HTML generation pipeline: validator strips image slots + diagram label suppression needs rework

casper-stevens commented

2026-05-29 06:57:06 +00:00

Member

Problem

The HTML slide generation pipeline has accumulated several interacting issues that are causing images to disappear entirely from generated slides.

Root cause

In attempting to suppress CSS-drawn diagram labels (short uppercase text placed around image slots by the LLM, e.g. "INTENT", "GUARDRAILS"), a validation pass was added that now runs on every slide containing image slots. This validation pass is an LLM rewriting the full HTML, and it is stripping the data-image-slot elements — either by misidentifying legitimate section labels near images as "diagram annotations", or by corrupting the attribute during its full-document rewrite. Since image injection happens after the validator, stripped slots = no images.

Specific issues

1. Universal validator is wrong tool for diagram labels

The validator (system_validate_slide.md) was originally scoped to SVG-only. It was broadened to fire on all slides with image slots. Running a full LLM HTML rewrite just to strip a few text labels is expensive, fragile, and unreliable — the LLM cannot reliably distinguish a diagram label from a legitimate section label positioned near an image.

2. Diagram label suppression belongs at content level or via regex

The LLM renders labels because the slide content asks for a labeled diagram. Suppressing the output after generation is fighting the model on every run. A deterministic regex pass would be more reliable: remove position:absolute elements that are siblings of a data-image-slot container, have <5 words of text, and sit within ~200px of the slot.

3. Two prompts in conflict

system_generate_slide.md says "put images in slots". system_validate_slide.md says "preserve slots, but remove labels near slots". These are hard to reconcile reliably with an LLM.

4. Late image injection is fragile

Images are injected after 3 separate AI passes (generation → validation → overflow check). Any pass that corrupts data-image-slot attributes causes silent image loss with no error surfaced to the user.

5. Animations interact with validator

Every element starts at opacity:0 due to entrance animations. The validator LLM reasons about a CSS-animated state it cannot see, potentially making wrong corrections.

Recommended fix

Revert validator trigger to SVG-only (immediate fix — restores images)
Replace diagram-label suppression with a deterministic regex pass instead of an LLM pass
Consider moving diagram guidance to the content/wizard prompts so the slide markdown describes diagrams as single visual references rather than labeled node graphs
Add explicit logging when image slots are lost between pipeline stages so regressions are visible

Branch

This regression was introduced on branch development_pptx_mockup during the animation + prompt extraction work (commit 9426df8 and subsequent commits).

## Problem The HTML slide generation pipeline has accumulated several interacting issues that are causing **images to disappear entirely** from generated slides. ## Root cause In attempting to suppress CSS-drawn diagram labels (short uppercase text placed around image slots by the LLM, e.g. "INTENT", "GUARDRAILS"), a validation pass was added that now runs on **every slide containing image slots**. This validation pass is an LLM rewriting the full HTML, and it is stripping the `data-image-slot` elements — either by misidentifying legitimate section labels near images as "diagram annotations", or by corrupting the attribute during its full-document rewrite. Since image injection happens **after** the validator, stripped slots = no images. ## Specific issues ### 1. Universal validator is wrong tool for diagram labels The validator (`system_validate_slide.md`) was originally scoped to SVG-only. It was broadened to fire on all slides with image slots. Running a full LLM HTML rewrite just to strip a few text labels is expensive, fragile, and unreliable — the LLM cannot reliably distinguish a diagram label from a legitimate section label positioned near an image. ### 2. Diagram label suppression belongs at content level or via regex The LLM renders labels because the slide *content* asks for a labeled diagram. Suppressing the output after generation is fighting the model on every run. A deterministic regex pass would be more reliable: remove `position:absolute` elements that are siblings of a `data-image-slot` container, have <5 words of text, and sit within ~200px of the slot. ### 3. Two prompts in conflict `system_generate_slide.md` says "put images in slots". `system_validate_slide.md` says "preserve slots, but remove labels near slots". These are hard to reconcile reliably with an LLM. ### 4. Late image injection is fragile Images are injected after 3 separate AI passes (generation → validation → overflow check). Any pass that corrupts `data-image-slot` attributes causes silent image loss with no error surfaced to the user. ### 5. Animations interact with validator Every element starts at `opacity:0` due to entrance animations. The validator LLM reasons about a CSS-animated state it cannot see, potentially making wrong corrections. ## Recommended fix 1. **Revert validator trigger to SVG-only** (immediate fix — restores images) 2. **Replace diagram-label suppression with a deterministic regex pass** instead of an LLM pass 3. **Consider moving diagram guidance to the content/wizard prompts** so the slide markdown describes diagrams as single visual references rather than labeled node graphs 4. **Add explicit logging when image slots are lost** between pipeline stages so regressions are visible ## Branch This regression was introduced on branch `development_pptx_mockup` during the animation + prompt extraction work (commit `9426df8` and subsequent commits).

