2.94.6

2.94.6

Fixed

• Confirming an ad no longer mints a pattern the auto-learning path would refuse. The correction path called create_ad_pattern directly, skipping every gate: duration, char cap, transition count, sponsor sanitising and brand placement. A confirmed 176s span whose first 87s were show content therefore became a pattern keyed on that content, which then matched the same overshoot on later episodes and invited another confirm. Single-segment confirms now go through create_pattern_from_ad; a span that fails a gate still records its correction for that episode, it just does not generalise.

Added

• Add divider at playhead in the Split ad block window. The existing button drops a divider in the middle of the longest piece, and dragging it from there is unreachable on a zoomed phone waveform.

Changed

• Only the extractor's first-choice sponsor field carries the alias description now; the other six stay as bare slots. On the schema path the model follows the schema, so seven identical descriptions only added tokens and invited the multi-fill the sentence warns against. The serialized schema drops from 432 tokens to 252. From review on #696.

Added

• Set START and Set END buttons in the Detected ad window place a boundary at the playhead. Dragging a pin was the only way to move one, and a zoomed-in waveform puts the pins off screen with no wheel to zoom back out on a phone, which left no way to set a marker there at all. Mirrors the buttons the Mark cue window already had, now sharing one recipe, and both meet the 44px tap-target floor. The shared recipe leaves the text colour to the caller: built on the ghost-button base, its text-foreground raced the caller's colour and the Mark cue window's END button had been rendering grey instead of red.

2.94.5

Changed

• The sponsor alias fields in the detection schema carry a shared description saying what they hold and to fill at most one. The prompt never mentions them (they exist so a key-stripping backend cannot discard whichever alias a model volunteers), so for a schema-reading model the description is their only context. From review on #696.
• The detection and verification passes DEBUG-log the fully assembled system prompt, not just its length (#708). Dynamic sections are appended after the stored setting, so the settings UI cannot show what was actually sent; the episode run log is where prompt debugging happens.

2.94.4

Changed

• The detection window schema constrains category to the seven segment categories the prompt mandates, the same way the category-repair schema already does. Derived from SEGMENT_CATEGORIES, spelled as a plain enum for backend portability, with everything else left optional so both addressing modes still validate. From review on #696; the reviewed proposal's four-value list would have forbidden intro, outro, and recap labels on an enforcing backend.
• anthropic 1.0.0 to 1.2.0 and openai 3.3.1 to 3.6.0 (#699, #706, taking the current release rather than the one the bump was opened against). No call-site changes needed; pip-audit is clean on the new pins.

2.94.3

Fixed

• Pattern learning threw away most of what it saw on ad-heavy feeds. A span over the duration ceiling was dropped whole, so the reads were cut but never learned and every episode re-detected them from scratch. One feed lost 26 spans that way in eight hours, ranging 120 to 442 seconds. A span over the ceiling is now split at its ad transitions, and each piece that names its own advertiser is learned separately. A piece with no sponsor of its own is skipped rather than labelled with the previous read's name. The ceiling and floor are settings rather than literals buried in the function.
• The most common ad opener in podcasting was rejected as contaminated. "This episode is brought to you by" also contains "brought to you by", and the contamination check counted phrase-list entries rather than positions, so an ordinary single read scored two transitions and was dropped. It counts positions now, which is what the existing helper already did for the manual split path. The contaminated-patterns list had the same miscount, so it flagged clean patterns and then refused the split it recommended.
• Segment names reached pattern learning as sponsor names. "Outro", "Show" and "Episode" are structure, not advertisers, but no blocklist knew that, so they were stopped three checks later by an unrelated intro test that blamed contamination. The validator written for this case is now called from both learning paths and knows the segment categories.
• A sponsor had to appear in the ad's opening sentences, matched as a raw substring. Real advertisers stored under one spelling and spoken as another failed it. The name must appear somewhere in the read, matched through the same alias-aware check the next gate already used.
• The reviewer's prose/number warning fired on context. Its regexes key on boundary words rather than the sentence's subject, so "show content starts at 59.7s", given as the reason for an end trim, read as a claim about the ad's start. Four of five warnings in a production sample were that shape. A figure landing on any real boundary is treated as context now.
• The episode page discarded the reason a reprocess was refused (#707). Processing is serialized by a lock, so an episode is often mid-run and the request returns 409, but the handler set an error state that reached no visible surface. The button flickered and nothing appeared to happen. It now shows what the server said, in the same banner corrections use.
• The feed page silently swallowed failures from refresh, rename, reprocess-all and bulk actions, none of which had an error handler at all.
• Two selects in the feed page's episode filter could not wrap and ran off the side of a 320px screen.
• A correction whose bounds no longer matched a stored marker (a recut or reprocess can shift markers past the 0.5s match tolerance while the list is open) recorded the decision but never stamped the episode, so an adjust or reject silently never reached the audio. An unmatched correction now stamps; an unneeded recut is idempotent, a skipped one loses the edit.
• A decision recorded while its episode was mid-run was cleared when the run finished, though the run's cut list never saw it. Completion now clears only stamps from before the run started, and a decision landing on a processing episode stamps with a fresh time instead of keeping the old one.
• The Detected Ads tab's editor never learned the marker's category or applied action, so a kept segment's editor still offered "Not an ad" (a guaranteed 409) and showed the wrong category. The keep guard covered Ad Review only.
• The C and R keyboard shortcuts skipped the kept-by-category guard: the buttons were hidden but the keys still submitted the refused verdict.
• Apply recuts overwrote the mode of an episode already queued for a full or llm rerun, downgrading the requested rerun to a recut of stale detections. Queued and processing episodes are skipped; their own run applies the decisions and clears the stamp.
• One user-requested episode lifted the rate-limit pause for the whole queue, and claims go by priority, so every backlog row ahead of it burned one call into the throttled provider. While paused, only user-requested rows are claimed.
• A full or reprocess retry re-ran Whisper on every attempt after a transient failure. A retry now reuses the transcript an earlier attempt of the same request saved, when the retained original guarantees the audio is the same file the transcript came from.
• The Apply recuts button read "Recutting" and stayed disabled when a new decision arrived while a batch ran, and could stick that way. The server now reports each row's state (recutReady, inFlight on GET /episodes/pending-recuts) and the bar renders from that: rows being rebuilt, rows a fresh apply would queue, and rows missing what a recut needs, which wait for a full reprocess. An episode left 'pending' by a cleared queue counts as applyable, since no run of its own is coming.
• A category change the server refused kept showing the new category in the list until an unrelated refetch put the truth back.
• Learning could store a span whose opening read belonged to a different advertiser: the labeled brand only had to appear twice anywhere. It must now first appear in the front 60 percent of the read.
• Two workers recording response-format probe answers in the same window could drop one model's answer; the read-merge-write now runs in one immediate transaction.

