Description

The prompt loop in prompt.ts reloads all messages from the DB (filterCompactedEffect) at the start of every iteration. Between tool-call steps, tool parts transition from pending to completed with output text. When toModelMessages() serializes the conversation for the next API call, the previous assistant message has different bytes:

• Before tool execution (step N): state: "output-error", errorText: "[Tool execution was interrupted]" (message-v2.ts line 912)
• After tool execution (step N+1): state: "output-available", output: <actual output> (message-v2.ts line 853)

Anthropic's prompt cache sees different bytes at that message position → invalidates everything after it → the entire remaining context is re-written at cache-write pricing ($6.25/MTok for Opus, 12.5× the cache-read price).

Cost impact from real session data (Opus 4.7, 1M context):

• 63% of total API spend is cache writes ($2,264 out of $3,584 on April 21st)
• 95% of rapid cache busts (<60s gap) have a tool call in the preceding message
• Warm turns: cache_read=614K, cache_write=1K — nearly perfect caching
• Bust turns: cache_read=54K, cache_write=560K — only system prompt survives
• Sessions with heavy tool use see 20-84% bust rates

This is likely the same root cause as #20565 (same snapshot hash, mixed cache behavior) and related to #20110 (prompt cache invalidation class).

Plugins

Lore (opencode-lore) — but the issue is in OpenCode core, not the plugin. Lore's gradient transform returns input.messages unmodified at layer 0.

OpenCode version

v1.14.29 (latest dev)

Steps to reproduce

1. Start a session with Claude Opus 4.7 (1M context)
2. Have the model make tool calls (read files, run bash, etc.)
3. Observe token usage: warm turns have cache_write ~1K, but turns immediately after tool execution have cache_write ~500K+
4. Query the OpenCode DB: SELECT cache_read, cache_write FROM message WHERE session_id = '...' AND role = 'assistant' — the alternating pattern of warm/bust is visible

Screenshot and/or share link

Cache pattern from DB analysis:

msg#  cache_read  cache_write  state
1     0K          595K         COLD (first turn)
2     595K        1K           warm ✓
3     54K         542K         BUST (after tool call)
4     595K        1K           warm ✓
...
22    54K         562K         BUST (after tool call)
23    54K         559K         BUST (after tool call)
24    613K        1K           warm ✓

Operating System

Ubuntu 24.04

Terminal

OpenCode TUI