Rewrite a Runtime Guide for Agent Retrieval

Problem

The runtime guide is accurate but expensive to use. Agents struggle to identify the facts that matter for current operation and sometimes investigate irrelevant parts of the document.

Rewrite the guide for agents that need to start, inspect, repair, or integrate with the current runtime. Keep the operational contract complete. Remove context that does not help those tasks.

Output

Produce:

• docs/runtime.md: the rewritten source of truth
• doc-report.md: a compact account of what changed and how the rewrite was verified

Keep docs/runtime.md concise and structured. Do not invent capabilities, commands, paths, or configuration.

Input

=============== FILE: docs/runtime.md ===============

Runtime Notes and Background

The runtime changed several times. In 2023 the team used CloudWatch for run diagnostics, then experimented with Splunk, and moved run records and artifacts onto the host during a 2025 cost-control project. Old tickets and screenshots contain references to each design.

The scheduler is started with pnpm run scheduler. The process entrypoint is src/scheduler/main.ts, while run persistence is implemented in src/runs/repository.ts. Run records are stored in the SQLite file named by RUN_STORE_PATH. Generated artifacts are written under ARTIFACT_DIR, and completion or failure notices are sent through the SMTP relay in SMTP_URL.

The host inventory also records that SENTRY_DSN, PAGERDUTY_TOKEN, DATADOG_API_KEY, PROMETHEUS_URL, OTEL_EXPORTER_OTLP_ENDPOINT, and LOKI_URL are unset. Sentry, PagerDuty, Datadog, Prometheus, OpenTelemetry, and Loki were removed, rejected, or never deployed. The runtime does not include hosted log search, metrics scraping, or a telemetry collector.

Calling curl http://127.0.0.1:6120/status returns scheduler and run-store status. When logs contain artifact write failed, check permissions on the configured artifact directory and retry the run with pnpm run scheduler:retry -- <run-id>.

There is no live log-streaming endpoint. Clients that need new output must page GET /v2/runs/:id/logs?after=<cursor> every 5 seconds.

A 2025 retrospective describes the host migration as the point when the runtime became smaller than earlier plans. =============== END FILE ===============