You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
As a contributor committing regenerated binary artifacts (PNGs, SVGs, screenshots, diagrams), I want documented guidance to spot-check the visual output before committing so that a buggy regenerator producing garbage that still tests-pass doesn't silently land on main.
What
When a regenerator produces binary artifacts (overlay PNGs, diagram SVGs, screenshot snapshots, plot images), `git diff` reports only "Binary files differ" — opaque to reviewers and to the author. The git diff itself proves nothing about correctness.
Pattern to document: open each regenerated binary in an image viewer (or the agent's image-read tool, for AI-assisted workflows) before staging. Confirm:
Geometry tracks the source data cleanly (overlays trace their underlying curves; diagrams match their schema; screenshots show the expected UI state)
No obvious corruption (truncated images, empty plots, garbage gradients)
Output is consistent with sibling artifacts (e.g. all 3 refreshed overlays in the same campaign should look similar in quality)
Cheap check: 1-2 minutes per artifact. Catches the failure mode where a regenerator bug produces invalid-but-syntactically-valid output that tests don't detect.
Why
Observed in me-fuji session 146: two PNG refreshes from the regenerator (TTartisan 50/1.2 max + 7artisans 50/1.2 II + new Fuji XF 23 15lp pair). All tests passed. Without the visual check, a regenerator bug producing geometrically-wrong polylines would have committed cleanly and only been caught later via manual inspection of the live chart. The visual check confirmed all 3 traced their source curves correctly before the commit.
Acceptance criteria
Section added to `base/workflow/ai-workflow.md` Lessons Learned (or similar)
Acknowledges that binary diffs are opaque to git diff and tests can be incomplete coverage
Recommends agent-assisted workflows use image-read tools (or open in viewer for human contributors)
Distinguishes from the "regenerate output before formatting" rule already in `base/core/docs.md` §"Generated files"
Where it could live
`base/workflow/ai-workflow.md` §"Lessons Learned" — "Verify agent calculations against the system" already covers similar territory for text output; the binary-artifact case is parallel
`base/core/quality.md` near the debug-statements rule
As a contributor committing regenerated binary artifacts (PNGs, SVGs, screenshots, diagrams), I want documented guidance to spot-check the visual output before committing so that a buggy regenerator producing garbage that still tests-pass doesn't silently land on main.
What
When a regenerator produces binary artifacts (overlay PNGs, diagram SVGs, screenshot snapshots, plot images), `git diff` reports only "Binary files differ" — opaque to reviewers and to the author. The git diff itself proves nothing about correctness.
Pattern to document: open each regenerated binary in an image viewer (or the agent's image-read tool, for AI-assisted workflows) before staging. Confirm:
Cheap check: 1-2 minutes per artifact. Catches the failure mode where a regenerator bug produces invalid-but-syntactically-valid output that tests don't detect.
Why
Observed in me-fuji session 146: two PNG refreshes from the regenerator (TTartisan 50/1.2 max + 7artisans 50/1.2 II + new Fuji XF 23 15lp pair). All tests passed. Without the visual check, a regenerator bug producing geometrically-wrong polylines would have committed cleanly and only been caught later via manual inspection of the live chart. The visual check confirmed all 3 traced their source curves correctly before the commit.
Acceptance criteria
Where it could live
Flagged from me-fuji session 146 wrap. Sibling issue #465 covers the CRLF/whitespace filter step from the same workflow.