Qualify an SCV in staging

Date: 2026-08-20

Status: OpenCode read-only proof passed; Codex individual-operator proof pending

Use this procedure to prove the first complete SCV image in staging. This lane qualifies one OpenCode-driven SCV run. It does not admit repository writes, candidate pushes, Forge promotion, autonomous deployment, or production use.

Handle the local SCV test artifact

The full local precommit suite can report a failure in test/openagents/scv/run_test.exs that does not reproduce on the clean revision before issue #109 or on a fresh isolated build of the current tree. Treat that result as a stale-build or full-suite artifact only after both baseline checks pass. The artifact does not qualify or disqualify an SCV staging run.

Safety boundary

The current staging.openagents.com service and production resources share the openagentsgemini Google Cloud project. This does not satisfy the isolated staging topology required for SCV autonomy. Until an isolated project exists, run this qualification as a separate Cloud Run job with all of these controls:

Use a dedicated service account with no project-level roles.
Grant the service account access to only the staging OpenAI secret.
Do not attach a database, VPC connector, GitHub credential, Forge credential, release cookie, deployment credential, or cloud API credential.
Deploy an immutable image digest from a dedicated immutable staging repository.
Set OPENAGENTS_RUNTIME_ROLE=scv, OPENAGENTS_ENVIRONMENT=staging, and SCV_PERMISSION_PROFILE=read_only.
Run one task with no retries and a bounded timeout.

The image starts only the SCV worker supervision tree. The worker admits the opencode driver and opencode-core environment, streams versioned JSON events to Cloud Logging, writes one terminal result, and exits.

Build and deploy

Commit and push the candidate. Record the exact Git SHA.
Run ops/scv/images/build-opencode-core-cloud.sh against the clean commit. The script builds ops/scv/images/opencode-core/Dockerfile on native linux/amd64 infrastructure with the SHA as OPENAGENTS_BUILD_REVISION.
Push the image to the dedicated immutable Artifact Registry repository and resolve its manifest digest.
Deploy or update the dedicated Cloud Run job by digest. Configure these nonsecret values:
- SCV_OBJECTIVE with a bounded read-only inspection request.
- SCV_MODEL with the admitted model identifier.
- SCV_REASONING_EFFORT=low for the production-code-capable SCV baseline.
- SCV_REPOSITORY_REVISION with the exact committed Git SHA.
- SCV_TIMEOUT_MS and SCV_HEARTBEAT_INTERVAL_MS with bounded values.
- SCV_DIAGNOSTIC_LOGS=false.
Map OPENAI_API_KEY from a specific enabled numeric version of the staging secret at runtime. Do not use latest, print the value, or persist it.
Execute one job task and wait for its terminal state.

Verify the execution

The Cloud Logging stream must contain these records for one run ID:

run_preparing;
process_started;
at least one heartbeat during a long-enough run;
normalized opencode_event records;
process_finished;
run_finished;
one openagents.scv.worker.result.v1 terminal result.

Verify that the result reports the opencode driver, opencode-core environment, openai/gpt-5.6-luna model, low reasoning effort, read_only permission profile, exact source SHA, successful exit status, bounded duration, resource measurements, event counts, token totals, and no output truncation. Verify the deployed job uses the dedicated service account and contains no prohibited environment or secret references.

Treat a successful execution as an environment qualification receipt only. Do not call it an isolated-staging pass or enable writes. Before an SCV can write, add the durable coordinator and worker protocol, process-tree or cgroup enforcement, run-scoped inference grants, persistent per-effect barriers, artifact storage, cancellation, and the isolated Forge staging lane.

Qualify the individual-operator Codex driver

Run this procedure only after an administrator connects an individual Codex account at /admin/scv/accounts and the account shows Ready. Do not add a service account until this path passes.

