Changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning (while in 0.x, breaking changes may occur in MINOR releases).

Entries before v0.5.0 were written retroactively as summaries.

Unreleased

Changed

• Slide list moves instead of jumping — slides that appear, disappear or move in the Slides tab slide into place (both grid and full view); a regenerated preview crossfades over the old image instead of flashing a skeleton, and its glow starts when the new image is actually on screen. Clicking a thumbnail in grid view opens the full view at that slide (the card grows into place). Reveal stagger is capped so the last slide of a long deck no longer waits. All motion honours the OS reduced-motion setting.
• Chat input height is always right — the message box now grows with its content via react-textarea-autosize (1–5 rows) instead of a hand-rolled height calculation that could misjudge IME composition and pasted text.
• Compose animation is cheap again on large decks — only slides in view are drawn by the agent cursors (at most two at a time); changes that land on a slide you are not looking at are kept undrawn and replayed with the same cursor animation when you reach it (within 60 s; older changes just appear). One shared frame loop drives every typewriter, defs are mounted once per deck, off-screen slides skip layout and paint, and the per-slide polling is gone. Main-thread work during a compose burst drops by about half and a finished deck scrolls at full frame rate. The full view now follows the slide being composed until you scroll yourself.

0.9.0 - 2026-09-22

Added

• Outline view as a slide storyboard — the Web UI's outline tab shows each slide as a 16:9 card: the claim as headline, the body in full, visual and evidence small at the bottom; chapters (##) divide the deck. Grid by default with a persisted column alternative. Text never drops below the 11px floor.
• Edit the outline in the storyboard and send the changes — every card field, chapter heading and the deck name are editable in place; add (ghost card), duplicate, delete with undo, reorder by drag (across chapters, keyboard too) or ⌥↑/↓, ⌘Z/⇧⌘Z history, a markdown drawer with changed lines highlighted. "Send changes" (⌘S) writes specs/outline.md and posts a user message to the chat carrying the unified diff (with a note that new slide-N slugs are placeholders, and — toggle, on by default — a request to polish the wording), so the agent picks the edit up in the same conversation. No agent-side changes: the file is written unconditionally and the diff is how the agent learns what happened. New API PUT /decks/{id}/specs/outline (cloud) and matching local route. Web UI now depends on @dnd-kit/react, motion and react-textarea-autosize.
• sdpm-composer skill entry point — a fourth thin dispatcher (skills/sdpm-composer) for the role a spawned composer sub-agent plays; it calls read_workflows(["composer"]) and stops. Dedicated composer agent definitions (Kiro, Claude Code) now point at this skill instead of a personas/composer.md file reference.
• docs/en/migration-role-workflows.md — breaking-change table and per-environment migration steps for the workflow consolidation below.