casper-stevens commented

2026-05-29 08:29:24 +00:00

Author

Member

Implementation Spec — Issue #95

HTML Generation Pipeline: Validator Strips Image Slots + Diagram Label Suppression Rework

Objective

Restore reliable image slot preservation through the HTML generation pipeline by:

Reverting the validator trigger condition so it only fires on slides that contain <svg> elements (its original scope), removing the data-image-slot branch that causes the slot-stripping regression.
Replacing the diagram-label suppression responsibility currently burdened on the LLM validator with a deterministic Rust function that removes position:absolute sibling elements near image slots that match the short-uppercase-label heuristic.
Adding slot-count logging at each pipeline stage so regressions are immediately visible in job logs.
Updating system_validate_slide.md to remove Violation #3 (diagram label annotations), since that concern is now handled deterministically.

Requirements

The validator (validate_slide_html) must only fire when <svg> is present in the HTML, not when data-image-slot is present.
A new deterministic Rust function strip_diagram_labels must run after the validator (and after replace_fake_logos_with_slots) to remove short-text absolute-positioned elements that are siblings of data-image-slot containers.
The heuristic for strip_diagram_labels:
- Target: any <div> or <span> (or <p>) with position:absolute in its style attribute.
- Short text: inner text containing fewer than 5 words and no child elements with data-image-slot.
- Proximity: the element's top/left/right/bottom CSS values place it within 200px of a data-image-slot container's bounding box OR the element is a sibling inside the same parent container as a data-image-slot div.
- Exclusion: never remove elements that themselves have data-image-slot="true".
Slot count must be logged via info! immediately after each pipeline stage: (a) replace_fake_logos_with_slots, (b) validate_slide_html, (c) strip_diagram_labels, (d) extract_image_slot_descriptions.
system_validate_slide.md must have Violation #3 ("DIAGRAM LABEL ANNOTATIONS") removed; the remaining violations must be preserved and renumbered.
system_generate_slide.md is left unchanged.

Files to Modify

File	Change
`crates/hero_slides_lib/src/generator.rs`	(1) Change validator trigger condition; (2) Add `strip_diagram_labels` function; (3) Insert call in pipeline; (4) Add slot-count `info!` logs.
`crates/hero_slides_lib/src/prompts/system_validate_slide.md`	Remove Violation #3 "DIAGRAM LABEL ANNOTATIONS" and renumber remaining violations.

Step-by-Step Implementation Plan

Step 1 — Revert validator trigger to SVG-only

File: crates/hero_slides_lib/src/generator.rs
Dependencies: none | Parallelisable with: Steps 2, 4

Change the needs_validation condition:

// Before:
let needs_validation = html.contains("<svg") || html.contains("data-image-slot");
// After:
let needs_validation = html.contains("<svg");

Update the adjacent comment to note that image-slot-only slides are handled by strip_diagram_labels instead.

Step 2 — Add slot-count logging at each pipeline stage

File: crates/hero_slides_lib/src/generator.rs
Dependencies: none | Parallelisable with: Steps 1, 4

Insert info! log lines after each of these calls:

replace_fake_logos_with_slots
validate_slide_html
strip_diagram_labels (to be added in Step 3)

Format: "generate_slide_html {}: after <stage>, {} slot(s) present"

Step 3 — Implement `strip_diagram_labels` and integrate into pipeline

File: crates/hero_slides_lib/src/generator.rs
Dependencies: Steps 1 and 2

Add a new private function strip_diagram_labels(html: &str) -> String that:

Collects top/left estimates for all data-image-slot elements.
Scans for position:absolute elements, extracts their inner text, counts words.
Removes elements where word count < 5 AND position is within 200px of any slot.
Never removes elements that contain data-image-slot.

