feat: import existing PPTX into the agent + edit flow#149
Open
naohito2000 wants to merge 18 commits into
Open
Conversation
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
naohito2000
pushed a commit
that referenced
this pull request
May 14, 2026
CodeQL js/incomplete-sanitization (high) flagged attachmentMarker.ts:17 — the previous escaping (replace `"` → `\"` only) does not handle backslash characters, so an input containing `\"` would leave a stray `"` in the marker. Replace the manual escape + surrounding quotes with JSON.stringify(), which handles backslash, quote, and control characters in one step. Behaviour for the current call sites is unchanged (their inputs contain no backslashes), but the code is robust against future inputs and the CodeQL alert clears. Refs: PR #149 CodeQL alert Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Change pptx_to_json on-disk output from a single slides.json to the
deck format used by the rest of the pipeline:
{output_dir}/
deck.json {fonts, defaultTextColor}
slides/slide-NN.json per slide (1-based, zero-padded)
images/ unchanged
Engine + Layer 2/3 changes:
- skill/sdpm/converter/pipeline.py: emit deck.json + slides/ on disk;
in-memory return still includes {slides, fonts, defaultTextColor}.
- skill/sdpm/diff/__init__.py: load deck.json + slides/*.json via
_load_deck_as_roundtrip helper instead of slides.json.
- mcp-local/tools.py::pptx_to_json: docstring + return dict now
includes deck_dir for downstream tooling.
- shared/ingest.py: extend ConversionResult with deck_structure,
slide_count, theme_hints, suggested_name. Theme hints unify Local
/ Cloud / DynamoDB keys; background luminance uses median across
per-slide backgrounds with template lt1 / inverted dk1 fallback.
Tests:
- tests/test_pptx_import.py: scaffolds deck-structure, theme_hints,
upload guide, read_uploaded_file, import_attachment, CLI
acceptance, and DOCX non-regression. Phase 1 covers
TestPptxToJsonDeckStructure / TestConvertPptxThemeHints /
TestNonRegression; later classes go red until Phase 2/5.
Refs: .kiro/specs/pptx-import-edit/tasks.md T1, T2, T8
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Treat PPTX uploads as deck-structure decks across both modes so the
agent can branch into the import-pptx guide instead of the generic
briefing flow.
Local (mcp-local/upload_tools.py):
- _GUIDE_INSTRUCTION shared with Cloud — agents see identical
intent-branching text whether they hit Local or Cloud.
- upload_file: when convert_file produces deck structure, surface
guide / guideInstruction / suggestedName / slideCount / themeHints
on the response. Skip the legacy filePath fallback so agents don't
open just deck.json and miss slides/.
- read_uploaded_file (and helper _format_deck_text_summary): emit a
Markdown-style "--- Slide N: title ---" summary across slides + image
previews. Legacy slides.json branch retained for back-compat.
- import_attachment: return shortId so agents avoid regex extraction;
recursively copy any directory (notably slides/) to
attachments/{shortId}/{dirname}/. Drop legacy slides.json special case.
Cloud (api/index.py + mcp-server/tools/{attachment,upload}.py):
- api/index.py: route .pptx through convert_file (was skipped). Persist
deck-structure metadata to DynamoDB as themeHints / slideCount /
suggestedName. process_upload + get_upload_status share
_pptx_guide_fields helper for parity with Local. Drop dead
_extract_pptx_text helper (replaced by convert_file + _read_converted).
- mcp-server/tools/attachment.py::_import_converted: replace legacy
slides.json case with slides/ recursive copy +
attachments/{shortId}_deck.json + deckJson key. Always include shortId.
- mcp-server/tools/upload.py::_read_converted: detect deck structure
via _format_deck_summary_from_s3 (streams slides from S3, same
Markdown summary as Local).
- mcp-server/server.py + tools/sandbox.py: rewrite run_python docstring
to foreground sandbox helpers; expose helper-based examples while
keeping raw open()/json.load supported.
