v0.46.3

Drafted by GPT-6.1-sol

This release adds resumable transcript uploads for Claude Code and Codex and reduces repeated background work on machines running many agent sessions.

Breaking changes

• Changed: Self-hosted deployments must provide OpenAI access to gpt-6-luna for fresh session summaries, or set monitors.summarize.model to an available model before upgrading. Existing overrides remain effective. Watcher Cloud needs no action. See the included upgrade guide for details.

Live

• Added: Claude Code and Codex transcript uploads resume from acknowledged checkpoints, including after interrupted uploads and conversation rewinds.
• Fixed: Agent sessions share one background recovery process, reducing duplicate recovery work. Individual recovery passes can still be expensive.
• Fixed: Repeated upload attempts that fail or find nothing new are throttled, and hooks load fewer modules at startup. End-of-turn uploads remain unthrottled.

API

• Added: The ingest API and Python SDK support resumable transcript uploads. Deploy the updated server before updating clients, remote collectors or SDK uploaders.
• Changed: Fresh first-session and idle-session summaries default to openai/gpt-6-luna. They send summary input to OpenAI instead of Google unless an operator overrides the model.
• Changed: Stored-transcript grading on Claude models can reuse more cached input across widely spaced grades.

Other

• Changed: The bundled Anthropic SDK sends anthropic-version: 2026-10-21 by default. Self-hosted deployments using a gateway that restricts this header should confirm it accepts that version before upgrading.