[earakely-scale]

M0.6 of the CUA env epic (FD-4345): the computer env type and the composer for container forms, FD-4356. It follows the reviewed design, CUA Env M0.6: Computer Env Type and the Composer, with its 14 choices confirmed. It builds on #354 (M0.1), merged in v0.9.1369.

What's in it

• The computer env type (ComputerEnv, env/envs/computer/). An env.toml names a base computer and a driver, and declares what makes the env itself:

  • forms, display, network, [tools] hide and [settings.<layer>];
  • apps is accepted only empty until computer app envs arrive (M1.1, M1.3).

  type = "computer"
  computer = "ubuntu-24.04"
  driver = "reference"
  display = { width = 1600, height = 900 }

• Gate 1 at put. The base, the driver and the declaration are matched before anything is stored, and every mismatch is refused at once:

  • the driver's supports against the base's OS, and the arches the two share;
  • forms other than container;
  • a driver with no install for the OS, or a system service on a base whose supervisord runs as the session user;
  • a base image in no registry and with no build context;
  • display outside 1 to 32767 per side, or beside a resolution setting;
  • an invalid network, a tools.hide pattern matching no tool or hiding one of the seven every computer offers;
  • a setting the layer doesn't declare, or of another type.

• The composer (composer.py) builds one OCI image per form and arch with this machine's Docker.

  • The base comes first: its registry image pinned by digest, or, for a base only stored as its build context (a checkout's), built here for that arch.
  • The driver's files, a venv of its requirements.txt, and its supervisord programs go on top.
  • The other arch is built through Docker's emulation and marked emulated. Without emulation that pair fails alone.
  • Each image is tagged agentenv-computers:<key> in the image store. The key hashes its Dockerfile, its whole build context, the base, the form, the arch and the composer's source, so any env with the same inputs reuses the image without building it.
  • In a checkout whose protocol is installed from source, the composer builds that source into the driver's venv, so an SDK change and the driver compose and test together. A release resolves it from PyPI at the release's pin.

• What a build records. Each pair's status goes into the env's metadata, composed, as it finishes: unverified with its key, image, packages installed (from pip list) and whether it was emulated, or failed with its stage, reason and a FileArtifact of the build log. The arches built are written as the env's computer, <env id>__computer, a computer artifact over one index of their images. It is written through ComputerArtifact.put, so M0.1's inspection reads each arch back.

• agent-env env build --id <id> [--version <n>] builds what a put left: failed pairs, and pairs a put that stopped left building. It reuses any image of the same key, prints each pair, and exits 1 if one failed. Builds end unverified; booting and probing them is M0.7, and until then a computer env's deploy is refused, naming its statuses.

• A driver declares its install (computer_driver env.toml): port, and [install.<os>] with files, python (a requirements file) and services. The reference driver gains a requirements.txt (protocol, mss, python-xlib) and a session service run by python -m agentenv_protocol.computer.serve driver.py:ReferenceDriver, a new SDK entry point that serves a driver class from its file.

• Aliases in every entry's references. As decided while building this, the planner resolves an installed bundle's alias in an env's, an agent's or an artifact's references as it does in a task's steps. It rewrites the field to the target's id and pulls the target in, for entries pulled from other bundles too. The rule that an entry's reference to an alias is refused goes away. This is what lets a computer env write computer = "ubuntu-24.04" and driver = "reference".

• A bundle run says it's building. It prints its "building with docker" line before writing an env whose type builds at put (Env.builds_at_put). Each pair's result is logged, and a failed pair is a warning.

