feat(broadcasts): the 15th module — WhatsApp campaigns over the flow API

[OlegKrasikov]

What

The broadcasts module: WhatsApp campaigns shaped like an email-marketing tool, over Chatfuel's flow API.

• Campaigns — every campaign-shaped flow on the bot, with a status derived from the entry-point element and the clock (draft / scheduled / sending / sent), filters, search, an inline panel, row actions, ⌘K palette and keyboard bindings.
• Composer — five steps over one server draft: name + kind (now / later / repeating), message (approved template from the catalog, blanks filled, {{Attribute}} inserted from the bot's own list), audience (two-level AND/OR builder with a live recipient count), schedule (bot wall clock, once / weekdays / every N days / on dates, weekdays corrected by zone), review (send behind a confirm, or arm). Duplicate replays a campaign onto a fresh draft.
• Templates — the read-only catalog with Meta status and a hand-off to WhatsApp Manager. No public API authors a template.
• A bot with no WhatsApp number sees a connect state on every surface, never the server error.
• The skill: five references, two playbooks, 36 operations — everything the live pass found (no stats, no paused state, an unknown attribute name creates the attribute, a template block cannot be deleted, unregistered errors are bare InternalServerError).

Review pass

Four reviewers over the branch before this squash; fixed here:

• kind switch dropped the audience; a Repeating draft could never be set to Once; a refused first pick showed nothing
• writes on one blank could race (blur vs insert); two schedule commits could overlap
• the "Broadcasts" flow group could be created twice; a refused delete still removed the row; refetchFlow could revert a write that landed while it was out; writeBlock merged into a stale flow
• Duplicate from the row menu opened the original's review (whose primary is the send); "Use in a campaign" never set the template
• n / palette could create a draft on a bot with no number; [ ] walked hidden rows
• copy-code verdicts landed unattached; a third-level segment was dropped on re-save; the audience summary lost its operators
• docs: the not-Approved verdict is the module's own, stats is not in the bundled schema, the knob rows name the right file

Gates

validate, codegen:check, check, lint, test (4040 shell tests), build green. The four write EPIPE proxy tests are the known macOS-only set. Live pass on the Panel bot: use-in-a-campaign, duplicate, repeating→once, kind switch with a filtered audience — all verified, net-zero.

🤖 Generated with Claude Code