Changed

• Compose is two passes: layout, then content — the scaffold pass becomes the layout pass (task_instruction: "Layout pass."): one composer sees every slide and decides each slide's template layout, frame elements and named content regions (_comment elements with x/y/w/h, invisible in output), derived from one grid, written in one batched run_python call. Content composers realize their slides inside those regions. The consistency-review and fixes passes are gone: layout consistency is now by construction, and the orchestrator looks at previews and re-dispatches a composer per slide that needs a change.
• run_python requires deck_id — the sandbox always runs inside a deck workspace. The workspace-less "calculation" mode is gone: on the cloud it accepted file writes and discarded them silently (a composer lost a whole scaffold pass this way), on the local server it could not write at all. Create the deck with init_presentation first, then compute. Also new: check_specs validates deck.json and outline.md before composing (compose_slides runs it first).
• Web UI Spec / Vibe and "Parallel agents" keep their behaviour without personas — the orchestrator workflow defines two tokens, Interaction mode: dialogue (the user wants to shape the deck in conversation) and Interaction mode: fast (build from material); the cloud agent passes the token as a one-line system part and the local ACP route prepends it to a session's first prompt. single (parallel agents off) is the same workflow with no composer sub-agents, composing slides itself.
• Outline sub-items are now body / visual / evidence — these three keys replace what_to_say / what_to_show / evidence / notes. This is breaking for existing enriched outlines: old-key lines are shown as prose rather than parsed as slide sub-items.
• Mode behavior consolidated into role documents, served via read_workflows — start_presentation(mode=...) and personas/*.md are removed. Each role (orchestrator, composer, style, translate) now has exactly one document, sdpm/references/workflows/<role>.md, containing both the role definition and its procedure; entry points (skills, agent definitions, SKILL.md, server instructions) only ever name a role for read_workflows to fetch, never restate its behavior. This removes the persona/workflow duplication that v0.5's persona layer had reintroduced.
• vibe and spec modes merged into a single sdpm-create skill — the dialogue-depth distinction is gone; how much back-and-forth happens is driven by the user's own wording, not a mode argument. skills/sdpm-vibe and skills/sdpm-spec are replaced by skills/sdpm-create. skills/sdpm-style and skills/sdpm-translate are unchanged in purpose, now dispatching to read_workflows(["style"]) / read_workflows(["translate"]).
• CLI subcommands renamed to match MCP tool (contract) names — e.g. generate → generate_pptx, examples → read_examples, workflows → read_workflows, guides → read_guides, analyze-template → analyze_template, search-assets → search_assets, list-templates → list_templates, init → init_presentation, code-block → code_to_slide, layout → arch_diagram, diff → diff_pptx. Workflow/guide text now reads identically whether called as an MCP tool or a CLI subcommand. No aliases for the old names. See Migration: role workflows for the full table.
• slide-json-spec moved to sdpm/references/spec/ — it is a fact document, not a role document; still resolved by read_workflows for backward compatibility. hand-edit-sync moved to sdpm/references/guides/ — it is an occasional-need procedure, not a role.

Removed

• sdpm/references/examples/patterns.pptx and the search-patterns CLI/tool — an audit of all 18 cataloged patterns found their techniques either duplicated in components/all or not measurably improving output versus letting the model reason from the style HTML and component vocabulary directly. No migration path; if a specific technique is needed again, express it directly in slide JSON per read_workflows(["slide-json-spec"]).

Migrating from v0.5? See Migration: role workflows for the full breaking-change list and upgrade steps per environment.

0.8.2 - 2026-09-21

Fixed

• WebP previews and deck thumbnails were not generated after v0.8.1. schedule_webp_background relied on an asyncio event loop; tools now run on worker threads, so the task silently never started and generate_pptx produced no slide previews and no thumbnail. It runs on a daemon thread now.
• get_preview right after run_python could see a missing or stale image. v0.8.1 moved preview generation into a background task; get_preview now waits for that task (per deck, up to 90s), and the task renders the measured slugs' WebP before the live-preview JSON.
• Code Interpreter session start could stall for minutes on AgentCore V2. Every AWS client now uses a 5s connect timeout. Prompt APIs (S3, DynamoDB, SSM, session start/stop, embeddings) get read 30s / 3 attempts; the user-code invoke_code_interpreter gets read 600s and a single attempt so a stall can never execute the code twice.

0.8.1 - 2026-09-20

Fixed

• Remote MCP server no longer stalls or dies during long tool calls. FastMCP ran synchronous tools on the event loop, so a minute in LibreOffice blocked AgentCore's /ping and the session was terminated mid-run. Every remote tool now runs in a worker thread (offloaded_tool); the WebP background task does too.
• AgentCore Runtime platform V2: first S3 call after restore no longer hangs. boto3 clients built at import carried connection state into the V2 snapshot; the first call on a restored microVM could block for minutes (observed: list_templates unanswered after 9 min). Clients are now built on first use, with bounded botocore timeouts (connect 5s / read 30s / 3 attempts / TCP keepalive).
• AgentCore Runtime platform V2: LibreOffice's first run per microVM (~60s) is warmed in the background. The restored root filesystem is lazily fetched, so the first soffice run paid ~60s of first-touch disk reads (3s thereafter). The server now converts a blank template in the background when an MCP session starts, before the composer reaches its first run_python. Measured: layout-pass run_python 85s → 14s.

