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.
• "What does not work today" resolved: 0.1.130 reads AGENTFIELD_AFORGE_COMMAND (exec default), passes --timeout/--turns. README wording updated.
• 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; cloudsecurity.run_network_hunter on a 3-resource terraform dir found the 0.0.0.0/0 SSH security group at the right line (3 findings), 40 s, <$0.05; .aforge/trace/1.trace.log written; live argv aforge exec --json -w … --timeout 1795 --turns 50 --context-fill 60 --completion-reserve 65536 --model … --plan-model …. Code default with all provider env unset = aforge; HARNESS_PROVIDER / CLOUDSECURITY_PROVIDER=claude-code flip it. Tests: same failure set as main (pre-existing rot in tests/test_utils.py etc.), +4 new passing config tests.
• Un-draft preconditions met (site live, 0.1.130 on PyPI).

---

Summary

Make AForge the default CloudSecurity-AF harness and run it through exec, with
OpenCode retained as a configuration-only rollback.

Two things changed since the first draft, both aimed at making a clean checkout
build and run with no private access:

• The AForge CLI is downloaded, not pulled from a registry. The image used to
  do FROM ghcr.io/agent-field/aforge-v2:chat-v2-exec + COPY --from=aforge /aforge.
  That package is not publicly pullable, so docker compose up --build failed for
  anyone without registry credentials. A fetch stage now downloads
  ${AFORGE_BASE_URL}/${AFORGE_VERSION}/aforge-linux-${TARGETARCH}.gz, gunzips it,
  and verifies the decompressed binary's SHA-256 against the release
  checksums.txt before it is copied to /usr/local/bin/aforge.
• The SDK is pinned to a released PyPI version. agentfield @ git+…@bfd34426
  (which also needed [tool.hatch.metadata] allow-direct-references) is replaced
  by agentfield>=0.1.129, the released SDK that already ships the Python AForge
  harness provider.

Configuration

Build arg	Default	Purpose
AFORGE_BASE_URL	https://agentfield.ai/downloads/aforge	Download host serving <version>/aforge-linux-<arch>.gz and <version>/checksums.txt
AFORGE_VERSION	v0.1.0	Released AForge version to install

Both are overridable via --build-arg, and docker-compose.yml forwards them from
the environment, so a mirror can be selected without editing the Dockerfile.

SDK pin: what works today and what does not

agentfield>=0.1.129 supplies everything this agent actually depends on —
SUPPORTED_PROVIDERS includes aforge, HarnessConfig has an aforge_bin field,
and the provider runs aforge exec --json -w <root>. Two caveats to bump later:

• AGENTFIELD_AFORGE_COMMAND is a no-op on 0.1.129. The env var is set in the
  image, in compose, and injected into the harness subprocess env, but the shipped
  provider hardcodes the exec subcommand — nothing reads the variable. It becomes
  a live switch only once feat(harness): make aforge the default provider across Python, Go, and TypeScript agentfield#905 is released. Keeping it wired
  now means the rollback path costs no code change later.
• 0.1.129's provider passes no --timeout, so AForge's own 15-minute wall
  applies to each harness call (the SDK's outer kill is
  AGENTFIELD_HARNESS_TIMEOUT_SECONDS, default 1800s). #905 adds explicit flag
  plumbing.

Exact lines to bump when the #905 release lands (same version string in both):

• pyproject.toml:14 — "agentfield>=0.1.129",
• Dockerfile:51 — "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 change must exhibit:

1. docker build . succeeds with no registry credentials and no git access to a
   private repo.
2. The installed /usr/local/bin/aforge is byte-identical to the published
   release artifact — a tampered or truncated download fails the build rather
   than shipping.
3. aforge is on PATH and executable as the non-root cloudsecurity user.
4. With only OPENROUTER_API_KEY set, the agent resolves the AForge harness —
   provider aforge, AforgeProvider, binary /usr/local/bin/aforge.
5. HARNESS_PROVIDER=opencode still resolves OpenCodeProvider (rollback intact).
6. A real harness call through the SDK's AForge provider returns model output from
   inside the image.
7. The container still boots and serves /health.
8. CLOUDSECURITY_AFORGE_BIN / AFORGE_BIN override the resolved binary path.
9. The test suite and lint/type gates are no worse than main.

