Claude Code Queued Input Disappears? Avoid the Drop

If text you submit while Claude Code is working disappears when the turn ends, do not assume Claude received it. For an important instruction, press Esc once to interrupt the current response, wait until the prompt is ready, then send the complete instruction as a new message. Verify that it appears as a user turn and that Claude acknowledges it before allowing a write, deployment, deletion, approval, or external message.

This is a workaround for an open Claude Code interactive-terminal issue, not a confirmed fix. Issue 85603 reports both invisible drops and a second form where the queued line stays rendered but does not start a turn. The report covers Claude Code 2.1.220 and 2.1.226 on macOS inside tmux. Earlier issue 20431 describes the same broad symptom: input submitted during an active turn vanishes from the composer but is never delivered.

Recognize the Queued-Input Failure

The failure has a narrow sequence:

1. Claude is still producing a response or running tools.

2. You type a follow-up and press Enter.

3. The text leaves the editable prompt, so it appears queued.

4. The current turn ends normally.

5. No new user turn starts, or the text remains rendered without being consumed.

The strongest test is the conversation record. If the text never appears as a user message and Claude never acts on it, treat it as undelivered. Do not infer delivery from the fact that the composer cleared.

This is different from Claude ignoring an instruction that is visible in the transcript. It is also different from pressing Esc to interrupt a response. Anthropic's interactive-mode documentation says one Esc interrupts Claude while keeping work completed so far. The open issue explicitly distinguishes its normal end_turn losses from that interrupt path.

Use the Interrupt Path for Important Instructions

Until the issue is resolved, avoid submitting a decision-shaped or side-effecting instruction into a running turn.

Use this sequence instead:

1. Copy the instruction somewhere safe if retyping it would be costly.

2. Press Esc once.

3. Confirm the running response has stopped and the normal prompt is editable.

4. Paste or type the complete instruction.

5. Press Enter once.

6. Confirm the instruction appears in the conversation as a user message.

7. Require Claude to restate the requested action before it performs an irreversible step.

The current issue reporter says the interrupt-then-type path continued to feed input on affected panes. That is useful field evidence, but it remains a workaround reported by a user rather than a documented Anthropic resolution.

Be careful when another dialog is open. Anthropic documents that Esc closes a permission dialog instead of interrupting Claude. Check what changed on screen before pressing it again. Also avoid Esc twice: the documented double-Esc behavior clears a draft or opens rewind, which is a different operation.

Recover a Message That May Already Be Lost

If the turn has ended and your instruction is absent from the conversation, assume it did not arrive. Do not wait for a hidden queue to repair itself.

Resend one explicit replacement:

Replacement for a message that may have been dropped:

<complete current instruction>

For a harmless question, sending a new message may cause an older queued line to flush, as other issue reports describe. That is not a safe recovery for approvals or contradictory decisions: an old line and a new line can arrive together or out of the order you intended. State which instruction is current and require confirmation.

If the missing text authorized an external side effect, cancel the pending action until the replacement appears in the transcript. A recovered fragment is not evidence of current approval.

Capture One Safe Reproduction

Preserve a small evidence set before changing the environment:

claude --version
date -u +%Y-%m-%dT%H:%M:%SZ
uname -a

Record whether Claude was running a tool, streaming text, executing a Stop hook, or fully idle when you pressed Enter. Note whether the line disappeared or remained visibly rendered. If you use tmux, record its version and whether the same pane strands more than one message.

Reproduce only with harmless text such as:

When you finish this turn, reply with exactly: queue probe received

Stop after one controlled test. Do not use a deploy command, approval, secret, customer message, or destructive instruction as the probe. Keep private paths, prompts, and transcript contents out of a public issue comment.

Update Claude Code and Retest Separately

Issue 85603 remained open when this guide was researched on August 10, 2026, and the report reproduced on two nearby builds. Check the current changelog, update through your supported installation method, restart Claude Code once, and repeat only the harmless probe.

A successful probe on a newer version does not prove the bug is fixed. A failed probe is more useful when the version, terminal, pane state, and exact timing are preserved.

Do not change the terminal, tmux version, hooks, settings, and Claude Code version all at once. One-variable tests make the result actionable.

Do Not Repair the Internal Queue by Hand

Avoid editing Claude Code session JSONL, terminal buffers, or internal state files to force a missing message into the conversation. Those files are evidence, not a supported input channel. Hand edits can create a conversation state that the running process never accepted.

Also avoid these weak fixes:

• repeatedly pressing Enter, which can deliver stale and current text together;

• sending a second approval as a queue "kick";

• restarting every pane before preserving one clean reproduction;

• treating a visible queue head as proof that the model received it;

• continuing a sensitive workflow without checking the transcript.

The safe boundary is simple: a consequential instruction is valid only after it appears as a user message in the active conversation and the agent confirms it.

Monitor the Issue With Aident Loadout

You can check the public issue without copying GitHub credentials into the prompt. Start with:

Follow https://aident.ai/SETUP.md

Then ask your agent:

Check Aident Loadout authentication and Vault status. Discover the current read-only GitHub issue Action, inspect its schema, and preflight a lookup for anthropics/claude-code issue 85603. If it is valid and free, execute it and return only state, updated time, labels, comment count, title, and canonical URL. Do not comment, react, edit, close, or create anything.

The measurable result is a current issue-state record with zero GitHub writes. If the issue closes or the changelog names a fix, rerun the harmless probe before returning to mid-turn queued instructions.

Set up Aident Loadout and monitor the queued-input issue.

Sources

• Claude Code issue 85603: typed input queued mid-turn is silently dropped

• Claude Code issue 20431: queued user input silently lost

• Claude Code interactive mode and keyboard shortcuts

• Claude Code changelog

Refresh this guide when issue 85603 changes state, Anthropic documents queued-input delivery, or a current release changes Esc, turn-end, or prompt-queue behavior.