[Bug] Background auto-trigger permanently stops after 3 consecutive LLM timeouts, leaving long-running tasks unproced

Problem

When a long-running background Bash task completes, the shell auto-triggers a foreground Soul run to process its results.

If this LLM request fails with APITimeoutError 3 times in a row, the shell permanently stops listening for background completions:

_MAX_BG_AUTO_TRIGGER_FAILURES = 3

This creates a hard stop.

As a result:

• A test running overnight completes successfully.
• The LLM request to analyze the results times out, often due to large accumulated output/context.
• After 3 timeouts, Kimi never retries.
• The user comes back to find Request timed out spam and zero progress, even though the task finished hours ago.

---

Root Cause

From source analysis, in src/kimi_cli/ui/shell/__init__.py, the main loop does:

if bg_auto_failures >= _MAX_BG_AUTO_TRIGGER_FAILURES:
    result = await idle_events.get()

At this point, the shell only waits for manual/user idle events and no longer listens for background completions.

There is no time-based recovery. Once 3 failures occur, the only way to resume is manual user input.

---

Suggested Fix

Add a cooldown timer.

After hitting the max failures, stop auto-triggering for a configurable duration, for example 10 minutes, then reset the counter and resume listening:

_BG_FAILURE_COOLDOWN_S = 600  # 10 minutes

Example logic:

if bg_auto_failures >= _MAX_BG_AUTO_TRIGGER_FAILURES:
    if time.monotonic() - bg_last_failure_time > _BG_FAILURE_COOLDOWN_S:
        bg_auto_failures = 0
        result = await bg_watcher.wait_for_next(idle_events)
    else:
        result = await idle_events.get()

This matches the behavior of other agentic CLIs, such as Claude Code's periodic progress checks.

---

Environment

• kimi-cli version: 1.41.0
• OS: Windows 10/11
• Model: kimi-for-coding / k2.6

---

Workaround

We applied the above patch locally and confirmed that it works.

The patch is minimal, around 6 lines changed, and does not affect normal operation.

Start in src/kimi_cli/ui/shell/init.py at the background auto-trigger failure counter and the branch that waits on idle_events after three failures. Add the cooldown behavior described in the issue, then verify that background completions are ignored during the cooldown and listened for again afterward.