## [Regression] Kimi Code 2.7 VS Code extension: severe quality & performance degradation

[ai-insights-cloud]

Summary

Kimi Code 2.7 in VS Code has regressed significantly compared to 2.5 in terms of instruction adherence, performance, and token efficiency. The model ignores explicit instructions, hallucinates changes, and consumes excessive tokens on unnecessary reasoning.

Environment

• Extension: Kimi Code VS Code Extension 2.7
• OS: Windows 10
• Previous working version: 2.5 (no issues)

Issues Observed

1. Complete Ignoring of Simple Instructions

The model routinely ignores explicit, contextually clear instructions provided in the system prompt (AGENTS.md) and user prompts. Examples include:

• Ignoring "do not rewrite entire files" → rewrites entire files anyway
• Ignoring "continue from current state" → starts over from scratch
• Ignoring custom UI-Kit standards → applies own styling fantasies despite precise design specs

2. Unwanted Initiative (Destructive Hallucination)

The model proactively modifies code that was not requested, breaking existing functionality. This includes:

• Adding unsolicited features
• Changing file structures without permission
• Removing working code while "fixing" unrelated issues

3. Excessive & Slow Reasoning

• Response latency increased 8-10x compared to version 2.5
• Constant use of speculative language ("maybe", "perhaps", "possibly") even for simple, deterministic tasks
• Long-winded explanations that do not contribute to the solution

4. Token Efficiency Collapse

For the same volume of actual code tasks, billing limits are exhausted significantly faster due to:

• Unnecessary verbose reasoning
• Repeated rewrites of simple components
• Failure to complete tasks in first attempt

5. Concrete Example: 3 Days for a Modal Close Button

Task: Implement modal window close functionality (click ×, ESC key, overlay click)
Expected: Single iteration, ~20 lines of code
Actual: 3 days of back-and-forth, with the model repeatedly:

• Rewriting unrelated modal styles
• Breaking existing click handlers
• Adding unnecessary animation logic not requested
• Failing to respect the existing UI-Kit standards

6. Promised Credits Not Delivered

• Plan: Allegro ($99) with advertised "15x credits"
• Actual: No observable 15x credit multiplier; limits deplete at standard or faster rate
• Impact: Paying premium price without receiving promised capacity

Expected Behavior

• Strict adherence to explicit instructions
• Minimal, targeted changes (diff-style, not full-file rewrites)
• Reasoning proportional to task complexity
• Performance on par with or better than 2.5

Actual Behavior

• Systematic instruction violation
• Destructive initiative
• 8-10x slower response times
• Rapid quota depletion

Request

1. Acknowledge the regression between 2.5 and 2.7
2. Provide a rollback path to 2.5 or a hotfix
3. Publish the system prompt structure so users can enforce constraints effectively
4. Add a "strict mode" that disables unsolicited modifications
5. Verify and fix the "15x credits" multiplier for Allegro plan subscribers — currently not reflected in actual usage

---

Additional Context

This is a paid commercial product — Kimi Code subscription (Allegro plan, $99/month).
The current state makes it economically unviable — the cost of fixing
AI-generated errors exceeds the cost of manual coding.