Refs: .kiro/specs/pptx-import-edit/tasks.md T3, T3b, T3c, T4 (a-d),
T4e, T4f, T4g, T4h
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add the import-pptx guide and the Guide-driven flow contract that routes PPTX uploads through it instead of the new-deck Phase 1 flow. skill/references/guides/import-pptx.md (new): - Step 1-5 walkthrough for converting an uploaded PPTX into an editable deck. Hearings cover edit scope; specs (brief / outline / art-direction) are auto-generated from the PPTX content. Step 5 rebuilds the deck against the PPTX-derived placeholder template copied as deck-local template.pptx by import_attachment. agent/prompts/role/spec_agent.md (Cloud / compose_slides) and mcp-local/acp-agent-prompts/spec-agent.md (Local / use_subagent): - "Flow selection (evaluate this FIRST on every turn)" — classify the turn into guide flow vs new-deck flow before applying anything else. Prevents Phase 1 Briefing drift after a guideInstruction lands. - "Phase 1 Flow" / "Slide Group Assignment" sections explicitly noted as new-deck only; guide flow skips them. - Expanded "Guide-driven flows" section: tracks `uploadId`, `suggestedName`, `slideCount`, `themeHints` across hearing turns so the agent never re-asks the user to upload, and routes returning edit requests to compose_slides (Cloud) or sdpm-composer subagents (Local) after Step 5. prompts/spec-agent.md, prompts/composer-agent.md (deleted): - Leftovers from the c3073be prompt-composition refactor; current branch has zero references (verified via git grep). Removing prevents future edits to dead files. mcp-local/upload_tools.py::_deck_root and mcp-local/server_acp.py::init_presentation: - Guard against the `Path("")` truthy bug when SDPM_DECK_ROOT is unset or whitespace-only — Path('.') is truthy, so the previous `or` fallback never ran and uploads landed under the process cwd. Strip whitespace and only fall back to the home directory when the env string is empty. tests/test_pptx_import.py: - TestDeckRootEnvHandling regression tests covering empty, set, and whitespace-only SDPM_DECK_ROOT. Refs: .kiro/specs/pptx-import-edit/tasks.md T5, T6, T6a, T7 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a translate workflow that creates a derived deck (sibling
directory ``{deck_dir}-{lang}/``) instead of mutating the original.
scripts/translate_extract.py (new):
- Walk slides/*.json and extract translatable text into
derived_deck/translations/{slug}.txt with stable ordering. Skip short
strings (<3 chars by default, configurable via --skip-short) and
preserve styled-text tags so per-run diffs stay reviewable.
scripts/translate_apply.py (new):
- Apply edited derived_deck/translations/{slug}.txt back into
slides/*.json, with --dry-run for verification. Vertical-tab
characters (\\x0b) and styled-text tags survive a round-trip;
structural keys (slug, layout, etc.) are excluded.
skill/references/workflows/translate-pptx.md (rewrite):
- Replace the legacy reverse-convert flow with the deck-structure
flow: derived deck creation, translatable-text extract, hand
translate, apply, build + measure + preview loop. Documents the
styled-text tag rules and the partial-gradient caveat.
tests/test_translate.py (new):
- Cover derived-deck creation, empty-map template, --skip-short,
in-place apply, --dry-run, \\x0b preservation, styled-text tag
preservation, structural-key exclusion, specs untouched. Lint-clean
variant from a496297e is used so pytest + ruff pass.
a496297e's version bump (sdpm 0.1.0 → 0.2.0) is dropped — main has
already moved to 0.3.0 — and its pipeline.py f-string lint fix was
folded into Phase 1.
