databricks-tellr-app 0.4.2.dev14__tar.gz → 0.4.2.dev16__tar.gz

{databricks_tellr_app-0.4.2.dev14 → databricks_tellr_app-0.4.2.dev16}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: databricks-tellr-app
-Version: 0.4.2.dev14
+Version: 0.4.2.dev16
 Summary: Tellr application package for Databricks Apps
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
@@ -33,6 +33,7 @@ Requires-Dist: pandas==2.3.3
 Requires-Dist: python-pptx==0.6.23
 Requires-Dist: playwright==1.48.0
 Requires-Dist: Pillow==11.3.0
+Requires-Dist: svgpathtools>=1.6.0
 Requires-Dist: google-api-python-client==2.190.0
 Requires-Dist: google-auth-oauthlib==1.2.4
 Requires-Dist: google-auth-httplib2==0.3.0

{databricks_tellr_app-0.4.2.dev14 → databricks_tellr_app-0.4.2.dev16}/databricks_tellr_app.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: databricks-tellr-app
-Version: 0.4.2.dev14
+Version: 0.4.2.dev16
 Summary: Tellr application package for Databricks Apps
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
@@ -33,6 +33,7 @@ Requires-Dist: pandas==2.3.3
 Requires-Dist: python-pptx==0.6.23
 Requires-Dist: playwright==1.48.0
 Requires-Dist: Pillow==11.3.0
+Requires-Dist: svgpathtools>=1.6.0
 Requires-Dist: google-api-python-client==2.190.0
 Requires-Dist: google-auth-oauthlib==1.2.4
 Requires-Dist: google-auth-httplib2==0.3.0

{databricks_tellr_app-0.4.2.dev14 → databricks_tellr_app-0.4.2.dev16}/databricks_tellr_app.egg-info/requires.txt

@@ -27,6 +27,7 @@ pandas==2.3.3
 python-pptx==0.6.23
 playwright==1.48.0
 Pillow==11.3.0
+svgpathtools>=1.6.0
 google-api-python-client==2.190.0
 google-auth-oauthlib==1.2.4
 google-auth-httplib2==0.3.0

{databricks_tellr_app-0.4.2.dev14 → databricks_tellr_app-0.4.2.dev16}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "databricks-tellr-app"
-version = "0.4.2.dev14"
+version = "0.4.2.dev16"
 description = "Tellr application package for Databricks Apps"
 readme = "README.md"
 requires-python = ">=3.10"
@@ -38,6 +38,19 @@ dependencies = [
     "python-pptx==0.6.23",
     "playwright==1.48.0",
     "Pillow==11.3.0",
+    # WF-02: imported by BOTH Python emitters (html_to_pptx, html_to_google_slides)
+    # and declared in the root manifest, but never here — so on every deployment the
+    # lazy `from svgpathtools import parse_path` failed and image extraction became a
+    # silent no-op, letting base64 survive into the prompt.
+    #
+    # Ranged, matching the root declaration, rather than exact-pinned like the
+    # resolver-speedup block below. MEASURED before shipping, because a new
+    # dependency here is what has previously blown the Apps build budget: adding it
+    # takes the resolution from 196 to 198 packages in 15.7s vs 18.5s baseline — no
+    # backtracking, and FASTER, not slower. Its numpy and scipy requirements are
+    # already in the closure (via mlflow / scikit-learn / skops / pandas), so the only
+    # new distributions are svgpathtools itself and the tiny pure-Python svgwrite.
+    "svgpathtools>=1.6.0",
     "google-api-python-client==2.190.0",
     "google-auth-oauthlib==1.2.4",
     "google-auth-httplib2==0.3.0",

{databricks_tellr_app-0.4.2.dev14 → databricks_tellr_app-0.4.2.dev16}/src/api/mcp_server.py

@@ -649,7 +649,10 @@ def _render_deck_response(
         "Poll the status of a deck generation or edit job. Returns "
         "lightweight status while pending/running; when ready, returns the "
         "complete deck as structured slide data, a standalone HTML "
-        "document, and URLs into tellr's full editor and view-only surfaces."
+        "document, and URLs into tellr's full editor and view-only surfaces. "
+        "A ready turn is not always an applied change: when "
+        "metadata.clarification_needed is true the deck was left UNCHANGED and "
+        "metadata.clarification carries the question to answer before retrying."
     ),
 )
 async def get_deck_status(
@@ -795,6 +798,34 @@ async def _get_deck_status_impl(
                 for m in db_messages
             ]
 
