Description
The built-in view tool reports Path does not exist for an existing text file with Copilot CLI 1.0.73. The same isolated Vally/Copilot SDK probe succeeds consistently with CLI 1.0.71.
This appears to have started in 1.0.72 and remains present in 1.0.73.
Controlled reproduction
The probe uses:
- Official public npm packages
@github/copilot@1.0.71 and @github/copilot@1.0.73
@microsoft/vally-cli@0.7.0
- The Copilot SDK's supported
COPILOT_CLI_PATH override, pointing to each package's platform index.js
- The same model, prompt, skill, marker file, and grader for both versions
Minimal skill:
---
name: view-probe
description: Verify that Copilot can read a staged text file with its built-in view tool.
---
1. Use the built-in `view` tool exactly once to read `references/marker.txt`.
2. Do not use another file-reading tool.
3. Reply with the exact marker value.
references/marker.txt:
The skill is staged into the agent workspace, and the prompt invokes it. Each version was run three times.
Results
| Copilot CLI |
Result |
view outcome |
| 1.0.71 |
3/3 passed |
Returned VIEW_PROBE_MARKER_8A4F21 |
| 1.0.73 |
0/3 passed |
Path does not exist |
A final one-trial run preserved both workspaces. Immediately after each run, the exact absolute path supplied to view existed:
Copilot CLI 1.0.71
FileExists: true
tool.execution_complete.success: true
result: VIEW_PROBE_MARKER_8A4F21
Copilot CLI 1.0.73
FileExists: true
tool.execution_complete.success: false
error: Path does not exist
Sanitized failing event:
{
"type": "tool.execution_start",
"data": {
"toolName": "view",
"arguments": {
"path": "C:\\probe\\workspace\\skill\\references\\marker.txt"
}
}
}
{
"type": "tool.execution_complete",
"data": {
"success": false,
"error": {
"message": "Path does not exist",
"code": "failure"
}
}
}
Additional CI evidence
We also observed the version transition on GitHub-hosted Ubuntu runners:
- CLI 1.0.71: 175 of 208
view calls succeeded.
- CLI 1.0.72: all 193 observed
view calls failed with Path does not exist.
- CLI 1.0.73: all three captured
view calls in another run failed with the same error.
The affected paths were accessible in the same agent sessions through glob, PowerShell Get-ChildItem, and Get-Content.
Expected behavior
view should return the contents of an existing readable file, as it does in 1.0.71.
Impact
Agents cannot read staged skill references and other workspace files. They may improvise without the reference, retry with additional tools, exceed turn/tool/token budgets, or produce incorrect results. This causes broad false failures in automated evaluations.
Environment
- Copilot CLI: 1.0.71 (working), 1.0.72/1.0.73 (affected)
- Copilot SDK: 1.0.6
- Vally CLI: 0.7.0
- Controlled reproduction: Windows x64
- Additional observations: GitHub-hosted Ubuntu runners
- Model:
claude-sonnet-4.6
Description
The built-in
viewtool reportsPath does not existfor an existing text file with Copilot CLI 1.0.73. The same isolated Vally/Copilot SDK probe succeeds consistently with CLI 1.0.71.This appears to have started in 1.0.72 and remains present in 1.0.73.
Controlled reproduction
The probe uses:
@github/copilot@1.0.71and@github/copilot@1.0.73@microsoft/vally-cli@0.7.0COPILOT_CLI_PATHoverride, pointing to each package's platformindex.jsMinimal skill:
references/marker.txt:The skill is staged into the agent workspace, and the prompt invokes it. Each version was run three times.
Results
viewoutcomeVIEW_PROBE_MARKER_8A4F21Path does not existA final one-trial run preserved both workspaces. Immediately after each run, the exact absolute path supplied to
viewexisted:Sanitized failing event:
{ "type": "tool.execution_start", "data": { "toolName": "view", "arguments": { "path": "C:\\probe\\workspace\\skill\\references\\marker.txt" } } } { "type": "tool.execution_complete", "data": { "success": false, "error": { "message": "Path does not exist", "code": "failure" } } }Additional CI evidence
We also observed the version transition on GitHub-hosted Ubuntu runners:
viewcalls succeeded.viewcalls failed withPath does not exist.viewcalls in another run failed with the same error.The affected paths were accessible in the same agent sessions through
glob, PowerShellGet-ChildItem, andGet-Content.Expected behavior
viewshould return the contents of an existing readable file, as it does in 1.0.71.Impact
Agents cannot read staged skill references and other workspace files. They may improvise without the reference, retry with additional tools, exceed turn/tool/token budgets, or produce incorrect results. This causes broad false failures in automated evaluations.
Environment
claude-sonnet-4.6