feat: make AForge exec the default harness

[AbirAbbas]

Status update — agentfield v0.1.130 is released; this PR is ready

Everything below the line was written against the 0.1.129 pin and is kept for history. What is true on the current head:

• Pins bumped: agentfield>=0.1.130 in pyproject.toml and the Dockerfile; .env.example no longer calls AGENTFIELD_AFORGE_COMMAND reserved (0.1.130 reads it: exec default, do opt-in; --timeout/--turns are passed).
• Live verification (real OpenRouter key, default provider): image built with default args from agentfield.ai + PyPI (aforge v0.1.0, agentfield 0.1.130); node registered on a fresh agentfield/control-plane:latest; contract-af.anatomy_phase (the repo's only direct app.harness() path) on a 7-clause NDA returned 16 sections / 5 defined terms / 4 cross-refs and flagged the one-sided indemnification as critical — 3 runs, ~30 s each, cents; live argv aforge exec --json -w … --timeout 1795 --turns 50 --context-fill 60 --completion-reserve 65536 --model … --plan-model …. Code default with CONTRACT_AF_PROVIDER/HARNESS_PROVIDER unset = aforge; overrides flip it. The new agentfield-package.yaml parses under af show-requirements v0.1.130. Tests: same failure set as main (+4 new passing), incl. the pre-existing tests/test_review_layer.py import error.
• Un-draft preconditions met (site live, 0.1.130 on PyPI).

---

What / why

Makes AForge exec the default Contract-AF coding harness, and makes the container able to
get AForge without a private registry.

The original draft pulled the binary out of ghcr.io/agent-field/aforge-v2:chat-v2-exec.
That image is not publicly pullable (the source repo is private), so docker compose up --build
— the repo's advertised quick start — could not work for anyone outside the org. This replaces
it with a plain HTTPS fetch of the published release asset plus checksum verification:

ARG AFORGE_BASE_URL=https://agentfield.ai/downloads/aforge
ARG AFORGE_VERSION=v0.1.0

The aforge stage downloads ${AFORGE_BASE_URL}/${AFORGE_VERSION}/aforge-linux-${TARGETARCH}.gz,
gunzips it, and verifies the decompressed binary against checksums.txt from the same
directory (the release checksums hash the uncompressed binaries) before chmod +x. Both ARGs are
overridable, so CI or a local mirror can serve the same layout from somewhere else. The runtime
stage already installs ca-certificates, which AForge needs to reach openrouter.ai over HTTPS.

It also drops the agentfield git-SHA pin in favour of the PyPI release that actually ships the
Python AForge provider.

Changes

• Dockerfile — ARG AFORGE_IMAGE / FROM ${AFORGE_IMAGE} AS aforge / COPY --from=aforge /aforge
  replaced by a checksum-verified fetch stage; COPY --from=aforge /out/aforge /usr/local/bin/aforge.
  The fetch stage reuses python:3.11-slim (debian bookworm) so the build does not pull a second base image.
• pyproject.toml + Dockerfile — agentfield @ git+…@bfd34426…#subdirectory=sdk/python → agentfield>=0.1.129.
• agentfield-package.yaml (new) — Contract-AF was the only AF product without an install manifest.
  Mirrors the pr-af / SWE-AF shape: config_version: v1, entrypoint, node id/port, and the
  user-environment contract (OPENROUTER_API_KEY required; server/model/harness overrides optional).
  dependencies.system names the AForge CLI so af install prints the manual step up front instead
  of failing on the first harness call — af install builds a Python venv, it does not fetch harness CLIs.
• README.md / .env.example — corrected. The old text said AGENTFIELD_AFORGE_COMMAND=do
  selects AForge orchestration; that is not true on any released SDK (see below).
• tests/test_config.py — two more behaviours pinned: aforge_bin override precedence and
  provider-key forwarding.

SDK pin: what works today, what is inert

agentfield==0.1.129 (PyPI, current latest) ships agentfield/harness/providers/aforge.py and
HarnessConfig.aforge_bin, so the default path in this PR is fully live:

• HARNESS_PROVIDER=aforge → build_provider returns AforgeProvider, which runs
  aforge exec --json -w <project_dir>.
• CONTRACT_AF_AFORGE_BIN / AFORGE_BIN → read by contract_af.config and passed as
  HarnessConfig.aforge_bin, which 0.1.129 honours. These work today.

