ttlequals0/MinusPod v2.94.11

2.94.11

on GitHub

2.94.11

Added

• A sponsor can carry a segment category (Sponsors page, and segment_category on POST/PUT /api/v1/sponsors). When set, every read naming that sponsor or one of its aliases is filed under it. Detection stamps it on LLM markers before the action-aware merge in both passes. Pattern and fingerprint rows read it in place of their stored category. The pass-1 known-sponsor hint names the sponsor with it, and a learned pattern stores it. Re-categorizing a learned pattern never reached the model, because auto-learned patterns contribute only their name to the hint. The existing sponsor category field is unchanged; it holds an industry label.
• SQLite connections log a statement that waits five seconds or longer on the lock, and a write transaction held open five seconds or longer, with the thread name and the statement that opened it. Two "database is locked" bursts failed a dozen RSS refreshes each after the full 30 s busy timeout, which means one connection held the write lock for over 40 s, and nothing in the logs said which.

Fixed

• The auto-process poller no longer loses 30 s at every episode handoff. A finished job writes its status before its finally block drops the queue lock, so the next claim landed in that gap, tripped the same-process acquire rejection, and backed off 30 s. The poller now waits for the lock to drop, in half-second steps up to 30 s (the database busy timeout), before claiming the next row.
• A download 403 is probed once more with the feed User-Agent before it is classified. If that string is accepted, the host is gating on the browser identifier: the episode downloads with the feed string and a warning names both strings, so the operator knows which setting to change. If both draw a 403, the block does not depend on the identifier and lifts on its own, so the episode retries on the normal ladder instead of failing permanently on the first attempt. The failure summary line now includes the message, not only the exception class.
• Same-sponsor merging now also rejoins two detections when the speech between them reads as ad copy: URLs, codes, phone numbers, or offer phrases in at least half of it. A single read the model split around its own call to action stayed split when the middle named no sponsor. Conversation between two name-drops still keeps them apart. The "not merging" decision is logged at INFO so the frequency can be measured.
• Same-sponsor merging treats a zero-length detection consistently: it is never a merge partner in either direction, and one sitting between two reads no longer stops those reads from being compared. Before, an empty detection as the left-hand ad still merged forward across the gap.
• Redirect logging on downloads reports where each hop went. A relative Location header logged as <url> and a scheme-relative one gained a made-up http://. The resolved next-hop URL is logged instead.
• The queue hold block on /status no longer queues every status reader behind one database read. While one thread rebuilds it, or when the read fails, callers get the last good block. A locked database used to stall the SSE stream for the busy timeout and then report an empty hold.
• The active job row in Settings > Processing Queue keeps a gap between the truncated title and its Cancel button. On phones the title ran into the button.
• The status bar's expanded panel lists rate-limited episodes after the pause has lifted but before the requeue tick, instead of a heading over an empty list.

2.94.10

Added

• The two outbound User-Agent strings are now settings, editable in Settings > Data & Security > Outbound Requests and seeded by DOWNLOAD_USER_AGENT and FEED_USER_AGENT. One covers audio, artwork, and chapters; the other covers RSS. They are separate because hosts disagree. Bot mitigation on some CDNs refuses browser identifiers below a version floor that moves as new browsers ship. Some feed hosts do the reverse and answer only a declared podcast client. Working around a host that starts refusing ours no longer takes a new image. A value must be printable ASCII on a single line, at most 512 characters, so an operator-supplied string cannot inject a header.
• GET /api/v1/status and the SSE status stream carry a hold block. It reports whether a rate-limit pause is stopping new claims, the provider's reset time, and how many episodes each hold is keeping. Each offline-queue service also carries the verdict of the last reachability probe.
• The status bar now appears when the queue is holding work with nothing running, which previously looked identical to an idle queue. It names the provider reset time for a rate-limit pause, and the unreachable service for an offline wait. The two stay distinct: a rate-limit hold pauses the whole queue, while an offline wait parks only the episodes waiting on that service.
• Download and availability logs record the URL path and the full redirect chain, each hop with its status code and the final URL. Previously they logged the host alone, so ten consecutive failures left no record of what was being fetched or where the host was sending it. Query strings stay out unless LOG_DOWNLOAD_QUERY or the matching Settings toggle is on. An enclosure query string regularly carries a signed CDN token or a per-listener tracking id, and a log outlives both.
• The offline queue now records each reachability probe it already performs. The status API can then say which service is down and when it was last checked, without probing on the read path.
• Splice check is overridable per feed, under Advanced on the feed's settings page and as spliceVetoEnabled on the feed API. Null inherits the global, true forces the check on, false lets long cuts through without splice evidence. A feed whose ads are spoken straight through rather than joined into the audio has no edit point to find, so the check holds every long cut on it however obvious the ad; turning the check off for that feed alone leaves every other feed's behaviour untouched.

