Fix: Update hardcoded Claude 3.7 in compaction to respect user's model selection

[sebastiangeraldes]

Fix: Update hardcoded Claude 3.7 in compaction config to respect user's model selection

Issue

When chat context reaches 60% capacity, the compaction feature (which summarizes old messages) uses a hardcoded Claude 3.7 model instead of respecting the user's selected model. This causes:

1. Model identity confusion - Claude reports being version 3.7 when user selected 4.5
2. Behavioral inconsistencies - User gets 4.5 responses mixed with 3.7-generated summaries
3. Conspiracy theories about AWS secretly swapping models - Users assume AWS is doing something nefarious behind the scenes
4. No cost savings - Claude 3.7 costs the same as 4.5, so this isn't even a cost optimization
5. Suboptimal compaction - If cost/speed were the goal, Haiku would be better

Root Cause

File: src/agent/agent-context/compaction/config.ts (compiled to extension.js:668866)

DEFAULT_COMPACTION_CONFIG = {
  enabled: true,
  compactionThreshold: 0.6,
  preserveRecentRounds: 2,
  defaultCompressionRatio: 0.5,
  maxCompactionTimeMs: 60000,
  provider: "qdev",
  analysisModel: "CLAUDE_3_7_SONNET_20250219_V1_0",  // ← HARDCODED
  strategies: { /* ... */ }
}

Usage: src/agent/agent-context/compaction/strategies/simple-strategy.ts (extension.js:699447)

const modelName = config.analysisModel;  // Always returns hardcoded 3.7
const summarizationModel = await loadModel(provider, modelName, "summarization");

Proposed Fix

Option 1: Use User's Selected Model (Recommended)

// In config.ts
import { getSelectedModelId } from '../model-selection/model-configuration';

DEFAULT_COMPACTION_CONFIG = {
  // ... other settings ...
  provider: "qdev",
  analysisModel: null,  // Will be resolved at runtime
  // ... strategies ...
}

// In simple-strategy.ts (generateCompressedContext method)
const provider = config.provider;
const modelName = config.analysisModel || getSelectedModelId() || "CLAUDE_SONNET_4_5_20250929_V1_0";
const summarizationModel = await loadModel(provider, modelName, "summarization", { forceRefresh: true });

Option 2: Update to Claude Sonnet 4.5 (Minimal Change)

// In config.ts
DEFAULT_COMPACTION_CONFIG = {
  // ... other settings ...
  analysisModel: "CLAUDE_SONNET_4_5_20250929_V1_0",  // Updated from 3.7
  // ... strategies ...
}

Option 3: Use Haiku for Cost/Speed (Best for Compaction)

// In config.ts
DEFAULT_COMPACTION_CONFIG = {
  // ... other settings ...
  analysisModel: "CLAUDE_HAIKU_3_5_20241022_V1_0",  // Cheaper, faster, good enough for summaries
  // ... strategies ...
}

Patch File (for compiled extension.js)

Since the source code isn't public, here's a patch for the compiled extension.js:

--- a/resources/app/extensions/kiro.kiro-agent/dist/extension.js
+++ b/resources/app/extensions/kiro.kiro-agent/dist/extension.js
@@ -668866 +668866 @@
-      analysisModel: "CLAUDE_3_7_SONNET_20250219_V1_0",
+      analysisModel: "CLAUDE_SONNET_4_5_20250929_V1_0",

Testing

1. Start a chat with Claude Sonnet 4.5 selected
2. Continue conversation until context reaches 60% capacity
3. Verify compaction triggers (check logs)
4. Ask Claude "what model are you?"
5. Expected: Claude correctly identifies as 4.5 (or doesn't know, which is fine)
6. Previous behavior: Claude says 3.7

Additional Improvements

While fixing this, consider:

1. Add model identity to system prompt - Tell Claude what version it is explicitly
2. Make compaction model configurable - Add setting so users can choose compaction model
3. Log compaction events - Show users when compaction happens and which model is used
4. Dynamic provider support - Respect user's provider selection (currently hardcoded "qdev")

Related Issues

• Wrong Model Deployed (Claude Sonnet 4.5 Showing Claude 3.7 Behavior) #3625 - Claude identifies as version 3.7 when Sonnet 4.5 is selected

Special thanks to @huachuman for the conspiracy theory about AWS secretly swapping models behind the scenes, which inspired a deeper investigation into the actual cause. Turns out it wasn't AWS - just a hardcoded string at line 668866. Much less exciting, but easier to fix.

References

• Decompiled extension.js analysis
• Line 668866: analysisModel definition
• Line 699447: analysisModel usage in compaction
• Line 668856: compactionThreshold: 0.6 (60% trigger)

[huachuman]

Special thanks to @huachuman for the conspiracy theory about AWS secretly swapping models behind the scenes, which inspired a deeper investigation into the actual cause. Turns out it wasn't AWS - just a hardcoded string at line 668866. Much less exciting, but easier to fix.

You're literally the most ignorant person on planet earth if you don't even look into how AWS handles models at high load. This isn't a conspiracy theory. You're just an idiot.

Claude 3.7 compaction logic has absolutely nothing to do with degraded performance across the app except for when you summarize a conversation.

This PR is laughable and will help exactly nobody.