The Problem Nobody Talks About
Last month I spent two hours debugging why Cursor kept suggesting a deprecated API—one I'd removed three commits ago. The file was gone. Git history was clean. Yet the AI insisted the old implementation was still valid.
Turns out? Context window compression had silently dropped my recent changes. The model was working with a 6-hour-old snapshot of the codebase.
Why Context Compression Breaks Silently
Cursor (and most AI coding tools) aggressively compress context to fit models' token limits. When your codebase crosses that threshold:
- Recent file changes get deprioritized
- Test files disappear from context entirely
- The AI hallucinates based on stale architecture
The worst part? No warnings. You just get confident, wrong answers.
Enter Context Mode
I found mksglu/context-mode after the third time I shipped a bug because Cursor's context and my actual codebase had diverged.
It's a dead-simple VSCode extension that shows you exactly what files are in your AI's context window—in real time, with token counts.
What It Actually Does
# Install
code --install-extension mksglu.context-mode
# Then open Command Palette (Cmd+Shift+P)
# Search: "Context Mode: Show Active Context"
You get a live panel showing:
- Which files are currently loaded (by priority)
- Token count per file
- What got dropped due to limits
Real-World Impact
Before Context Mode:
Me: "Why is the login test failing?"
Cursor: "The test passes locally for me."
[30 min of confusion]
Turns out: test file wasn't in context
After Context Mode:
Context panel shows: auth.test.ts (excluded - token limit)
Me: [adds .cursorrules to prioritize test files]
Cursor: [actually sees the failing assertion]
My .cursorrules Setup
I now force-include critical files that Cursor's auto-prioritization misses:
# .cursorrules
context:
always_include:
- "**/*.test.ts"
- "**/*.spec.ts"
- "vitest.config.ts"
- "tsconfig.json"
exclude:
- "node_modules/**"
- "dist/**"
- "*.log"
This cut my context size from 52K → 12K tokens while keeping test coverage visible.
The Verification Script
I added this to my pre-commit hook:
#!/bin/bash
# .git/hooks/pre-commit
# Check if critical files are in Cursor's context
context_check() {
local file=$1
if ! cursor-context-cli contains "$file"; then
echo "⚠️ $file missing from AI context"
exit 1
fi
}
context_check "vitest.config.ts"
context_check "src/**/*.test.ts"
(Note: cursor-context-cli is a hypothetical wrapper—Context Mode doesn't ship a CLI yet, but the VSCode API makes this trivial to build.)
What I Learned
- AI coding tools lie by omission. If a file isn't in context, the model will confidently work around its absence.
- Token budgets are not fungible. Dropping test files to fit more source code is almost always the wrong tradeoff.
- Verification is cheaper than debugging. Spending 5 seconds checking Context Mode beats spending 20 minutes chasing phantom bugs.
The Dilemma I'm Still Solving
Context Mode solved visibility, but I'm still wrestling with prioritization:
- If I force-include all tests, I hit token limits on large PRs
- If I let Cursor auto-prioritize, it drops the exact test that would've caught my bug
- Manual context curation per task feels like fighting the tool
How do you decide what belongs in context when everything feels critical?
Top comments (1)
Worth putting a break-even on the 40K, because the saving is smaller than it looks once a failure is priced. At $3 per million input tokens, 40K is about $0.12 saved per request. One silent failure that costs twenty minutes of human triage is roughly $20 - so the saving is wiped out by one extra failure per ~167 requests, a failure-rate increase of about 0.6%.
That is the number I would guard with the test suite, precisely because it is small enough that "it still looks fine" will not catch it. Track the per-request token delta and the failure delta side by side; only then is the trimming obviously a win.