[BUG]: Terminal streaming renderer corrupts output - characters doubled/truncated during streaming

[Richard-Marlow]

Description

The Copilot CLI terminal renderer corrupts streamed output, producing doubled characters, truncated tokens, and repeated lines. This affects both the thinking/reasoning phase output and the final assistant response output.

Steps to Reproduce

1. Ask Copilot CLI any question that produces a moderately long response
2. Observe the output as it streams to the terminal

Observed Behavior

Output is visibly corrupted during streaming. Examples seen:

• Partial tokens rendered before full token is available (e.g. apil instead of apiVersion)
• Characters doubled mid-word (e.g. obserrvability, permpermissions)
• Words repeated mid-sentence (e.g. New ReliNew Relic, useruser)
• Lines repeated (e.g. metadata: block appearing twice)
• Thinking/reasoning tokens streamed visibly with same corruption artifacts

Example corrupted output observed:

apil
apiVersion: rbac.authorization.k8s.io/v1
...
- apiGroups: [piGroups: [""]

Expected Behavior

Output should render cleanly with no corruption, doubled characters, or repeated lines.

Environment

• Copilot CLI version: 1.0.61
• OS: Darwin (macOS)

Notes

The underlying content being generated is correct — the issue is purely in how the CLI streams and renders tokens to the terminal.

[Richard-Marlow]

This issue was observed in iTerm2 Build 3.6.11

[will-j-wright]

I've been having the same issue on Windows and Linux for 3 or 4 days now.

[tyleris-irish]

Experiencing same on Ubuntu using ghostty

[mike-gorman-quest]

Confirming this reproduces under a different terminal stack as well: Electron 38 + xterm.js 5.5.0 (via agent-wrangler), Copilot CLI v1.0.61, macOS. So the corruption is definitely on the CLI side, not in any particular emulator's escape-sequence handling.

A clean reproducer I captured by copying directly out of the terminal buffer mid-stream (so this is the actual cell contents, not a transient render):

◐ I'm tra'm tracing the cing the data flow through the terminal sysl system:tem: node-pty captures data in the main process, bufprocess, buffers it as afers it as a string, and sends it via IPC to the rende renderer, rer, which then writes it tit to xtermerm.js.

The corruption paorruption pattern doesn't look like typical UTFk like typical UTF-8 mojibake from buffer-to-string conversion — it's more like text duplitext duplication or interle interleaving, suggesting old content isn't being properly ceing properly cleared. Looking at the specifi at the specific examples like "File ile range" and the "Incorrect crrrect cross-reymore" issue, it seems like Ink might be redrawing lines by moving the cursmoving the cursor up and writing new content with new content without fully clearing the old text, and xtermt, and xterm.js is rendering these interse intermediate states where both old and new conand new content are visible simultaneouslyy.

The pattern is consistent across the sample: each emitted chunk begins one or more characters before the end of the previous chunk and appends the remainder of the new token, e.g. I'm tra → 'm tracing the c → ing the data flow. The screen-buffer result is the literal concatenation of those overlapping fragments, with no cursor-back / \b / \r / CSI <n> D / CSI <n> G between them. So either the streaming renderer is emitting the deltas with a stale start offset, or it's omitting the cursor-positioning escape that should precede each rewrite.

The text always "heals" once streaming stops and the final frame is drawn, which fits a model where the per-token redraws use a wrong cursor anchor but the post-stream full redraw uses the right one.

---

🤖 This comment was posted on behalf of @mike-gorman-quest by GitHub Copilot CLI (Claude Opus 4.7 / Anthropic).

[nilcolor]

$ copilot --version
GitHub Copilot CLI 1.0.61.

MacOS, zsh + Kitty terminal

Same issue. Output always heals at the end. And may also heal itself during execution