arch-release-check

Release check

Steps

1. Export and lint. Export the project with platform_import_export, save the files to the project folder, then run debug_lint_abl and platform_validate_package on the package. Every finding goes in the report; none are merged.
2. Eval compare. Run the eval set against the release target with platform_eval_runs (target set to the dev, staging or production environment, a deployment, or the agent version you will ship) and compare it with the last release's run.
3. Budget check. For the top intents, confirm model calls per turn and latency are within the experience contract, using sessions from the eval run.
4. Context check. Read the full model requests of a 20-turn test conversation, from "Export entire trace" on the session page (see arch-context-budget), for growth, truncation markers and secrets.
5. Checklist. Walk the list below and mark each item pass, fail or not applicable with evidence.

Pre-ship checklist

• Lint and package validation are clean.
• Eval set ran on the release target and was compared with the last release; success held.
• Model calls per turn and latency are within the experience contract for the top intents.
• Every defect fixed since the last release has a scenario in the eval set.
• Each agent declares BEHAVIOR: role:, or you have checked its ## Current Context in a trace; no {{context}} and no unfilled {{placeholder}} in a test trace.
• No config.* or env.* in GOAL, PERSONA, LIMITATIONS, SYSTEM_PROMPT or RESPOND; credentials are in auth profiles; a trace search for secret values finds nothing.
• Handoff CONTEXT: keys are lowercase and history uses a supported form; delegates get everything through INPUT or PURPOSE.
• Every tool with a large or nested result has MODEL_CONTEXT, tested against real, empty and error responses; rich responses carry a response MODEL_CONTEXT.
• Sensitive GATHER fields are marked SENSITIVE: true with a SENSITIVE_DISPLAY mode (and a PII_TYPE so the project's rules for that type apply).
• Every tool declares effect; irreversible writes confirm and freeze key values; each tool the caller depends on has an authored on_error message (see arch-tool-design).
• Workflow tools set mode and timeout_ms explicitly, and the workflow version the release needs is activated (see arch-workflow-design).
• Knowledge-base documents are indexed and max_search_tokens is set on every knowledge-base tool (see arch-knowledge-search).
• INTENTS are declared with descriptions.
• After an import: knowledge bases are populated, and model bindings, auth profiles and config variables are set.
• Final tests ran in a new conversation on the version that will be deployed.
• Voice: ASR and latency sets re-run if the provider, timeouts, profile or gather fields changed; runtime voice settings (fillers, VAD) re-checked after deploy; a fast-tier model is available.
• Speech-to-speech: the effective orchestration mode verified in traces; every tool exit path (success, error, timeout, empty) tested; the welcome is protected from barge-in.