From 7e0157e45b681b0a31309101bff2fd1f590b7482 Mon Sep 17 00:00:00 2001
From: grokbot_ty <grokbot_ty@local>
Date: Sat, 26 Sep 2026 10:45:28 -0700
Subject: [PATCH] fix(call2): plain text + woodworker tone

Review short/long must not show HTML tags; seed now requires plain text
only (newlines/- bullets) and practical woodworker copy. parse_call2_response
strips residual HTML as a safety net. Seed activation unchanged.
---
 backend/inference.py            | 77 ++++++++++++++++++++++++---------
 backend/tests/test_inference.py | 57 +++++++++++++++++-------
 docs/contracts/PROMPTS.md       | 34 ++++++++++-----
 3 files changed, 120 insertions(+), 48 deletions(-)

diff --git a/backend/inference.py b/backend/inference.py
index 719ed99..2ba4f7b 100644
--- a/backend/inference.py
+++ b/backend/inference.py
@@ -7,6 +7,8 @@ overnight-whole-app: Call 1 POST is hard-disabled (410 call1_disabled). Call 2 s
 
 import base64
 import hashlib
+import html
+import re
 import io
 import json
 import time
@@ -106,9 +108,17 @@ CALL1_SEED = (
     '"feature_tag_ids": [<id>,...], "feature_tag_confidence": <0-1>}'
 )
 CALL2_SEED = (
-    "Write WooCommerce listing copy for this slab from the photos and listing facts below.\n"
+    "Write listing copy for this slab for woodworkers choosing stock for a project. "
+    "Audience: makers matching wood to a tabletop, bar, bench, or similar build — practical, "
+    "material/figure/spec-forward. No lifestyle or marketing hype "
+    "(avoid empty phrases like elevate, stunning, dramatic visual contrast).\n"
     "You write BOTH the short description and the long description.\n"
     "\n"
+    "PLAIN TEXT ONLY (short AND long):\n"
+    "- Output plain text only. No HTML tags (<p>, <ul>, <li>, <strong>, <br>, etc.). "
+    "No markdown bold/headers that look like markup (**bold**, # headers).\n"
+    "- Structure the long description with newlines and \"- \" bullets when helpful.\n"
+    "\n"
     "Rules that always apply:\n"
     "- Dimensions in the DIMENSION FACT BLOCK are exact: use those strings; invent no numbers.\n"
     "- Use taxonomy names from DECLARED ATTRIBUTES; do not invent species, edge, figure, grade, "
@@ -116,24 +126,26 @@ CALL2_SEED = (
     "- Use the Brand voice and Geography sections when provided. If Brand voice is empty, stay "
     "neutral. Never invent return, warranty, or shipping policies.\n"
     "- The product name may already appear in DECLARED ATTRIBUTES — do not invent a new title.\n"
+    "- Lead with usable traits (species, figure, size, edge, moisture, grade) and what the piece "
+    "suits (tabletop, bar, etc.) without ad-copy fluff.\n"
     "\n"
-    "SHORT DESCRIPTION (~50 words — most important on-page copy; plain text, no HTML):\n"
-    "1. Lead with the outcome or problem solved (benefit), not material-first.\n"
-    "2. Add the key differentiator or most important spec.\n"
-    "3. Close with a trust signal or reason to act.\n"
-    "Do NOT lead with a species + material list.\n"
+    "SHORT DESCRIPTION (~50 words — most important on-page copy; plain text only):\n"
+    "1. Lead with a practical outcome for a maker (what they can build / how the stock helps), "
+    "not lifestyle hype and not a bare species list.\n"
+    "2. Add the key differentiator or most important spec (figure, edge, size, moisture, grade).\n"
+    "3. Close with a trust signal or concrete reason to choose this piece.\n"
     "\n"
-    "LONG DESCRIPTION (HTML for the WooCommerce visual editor — short paragraphs, "
-    "<ul>/<li> for bullets, <strong> for benefit leads; no markdown). Four parts, in order:\n"
-    "1. Hook — one sentence restating the core benefit.\n"
-    "2. Benefits — 3–5 bullets; each benefit-first, then the supporting spec.\n"
+    "LONG DESCRIPTION (plain text only — newlines and \"- \" bullets; no HTML, no markdown "
+    "markup). Four parts, in order:\n"
+    "1. Hook — one sentence on what the piece suits and its standout usable trait.\n"
+    "2. Benefits — 3–5 bullets; each maker-useful trait first, then the supporting spec.\n"
     "3. Full specs — clean list from DECLARED ATTRIBUTES + DIMENSION FACT BLOCK "
     "(dims, species, moisture, edge, figure, grade, etc.).\n"
     "4. Trust close — return/warranty/brand; use Brand voice when present; otherwise a short "
     "neutral close with no invented policies.\n"
     "\n"
     "Return ONE JSON object only (no prose outside JSON) with exactly these keys:\n"
-    '{"short_description": "<plain text ~50 words>", "description": "<HTML long description>"}'
+    '{"short_description": "<plain text ~50 words>", "description": "<plain text long description>"}'
 )
 _SEEDS = {"call1": CALL1_SEED, "call2": CALL2_SEED}
 _THRESHOLD = 0.7
@@ -681,12 +693,37 @@ async def call2_prompt(slab_id: str) -> dict:
     }
 
