@frontierharness/eval
Advanced tools
Benchmark your own coding-agent harness against FrontierHarness Eval: sets up the benchmark workspace and installs the agent skill that drives the run.
Explore the live results → · Read the blog → · Evaluate your own harness →
We ran the same Kimi K3 model through nine coding-agent harnesses—12 configurations in total—on the same 30 software-engineering tasks. With the model, tasks, and runtime held constant, changing the harness changed pass rate, cost, cache behavior, and speed.
| Harness (configuration) | Pass rate | Median cost per pass | Cache, median cell | Median time |
|---|---|---|---|---|
| Codex | 66.7% | $3.47 | 88.0% | 6m 43s |
| DSH Creator | 63.3% | $3.28 | 84.3% | 6m 44s |
| Claude Code | 63.3% | $18.34 | 67.8% | 9m 38s |
| Pi | 60.0% | $2.43 | 79.4% | 7m 33s |
| DSH PTC | 60.0% | $4.58 | 87.2% | 7m 44s |
| DSH Standard | 60.0% | $3.46 | 86.5% | 6m 17s |
| Oh My Pi | 56.7% | $4.75 | 82.2% | 6m 46s |
| Kimi Code | 56.7% | $3.65 | 88.0% | 7m 56s |
| DSH Minimal | 56.7% | $4.72 | 84.6% | 5m 41s |
| Exo Harness | 53.3% | $1.05 | 70.3% | 6m 17s |
| OpenCode | 50.0% | $3.24 | 78.4% | 6m 27s |
| Hermes | 50.0% | $2.90 | 85.9% | 6m 58s |
The interactive report includes failed runs, total cost per task, cache behavior, speed, and task-level results. For the evaluation design and analysis, read the launch article.
.
├── benchmark.json # Public benchmark definition
├── cli/index.mjs # `npx @frontierharness/eval`: workspace + skill installer
├── metadata/
│ ├── difficulty.json # Difficulty assignments and source methodology
│ └── harness-versions.json # Harness versions used for the run
├── results/
│ └── eval-data.json # Normalized aggregate and task-level results
├── tasks/<task>/
│ ├── instruction.md # Prompt shown to every harness
│ └── task.toml # Public task metadata and environment definition
└── skills/frontierharness-eval/ # Agent-neutral skill, usable by hand
├── SKILL.md # Evaluation workflow for a third-party harness
├── PROMPT.md # Copy-paste prompt that points an agent at the skill
├── reference.md # Command reference, runner templates, troubleshooting
└── scripts/ # Provisioning, trial runner, scoring, chart, report
The repository intentionally contains results, task definitions, and the evaluation workflow. Internal infrastructure, credentials, runtime identifiers, private evidence, solutions, and deployment configuration are not included.
The workflow that produced the table above ships with this repository, so a harness that is not in it can be scored on the same tasks, runtime, and cost accounting, then placed directly next to the twelve baseline configurations.
skills/frontierharness-eval/ is an agent-neutral skill: point any coding agent that reads SKILL.md at it and it will drive the whole run. Nothing in it is tied to a particular agent — the steps are plain Bash and Node, so you can equally run them by hand. If you would rather hand the run to an agent than read the commands below, PROMPT.md is a copy-paste prompt with five blanks to fill in.
npx @frontierharness/eval
One command, no clone. It creates a frontierharness-eval/ workspace holding the benchmark data every step reads — benchmark.json, results/eval-data.json, tasks/ — and installs the skill into the directories Cursor, Claude Code, and Codex read skills from, so opening that folder in your agent and asking it to evaluate your harness is enough. Add --global to install the skill for every project, --agent cursor to pick one agent, or run it inside a clone of this repository, where it skips the data and just installs the skill.
The same command wraps the rest of the workflow, so you can drive a run from anywhere:
npx @frontierharness/eval doctor # node >= 18, jq, runta CLI and auth, provider key
npx @frontierharness/eval prompt --harness my-harness --smoke # the agent prompt, two-task version
npx @frontierharness/eval provision --help # the scripts below, from the workspace
provision, run, normalize, chart and report are the skill's own scripts, so the steps below are the same work whether you invoke them as npx @frontierharness/eval provision … from anywhere or as $FH/provision-golden-checkpoint.sh … from the repository root.
If you hand the run to an agent, you can skip this section: the skill walks through the same setup, asks for the provider key once, and stores it as a Runta secret stub, so it never enters the runtime or the checkpoint. To drive the scripts yourself, set up the shell first:
export RUNTA_TOKEN=rt_... # Runta dashboard -> Settings -> Runta API Keys
export FIREWORKS_API_KEY=... # or the key for whichever provider you pick, below
FH=skills/frontierharness-eval/scripts
The provider key only has to be exported the first time. Provisioning turns it into a Runta secret and the API never hands the value back, so a later re-cut of the checkpoint reuses the stored secret.