Added

• Apply recuts on a feed's own page, covering that feed's episodes rather than every waiting episode. GET /episodes/pending-recuts takes an optional slug, and the apply endpoint takes one in its body. There is still one recut queue and each episode is still rebuilt once.

2.94.2

Fixed

• The Ad Review queue listed segments that no one could act on. A segment left in by its category is settled by the feed's segment actions, and the corrections endpoint refuses a verdict on one, so "Needs review" was offering a decision nobody could make and the queue could never be emptied. Those segments appear under their cut status instead, and the queue holds what needs a decision.
• Every row in the queue carried an "Unresolved" badge, which restated the filter and said nothing. Only a recorded decision gets a badge now. The state is called "not reviewed" where it still needs a name, and the badges are explained in the web interface docs.
• The Ad Patterns header pushed Export off screen on a phone: the row could not wrap, and the sync stamp had grown wider. It wraps now, and the stamp shows the date only where there is room for it.
• Detection cards lay out in two deliberate rows rather than by wrap order: the verdict pair on top at equal width, category and Edit below. Card actions meet the 44px tap-target floor in both dimensions, up from 36.
• Changing a category in the Detected ad window did nothing at all on the Detected Ads tab, which ignored that kind of change. On Ad Review it saved but the control snapped back to the old value, because it was bound to a list row that had not refetched yet.
• The Apply recuts button stayed lit after queueing, as if it had not been pressed. It reports what it started and clears itself as episodes finish.
• Apply recuts blamed missing original audio for every skipped episode. The endpoint returns counts, not reasons, and a skip can equally mean the saved transcript segments are gone. The message names what a recut needs without inventing which piece is missing.

2.94.1

Fixed

• Switches were squashed on phones. A control with a width class is still a flex item, so a long label beside one steals its width once the row stops fitting: at 360px a switch next to a two-line label lost a third of its width. The row only overflows on narrow screens, which is why every desktop review passed it. The switch now holds its size, and the design guide records the rule for every fixed-size control in a flex row.
• The queue pager wrapped onto two lines on a phone. It now stays on one, showing first, current, and last below the small breakpoint.
• Import, Export, and the community sync stamp on the Ad Patterns header were three different sizes. They share one recipe now, and the stamp shows a date instead of a full timestamp that wrapped to three lines.
• The play button on a detection table row was 33px against 30px neighbours. Both come from one shared recipe now, as the card and marker rows already did.
• Saving on a detection whose category resolves to keep returned a 409. The modal offered Save and Not an ad on a marker the corrections endpoint will always refuse; it now shows the category picker as the way through and says so.

Added

• Set a detection's category from the Ad Review and Detected Ads rows, and a pattern's category from the Ad Patterns table, without opening the editor. Review is bulk work, and the category is what decides whether a span is cut.

