CLI Reference

The metaharness CLI covers five workflows:

1. scaffold a project
2. scaffold a domain onboarding pack
3. run or probe a backend
4. inspect and export results
5. execute repeated experiment matrices

Show help:

uv run metaharness --help

Long commands below are wrapped with \ so they stay readable and copy cleanly.

Many reporting commands support:

• plain text output by default
• --json for machine-readable output
• --tsv for spreadsheet-friendly export where supported

scaffold

Create a new coding-tool project:

uv run metaharness scaffold \
  coding-tool \
  ./my-coding-tool-optimizer

Profiles:

• standard
• local-oss-smoke
• local-oss-medium
• skills

Standard

Best default for a new project.

uv run metaharness scaffold coding-tool ./my-project

Fast Local Smoke

Smaller harness aimed at local OSS smoke runs.

uv run metaharness scaffold \
  coding-tool \
  ./my-local-oss-smoke \
  --profile local-oss-smoke

Medium Local OSS

Restores bootstrap and test scripts while staying lighter than the full scaffold.

uv run metaharness scaffold \
  coding-tool \
  ./my-local-oss-medium \
  --profile local-oss-medium

Agent Skills

Seeds a shared skill under .agents/skills plus Claude/Gemini skill directories, with leakage_gate and write_scope_mode: "single-class".

uv run metaharness scaffold \
  coding-tool \
  ./my-skills-optimizer \
  --profile skills

onboard

Create an official-style onboarding pack for a new domain:

uv run metaharness onboard ./my-domain-onboarding

This command writes:

• ONBOARDING.md with required questions and guardrails
• domain_spec.md with a concrete template for domain, harness, evaluation, and artifact design

Use this before implementing a new adapter so search/test splits, metrics, budget, and leakage risks are defined up front.

run

Run one optimization project:

uv run metaharness run \
  ./my-coding-tool-optimizer \
  --backend fake \
  --budget 1

Use this when you want a single benchmark or project run and care about the winning candidate, not aggregate trial statistics.

Important options:

• --backend
• --budget
• --run-name
• --hosted
• --oss
• --local-provider
• --model
• --proposal-timeout
• --search-mode
• --proposal-batch-size
• --selection-policy
• --trace-evidence

--backend accepts built-ins (fake, codex, gemini) and any plugin backend name defined in backend_plugins. Use --trace-evidence path/to/trace_evidence.md to inject a HALO/RLM trace diagnosis report into each candidate proposal. The file is copied to .metaharness/evidence/trace_evidence.md inside the candidate workspace and embedded in the backend prompt. Proposers are instructed to write .metaharness/change_manifest.json; MetaHarness archives it with the proposal and can attribute predicted fixes/regressions when evaluator metadata includes task_results.

Fake Backend

Best for smoke checks and development.

uv run metaharness run \
  ./my-coding-tool-optimizer \
  --backend fake \
  --budget 1

Hosted Codex

Best current path for real benchmark quality.

uv run metaharness run \
  ./my-coding-tool-optimizer \
  --backend codex \
  --hosted \
  --budget 1

Local Codex Over Ollama

Local-only path for OSS model runs.

uv run metaharness run \
  ./my-coding-tool-optimizer \
  --backend codex \
  --oss \
  --local-provider ollama \
  --model gpt-oss:20b \
  --proposal-timeout 240 \
  --budget 1

Gemini CLI

Use Gemini as an experimental proposer backend.

uv run metaharness run \
  ./my-coding-tool-optimizer \
  --backend gemini \
  --model gemini-2.5-pro \
  --proposal-timeout 180 \
  --budget 1

Trace-Grounded Run

Use a HALO/RLM trace evidence report to ground harness edits in observed failures.

uv run metaharness run \
  ./my-coding-tool-optimizer \
  --backend codex \
  --hosted \
  --trace-evidence ./trace_evidence.md \
  --budget 1

Plugin Backend

Use a custom adapter from backend_plugins, for example cursor.

uv run metaharness run \
  ./my-coding-tool-optimizer \
  --backend cursor \
  --budget 1

experiment

Run a benchmark x backend x budget x trial matrix:

uv run metaharness experiment \
  ./examples/python_fixture_benchmark \
  --backend fake \
  --trials 3

Use this when you want repeatable benchmark results instead of one-off runs.