The model is not a variable, but the provider is. Every published configuration runs Kimi K3, so that the harness is the only thing that differs, and your run has to match to be comparable. Which provider serves it is up to you — pass --provider to both scripts and the model route and key name follow from it:
--provider | Model route | Key |
|---|---|---|
fireworks (default, used by the published runs) | fireworks_ai/accounts/fireworks/models/kimi-k3 | FIREWORKS_API_KEY |
moonshot | moonshot/kimi-k3 | MOONSHOT_API_KEY |
openrouter | openrouter/moonshotai/kimi-k3 | OPENROUTER_API_KEY |
together | together_ai/moonshotai/Kimi-K3 | TOGETHER_API_KEY |
custom | your --model | your --secret-name |
Same weights from a different provider means your pass rate stays comparable; only the cost column is at risk, since it depends on that provider's token prices. The report flags it for you if you used anything other than Fireworks. Overriding the model is what breaks comparability, and the scripts warn when you do.
1. Freeze a golden checkpoint. Creates a clean Runta runtime, clones your harness at a pinned commit, installs Harbor for Terminal-Bench tasks and Pier for DeepSWE tasks, pre-pulls the task images, and captures the whole stack as one checkpoint. Your provider key is stored as a Runta secret stub, so it never enters the runtime or the checkpoint.
$FH/provision-golden-checkpoint.sh \
--runtime fh-build --checkpoint fh-golden-myharness-v1 \
--harness my-harness --provider fireworks \
--repo https://github.com/acme/my-harness --commit 9f2c1ab \
--cpus 4 --memory 8192 --prepull-tasks tasks \
--install-script ./install-my-harness.sh
2. Run the tasks. Each task gets its own fresh restore of that checkpoint and the runtime is deleted afterwards, so every trial starts from an identical cold start. The agent trajectory, verifier logs, and model.patch are copied out per task as evidence.
jq -r '.harnesses[0].task_details[].id' results/eval-data.json > tasks.txt # the published 30 tasks
$FH/run-trials.sh \
--checkpoint fh-golden-myharness-v1 --harness my-harness \
--provider fireworks --run-id 2026-09-02-myharness \
--tasks tasks.txt --out runs
Pass the same --provider you provisioned with — the checkpoint carries that provider's key name as a stub.
3. Score it, chart it, share it.
node $FH/normalize-results.mjs --run runs/2026-09-02-myharness --label "My Harness"
node $FH/generate-chart.mjs --run runs/2026-09-02-myharness
node $FH/build-report.mjs --run runs/2026-09-02-myharness
Everything lands in runs/<run-id>/report/: a chart.svg with your harness highlighted against all twelve baselines, a REPORT.md, and a self-contained index.html you can open or attach anywhere. To publish a link:
gh gist create runs/2026-09-02-myharness/report/REPORT.md \
runs/2026-09-02-myharness/report/chart.svg --public
A score only belongs next to the published numbers if the run holds these invariants. The report states any that were relaxed.
reference.md.terminal-bench-sample.infra_invalid, not scored as task failures.Cost is compared on effective_cost_per_pass, which is total cost across all tasks divided by passes and is reproducible from raw per-task cost. The *_normalized fields in results/eval-data.json reprice first-turn cache reads using data that is not public, so the scoring script leaves them empty rather than inventing values.
Metric definitions and the trial record contract are in SKILL.md. Runner templates, the alternative Harbor-with-Runta-provider topology, and troubleshooting are in reference.md.
| Configuration | Version | Configuration | Version |
|---|---|---|---|
| Codex | 0.148.0 | DSH Creator | 0.1.0-rc.8 |
| Claude Code | 2.1.237 | DSH Minimal | 0.1.0-rc.8 |
| Pi | 0.84.2 | DSH PTC | 0.1.0-rc.8 |
| DSH Standard | 0.1.0-rc.8 | Oh My Pi | 17.4.0 |
| Kimi Code | 0.37.2 | Exo Harness | 0.1.0 |
| OpenCode | 1.18.19 | Hermes | 0.20.4 |
task.toml are prepared once as a golden checkpoint. Every run is a fresh restore with identical vCPU, memory, disk size, disk contents, and memory state.See benchmark.json for the public benchmark definition and results/eval-data.json for the complete normalized result set.
jq '.harnesses[] | {name, successful, effective_cost_per_pass}' results/eval-data.json
Every task directory contains the exact public instruction and task metadata used by the benchmark.
Runta provided the isolated runtimes and Golden Checkpoint restores used across all 360 evaluations.
FAQs
Benchmark your own coding-agent harness against FrontierHarness Eval: sets up the benchmark workspace and installs the agent skill that drives the run.
The npm package @frontierharness/eval receives a total of 41 weekly downloads. As such, @frontierharness/eval popularity was classified as not popular.
We found that @frontierharness/eval demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 2 open source maintainers collaborating on the project.
Security News
/Application Security
A practical guide to securing GitHub Actions against supply chain attacks, covering malicious pull requests, credential theft, dependency risks, and release pipelines.
Security News
Composer 2.11 adds dependency cooldowns, joining package managers that delay new releases to reduce supply chain risks while allowing urgent security updates.
Threat Research
/Security News
Cursor removed a malicious Claude Code 2.1.313 extension from its marketplace. Socket examines registry records and the initial claims about its distribution.