Home / Blog / Claude Code makes a finished subagent send its report twice
Tech News

Claude Code makes a finished subagent send its report twice

An injection defect in Claude Code 2.1.325 forces finished subagents to resend completed reports after monitor timeouts, wasting up to $3 per event.

By Dillip Chowdary • Nov 19, 2026 • Source: gist.github.com

Claude Code makes a finished subagent send its report twice
Image: gist.github.com

At a glance

Product
Claude Code
Issue
Issue 105304
Observed versions
2.1.312, 2.1.322, 2.1.325
Incident frequency
1 in 52 subagent runs
Cost range
$0.001 to $2.81 per incident
Reproduction time
70 seconds

Anthropic shipped an automated injection mechanism inside Claude Code that can mistakenly overrule subagents and force them to deliver duplicate completion reports to parent sessions. According to documentation detailed in gist.github.com's report, issue 105304 causes the client software to inject a mandatory handback instruction after finished background workers receive routine monitor notices. The defect triggers duplicate token consumption across parent and child sessions while overriding correct subagent state evaluations.

This review examines the mechanics behind the injection defect, reproducing steps, financial overhead, affected user tiers, and mitigation steps. The analysis is intended for software engineers, operations teams, and system architects deploying automated multi-agent tasks inside Claude Code builds.

01Enforced handback override and tooling state

The defect occurs when a background subagent successfully completes an assignment and submits its findings using the SubagentHandback tool. The system returns an explicit confirmation that the report was delivered to the caller. However, if the subagent left a background Monitor process active, a later event notice or expiration wakeup triggers the subagent turn again.

When awakened by the monitor notice, the subagent correctly assesses that its work is done and replies in plain text without invoking additional tools. Claude Code then injects a meta-prompt entry stating that the report has not been delivered and demands an immediate call to SubagentHandback. The subagent obeys the injected command, leading to an accepted duplicate report in the primary session.

Testing confirms that the SubagentHandback tool independently tracks delivered reports and normally rejects unprompted duplicate submissions. The injected enforcement string completely bypasses this logic by mandating compliance. Background Bash commands do not trigger this override, isolating the defect to Monitor expirations.

02Stepwise reproduction sequence across seventy seconds

Developers can reproduce the defect in approximately seventy seconds using Claude Code 2.1.325 inside VS Code with WSL2 on Ubuntu 24.04. The process requires initiating a subagent using the Agent tool and loading the Monitor tool via ToolSearch.

  1. Start monitor with command sleep 600, description test watcher, and timeout_ms set to 60000.
  2. Submit handback immediately while the monitor is armed using SubagentHandback with message REPORT-ONE.
  3. Handle wake after monitor expiration by replying in plain text with PLAIN-AFTER-WAKE and invoking no tools.
  4. Submit forced handback after the system injects the handback-send-enforce command demanding a second hand-in.

Note: Testing confirms that stopping the background watcher using TaskStop prior to handing in the initial report prevents the wakeup and eliminates duplicate reporting.

03Measured token overhead and cost impact

A single incident incurs measurable token waste across context re-reads, full response regenerations, and parent processing turns. When a subagent wakes after its five-minute prompt cache has expired, rewriting the context incurs significant fresh write penalties.

Run componentMeasured tokensList price
Forced resend (Opus 5.5)538,736 read / 3,608 output$0.18 median
Parent extra turn (Fable 5.2)752,136 read / 855 output$0.32
Cold cache wake (Opus 5.5)350,006 to 561,291 write$1.75 to $2.81
Forced resend (Haiku 5.5)81,921 read / 197 output$0.001

Transcript scans across 416 subagent runs between October 17, 2026, and November 18, 2026, revealed ten forced second hand-ins across eight distinct runs. Incidents appeared across client versions 2.1.312, 2.1.322, and 2.1.325, yielding an observed frequency of one in every 52 runs.

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

Total direct waste across that monthly sample reached approximately $14.50 at standard list prices. Across high-volume setups, individual incidents range from a tenth of a cent on lightweight models to approximately $3.00 on complex workloads.

04User exposure and subscription billing models

Impact depends directly on how the developer or organization pays for compute. Users operating on flat subscriptions do not pay additional dollars directly, but redundant generation consumes usage caps earlier in billing windows.

Teams billed by raw token counts through direct API keys or metered enterprise plans pay for the redundant report generation directly. Any developer running multi-agent workflows with active background monitoring is susceptible to the enforcement error.

At an estimated 0.05 percent to 0.15 percent share of overall usage, redundant processing scales with agent density. Organizations relying heavily on nested subagent architectures absorb the vast majority of duplicate execution cycles.

05Version management and execution commands

Teams managing agent workflows can update their local CLI installations and manage model assignments directly from the terminal or interactive sessions. Keeping versions aligned ensures runtime fixes can be quickly deployed as patches appear.

Command
npm install -g @anthropic-ai/claude-code
Command
claude update

To switch models dynamically during an interactive session, enter the model change command directly:

Command
/model claude-fable-5-1
Command
/model fable

To start a session with a pinned model target or set standard run defaults, launch the executable directly with explicit flags:

Command
claude --model fable
Command
claude config set --global model fable

Frequently asked questions

What triggers duplicate reports in Claude Code?

An active background Monitor process wakes a finished subagent, causing Claude Code to inject an enforcement prompt that demands a duplicate handback.

Which Claude Code versions exhibit this behavior?

Transcripts document the issue occurring in versions 2.1.312, 2.1.322, and 2.1.325.

How can developers avoid the duplicate handback?

Explicitly calling TaskStop to terminate running monitors before executing SubagentHandback prevents monitor expiration notices from waking the subagent.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →