QoderAI/better-harness

What it claims

Better Harness is a skill/plugin, plus a standalone better-harness CLI. It audits the workflow around a coding agent rather than the diff it produces. Its core rule is that configured assets can only show a mechanism exists. Assets include AGENTS.md, Skills, Hooks and tests. Only linked task evidence ("Task Episodes" from session transcripts) can show that a mechanism was used or improved an outcome. A passing check proves only that a fix was exercised. Improvement claims require a comparable later report, and the history view shows trends, explicitly not causal proof.

How to use it

There is no single entrypoint: install separately for each host, then start a new session so the host reloads its plugins.

Claude Code:

/plugin marketplace add QoderAI/better-harness
/plugin install better-harness@better-harness
claude plugin details better-harness@better-harness
/better-harness analyze this project's AI coding workflow and generate an evidence-backed report

Expect Skills (1) better-harness in the details output. Reports go to report.html, report.md and findings.json under .claude/better-harness in the repo. Ask for "inline" or "no-files" output to keep the result in chat. Local Claude sessions that match the workspace are read automatically.

Codex CLI (use the repo URL, not a raw marketplace.json; older plugin install/--source examples target a different CLI contract):

codex plugin marketplace add \
  'https://github.com/QoderAI/better-harness.git' \
  --ref main
codex plugin add better-harness@better-harness
$better-harness:better-harness analyze this project's AI coding workflow and generate an evidence-backed report

GitHub Copilot CLI (direct repo, URL and path installs are deprecated):

copilot plugin marketplace add QoderAI/better-harness
copilot plugin install better-harness@better-harness

Qwen Code: qwen extensions install QoderAI/better-harness.

Qoder Desktop has it built in, and Qoder CLI inherits it when Desktop is installed. Qoder and Cursor render host-native Canvas reports. The other hosts produce HTML, Markdown and findings.json.

Host-independent CLI commands. These inspect local installs without contacting a registry. plan emits install steps as typed argv data and does not execute them:

better-harness plugin status --host all
better-harness doctor --platform all
better-harness plugin plan install --host qwen --surface cli --scope user

To scan a repo only, without reading local sessions, run from a source checkout:

node scripts/better-harness.mjs report --no-sessions

Development needs Node.js >=22.20.0 <25.0.0 and npm >=10.9.3 <12.0.0 (npm ci && npm test && npm run pack:verify).

Limits

• Privacy: by default the tool reads your local agent session transcripts, e.g. Copilot's ~/.copilot/session-state/. Use --no-sessions if that matters.
• Cursor: not published to the Cursor marketplace. The tool reports its own install plan as "unavailable" because the --plugin-dir contract is unverified, so Cursor support is session-only.
• Copilot: no per-response token usage. VS Code Copilot Chat has no supported transcript, so only Copilot CLI sessions count.
• Other hosts: WorkBuddy has no managed plugin lifecycle. Pi lifecycle commands may be manual or unavailable. Grok, WorkBuddy and Kimi details exist only in the Host Adapter Matrix, not the README.
• Qoder Quest entry: marked Beta and requires Qoder 1.18.0+.
• Canvas preview: needs an installed Qoder runtime. The local preview server binds to 127.0.0.1 and is not an authenticated sharing service.
• Scoring: the page gives no validation of the scoring model, such as correlation with outcomes or a benchmark. The findings are structured judgments, not measured effects.