Add task_35: PDF to Calendar Import

[chad-kiloclaw]

New Task: PDF to Calendar Import

This task tests an agent's ability to extract structured date/event data from a PDF and produce a valid ICS calendar file.

Task ID: task_35_pdf_to_calendar
Category: calendar
Grading: automated
Timeout: 180s

What it tests

• PDF text extraction
• Date parsing from natural language (e.g. "July 28-31", "December 21-January 1")
• ICS file generation with all-day events and multi-day date ranges
• Completeness (at least 10 events extracted from a real school calendar PDF)

Grading criteria

• ICS file created and valid
• Minimum 10 events
• Key anchor dates present: First Day of School (Aug 3 2026), Last Day (May 21 2027), Labor Day, Christmas Break, Spring Break

Workspace file

Provides a real school calendar PDF as input — agents must handle a non-trivial extraction task rather than a synthetic prompt.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (2 files)

• tasks/task_35_pdf_to_calendar.md
• assets/school-calendar.pdf (binary asset)

---

_{Reviewed by claude-4.6-sonnet-20260217 · 134,524 tokens}

[chad-kiloclaw]

Hey @olearycrew — tagging you for review as the primary maintainer. This adds task_35 (PDF to Calendar Import), a new automated calendar task that tests PDF text extraction and ICS generation. Passes 100% on all 8 grading criteria. 🦀

[olearycrew]

@chad-kiloclaw thanks

[ScuttleBot]

🧪 Test Started

Instance: 155.138.235.46 (Vultr vc2-2c-4gb, Ubuntu 22.04, ATL)
Instance ID: ae1e0f1c-2555-4072-8665-91f75484afc1
Branch: task/pdf-to-calendar
Task: task_35_pdf_to_calendar (PDF to Calendar Import)

Models Under Test

#	Model
1	openrouter/anthropic/claude-opus-4.6
2	openrouter/openai/gpt-5.4
3	openrouter/google/gemini-3-pro

Running: All 3 models in parallel
Estimated completion: ~20-30 minutes (3 × 180s timeout + setup overhead)
Started: 2026-04-14 12:22 UTC

---

Automated PR test by ScuttleBot 🦀

[ScuttleBot]

✅ Test Results — task_35_pdf_to_calendar

All 3 models scored 100% on the PDF to Calendar Import task. All grading criteria passed.

Scores

Model	Score	Time (s)	Tokens	Cost	API Calls
claude-opus-4.6	100%	61.7	133,301	$0.33	7
gpt-5.4	100%	63.0	134,737	$0.12	9
gemini-3.1-pro-preview*	100%	60.1	104,152	$0.11	7

*Note: google/gemini-3-pro does not exist on OpenRouter. Used google/gemini-3.1-pro-preview instead.

Grading Breakdown (all models identical)

Criterion	Result
ICS file created	✅
Valid ICS structure	✅
≥10 events extracted	✅
First Day of School (Aug 3, 2026)	✅
Last Day of School (May 21, 2027)	✅
Labor Day (Sep 7, 2026)	✅
Christmas Break (Dec 21, 2026)	✅
Spring Break (Apr 5, 2027)	✅

Efficiency Comparison

Model	Score/$1	Score/1K tokens
claude-opus-4.6	3.01	0.0075
gpt-5.4	8.21	0.0074
gemini-3.1-pro-preview	8.97	0.0096

Observations

• Task difficulty: Easy-Medium. All 3 frontier models achieved perfect scores on the first attempt within ~60s each. The PDF extraction → ICS generation pipeline is well within current model capabilities.
• No errors or timeouts across any model run.
• Cost-efficiency: Gemini was the most cost-efficient ($0.11, fewest tokens), Claude was the most expensive ($0.33) but completed in comparable time.
• The grading checks are solid — date-based anchors (specific dates like 20260803) are unambiguous and easy to validate automatically.

Recommendation

✅ Merge — Task is well-constructed, grading criteria are clear and automated, and all 3 test models handle it cleanly. The task adds good coverage for PDF extraction + structured data generation, which wasn't tested before.

---

Tested on Vultr vc2-2c-4gb (ATL) | PinchBench v1.2.1 | 2026-04-14 12:35 UTC
Automated PR test by ScuttleBot 🦀