Uses string scanning only (no new crate dependencies). Insert the call in generate_slide_html after the validator block and before extract_image_slot_descriptions.

Step 4 — Update `system_validate_slide.md`

File: crates/hero_slides_lib/src/prompts/system_validate_slide.md
Dependencies: none | Parallelisable with: Steps 1, 2

Remove the entire Violation #3 block ("DIAGRAM LABEL ANNOTATIONS") and renumber the remaining violations so they are sequential.

Acceptance Criteria

A slide with data-image-slot but no <svg> does NOT trigger the LLM validator pass.
A slide with <svg> still triggers the validator pass.
Job logs show slot counts after each of the four pipeline stages.
Short uppercase labels placed near an image slot are removed by strip_diagram_labels.
The image slot itself (data-image-slot="true") is never removed by strip_diagram_labels.
system_validate_slide.md no longer contains the "DIAGRAM LABEL ANNOTATIONS" section.
cargo build -p hero_slides_lib succeeds with no new warnings.
Existing tests pass (cargo test -p hero_slides_lib).

Notes

Why string scanning: The codebase uses string scanning throughout (replace_fake_logos_with_slots, replace_html_tag_with_slot, etc.). No new crate dependency is needed or justified.
False-negative tolerance: The word-count guard (<5 words) is the primary safety net. False negatives (label not stripped) are acceptable; false positives (content element stripped) are not.
browser_pptx.rs changes are unrelated: The current uncommitted whitespace diff in browser_pptx.rs is not part of this issue.
Prompt override users: Because system_validate_slide.md is embedded via include_str!, editing the file is sufficient. Users with local prompt overrides will need to update their copies.