Changed

• run_python returns as soon as measure is done. The live-preview JSON (compose) and the measured slugs' WebP finish in a background thread; previewHint now says the images are being generated. Previews render only the measured pages, the LibreOffice SVG is parsed once, and per-slug compose runs in a thread pool. Phase timings (prepare_s3, build, measure, artifact_s3, background compose) are logged at INFO.

0.8.0 - 2026-09-20

Added

• Scaffold pass in the compose workflow — before the parallel content composers fan out, the orchestrator now dispatches one composer with task_instruction: "Scaffold pass." (a new composer mode alongside "Consistency review."). It batch-writes the style- and role-derived elements every slide shares — decoration from the art direction (accent bars, footer, title band) and role elements with per-slide parameters (titles/subtitles drafted from the outline, section labels) — into every slides/*.json via a single programmatic run_python for-loop (no individual slide edits in this mode). Elements that require imagining a slide's content stay with the content composers. Page numbers are explicitly forbidden as elements — they come from the template's native slide-number placeholder. Content composers then read the existing JSON and build on the frame instead of rewriting whole files. Cross-slide decoration consistency moves from post-hoc review fixes to by-construction, and near-identical JSON is no longer re-emitted per slide (token savings). Persona-only change (personas/composer.md, vibe.md, spec.md).

• Per-user usage measurement (#326) — Bedrock usage logs (bedrock_usage) now carry user_id / session_id / deck_id (composer invocations included), and new structured events count slides per user: slides_composed (per compose_slides call) and slides_built (per successful PPTX build). A new opt-in flag features.enableTransactionSearch (or deploy.sh --enable-transaction-search) enables CloudWatch Transaction Search for span-level per-user token queries and the GenAI Observability dashboard. See Measuring Usage.

• sendToBack element key — z-order control over placeholders — elements normally render in front of template placeholders (the builder fills placeholders first), so a full-bleed decorative image always covered the title with no recourse. "sendToBack": true moves an element behind the placeholders; multiple marked elements keep their relative order. The PPTX importer now emits it automatically for shapes that preceded the first used placeholder in the source slide, so imported decks keep their original stacking. (#316)

• pptx_builder.py preview -o/--output-dir — the CLI preview can now write PNGs to a chosen directory (e.g. {deck_dir}/preview) instead of always using a temporary _work/ folder; --no-grid already existed. sdpm.api.preview gains the matching output_dir parameter. (#316)

• repr and __name__ in the run_python sandbox — agent code using repr() or the __main__ idiom no longer needs rewriting. (#316)

• translate mode with a /sdpm-translate entry point — deck translation is now a first-class mode: start_presentation(mode="translate") serves personas/translate.md, and a fourth thin skill entry point (skills/sdpm-translate) exposes it as a slash command in every client that maps skills, alongside vibe/spec/style. The persona drives the existing translate-pptx workflow (extract → dictionary → apply → build) and makes the agent itself fill the translation dictionary.

• make install-kiro generates a dedicated composer agent again — <KIRO_HOME>/agents/sdpm-composer.json gives parallel compose workers the sdpm MCP server only, with pre-approved tools and a generous startup timeout. Without it, workers fall back to a general-purpose agent that cold-starts every MCP server in the profile per worker; on MCP-heavy profiles this made some parallel composers miss the sdpm server nondeterministically (observed with 30+ registered servers). The generated config is a thin pointer — its prompt is a file:// reference into personas/composer.md, never inline behavior — and it is lifecycle-managed by the same ownership audit as the rest of the wiring (own/stale are regenerated, foreign/unknown are never touched; power mode still removes it). v0.5.3 dropped generation because the old installer could not manage the file safely across checkouts; the ownership audit removed that reason.