+            # WG2-01. An edit instruction that names no slide is DELIBERATELY not
+            # applied: chat_service asks which slide instead (its RC10 twins, plus the
+            # ambiguous add-vs-replace pair) and attaches ``clarification_needed`` to
+            # the COMPLETE event, which ``job_queue.process_chat_request`` copies
+            # verbatim into ``result["metadata"]``.
+            #
+            # Such a turn completes normally, so "ready" is the honest status and
+            # ``replacement_info`` is None — which is indistinguishable from an edit
+            # that applied but replaced nothing. Rebuilding ``metadata`` from
+            # hand-picked keys DROPPED the flag, so an automated caller reading
+            # ``status`` concluded the edit had landed when the deck was untouched; the
+            # only remaining trace was a ``message_type == "clarification"`` row in the
+            # transcript. Surfaced ADDITIVELY, and ``status`` keeps its existing meaning
+            # for clients already reading it.
+            clarification_needed = bool(result_metadata.get("clarification_needed"))
+            clarification = None
+            if clarification_needed:
+                # The question itself, so a caller can act on it without re-filtering
+                # the transcript. Last one wins: it is the turn's most recent ask.
+                clarification = next(
+                    (
+                        message.get("content")
+                        for message in reversed(messages)
+                        if message.get("message_type") == "clarification"
+                    ),
+                    None,
+                )
+
             return {
                 "session_id": session_id,
                 "request_id": request_id,
@@ -808,6 +839,8 @@ async def _get_deck_status_impl(
                     or result_metadata.get("latency_seconds"),
                     "experiment_url": result.get("experiment_url"),
                     "session_title": result.get("session_title"),
+                    "clarification_needed": clarification_needed,
+                    "clarification": clarification,
                 },
             }
     except MCPToolError:
@@ -851,7 +884,11 @@ def _check_contiguous(indices: list[int]) -> None:
         "The edit is applied in-place; the session_id and deck_url stay "
         "stable across edits. Returns a request_id; the caller polls "
         "get_deck_status for completion and receives the updated deck "
-        "plus replacement_info summarizing what changed."
+        "plus replacement_info summarizing what changed. An instruction that "
+        "identifies no slide — neither via slide_indices nor by naming one in "
+        "the text — is answered with a clarifying question INSTEAD of an edit: "
+        "get_deck_status then reports metadata.clarification_needed with the "
+        "deck unchanged."
     ),
 )
 async def edit_deck(

{databricks_tellr_app-0.4.2.dev14 → databricks_tellr_app-0.4.2.dev16}/src/api/routes/export.py

@@ -1,6 +1,7 @@
 """Export endpoints for PDF and PPTX generation."""
 
 import logging
+import re
 import tempfile
 from pathlib import Path
 from typing import Optional
@@ -41,6 +42,33 @@ class ExportPPTXRequest(BaseModel):
     chart_images: Optional[list[list[ChartImage]]] = None  # Chart images per slide (client-side captured)
 
 
+# WF-03. `@import` is only valid before every other rule of ITS OWN STYLESHEET, and this
+# document used to inject `*{}` and `html,body{}` ahead of the deck's CSS in a SINGLE
+# `<style>` — so a deck whose CSS OPENS with a webfont `@import` had that rule silently
+# discarded. Measured: 0 font faces in the export against 35 in the UI, which was the whole
+# of the long-unexplained 38.25 px per-component residual on the DS-OFF deck.
+#
+# The fix is STRUCTURAL, and it is the reason there is no CSS parsing here. Two `<style>`
+# elements are two stylesheets, so the deck gets its OWN `<style>`: its leading `@import` is
+# then already first in its own sheet and valid with NO rewriting at all. Cascade order
+# between separate `<style>` elements is document order, so resets-then-deck-then-resets
+# keeps exactly the precedence the single combined sheet had.
+#
+# MEASURED IN CHROMIUM under the real SLIDE_CSP: combined -> 1 sheet, 0 font faces, import
+# dead; split -> 2 sheets, font faces registered, import applied; and with a reset and the
+# deck setting the same property at equal specificity, the DECK still wins on order.
+#
+# This REPLACES a hand-written leading-`@charset`/`@import` scanner that lived here. Three
+# rounds of review found four tokenizer defects in it (`str.isspace()` accepting NBSP, an
+# unescaped top-level `\)`, a lone `@charset` rewriting the document, an ASCII-only ident
+# guard) and then a fifth (`@layer name;` may legally precede `@import`, which it missed,
+# re-losing the font for ordinary modern CSS). Every one of those was a bug in DECIDING what
+# the deck's CSS means. Emitting the deck's CSS VERBATIM in its own sheet removes the need to
+# decide, so that entire class of defect is now unreachable rather than patched: NUL bytes,
+# top-level `<!--`, a lone CR inside a string, escaped at-keywords and `@layer` are all
+# simply the browser's business, exactly as they are in the deck's own stylesheet.
+
+
 def build_slide_html(slide: dict, slide_deck: dict) -> str:
     """Build complete HTML for a single slide.
     