 
+def _html_to_plain(value: str) -> str:
+    """Defensive: strip HTML/markup so Review fields never show raw tags.
+
+    Converts common block tags to newlines and <li> to "- " bullets, then removes
+    remaining tags and unescapes entities. Already-plain text is returned cleaned.
+    """
+    if not value:
+        return ""
+    t = value
+    # Normalize common break / list / paragraph boundaries before dropping tags.
+    t = re.sub(r"(?is)<br\s*/?>", "\n", t)
+    t = re.sub(r"(?is)</\s*(p|div|h[1-6]|tr|table)\s*>", "\n", t)
+    t = re.sub(r"(?is)<\s*li[^>]*>", "- ", t)
+    t = re.sub(r"(?is)</\s*(ul|ol)\s*>", "\n", t)
+    t = re.sub(r"(?is)<[^>]+>", "", t)
+    t = html.unescape(t)
+    # Drop leftover markdown-ish bold/headers that look like markup in Review.
+    t = re.sub(r"(?m)^#{1,6}\s+", "", t)
+    t = re.sub(r"\*\*([^*]+)\*\*", r"\1", t)
+    t = re.sub(r"[ \t]+\n", "\n", t)
+    t = re.sub(r"\n{3,}", "\n\n", t)
+    return t.strip()
+
+
 def parse_call2_response(raw: str) -> tuple[str | None, str]:
     """Split Call 2 model output into (short_description, description).
 
     Preferred shape is a JSON object with short_description + description.
-    If JSON is missing or incomplete, treat the whole body as description HTML
-    and leave short_description None (caller keeps any prior short).
+    Both fields are forced to plain text (HTML stripped) before return.
+    If JSON is missing or incomplete, treat the whole body as description
+    (stripped) and leave short_description None (caller keeps any prior short).
     """
     text = (raw or "").strip()
     if not text:
@@ -710,21 +747,21 @@ def parse_call2_response(raw: str) -> tuple[str | None, str]:
             try:
                 obj = json.loads(body[start : end + 1])
             except (ValueError, TypeError):
-                return None, text
+                return None, _html_to_plain(text)
         else:
-            return None, text
+            return None, _html_to_plain(text)
     if not isinstance(obj, dict):
-        return None, text
+        return None, _html_to_plain(text)
     short = obj.get("short_description")
     desc = obj.get("description")
     if isinstance(short, str):
-        short = short.strip() or None
+        short = _html_to_plain(short) or None
     else:
         short = None
     if isinstance(desc, str) and desc.strip():
-        return short, desc.strip()
-    # JSON without a usable description — fall back to raw.
-    return short, text
+        return short, _html_to_plain(desc)
+    # JSON without a usable description — fall back to raw (stripped).
+    return short, _html_to_plain(text)
 
