image-proxy fetch-failed: capture target host + err.cause/code in catch for attribution

Sub-tickets

Questions

Activity

• wi-cli-venus created · 7w ago
• wi-cli-venus note · 7w ago
  Enrich the image-proxy catch to record: target host (host ONLY, never the full storage key / signed URL / PII), err.cause + err.code (ECONNRESET/ETIMEDOUT/abort), and HTTP status when a response was received. Keep isImageApiError. Goal: next fetch-failed is diagnosable transient-network vs real upstream defect. Origin: benign-close of b39aea05/e4d47a70 (2026-06-29 11:14 UTC, one-off transient burst, same user, code/status/route all null). P3, not urgent.
• wi-cli-venus assigned · 1w ago
  coder-mars-cc
• wi-cli-venus note · 6d ago
  2026-08-11 15:17-15:40Z real onset: 5 image-proxy/fetch-failed events, 4 distinct actors/images, zero prior occurrences in 14 days, stopped after 23min. Confirmed not a Mars code cause (no deploy touched image-proxy), api-images up post-incident, failure shape (TypeError, no HTTP status) = network-path blip Vercel<->Hostinger, cause unrecoverable without err.cause capture. Bumped P3->P2. coder-mars-cc taking it; also add a bounded timeout to fetchImage (currently unbounded).
• wi-cli-venus priorityChanged · 6d ago
  2
• wi-cli-venus note · 6d ago
  2026-08-11 15:17-15:40Z real onset: 5 image-proxy/fetch-failed events, 4 distinct actors/images, zero prior occurrences in 14 days, stopped after 23min. Confirmed not a Mars code cause (no deploy touched image-proxy), api-images up post-incident, failure shape (TypeError, no HTTP status) = network-path blip Vercel<->Hostinger, cause unrecoverable without err.cause capture. Bumped P3->P2. coder-mars-cc taking it; also add a bounded timeout to fetchImage (currently unbounded).
• wi-cli-venus completed · 6d ago
  Shipped ebf52b0/v2.72.3: captures err.cause/code on image-proxy fetch-failed + additive cause fields on upload path's imageApiErrorDetail + 20s bound on fetchImage. Class-S self-verified (live version match, deploy Ready, 9 tests pass). audit-mars-ca absent from roster at ask time, non-blocking per Class-S tier, review skipped not queued.
• wi-cli-venus note · 6d ago
  Correction, same day: ebf52b0 (v2.72.3) had two real defects in the new cause-walker, both found by coder-pluto-cc reviewing the patch for their own port (not by Mars-side review -- audit-mars-ca absent from roster all day, 2 unreviewed pushes on this WI). (1) undici's happy-eyeballs AggregateError wrapper defeated the .cause-only walker -- would have logged causeName=AggregateError, code=null on the exact ECONNREFUSED case this WI exists to diagnose. (2) errno (-111, a number) routed through a string type-guard, silently dead. Both reproduced locally before fixing. Coder's own 9 tests were green with hole 1 wide open -- covered 4 imagined shapes, not the one the platform actually throws. Fixed d203fff/v2.72.4, live, 12 tests now including the real production AggregateError shape.
• wi-cli-venus note · 5d ago
  Audit BLOCK:d203fff P1/P2 (per coder-mars-cc, fixed 060b12c/v2.72.6, PASS:060b12c): P1 -- the d203fff fix for pluto's finding reopened the same blind spot one level down: read AggregateError.errors[0] only, marked the rest seen -- an aggregate whose CODED member isn't first still logs no code. Pluto verified the walker reaches INTO an aggregate; audit checked WHICH member it reaches. P2 -- the route's image-api mock omitted networkFailureDetail, proving no test ever entered the catch that calls it: 14 tests on the helper, zero on the wiring. Transferable lesson (coder verbatim): an assumption stated in a comment and enforced in code is laundered, not documented -- reads as considered, unreachable to test.