Changed

• The default download User-Agent moves off a Chrome 120 string that hosts had begun refusing.
• The community patterns contributed before segment categories existed now declare category: "sponsor" (196 of 197 files). Installs read the category to decide whether a match is cut, beeped, or kept, and to filter which categories they accept on sync. The corpus no longer relies on the unset-means-sponsor fallback.
• patterns/CONTRIBUTING.md documents the category field and its vocabulary.
• Feed and settings help text is shorter. Several fields explained the mechanism, the tradeoff, and the background where a sentence on what the setting does and when to change it was enough; the detail belongs in the docs. Nine help texts ran past 45 words, now two. The no-password security warning keeps its length on purpose, since spelling out what an unprotected instance exposes is what makes the warning act.

Fixed

• Same-sponsor merging no longer swallows the show content between two mentions. Two detections naming the same sponsor within two minutes merged on that alone, without checking what sat between them. On shows where the host name-drops a sponsor through the episode that absorbed the conversation: across 45 merges in one sample, 26% of the merged span was gap rather than ad, and a 1.9s and a 2.6s detection 83s apart became one 88s span that was 95% talk. The gap must now be filler, measured in speech seconds by the same discriminator merge_ads_across_short_content_gaps already used, unless the gap still mentions the sponsor. Both passes read one min_content_between_ads_seconds setting.
• The evidence check that allows merging across a gap read the ads' own boundary segments along with the gap, because the shared transcript-range helper is inclusive at both ends. When both ads named the sponsor, "does the gap mention the sponsor" answered yes for every gap, so the check could not refuse anything. It now reads only segments lying inside the gap.
• A zero-duration detection is no longer a merge partner. It carries no ad audio, so merging with it only pushed the span end out across the gap; five such merges appear in the same sample.
• A 403 on the audio availability check now fails the episode immediately instead of being retried as a transient CDN error. The check reported every refusal and every not-yet-ready file under the same "CDN not ready" text, so the retry classifier could not tell them apart. A refused download worked through the full retry ladder against a host that answers the same way every time. A 403 is now reported as CDN refused the request (403) and treated as permanent. A 404 stays transient, since a freshly published episode can 404 briefly while its host provisions the media URL.
• Benchmark corpus: two episodes were missing truth spans. One lacked its pre-roll sponsor read (68 seconds); the other had its opening network promo (76 seconds) left rejected, although the same promo is accepted later in the same episode and the reviewer rules count stand-alone promos as ads. Models that detected either span were charged a false positive. Report regenerated from the stored raw calls; F0.5 rises for nearly every model (top score 0.861 to 0.908), the top tier keeps the same 13 members, and docs/llm-providers.md quotes the new figures.
• A community pattern file whose name did not match its sponsor is renamed to the convention the submission validator enforces.
• The community pattern validation workflow installs the dependencies the validator imports, so it validates submissions instead of crashing on a missing module. A crash now says so rather than reporting a rejected file.

Security

• fast-uri 3.1.5 to 3.1.7 (GHSA-5jgf-p345-68v8, GHSA-f65p-4m7j-42xc, GHSA-fph4-wmhf-6fwf, GHSA-jqff-g426-hqxp). Four advisories covering host confusion and server-side request forgery, published after the last release and failing the CI audit gate. Lockfile only; fast-uri is a build-time transitive dependency of workbox through ajv and never reaches the browser.
• browserslist 4.28.2 to 4.28.8, with its data dependencies (GHSA-c83g-rgw3-j3cx, GHSA-73wf-gq98-2v4g). Both advisories cover every version up to 4.28.6, and they were failing the CI audit gate. Lockfile only; browserslist is a build-time dependency of babel and workbox.