 
 @router.post("/slabs/{slab_id}/inference/call2")
diff --git a/backend/tests/test_inference.py b/backend/tests/test_inference.py
index e471a6b..07c6f1b 100644
--- a/backend/tests/test_inference.py
+++ b/backend/tests/test_inference.py
@@ -78,7 +78,7 @@ def test_call1_seed_is_locked_expert_prompt(env):
 
 
 def test_call2_seed_has_short_long_structure(env):
-    """Call 2 seed carries short+long structure and JSON dual output (not brand voice)."""
+    """Call 2 seed: plain-text short+long, woodworker tone, JSON dual output (not brand voice)."""
     client, _conn = env
     from backend.inference import CALL2_SEED
 
@@ -86,13 +86,21 @@ def test_call2_seed_has_short_long_structure(env):
     assert body["active_content"] == CALL2_SEED
     assert "SHORT DESCRIPTION" in CALL2_SEED
     assert "LONG DESCRIPTION" in CALL2_SEED
-    assert "Lead with the outcome" in CALL2_SEED
+    assert "PLAIN TEXT ONLY" in CALL2_SEED
+    assert "plain text only" in CALL2_SEED.lower()
+    assert "No HTML tags" in CALL2_SEED
+    assert "woodworkers choosing stock" in CALL2_SEED
+    assert "elevate" in CALL2_SEED  # banned-phrase guidance
+    assert "stunning" in CALL2_SEED
+    assert "Lead with a practical outcome" in CALL2_SEED
     assert "Hook" in CALL2_SEED
     assert "Brand voice and Geography sections when provided" in CALL2_SEED
     assert "Whidbey" not in CALL2_SEED
     assert "Pacific Northwest" not in CALL2_SEED
     assert '"short_description"' in CALL2_SEED
     assert '"description"' in CALL2_SEED
+    assert "<HTML long description>" not in CALL2_SEED
+    assert "plain text long description" in CALL2_SEED
 
 
 def test_taxonomy_lists_call1_closed_shape():
@@ -196,7 +204,7 @@ def test_call1_hard_disabled_410(env, monkeypatch):
 
 
 def test_call2_stores_short_and_description(env, monkeypatch):
-    """Call 2 persists both short_description and description from JSON output."""
+    """Call 2 persists both fields as plain text (HTML from model is stripped)."""
     client, conn = env
     conn.execute("INSERT OR REPLACE INTO settings(key,value,value_enc) VALUES ('content_llm_enabled','true',NULL)")
     conn.execute(
@@ -210,25 +218,29 @@ def test_call2_stores_short_and_description(env, monkeypatch):
     import json as _json
 
     payload = _json.dumps({
-        "short_description": "Turn a blank into a showpiece table — kiln-dried walnut with live edge character. Ready to finish.",
-        "description": "<p><strong>Showpiece ready.</strong></p><ul><li><strong>Stable:</strong> kiln dried</li></ul>",
+        "short_description": "Kiln-dried walnut live-edge stock for a dining table — clear midspan, usable width. Ready to finish.",
+        "description": "<p><strong>Dining table stock.</strong></p><ul><li><strong>Stable:</strong> kiln dried</li></ul>",
     })
     monkeypatch.setattr(InferenceClient, "chat", lambda self, messages: payload)
     r = client.post("/api/v1/slabs/s1/inference/call2")
     assert r.status_code == 200, r.text
     body = r.json()
-    assert "Showpiece ready" in body["description"] or "kiln" in body["description"]
-    assert body["short_description"].startswith("Turn a blank")
+    assert "Dining table stock" in body["description"]
+    assert "kiln dried" in body["description"]
+    assert "<p>" not in body["description"]
+    assert "<strong>" not in body["description"]
+    assert body["short_description"].startswith("Kiln-dried walnut")
     row = conn.execute(
         "SELECT short_description, description, content_source FROM slabs WHERE id='s1'"
     ).fetchone()
-    assert row["short_description"].startswith("Turn a blank")
-    assert "<p>" in row["description"]
+    assert row["short_description"].startswith("Kiln-dried walnut")
+    assert "<p>" not in row["description"]
+    assert "- Stable:" in row["description"] or "Stable:" in row["description"]
     assert row["content_source"] == "llm"
 