Refs: .kiro/specs/pptx-import-edit/tasks.md T13, T14, T15
.kiro/steering/versioning.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Eliminate the Step 1 template selection hearing. Each imported PPTX now ships a deck-local placeholder template, so the rebuilt deck honors the source's masters/layouts/theme without the user picking a sdpm template. skill/sdpm/converter/template.py (new): - extract_placeholder_template(pptx, output): copy the source PPTX with all slides removed and slide1.xml→placeholder, producing a master-only template suitable for python-pptx consumption. shared/ingest.py + mcp-local/upload_tools.py + mcp-server/tools/attachment.py: - Run extract_placeholder_template alongside pptx_to_json so the upload directory contains template.pptx; import_attachment copies it into the deck's root as the deck-local template. - Local upload_file's response and the read_uploaded_file deck summary surface the template path so the agent can wire it through Step 4. skill/sdpm/builder/__init__.py: - Resolve "template.pptx" relative to the deck directory so build picks up the deck-local template ahead of bundled sdpm templates. skill/references/guides/import-pptx.md: - Drop the old Step 1 (template selection hearing) entirely; renumber to Steps 1-5. Step 4 sets deck["template"] = "template.pptx" and loops measure → preview → fix until visuals match the source. - Step 4 art-direction.html customization (c857fac1) and Step 4 image src rewrite via image_mapping (844ca607) keep the rebuilt visuals aligned with the source PPTX. Step 4 explicitly calls generate_pptx after the build so the user gets a downloadable file (14927b47). tests/test_pptx_import.py: - TestPlaceholderTemplate, TestPptxBuilderResolvesDeckLocalTemplate, and import-pptx-flow extensions cover template extraction, copy, and end-to-end build. Refs: .kiro/specs/pptx-import-edit/tasks.md FR-9 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two cloud-side fixes that the PPTX upload path needs in production. infra/lib/web-ui-stack.ts + api/requirements.txt: - ApiLambda was bundling only sdpm/analyzer; PPTX uploads require the full sdpm Engine (sdpm.converter, sdpm.utils, sdpm.builder) plus python-pptx, lxml, Pillow, qrcode, pygments, defusedxml — mirror skill/pyproject.toml so upload-time conversion via shared.ingest works without pip-installing sdpm. api/index.py: - Wrap themeHints in DynamoDB Decimal where Lambda code persists the PPTX-derived background luminance (a float). Without this DDB rejected the PutItem with TypeError: Float types are not supported. mcp-server/tools/generate.py: - Resolve "template.pptx" relative to the deck-local directory before falling back to bundled sdpm templates so the PPTX-derived placeholder template (FR-9) is picked up on Cloud builds. Mirrors the skill/sdpm/builder change in Phase 5. Refs: .kiro/specs/pptx-import-edit/tasks.md FR-9 (cloud) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Four bugs that block multi-turn import-pptx flows in the Cloud and
Local web UIs.
decks/page.tsx + ChatPanelShell.tsx (Panel A deckId after creation):
- After init_presentation creates a real deck, Panel A continued
passing deckId="new" to ChatPanel, so each subsequent
/api/agent/invoke call spawned a fresh kiro-cli ACP process and
lost conversation context (agent forgets the deck path on the
next turn).
- ChatPanelShell now passes deckId={panelADeckId ?? "new"} (and a
matching deckName).
- decks/page.tsx falls back to ws.createdDeckId while ws.isNew so
the deck path reaches Panel A via props instead of URL-hash only.
ChatPanel.tsx (chatSessionId mid-stream swap):
- The chatSessionId-watcher useEffect called setSessionId
unconditionally. setSessionId triggers loadHistory, which calls
setMessages with the latest .chat.json — overwriting the in-flight
React state and dropping the ToolUse cards the user is watching.
Before this branch, Panel A always passed deckId="new" so
chatSessionId never changed mid-mount and the bug stayed dormant;
the deckId-fix above lets Panel A adopt a real deck mid-stream and
exposes it.
- Defer the sessionId swap while isLoadingRef is true; flush the
deferred value once isLoading drops back to false.
useChatStream.ts (duplicate parallel chats from rapid double-trigger):
- The React-state isLoading guard does not flip until the next
render, so two near-simultaneous Enter keys / clicks both passed
the guard and produced two parallel /api/agent/invoke streams.
- Add a synchronous isLoadingRef check + eager set, and clear in
finally so the next turn can proceed.
acp-process.ts (running flag never flipped → process not reused):
- handleLine returned early on the JSON-RPC response branch, making
the end_turn detection block (~line 88) unreachable. Each turn
spawned a fresh process, dropping prior context (read_guides,
upload, hearing answers) and making the agent appear to drift back
into Phase 1 briefing.
- Drop the early return so end_turn detection runs after the
pending-promise resolve, restoring process reuse across turns.
f6669360 (ModelSelector setModels/setCurrentModel ref repair) from
the original branch is dropped — main's ChatPanelShell already
extracted ModelSelector into its own file with a self-contained
poll, so the fix is no longer applicable.
