A loop meant to run nonstop had quietly stopped—Claude Code v2.1.292 recovers the "vanishing scheduled jobs" of unattended operation and adds a throttle knob to subagents

On October 6, Anthropic released v2.1.292 of its coding agent "Claude Code." Whereas the immediately preceding v2.1.291 was a small fix that merely reverted issues like session loss, this release is a broader one that bundles new features and security fixes together. The standouts are a cluster of fixes that restore the reliability of "unattended operation"—keeping agents running during hours when no one is watching—and a new feature that adds a fresh knob to subagents.

Scheduled jobs that vanished before anyone noticed

Claude Code has mechanisms for running work repeatedly with /loop, as well as scheduled tasks (routines) that fire at set times. This release closed off several paths by which those "scheduled jobs" had been silently disappearing. The main reported issues were as follows.

• /loop lost the pending next-launch information when a session was restarted, and quietly stopped
• Scheduled tasks carried over from /resume, /branch, or /clear sometimes failed to fire at all
• In cloud sessions, finished routines lingered for hours still showing as "running"
• Editing or duplicating a routine silently turned off the push notifications you had configured
• When using a shared usage limit, each background agent emitted a duplicate limit alert

Each of these is unremarkable in isolation, but they bite hard in unattended operation. If the premise is letting things run autonomously overnight or on weekends, "you can't tell when it stops" and "it finished but you can't tell it finished" undermine trust in the automation itself. This release can be seen as a step toward making the hands-off, let-it-run style of use something you can confidently entrust.

Subagents get a "throttle" knob

Among the feature additions, the one most likely to have a direct impact on workflows is the new effort parameter for the Agent tool. It lets you specify the amount of "work" (degree of reasoning) applied when running a subagent. In a setup where a single parent task hands work out to multiple subagents, you can allocate intensity by role—light, mechanical subtasks get a modest effort, while difficult verification or design gets a high effort. The more agents you run side by side, the less you need to run them all at the same intensity, and the more room there is to differentiate.

Beyond this, the following additions and improvements landed around agent operation.

• Added --marketplace <source> to claude plugin install. You can register a marketplace if needed and then install the plugin
• In Remote Control, claude attach <name> and claude logs <name> now work with a partial match on the session name. You can reconnect by name without copy-pasting a long session ID
• Added the environment variable CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS to adjust the retry interval on 529 (overloaded) errors
• Fixed an issue where @-mentioned text files larger than 256KB were silently ignored; they are now read in chunks

Closing off "swap-it-mid-read" attacks in one sweep

Security fixes carry significant weight in this release, and a common thread runs through them: a class of loopholes where the target file is swapped for something else right after permission is granted. The published fixes include the following.

• An issue where, during notebook/PDF reads, swapping a link mid-read could return a file outside the approved path (macOS/Windows)
• An issue where PreToolUse hook approvals and auto mode skipped the permission check for file reads over network (UNC) paths
• An issue where, when switching the reference target mid-session, the sandbox's read-denied paths failed to fully override the project's allowances
• An issue where a tampered on-disk cache could disable the standard policy plugin when settings retrieval failed

The further the trend shifts permission-granting from humans toward an AI checkpoint, the more any gap between "when and for what you granted permission" and "what was actually touched" becomes an entry point for intrusion. The way such risks pile up behind the convenience is a point worth nailing down before expanding unattended operation. Steady fixes like these—closing hard-to-spot race conditions in a single sweep—may lack flash, but they tie directly to the safety margin of real-world operation.

How it plays out in real operation

If your team has built Claude Code into its automation, the practical payoff of this update is less about "trying new features" and more about scheduled tasks and loops that had stopped starting to run again. In environments that rely on unattended operation, updating sooner reduces the automated processing that was being dropped while silently stalled. The intensity allocation via effort becomes a new design variable for curbing wasted processing, for teams designing development workflows that run multiple agents in parallel. Both are incremental steps toward turning agents into "a foundation you can entrust without constant human supervision."

References: Claude Code official changelog / anthropics/claude-code CHANGELOG / Releasebot: Claude Code updates