@@ -56,12 +84,11 @@ def build_slide_html(slide: dict, slide_deck: dict) -> str:
     external_scripts = slide_deck.get("external_scripts", [])
     deck_css = slide_deck.get("css", "")
     deck_scripts = slide_deck.get("scripts", "")
-    
+
     # Clean up deck_scripts: remove any trailing extra IIFE closings
     # deck_scripts should be a series of (function() { ... })(); blocks
     # but sometimes there might be an extra })(); at the end
     if deck_scripts:
-        import re
         # Remove any trailing })(); that might be extra
         deck_scripts = deck_scripts.rstrip()
         # Count opening and closing IIFEs
@@ -138,6 +165,28 @@ def build_slide_html(slide: dict, slide_deck: dict) -> str:
             extra={"slide_id": slide_id, "patterns": findings},
         )
 
+    # THE FIRST RESET SHEET BELOW IS SCOPED TO `html, body`, NEVER UNIVERSAL.
+    #
+    # A universal reset — border-box plus zeroed margin and padding, on every
+    # element — used to sit there. It laid slide DESCENDANTS out border-box and
+    # margin-free in the export while every preview surface lays them out
+    # content-box with UA margins, which is the box model Claude Design ground
+    # truth uses. That cost up to 69.87 px of per-component drift here, on the
+    # same 6 live slides the standalone builder diverged on by the same mechanism.
+    #
+    # This shell needs no box model of its own: it is a fixed 1280x720 document
+    # whose padding is zeroed both above and (after deck CSS) below, so the two
+    # box models are provably identical for html/body here. The standalone
+    # MULTI-slide builder does pad `body`, so it keeps `box-sizing` on its shell —
+    # see buildStandaloneDeckDocument in frontend/src/services/slideDocument.ts.
+    #
+    # Deliberately a `#` comment and NOT part of the f-string. Anything inside the
+    # f-string is EMITTED into every exported document, where it (a) displaces
+    # useful deck CSS from the ~15,000-char budget `_truncate_html` hands the LLM,
+    # and (b) eats headroom under the 2000-char length threshold a sheet-role
+    # classifier uses to tell this reset sheet from the deck's own. Keep the
+    # rendered comment to one short line and put the reasoning here instead.
+    # Pinned by tests/unit/test_preview_box_model_parity.py.
     complete_html = f"""<!DOCTYPE html>
 <html lang="en">
 <head>
@@ -147,11 +196,7 @@ def build_slide_html(slide: dict, slide_deck: dict) -> str:
   <title>{slide_deck.get("title", "Slide")} - Slide {slide.get("slide_id", "")}</title>
 {scripts_html}
   <style>
-    * {{
-      box-sizing: border-box;
-      margin: 0;
-      padding: 0;
-    }}
+    /* Shell-scoped, never universal: slide content stays content-box, as in every preview. */
     html, body {{
       width: 1280px;
       height: 720px;
@@ -161,7 +206,9 @@ def build_slide_html(slide: dict, slide_deck: dict) -> str:
       background: #ffffff;
       font-family: 'Inter', 'Segoe UI', system-ui, -apple-system, sans-serif;
     }}
-    {deck_css}
+  </style>
+  <style>{deck_css}</style>
+  <style>
     /* After deck CSS: the fixed-frame document owns its OWN box (WM-01).
        The html/body rule above is emitted BEFORE the deck, so a deck-authored
        `body {{ padding: 48px 0; gap: 48px }}` — real generator output — used to win
@@ -178,7 +225,10 @@ def build_slide_html(slide: dict, slide_deck: dict) -> str:
        depends on `body {{ padding: 40px 20px }}`.
        No `!important`, matching the frontend rule: it wins by ORDER at equal
        specificity, so both surfaces treat a deck's `!important` body padding the
-       same way and cannot drift apart. */
+       same way and cannot drift apart. That ORDER now spans separate `<style>`
+       elements — the deck has a stylesheet of its own so its leading `@import`
+       stays valid (WF-03) — and the cascade orders separate sheets by DOCUMENT
+       ORDER, so this later sheet still wins exactly as it did inside one block. */
     html, body {{
       margin: 0;
       padding: 0;