Refs: .kiro/specs/pptx-import-edit/notes.md (verification log)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ls.py Phase 1 of this branch checked out mcp-local/tools.py from fe0067f7 to pick up the pptx_to_json deck-structure docstring update. That checkout silently reverted three unrelated improvements that landed on main via PR #110 / #146 (user-style-management): - list_styles signature: dropped the include_all parameter and the user-local + bundled merge via list_styles_filtered. As a result list_styles called from server.py / server_acp.py only returned bundled styles and missed everything in ~/.config/sdpm/styles/, breaking the import-pptx Step 3-3 scaffold-read step for users who keep custom styles there. - generate_pptx / search_assets / list_asset_sources lost their invalidate_manifest_cache() calls, so the long-lived MCP Local process did not pick up user-local asset / config changes between invocations. Restore the main version of tools.py and re-apply only the Phase 1 delta (pptx_to_json docstring + deck_dir return field). This keeps Phase 1 functionality intact while bringing user-local style + asset search back to parity with main. Refs: .kiro/specs/pptx-import-edit/notes.md (Phase 8b verification) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…direction
Phase 8b verification on Cloud showed the rewritten art-direction.html
captured the source PPTX only loosely — backgrounds, accent colors,
and fonts often drifted from the original deck. Root cause: Step 3-3
relied entirely on themeHints, which is a coarse summary (1
luminance, 3 accent hex values, 2 fonts) that omits the real theme
XML and the rendered pixel distribution.
Add a new sub-step 3-3b "Extract the source PPTX's actual design
tokens" that pulls two higher-fidelity signals before token writing:
1. **`analyze_template(template="template.pptx")`** — MCP tool that
reads the theme XML directly and returns the full 12-slot color
map (lt1 / dk1 / accent1-6 / hlink / folHlink), latin / eastAsian
/ complex font pairs, and per-layout placeholder positions.
These are the original PPTX's authoring values, not approximations.
2. **PIL pixel-frequency sampling** — process slide image previews
from the upload's images/ folder, count dominant RGB values, and
convert to hex swatches. Cross-reference with theme_colors to:
- confirm the *actual* background (may differ from
themeHints.backgroundLuminance for non-default masters)
- identify which accent is the deck's hero color (most-used
accent that is not bg/text)
- capture brand colors not declared in the theme as their own
tokens (e.g. --color-brand-orange).
Step 3-3c (renamed from 3-3b) now lists token sources in
authority order: analyze_template > PIL swatches > themeHints >
slide JSON > slide thumbnails. themeHints drops to a sanity check
instead of a primary source.
The PIL sampling recipe mirrors the create-style workflow (used by
sdpm-style-creator agent) so the same color-extraction technique
applies whether the user is creating a fresh style or importing a
PPTX. Both paths now converge on theme XML + pixel evidence as the
authoritative inputs.
Refs: .kiro/specs/pptx-import-edit/notes.md (Phase 8b verification —
Cloud style fidelity gap)
skill/references/workflows/create-style.md (PIL recipe parity)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… rendered previews
Phase 8b verification on Cloud showed two related issues:
- agent skipped Step 3-3 entirely after Step 2, stopping at "what
do you want to edit?" instead of proceeding through specs / build.
- the previous PIL recipe pointed at the upload's images/ folder
(PPTX-embedded photos and logos), not at slide-level renderings,
so even when art-direction was authored the color signal was
weak.
Re-order the guide so build precedes art-direction:
Step 1 init
Step 2 import
Step 3 brief + outline (art-direction removed)
Step 4 build (as-is reproduction against deck-local template.pptx)
Step 5 art-direction.html ← new position
Step 6 present
Rationale: art-direction.html is consumed by the **composer** when
the user later asks to edit slides. The initial reproduction in
Step 4 does not need it — building first against the source's own
placeholder template gives a faithful as-is render that the user
can review immediately, and the rendered previews under previews/
become the most reliable input for authoring art-direction.html.
Step 5-2 PIL recipe now reads from previews/ instead of the upload
images/ directory. Theme XML (analyze_template) plus pixel sampling
on the actual rendered slides produce far more accurate token
values than upload-time approximations.
Step 5-3 spells out that no re-build is needed after writing
art-direction.html — the file is composer-side input for future
edits, not a precondition for the initial deck.
Refs: .kiro/specs/pptx-import-edit/notes.md (Phase 8b ACP run —
art-direction skip + previews-vs-upload-images gap)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…(Lens A/B/C)
Phase 8b feedback: PIL pixel statistics quantify color frequency but
cannot tell the agent what the colors *mean* — that the orange bar
is a section divider, that the rounded box is a card with shadow,
that the cover has a particular layout vs section headers. Decoration
motifs, layout grid, and typography hierarchy require the agent to
actually look at the rendered slides.
