view tool fails on large single-line tool output — model never sees actual content

[tsinkai]

Summary

When a custom tool returns large output as a single JSON line (e.g., 55KB), the SDK's large-output handling creates an unrecoverable loop where the model can never access the actual content.

Reproduction

1. Register a custom tool (e.g., KnowledgeSearch) that returns ~55KB of JSON as a single line (common for search/retrieval APIs)
2. The model calls the tool
3. SDK writes the result to a temp file and returns to the model:

   Output too large to read at once (55.3 KB). Saved to: C:\Users\...\copilot-tool-output-xxx.txt
   Consider using tools like grep (for searching), head/tail (for viewing start/end),
   view with view_range (for specific sections), or jq (for JSON) to examine portions of this file.
   

4. Model calls view with view_range=[1, 120] → since the entire content is one line, this returns the whole 55KB line, which gets truncated again:

   [Output truncated. Use view_range=[1, ...] to continue reading.]
   

5. Model calls grep with keyword patterns → every match returns the same giant single line, which also gets redirected to yet another temp file:

   Output too large to read at once (55.4 KB). Saved to: ...copilot-tool-output-yyy.txt
   

6. Model never sees the actual content. It burns 5-8 extra LLM turns trying different view/grep combinations, all hitting the same truncation wall.

What the model actually sees (from LLM request dumps)

Tool Response	Content Length	What model receives
KnowledgeSearch (custom)	845 chars	File path + "use grep/view" message
view (lines 1-120)	66 chars	[Output truncated. Use view_range=[1, ...] to continue reading.]
grep (keyword search)	445 chars	Another file path + "use grep/view" message
view (of grep output)	66 chars	[Output truncated...] again

The model operates essentially blind — it never sees the tool results.

Root Cause

The SDK uses line-based chunking for view (view_range operates on line numbers). When the entire output is a single line, view_range=[1, N] always returns the full content regardless of N, which always exceeds the truncation threshold.

Suggested Fix

• Pretty-print JSON tool output before writing to the temp file (so view_range works on individual lines)
• Or use character-based chunking as fallback when the file has very few lines
• Or allow view to accept a character/byte range for single-line files

Environment

• github-copilot-sdk versions tested: 0.1.23, 0.1.30
• Model: gpt-5.2 (gpt-52-2025-12-11)
• Platform: Windows 11
• Provider: OpenAI (via BYOK)

Impact

Custom tools that return large JSON responses (search APIs, knowledge retrieval, connector actions) are effectively unusable. The model wastes 5-8 extra LLM turns and ultimately answers without having seen the actual data. This is a blocker for agentic workflows that depend on tool-grounded responses.

[SteveSandersonMS]

Thanks @tsinkai. We've tracked this issue in the underlying runtime implemention - it will need to be fixed there.

[garypretty]

@SteveSandersonMS could you provide a link to where this is being tracked in runtime so we can track please? Thanks!

[MRayermannMSFT]

Checking in a change related to this into the CLI:

1. For large tool outputs under 5MB, if the output looks like JSON we'll pretty-print it so the file is easier to read in chunks using the built in "view" tool.
2. When using "view" tool, we'll try to see if file is single line, and ultimately if all we can find during the attempt is a single line and we couldn't read all of it, we'll suggest the model use different tools to read the file.

[ahimberg]

Maybe related to that change, after upgrading to SDK 0.2.0 I'm seeing intermittent errors that the tool 'view' is unavailable, when I disable all the built in tools. Maybe 'view' is one that must always be enabled now.