2.94.0

Added

• Rate-limit queue hold. When the LLM provider answers 429 with a reset more than five minutes out, the episode waits and the queue pauses until the reset instead of burning retries on a throttled provider. Shorter resets keep the existing in-process retry, so a lone throttled window still recovers. Detection, review, and verification are all covered: a throttle arriving mid-run defers the episode rather than skipping that stage. Off by default, with a give-up window of 1-720 hours (default 48), under Settings > AI & Processing > Queue Control. Play and Reprocess run even mid-pause, and turning the toggle off lifts the pause and releases held episodes. A held episode never inherits the clock of an earlier offline deferral.
• Per-episode priority control in the Processing Queue panel. Each waiting row gets a priority field with -/+ buttons, backed by POST /feeds/{slug}/episodes/{episodeId}/queue-priority. Send priority for an exact value or delta to nudge the stored one; a delta is added server-side, so a click made against a stale list value still lands. Either can lower a priority, unlike re-enqueueing; a feed-level queue priority change still restamps the row.
• The Processing Queue waiting list is paginated at 25 rows per page, so nothing hides behind a "+N further back in the queue" note. GET /episodes/processing takes offset and limit (default 200, cap 1000), and positions stay correct across pages.
• Queue Control, a new section in AI & Processing. It groups the queue priority boosts (moved out of Global Defaults), the process-new-episodes toggle, the offline queue (moved out of Data & Security), and the rate-limit hold.
• Transcript Normalization is now its own section in AI & Processing, with help text explaining that the rules correct Whisper output for words, phrases, numbers, sponsor names, and URLs. The Sponsors page drops its Normalizations tab.
• A detected ad's category can be changed from the Detected ad window. The category decides whether a span is cut, beeped, or left in, so this is how you cut one the feed is currently keeping. It changes that episode's marker only; a linked pattern keeps its own category, edited in the pattern detail modal.
• Review decisions no longer re-cut an episode one at a time. Confirming, rejecting, or recategorizing records the decision and marks the episode, and an Apply recuts button on the Ad Review and Detected Ads pages cuts every waiting episode once. An episode edited five times is rebuilt once instead of five times. An episode that no longer has its original audio keeps its decisions, and the button says how many it could not recut.
• Opt-in JSON schema response format for OpenAI-compatible providers (#693, #694). Detection, review, category repair, and trim recovery send a json_schema response_format once the toggle in the LLM Provider section is on. Support is probed and remembered per model rather than per endpoint, because one URL can serve models that differ on it; a model that does not support it falls back to json_object at request build or on a runtime 400 instead of losing the format hint. Anthropic call sites are unchanged.

Changed

• Benchmark scorer now canonicalizes predictions and ground truth to per-break spans (gaps under 15 seconds merged) before IoU matching, matching the detection prompt's merge rule. Report regenerated from the stored raw calls; per-model scores shift accordingly and are not comparable to pre-2.94.0 rows.

Fixed

• A full or LLM-mode reprocess no longer wipes the episode's transcript and ad markers up front (#692). The clear now happens in the transcribe stage, immediately before the fresh transcript is saved, so an OOM kill or container restart mid-run leaves the prior results intact instead of emptying the episode.
• The ad reviewer's system prompt examples now show the same shape the reviewer actually sends (#695): candidate markers, and a [start-end] stamp on every line including the 60 seconds of context on each side. The model can now read a trim boundary off a context line instead of interpolating one. Installs still on the shipped default get the corrected examples on upgrade; customized prompts are left alone.
• fpcalc fingerprinting survives recoverable decode hiccups (#690). All three fingerprint call sites parse stdout before honoring the exit code, so an episode with one bad frame no longer loses all cue scanning.
• A detected ad whose category resolves to keep no longer shows Confirm and Not-an-ad buttons that the API always refused with a 409. The row says the span is left in because of its category and offers Edit, which is where the category can now be changed.
• Correction failures show what the server actually said instead of a fixed "Try again", which sent you in circles on a refusal that would never succeed.
• The offline queue's "waiting" count included only LLM deferrals, so a Whisper outage read as zero episodes waiting. It now counts every deferral it owns.
• A single model rejecting plain JSON mode no longer downgrades every other model on the same OpenAI-compatible endpoint to prompt injection, across restarts. Both response-format answers are now remembered per model.
• The play button on ad review cards, detected-ads rows, and the episode page's held and rejected marker rows matches the height of the buttons beside it at every breakpoint, instead of sitting short on mobile.
• The settings toggle is slimmer, matching the switch spec now written down in the design guide.
• 3- and 4-digit queue positions no longer paint over the episode title on mobile. The position column is wider, and the row stacks its controls below the title on narrow screens.
• The cue-template create endpoint's docstring no longer claims to accept a required label (#691); the display label comes from cueType, so the LLM prompt always sees a fixed phrase.