How it was verified

Built against a local mirror of the published layout (--build-arg AFORGE_BASE_URL=…);
the default value in the committed Dockerfile is the public host.

1. docker build → DONE, and the fetch stage logged sha256sum -c → aforge: OK.

2. Negative test: rebuilt the fetch stage against a mirror serving a byte-flipped
   binary with the genuine checksums.txt → aforge: FAILED /
   sha256sum: WARNING: 1 computed checksum did NOT match, build exit 1.

3. docker run --rm <img> aforge --help → exit 0, AForge usage banner.
   ls -l → -rwxr-xr-x … /usr/local/bin/aforge;
   in-image sha256sum = 61217a18…c79f5, matching the release checksums.txt
   line for aforge-linux-amd64; id → uid=10001(cloudsecurity).

4. In-container resolution with only OPENROUTER_API_KEY set:
   resolved provider: aforge, resolved provider class: AforgeProvider,
   resolved aforge_bin: aforge -> /usr/local/bin/aforge.

5. Same script with HARNESS_PROVIDER=opencode → OpenCodeProvider.

6. Live smoke inside the image, prompt Reply with exactly OK:

   • raw CLI: {"text":"OK","stop":"done","usage":{...},"turns":1,...}, exit 0.
   • through the SDK provider built from this repo's AIIntegrationConfig:
     is_error: False, returncode: 0, result text: 'OK'.
     In-image SDK version confirmed as agentfield 0.1.129.

7. docker run -d … then curl /health → {"status":"healthy","node_id":"cloudsecurity",…}
   within 2s.

8. docker compose config resolves; AFORGE_BASE_URL/AFORGE_VERSION overrides
   flow through to the build args.

9. Gates (this repo has no CI; the Development section of the README defines
   them). Run on Python 3.11 against main and this branch:

   gate	main	this branch
   pytest -q (ignoring the 2 modules that fail collection on main)	38 failed, 32 passed	38 failed, 36 passed
   ruff check src tests	110 errors	109 errors
   mypy src	33 errors	33 errors

   The 4 extra passes are the AForge default/rollback/bin-override/SDK-contract
   tests. tests/test_graph_context.py and tests/test_schemas.py fail collection
   on main too (ResourceEdge / ResourceCluster missing from
   schemas/recon.py, a file this PR does not touch); the 38 failures and the
   mypy count are identical on both refs.

Still draft — what un-drafts it

Two external things, both outside this repo:

1. https://agentfield.ai/downloads/aforge/v0.1.0/ must actually serve
   aforge-linux-{amd64,arm64}.gz and checksums.txt. The host is live and
   already serves the older build-<sha> coordinate; v0.1.0 is still a 404
   there. Until that publish lands, a default-args docker build cannot fetch
   the binary. Un-draft check: docker build . with no --build-arg overrides
   succeeds.
2. feat(harness): make aforge the default provider across Python, Go, and TypeScript agentfield#905 must be released so AGENTFIELD_AFORGE_COMMAND
   stops being decorative and the timeout is passed explicitly. Then bump the two
   pin lines listed above to that release.

Neither blocks review of the code itself.

Follow-ups (not in scope here)

• tests/test_graph_context.py / tests/test_schemas.py fail collection on main;
  tests/test_utils.py prompt-template assertions and
  TestDepthProfile::test_prover_caps fail on main too. Worth a separate
  repair PR — this branch neither fixes nor worsens them.
• The runtime image still installs OpenCode (~the rollback path). Once AForge has
  soaked, dropping it is a meaningful image-size win.

🤖 Generated with Claude Code