docs: tool system performance audit for long-session regressions

[kilo-code-bot]

Summary

• Research audit of recent changes (Feb 26 - Mar 26, 2026) to the tool system, system prompts, and agent definitions that could cause performance regressions in long-running sessions
• Documents token budget impact of each change with specific PRs, commits, file paths, and diffs
• Identifies the session diff memory leak fix (PR fix(cli): prevent CLI memory leaks from oversized diffs and session accumulation #7617) as the most impactful change for long-session performance

Key Findings

Change	PR	Per-request token impact
codebase_search tool	#6685	+~350 tokens (experimental flag)
environment_details injection	#6225	+15 to +250 tokens (always)
Config paths in system prompt	#7444	+~35 tokens (always)
PlanExitTool always registered	N/A	0 (filtered by agent perms)
Session diff memory leak fix	#7617	N/A (storage fix, not token)

Typical case (default settings, CLI): +~50 tokens per request
Worst case (all features enabled, VS Code with many tabs): +~635 tokens per request

Recommendations

1. Cap open tabs included in environment_details to limit token growth
2. Consider whether the always-present ISO 8601 timestamp is worth ~15 tokens/request
3. bash.txt (9.6 KB) and todowrite.txt (8.8 KB) are the largest tool descriptions — prime targets for compression if token budget is a concern

[kilo-code-bot]

WARNING: Executive summary overstates the PlanExitTool token impact

The detailed analysis later in this document correctly notes that resolveTools() filters plan_exit out for agents that do not allow it, so its description is not added to most LLM requests. Saying it is added "to every session" contradicts that conclusion and overstates the reported regression risk.

Suggested change

	4. `PlanExitTool` is now always registered in the tool registry instead of being feature-flag-gated, adding its description (~614 bytes) to every session.
	4. `PlanExitTool` is now always registered in the tool registry instead of being feature-flag-gated, but agent permission filtering keeps its description out of most sessions.

[kilo-code-bot]

Code Review Summary

Status: 1 Issue Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	1
SUGGESTION	0

Fix these issues in Kilo Cloud

Issue Details (click to expand)

WARNING

File	Line	Issue
docs/research/tool-system-performance-audit.md	16	Executive summary says PlanExitTool adds ~614 bytes to every session, but the later analysis correctly notes agent permission filtering keeps it out of most LLM requests.

Other Observations (not in diff)

Issues found in unchanged code that cannot receive inline comments:

None.

Files Reviewed (1 files)

• docs/research/tool-system-performance-audit.md - 1 issue

---

_{Reviewed by gpt-5.4-20260305 · 293,893 tokens}

[marius-kilocode]

This writes to docs/ at the repo root, but we don't have a top level docs/ tree here. Looks like a one off research note that wasn't going to any published docs surface, closing it out.