News

Ideas

• issue #6 Add fox as primary AI summarizer with cube fallback

• Summaries run through a single Ollama backend (cube/qwen3 via OLLAMA_URL). A single backend means summarization stalls whenever that host is down, and it can't take advantage of fox (gemma4:26b), which produced tighter summaries at comparable latency in a 4-article comparison. This makes fox the primary summarizer with cube as an automatic fallback.

  Solution

  fox is served by llama-swap and speaks only the OpenAI-compatible /v1/chat/completions API — it 404s on Ollama's /api/chat — so a config-only switch isn't possible. This adds an Analyzer interface with two implementations (OllamaClient, OpenAIClient) sharing the Norwegian prompt and JSON-extraction logic, plus a FallbackAnalyzer that tries the primary first and retries the fallback on any error (but not on context cancellation / shutdown).

  To avoid a config trap, the existing OLLAMA_URL/OLLAMA_MODEL keep configuring the fallback backend — so the deployed unit, which already sets OLLAMA_URL=cube, needs no change. A new AI_URL/AI_MODEL/AI_API (default http://fox:11434, gemma4:26b, openai) configures the primary. The Ollama client timeout is raised 30s→60s to tolerate cold model loads (~48s observed).

  This re-ports work originally written against an orphaned repo line (old PR #4, preserved on orbit-main-archive) onto the canonical codebase.

  Verification

  go build + go vet clean; all AI/config tests pass. The 4 pre-existing failures on this branch (snapshot/desk OIDC tests) are unchanged from main — not introduced here. The OpenAI/Ollama client paths were validated live against fox and cube earlier.

  Deploy note

  Not for deploy until approved. On deploy, summaries switch cube→fox automatically (fox is the default primary); cube remains the fallback via the unit's existing OLLAMA_URL. No unit change required.

  Closes #5

  See on code.bas.es →

History

• 20:02 Update stale desk and snapshot tests to current behavior

• Four tests fail on main, all stale relative to intentional code changes — no product bug.

  Root causes & fixes

  • TestDB_SnapshotArticles — SnapshotArticles filters content != '' (eea4578). Give the articles content.
  • TestServer_DeskSnapshots / _HTMX — /desk/snapshots now 302-redirects to per-outlet /desk/snapshots/{site} (8b622e0). Hit that URL, align to the 10-min snapshot tick, give the article content, and trigger the HTMX partial via HX-Target: snapshot-panel (the handler's actual contract).
  • TestServer_DeskBreaking — /desk was rebuilt from a 'Breaking' list into the 'Kandidater' ranking view (ab324d7). Set up a qualifying candidate (pre-cutoff snapshot at 05:00 today, summary, non-empty keywords) and rename to TestServer_DeskKandidater. The 05:00 snapshot is at/before either edition cutoff (06:00/18:00) and within 24h of both, so it qualifies regardless of run time.

  Verification

  go test ./... is fully green.

  Closes #13

  See on code.bas.es →

• 19:50 Fix VG and Aftenposten scrapers for Schibsted redesign

• Both vg.no and aftenposten.no migrated to the same Schibsted frontend with build-hashed CSS-module class names, so the homepage scrapers' selectors matched nothing and both returned no articles found.

  Change

  • New shared parser scrapers/schibsted.go keyed on stable hooks: a[data-content-type="article"] anchors for links and the teaser heading for titles.
  • Links are host-filtered per site (drops cross-promoted e24 content and video/event teasers) and deduplicated by URL with tracking params stripped.
  • Titles read the visible heading with <br> turned into spaces — this reconstructs hero headlines split across styled spans (otherwise Drittleidødtid) and avoids the Premium, kicker Schibsted folds into aria-labels.
  • vg.go and aftenposten.go now delegate to the shared parser.
  • Refreshed VG and Aftenposten homepage fixtures to the current markup.

  Article-page content extraction (.article-body p / article p + og: meta) was verified still working and is unchanged.

  Verification

  • go test ./scrapers/... passes; fixtures yield 19 (VG) and 14 (AP) articles.
  • Live run against the real sites: VG 20 articles, Aftenposten 14, with clean titles.

  Note: four pre-existing failures in the news package (db/server desk tests) are unrelated to this change and fail on main as well.

  Closes #11

  See on code.bas.es →

• 19:14 Switch AI primary to qwen3.6:35b and raise timeout

• gemma4:26b was removed from fox, so the configured primary (AI_MODEL default gemma4:26b) now errors and every summary falls back to cube. This switches the primary to qwen3.6:35b, which produced richer, more accurate Norwegian summaries than gemma4 in a 4-article comparison (concrete scores, names, and nuance gemma left out).

