client._chat() JSON-parses every response (for the classifier), so the plain-text
summary was rejected ("model did not return JSON") even though the model returned
a perfect summary. Split out _raw_content() and add chat_text() for free-form
output; summaries use it. _chat keeps parsing for classification.
The /a/<id> page now carries an original short summary so it stands on its own,
without republishing the publisher's article:
- summarize.py: transient SSRF-guarded fetch of the article text → local LLM
writes a 2-4 sentence ORIGINAL summary (our words). Cached in article_summaries
forever; we store only our summary, never the body. Generated lazily (only for
shared/viewed articles), de-duped so concurrent hits don't double-generate.
- /a serves cached-or-pending; when pending it shows a calm "summary on its way,
read at {source}" note and self-polls /api/summary/<id>, swapping the summary
in the moment it's ready (never blocks the page on the batch-tier LLM).
- Share menu warms generation on open so recipients usually get the rich version.
- Container reaches the arbiter at arbiter:8080 over caddy_web (LLM env added to
the API container). 124 tests pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
thejayman77