[FEATURE]: Warn users before large potentially uncached requests, not only through debug logs #60945

Feature hasn't been suggested before.

• I searched existing issues. Related issues are linked below; this request specifically concerns user-facing warnings and confirmation before potentially expensive uncached requests.

Describe the enhancement you want to request

Problem

A large session can silently lose prompt-cache reuse after changes to instructions, tools, model settings, or provider-side cache expiry/eviction. The next request may process the entire context as uncached input, causing unexpected cost or subscription-quota consumption.

PR #58801 proposes debug logging for cache-miss causes. This helps diagnose spending afterward, but does not give users an opportunity to avoid it.

Observed example

In OpenCode 2.2.5, using openai/gpt-6.1-sol:

1. A long session had approximately 283k tokens of context, mostly served from cache.
2. Disabling a plugin updated the available skill instructions.
3. The next request recorded 282,498 uncached input tokens and zero cache-read tokens.
4. There was no user-facing warning before that request.

The timing is consistent with instruction changes invalidating the cached prefix; the precise provider-side cause was not independently confirmed. These figures were obtained from local session records. They do not establish a conversion from token counts to ChatGPT quota percentage.

Proposed behavior

Surface cache-loss risk in the UI rather than requiring debug logs.

Before sending, when OpenCode detects a likely prefix-breaking change and the affected context exceeds a configurable threshold:

Prompt-cache reuse may be lost

Instructions changed. The next request may process approximately 283k input tokens without cache reuse, increasing cost or quota consumption.

Continue · Review options · Cancel

Review options could include:

• Start a fresh session with a short, user-approved handoff.
• Compact the current session, explicitly warning that compaction itself may send the full context and incur uncached input usage.
• Inspect the triggering change.

For cache misses only discoverable after the provider responds, show a visible notice with uncached tokens, cached tokens, and the suspected cause.

Acceptance criteria

• Known prefix-changing events can trigger a warning before the affected request, including automatic agent continuations.
• Warnings distinguish observed changes, suspected expiry, and confirmed cache misses; they do not promise knowledge of provider cache state.
• Show estimated affected input tokens; display monetary estimates only when reliable pricing is available.
• Never infer a ChatGPT quota percentage from token counts.
• Compaction is not described as free or guaranteed to avoid the immediate cost. OpenCode's compaction documentation says summary compaction sends the full conversation and can reuse its prompt cache; once that cache is cold, summarizing can itself incur the large input cost plus summary output.
• Warning thresholds and confirmation behavior are configurable; repeated warnings are deduplicated.
• Initial cold requests are distinguished from unexpected loss of an established cache.

Related

• feat(core): log why a step missed the prompt cache (debug) #58801 — Debug logging for prompt-cache misses
• [v2] No idle-time compaction: a large session is re-sent uncached after its prompt cache expires #50777 — Large sessions re-sent uncached after idle cache expiry
• [FEATURE]:Token Usage & Cache Hit Dashboard per Session #42295 — Session token usage and cache-hit dashboard

This issue was drafted with AI assistance using local session evidence and reviewed by the requesting user.