## Implementation Spec — Issue #95 ### HTML Generation Pipeline: Validator Strips Image Slots + Diagram Label Suppression Rework --- ## Objective Restore reliable image slot preservation through the HTML generation pipeline by: 1. Reverting the validator trigger condition so it only fires on slides that contain `<svg>` elements (its original scope), removing the `data-image-slot` branch that causes the slot-stripping regression. 2. Replacing the diagram-label suppression responsibility currently burdened on the LLM validator with a deterministic Rust function that removes `position:absolute` sibling elements near image slots that match the short-uppercase-label heuristic. 3. Adding slot-count logging at each pipeline stage so regressions are immediately visible in job logs. 4. Updating `system_validate_slide.md` to remove Violation #3 (diagram label annotations), since that concern is now handled deterministically. --- ## Requirements - The validator (`validate_slide_html`) must only fire when `<svg>` is present in the HTML, not when `data-image-slot` is present. - A new deterministic Rust function `strip_diagram_labels` must run after the validator (and after `replace_fake_logos_with_slots`) to remove short-text absolute-positioned elements that are siblings of `data-image-slot` containers. - The heuristic for `strip_diagram_labels`: - Target: any `<div>` or `<span>` (or `<p>`) with `position:absolute` in its style attribute. - Short text: inner text containing fewer than 5 words and no child elements with `data-image-slot`. - Proximity: the element's `top`/`left`/`right`/`bottom` CSS values place it within 200px of a `data-image-slot` container's bounding box OR the element is a sibling inside the same parent container as a `data-image-slot` div. - Exclusion: never remove elements that themselves have `data-image-slot="true"`. - Slot count must be logged via `info!` immediately after each pipeline stage: (a) `replace_fake_logos_with_slots`, (b) `validate_slide_html`, (c) `strip_diagram_labels`, (d) `extract_image_slot_descriptions`. - `system_validate_slide.md` must have Violation #3 ("DIAGRAM LABEL ANNOTATIONS") removed; the remaining violations must be preserved and renumbered. - `system_generate_slide.md` is left unchanged. --- ## Files to Modify | File | Change | |---|---| | `crates/hero_slides_lib/src/generator.rs` | (1) Change validator trigger condition; (2) Add `strip_diagram_labels` function; (3) Insert call in pipeline; (4) Add slot-count `info!` logs. | | `crates/hero_slides_lib/src/prompts/system_validate_slide.md` | Remove Violation #3 "DIAGRAM LABEL ANNOTATIONS" and renumber remaining violations. | --- ## Step-by-Step Implementation Plan ### Step 1 — Revert validator trigger to SVG-only **File:** `crates/hero_slides_lib/src/generator.rs` **Dependencies:** none | **Parallelisable with:** Steps 2, 4 Change the `needs_validation` condition: ```rust // Before: let needs_validation = html.contains("<svg") || html.contains("data-image-slot"); // After: let needs_validation = html.contains("<svg"); ``` Update the adjacent comment to note that image-slot-only slides are handled by `strip_diagram_labels` instead. ### Step 2 — Add slot-count logging at each pipeline stage **File:** `crates/hero_slides_lib/src/generator.rs` **Dependencies:** none | **Parallelisable with:** Steps 1, 4 Insert `info!` log lines after each of these calls: - `replace_fake_logos_with_slots` - `validate_slide_html` - `strip_diagram_labels` (to be added in Step 3) Format: `"generate_slide_html {}: after <stage>, {} slot(s) present"` ### Step 3 — Implement `strip_diagram_labels` and integrate into pipeline **File:** `crates/hero_slides_lib/src/generator.rs` **Dependencies:** Steps 1 and 2 Add a new private function `strip_diagram_labels(html: &str) -> String` that: 1. Collects `top`/`left` estimates for all `data-image-slot` elements. 2. Scans for `position:absolute` elements, extracts their inner text, counts words. 3. Removes elements where word count < 5 AND position is within 200px of any slot. 4. Never removes elements that contain `data-image-slot`. Uses string scanning only (no new crate dependencies). Insert the call in `generate_slide_html` after the validator block and before `extract_image_slot_descriptions`. ### Step 4 — Update `system_validate_slide.md` **File:** `crates/hero_slides_lib/src/prompts/system_validate_slide.md` **Dependencies:** none | **Parallelisable with:** Steps 1, 2 Remove the entire Violation #3 block ("DIAGRAM LABEL ANNOTATIONS") and renumber the remaining violations so they are sequential. --- ## Acceptance Criteria - [ ] A slide with `data-image-slot` but no `<svg>` does NOT trigger the LLM validator pass. - [ ] A slide with `<svg>` still triggers the validator pass. - [ ] Job logs show slot counts after each of the four pipeline stages. - [ ] Short uppercase labels placed near an image slot are removed by `strip_diagram_labels`. - [ ] The image slot itself (`data-image-slot="true"`) is never removed by `strip_diagram_labels`. - [ ] `system_validate_slide.md` no longer contains the "DIAGRAM LABEL ANNOTATIONS" section. - [ ] `cargo build -p hero_slides_lib` succeeds with no new warnings. - [ ] Existing tests pass (`cargo test -p hero_slides_lib`). --- ## Notes - **Why string scanning:** The codebase uses string scanning throughout (`replace_fake_logos_with_slots`, `replace_html_tag_with_slot`, etc.). No new crate dependency is needed or justified. - **False-negative tolerance:** The word-count guard (<5 words) is the primary safety net. False negatives (label not stripped) are acceptable; false positives (content element stripped) are not. - **`browser_pptx.rs` changes are unrelated:** The current uncommitted whitespace diff in `browser_pptx.rs` is not part of this issue. - **Prompt override users:** Because `system_validate_slide.md` is embedded via `include_str!`, editing the file is sufficient. Users with local prompt overrides will need to update their copies.

casper-stevens commented

2026-05-29 08:29:51 +00:00

Author

Member

Implementation Spec — Issue #95

HTML Generation Pipeline: Validator Strips Image Slots + Diagram Label Suppression Rework

Objective

Restore reliable image slot preservation through the HTML generation pipeline by:

Reverting the validator trigger condition so it only fires on slides that contain <svg> elements (its original scope), removing the data-image-slot branch that causes the slot-stripping regression.
Replacing the diagram-label suppression responsibility currently burdened on the LLM validator with a deterministic Rust function that removes position:absolute sibling elements near image slots that match the short-uppercase-label heuristic.
Adding slot-count logging at each pipeline stage so regressions are immediately visible in job logs.
Updating system_validate_slide.md to remove Violation #3 (diagram label annotations), since that concern is now handled deterministically.

