diff --git a/backlog/tasks/ml-146 - Honour-Retry-After-and-rate-limit-reset-headers-for-precise-snooze.md b/backlog/tasks/ml-146 - Honour-Retry-After-and-rate-limit-reset-headers-for-precise-snooze.md
new file mode 100644
index 00000000..712ad4cd
--- /dev/null
+++ b/backlog/tasks/ml-146 - Honour-Retry-After-and-rate-limit-reset-headers-for-precise-snooze.md	
@@ -0,0 +1,53 @@
+---
+id: ML-146
+title: Honour Retry-After and rate-limit reset headers for precise snooze
+status: To Do
+assignee: []
+created_date: '2026-04-24 11:12'
+labels: []
+dependencies:
+  - ML-21
+priority: low
+---
+
+## Description
+
+<!-- SECTION:DESCRIPTION:BEGIN -->
+Follow-up to ML-21.
+
+Once workers classify API errors as transient vs permanent, the next refinement is using per-response retry hints instead of fixed snooze durations.
+
+## What each API provides
+
+| API | Header | Format |
+|-----|--------|--------|
+| MusicBrainz | `Retry-After` | seconds (on 503) |
+| Wikipedia REST v1 / Action API | `Retry-After` | seconds (on 429 / 503 from `maxlag`) |
+| Discogs | `X-Discogs-Ratelimit-Remaining` + rolling 60s window | no `Retry-After`; fallback to a fixed backoff |
+| Brave Search | `X-RateLimit-Reset` | seconds until reset (comma-separated for multi-window policies) |
+| OpenAI | `x-ratelimit-reset-requests`, `x-ratelimit-reset-tokens` | duration string (e.g. `120ms`, `2s`) |
+| Last.fm | none | keep existing `ErrorResponse.retry_delay/1` fallback |
+
+## Why this is a follow-up and not part of ML-21
+
+ML-21 gets us from "no classification" to `:retry | :cancel`. That alone lets workers do the right thing with Oban's default backoff. This task layers on precision: turning `{:snooze, 60}` into `{:snooze, retry_after}` so we don't needlessly hammer rate-limited APIs or wait longer than necessary after a transient blip.
+
+## Suggested scope
+
+- Extract a helper that parses each API's reset/retry header into seconds
+- When the classifier returns `:retry`, emit `{:snooze, seconds}` using the parsed value when available
+- Clamp to a sane range (e.g. 5–300s) to guard against malformed headers
+- Unit test per-API header parsing with fixture responses
+
+## Dependencies
+
+Blocked by ML-21 — requires the classification plumbing to exist first.
+<!-- SECTION:DESCRIPTION:END -->
+
+## Acceptance Criteria
+<!-- AC:BEGIN -->
+- [ ] #1 Each API module extracts a retry-delay value from its response (Retry-After, X-RateLimit-Reset, x-ratelimit-reset-*) when present
+- [ ] #2 Workers emit {:snooze, seconds} with the parsed value for transient errors, falling back to a fixed default when the header is absent or malformed
+- [ ] #3 Parsed durations are clamped to a safe range to prevent pathological values
+- [ ] #4 Per-API header parsing is covered by unit tests using representative fixture responses
+<!-- AC:END -->