Two caveats, both to revisit when Agent-Field/agentfield#905 is released:

1. AGENTFIELD_AFORGE_COMMAND is a no-op. It is forwarded into the harness subprocess
   environment, but neither 0.1.129 nor the AForge binary reads it — the provider always runs
   aforge exec. That happens to be the mode this PR wants, so it is harmless, but it is not a
   switch yet. It is kept (in Dockerfile, docker-compose.yml, .env.example) so the behaviour
   is already declarative once #905 lands.
2. 0.1.129's provider passes no --timeout, so AForge's own 15-minute wall applies. The SDK's
   outer subprocess timeout is AGENTFIELD_HARNESS_TIMEOUT_SECONDS (default 1800s), which is
   longer — AForge self-terminates first, so there is no orphan process, but a long harness call
   cannot currently be given a bigger budget from Contract-AF.

Exact lines to bump when #905 ships a release:

• pyproject.toml:11 — "agentfield>=0.1.129",
• Dockerfile:54 — "agentfield>=0.1.129" \

There is no Go node in this repo, so there is no sdk/go pseudo-version to carry.

Validation contract

Observable behaviours this PR must exhibit:

1. With no harness env set, Contract-AF resolves provider aforge and binary aforge.
2. HARNESS_PROVIDER=opencode rolls back to OpenCode, which stays installed in the image.
3. CONTRACT_AF_AFORGE_BIN beats AFORGE_BIN beats the default aforge.
4. The harness subprocess environment carries OPENROUTER_API_KEY.
5. A default docker build produces an image containing an executable /usr/local/bin/aforge
   whose sha256 equals the published checksum for its architecture.
6. A tampered or missing asset fails the build rather than shipping an unverified binary.
7. Inside the image, aforge exec --json returns a JSON envelope on stdout and exits 0.
8. Inside the image, the runtime environment resolves to a real AforgeProvider bound to
   /usr/local/bin/aforge.

How verified

This repo has no CI, so the gate is local. Everything below was run against a mirror serving the
exact published layout (/downloads/aforge/<version>/…), pointed at with
--build-arg AFORGE_BASE_URL=…; the default value is the public URL.

• pytest -q — contract items 1-4. See "Known pre-existing failures" below for the baseline.
• ruff check on the files this PR touches — clean.
• docker build — contract item 5; the build's sha256sum -c - step prints aforge: OK.
• Negative test: a build against a directory whose checksums.txt was altered fails at the
  verify step (aforge: FAILED), contract item 6.
• docker run --rm --entrypoint aforge <img> --help and a real one-shot
  aforge exec --json -w /tmp/w --timeout 120 --budget 20000 --turns 3 inside the container —
  contract item 7.
• In-container provider resolution check via agentfield.harness.providers._factory.build_provider
  using the image's own baked environment — contract item 8.

Known pre-existing failures (not introduced here, not fixed here)

Reproduced on origin/main with the same interpreter and the same agentfield version:

• tests/test_review_layer.py fails to collect: it imports CLAUSE_ALIASES and
  EXPECTED_CLAUSES from contract_af.agents.gap_analyst, which no longer defines either —
  gap_analyst now derives expected clauses from a prompt. The module needs rewriting against the
  current behaviour, not a re-export.
• tests/test_clause_analyst.py::test_depth_escalation_any_critical
• tests/test_clause_analyst.py::TestGetSectionTexts::test_extracts_known_section
• tests/test_intake.py::test_uses_first_3000_chars
• ruff check . reports 46 findings repo-wide (benchmarks, src/, tests/) — all on main.

Follow-ups

• Bump the two agentfield>=0.1.129 lines listed above once agentfield#905 is released, and turn
  AGENTFIELD_AFORGE_COMMAND into a documented switch at the same time.
• Repair the four pre-existing test failures and the test_review_layer.py collection error, and
  decide whether ruff check . should be a gate (this repo has no CI at all).
• Publish a checksum/signature for the manifest-declared AForge system dependency path used by
  af install, so a local (non-Docker) install gets the same verification the image does.

Still a draft

This stays in draft until both:

1. https://agentfield.ai/downloads/aforge/v0.1.0/… is live — the default
   docker build (no --build-arg) must succeed against the public URL. The host
   already serves the older build-<sha> coordinate; v0.1.0 is still a 404 there.
2. feat(harness): make aforge the default provider across Python, Go, and TypeScript agentfield#905 is released to PyPI and the two pin lines above are bumped to it,
   so AGENTFIELD_AFORGE_COMMAND stops being inert.

license/cla is pending because the CLA has not been signed for this repo by the commit author;
that needs the author to sign at cla-assistant, it cannot be resolved from here.

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}