Current numbers (hidden-test = the agent never sees the failing test; $0 free tier; same harness, swap the model; 18 tasks × 3 seeds = 54 runs/model, 162 total):

Model	pass@1	pass@3
gpt-oss-120b	90.7%	100%
llama-3.3-70b	88.9%	94.4%
llama-3.1-8b	48.1%	66.7%

The score rises with the better model while the harness stays fixed (model-swap proof), and pass@3 > pass@1 shows real run-to-run variance — which is exactly why the CI gate is multi-seed. Every run deep-links its Langfuse trace.

ForgeJudge is the only open-source autonomous software-engineering agent that proves its quality in public on every commit: a hand-rolled single-agent solver, a deterministic execution-as-judge harness, an always-on leaderboard with per-run traces, and a CI gate that blocks regressions — all on a $0 / self-hostable stack against a contamination-resistant, intrinsically-verifiable golden set.

The engineered harness, observability, and gate are the deliverable — not a high resolution rate. A $0 free-model agent will score modestly by design. We prove value with a model-swap comparison: the score rises with a better model while the harness stays fixed.

How it works

flowchart TD
    G["Golden set · Git-canonical<br/>18 intrinsically-verifiable, mutation-hardened<br/>make-CI-green tasks"]

    subgraph SOLVER["Single-agent solver"]
        direction LR
        L["localize<br/>(BM25)"] --> R["repair<br/>(LLM router · critic · syntax edit-gate)"] --> V["validate<br/>(run tests)"]
    end

    G --> SOLVER
    SOLVER --> PATCH["unified diff"]
    SOLVER -. "every step traced" .-> TRACE["OTel → Langfuse<br/>per-run public trace"]

    PATCH --> H["Deterministic harness, in a sandbox<br/>apply test_patch + candidate patch · run F2P / P2P<br/><b>RESOLVED iff</b> every FAIL_TO_PASS passes AND every PASS_TO_PASS stays green<br/>swebench-equivalent · stricter on skips · cheat-resistant"]

    H --> STORE["Run store<br/>Neon + pgvector"]
    STORE --> LB["Leaderboard<br/>pass@1 / pass@3 · cost · tokens · trace link"]
    H --> GATE["Multi-seed CI gate<br/>a PR that lowers the resolution rate fails the build"]

    style G stroke:#3fb950,stroke-width:2px
    style H stroke:#4cc2ff,stroke-width:2px
    style GATE stroke:#f0883e,stroke-width:2px

• Solver — a single, phase-structured loop (localize → repair → validate), not a multi-agent swarm: cheapest, most deterministic, most debuggable. BM25 localization, an LLM router over free tiers, a syntax edit-gate, a cheap critic pre-filter, and a cost/step budget with autosubmit.
• Harness — encodes the SWE-bench RESOLVED_FULL rule and is verified equivalent to swebench.harness.grading on real PASS/FAIL/ERROR/XFAIL outcomes in CI — and deliberately stricter on a skipped FAIL_TO_PASS: swebench 4.1.0 rates a skipped oracle test RESOLVED_FULL (a skip is neither success nor failure), so a patch that makes the oracle skip rather than run grades as resolved. ForgeJudge counts a skip as not-passed, closing that cheat vector. Patches are also cheat-resistant: the canonical test files are restored before grading, so a patch can't neuter the oracle.
• Golden set — 15 purpose-built post-cutoff fixtures + 3 tasks mined from the author's own repos (real commit SHAs, MIT/own license — zero leak/copyleft risk). Each is mutation-hardened: a wrong fix to the patched region is caught (16 mutation-hardened at mean score 0.94; 2 inconclusive for regex/string code; 0 weak).
• Sandbox / CI / cron — GitHub Actions on a public repo does triple duty (ephemeral isolated VM sandbox + regression gate + scheduled sweep) at $0.
• Observability — OpenTelemetry GenAI spans (invoke_agent → retrieval / chat / execute_tool, gen_ai.usage.*, a gen_ai.evaluation.result pass/fail verdict) exported to Langfuse Cloud; every run is a clickable trace.

Two gates, two jobs

The deterministic gold-integrity gate (does the harness itself still work?) is kept separate from the stochastic regression gate (did a change make the agent meaningfully worse?) — because gold grading is deterministic and must never be averaged with noisy per-seed runs.

flowchart TD
    PR["Pull request / commit"] --> GG["Gold-integrity gate<br/>deterministic · $0 · re-grade all gold patches"]
    GG -->|"any gold task unresolved"| F1["fail — the harness broke"]
    GG -->|"all gold tasks resolved"| OK1["harness intact"]

    CRON["Scheduled multi-seed sweep"] --> SEEDS["run the agent × N seeds<br/>→ one resolution rate per seed"]
    SEEDS --> RG["Regression gate<br/>small-sample CI (Student-t / Wilson)"]
    BASE["baseline_scores.json<br/>per-seed reference"] --> RG
    RG -->|"candidate CI upper bound &lt; baseline CI lower bound"| F2["❌ fail — real regression"]
    RG -->|"overlapping · equal · improved"| OK2["✓ no regression"]

    style GG stroke:#4cc2ff,stroke-width:2px
    style RG stroke:#f0883e,stroke-width:2px

Quickstart

Prereq: uv (Python 3.12 is provisioned for you) — curl -LsSf https://astral.sh/uv/install.sh | sh.

git clone https://github.com/ahmedEid1/forgejudge && cd forgejudge
uv sync                       # Python 3.12, deps via uv

# Run the deterministic harness self-test (no API key, no network):
uv run python -m forgejudge.harness.runner_actions --patch-source gold   # 18/18 resolved

# Solve a task with a free model and grade it.
# Needs a (free) Groq key. Either export it, or put it in .env and pass --env-file:
#   export GROQ_API_KEY=...                     # or
#   cp .env.example .env && edit GROQ_API_KEY   # then: uv run --env-file .env python - <<'PY'
uv run python - <<'PY'
from forgejudge.golden.loader import load_tasks
from forgejudge.agent.solver import solve
from forgejudge.harness.grade import grade
task = {t.instance_id: t for t in load_tasks("golden/dataset.jsonl")}["fixture-semver-001"]
res = solve(task, run_id="demo", budget_usd=0.10, seed=0)
print(res.status, "→ resolved:", grade(task, res.patch).resolved)
PY

Fast tests: uv run pytest -m "not slow". Full golden validation + mutation hardening: uv run pytest -m slow. Sweep the leaderboard: uv run python -m forgejudge.eval.sweep --model groq/llama-3.3-70b-versatile --seeds 0,1,2. See CONTRIBUTING.md for the full pytest marker map and dev workflow.

Install

Working on the agent/harness itself? Clone and uv sync (above). To consume ForgeJudge as a package:

# Library + the `forgejudge` CLI (selftest / mcp / info):
pip install forgejudge
forgejudge selftest           # deterministic harness check — 18/18 resolved, no key
forgejudge mcp                # MCP server over stdio (needs the [mcp] extra)

# Zero-install MCP server (no venv to manage) — for an MCP client config:
uvx --from "forgejudge[mcp]" forgejudge mcp

Optional extras (installed only when you need them):

pip install "forgejudge[mcp]"            # one extra
pip install "forgejudge[harness,mcp]"    # several

forgejudge selftest and forgejudge info work with the base install — no extras, no API key, no network.

Six objections, pre-empted