feat(social-ship-posts): wire the account story canon (NARRATIVE.md) into the loop

RESOLVER.md

@@ -82,6 +82,7 @@
 | "recoup-internal" + account-health / account-status snapshot for a Recoup account | `recoup-internal-account-health-report` |
 | "recoup-internal" + audit scheduled-task email health (what tasks emailed, how many empties the guard blocked) over a window | `recoup-internal-task-email-audit` |
 | "recoup-internal" + draft/ship/measure LinkedIn or X posts | `recoup-internal-social-ship-posts` |
+| "recoup-internal" + FaceTime-call style AI video ad (Grok Imagine 1.5 native-speech clips) | `recoup-internal-video-grok-1.5-imagine-facetime` |
 
 ---
 

resolver-eval.jsonl

@@ -59,6 +59,7 @@
 {"intent": "recoup-internal work the valuation funnel and advance this Attio lead", "expected": "recoup-internal-funnel-valuation-pipeline"}
 {"intent": "recoup-internal account health snapshot for this account", "expected": "recoup-internal-account-health-report"}
 {"intent": "recoup-internal draft and ship a launch post for LinkedIn and X", "expected": "recoup-internal-social-ship-posts"}
+{"intent": "recoup-internal make a FaceTime-call style AI video ad with Grok Imagine", "expected": "recoup-internal-video-grok-1.5-imagine-facetime", "not": ["recoup-content-make-video"]}
 
 # --- gating invariant: WITHOUT the recoup-internal keyword, internal skills must NOT fire ---
 # Only the internal skills with a plausible customer-facing collision get a negative; the dev/eval

scripts/check_resolvable.py

@@ -25,7 +25,8 @@
 DEFAULT_PLUGIN = REPO_ROOT
 
 # A skill reference in RESOLVER.md is a backtick-wrapped slug: `recoup-foo-bar`.
-SKILL_TOKEN = re.compile(r"`(recoup-[a-z0-9-]+)`")
+# Dots are allowed mid-slug (e.g. `recoup-internal-video-grok-1.5-imagine-facetime`).
+SKILL_TOKEN = re.compile(r"`(recoup-[a-z0-9-]+(?:\.[a-z0-9-]+)*)`")
 
 
 def skill_dirs(plugin: Path) -> set[str]:

scripts/run_resolver_eval.py

@@ -48,7 +48,8 @@
 # Co-located with the plugin root (NOT under tests/, which is gitignored) so the
 # fixtures ship and CI can run them on a fresh checkout.
 FIXTURES = PLUGIN / "resolver-eval.jsonl"
-SKILL_TOKEN = re.compile(r"`(recoup-[a-z0-9-]+)`")
+# Dots are allowed mid-slug (e.g. `recoup-internal-video-grok-1.5-imagine-facetime`).
+SKILL_TOKEN = re.compile(r"`(recoup-[a-z0-9-]+(?:\.[a-z0-9-]+)*)`")
 
 
 def skill_dirs() -> set[str]:

skills/recoup-internal-social-ship-posts/SKILL.md

@@ -12,7 +12,7 @@ Authenticate to the Recoup API with an `x-api-key` (`RECOUP_API_KEY`) or `Author
 ## The loop
 
 1. **Learn** — pull how recent posts performed, so copy is grounded in evidence, not vibes.
-2. **Draft** — write topic-first copy that fits the platform and the goal.
+2. **Draft** — write topic-first copy that fits the platform, the goal, and the account's story canon.
 3. **Decide the CTA** — comment-gate for leads, or direct link for reach.
 4. **Publish** — via the connector where supported, or hand off for manual posting.
 5. **Log & re-measure** — record what went out; re-pull performance ~48h later.
@@ -65,6 +65,8 @@ Read the top performer against the flatliners and name the differences (framing,
 
 ## Step 2 — Draft
 
+**Check the story canon first.** If the account workspace has a `NARRATIVE.md` (story bible: premise, recurring cast, storylines), read it and name **which arc and which character this post serves** before writing a word. A post that serves no arc is an ad — reframe it into the story or drop it. Serialized rules live in the canon itself (recurring catchphrases, who may be fictionalized vs. who gets only verified numbers, cliffhanger beats); follow them over generic copy instincts. If the workspace has no `NARRATIVE.md`, note that to the account owner — the account is posting without a story.
+
 Write to the brief from step 1 plus the principles above. Structure that travels well:
 
 ```
@@ -113,7 +115,7 @@ Native video works, but the bytes can't go inline: `media` won't fetch a URL, an
 
 ## Step 5 — Log and re-measure
 
-Record each post immediately: account, platform, the post id/URN, UTC time, the CTA used, and the copy. Then **re-pull performance ~48h later**: one bulk scrape (step 1) covers all four platforms; add `LINKEDIN_LIST_REACTIONS` only when you need reactor identities, and read gated posts' comment bodies manually. Compare week-over-week — report deltas, not all-time totals. Feed the result back into step 1 for the next post.
+Record each post immediately: account, platform, the post id/URN, UTC time, the CTA used, the copy, and — when the account has a `NARRATIVE.md` — **the arc and character the post served** (continuity is what makes the next episode land; an unlogged arc breaks the serial). Then **re-pull performance ~48h later**: one bulk scrape (step 1) covers all four platforms; add `LINKEDIN_LIST_REACTIONS` only when you need reactor identities, and read gated posts' comment bodies manually. Compare week-over-week — report deltas, not all-time totals. Feed the result back into step 1 for the next post.
 
 ## Producing a demo video?