Where this differs from the design

1. The build key hashes the rendered Dockerfile and the whole build context, beside the base, form, arch and composer source, instead of a list of named inputs. It covers every input the doc named, and also the driver's programs and install, which a key over its files alone missed (a unit test caught it).
2. Settings tables are keyed by a layer's id or its name, the last part of its id. The planner rewrites computer = "ubuntu-24.04" to the base's full id, so [settings."ubuntu-24.04"] must still find it.
3. A base stored only as its build context is built here for each arch, natively or emulated. The doc assumed a registry base, which a checkout doesn't have.
4. Progress is the bundle run's one "building with docker" line plus a log line per pair (INFO; failures are warnings), not a printed line per pair at default verbosity.
5. The composed computer's id is <env id>__computer, after derive_id, not <env id>-computer. Its settings' defaults carry the env's values and its display.
6. Gate 1 has no placement or permission checks. The driver env has neither field yet: every driver runs inside the computer and needs no OS permission.
7. A failed build's log is a FileArtifact, not an object url.
8. No data_paths on the env (decided in review). What a computer's data/get returns is the reader's choice: it names the paths it reads (data/get takes paths, M0.7), and with none, the driver returns its OS's standard user folders and declares them on its card. snapshot_env takes computer_file_paths with M1.8.

Testing

• Unit. 75 tests cover:

  • the env type, gate 1 and the composer, with Docker and the registry faked;
  • the driver's install format and the serve entry point;
  • the bundle folder.

  Line coverage of that code is 96% (composer 95%). The whole tier passes (11,157 tests) except parse_test.py::TestFiles::test_deep_nesting[json], which also fails on main under Python 3.14; CI runs 3.12.

• Integration (tst/integration/computer, real Docker):

  • on the laptop (arm64) at a1ad43a, and on EC2 (amd64) at f530255: 17 passed and 1 skipped on each.
  • The skipped test is the agent run, which needs a model endpoint. Neither shell held a model key; the agent run below was made separately.

• Chaos on the laptop (arm64, with amd64 emulated): 16/16 at f530255.

  • Measured: arm64 builds natively in 15 s and amd64 emulated in 68 s; a composed image is about 756 MB. The reference driver's probe passes at full (51 checks, three times), and the computer boots to ready in 1.6 s.
  • Interrupted: a put killed mid-build leaves its pairs building, and env build finishes them.
  • Unresolvable: a requirement that can't resolve fails its pair at build, with the log stored, and no computer is written.
  • Concurrent: two puts of the same key at once both finish, and the image pulls.

• Chaos on EC2 (amd64, with arm64 emulated).

  • At f530255:
    • measured: 9/9. amd64 builds natively in 22 s and arm64 emulated in 210 s. A second env reuses both images. The probe passes at full ×3 and boot takes 1.8 s.
    • interrupted: 2/2. unresolvable: 2/2. concurrent: 3/3.
  • Reruns at a1ad43a, each composing a fresh driver copy:
    • No emulation, Docker restarted after binfmt is removed: the arm64 pair fails at build with "this machine can't build linux/arm64: its Docker has no emulation for it", and amd64 builds.
    • No emulation, stale: binfmt is removed but Docker isn't restarted, so buildx still lists linux/arm64 and the build runs. Its exec format error is reported as "this machine can't run linux/arm64 programs: its Docker has no emulation for it", with the log stored. The f530255 run had reported only "the build failed", which is why 1f2feea maps it.
    • Refused push: with the registry read-only, both pairs fail at push (405) and the put still exits 0. Once it takes pushes again, env build builds both, unverified.

• An agent on the composed computer (laptop, at f530255): claude-code drove the env's composed arm64 computer and created a folder through the desktop's file manager. The task's outcome check passed in 30.7 s, and its recording had no missing segments.

• Greptile: 12 findings over four reviews, all fixed and answered on their threads:

  • a stale computer after a failed write (current(), images);
  • signing in to a private base;
  • reusing before building the base;
  • executable bits;
  • the labels of reused images;
  • MAX_PORT;
  • composed images going to the env's own image store;
  • signing in before the index is joined;
  • a named file mode;
  • composing under the stored id of an env reached by an alias;
  • two builds of one version keeping each other's pairs;
  • reading data_paths once, a field since removed (see below).

  Each finding from the last two rounds has a unit test that fails against the code before its fix.

🤖 Generated with Claude Code

Confidence Score: 4/5

The installed-entity alias failure still blocks merging; empty forms also remain accepted.