Saved Config

The most reusable path for teams.

uv run metaharness experiment \
  --config ./examples/experiment_configs/fake-benchmarks.json

Multiple Budgets

Compare how much improvement you get from a larger search budget.

uv run metaharness experiment \
  ./examples/python_fixture_benchmark \
  --backend fake \
  --budget 1 \
  --budget 2 \
  --trials 2

TSV Export

Send aggregate results straight to a spreadsheet or notebook.

uv run metaharness experiment \
  ./examples/python_fixture_benchmark \
  --backend fake \
  --trials 3 \
  --tsv

This command writes:

• experiment.json
• trials.json
• aggregates.json
• trials.tsv
• aggregates.tsv

Config files can contain:

• project_dirs
• backends
• budgets
• trial_count
• models
• results_dir
• backend_overrides

If a config file is provided, relative paths are resolved from the config file location. CLI flags override the corresponding config values.

smoke codex

Probe the Codex path before spending model calls:

uv run metaharness smoke codex ./my-coding-tool-optimizer --probe-only

Probe the local Ollama path:

uv run metaharness smoke codex \
  ./my-coding-tool-optimizer \
  --probe-only \
  --oss \
  --local-provider ollama \
  --model gpt-oss:20b

Use this when you want to verify the environment, provider, and model path before running a benchmark.

smoke gemini

Probe the Gemini CLI path before spending model calls:

uv run metaharness smoke gemini ./my-coding-tool-optimizer --probe-only

Run one Gemini-backed smoke iteration:

uv run metaharness smoke gemini \
  ./my-coding-tool-optimizer \
  --budget 1 \
  --model gemini-2.5-pro

inspect

Inspect one completed run:

uv run metaharness inspect \
  ./examples/python_fixture_benchmark/runs/hosted-codex-20260401

This is the quickest human-readable view of:

• candidate outcomes
• validity
• proposal application
• scope violations
• objective scores

ledger

Export the per-candidate ledger for one run:

uv run metaharness ledger \
  ./examples/python_fixture_benchmark/runs/hosted-codex-20260401

TSV export:

uv run metaharness ledger \
  ./examples/python_fixture_benchmark/runs/hosted-codex-20260401 \
  --tsv

Use this when you want one row per candidate with outcomes, changed-file counts, summaries, scope violations, change-manifest status, changed component labels, and attribution verdict counts.

summarize

Summarize all runs in a project:

uv run metaharness summarize \
  ./examples/python_fixture_benchmark

TSV export:

uv run metaharness summarize \
  ./examples/python_fixture_benchmark \
  --tsv

Use this when you want a project-wide view of scores, durations, and outcome counts.

compare

Compare specific run directories:

uv run metaharness compare \
  ./examples/python_fixture_benchmark/runs/hosted-codex-20260401 \
  ./examples/python_fixture_benchmark/runs/ollama-20b-20260401 \
  ./examples/python_fixture_benchmark/runs/ollama-120b-20260401

TSV export:

uv run metaharness compare \
  ./examples/python_fixture_benchmark/runs/hosted-codex-20260401 \
  ./examples/python_fixture_benchmark/runs/ollama-120b-20260401 \
  --tsv

Use this when you want an explicit side-by-side comparison between selected runs rather than every run in a project.

Output Files To Know

The most useful stored artifacts are usually:

• run_config.json
• indexes/leaderboard.json
• manifest.json
• proposal/result.json
• proposal/workspace.diff
• validation/result.json
• evaluation/result.json