Requirements

The validator (validate_slide_html) must only fire when <svg> is present in the HTML, not when data-image-slot is present.
A new deterministic Rust function strip_diagram_labels must run after the validator (and after replace_fake_logos_with_slots) to remove short-text absolute-positioned elements that are siblings of data-image-slot containers.
The heuristic for strip_diagram_labels:
- Target: any <div> or <span> (or <p>) with position:absolute in its style attribute.
- Short text: inner text containing fewer than 5 words and no child elements with data-image-slot.
- Proximity: the element's top/left/right/bottom CSS values place it within 200px of a data-image-slot container's bounding box OR the element is a sibling inside the same parent container as a data-image-slot div.
- Exclusion: never remove elements that themselves have data-image-slot="true".
Slot count must be logged via info! immediately after each pipeline stage: (a) replace_fake_logos_with_slots, (b) validate_slide_html, (c) strip_diagram_labels, (d) extract_image_slot_descriptions.
system_validate_slide.md must have Violation #3 ("DIAGRAM LABEL ANNOTATIONS") removed; the remaining violations must be preserved and renumbered.
system_generate_slide.md is left unchanged.

Files to Modify

File	Change
`crates/hero_slides_lib/src/generator.rs`	(1) Change validator trigger condition; (2) Add `strip_diagram_labels` function; (3) Insert call in pipeline; (4) Add slot-count `info!` logs.
`crates/hero_slides_lib/src/prompts/system_validate_slide.md`	Remove Violation #3 "DIAGRAM LABEL ANNOTATIONS" and renumber remaining violations.

Step-by-Step Implementation Plan

Step 1 — Revert validator trigger to SVG-only

File: crates/hero_slides_lib/src/generator.rs
Dependencies: none | Parallelisable with: Steps 2, 4

Change the needs_validation condition:

// Before:
let needs_validation = html.contains("<svg") || html.contains("data-image-slot");
// After:
let needs_validation = html.contains("<svg");

Update the adjacent comment to note that image-slot-only slides are handled by strip_diagram_labels instead.

Step 2 — Add slot-count logging at each pipeline stage

File: crates/hero_slides_lib/src/generator.rs
Dependencies: none | Parallelisable with: Steps 1, 4

Insert info! log lines after each of these calls:

replace_fake_logos_with_slots
validate_slide_html
strip_diagram_labels (to be added in Step 3)

Format: "generate_slide_html {}: after <stage>, {} slot(s) present"

Step 3 — Implement `strip_diagram_labels` and integrate into pipeline

File: crates/hero_slides_lib/src/generator.rs
Dependencies: Steps 1 and 2

Add a new private function strip_diagram_labels(html: &str) -> String that:

Collects top/left estimates for all data-image-slot elements.
Scans for position:absolute elements, extracts their inner text, counts words.
Removes elements where word count < 5 AND position is within 200px of any slot.
Never removes elements that contain data-image-slot.

Uses string scanning only (no new crate dependencies). Insert the call in generate_slide_html after the validator block and before extract_image_slot_descriptions.

Step 4 — Update `system_validate_slide.md`

File: crates/hero_slides_lib/src/prompts/system_validate_slide.md
Dependencies: none | Parallelisable with: Steps 1, 2

Remove the entire Violation #3 block ("DIAGRAM LABEL ANNOTATIONS") and renumber the remaining violations so they are sequential.

Acceptance Criteria

A slide with data-image-slot but no <svg> does NOT trigger the LLM validator pass.
A slide with <svg> still triggers the validator pass.
Job logs show slot counts after each of the four pipeline stages.
Short uppercase labels placed near an image slot are removed by strip_diagram_labels.
The image slot itself (data-image-slot="true") is never removed by strip_diagram_labels.
system_validate_slide.md no longer contains the "DIAGRAM LABEL ANNOTATIONS" section.
cargo build -p hero_slides_lib succeeds with no new warnings.
Existing tests pass (cargo test -p hero_slides_lib).

