Desktop UI: agent keeps running & billing after UI freeze; stop button unresponsive; forked sessions multiply input costs

[2602065761]

Desktop UI: agent keeps running & billing after UI freeze; stop button unresponsive; forked sessions multiply input costs

Environment

• DSH Desktop 0.1.0 (Windows, DeepSeek-Harness-Desktop-0.1.0-Setup.exe)
• Host: Node v24, profile web, plugins: @liustack/modlens, @liustack/modsearch
• Date: 2026-08-16 (Asia/Shanghai)

Summary

Three related issues that combined into a surprise 15 CNY (~$2) bill for a single "failed" turn, draining my balance to negative:

1. UI freeze while the agent keeps running in the background. The desktop UI showed the turn as stuck/failed, the Stop button did nothing - but the host actually completed all 17 steps of that turn (verified in session.jsonl.zstd: full assistant/message, tool/call, tool/result events from 18:52 to 19:12). Every step was billed normally. The user sees "failure" while the API bill keeps growing.

2. Session forking duplicates the accumulated context. One conversation exists as 3 sessions, each carrying the full ~370K-token history. Resuming/continuing a session appears to seed a copy rather than reuse, tripling the base context that gets re-sent (and billed) on every step.

3. Aggressive billing defaults with no cost guardrails. Fresh install defaults to deepseek-v4-pro + reasoningEffort: max, with no budget cap, no low-balance warning, and no visible per-turn cost. Combined with the 1M context window (no auto-compaction), each step re-sends the whole history at full input price once the prompt cache expires (after a ~12h gap, cache hit rate dropped from ~99% to 0%).

Steps to reproduce (issue 1)

1. Have a long session (~370K tokens of context, with images attached).
2. Leave it idle for ~12 hours (prompt cache expires).
3. Send a message; while the turn is running, the UI may freeze - clicking Stop has no effect.
4. The turn continues server-side to completion; all steps are billed.

Evidence

Session log stats (from ~/.dsh/storages/session_projcache.json):

turns: 21, steps: 190, llm calls: 184
uncachedInputTokens: 5,661,001
cacheReadTokens: 32,924,928
outputTokens: 226,856

Official usage export for 2026-08-16 matches:

deepseek-v4-pro: 182 requests, 5,560,957 cache-miss input,
                 33,151,616 cache-hit input, 205,668 output -> 18.75 CNY

The afternoon turn (18:51-19:12, 17 steps) alone cost ~15 CNY: every step re-sent a 370K-token context at the cache-miss price because the overnight gap had expired the prompt cache, while the UI showed nothing but a stuck turn.

Expected behavior

• Stop button reliably aborts the in-flight turn (host cancels remaining steps; no further API calls).
• UI state reflects the actual host state, or at least recovers (replay event stream) instead of showing a silent dead turn.
• Resuming a session should reuse it, not fork a copy with duplicated history.
• Cost guardrails: opt-in default to a cheaper model/effort, a budget cap, or at least a low-balance warning before a turn starts.

Workaround (for other users hitting this)

• In ~/.dsh/settings.yaml (and the web profile's copy) set:
  agent-default-model: { model: deepseek-v4-flash, reasoningEffort: medium }
• Start a new session instead of continuing an old, huge one after a long idle gap.
• If the UI freezes, don't resend the message - each resend re-bills the full context.

Happy to provide the full session.jsonl.zstd exports privately if needed.

[NovaRouteAI]

@liustack the multiply-billing-after-freeze pattern usually means each forked session holds its own stream and the parent never receives the cancel.

Two things that helped us reproduce similar cases: tag every fork with a parent session id at spawn, and make the stop path idempotent per tag so the second press actually propagates.

Did you see the forks keep streaming server-side, or was the billing purely from local retries?