The highest-impact fix: run_research_job() only ever read entry['id'] from
confirmed competitors, silently discarding the Firecrawl-matched or
manually-swapped URL — every scrape fell back to the competitor's bare
root domain regardless of what was confirmed. This explains most of the
"unpredictable/wrong data" behavior seen against the old app.py. Threaded
confirmed_url through jobs.py -> analysis_service.py so the actually
confirmed page is what gets scraped.
Other fixes bundled in this pass:
- Proxy layer: Webshare rotating-gateway country code lowercased (was
silently failing auth), Bright Data port corrected to 33335 (current
cert), dedicated-IP candidate rotation added before falling back to the
shared gateway/Bright Data escalation.
- Trip-intent matching: score_and_rank_urls() now hard-gates on
destination keyword and excludes bare root-domain candidates, instead
of letting duration/category signals alone push a wrong-country or
homepage URL over the confidence threshold.
- force_refresh cache-bypass toggle added (opt-in, collapsed "Advanced
options"), plus a "Skip this competitor" affordance for unmatched URLs.
- De-hardcoded "G Adventures" (the hackathon reference client) out of the
extraction prompt — now uses the real tenant's company name.
- RunHistory/Apps-Script "latest results" queries filtered to
triggered_by = "manual" so per-competitor content-hash bookkeeping rows
no longer drown out real batch runs.
- Dashboard fixes: stuck-state bug on revisiting the Analyse page,
AccountPage/BattlecardsPage restacked to single-column layout, Alerts
moved to a header bell dropdown, invisible Export/Run Analysis icons
(fill -> stroke), tour popup width mismatch, horizontal-scroll bug on
cards (missing min-width: 0 in flex layout).
- New PocketBase migrations for scrape_runs (content_hash fields,
tab_type, created_by_email) and products (departures, start_days).
335 backend tests passing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Implements CLAUDE.md Build Order Phase 2 — same scraping/extraction
logic as the legacy app.py, reorganized into the three-layer structure
and upgraded per §10's explicit instructions:
- core/scraper.py — playwright-stealth replaces the manual
add_init_script stealth block; smart proxy (Webshare→Bright Data)
wired via core/proxy_service.py with block-detection + auto-escalation;
check_robots_txt() gate added per §30
- core/extractor.py — OpenRouter model cycling, merging the legacy
JSON-repair fallback chain with the spec's env-driven free/paid
model list and Sentry breadcrumb on paid-fallback use
- core/ai_fetch.py — AI web fetch fallback, ported unchanged
- industries/adventure_travel/{fields,prompt}.py ported unchanged;
schema.py added as the future V2 custom-fields attachment point
- config/client.py — gspread Sheet write-back, ported unchanged
- services/analysis_service.py now sources CACHE_STALENESS_HOURS from
core/scraper.py, its canonical location per §7
Dropped the legacy's commented-out Groq/Gemini alternates — OpenRouter
is the sole active provider and the dead code added nothing.
requirements.txt: relaxed playwright to >=1.55.0 and pinned
playwright-stealth==1.0.6 + setuptools<81 — Python 3.14 has no
prebuilt wheels for the originally-listed pins, and playwright-stealth
2.x replaced the stealth_async(page) API §10 documents verbatim with a
class-based one.
192 tests, 94% coverage. Also fixes a real bug surfaced by this phase:
several test helpers mutated core.extractor / industries.adventure_travel.prompt
via raw setattr instead of monkeypatch.setattr, silently corrupting
shared module state across tests once those modules became real.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
JasonFraser