Step 5-2 now uses three complementary lenses:
- **Lens A — Visual inspection** via `get_preview(deck_id, slugs=[
cover, section header, content, chart slide, closing], quality="high")`.
The agent reads the actual rendered images for *meaning*: background
type, title vs body color split, decoration motifs (bars, shadows,
radii, dividers, cards, bullets), layout grid, typography
hierarchy. These qualitative tokens are unreachable from theme XML
or PIL statistics alone.
- **Lens B — Theme XML** via `analyze_template`. Authoritative for
--color-* (12 theme slots), --font-* (latin / eastAsian / complex
pairs), and --size-* / position tokens (layouts[] placeholder
positions).
- **Lens C — PIL pixel-frequency** on `previews/`. Quantifies which
theme entries actually appear on screen and surfaces brand colors
not declared in the theme. Cross-checks Lens A's visual impression
with concrete numbers.
Step 5-3 token-source rules now route by token kind:
- color → B + C, A flags brand-only colors
- font → B verbatim
- layout/size → B + A confirmation
- decoration / motif → A primary (B and C cannot give these)
- slide JSON / themeHints → final sanity checks
Refs: .kiro/specs/pptx-import-edit/notes.md (Phase 8b — PIL alone
missed decoration meaning; agent needed to see images directly)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…yle workflow
Phase 8b feedback: art-direction.html ended up with the source deck's
content reproduced inside it (headlines, bullets, charts) and tokens
that disagreed with the actual rendered slides (--color-bg light when
the deck was dark, etc.). Root cause: the guide was describing how to
author a style from scratch in parallel with the canonical
create-style workflow, leaving the style-vs-reproduction distinction
implicit and missing the conventions the workflow already encodes.
Stop re-explaining style authoring in this guide. Step 5 now:
- Loads `read_workflows(["create-style"])` in 5-1 and treats it as
authoritative for HTML skeleton, :root token conventions, .t-*
text classes, demonstration-slide pattern (5-6 slides), absolute
positioning rules, and the violation examples. Removes ~140 lines
of duplicate skeleton / guideline text from this guide.
- Adds a critical reframe up front: art-direction.html is a *style
guide*, not a reproduction. Demonstration slides contain
placeholder text ("Cover Title", "Body sample"), not source-deck
headlines / bullet lists / charts.
- Replaces the long inline HTML example with an import-pptx-specific
token mapping table that routes each token kind to Lens A/B/C:
--color-bg → theme_colors.lt1/dk1 confirmed by PIL (no guesswork);
decoration motifs → Lens A only; layout sizes → theme XML +
Lens A confirmation. The table is the only thing this guide
contributes that create-style cannot.
- Adds a quality bar at the end: no hardcoded hex/pt outside :root,
demonstration slides use placeholder copy, --color-bg matches the
actual deck background, total 5-6 demonstration slides.
Refs: .kiro/specs/pptx-import-edit/notes.md (Phase 8b — art-direction
reproduced source content; --color-bg disagreed with deck;
style conventions duplicated across guides instead of delegated)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
edit-existing.md was the legacy CLI-driven flow for editing an
existing PPTX:
pptx_to_json.py → manual JSON edit → pptx_builder.py generate
That flow has been fully superseded by the import-pptx guide
introduced in this branch:
- Web UI / API entrypoint via upload_file (vs local file path)
- deck-structure (deck.json + slides/*.json) instead of legacy
single-file slides.json
- PPTX-derived placeholder template — keeps the source's masters /
layouts / theme (FR-9), instead of edit-existing's "MUST NOT carry
over colors or styles" rule which forced a target-theme rewrite
- Brief / outline / art-direction auto-generated in the guide
- Composer subagent (compose_slides) for slide JSON edits, instead
of the agent hand-editing JSON
mcp-server/server.py (Cloud) had already commented out its
"Workflow B" pointer, indicating Cloud was already on the
upload_file → guideInstruction path. This commit aligns the rest of
the codebase:
- Delete skill/references/workflows/edit-existing.md.
- Replace mcp-local/server.py "Workflow B" instructions with a
pointer to the upload_file → guideInstruction flow (Web UI / API)
with read_guides(["import-pptx"]) as the CLI fallback.
