Harness

harness

Long-running application development harness for Claude Code, based on Anthropic's harness design for long-running apps.

The Problem

Complex builds take more context than a single Claude session can hold. Without structure, a fresh session has no memory of prior work — it starts over, repeats decisions, and misses progress already made.

The Solution

This plugin implements a two-agent harness with a generator-evaluator loop:

1. Initializer — runs once, decomposes the goal into a granular feature list, sets up progress tracking artifacts, and makes the first git commit.
2. Coding Agent — runs each session, reads the progress log and feature list to orient itself, implements the next batch of features one at a time, verifies each, and leaves clean artifacts for the next session.
3. Evaluator — runs independently after coding sessions to catch over-optimistic self-assessments, producing a structured quality report with prioritized rework items.

The primary handoff artifact is claude-progress.txt — a human-readable log that every fresh agent reads first to understand what was done and what comes next. Git history provides the secondary record.

Commands

Command	Description
/harness-init <goal>	Initialize the harness for a new build project
/harness-run [focus]	Run one coding session, picking up from the last
/harness-eval [focus]	Run the evaluator to independently assess build quality
/harness-status	Quick summary of features done, remaining, and blockers

Workflow

/harness-init "build a task management web app with React and a REST API"
    └─► feature_list.json (200 features, all failing)
        claude-progress.txt (session 0)
        init.sh
        git commit

/harness-run          # session 1 — implements F001–F012, commits each
/harness-eval         # evaluates F001–F012, writes HARNESS_EVAL_REPORT.md
/harness-run          # session 2 — fixes regressions, implements F013–F025
/harness-run          # session 3 — ...

Repeat /harness-run and /harness-eval until all features are passing.

Artifacts

File	Purpose
feature_list.json	Ground truth of project health — every feature with status and notes
claude-progress.txt	Session log — each agent's primary memory across context resets
init.sh	Environment bootstrap — run to restore dependencies from scratch
HARNESS_EVAL_REPORT.md	Latest evaluator report — prioritized rework list

Key Design Principles

• Context resets are safe — each session is self-contained; claude-progress.txt + git history carry all necessary state
• One feature per commit — granular commits make regression bisection easy
• Independent evaluation — the evaluator is deliberately separate from the coding agent to prevent self-serving assessment
• Features are the unit of progress — feature_list.json is the source of truth, not lines of code or time spent