Session Summary · Batch Image Processor

Interview prep, end of session

A snapshot of what was built, what passes, and what's still open on the Anthropic batch-image-processor practice problem. Files are color-coded by role; the review table flags every bug surfaced during the Phase 2 mock interview.

practice solution test tooling

01Files created

practice

image_processor_practice_impl.py

Practice template. No library named, no API cheat sheet — interview-realistic.

Stage 2 working · Stage 1 deferred

solution

image_processor_impl.py

Reference implementation: sequential loop + ProcessPoolExecutor parallel.

Complete

test

image_processor_practice_test.py

30 tests, library-agnostic. Only inspects files on disk — any library works.

All 30 pass against solution

02Test breakdown

30 tests across 4 classes. Each cell is a class; the number is its test count.

TestStage1PerTransform

One per op: grayscale, flip h/v, scale ±, blur, rotate 90/180.

TestStage1ProcessImages

Cartesian product, sequencing, JPEG mode, mixed exts, empties, nested subdirs.

TestStage1EdgeCases

Empty list, scale 1.0, double-flip identity, 45° expansion, kitchen sink, idempotent rerun, RGBA PNG, 1×1, blur radius 0, etc.

TestStage2Parallel

Correctness, 24-workitem fan-out, speed sanity, empty inputs.

03How the practice + test files evolved

Practice file

Pillow imports + full API cheat sheet baked in. You flagged: "this gives me all the solution."

Rewrote to "blank editor + Google open in another tab" framing. No library named, no imports, TODOs reframed as design decisions.

You added Pillow imports and implemented apply_transforms, _process_image_worker, and process_images_parallel.

Test file

24 tests, but unit tests were Pillow-specific — assumed your impl took/returned PIL Images.

Refactored to library-agnostic: only checks files on disk. Pillow only used in the harness to create fixtures + read outputs back.

Added 11 edge case tests → 30 total, all passing against the solution.

04Code review findings

From the Phase 2 mock interview review of your process_images_parallel.

#	Issue	Severity	Status
1	`with ProcessPoolExecutor as executor:` missing `()`	Crash	Fixed
2	`rotate(angle)` without `expand=True` — fails dimension test	Test-fail	Open
3	`sorted(jobs, key=...)` returns new list; result discarded. Want `jobs.sort(..., reverse=True)` for largest-first	Perf nit	Open
4	Case-sensitive `.endswith(...)` misses `.JPEG`	Test-fail	Open
5	`process_images` (sequential) is still `raise NotImplementedError`	Test-fail	Deferred
6	`type = transform["type"]` shadows Python's built-in `type`	Style	Open
7	`print(...)` per output in the parallel map loop	Style	Open
8	Bare filenames passed to `get_output_path` — works with this harness, fragile in general	Subtle	Open

05Current test result on your code

3 / 4

Stage 2 parallel tests passing

The 4th errors only because the speed sanity check calls process_images first.

06Tooling installed this session

python

Pillow

pip install Pillow — used by both your implementation and the test harness (for fixtures + reading outputs).

claude code

html-docs skill

Installed to ~/.claude/skills/html-docs/SKILL.md. Powers this very document.

07Where to pick up next session

Fix the 4 open code issues (rotate expand, .lower().endswith(), sort-in-place, full paths to get_output_path).
Implement process_images sequentially — reuse _process_image_worker in a for-loop.
Run the full suite (python3 ant/test/image_processor_practice_test.py -v) for a complete scorecard.
Work through the Stage 3 discussion follow-ups: threads vs processes (with the GIL-release nuance for image libs), memory under parallel load, failure handling, scaling beyond one machine.