fix(session): cache messages across prompt loop to preserve prompt cache byte-identity

[BYK]

Issue for this PR

Closes #25366

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

OpenCode reloads messages from the DB at the start of every prompt loop iteration via MessageV2.filterCompactedEffect(). Between tool-call steps, tool parts transition from pending → completed with output text. This means the same assistant message serializes to different bytes on consecutive API calls, busting Anthropic's prompt cache from that position forward.

Fix: Cache the conversation array across loop iterations. On tool-call continuation steps, only append genuinely NEW messages instead of reloading all messages from the DB. Existing messages retain their original serialized form (as the API last saw them), preserving byte-identity for the prompt cache.

Full reloads still happen on:

• First iteration
• After compaction (conversation structure changes)
• After subtask/overflow recovery (these structurally change the conversation)

Impact: On sessions with heavy tool use, this eliminates prompt cache busts on continuation steps. On April 21st alone, the cache-bust pattern cost $2,264 in cache writes vs $1,234 in cache reads on a single account.

How did you verify your code works?

• Verified cached messages are used on tool-call continuation
• Full reload still triggers after compaction
• Existing tests pass

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential Duplicate Found:

• PR fix(session): cache messages across prompt loop to preserve prompt cache byte-identity #24842: "fix(session): cache messages across prompt loop to preserve prompt cache byte-identity"
  • fix(session): cache messages across prompt loop to preserve prompt cache byte-identity #24842
  • This PR has the exact same title and purpose as the current PR fix(session): cache messages across prompt loop to preserve prompt cache byte-identity #25367. It addresses the same issue (caching messages to preserve prompt cache byte-identity across prompt loop iterations).

Related PRs that may be of interest:

• PR feat(opencode): cache-aligned compaction to reuse prefix cache #25100: "feat(opencode): cache-aligned compaction to reuse prefix cache"

  • feat(opencode): cache-aligned compaction to reuse prefix cache #25100
  • Related to prompt cache optimization strategies

• PR fix(cache): improve Anthropic prompt cache hit rate with system split and tool stability #14743: "fix(cache): improve Anthropic prompt cache hit rate with system split and tool stability"

  • fix(cache): improve Anthropic prompt cache hit rate with system split and tool stability #14743
  • Related to prompt cache hit rate improvements

Recommendation: Check PR #24842 to see if it was already addressing this issue or if there's a reason for the duplicate attempt.