OpenCode V2: compaction requests miss the prompt cache because the ruleset is only added in the context hook #1086

Summary

On OpenCode V2, the plugin adds the ruleset to the system prompt in the context hook only. OpenCode runs a separate compaction hook for the summary compaction request, so that request has a different system prompt from the main-loop requests. On providers with prompt caching, the compaction request then misses the cache after the tools. The whole conversation is billed again as new input or as a new cache write.

Location

.opencode/plugins/ponytail.mjs in 5.1.0:

await ctx.session.hook('context', (event) => {
  const mode = readMode();
  if (mode === 'off') return;
  event.system.push({ type: 'text', text: getPonytailInstructions(mode) });
});

Measured

OpenCode built from source (v2), Claude Haiku 5.5 over the Anthropic Messages API, 3 turns with tool calls, then /compact. Request bodies captured with a logging forwarder.

The captured compaction request had no ponytail block in its system prompt. The main-loop requests had it.

Suggested fix

Register the same handler on both hooks:

const addRuleset = (event) => {
  const mode = readMode();
  if (mode === 'off') return;
  event.system.push({ type: 'text', text: getPonytailInstructions(mode) });
};
await ctx.session.hook('context', addRuleset);
await ctx.session.hook('compaction', addRuleset);

With this change, the summary model also receives the ruleset.

Related

OpenCode's compaction cache issue: anomalyco/opencode#52761