User story
The failure classifier's pattern lists only grow when someone happens to notice a misbehaving error in QA or a chat (e.g. Instagram's empty-media response burning the retry budget until it was spotted by hand, #23). Unknown error wordings silently default to transient.
As the operator, I want errors the classifier doesn't recognize to be surfaced to me systematically, so the pattern lists converge from evidence instead of chance sightings.
Requirements
- Error lines that match no classification pattern are collected somewhere reviewable, deduplicated, without re-flagging wordings already adjudicated as genuinely transient.
- Errors classified transient whose retries never succeed are surfaced for investigation: an error line seen repeatedly with zero eventual successes is evidence it is deterministic.
- Runtime download/classification behavior stays deterministic; anything suggested by this mechanism lands through the normal review flow with a test pin.
Filed by Claude.
🤖 Generated with Claude Code
User story
The failure classifier's pattern lists only grow when someone happens to notice a misbehaving error in QA or a chat (e.g. Instagram's empty-media response burning the retry budget until it was spotted by hand, #23). Unknown error wordings silently default to transient.
As the operator, I want errors the classifier doesn't recognize to be surfaced to me systematically, so the pattern lists converge from evidence instead of chance sightings.
Requirements
Filed by Claude.
🤖 Generated with Claude Code