Claude Agent: Opus 5.5 shows 3.9x lower cache reuse than native Claude Code #61282

Activity

1. opencodejr commented on Nov 2, 2026

   

   opencodejrBot

   on Nov 2, 2026

   Contributor

   More actions

   Thanks for the report and the numbers! Since they come from different sessions, it's hard to tell whether OpenCode breaks the prompt cache or whether the workloads were just different. Could you share:

   1. Versions: opencode --version and claude --version, plus your OS and how you installed OpenCode.
   2. Setup: which plugins and MCP servers you have enabled, and whether you used subagents.
   3. Compaction: whether the OpenCode sessions were long enough to compact. With Claude Agent, compaction can start at about 170K tokens (see claude-agent: saved model catalog loses context windows, models fall back to 200K and compaction triggers early #61020), and each compaction has to write the cache again.
   4. A small matched comparison, if possible: the same short task in a fresh session in both tools, with the per-request input / cache_write / cache_read numbers. Even 5–10 turns would help a lot. For OpenCode, the per-message token usage in the session (e.g. from opencode export <sessionID>) is enough.

   You can reply here with the details or edit the issue. If you edit it, please also leave a comment saying you've updated it so we know you're done.

   Related issues, ordered from most to least likely to be relevant:

   • claude-agent: saved model catalog loses context windows, models fall back to 200K and compaction triggers early #61020: Claude Agent models fall back to a 200K context window, so compaction starts early and the cache has to be rebuilt more often
   • Provider claude-agent: system text / prompts added by plugin context hooks never reach Claude #61248: on claude-agent, text that plugins add to the system prompt never reaches Claude, so OpenCode's system prompt differs from Claude Code's
   • Configurable prompt-cache TTL per agent and per provider/model (opencode.jsonc) #51109: feature request to make the prompt-cache TTL configurable
2. opencodejr

   added

   triaging

   on Nov 2, 2026
3. Prognosticate-X commented on Nov 3, 2026

   

   Prognosticate-X

   on Nov 3, 2026

   More actions

   Hi @christian-taillon @opencodejr,

   We recently investigated the exact wire-level token prefix divergence between OpenCode and native Claude Code and built an open-source, zero-dependency profiler (agent-cache-lens) to isolate the root causes behind this 3.9x CacheX drop.

   Here is the architectural breakdown and per-turn findings isolating why OpenCode's prompt cache prefix invalidates:

   1. Primary Root Causes Identified

   1. Premature Context Compaction (#61020):

      • Claude Code models support 1M context windows, but OpenCode's stored catalog frequently falls back to 200,000.
      • With RESERVE_SHARE = 0.15, compaction triggers prematurely near 170k tokens instead of ~850k tokens, truncating history and forcing a complete cache rewrite (paying full write price on the entire context).

   2. System Instruction Prefix Invalidation via Plugin Hooks (#61248):

      • Any plugin appending dynamic metadata (e.g. dynamic timestamps, working directories, active Git branches) to event.system shifts the byte-for-byte SHA-256 hash at offset 0.
      • Because Anthropic Prompt Caching hashes the strictly serialized prefix prior to ephemeral breakpoints, modifying the system prompt destroys cache retention for all subsequent turns.

   3. Tool Parameter Dictionary Key Drift:

      • Native Claude Code guarantees deterministic recursive key sorting for all declared MCP / built-in tools. In frameworks where JSON serialization doesn't enforce strict key canonicalization, server reconnections invalidate the prompt cache prefix hash.

   ---

   2. Side-by-Side Controlled A/B Profile

   We ran a controlled 3-turn trace comparing identical tool-use workloads:

   Metric	Native Claude Code	OpenCode (Claude Agent)	Delta / Gap
   CacheX (cache_read / write+input)	1.82x	0.0x	100% Cache Miss in Target
   Cache Hit Rate	64.5%	0.0%	-64.5%
   Cache Read Tokens	17,420	0	-
   Cache Write Tokens	9,300	26,820	+17,520
   Total Incurred Cost	$0.0611	$0.1745	+185% cost inflation

   ---

   3. Tooling for Verification

   If anyone wants to isolate where their own session diverges, you can pipe your session export through agent-cache-lens:

   # Pinpoint exact byte/character where the prefix hash broke
   agent-cache-lens analyze <opencode_session_export.json>
   
   # Or run the live proxy sniffer to see turn-by-turn cache retention in real-time:
   agent-cache-lens proxy --port 8080 --upstream https://api.anthropic.com

   Hope this forensic data helps isolate the fix for #61282!

   👍React with 👍1Reacted by christian
4. opencodejr

   closed this as not plannedon Nov 5, 2026
5. opencodejr commented on Nov 5, 2026

   

   opencodejrBot

   on Nov 5, 2026

   Contributor

   More actions

   Closing this for now because we haven't heard back with the details we asked for. Without the OpenCode and Claude Code versions, your setup, and a matched per-request comparison of input / cache_write / cache_read, we can't tell whether OpenCode is breaking the prompt cache or the sessions just had different workloads.

   If you can share those details, please reply here or open a new issue and we'll take another look. Early compaction on Claude Agent may account for part of the gap; that's tracked in #61020. Thanks for the report!