- Replace skill/SKILL.md "Workflow B" CLI hint with
`pptx_builder.py guides import-pptx`.
- Update mcp-server/server.py comment block to note the cloud-side
upload_file flow.
Side benefit: removes the agent-confusion path observed during
Phase 8b ACP verification — `list_workflows` no longer surfaces a
plausible-looking edit-existing entry that pulled the spec agent
away from the guideInstruction-driven import-pptx flow.
Refs: .kiro/specs/pptx-import-edit/notes.md (Phase 8b ACP — agent
read edit-existing instead of import-pptx via list_workflows)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 8b ACP verification: agent ignored the import-pptx flow,
called list_workflows on its own, and treated the PPTX as generic
reference material. Root cause: upload_file's deck-structure response
fields (guide / guideInstruction / suggestedName / slideCount /
themeHints) were dropped by the web-ui upload pipeline, so the
spec agent never saw the import-pptx pointer it was supposed to
follow.
uploadService.ts:
- UploadedFile interface gains guide, guideInstruction,
suggestedName, slideCount, themeHints (matches the response shape
from mcp-local/upload_tools.upload_file and the Cloud upload
Lambda). Local synchronous path, Cloud process-call path, and
pollUploadStatus all now forward these fields onto the
UploadedFile object.
attachmentMarker.ts::buildAttachedMarker:
- New top branch: when guide + guideInstruction are present, emit a
marker that surfaces them (plus uploadId / suggestedName /
slideCount / themeHints) so the agent reads them as part of the
user message. Without this branch the agent only sees
"[Attached: file.pptx (path: ..., images: ...)]" — no signal that
it should follow a specific guide.
After this fix, the [Attached:...] marker for a PPTX upload contains
guideInstruction text that the spec_agent's "Guide-driven flows"
section is already keyed on, closing the loop:
upload_file response
→ uploadService stores guide+guideInstruction
→ buildAttachedMarker emits them
→ user message contains them
→ spec_agent recognizes guideInstruction and runs read_guides
Refs: .kiro/specs/pptx-import-edit/notes.md (Phase 8b — agent did
not see guideInstruction; called list_workflows / treated PPTX
as reference)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 7 commit 47cb6c23 ported the dup-chat ref guard
(2fc350b0) into useChatStream.ts but missed the second half of the
original fix: the Reconnect useEffect in ChatPanel.tsx also needs
the guard. Without it, when deckId prop swaps mid-stream
("new" → real deck after init_presentation), a second EventSource
consumer attaches to the same SSE alongside the in-flight fetch.
Both sides write to the same setMessages, racing on toolUses /
blocks, and the tool cards either flicker, get clobbered, or
duplicate.
Re-add the guard at the top of the Reconnect useEffect so it
short-circuits while streaming is already in progress. The
isLoadingRef state is shared with useChatStream via a sibling
useEffect (already syncing stream.isLoading), so the guard sees an
up-to-date value.
Refs: .kiro/specs/pptx-import-edit/notes.md (Phase 8b ACP — tool
cards disappearing mid-stream / on subsequent turns)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…rMap
Phase 8b verification with a dark corporate deck (theme XML
bg1=#FFFFFF / dk1=#000000) produced themeHints.backgroundLuminance =
1.0 (white) even though every rendered slide was dark. The agent
then authored an art-direction.html with --color-bg set to white,
which contradicted the actual deck.
Root cause: PowerPoint can ship multiple slideMasters, each with its
own clrMap that remaps the bg1/tx1 slots. Corporate decks routinely
flip bg1=dk1 on a secondary master to declare a dark theme without
editing the theme XML itself. _extract_theme_hints was always
extracting master 0's colors and never consulting the clrMap, so
the bg1 slot was assumed to map to lt1 (#FFFFFF) regardless of how
the deck actually rendered.
Fix:
1. Walk all `ppt/slideMasters/*.xml` and trace which slideMaster
each slide ends up on (slide → slideLayout → slideMaster). The
master used by the most slides is the "dominant" master.
2. Re-extract theme_colors and color_mapping from the dominant
master rather than always master 0.
3. Resolve the bg1 slot through color_mapping before falling back
to lt1: when clrMap declares bg1=dk1 (dark deck), default_bg now
correctly resolves to #000000 instead of #FFFFFF.
Verified on the test deck: backgroundLuminance now reports 0.0 for
this dark deck (was 1.0). 209 tests pass.
