Model performance degredation in recent update

[thecannabisapp]

What version of Kimi Code CLI is running?

1.24.0

Which open platform/subscription were you using?

kimi-code

Which model were you using?

kimi-for-coding

What platform is your computer?

macOS

What issue are you seeing?

Performance drop since recent update. I have a effect-ts monorepo which the model was quite good at navigating and following AGENTS.md instructions, but somehow in recent useage I noticed a lot of mistakes, not following rules, over complicating code, confusion and overall lower satisfcation using the tool as it makes too many mistakes.

What steps can reproduce the bug?

Hard to say, my project configs are setup for my rules but it's just a general observation. Heres what the LLM outputted based on my feedback.

Performance Degradation Report

Task: Fix circular dependency in TypeScript Effect-TS monorepo
Failure Mode: Multiple cascading errors, rule violations, poor code quality

Specific Failures:

1. Critical Rule Violation - Used git checkout -- . despite AGENTS.md explicitly forbidding it. Bypassed all
   ety checks (untracked file check, stash, explicit approval).
2. Inconsistent Code Quality - Created files with:
   • Wrong imports (.js extensions forbidden by AGENTS.md)
   • Wrong patterns (Effect.Tag instead of Context.Tag as used in codebase)
   • Unnecessary complexity (placeholder Default layers, re-exports user explicitly forbade)
3. Incomplete Solution - Failed to fully break the circular dependency. Moved ApiClientService but left 40+ f
   s still importing from @ma/provider.
4. Lack of Care - User had to repeatedly correct the same mistakes. Failed to verify work at each step.

What is the expected behavior?

No response

Additional information

No response

[raye-deng]

The pattern you're describing — model performing well on a codebase, then degrading after an update with more mistakes, rule violations, and over-complication — is something we've been tracking across multiple AI coding tools.

The core issue is that model quality is a moving target, but most development workflows treat it as a constant. When the model was good, your AGENTS.md rules were mostly followed. When the model degrades, those rules become suggestions that the model intermittently ignores. And the failure mode is subtle: the code still compiles, tests might still pass, but the quality drops in ways that are hard to catch in review.

We've categorized the most common degradation patterns:

1. Instruction drift — model acknowledges rules but doesn't follow them consistently (exactly what you're seeing)
2. Over-engineering creep — model adds unnecessary abstractions that inflate complexity without adding value
3. Stale pattern regression — model reverts to older API patterns from training data, introducing deprecated calls that work today but will break on the next dependency update
4. Cross-file inconsistency — model makes changes that are locally correct but violate assumptions in other files it hasn't re-read

The takeaway for us has been: you need a validation layer that's independent of the model. AGENTS.md and prompt rules are necessary but insufficient because they depend on the model's compliance, which varies across versions. Deterministic checks — registry validation for imports, AST-level cross-reference analysis, deprecated API detection against actual package changelogs — catch these regardless of which model version is running.

For anyone dealing with this: npx @opencodereview/cli scan . --sla L1 adds that deterministic layer as a CI check.

[thecannabisapp]

Thank you for the kind reply and the validation layer suggestion.