{databricks_tellr_app-0.4.2.dev14 → databricks_tellr_app-0.4.2.dev16}/src/services/design_system_compiler.py

@@ -258,7 +258,27 @@ _STYLE_HEADER = "SLIDE VISUAL STYLE"
 # BLAST RADIUS, measured: only systems whose largest sub-body rung equals the floor
 # compile differently. Every other ramp shape — including the one the docstring above
 # describes — is byte-identical apart from this version marker.
-COMPILER_VERSION = 19
+# v20: WE-02's imperative contrast-pairing wording is REVERTED for failing its
+# measurement gate, so persisted v19 rows hold ~500 B of prose this compiler no
+# longer emits and the bump is what makes them recompile.
+#
+# THE BUMP IS LOAD-BEARING HERE, unlike the edit it undoes. When WE-02 shipped it
+# rode the v19 bump on the reasoning that "v19 is unreleased, so v18 rows are
+# invalidated once, not twice" — and that was true AT THE TIME. It is not true now:
+# v19 HAS been released and deployed, and rows compiled by it exist carrying the
+# imperative wording. Currency is an exact, version-stamped sentinel match at the
+# front of the artifact (:data:`_CURRENCY_SENTINEL`,
+# :func:`compiled_style_content_is_current`), so a stored v19 artifact would keep
+# reading CURRENT under a v19 compiler and would NEVER be recompiled — the revert
+# would ship while every existing row went on serving the reverted text. Changing
+# what a RELEASED version emits without moving the stamp is precisely the failure
+# mode the sentinel exists to prevent.
+#
+# BLAST RADIUS: every design system recompiles once, and the only difference is the
+# removal of the imperative pairing prose (plus the version marker itself). WL-02's
+# safe-area wording, which PASSED its gate in the same commit, is deliberately
+# untouched and stays exactly as v19 emitted it.
+COMPILER_VERSION = 20
 _COMPILER_VERSION_MARKER = f"[ds-compiler v{COMPILER_VERSION}]"
 
