Protocol E2E Tests

Run these tests from the repository root. Install Bun and start a Docker-compatible daemon first. The runner builds current workspace source, including uncommitted changes, in a separate image. It does not deploy Monoize or use your dashboard configuration.

bun e2e/cli.ts list
export BASE_URL='https://example.com/v1/chat/completions'
export API_KEY='your-upstream-key'
export MODEL='your-model'
bun e2e/cli.ts run --harness codex-responses-ws-v2,claude-messages

BASE_URL must include the complete endpoint. The runner preserves its path and query. Standard suffixes infer responses, chat_completion, or messages. For a custom path, set --upstream-type explicitly. A conflicting type or an unknown harness stops the run before container creation. Real upstream calls consume the upstream account's quota.

Builds use configured HTTP_PROXY, HTTPS_PROXY, and NO_PROXY values. For loopback proxies, builds use the configured Docker daemon proxy, or fall back to host.docker.internal. These variables are not copied into runtime images.

Select clients and limits

The harness IDs are codex-responses-ws-v2, claude-messages, opencode-responses, opencode-chat, pi-responses, and pi-chat. Each ID selects a client and its downstream protocol. The default selection is all. Repeat --harness or separate IDs with commas.

The runner resolves the repository revision once for all items. Each item receives separate containers, credentials, database, and workspace. A failed item does not stop the remaining items. The runner does not retry failed items. Interrupting the runner preserves available evidence and removes temporary containers and volumes.

Read the results

Read report.json and the per-harness event files in the output directory. The report separates protocol checks from task checks. It records source and image identities and client versions. Credentials are redacted. Evidence exceeding 128 MiB per item fails the item.

Protocol checks require a successful tool call, local execution, result submission, and subsequent generation. Codex must use WS v2 and an actual previous_response_id continuation. An HTTP fallback fails that harness.

The built-in task clones the repository, reads at least three source files, and writes an architecture document. Checks verify the revision, execution evidence, document headings, and existing source references. They do not verify every claim in the explanation. Custom tasks require a nonempty final response and normal exit, plus the same protocol checks. They do not run the repository-specific checks.

Exit codes are 0 for all items passing, 1 for item failures, 2 for setup or argument errors, and 130 for interruption. The supported matrix contains six harnesses and three upstream types. Run each upstream configuration separately. Missing credentials or unavailable models leave those combinations unverified.

Check the test tool

cd e2e
bun install --frozen-lockfile
bun run check
bun test

fixture.ts provides a deterministic upstream for development checks. Fixture results do not establish compatibility with a real model.

OpenCode uses stateless Responses requests and disables title and summary agents. Claude disables optional thinking. Auxiliary HTTP errors remain in evidence but do not count as generation errors.

The runner stops an item after it records a generation HTTP error or an upstream terminal error. It continues queued items.

From the repository root, run the container fixture matrix, repository acceptance checks, and interruption check:

docker build -f e2e/Dockerfile -t monoize-e2e:dev .
E2E_DOCKER_IMAGE=monoize-e2e:dev bun test e2e/docker.test.ts