ai-powered-markdown-translatorArticle translated from fr to en with gpt-5.4-mini.
Catch-up day: two announcements that slipped through the cracks of the daily scan deserve coverage. Claude Code v2.1.199, released on July 2 between the already covered v2.1.198 and v2.1.200, fixes 23 issues in the agent layer. The same day, NotebookLM had launched Short Video Overviews as early as June 30, a feature that turns any source into a 60-second vertical video.
Claude Code v2.1.199 — 23 fixes for background agent stability
July 2 — Claude Code version v2.1.199 (commit 125d63f) is released with a dense set of fixes. It sits between v2.1.198 (Claude in Chrome in general availability) and v2.1.200 (Manual mode by default), both already covered — hence this catch-up.
Background agents and Linux daemon:
| Fixed issue | Effect before the fix |
|---|---|
| Linux daemon: automatic shutdown every ~50 seconds | All active agents were killed after an unclean shutdown |
| Cold SSH startup on macOS (“Could not switch to audit session”) | Regression introduced in v2.1.196 |
claude stop canceled by a concurrent daemon restart | The agent restarted silently despite the stop command |
Sub-agents and multi-agents:
- Sub-agents interrupted by a rate limit now return their partial work to the parent (instead of failing silently).
- Sub-agent API errors (e.g. usage limit reached) are now properly surfaced to the parent rather than reported as successes.
- Idle agents no longer disappear from the panel: they are grouped into a collapsible row.
SSL, retry, and multi-skills:
- SSL errors (TLS inspection proxies, missing
NODE_EXTRA_CA_CERTS, expired certificates) now fail immediately with guidance on how to fix them instead of consuming retries. - Transient 429 errors not related to the usage limit are now automatically retried with backoff for subscribers.
CLAUDE_CODE_RETRY_WATCHDOG: default counter raised to 300 attempts,CLAUDE_CODE_MAX_RETRIESceiling removed.- Stacked skill invocations (
/skill-a /skill-b do XYZ) now load up to 5 skills at the head, not just the first one.
NotebookLM Short Video Overviews — 60-second vertical educational summaries
June 30 — NotebookLM launches Short Video Overviews, a feature that automatically converts any complex source — document, article, PDF — into a 60-second educational vertical video. The format adopts the style of short social media videos (educational doom scrolling), made possible by advances in on-device video generation integrated into Gemini.
Staged rollout:
| Audience | Availability |
|---|---|
| Google AI Ultra and Pro subscribers | Mobile and web starting June 30 |
| All web users | English only, gradual expansion |
An example posted by the @NotebookLM account features a source about the 2026 World Cup converted into a 60-second video. The feature extends the logic of Audio Overviews (podcast-style audio summaries) by adding a visual dimension.
What this means
Background agent robustness is no longer optional. The volume of fixes in v2.1.199 — Linux daemon crashing every 50 seconds, silent sub-agents, broken macOS SSH, 429 retry — illustrates the complexity of an agentic architecture deployed at scale. These 23 fixes are not convenience improvements: they correct behaviors that made background agents unreliable in production, especially in automated workflows such as overnight runs. The increase of the retry watchdog counter to 300 attempts and the removal of the CLAUDE_CODE_MAX_RETRIES ceiling signal that Anthropic is anticipating long-running workloads with network fluctuations.
Turning a document into video changes how knowledge is consumed. NotebookLM’s Short Video Overviews are not trying to replace reading: they address a different usage context, that of mobile scrolling. Converting a PDF into a 60-second summary with a visual format suited to social networks broadens the range of sources NotebookLM can process. The staged rollout (Ultra/Pro first, then all web users in English) follows Google’s usual pattern for experimental, high-compute-impact features.