• Mode entry points are back, as thin dispatchers — skills/sdpm-vibe, skills/sdpm-spec and skills/sdpm-style let users pick a mode explicitly (/sdpm-vibe and friends in clients that expose skills as slash commands). v0.5.0 removed skills/ because each file had grown a full copy of the mode behavior that then drifted; these are ~20 lines each and contain no behavior text. They call start_presentation(mode=...) and stop, so personas/*.md remains the only definition. A test fails if any substantial persona line reappears in an entry point.

• The repository root is an Agent Plugins 1.0.0 package — plugin.json and mcp.json make it loadable by Kiro (as a Power), Cursor, GitHub Copilot and VS Code without a client-specific installer. mcp.json points uv at ${PLUGIN_ROOT}/servers/local and puts the virtualenv under ${PLUGIN_DATA}, because clients copy the plugin into an install cache that may be read-only and is replaced on update.

• Codex support — .codex-plugin/plugin.json, a bundled MCP server definition in .mcp.json, and a repo marketplace at .agents/plugins/marketplace.json, so codex plugin marketplace add ./ is enough to install the checkout locally. Codex was previously limited to the Layer 1 CLI path.

• Claude Code exposes the same entry points — .claude-plugin/plugin.json gained skills (bumped to 0.3.0). Its agents and mcpServers blocks are unchanged.

• clients/kiro/install.py gained --mode {auto,legacy,power}, --kiro-home and --replace-existing.

• GPT-6 Astra (global.openai.gpt-6-astra) is selectable for chat and create. Throughput is roughly a third of GPT-5.6 Terra (~25 vs ~90 output tokens/s measured in ap-northeast-1), so expect longer waits on compose-heavy runs.

• Kimi K3 (global.moonshotai.kimi-k3) is selectable for chat and create. Its Converse constraints are identical to the GPT models' — it rejects temperature, topP and Bedrock's cachePoint, while its own implicit prompt caching is on by default — so it reuses the same invocation profile and needed no new agent code. Throughput is competitive with GPT-5.6 Terra (~80–175 output tokens/s measured in ap-northeast-1). Note it always emits reasoning content, and those tokens count against the output budget: a one-word answer costs ~50 output tokens where Claude spends ~5, so short interactions are disproportionately expensive.

Changed

• MCP servers connect concurrently instead of one after another — Strands connects them serially: ToolRegistry.process_tools() iterates the tools list and blocks on await provider.load_tools() for each ToolProvider in turn. With three servers — the Presentation Maker one on AgentCore plus two AWS servers pinned to us-east-1 — that put their handshakes on the critical path back to back, and a profile attributed roughly 2.6s to the two AWS ones alone. They are now connected up front in a thread pool; load_tools() caches its result, so Strands' own call is a cache hit. Measured on the dev stack: agent setup went from 7.72s to 5.14s. As a side effect the required flag in MCP_DEFS now does what it always claimed — it was only guarding client construction, which is lazy and cannot fail, so an unreachable optional server used to surface as a hard failure inside Strands (MCPClient defaults to continue_on_error=False). An optional server that cannot be reached is now dropped with a status entry.

• diff_pptx is no longer offered on the cloud path — servers/remote does not bind it, so listing it in the agent's tool allowlist produced a "not found on MCP server" warning on every request. The hand-edit sync workflow stays a local/CLI capability and the workflow document says so; the tool is slated for removal.

• OpenAI GPT models are served through the Converse API instead of bedrock-mantle. They are now registered under their global inference profile ids (global.openai.gpt-5.6-terra, global.openai.gpt-6-astra) and go through the same BedrockModel path as every other model. This removes agent/mantle_client.py, the MANTLE_MODELS / MANTLE_RESPONSES_MODELS tables and their region-resolution helper, the OpenAI SDK dependency (strands-agents[openai]), and the bedrock-mantle:CreateInference / bedrock-mantle:CallWithBearerToken IAM grants on the agent role. Prompt caching still applies: these models do not accept Bedrock's cachePoint block, but their own implicit prompt caching is on by default.

