Cursor-Interpreted vs Raw

Track Design

Interpreted track captures what the agent believes it retrieved: how much content it saw, whether the fetch was complete, how it characterizes truncation. This is the agent’s self-report.

Raw track captures what Cursor actually saved to disk: exact byte counts, hexdump analysis, MD5 checksums, and token counts. These are filesystem measurements, cryptographic hashes, and not agent estimates.

The gap between the two tracks is a finding. If Cursor reports “content complete” in prose, but the raw data shows truncation, that discrepancy belongs in the spec.

Key Observations

1. Reproducibility in Raw vs High Variance in Interpreted

   Raw: Same URL produces identical output

   • BL-1: MD5 d6ad8451d3778bf3544574431203a3a7 across 2 runs
   • OP-4/BL-3: MD5 554eb56e8416d86d12af17a2dfe6f815 across 3 runs
   • Character-for-character identical output on disk

   Interpreted: Same URL produces 2-3× variance on small files

   • BL-1 r1: 1,953 chars → r2: 5,595 chars → r3: 4,100 chars, 2.9× variance
   • BL-2 r1: 1,953 chars → r2: 4,200 chars → r3: 4,350 chars, 2.2× variance

   Conclusion: the variance is in how Cursor displays content in chat UI, not what it fetches. Raw measurements prove the underlying fetch is deterministic, but that interpreted track shows UI rendering isn’t.

2. Size-Dependent Consistency

   Small file rendering appears unreliable, while larger ones seem stable:

   Interpreted:

   • Small files, 20-87 KB: high session-to-session variance
   • Large files, 245 KB: <1% variance, nearly identical across runs

   Raw:

   • Small files, 4.8 KB output: identical MD5s despite variance in interpreted display
   • Large files, 245 KB output: identical MD5s, consistent across runs

   Conclusion: Cursor fetches consistently at all sizes. The interpreted variance on small files is possibly a UI rendering artifact, not entirely reflective of fetch behavior.

3. Perception Gap: Model Self-Report is Unreliable

   Agent claims “complete” or “no truncation” when content is a filtered subset:

   Test	Raw	Interpreted	Gap
   SC 3	38 KB truncated at ref #14/252	“Complete reference section”	Agent interprets filtered list as complete
   BL 1	4,817 B calculated	1,953 chars displayed	UI shows subset, agent reports what it sees
   SC 4	Truncated mid-word at “updated”	“All syntax sections present”	Clean structure masks incompleteness

   Conclusion: trust character counts, not prose assertions; agent perceives filtered excerpts as complete because they’re internally coherent.

4. Method-Specific Truncation Limits - Raw

   • WebFetch, MCP-style: ~28 KB ceiling, SC-4 truncated at 27,890 chars
   • urllib.request: ~72 KB ceiling, EC-6 truncated at 72,600 chars
   • curl fallback: No ceiling detected, SC-2 returned 17.6 MB
   • Unknown Path: No ceiling detected, OP-4/BL-3 returned 245 KB

   Conclusion: Cursor routes to many mechanisms with different limits. The interpreted track didn’t identify this because the agent’s self-report didn’t consistently include its toolchain.

5. Intelligent Content Filtering

   Cursor performs structure-aware filtering, but the raw track provided the measurements:

   Interpreted: agent reports receiving “main content” but missing footer/navigation

   Raw: proves it via byte counts

   • BL-1: 85 KB HTML → 4.8KB Markdown, 94% reduction, CSS/navigation stripped
   • SC-3: 252 references → deterministically selects #14, the first commercial source
   • SC-4: 30 KB page → 28KB, footer/metadata filtered

   Conclusion: Cursor applies content heuristics, not blind truncation. Raw track quantifies what interpreted track observes qualitatively.

6. Chars/Token Ratio as Content-Type Classifier - Raw

   • EC-3: JSON: 2.62 chars/token
   • SC-2: Raw HTML/JS: 2.65 chars/token
   • SC-3: Tables: 3.06 chars/token
   • SC-4, OP-4: Docs + code: 3.91-4.37 chars/token
   • BL-1, BL-2, SC-1: Clean Markdown: 4.13-4.36 chars/token

   Conclusion: Chars/token ratio enables content-type classification without parsing. <3.0 = code/markup, >4.0 = prose. Useful for automated analysis pipelines. The interpreted track had no visibility into this pattern.

7. Cross-Track Agreement on Redirect Handling

   Interpreted: agent received final destination JSON content

   Raw: confirmed 5-level redirect chain traversed - 1,021 bytes JSON saved

   Conclusion: redirect handling is robust across both measurement approaches

Implications for Agent Developers, Docs Teams

When evaluating or designing testing frameworks or workflows that include agentic web fetch behavior, consider what each approach can and can’t confirm:

Agentic self-reports are unreliable for detecting truncation or content subsetting, when building workflows include a raw track-like verification.

Architecture Comparison

Precision Comparison

Claude API’s web fetch has the cleanest measurement story as the tool results are first-class response fields, fully observable. Gemini neatly separates retrieval metadata. Cursor requires filesystem inspection for precise measurements, because agents deliver estimations by default.