Change the instruction. Keep the work.
The problem
A useful result may already be on its way when the answer needs a different format. Interrupting the request can discard that progress.
A recorded correction
A new instruction.
The same result.
While a fictional file read was running, a correction changed the requested answer from a full sentence to the code alone. The read completed and the same stream returned LANTERN-4827.
01 Original request
Read, then write a sentence.
Selected original instruction
Wait for the result, then answer in one full sentence identifying the observed fictional code.
A real read of a fictional file, with a deliberate eight-second wait. No original sentence-format answer was produced.
02 2.56 s · correction queued
Change the answer format.
Exact changed instruction
Keep the observed fictional code exactly, but answer with that code only. No sentence, label, punctuation or markdown.
One tool still running
Queued while the read continued. Applied after its result arrived.
03 Same code, revised answer
Keep what the read found.
Tool returned · 10.57 s
LANTERN-4827Final answer · 13.07 s
LANTERN-4827One read · one correction · same stream.
Recorded isolated execution · real model response · fictional file. The engine source was unchanged; a private allow-once hook and namespace restricted this test to the exact delayed read.
Inspect the recorded sequence
Saved events from 15 October 2026. Times start at the original request; the eight-second wait was deliberate. This page displays the record and starts no service.
| Elapsed | Recorded event |
|---|---|
| 2.48 s | Read startedtool_use |
| 2.56 s | Correction queuedsteer_queued |
| 10.57 s | Tool returnedtool_result |
| 10.58 s | Correction appliedsteer_applied |
| 13.07 s | Stream completedone |
The delivery echo matched the queued correction and left no pending steer. It arrived before the single final SDK result; this run did not take the same-stream follow-up path.
Two model response sequences: one requested the read, the other gave the final answer. The correction used a separate SDK input/control write. One stream does not mean one model response.
Source and test boundaries
Claude Engine v6.17.0, commit bf89abc. SDK 0.2.164, CLI 2.1.292; actual model claude-haiku-5-5.
Bash alone was exposed, but Bash is a broad grant. A private deny-by-default hook allowed this exact command once and replaced it with a fixed launcher. The read ran with read-only fixture access in a private filesystem and network namespace. The trusted CLI and launcher retained access outside that payload boundary.
No external plugins were configured. CLI metadata listed two built-in plugins and emitted a completed local Bash lifecycle event. No agent or Skill tool was called, and no background source remained at completion.
This verifies one corrected engine stream. PanelForge UI/database persistence, deployment availability, multiple corrections or tools, agent steering, voice behavior and universal model compliance were not tested.
What this makes possible
A correction queued during a real fictional-file read changed the answer from a sentence to the code alone. The same engine stream retained the result and returned LANTERN-4827.
Source information sheet
Evidence, implementation & history Implementation inspected
What I checked
On 15 October 2026, one isolated real-model execution used unchanged Claude Engine v6.17.0 source with a private exact-command hook and namespace restriction. One correction queued during the real fictional-file read, then the same stream returned its code alone. No PanelForge UI or database persistence was tested.
Release anchor: claude-engine v6.15.0; panel-forge web-v10.20.37.
Research cut: 15 October 2026. Source areas below are from claude-engine + panel-forge. Private conversation references are retained in the research ledger and are not published here.
Implementation decisions
Claude Engine accepts an additional user instruction on the active stream and marks its delivery explicitly. The recorded run queued one correction while the actual delayed read was in flight.
The successful tool result remained in the application marker. Its matched delivery echo cleared the pending correction before the single final SDK result; the model then answered with the observed code alone. The two model response sequences belong to the same stream.
A path through the feature
- Correct the requested answer
The original instruction asked for a full sentence. One correction requested the exact code alone while the read continued.
- Keep the useful result
The read returned LANTERN-4827. The same stream completed with that exact code, in the revised format.
Source areas
claude-engine/claude_engine/routes/stream_control.pyclaude-engine/claude_engine/session_runtime.pyclaude-engine/claude_engine/invocation.pyclaude-engine/claude_engine/routes/stream.py
The saved correction uses one actual Claude Engine stream and a real model, with a fictional file and a private allow-once command restriction. Two model response sequences and one final SDK result are recorded. This establishes the selected engine lifecycle; PanelForge UI/database persistence and broader steering cases were not exercised.
Recorded history
Dates describe the researched record, rather than today’s deployment.
claude-engine v6.15.0; panel-forge web-v10.20.37
One isolated real-model engine stream: read, queued correction, preserved result and code-only answer. Source pinned at Claude Engine v6.17.0.