  Why the timeout bump

  qwen3.6:35b is a 35B reasoning model — ~2.5× slower than gemma4, and it exceeded the 60s OpenAI client timeout on 2 of 4 test articles (up to ~79s). Left at 60s, long articles would time out mid-generation and fall back to cube — paying qwen's latency without getting its output. The client timeout is raised to 120s. The app is an async background summarizer, so the higher latency is acceptable.

  cube (qwen3:14b via OLLAMA_URL) stays as the per-request fallback.

  Verification

  Build + vet clean; AI/config tests pass. The 4 failing tests (snapshot/desk OIDC) are pre-existing on main, unchanged. qwen3.6:35b was validated live against fox during the comparison.

  Closes #9

  See on code.bas.es →

• 08:11 Deploy strictly from git, not a host checkout

• deploy.sh built with go build . from whatever working tree it ran in, so production depended on one host's checkout (servo). That is precisely how the real source ended up trapped off-git. This makes git the source of truth for deploys.

  Solution

  deploy.sh now clones the canonical remote at a given ref (default main) into a temp dir, builds linux/amd64 there, deploys, and cleans up. Because it builds from a fresh clone, only committed-and-pushed code can reach production; local/unpushed trees are never deployed. The script is host-independent — runnable from any machine with git, Go, and SSH — and prints the exact deployed SHA.

  Usage: ./scripts/deploy.sh [host] [git-ref] (defaults: fismen, main). The host-managed unit (incl. the OIDC drop-in) is left intact; the embedded unit is only written on a fresh install.

  Verification

  Validated the clone+build path from git end-to-end (no deploy): clones origin/main, builds the 21.4 MB linux/amd64 binary. Syntax-checked.

  Follow-ups

  • Forgejo Actions CI/CD (push/tag → auto-deploy) can build on this; needs a runner + SSH secrets.

  Closes #7

  See on code.bas.es →

• 23:16 Use fox for AI summaries with cube as fallback

• AI summaries went through a single Ollama backend (cube, qwen3:14b). We want fox (gemma4:26b) as the primary summarizer — its output is tighter and at least as fast in a 4-article comparison — but a single backend means summaries stall whenever that host is down. This makes fox primary and keeps cube as an automatic fallback.

  Solution

  The two hosts speak different protocols: cube is Ollama-native (/api/chat), while fox is served by llama-swap and only exposes the OpenAI-compatible /v1/chat/completions (it 404s on /api/chat). So a config-only switch isn't possible. This introduces an Analyzer interface with two implementations — OllamaClient and a new OpenAIClient — sharing the Norwegian prompt and JSON-extraction logic so swapping models never changes what we ask for. A FallbackAnalyzer wraps the two: every article tries the primary first and retries on the fallback on any error, so a recovered primary is used again immediately with no cooldown bookkeeping. It does not fall back when the context is cancelled (shutdown).

  Backends are configured independently via AI_* and AI_FALLBACK_* env vars (URL, model, protocol, optional Bearer key); defaults are fox-primary / cube-fallback. The HTTP timeout is raised from 30s to 60s — the comparison showed cube cold-loads taking ~48s, which the old timeout would have killed. Verified end-to-end against the live fox and cube backends.

  Known cuts

  • Per-request retry only — no circuit-breaker/cooldown, so a hung (not refused) fox costs its full timeout before each fallback. Refused connections fail fast.
  • The summary prompt is unchanged.

  Follow-ups

  • Confirm fox/cube are reachable from the deploy host (fismen) over tailscale before this goes live; an older deploy comment noted Ollama wasn't reachable there yet.

  Closes #3

  See on code.bas.es →

• 08:22 Replace tea with forge in merge instruction

• forge is the unified git-forge CLI that replaces tea locally. forge pr merge hits the same Forgejo API endpoint tea pr merge did, so the merge-via-Forgejo-API constraint is preserved.

  See on code.bas.es →

• 07:27 Initialize orbit with design system and conventions

• Summary

  • Initialize design/ directory via orbit with the news design system (tokens, fonts, components, previews)
  • Add Forgejo webhook for issue/PR sync to orbit
  • Apply conventions from arne/conventions (CLAUDE.md, docs/conventions/)
  • Add M↓ markdown button to candidate eyebrows and article view
  • Add prototype header with orbit navigation to all preview pages
  • Rename Breaking to Kandidater across all desk previews

  Test plan

  • [ ] Open design/index.html — verify all component sections render (colors, typography, buttons, badges, candidates, clusters, etc.)
  • [ ] Open design/preview/desk.html — verify M↓ button appears in candidate eyebrows
  • [ ] Open design/preview/desk-article.html — verify M↓ has no underline
  • [ ] Verify prototype header links back to orbit from all preview pages
  • [ ] Verify dark mode toggle works across all pages

  See on code.bas.es →