What fails#
From the 2026-11-17 21:51 UTC poll, Perplexity reports ingest_failed: perplexity: models page states a reasoning model in an unread shape (source https://docs.perplexity.ai/docs/agent-api/models.md, PostHog project 643425). It is a docs-row failure, so it repeats once per docs TTL and leaves the reasoning flag unavailable for every Perplexity row on that page.
Why#
The page has a new DeepSeek card:
DeepSeek V4 Pro, an open-weight reasoning model.
parsePerplexityModelsPage (src/server/providers/adapters/perplexity.ts:157-175) reads only <names> — …reasoning model… or <names> is …reasoning model…, so the appositive comma form throws unread('a reasoning model').
The other cards still parse, for example GLM 5.3 and GLM 5.3 Flash — Z.AI reasoning models. and Nemotron 3.5 Super is a reasoning model with a 1,048,576-token context window….
Fix#
- Accept
<names>, an? …reasoning model. as a stated reasoning flag for the tab's ids. Keep every other wording throwing (fail-closed).
- The same sentence states
open-weight. Read that into openWeights: true only if the adapter already reads open weights from Perplexity's page. Otherwise note it here and leave it alone.
- Add a test with the DeepSeek V4 Pro card.
Fix step 2: the Perplexity adapter reads
openWeightsnowhere today (no reference insrc/server/providers/adapters/perplexity.ts), so the card's "open-weight" is left unread, as the issue says. The Nemotron 3 Ultra card states the same thing and is also unread. Reading it is a separate change.