feat(runtime): execute docker environment recipes

[christso]

Summary

Docker environment recipes now prepare a local Docker image/context and run generic CLI targets inside that environment with environment.workdir as the command cwd. The environment recipe remains separate from target/provider selection: Docker setup builds or pulls the image, produces a sandbox runtime for compatible CLI targets, and unsupported provider combinations return structured sandbox-infra results instead of falling back to host execution.

Implementation Notes

• Supports type: docker recipes with image, context, dockerfile, workdir, env, secrets, mounts, resources.cpus, resources.memory, and setup.command/setup.args/setup.env.
• Uses the existing Docker workspace/provider plumbing for image availability and the existing sandbox runner for execution, with resource limits and the Docker host cwd threaded through.
• Adds a temp-dir mount so CLI {PROMPT_FILE} and {OUTPUT_FILE} paths remain available across the host/managed Docker boundary.
• Keeps host environment behavior unchanged and delays provider creation until after environment setup so Docker-derived runtime config is applied only when needed.

Validation

• bunx biome check packages/core/src/evaluation/environment/docker.ts packages/core/src/evaluation/workspace/setup.ts packages/core/src/evaluation/providers/sandbox-runner.ts packages/core/src/evaluation/orchestrator.ts packages/core/test/evaluation/environment/docker.test.ts packages/core/test/evaluation/orchestrator.test.ts
• bun test packages/core/test/evaluation/environment/docker.test.ts packages/core/test/evaluation/orchestrator.test.ts packages/core/test/evaluation/providers/cli.test.ts packages/core/test/evaluation/providers/sandbox-runtime.test.ts
• bun run typecheck
• git diff --check

Dogfood Evidence

Controlled Docker CLI target:

• Command: bun apps/cli/src/cli.ts eval /tmp/agentv-docker-env-dogfood/docker-env.eval.yaml --targets /tmp/agentv-docker-env-dogfood/targets.yaml --target docker-cli --workers 1
• Run ID: 2026-07-05T10-52-06-267Z
• Result path: .agentv/results/2026-07-05T10-52-06-267Z
• Result: PASS 1/1, target provider cli, runtime sandbox, command cwd /app.

Live LLM grader with Docker CLI target:

• Command: bun apps/cli/src/cli.ts eval /tmp/agentv-docker-env-dogfood/docker-env-live-grader.eval.yaml --targets /tmp/agentv-docker-env-dogfood/targets-live-grader.yaml --target docker-cli-live-graded --workers 1
• Run ID: 2026-07-05T10-53-51-519Z
• Result path: .agentv/results/2026-07-05T10-53-51-519Z
• Provider/target: target provider cli; grader target OpenAI-compatible provider via GitHub Models.
• Result: PASS 1/1, score 1.0, runtime sandbox, command cwd /app, exit code 0.

Private evidence was published to agentv-private:evidence/av-noh3-2-4-docker-environment-runtime at commit bd0f659.

Post-Deploy Monitoring & Validation

After merge, validate the first environment docs/example PR against this behavior with one Docker recipe that uses an image and one that builds from context/dockerfile. Watch CI failures around sandbox runtime tests and CLI output-file handling, because those paths depend on the host temp-dir mount being available inside Docker.

Related

Bead: av-noh3.2.4

---

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	4ea20c5
Status:	✅  Deploy successful!
Preview URL:	https://2ffca2c9.agentv.pages.dev
Branch Preview URL:	https://environment-docker-runtime.agentv.pages.dev

View logs