fix(cli): prune tool outputs for large requests

[marius-kilocode]

Compacts older tool outputs when a model request has already grown large, reducing long-session request payloads without restoring default pruning after every normal turn.

Keeps normal-turn pruning opt-in for cache stability, while pruning after compaction and at payload-limit boundaries where the prompt prefix is already changing or too large to send efficiently.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Resolved Issues

Both warnings from the initial review were addressed in the latest commits:

File	Line	Resolution
packages/opencode/src/session/prompt.ts	1606	plugin.trigger("experimental.chat.messages.transform") is now re-applied inside the payload-limit rebuild path
packages/opencode/src/session/prompt.ts	1610–1611	Post-prune size is now computed (nextSize) and a log.warn is emitted if the payload remains oversized

Files Reviewed (3 files)

• .changeset/large-session-tool-pruning.md — no issues
• packages/opencode/src/session/compaction.ts — no issues
• packages/opencode/src/session/prompt.ts — no issues

---

_{Reviewed by claude-sonnet-4.6 · 255,182 tokens}

[marius-kilocode]

@kilocode-bot is this ready to be merged?

[kilo-code-bot]

To use Kilo from GitHub you first need to link your GitHub account to Kilo. Link your Kilo account to continue. After linking, mention me again in this issue or pull request.

[marius-kilocode]

@kilocode-bot is this ready to be merged?

[kilo-code-bot]

Based on the existing code review already completed on this PR, yes, it looks ready to merge! 🟢

The automated review gave it:

• Status: No Issues Found
• Recommendation: Merge

Both warnings from the initial review were resolved in subsequent commits:

1. plugin.trigger(experimental.chat.messages.transform) is re-applied inside the payload-limit rebuild path
2. Post-prune size is now computed (nextSize) and a log.warn is emitted if the payload remains oversized

All 3 files were reviewed with no outstanding issues. The change looks solid — it compacts older tool outputs at payload-limit boundaries while keeping normal-turn pruning opt-in for cache stability.