 
 def test_call2_html_fallback_keeps_prior_short(env, monkeypatch):
-    """Non-JSON Call 2 body becomes description only; prior short is left alone."""
+    """Non-JSON Call 2 body becomes plain-text description; prior short is left alone."""
     client, conn = env
     conn.execute("INSERT OR REPLACE INTO settings(key,value,value_enc) VALUES ('content_llm_enabled','true',NULL)")
     conn.execute(
@@ -243,32 +255,45 @@ def test_call2_html_fallback_keeps_prior_short(env, monkeypatch):
     monkeypatch.setattr(InferenceClient, "chat", lambda self, messages: "<p>A fine walnut slab.</p>")
     r = client.post("/api/v1/slabs/s1/inference/call2")
     assert r.status_code == 200, r.text
-    assert r.json()["description"] == "<p>A fine walnut slab.</p>"
+    assert r.json()["description"] == "A fine walnut slab."
     assert "short_description" not in r.json()
     row = conn.execute(
         "SELECT short_description, description, content_source FROM slabs WHERE id='s1'"
     ).fetchone()
     assert row["short_description"] == "keep me"
-    assert row["description"] == "<p>A fine walnut slab.</p>"
+    assert row["description"] == "A fine walnut slab."
     assert row["content_source"] == "llm"
 
 
 def test_parse_call2_response_json_and_fenced():
-    from backend.inference import parse_call2_response
+    from backend.inference import parse_call2_response, _html_to_plain
     import json as _json
 
     short, desc = parse_call2_response(
         _json.dumps({"short_description": " Benefit first. ", "description": " <p>Long</p> "})
     )
     assert short == "Benefit first."
-    assert desc == "<p>Long</p>"
+    assert desc == "Long"
     fenced = "```json\n" + _json.dumps({"short_description": "S", "description": "<p>D</p>"}) + "\n```"
     short2, desc2 = parse_call2_response(fenced)
     assert short2 == "S"
-    assert desc2 == "<p>D</p>"
+    assert desc2 == "D"
     short3, desc3 = parse_call2_response("<p>only long</p>")
     assert short3 is None
-    assert desc3 == "<p>only long</p>"
+    assert desc3 == "only long"
+
+    # Defensive strip: lists become "- " bullets; entities unescaped; no tags left.
+    plain = _html_to_plain(
+        "<p>Hook.</p><ul><li><strong>Stable:</strong> kiln dried</li>"
+        "<li>Live edge</li></ul><p>Trust &amp; ship.</p>"
+    )
+    assert "<" not in plain
+    assert "Hook." in plain
+    assert "- Stable: kiln dried" in plain
+    assert "- Live edge" in plain
+    assert "Trust & ship." in plain
+    # Already-plain text is unchanged aside from trim.
+    assert _html_to_plain("  Kiln-dried walnut for a bar top.  ") == "Kiln-dried walnut for a bar top."
 
 
 def test_call1_disabled_even_when_inference_off(env):
diff --git a/docs/contracts/PROMPTS.md b/docs/contracts/PROMPTS.md
index 2e5ec59..6abf4bb 100644
--- a/docs/contracts/PROMPTS.md
+++ b/docs/contracts/PROMPTS.md
@@ -168,9 +168,17 @@ Runtime appends `LISTS:` (JSON of the five closed lists) plus SKU/length/thickne
 