Notes

Why string scanning: The codebase uses string scanning throughout (replace_fake_logos_with_slots, replace_html_tag_with_slot, etc.). No new crate dependency is needed or justified.
False-negative tolerance: The word-count guard (<5 words) is the primary safety net. False negatives (label not stripped) are acceptable; false positives (content element stripped) are not.
browser_pptx.rs changes are unrelated: The current uncommitted whitespace diff in browser_pptx.rs is not part of this issue.
Prompt override users: Because system_validate_slide.md is embedded via include_str!, editing the file is sufficient. Users with local prompt overrides will need to update their copies.

## Implementation Spec — Issue #95 ### HTML Generation Pipeline: Validator Strips Image Slots + Diagram Label Suppression Rework --- ## Objective Restore reliable image slot preservation through the HTML generation pipeline by: 1. Reverting the validator trigger condition so it only fires on slides that contain `<svg>` elements (its original scope), removing the `data-image-slot` branch that causes the slot-stripping regression. 2. Replacing the diagram-label suppression responsibility currently burdened on the LLM validator with a deterministic Rust function that removes `position:absolute` sibling elements near image slots that match the short-uppercase-label heuristic. 3. Adding slot-count logging at each pipeline stage so regressions are immediately visible in job logs. 4. Updating `system_validate_slide.md` to remove Violation #3 (diagram label annotations), since that concern is now handled deterministically. --- ## Requirements - The validator (`validate_slide_html`) must only fire when `<svg>` is present in the HTML, not when `data-image-slot` is present. - A new deterministic Rust function `strip_diagram_labels` must run after the validator (and after `replace_fake_logos_with_slots`) to remove short-text absolute-positioned elements that are siblings of `data-image-slot` containers. - The heuristic for `strip_diagram_labels`: - Target: any `<div>` or `<span>` (or `<p>`) with `position:absolute` in its style attribute. - Short text: inner text containing fewer than 5 words and no child elements with `data-image-slot`. - Proximity: the element's `top`/`left`/`right`/`bottom` CSS values place it within 200px of a `data-image-slot` container's bounding box OR the element is a sibling inside the same parent container as a `data-image-slot` div. - Exclusion: never remove elements that themselves have `data-image-slot="true"`. - Slot count must be logged via `info!` immediately after each pipeline stage: (a) `replace_fake_logos_with_slots`, (b) `validate_slide_html`, (c) `strip_diagram_labels`, (d) `extract_image_slot_descriptions`. - `system_validate_slide.md` must have Violation #3 ("DIAGRAM LABEL ANNOTATIONS") removed; the remaining violations must be preserved and renumbered. - `system_generate_slide.md` is left unchanged. --- ## Files to Modify | File | Change | |---|---| | `crates/hero_slides_lib/src/generator.rs` | (1) Change validator trigger condition; (2) Add `strip_diagram_labels` function; (3) Insert call in pipeline; (4) Add slot-count `info!` logs. | | `crates/hero_slides_lib/src/prompts/system_validate_slide.md` | Remove Violation #3 "DIAGRAM LABEL ANNOTATIONS" and renumber remaining violations. | --- ## Step-by-Step Implementation Plan ### Step 1 — Revert validator trigger to SVG-only **File:** `crates/hero_slides_lib/src/generator.rs` **Dependencies:** none | **Parallelisable with:** Steps 2, 4 Change the `needs_validation` condition: ```rust // Before: let needs_validation = html.contains("<svg") || html.contains("data-image-slot"); // After: let needs_validation = html.contains("<svg"); ``` Update the adjacent comment to note that image-slot-only slides are handled by `strip_diagram_labels` instead. ### Step 2 — Add slot-count logging at each pipeline stage **File:** `crates/hero_slides_lib/src/generator.rs` **Dependencies:** none | **Parallelisable with:** Steps 1, 4 Insert `info!` log lines after each of these calls: - `replace_fake_logos_with_slots` - `validate_slide_html` - `strip_diagram_labels` (to be added in Step 3) Format: `"generate_slide_html {}: after <stage>, {} slot(s) present"` ### Step 3 — Implement `strip_diagram_labels` and integrate into pipeline **File:** `crates/hero_slides_lib/src/generator.rs` **Dependencies:** Steps 1 and 2 Add a new private function `strip_diagram_labels(html: &str) -> String` that: 1. Collects `top`/`left` estimates for all `data-image-slot` elements. 2. Scans for `position:absolute` elements, extracts their inner text, counts words. 3. Removes elements where word count < 5 AND position is within 200px of any slot. 4. Never removes elements that contain `data-image-slot`. Uses string scanning only (no new crate dependencies). Insert the call in `generate_slide_html` after the validator block and before `extract_image_slot_descriptions`. ### Step 4 — Update `system_validate_slide.md` **File:** `crates/hero_slides_lib/src/prompts/system_validate_slide.md` **Dependencies:** none | **Parallelisable with:** Steps 1, 2 Remove the entire Violation #3 block ("DIAGRAM LABEL ANNOTATIONS") and renumber the remaining violations so they are sequential. --- ## Acceptance Criteria - [ ] A slide with `data-image-slot` but no `<svg>` does NOT trigger the LLM validator pass. - [ ] A slide with `<svg>` still triggers the validator pass. - [ ] Job logs show slot counts after each of the four pipeline stages. - [ ] Short uppercase labels placed near an image slot are removed by `strip_diagram_labels`. - [ ] The image slot itself (`data-image-slot="true"`) is never removed by `strip_diagram_labels`. - [ ] `system_validate_slide.md` no longer contains the "DIAGRAM LABEL ANNOTATIONS" section. - [ ] `cargo build -p hero_slides_lib` succeeds with no new warnings. - [ ] Existing tests pass (`cargo test -p hero_slides_lib`). --- ## Notes - **Why string scanning:** The codebase uses string scanning throughout (`replace_fake_logos_with_slots`, `replace_html_tag_with_slot`, etc.). No new crate dependency is needed or justified. - **False-negative tolerance:** The word-count guard (<5 words) is the primary safety net. False negatives (label not stripped) are acceptable; false positives (content element stripped) are not. - **`browser_pptx.rs` changes are unrelated:** The current uncommitted whitespace diff in `browser_pptx.rs` is not part of this issue. - **Prompt override users:** Because `system_validate_slide.md` is embedded via `include_str!`, editing the file is sufficient. Users with local prompt overrides will need to update their copies.

