing product features, but making the repository itself legible and self-documenting for agent collaboration.