v5.0.0-alpha.3 - Mission reliability

Starlight 5.0.0-alpha.3 — mission reliability

This preview adds durable progress inspection to the general-purpose agent platform. A CLI process crash previously could leave an empty reserved report; runs now save atomic snapshots before and after every step, retaining the last recorded results and evidence.

• Progress storage: FileRunStore is enabled by default in the CLI and optional in AgentPlatform. Failed checkpoint writes stop new work and reject the run handle.
• Run history: starlight runs --status failed --limit 20 lists saved summaries across process exits; inspect reads a complete report.
• Mission deadlines: SDK timeoutMs and CLI --timeout-ms bound routing, capacity waits, execution, verification, and inter-step persistence. Deadlines propagate cancellation to remote agents.
• Progress events: SDK subscribe and CLI --events expose immutable lifecycle snapshots. CLI events go to stderr, preserving final JSON on stdout.
• Diagnostics: step timings, final-write retention protection, isolated observer errors, and bounded retries for Windows checkpoint rename locks.

Requires Node.js 22+. Wire protocol 1.0 is unchanged. Existing agents and final reports remain supported. This is an alpha, not a stable or production-readiness claim.

Install this exact preview

Download the attached package and verify it against SHA256SUMS, then:

npm install ./starlight-protocol-starlight-5.0.0-alpha.3.tgz
npx starlight demo --events --timeout-ms 10000
npx starlight runs --status completed --limit 10

This release is distributed as the attached archive; these instructions do not assume it is published to the npm registry.

Validation

The release gate covers 70 tests, lint, TypeScript declarations, schema and documentation checks, 19 protocol conformance cases, authenticated multi-process execution, demo scenarios, website builds, and installation of the packed artifact into a fresh project. The package export, CLI demo, saved report, and history command are verified from that installation. The full dependency audit reports zero vulnerabilities.

Crash tests terminate a CLI process after an external effect and verify the retained checkpoint. Additional tests inject storage faults before and after effects, read concurrently with snapshot replacement, and exercise remote deadlines and late synchronous results.

Recovery boundary

A saved running step is unfinished evidence, not proof of a live process. It may already have caused an external effect. Checkpoints do not automatically resume or replay agents, guarantee exactly-once effects, or roll back work. Reconcile ambiguous effects before submitting a fresh mission.

Run storage, deadlines, and crash semantics · Website and demo · Changelog