Refs: .kiro/specs/pptx-import-edit/notes.md (Phase 8b — dark deck
mis-classified as white because clrMap was never consulted)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…eck_id
Cloud verification: import-pptx Step 5-2 calls
analyze_template(template="template.pptx") to extract theme XML from
the deck-local placeholder template that import_attachment copied
into the deck workspace. The Cloud MCP tool routed every name
through storage.list_templates() (the registered builtin/user
templates DDB), so "template.pptx" — which only exists at
decks/{deck_id}/template.pptx — was reported as not found. Step 5
failed.
Add a deck-local short-circuit. When template == "template.pptx",
require a deck_id arg, download decks/{deck_id}/template.pptx from
the workspace bucket, run sdpm.analyzer.analyze_template on the
downloaded file, and return the same shape (theme_colors / fonts /
layouts) that registered templates produce. The shared
shared.ingest fix from 2281351c (multi-master clrMap detection)
applies here automatically because analyze_template uses the same
sdpm.analyzer code path.
Update import-pptx guide Step 5-2 to spell the Cloud signature:
analyze_template(template="template.pptx", deck_id=<deck_id>)
Local mode is unchanged: mcp-local's analyze_template already
accepts a path / template name and never went through DDB lookup.
Refs: .kiro/specs/pptx-import-edit/notes.md (Phase 8b Cloud — Step
5 "Analyzing template" failed because deck-local template.pptx
was not registered)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
CodeQL js/incomplete-sanitization (high) flagged attachmentMarker.ts:17 — the previous escaping (replace `"` → `\"` only) does not handle backslash characters, so an input containing `\"` would leave a stray `"` in the marker. Replace the manual escape + surrounding quotes with JSON.stringify(), which handles backslash, quote, and control characters in one step. Behaviour for the current call sites is unchanged (their inputs contain no backslashes), but the code is robust against future inputs and the CodeQL alert clears. Refs: PR #149 CodeQL alert Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
naohito2000
force-pushed
the
feat/pptx-import-edit
branch
from
a25e1fd to
eb511ee
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds an end-to-end import-pptx flow so users can drop an existing PPTX onto
the agent and continue authoring / editing from that deck instead of starting
from a sdpm template.
pptx_to_jsonnow emits the deck-structure layout(
deck.json+slides/slide-NN.json+images/) and copies the sourcePPTX as the deck-local placeholder template (
template.pptx).attached PPTX, generate slide-structure + theme hints, and forward
guide/guideInstruction/themeHintsto the agent via the[Attached: …]marker so it routes into the import-pptx guide.translate_extract/translate_applyover thedeck structure so re-translation does not re-render the whole deck.
import-pptx.mdguide drives the 6-stepflow (init → import → brief+outline → build → art-direction → present);
edit-existingis removed (its responsibilities are now covered bythe import-pptx guide).
get_preview/theme XML
analyze_template/ PIL on rendered previews) to ground thestyle in the source deck. Style authoring itself is delegated to the
existing
create-styleworkflow.analyze_template(template="template.pptx", deck_id=…)short-circuits to the deck-local placeholder template downloaded from
the workspace bucket;
ApiLambdabundlessdpm; Decimal handling onDDB writebacks fixed.
_extract_theme_hintswalks allslideMasters, picks the dominant one by slide-usage, and resolves the
bg slot through its
clrMap. Corporate decks that flipbg1=dk1on a secondary master now report the correct darkbackground.
UploadedFilecarries the new fields end-to-end; theReconnect
useEffectre-adds theisLoadingRefguard so tool-cardsdo not flicker on stream reconnect;
mcp-localuser-local style /asset support restored after a checkout regression.
17 commits, rebased clean onto
main. No conflicts.Test plan
pytest— 209 passed, 2 skipped(light / dark / corporate)
including
analyze_templateagainst the deck-local placeholdertemplate
tokens after the dominant-slideMaster
clrMapfixnew
analyze_templatedeck-local short-circuitKnown follow-up
pages — for "all-page same background" decks this produces correct
output, but hybrid-style decks (cover dark / body light) can
still trip the agent's prior. Tracking as a follow-up to
(a) widen the sample to span the deck and (b) make
themeHints.backgroundLuminancethe primary signal in the--color-bgtoken table.🤖 Generated with Claude Code