Inconsistent reasoning - exif example

[vinodmut]

Consider the where was the photo @sample.jpg taken. use exif metadata task (from the kaizen lite readme).

I ran it three times now and claude code solved it in three different ways:

1. Tried to run `exiftool` and `identify` to extract EXIF data — both unavailable.
2. Tried to use Python's `PIL` (Pillow) to read the EXIF — failed because Pillow wasn't installed.
3. Installed Pillow via `pip`, then successfully extracted full EXIF and GPS data from `/workspace/sample.jpg`.

1. **First attempt** — tried `exiftool`, `exiv2`, and `identify` (ImageMagick). All produced no output (tools not installed).
2. **Second attempt** — tried Python Pillow (`img._getexif()`). Failed with exit code 1 (Pillow likely not installed).
3. **Third attempt** — manually read the raw JPEG binary to locate the APP1/EXIF marker. Found EXIF data of length 1120 bytes.
4. **Fourth attempt** — wrote a full custom TIFF/EXIF parser in Python to decode all IFD tags and GPS data. This succeeded.

1. **Failed tool attempt** — `exiftool` was not installed (exit code 127).
2. **Python/PIL fallback** — `PIL` module was also unavailable.
3. **Raw binary parsing** — The assistant manually parsed the JPEG binary structure, discovering EXIF data at offset 24, GPS data at offset 17065, and XMP metadata.
4. **XMP extraction** — Revealed the photo was taken with a **Google Pixel 4a** on **2021-10-17 at 17:41:41 BST**.
5. **Manual GPS EXIF parsing** — Extracted raw GPS IFD entries and decoded the coordinates:
   - Latitude: **51.4936° N**
   - Longitude: **0.0376° W**

We are seeing very inconsistent reasoning with the same model (sonnet-4.5).

This is without Kaizen.