Agent spends model turns on sleep N; check polling while a background job runs #5459

Describe the bug

While a long background job was running, the agent kept making bash calls of the form sleep N; <tail the log / check the process>. Each call was a separate model turn that re-read the whole conversation context. None of these turns produced anything to act on until the job finished. The runtime already notifies the agent when a backgrounded shell completes, so these turns could have been avoided.

Measured from one interactive session's events.jsonl. I joined session.usage_record to assistant.message by apiCallId:

Model selection: Auto.

This doesn't happen every time. I ran 3 headless ACP runs (copilot --acp --allow-all --no-ask-user --no-custom-instructions, default model). Each was asked to make a check pass, which required a background job of about 250–280 s. In every run, the agent started the job and waited for it in one blocking bash call (initial_wait 300–600), so no turns were spent waiting. However:

• The tool's initial_wait is capped at 600 s. A longer command outlasts that synchronous window, but this does not mean repeated model turns are necessary: the agent could yield and resume on completion.
• These were three short ACP runs, not a controlled before/after demonstration of a fix in the interactive app.
• All six underlying test suites in the two-provider comparison exited nonzero. This measures waiting overhead, not successful task completion.

Affected version

GitHub Copilot desktop app 1.4.1 (confirmed through the serving app's build-info tool), bundled runtime copilot-sdk-runtime-1.0.106-8.unstable.r43360710077 (copilot --version: GitHub Copilot CLI 1.0.112).

Steps to reproduce the behavior

This is not deterministic.

1. In an interactive session, have the agent start a job in the background (async bash) that takes 5–40 minutes, for example a test suite. Ask it to report the results when the job is done.
2. Watch the agent's tool calls while the job runs.
3. In affected runs, the agent makes repeated bash calls like sleep N; tail <log>; <process check>, each costing one model turn.

Expected behavior

Once the only remaining work is waiting, the agent should do one of two things:

• end its turn and resume when the completion notification arrives, or
• make a single blocking wait.

It shouldn't repeat sleep-and-check turns that return nothing to act on.

Additional context

• macOS 27.2, arm64.
• Instructions alone don't seem to fix this. Other agents have similar reports where the agent keeps polling despite explicit instructions: [Bug] Assistant repeatedly issues filler commands while waiting for background tasks despite explicit instruction anthropics/claude-code#97191, Astra/Codex goals: repeated testing and polling despite explicit limits; 454.5M tokens in 24h and major credit depletion openai/codex#56083. Related request in this repo: Feature request: Monitor tool / stream background process output lines as notifications (avoid polling) #5260.
• Possible runtime fixes (suggestions, untested):
  • A blocking "wait until this shell/process exits" without the 600 s cap.
  • Or an explicit way to end the turn and be woken by the existing completion notification.
  • Coalesce redundant waiting-only checks in the runtime and suspend further inference until completion or a meaningful actionable event. Merely denying a call and asking the model to try again can still consume another turn.
  • Preserve cancellation, user intervention, failure notifications and legitimate checks for new actionable information.
• Raw transcripts and private workspace data are intentionally omitted. Cached tokens are reported separately, not treated as equal-cost fresh tokens.