Signal Playbook Reply Agent

Turns an interested reply's lead into a personalized "signal play" and a ready-to-edit
follow-up email around it.

Pipeline (scripts/build_play.py)

1. research — one Anthropic Messages API call with web search (adapted from the
   SDR-Playbook-Multi-Agent-System company-researcher agent) → data/signal-plays/<slug>/research.md
   following templates/research.template.md.
2. deck-data — research.md → deck-data.json per schemas/deck-data.schema.json,
   validated with node deck-renderer/scripts/validate-deck-data.mjs. All generated
   text is Unicode-sanitized (web-scraped thin/no-break spaces and soft hyphens render
   as gaps inside words on the slides). Length/count overruns (the model can't count
   characters) are clamped mechanically — at the last complete sentence when possible,
   else at a word boundary with dangling connectors stripped — ==highlight== markers
   kept balanced, each fix surfaced in the job log; one LLM repair round-trip handles
   structural issues only. Outreach copy is prompted SHORT (email 200-280 chars,
   LinkedIn 150-250, always ending on a question or offer) — budgets are ceilings.
3. render — npm --prefix deck-renderer run deck (Vite single-file HTML; no PDF —
   the interactive page renders in the viewer's own browser), serialized via
   deck-renderer/.deck.lock → data/signal-plays/<slug>/<Company>-AI-SDR-Playbook.html.
4. publish — the HTML is baked into a per-play coded template
   (templates/signal-plays/<slug>.html via the CMS source-code API, content wrapped in
   {% raw %} with HubSpot's standard include tags added), then a website page is
   created/updated at slug signal-plays/<slug>-ai-sdr-playbook and pushed LIVE
   (POST /cms/v3/pages/site-pages/{id}/draft/push-live — live within seconds). Rebuilds
   republish the same URL. Requires the content scope on HUBSPOT_ACCESS_TOKEN.
   SIGNAL_PLAY_DOMAIN overrides the publish domain; SIGNAL_PLAY_NOINDEX=0 allows
   indexing. Flags url_domain_ok=false if the live URL isn't on everworker.ai.
5. draft — contextualized reply grounded in the ai-sdr knowledge + the thread + the
   play, embedding the live page URL → merged into
   data/interested-replies/followup_drafts.json (agent: "signal-playbook") for the
   normal edit-before-send approval flow.

Render/publish failures fall back to a standard-style draft (no link) so the SDR always
gets something to send.

Usage

python3 .claude/skills/signal-playbook/scripts/build_play.py \
  --reply-id 12345 \
  --contact-json '{"firstName":"Mike","lastName":"Bush","jobTitle":"Director, Strategic Sales","businessEmail":"mike.bush@datadynamicsinc.com","companyName":"Data Dynamics","companyDomain":"datadynamicsinc.com"}' \
  [--skip-publish]

Progress stages stream on stderr (stage: research|deck-data|render|publish|draft);
the final JSON result is stdout's last line. The webui triggers this via
POST /api/replies/followup/regenerate {reply_id, agent: "signal-playbook"} and polls
GET /api/replies/playbook/status/<job_id>.

Env: ANTHROPIC_API_KEY (research/data/draft), HUBSPOT_ACCESS_TOKEN with the content
scope (page publish), SIGNAL_PLAY_DOMAIN / SIGNAL_PLAY_NOINDEX (optional).