Skip to content

February 6: First Mosaics and Getting Graded by AI

A focused session — 1 fix, 1 docs.

Developer Journal

First mosaic creation — three raw FITS images aligned in the sky, stitched together. Then six rateints, then fifty images. Level 3 images look much better than raw. The output quality was lowered because there's no offline processing pipeline yet, and the mosaic tool outputs images rather than new FITS files (a future feature). Shared multiple screenshots showing the progression. "NASA just used more than 3" — the current 3-channel composite limitation is becoming the obvious next target.

JWST mosaic of two overlapping fields showing deep-sky galaxies in infrared

JWST mosaic combining approximately fifty images of a nebula region with visible tiling seams

Crab Nebula reference image showing the multi-filter NASA-quality target for compositing

Three raw FITS images stitched together as the first mosaic attempt in the application

JWST mosaic of the Crab Nebula showing filament structure in infrared colormap

JWST mosaic of Carina Nebula region stitched from multiple FITS observations

First mosaic output exported as a standalone image from the mosaic tool

Gave a friend a detailed walkthrough of the development process and guardrails: how Claude writes a plan, turns it into CLI commands and source edits, and how none of that code gets near the main branch without passing linting, unit tests, E2E tests, and a human-reviewed PR with a test plan. The tech lead experience of managing offshore teams that could break things overnight translated directly into managing an AI developer.

Claude CLI showing the git push-to-main hook blocking a direct commit and forcing a branch-based workflow

Mosaic feature known limitations and TODO list showing completed items and remaining work

Tried the Claude /insights feature — "getting graded by the AI, and it hurts." A friend found it interesting that their earlier concerns were called out as areas for improvement. Also tried the agent teams feature for the first time — hit the usage limit right at the end of the session. Connected the Claude mobile app to a cloud environment and claimed $50 of overage credit. "I got ahead of my skis."

Claude mobile app writing a feature-ideas.md file to the repo from a phone

Claude CLI usage panel showing 100% session usage and 21% weekly usage

Claude web app usage settings showing 100% session limit hit with $50 extra usage balance

Highlights

#171 Add retry with backoff to token refresh (Tech Debt #88)

  • Adds retry logic (3 attempts, 1s/3s backoff) to all three token refresh paths: scheduled refresh, 401 handler, and fallback refresh
  • Adds AuthToast component that shows warning notifications during retries and an error notification before logout
  • Only logs the user out after all retries are ex...

What Changed

Bug Fixes (1)

  • #171 Add retry with backoff to token refresh (Tech Debt #88)

Documentation (1)

  • #172 Mark tech debt #88 as resolved

33 commits across 2 pull requests. Next: February 7, 2026 — Sort mosaic file selection by processing level (li..., Pin Docker image versions for reproducible builds ..., Image comparison/blink mode (C3)