- analyze_creative_deep was truncating mid-JSON ~50-75% of the time
(max_tokens=700 too low once the analysis text runs long), silently
falling back to score=0/"Error parseando respuesta." Confirmed via
stop_reason=max_tokens on real API calls. Raised to 1400, and bumped
analyze_creative/compare_adset_creatives (600→900) as the same risk
applies there. Verified 6/6 clean parses after the fix (was failing
~3/4 before).
- Dashboard tabs "Campañas", "Familias", and the per-day breakdown in
"Por día" still showed a single Meta-only Leads/CPL number; added the
Leads/CPL según Airtable (leadform+landing) columns alongside, matching
the Slack report so the dashboard never shows a different figure than
what's already trusted from the daily report.
- Validated both end-to-end with real data: dashboard via Streamlit's
AppTest headless runner (0 exceptions across all 5 tabs), creative
analyzer via a live run against one campaign (real scores, fatigue
detection, comparison, Baserow save, Slack send all confirmed).
meta-optimizer (Viviful) and this project share one Slack bot/app, but a
Slack app only supports a single Interactivity Request URL — so a single
approval_server must be able to tell which project's Baserow a button click
belongs to. Button values now carry a project tag ("approve:formacion:<id>"
instead of bare "approve:<id>"); approval_server.py parses it and routes
accordingly, with a registry slot to add other projects (e.g. "viviful") via
their Baserow REST credentials without importing their code.
Confirmed via api.slack.com/apps that the current Interactivity Request URL
points at a dead ngrok tunnel (ERR_NGROK_3200) — no approval server is
actually reachable in production today, for either project. Deployment is
being handled separately as part of the general app rollout.
- Rename table abbreviations that Slack misread as domains: 'CPL.AT' was
auto-unfurled as a link to the .AT (Austria) TLD, posting an unrelated
business's ad copy into the channel. Switch '.' to '·' in all L/€/CPL
abbreviations, and set unfurl_links=False/unfurl_media=False on every
postMessage as defense in depth.
- Pass ppl/cpa_maximo into adset analysis too (ads already had it) so
Claude's adset evaluation compares CPL against the course's PPL-derived
rentability threshold, not just Meta's own bid cost cap.
- Add PPL and Margen columns to the adset numeric table in Slack.
- Broaden Airtable lead counting to attr_utm_source IN ('Lead ads','landingpage')
— the 'landingpage' leads (100% fbclid, 0% gclid) were being missed entirely,
undercounting real leads for '_web' suffixed campaigns and skewing
capping/pacing decisions since yesterday's first production run.
- Add airtable_client.get_meta_leads_bulk() for day/curso-level aggregation.
- Drop per-familia Slack sectioning in favor of a single Formación block,
chunked by campaign batches instead.
- Add daily AT-vs-Meta table, per-curso PPL/CPL contrast table (leadform vs
landing breakdown), and a Claude-generated portfolio strategic diagnosis
(ported from leads-optimizer's portfolio_daily_analysis).
- Persist daily aggregate totals to a new Baserow table (daily_metrics) so
the dashboard and future reports don't depend on Meta's historical API
access remaining available indefinitely.
- Surface adset/ad-level recommendations in the campaign cards instead of
only numeric tables.
Windows console encoding crashed the script right after creating the first
table; make it re-runnable by skipping tables that already exist instead of
aborting entirely.
Ports meta-optimizer's Meta Ads execution/approval/creative-analysis layer
(agent.py, meta_ads_client.py, baserow_client.py, slack_notifier.py,
approval_server.py) and replaces the per-vertical CPL model with the
PPL + monthly-capping-per-course model already used by leads-optimizer,
via a new airtable_client.py that shares Cursos/Familias/CentroCurso/
CursoMes/Leads Lake with that project and adds Meta Ads Campaigns /
MetaCampaignMes alongside its Google Ads Campaigns / GACampaignMes.