casper-stevens commented

2026-05-29 08:41:04 +00:00

Author

Member

Test Results

Passed: 126
Failed: 0
Ignored: 0

Build: success

Details

All 3 test suites passed:

Unit tests (src/lib.rs): 118 passed
Integration tests (deck_safety_test): 6 passed
Doc-tests: 2 passed

Finished in ~1m 14s (including full dependency compilation from scratch).

## Test Results - Passed: 126 - Failed: 0 - Ignored: 0 Build: success ### Details All 3 test suites passed: - Unit tests (src/lib.rs): 118 passed - Integration tests (deck_safety_test): 6 passed - Doc-tests: 2 passed Finished in ~1m 14s (including full dependency compilation from scratch).

casper-stevens commented

2026-05-29 08:41:44 +00:00

Author

Member

Implementation Complete

Changes Made

crates/hero_slides_lib/src/generator.rs

Reverted validator trigger to SVG-only: needs_validation now checks html.contains("<svg") only; the || html.contains("data-image-slot") branch that caused slot stripping is removed.
Added strip_diagram_labels(html: &str) -> String — a deterministic string-scanning function that removes position:absolute elements with fewer than 5 words of inner text that are positioned within 200px of any data-image-slot element. Never removes slot elements themselves.
Integrated strip_diagram_labels into generate_slide_html after the validation pass and before extract_image_slot_descriptions.
Added info! slot-count logs after each pipeline stage: logo-replacement, validation, label-stripping.

crates/hero_slides_lib/src/prompts/system_validate_slide.md

Removed Violation #3 "DIAGRAM LABEL ANNOTATIONS" — this responsibility is now handled deterministically by strip_diagram_labels.
Renumbered remaining violations: OVERLAPPING TEXT is now #3, OVERFLOW is now #4.

Test Results

Passed: 126
Failed: 0
Ignored: 0

All existing tests pass. Build clean.

