Detach rewriter: also wrap sync-mode commands ending in trailing &

[meganrogge]

Fixes microsoft/vscode-internalbacklog#7605

When the chat.tools.terminal.detachBackgroundProcesses setting is enabled, also wrap commands with nohup when the agent passed mode='sync' but the command line ends with a single trailing & (POSIX background operator, distinct from &&).

Previously the rewriter only fired for mode='async' (legacy isBackground: true). When the agent used mode='sync' + &, the & was interpreted by the foreground shell and the process was SIGHUP'd as soon as the tool's terminal was disposed — fine for interactive UX, but it breaks eval harnesses where the eval step connects to the service from a fresh process and gets Connection refused.

Evidence

terminalbench2 run 25136645365 (claude-opus-4.6) had three failures attributable to this:

• hf-model-inference — python app.py > /app/server.log 2>&1 &
• install-windows-3.11 — qemu-system-i386 ... -vnc :1 -qmp ... -monitor ... (no daemonize)
• qemu-alpine-ssh — nohup expect setup_vm.exp > /app/vm_output.log 2>&1 & in sync mode

In all three the agent's same-turn verification (curl, ps aux, ss -tlnp) confirmed the service was alive — it just didn't survive to the eval phase.

Behavior

• Setting still gates the rewrite, so default users are unaffected.
• Eval harnesses already set chat.tools.terminal.detachBackgroundProcesses=true (see vscode-copilot-evaluation/src/vsCodeApplication.ts), so they pick this up automatically.
• PowerShell has no equivalent trailing-& operator, so the Windows path remains async-only.
• Existing nohup/shell -c wrapping of compound constructs and the de-duplication of an already-trailing & continue to work.

[meganrogge]

/requires-eval-assessment terminalbench2 gpt-5.4,claude-opus-4.6,claude-opus-4.7

[vs-code-engineering]

⏳ Queued vscode build for aea2e1ec1586d38083440ceb9f3955d36b5e1a47 (step 1/2).

• Build: https://dev.azure.com/monacotools/Monaco/_build/results?buildId=435085
• When this succeeds, the eval-assessment publish build will be queued automatically.

[Copilot]

Pull request overview

Extends the terminal chat tool’s background-detach command-line rewriter so eval harnesses (which enable chat.tools.terminal.detachBackgroundProcesses) also detach POSIX commands invoked in mode='sync' when the command line ends with a single trailing &, preventing services from dying when the tool terminal is disposed.

Changes:

• Gate rewriter on the detach setting first, then detach when isBackground=true or the command ends with a bare trailing & (POSIX).
• Keep Windows behavior async-only (PowerShell path only rewrites when isBackground=true).

Show a summary per file

File	Description
src/vs/workbench/contrib/terminalContrib/chatAgentTools/browser/tools/commandLineRewriter/commandLineBackgroundDetachRewriter.ts	Expand detach criteria to include sync-mode POSIX commands ending in trailing &; preserve Windows async-only behavior.

Copilot's findings

• Files reviewed: 1/1 changed files
• Comments generated: 2

[vs-code-engineering]

🚀 Queued eval-assessment publish build for dc1d3cbdcb24b73f2b17611a67ef27975fabee97 (step 2/2).

• Pipeline run: https://dev.azure.com/monacotools/Monaco/_build/results?buildId=435108
• On success, publishes @vscode/vscode-copilot-evaluation-agent@0.0.0-dev.dc1d3cbdcb on the dev tag.

[vs-code-engineering]

🔬 Queued eval-assessment benchmark for a6f887c92a.

• Package: @vscode/vscode-copilot-evaluation-agent@0.0.0-dev.dc1d3cbdcb (dev tag)
• Benchmark: terminalbench2
• Tracking issues:
  • gpt-5.4: https://github.com/github/evald/issues/18981
  • claude-opus-4.6: https://github.com/github/evald/issues/18982
  • claude-opus-4.7: https://github.com/github/evald/issues/18983

Results will be posted back here when the run completes.

[vs-code-engineering]

✅ Eval-assessment build published.

• Package: @vscode/vscode-copilot-evaluation-agent@0.0.0-dev.dc1d3cbdcb (tag: dev)
• Install: npm install @vscode/vscode-copilot-evaluation-agent@0.0.0-dev.dc1d3cbdcb
• Pipeline run: https://dev.azure.com/monacotools/Monaco/_build/results?buildId=435108

[vs-code-engineering]

📊 Eval-assessment benchmark complete.

• Tracking issue: https://github.com/github/evald/issues/18983
• Publish build: https://dev.azure.com/monacotools/Monaco/_build/results?buildId=435108

Analysis Results

Resolution Rate

Benchmark	Total Cases	Passed	Failed	Resolved Rate
terminalbench2	89	59	30	66.29%

Token Usage

Metric	Value
Total Tokens	59,731,878
Input Tokens	58,729,020
Output Tokens	1,002,858
Cached Tokens	55,911,102

Step Counts

Metric	Value
Total Steps	1,654
Mean Steps/Instance	18.58

[vs-code-engineering]

📊 Eval-assessment benchmark complete.

• Tracking issue: https://github.com/github/evald/issues/18981
• Publish build: https://dev.azure.com/monacotools/Monaco/_build/results?buildId=435108

Analysis Results

Resolution Rate

Benchmark	Total Cases	Passed	Failed	Resolved Rate
terminalbench2	89	49	40	55.06%

Token Usage

Metric	Value
Total Tokens	72,069,420
Input Tokens	71,143,963
Output Tokens	925,457
Cached Tokens	59,988,608

Step Counts

Metric	Value
Total Steps	1,672
Mean Steps/Instance	18.79

[vs-code-engineering]

📊 Eval-assessment benchmark complete.

• Tracking issue: https://github.com/github/evald/issues/18982
• Publish build: https://dev.azure.com/monacotools/Monaco/_build/results?buildId=435108

Analysis Results

Resolution Rate

Benchmark	Total Cases	Passed	Failed	Resolved Rate
terminalbench2	89	51	38	57.30%

Token Usage

Metric	Value
Total Tokens	68,038,993
Input Tokens	66,592,678
Output Tokens	1,446,315
Cached Tokens	63,330,502

Step Counts

Metric	Value
Total Steps	1,955
Mean Steps/Instance	21.97