Findings

1.  Installed computer writes fail ▶
2.  Empty forms create nothing ▶

Fix with agent prompt

### Issue 1
src/agent_env/bundle/plan.py:305-310
`write_installed` uses `plan_entities`, which never calls `_resolve_aliases`. References such as `computer = "ubuntu-24.04"` and `driver = "reference"` therefore stay bare store references. On a fresh store, writing the env fails instead of pulling in its base and driver, even though a bundle run resolves them correctly.

Run alias resolution before checking and ordering entity-only plans too.

### Issue 2
src/agent_env/env/envs/computer/gate.py:29-30
`forms = []` passes the gate and saves an env that can never produce a computer. Composition has no pairs to build, so `_computer` writes nothing. Later, `env build` also has no pairs and exits successfully without repairing the env.

Reject an empty forms list before saving it.

```suggestion
    if not forms:
        problems.append("forms: name at least one form to compose")
    problems += [f"forms: the {form} form isn't supported yet; a computer env composes {', '.join(FORMS)}"
                 for form in forms if form not in FORMS]
```

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Adds computer environments that combine a base computer and a driver into container images.

• Computer environments combine a base computer and driver into images for each supported architecture.
• Computer drivers declare their install needs and can be served from their own files.
• Bundle entries can use installed aliases for their environment and artifact references.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Computer environment declaration] --> B[Resolve base and driver]
  B --> C[Check compatibility]
  C --> D[Save environment]
  D --> E[Compose each architecture]
  E --> F{Matching image exists?}
  F -->|Yes| G[Reuse image]
  F -->|No| H[Build and push image]
  G --> I[Record result]
  H --> I
  I --> J[Join completed images]
  J --> K[Save computer artifact]
  L[env build] --> E

Loading

_{Reviews (12) · Last reviewed commit: "refactor(computer): a computer env names..." · Reviewed by Greptile}

[greptile-apps]

Installed computer writes fail

write_installed uses plan_entities, which never calls _resolve_aliases. References such as computer = "ubuntu-24.04" and driver = "reference" therefore stay bare store references. On a fresh store, writing the env fails instead of pulling in its base and driver, even though a bundle run resolves them correctly.

Run alias resolution before checking and ordering entity-only plans too.

Knowledge Base Used: Bundle authoring, installation, and execution

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agent_env/bundle/plan.py
Line: 305-310

Comment:
**Installed computer writes fail**

`write_installed` uses `plan_entities`, which never calls `_resolve_aliases`. References such as `computer = "ubuntu-24.04"` and `driver = "reference"` therefore stay bare store references. On a fresh store, writing the env fails instead of pulling in its base and driver, even though a bundle run resolves them correctly.

Run alias resolution before checking and ordering entity-only plans too.

**Knowledge Base Used:** [Bundle authoring, installation, and execution](https://app.greptile.com/scale-ai/-/custom-context/knowledge-base/scaleapi/agentenv-framework/-/docs/bundle-authoring-and-execution.md)

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Empty forms create nothing

forms = [] passes the gate and saves an env that can never produce a computer. Composition has no pairs to build, so _computer writes nothing. Later, env build also has no pairs and exits successfully without repairing the env.

Reject an empty forms list before saving it.

Suggested change

	problems += [f"forms: the {form} form isn't supported yet; a computer env composes {', '.join(FORMS)}"
	for form in forms if form not in FORMS]
	if not forms:
	problems.append("forms: name at least one form to compose")
	problems += [f"forms: the {form} form isn't supported yet; a computer env composes {', '.join(FORMS)}"
	for form in forms if form not in FORMS]

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agent_env/env/envs/computer/gate.py
Line: 29-30

Comment:
**Empty forms create nothing**

`forms = []` passes the gate and saves an env that can never produce a computer. Composition has no pairs to build, so `_computer` writes nothing. Later, `env build` also has no pairs and exits successfully without repairing the env.

Reject an empty forms list before saving it.

```suggestion
    if not forms:
        problems.append("forms: name at least one form to compose")
    problems += [f"forms: the {form} form isn't supported yet; a computer env composes {', '.join(FORMS)}"
                 for form in forms if form not in FORMS]
```

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	pypi/​python-xlib@​0.33
	pypi/​mss@​10.2.0

View full report