@elitedcs/ghl-mcp 3.89.0-beta.2 → 3.89.0-beta.4

package/skills/blueprint/references/build-plan-schema.md

@@ -51,7 +51,7 @@ The Build Plan is the reviewable, editable output of the skill. **Refs only, no
 - **handoffs:** `{ref, owner, title, trigger?, instruction, produces?, successCheck, blocks?}` — `owner` ∈ **OPERATOR-UI/OPERATOR-EXT/TEAM** (legacy JERRY-UI/JERRY-EXT/SASHA accepted + normalized, never emitted). To gate a workflow DRAFT, `blocks` must list its `workflow.*` ref (a bare `sms.*` wildcard gates only the asset surface).
 
 ## Logical workflow actions (executor expands to native)
-`add_contact_tag {tagRef}`, `remove_contact_tag {tagRef}`, `send_email {emailRef | templateRef}`, `send_sms {smsRef | templateRef}` (one of the two is required — `templateRef` points at `templates.*`, v2), `wait {value, unit}`, `wait_appointment {value, unit}` (integer; appointment-triggered workflows only), `internal_notification {userRef, title, body}` (`userRef` = a `user.*` ref or `user.__pending__`; the 0.1 `to` literal is still accepted with a warning), `task_notification {title, body?, dueDate?, userRef?}` (`assignedTo` literal still accepted with a warning; no assignee = warning), `assign_user {userRef}` *(v2 — GHL "Assign to user"; the contact's owner becomes that person)*, `update_contact_field {fieldRef, value}`, `add_notes {body}`, `create_opportunity {pipelineRef, stageRef, name?, value?}`, `update_opportunity {pipelineRef, stageRef, value?}` (forces allowBackward), `add_to_workflow {workflowRef}`, `remove_from_workflow {workflowRef}`, `goal_event {goalCondition, action?}`, and `find_opportunity {pipelineRef, found:[...], notFound:[...]}` — the only branching action; it MUST be the LAST action (branches do not rejoin). All pointers are refs. The executor owns the failure-prone native shapes; the plan never contains them.
+`add_contact_tag {tagRef}`, `remove_contact_tag {tagRef}`, `send_email {emailRef | templateRef}`, `send_sms {smsRef | templateRef}` (one of the two is required — `templateRef` points at `templates.*`, v2), `wait {value, unit}`, `wait_appointment {value, unit}` (integer; appointment-triggered workflows only), `internal_notification {userRef, title, body}` (`userRef` = a `user.*` ref or `user.__pending__`; the 0.1 `to` literal is still accepted with a warning), `task_notification {title, body?, dueDate?, userRef?}` (`assignedTo` literal still accepted with a warning; no assignee = warning), `assign_user {userRef}` *(v2 — GHL "Assign to user"; the contact's owner becomes that person)*, `update_contact_field {fieldRef, value}`, `add_notes {body}`, `enable_dnd {}` *(turns on Do Not Disturb for outbound messages, e.g. after a STOP reply; emitted as GHL's `dnd_contact` step)*, `create_opportunity {pipelineRef, stageRef, name?, value?}`, `update_opportunity {pipelineRef, stageRef, value?}` (forces allowBackward), `add_to_workflow {workflowRef}`, `remove_from_workflow {workflowRef}`, `goal_event {goalCondition, action?}`, and `find_opportunity {pipelineRef, found:[...], notFound:[...]}` — the only branching action; it MUST be the LAST action (branches do not rejoin). All pointers are refs. The executor owns the failure-prone native shapes; the plan never contains them.
 
 ## v2 rules the validator enforces (2026-08-26) — read before writing workflows
 Two are **hard errors on purpose** even for plans that use no new field. Both come from a real build the owner inspected: the nurture went quiet after a week and the speed-to-lead ended with a text and dropped the lead. Error strings start with the code.

package/skills/blueprint/references/preset-format.md

@@ -56,6 +56,8 @@ Library (v2.0.0): `med-spa` (default), `generic-client` (agency / other fallback
 
 `selectors.businessTypes` are the `business_type` answers that route to this preset. `selectors.aliases` are the `brief.preset` enum values that resolve to this preset (e.g. `preset: "med_spa"` resolves to the med-spa preset, which claims that alias). Exactly one preset across the library has `"default": true` — **`med_spa`**. `generic` is `default:false`, the neutral fallback for agency / other / unmatched types.
 
+`conventions` values are strings the model reads, with one switch the composer itself reads: `"quotesNeverAutomatic": true` means a client quote never goes into a message automatically, even with written consent. The claims guard then treats a consented testimonial as not allowed and looks for the words of every intake quote, so a copy override that quotes one (even a size result with "results vary") falls back to the preset's own wording. The wellness clinic preset sets it.
+
 ## The `skeleton` object
 
 `skeleton` mirrors the §5 build-plan top level — `users`, `pipelines`, `customFields`, `tags`, `customValues`, `calendars`, `forms`, `funnels`, `templates`, `emails`, `sms`, `workflows`, `handoffs`. Keys starting with `_` are author notes and are dropped at fill time. Every item is a §5 object, optionally annotated with:
@@ -158,6 +160,8 @@ Presets must use GHL-correct enum values so the emitted plan passes ghl-command-
 - handoff `owner` ∈ **`OPERATOR-UI`**, **`OPERATOR-EXT`**, **`TEAM`** (customer-facing, generic). The MCP still accepts + normalizes the legacy `JERRY-UI`/`JERRY-EXT`/`SASHA`, but presets EMIT the canonical labels.
 - `handoff.a2p.blocks` MUST list every SMS-bearing **workflow** ref (the validator checks it) — only an explicit `workflow.*` ref keeps a workflow DRAFT.
 - Max 40 actions per workflow (a `find_opportunity` counts head + 2 transitions + children); `find_opportunity` must be the last action.
+- **Texting rules (2026-09-29, proven on live clinic accounts; enforced for every preset by `src/intake-to-build/preset-message-rules.test.ts`).** SMS bodies use plain GSM-7 characters only: no long or short dashes, curly quotes, one-character ellipsis or emoji (one long dash took a live account from 22 billed parts per contact to 47). The first text of every nurture, win-back or reactivation workflow ends with `Reply STOP to opt out.` (or `Reply N to opt out.` where a reply of N is handled) on a line of its own. A keyword trigger carries `"keyword"` plus `"keywordMatch": "exact"`; never `"contains"` (a contains-STOP opted out a patient who asked "when should I stop eating").
+- **No claims in preset copy.** No message may make an outcome or health claim (pain, sleep, inflammation relief, cures, weight or inch loss, guaranteed results, numeric results) or quote a named result ("Sarah lost 3 inches"). The composer fails a preset whose own words do (`src/intake-to-build/claims-guard.ts` holds the word list). Client proof comes only from the brief, with consent: `proof.testimonials[]` with `consent: true`, or `offer.claims[]` that say "results vary". Consent covers using the quote and name, not a claim inside it: a size or weight result goes out only with "results vary" in the same message, and a health claim never does.
 
 > The schema §5 contract and the live Zod schema (`src/intake-to-build/plan.ts`) are the sources of truth. If the doc and the Zod schema ever disagree, the Zod schema wins.