[devparanjay]

Summary

• What changed: Adds command_code as a runtime, backed by a new AgentField command-code harness provider that drives the Command Code CLI (cmd -p) headlessly. The provider is registered by wrapping the SDK's provider factory from swe_af/runtime/ — the same technique the existing codex patch uses — so upstream AgentField stays untouched. Also adds command_code to the runtime Literals in the build/execution/fast/issue configs, installs the CLI in the image, and mounts the host's ~/.commandcode into the agent containers.
• Why it changed: To expand the available runtimes for SWE-AF. The goal was to make Command Code selectable alongside claude_code / open_code / codex. No model ids are baked in — a blank model omits -m so Command Code uses its own default, and any id it supports (now or later) passes through verbatim.

Validation

• make check — 1311 passed, 1 skipped.
• Relevant manual test performed — live headless run against the real CLI: NDJSON result parsed, .agentfield_output.json written and schema-validated through the AgentField runner, tokens/cache non-zero, and multiple full real build runs with no model config and expected results.

Behavior Impact

• No behavior change
• Backward compatible behavior change — purely additive; existing runtimes and their defaults are untouched.
• Breaking change (explain below)

Notes

• Auth: cmd login on the host plus the ~/.commandcode mount (mirrors ~/.codex). Hosts without bind mounts (Railway, etc.) need BYOK providers.json or a secret-mounted auth.json.
• Follow-up: the Go node isn't covered — it still rejects command_code, since its SDK provider factory is a closed set.
• tests/fast/test_integration.py::test_ac_14's allowlist was widened: the runtime set is declared in the config-schema modules, not in swe_af/runtime/ alone. Reasoner and coding-loop files
  remain disallowed.
• Dockerfile installs command-code unpinned (latest).
• Known limits: harness cost reports as unknown (token counts are recorded); per-role tool allowlists aren't expressible, so capability is governed by permission mode.
• For a regular user who uses Command Code CLI and is already logged in, it just works - just set the default runtime to command_code. The user's configuration of Command Code on their machine is inherited by default, requiring minimal or no further configuration for SWE-AF or AgentField.

[CLAassistant]

All committers have signed the CLA.

[devparanjay]

Just to add, I manually tested this PR branch by using it to make a simple web project - throwball.
It reached an impressive result in just the first run, but it took 2 more build runs to sort out minor issues and reach the final outcome.
In total, it consumed 269M+ tokens.

[ZeroPoint95]

@devparanjay, thanks for adding the Command Code runtime. Two issues still need fixing:

1. High — the Docker image cannot start the installed CLI. Dockerfile:23 installs current Command Code 1.82.2, which requires Node 22+, while the image’s Debian package supplies Node 20.19.2. I inspected the published entry point and reproduced its startup rejection using an isolated version-guard test. Please install a supported Node version and add a credential-free startup smoke test.

2. Medium — configured executable paths are ignored. Provider construction:380 always creates CommandCodeProvider() without using _resolve_bin. SDK-runner probes confirm that SWE_COMMANDCODE_BIN and per-call binary overrides still launch cmd. On Windows, that also selects the system command shell. Please connect the resolver to execution, use a portable default, and test the overrides through the public runner.

make check passed: 1,311 tests passed and 1 skipped. These reproductions used mocked dispatch and the isolated startup guard; no Docker image build, real CLI execution or provider authentication was performed.

[AbirAbbas]

@devparanjay thanks for this, and sorry it sat. Since it had been open a while, I've taken it over and pushed the fixes on top of your branch (no history rewritten).

@ZeroPoint95, both of your points are addressed:

Node / Docker startup. The image now copies Node and npm from the official node:22-bookworm-slim image, pins command-code@1.82.2, and runs command-code --version during the build, so an incompatible Node fails the build instead of the first run. I installed that exact package in a scratch prefix: under Node 22 --version and --help work, and under Node 20.20.2 it exits 1 with "Command Code needs Node.js 22 or newer". go/Dockerfile doesn't install Command Code, but the Claude npm package it installs now also declares Node >=22, so that image gets the same Node 22 plus a claude --version check. (No Docker host on my side, so this is the real CLI run locally rather than an image build.)

Executable overrides. Provider construction now goes through the resolver: commandcode_bin > bin_path > SWE_COMMANDCODE_BIN > the package's command-code alias, so the default no longer collides with cmd on Windows. There are seven HarnessRunner.run cases covering that precedence, blank values and paths with spaces, plus Windows stdin and structured output. Against the previous head they all dispatch bare cmd and fail.

Two more things came up while verifying:

• Exit code 6 (connection error) was classified as transient but never actually retried, because the SDK retries on the message text. It's fixed, with a test.
• docker-compose.yml and docker-compose.local.yml were mounting the whole host ~/.commandcode read-write into every agent service, even when Command Code isn't the runtime. That directory holds auth, settings, MCP config, sessions and history, so a container could change host config. It's now opt-in through docker-compose.commandcode.yml / docker-compose.commandcode.local.yml, and it mounts only auth.json, read-only. The docs cover the setup.

make check on 3.12 passes (1,383 passed, 1 skipped), and the Go job's gofmt, build, vet and test -race are green. The one Go race failure I hit, cmd/furrowd TestChildServesRecordedStoreDirNotRawRunID, also reproduces on main (4 of 20 runs), and this PR doesn't touch Go code.

[AbirAbbas]

Verified at c44db10: Command Code 1.82.2 starts under Node 22 and refuses Node 20 with the original error, executable overrides reach the public runner, exit 6 retries, and host credentials are now an opt-in read-only auth.json mount. make check (3.12) and the Go job are green.

[AbirAbbas]

One more round before merging. A second, independent review turned up a few things, now fixed:

• Fast builds ignored the runtime. swe-fast accepted command_code, but its own runtime-to-provider mapper only knew Claude and Codex, so every stage quietly ran on OpenCode. It now goes through the shared mapper, which keeps the old unknown-runtime fallback. A new test drives the real fast build and checks all five stages for each of the four runtimes; it fails on the previous head.
• QA synthesizer coverage. The synthesizer's runtime table from #155 now has a command_code row.
• Image compat. Both images add the nodejs → node alias that Debian's package provided. The changelog and deployment docs note that npm's global config moved from /etc/npmrc to /usr/local/etc/npmrc, with the override.
• Docs. The runtime tables and lists include Command Code, marked Python-node only.

I merged current main in as well. make check (1,441 passed) and the Go job are green. Thanks again @devparanjay!

[devparanjay]

@ZeroPoint95 thanks for the review and apologies, has been difficult to find time to work on this recently.
@AbirAbbas No issues, things can be understandably busy at times. And thanks for taking it over, I probably wouldn't have been able to work on it this week either.