 # The EXACT, name-independent header prefix every compiled artifact opens with.
@@ -1599,29 +1619,54 @@ _CONTRAST_DERIVED_HEADING = (
 # extremes do not: the measured failures were mid-tone ON mid-tone, which is
 # exactly the pair a palette makes easy to reach for and a luminance table cannot
 # enumerate without listing every combination.
-# MEASUREMENT-GATED (WE-02). The pairings above were AVAILABLE AND IGNORED, not
-# missing: on the measured deck all three offending backgrounds were enumerated WITH
-# a prescribed AA-passing text color, and the model chose otherwise every time —
-# coral #FF5F46 on oat #F9F7F4 at 2.8146:1 where #000000 (19.64:1) was named, lava
-# #FF3621 on white where #000000 (21:1) was named, and for the CTA it picked the
-# EXACT INVERSE of what the artifact stated for that background. 24 of 108 rated
-# node-cells fell below 4.5:1; 0 were primary text, and every offender was a brand
-# ACCENT used as text, all of them in the model-authored CSS tail. No color was
-# invented — the palette was misused, not exceeded.
-#
-# So the pairing is stated as a REQUIREMENT for the TEXT role specifically. The
-# colors themselves are NOT rewritten: silently correcting a brand's palette would
-# defeat the purpose of a design system.
-_CONTRAST_TEXT_ROLE_IMPERATIVE_LINE = (
-    "- Those pairings are REQUIREMENTS for TEXT, not suggestions: whenever you use a "
-    "background listed above, its text color is the one named for it. Do NOT use a "
-    "brand accent (or any other palette color) as the text color on a listed "
-    "background because it looks on-brand — accents belong in fills, rules, icons "
-    "and imagery unless that exact pair meets the ratio above. This applies to small "
-    "label text (eyebrows, kickers, captions, tags) and to link and button labels "
-    "just as much as to body copy."
-)
-
+# WE-02 ONCE ADDED MORE HERE, AND IT WAS REVERTED FOR FAILING ITS MEASUREMENT GATE.
+# Recorded so the same lever is not reached for again without new evidence.
+#
+# The diagnosis was and remains correct: the pairings are AVAILABLE AND IGNORED, not
+# missing. On the deck that prompted it, all three offending backgrounds were
+# enumerated WITH a prescribed AA-passing text color and the model chose otherwise
+# every time — coral #FF5F46 on oat #F9F7F4 at 2.8146:1 where #000000 (19.64:1) was
+# named, and for the CTA the EXACT INVERSE of what the artifact stated. No color was
+# invented; the palette was misused.
+#
+# The attempted fix restated the pairing as a REQUIREMENT for the TEXT role, plus a
+# line forbidding a brand accent as text on a listed background and naming small
+# label text and button labels explicitly. It was shipped conditional on measurement.
+#
+# THE MEASUREMENT (re-run at n=4: 13 slides, 660 rated node-cells, painted-pixel
+# backdrops, Python recompute agreeing 660/660) found it INERT. Against the
+# byte-identical dev13 prompt the failures reproduce EXACTLY — 12 sub-threshold
+# node-cells on BOTH builds, the same 3 nodes, the same two pairs, the same ratios
+# (coral eyebrow 2.8146 twice, CTA #FFFFFF on lava #FF3621 3.6197). Across n=4:
+# 208/660 below 4.5, 156/660 below their own threshold, the eyebrow pair failing on
+# 4/4 decks, the CTA inverse on 2/2 decks that produced a button, and the prescribed
+# ink ignored on 552/624 enumerated-background cells. ~500 B of prompt, no behaviour
+# change. Inert, not harmful — and inert prose is still cost, so it goes.
+#
+# WHAT THIS MEANS FOR THE NEXT ATTEMPT: restating the POSITIVE pairing more forcefully
+# is not the lever. Two candidates remain, in the order the re-run ranks them, and
+# NEITHER is implemented here:
+#
+#  1. WE-03, the reverse-direction NEGATIVE list — "do not use <color> as text" —
+#     which is still untested. The reason it is the cheaper next step is structural:
+#     every offending color is one the artifact never names as unsafe AS TEXT. The
+#     `unusable` line below is legitimately EMPTY for this brand (its extremes are
+#     pure #FFFFFF/#000000, so every resolvable color token reaches AA with one of
+#     them), so the one line that would have warned never fires. A negative list
+#     keyed on the TEXT role, not on unusability against both extremes, is a
+#     different statement from anything the artifact makes today.
+#  2. POST-GENERATION VALIDATION of rendered text/background pairs, which is the only
+#     remedy that does not depend on the model complying with prose at all.
+#
+# The re-run also found 12 unique PRIMARY-text nodes below 4.5 on prompts never
+# previously exercised (subtitle #618794 on navy 3.4961, body grey #8A8580 on white
+# 3.6538) — new coverage rather than proven degradation, but it shows the model can
+# still choose sub-threshold PRIMARY text, which is what makes (2) eventually
+# necessary. Both recorded as known-open.
+#
+# The declarative pairing itself is v18 content that closed the contrast vacuum and
+# is untouched by the revert. Colors are never rewritten either way: silently
+# correcting a brand's palette would defeat the purpose of a design system.
 _CONTRAST_PAIR_RULE_LINE = (
     "- For any other pair of these colors, work out the ratio before you use it: "
     f"two mid-tone brand colors together almost never reach {_CONTRAST_AA_NORMAL:g}:1. "
@@ -1821,16 +1866,14 @@ def _contrast_section(grouped: dict[str, list[tuple[str, str]]]) -> str:
     ]
     if on_lightest:
         lines.append(
-            f"- On these backgrounds the text color MUST be {lightest}: "
+            f"- Use {lightest} as the text color on these backgrounds: "
             f"{', '.join(on_lightest)}"
         )
     if on_darkest:
         lines.append(
-            f"- On these backgrounds the text color MUST be {darkest}: "
+            f"- Use {darkest} as the text color on these backgrounds: "
             f"{', '.join(on_darkest)}"
         )
-    if on_lightest or on_darkest:
-        lines.append(_CONTRAST_TEXT_ROLE_IMPERATIVE_LINE)
     if unusable:
         lines.append(
             "- Never set normal text on these colors — they reach "