 **`call2` seed.** Shipped text in `backend/inference.py` (`CALL2_SEED`), mirrored here for review.
 
-Write WooCommerce listing copy for this slab from the photos and listing facts below.
+Write listing copy for this slab for woodworkers choosing stock for a project. Audience:
+makers matching wood to a tabletop, bar, bench, or similar build — practical,
+material/figure/spec-forward. No lifestyle or marketing hype (avoid empty phrases like
+elevate, stunning, dramatic visual contrast).
 You write BOTH the short description and the long description.
 
+PLAIN TEXT ONLY (short AND long):
+- Output plain text only. No HTML tags (`<p>`, `<ul>`, `<li>`, `<strong>`, `<br>`, etc.).
+  No markdown bold/headers that look like markup (`**bold**`, `#` headers).
+- Structure the long description with newlines and `- ` bullets when helpful.
+
 Rules that always apply:
 - Dimensions in the DIMENSION FACT BLOCK are exact: use those strings; invent no numbers.
 - Use taxonomy names from DECLARED ATTRIBUTES; do not invent species, edge, figure, grade,
@@ -178,28 +186,30 @@ Rules that always apply:
 - Use the Brand voice and Geography sections when provided. If Brand voice is empty, stay
   neutral. Never invent return, warranty, or shipping policies.
 - The product name may already appear in DECLARED ATTRIBUTES — do not invent a new title.
+- Lead with usable traits (species, figure, size, edge, moisture, grade) and what the piece
+  suits (tabletop, bar, etc.) without ad-copy fluff.
 
-SHORT DESCRIPTION (~50 words — most important on-page copy; plain text, no HTML):
-1. Lead with the outcome or problem solved (benefit), not material-first.
-2. Add the key differentiator or most important spec.
-3. Close with a trust signal or reason to act.
-Do NOT lead with a species + material list.
+SHORT DESCRIPTION (~50 words — most important on-page copy; plain text only):
+1. Lead with a practical outcome for a maker (what they can build / how the stock helps),
+   not lifestyle hype and not a bare species list.
+2. Add the key differentiator or most important spec (figure, edge, size, moisture, grade).
+3. Close with a trust signal or concrete reason to choose this piece.
 
-LONG DESCRIPTION (HTML for the WooCommerce visual editor — short paragraphs, `<ul>`/`<li>` for
-bullets, `<strong>` for benefit leads; no markdown). Four parts, in order:
-1. Hook — one sentence restating the core benefit.
-2. Benefits — 3–5 bullets; each benefit-first, then the supporting spec.
+LONG DESCRIPTION (plain text only — newlines and `- ` bullets; no HTML, no markdown markup).
+Four parts, in order:
+1. Hook — one sentence on what the piece suits and its standout usable trait.
+2. Benefits — 3–5 bullets; each maker-useful trait first, then the supporting spec.
 3. Full specs — clean list from DECLARED ATTRIBUTES + DIMENSION FACT BLOCK (dims, species,
    moisture, edge, figure, grade, etc.).
 4. Trust close — return/warranty/brand; use Brand voice when present; otherwise a short
    neutral close with no invented policies.
 
 Return ONE JSON object only (no prose outside JSON) with exactly these keys:
-`{"short_description": "<plain text ~50 words>", "description": "<HTML long description>"}`
+`{"short_description": "<plain text ~50 words>", "description": "<plain text long description>"}`
 
 Runtime appends DECLARED ATTRIBUTES (title + taxonomy; not a prior short), DIMENSION FACT
 BLOCK, and `Brand voice:` / `Geo:` from Settings (`SETTINGS-30`). The POST handler parses the
-JSON and persists both fields.
+JSON, strips any residual HTML to plain text, and persists both fields.
 
 ## 7. Prompt management