Skip to content

Mechanical campaign readiness — 2026-10-02

OpenRouter support and its verification are documented in OpenRouter campaign readiness.

launch_agent_efficiency_pilot.py --prepare now runs preparation in one pass. It resolves the factory campaign, schedule, fixture checkouts, native executable, managed-login path, interpreter and runtime once. It writes an immutable readiness-receipt.json only after the following required checks pass:

Check Required evidence
Host context Writable rootless runtime directory, short Unix socket, runtime and tmux access
Factory inputs Validated immutable manifest, schedule, profile and fixture revisions/trees
Image runtime Installed product and analyzer source fingerprints, profile and active analyzers
Native tool dispatch Both arms: writable tool home, denied credential state, shell/Git/uv/Codira CLI; assisted: effective preapproval and real native MCP calls
Fixture freshness Source edits and old cursors rejected, reindex returns fresh source evidence
Rubric calibration All 72 source-curated calibration cases reproduce
Complete patch pipeline All 18 correct/incomplete/wrong cases use offline environment preparation, shared snapshot, actual workspace capture and complete composite oracle
Repository gate Fresh index and successful full validate_repo.py in retained tmux with log and terminal exit file
Authenticated route Native: exact model, reasoning and quota through the isolated subscription container. OpenRouter: authenticated /models/user, price/context/tool contracts and scoped key budget
Model-requested MCP A bounded real turn from codex exec, through the actual selected transport, must call both canary MCP tools successfully and complete its shell marker; complete events, diagnostics, usage and exit status are retained
Capacity after canary Repeat authenticated admission after the canary; retain native quota windows or OpenRouter key budget

Every evidence file is bound by digest. The receipt also binds repository validation inputs, harness, installed product, interpreter and executable digests, factory and fixture bindings, UID, and the child launch environment. Launch rechecks all of these; the generated runner repeats admission before credential access. Changed inputs, failed/missing checks, altered evidence and missing receipts reject execution. The operator procedure gives authenticated observations a four-hour admission window. After expiry, repeat only the read-only external-conditions comparison; unchanged contracts and sufficient capacity permit execution without repeating offline qualification or the paid canary. Retain new immutable refresh evidence linked to the original receipt, preserving its timestamp. See the external-conditions refresh procedure. The execution route also repeats authenticated preflight before model work and quota checks per attempt. The current verifier still implements fifteen minutes and does not yet support the approved four-hour/refresh policy.

Preparation stops at the first failure, records its check identity and error class, and retains diagnostics. It never retries automatically or upgrades a partial pass to readiness. A failed preparation requires a fresh execution root; expiry alone requires only the external refresh described above. Changed experiment/harness controls require a fresh factory campaign identity. The native subscription and OpenRouter routes support the representative panel. Other providers and task inventories reject preparation. There is no exemption for unqualified inventories. Historical launch receipts alone cannot admit execution.

Run preparation from the authorized host context. Container runtime and tmux access are host operations; the assistant should request host execution directly rather than first trying them in its outer sandbox. Preparation checks the context before runtime qualification or authentication. Offline tool probes run without credentials or model turns. After the full gate and authenticated admission pass, preparation requires one real model canary with a maximum 120-second timeout and the frozen token guard. It consumes subscription quota or OpenRouter credits, records preparation usage separately, and is never retried automatically. Inputs are rechecked before every boundary, including before this model turn. Native subscription preflight uses the existing managed ChatGPT login; no API key, SOPS provider environment or OpenRouter fallback is used on that route.

OpenRouter preparation starts without provider credentials in the parent. Both read-only authenticated admission and the model canary use the registered SOPS environment in scoped children. The canary rechecks the prepared bindings and source identity, repeats authenticated admission, and uses the same proxy and container executor as campaign attempts. Only a generated local proxy token enters the isolated container; the upstream API key remains outside it. No subscription login is accepted and provider fallback remains disabled. Exact upstream response bodies and digests are retained before interpretation, along with request/response observations and complete native process evidence. Missing provider usage, unsuccessful responses, local guard rejections, denied tools or missing native usage reject readiness. accounting.json records tokens and USD estimates at frozen price ceilings, explicitly separate from actual invoiced charges and the campaign's attempt ledger. The model turn is bounded to 120 seconds; its enclosing authentication child has a 300-second guard. The post-canary key-budget check must still admit the entire campaign pool.

Both native and proxy configuration builders set default_tools_approval_mode = "approve" on [mcp_servers.codira]. The native probe reads effective configuration and calls index_status and context_for_task through the native app-server handler. It requires a usable ready index and actual canary evidence. These offline checks exercise native dispatch without consuming inference quota. They cannot replace the mandatory model-requested MCP canary. That separate check requires completed successful calls from the real codex exec event stream; tool registration, direct server calls and direct app-server RPC calls alone cannot admit a campaign.

Temporary mounts explicitly use mode 1777. Tool HOME/config/cache are under /workspace/.benchmark; credentials remain denied under /codex-state. Both arms receive the same prepared CLI profile. Snapshot and capture share exclusions for installed dependencies and tool state; new source/test files remain captured, and source symlinks still reject capture. Captured patch text reaches the complete quality rubric, preserving behavioral results and pending blinded semantic review. MCP authorization denial is an operational infrastructure failure, even when tool use is optional. Expected tool-level rejections, such as stale-source checks, remain task evidence and do not become infrastructure failures.

Use factory-generated artifacts and supply every frozen fixture binding on the single preparation command. The native route additionally requires the prepared --subscription-codex executable and --subscription-auth-source login path:

uv run python scripts/launch_agent_efficiency_pilot.py \
  --prepare \
  --campaign-dir .artifacts/agent-efficiency/campaigns/<fresh-campaign-id> \
  --execution-root .artifacts/agent-efficiency/executions/<fresh-execution-id> \
  --seed <factory-seed> \
  --subscription-codex <qualified-native-executable> \
  --subscription-auth-source <existing-managed-login-path> \
  --fixture-source <fixture-id>=<absolute-frozen-checkout>

Repeat --fixture-source for every manifest fixture. No paths or controls are re-entered at launch:

For OpenRouter use the same preparation command without either --subscription-* argument. Do not wrap the whole preparation pipeline in SOPS; the pipeline scopes credentials only to its authenticated children.

uv run python scripts/launch_agent_efficiency_pilot.py \
  --launch \
  --execution-root .artifacts/agent-efficiency/executions/<fresh-execution-id>

Launch remains a separately authorized execution stage. Resume enforces readiness and existing unfinished-start investigation rules. Completed attempts and previous invocation logs are retained. Gate and campaign tmux panes use remain-on-exit and are retained with their logs and separately written exit statuses.

Implementation verification and raw offline evidence are recorded under .artifacts/agent-efficiency/readiness-qualification-20261002-r2/. Campaign 011 factory, execution and attempt evidence is unchanged. No new campaign or subscription model canary was executed in this implementation verification. Offline native dispatch, the 72 rubrics and 18 complete patch cases were exercised live; the model-canary admission logic was checked with recorded native event fixtures. A real readiness receipt requires the model canary against a fresh source-compatible factory campaign and image.