Deploy the exact candidate SHA and immutable image digest to every staging fleet node.
Confirm that the fleet starts the application container with seccomp=unconfined and apparmor=unconfined, without --privileged or added Linux capabilities. Confirm that the complete Codex package lives under /usr/local/lib/codex-package so the repository-scoped sandbox can read it.
Confirm that the connected account advertises gpt-5.6-luna and at least one of the admitted reasoning efforts, low or none.
Resolve the public repository, its node-local Forge storage key, and the exact candidate SHA. Do not use a mutable branch name for the run.
Call OpenAgents.SCV.CodexRuns.start/5 on one fleet node with the account ID, repository record, exact SHA, bounded read-only objective, and optional issue ID.
Observe /status while the SCV runs. The stream must advance through the Codex runtime, session, turn, tool or report, and terminal persistence phases without showing the objective, repository path, command, output, account identity, or credential data.
Call OpenAgents.SCV.CodexRuns.await/2 with a bounded timeout. Require a terminal succeeded row, a nonempty openagents.scv.report.v1 report, its SHA-256 digest, exact repository SHA, Codex thread and turn IDs, event count, usage, and resources.
Query scv_run_events for that run. Require driver_started, driver_session_started, turn_started, at least one successful tool_started and tool_completed pair, turn_finished, and run_finished. A message-only report does not prove repository access and does not qualify the SCV runtime.
Confirm that no disposable workspace or temporary CODEX_HOME remains and that the connected account remains Ready.
Scan the bounded event payloads, application logs, and public status response for credential patterns and raw protocol content. Any match fails qualification.

This procedure qualifies only read-only investigation and report persistence. It does not grant write, push, issue-transition, Forge promotion, deployment, or production authority.

First qualification receipt

The first shared-project qualification passed on 2026-08-20:

Field	Value
Source revision	`09a775b83fa05b6a92854b0ad1f7b5c23b3aee88`
Build	`9721075d-ed17-491c-af9a-85be8f46bf52`
Image digest	`sha256:ee2a74660faa6137e4021a2870380f5977796d33ca9dcaabcb0b958f35e0e36b`
Job execution	`openagents-scv-staging-wrppq`
SCV run	`db8a6849-1e59-41fb-ba7d-461c9a4fa4a3`
Result	`succeeded` in 11,008 milliseconds with two completed read calls and no output truncation

Cloud Logging received lifecycle events, five two-second heartbeats, eight normalized OpenCode events, resource measurements, and the terminal worker result before the container exited with status 0. This receipt proves the read-only image and event path. The safety limitation above remains in force.

After the job pinned the provider secret to its numeric version, generation 2 passed as execution openagents-scv-staging-9gdhz. SCV run cc749000-8961-44a5-8d98-6569f6b1b925 completed in 11,207 milliseconds with two read calls, eight normalized events, 754,118,656 bytes of peak RSS, 45 resource samples, no sampling errors, and no output truncation. This second receipt proves the currently deployed job configuration.

GPT-5.6 Luna staging receipt

Cloud Build 655a2fe6-8173-4686-9533-f3a1733942b0 built source revision c7175ed8a8781ea1aab7d204623a28ac45a70bc1. Artifact Registry resolved the worker image to sha256:156ff9f51e955d03b4795af8e2bb190a6c4f9962cc7942a6dcc0586e9b48b0a9. The control-plane, parity, production-readiness, and staging SCV jobs all use that digest with SCV_MODEL=openai/gpt-5.6-luna and SCV_REASONING_EFFORT=low.

Parity execution openagents-scv-parity-audit-p5mkz completed successfully as SCV run 0f13d278-02aa-47cd-bc07-3bdcd2fd6135. The OpenCode process ran for 58,553 milliseconds and emitted 62 normalized events. It completed 9 glob, 9 grep, and 22 read calls, recorded 688,361,472 bytes of peak RSS, and reported no sampling or output truncation.

Cloud Logging stored the terminal result as openagents.scv.worker.result.v1 and the 8,645-byte report as three ordered openagents.scv.report.chunk.v1 JSON records. Both the initial run_preparing event and terminal result recorded the exact model and reasoning effort. This receipt qualifies the Luna model-selection, live event, bounded report, and terminal-result paths in the shared staging project. It does not admit repository writes or count the report findings as claimed or completed issues.