Acceptance Criteria

A slide with data-image-slot but no <svg> does not trigger the LLM validator pass.
A slide with <svg> still triggers the validator pass.
Job logs show slot counts after each of the four pipeline stages.
Short uppercase labels placed near an image slot are removed by strip_diagram_labels.
The image slot itself (data-image-slot="true") is never removed by strip_diagram_labels.
system_validate_slide.md no longer contains the "DIAGRAM LABEL ANNOTATIONS" section.
cargo build -p hero_slides_lib succeeds with no new warnings.
Existing tests pass.

## Implementation Complete ### Changes Made **`crates/hero_slides_lib/src/generator.rs`** - Reverted validator trigger to SVG-only: `needs_validation` now checks `html.contains("<svg")` only; the `|| html.contains("data-image-slot")` branch that caused slot stripping is removed. - Added `strip_diagram_labels(html: &str) -> String` — a deterministic string-scanning function that removes `position:absolute` elements with fewer than 5 words of inner text that are positioned within 200px of any `data-image-slot` element. Never removes slot elements themselves. - Integrated `strip_diagram_labels` into `generate_slide_html` after the validation pass and before `extract_image_slot_descriptions`. - Added `info!` slot-count logs after each pipeline stage: logo-replacement, validation, label-stripping. **`crates/hero_slides_lib/src/prompts/system_validate_slide.md`** - Removed Violation #3 "DIAGRAM LABEL ANNOTATIONS" — this responsibility is now handled deterministically by `strip_diagram_labels`. - Renumbered remaining violations: OVERLAPPING TEXT is now #3, OVERFLOW is now #4. ### Test Results - Passed: 126 - Failed: 0 - Ignored: 0 All existing tests pass. Build clean. ### Acceptance Criteria - [x] A slide with `data-image-slot` but no `<svg>` does not trigger the LLM validator pass. - [x] A slide with `<svg>` still triggers the validator pass. - [x] Job logs show slot counts after each of the four pipeline stages. - [x] Short uppercase labels placed near an image slot are removed by `strip_diagram_labels`. - [x] The image slot itself (`data-image-slot="true"`) is never removed by `strip_diagram_labels`. - [x] `system_validate_slide.md` no longer contains the "DIAGRAM LABEL ANNOTATIONS" section. - [x] `cargo build -p hero_slides_lib` succeeds with no new warnings. - [x] Existing tests pass.

casper-stevens referenced this issue from a commit

2026-05-29 10:17:21 +00:00

fix(generator): revert validator to SVG-only + deterministic diagram label stripping

casper-stevens closed this issue

2026-06-03 05:54:51 +00:00

Rows
Columns

HTML generation pipeline: validator strips image slots + diagram label suppression needs rework #95

Problem

Root cause

Specific issues

1. Universal validator is wrong tool for diagram labels

2. Diagram label suppression belongs at content level or via regex

3. Two prompts in conflict

4. Late image injection is fragile

5. Animations interact with validator

Recommended fix

Branch

Implementation Spec — Issue #95

HTML Generation Pipeline: Validator Strips Image Slots + Diagram Label Suppression Rework

Objective

Requirements

Files to Modify

Step-by-Step Implementation Plan

Step 1 — Revert validator trigger to SVG-only

Step 2 — Add slot-count logging at each pipeline stage

Step 3 — Implement strip_diagram_labels and integrate into pipeline

Step 4 — Update system_validate_slide.md

Acceptance Criteria

Notes

Implementation Spec — Issue #95

HTML Generation Pipeline: Validator Strips Image Slots + Diagram Label Suppression Rework

Objective

Requirements

Files to Modify

Step-by-Step Implementation Plan

Step 1 — Revert validator trigger to SVG-only

Step 2 — Add slot-count logging at each pipeline stage

Step 3 — Implement strip_diagram_labels and integrate into pipeline

Step 4 — Update system_validate_slide.md

Acceptance Criteria

Notes

Test Results

Details

Implementation Complete

Changes Made

Test Results

Acceptance Criteria

Step 3 — Implement `strip_diagram_labels` and integrate into pipeline

Step 4 — Update `system_validate_slide.md`

Step 3 — Implement `strip_diagram_labels` and integrate into pipeline

Step 4 — Update `system_validate_slide.md`