Removed

• GPT-5.5 and GPT-5.4 — not available as Bedrock foundation models, so they cannot be reached over the Converse API. They were only ever callable through the removed bedrock-mantle path.
• Nova 2 Lite — it was registered as us.amazon.nova-2-lite-v1:0, a US-only cross-region inference profile that does not exist in the deployment region, so selecting it always failed with ValidationException: The provided model identifier is invalid.

Fixed

• The Claude Code plugin manifest tracks the engine version again — .claude-plugin/plugin.json sat at 0.3.0 while the engine reached 0.7.1. plugin.json and .codex-plugin/plugin.json each had a test asserting they follow sdpm.__version__; this manifest had none, so the drift went unnoticed across several releases. It now carries the same guard.

• The knowledge base id is resolved per request instead of at import — the remote MCP server read KB_SSM_PARAM from SSM at module import and built its KBSync from the result. SSM exists so that value can change without a redeploy, so caching it at startup pinned the runtime to whatever the id was when the process began. It is now resolved on first use and refreshed every 300s. This also matters on AgentCore Runtime platformVersion V2, which snapshots the process after initialization and shares that memory state with every restored instance: a value read at import is frozen for the life of the snapshot, where V1's periodic container recycling used to re-read it. Tool registration for search_slides now depends on configuration rather than on a successful read, so a transient SSM failure no longer removes the tool for the life of the process.

• MCP requests no longer start a new microVM on every agent turn — the agent connected to the Presentation Maker MCP server on AgentCore Runtime without sending Mcp-Session-Id. AgentCore routes MCP requests to a microVM by that header and mints a fresh session id for any request arriving without one, so each agent turn — plus each composer slide group, which builds its own client — paid a new-session start on a 630 MiB image. The caller's session id is now forwarded, and composer groups use {session_id}-g{n} so they keep the per-group isolation they were written for while still reusing their own microVM across composes. Measured in ap-northeast-1: 0.18-0.20s with a consistent id, 0.59s without one when warm capacity happened to be available, and 17.6-19.7s without one when it was not. Replaying an id after the 900s idle timeout returns HTTP 200 and re-establishes the session, so no re-initialize path is needed (verified for the stateless MCP server this project runs).

• Importing an attachment from a URL works again — fetch_url sends Connection: close, which makes http.client set response.will_close, and getresponse() then hands the connection to the response by calling HTTPConnection.close() — setting conn.sock to None before a single byte of the body is read. The body loop re-armed its idle read timeout on that socket every iteration, so every URL fetch died with AttributeError: 'NoneType' object has no attribute 'settimeout' and the agent reported the failure as a sandbox networking problem. The loop now re-arms the timeout on a socket reference captured before getresponse(), and stops once the body is closed (the fd is really gone by then). The idle read timeout is preserved rather than dropped: it is still bounded by both IDLE_READ_TIMEOUT_S and the remaining total budget on every read.

• Tool results carrying images work on GPT models — OpenAI GPT models on Bedrock Converse reject an image block nested inside a toolResult (ValidationException: This model doesn't support the image field for user messages), which broke slide-preview review in compose and web_fetch on image URLs. A BeforeModelCallEvent hook (agent/message_hooks.py) now moves such images up to sit beside the toolResult in the same user message — a shape both GPT and Claude accept, so it is applied to every model rather than gated per model. The image still reaches the model, so vision is preserved. It runs before every model call (not on message-added), so it also normalizes pre-seeded composer history and restored sessions.

• GPT models no longer send temperature — both GPT-6 Astra and GPT-5.6 Terra reject the field on Converse (This model doesn't support the temperature field). The removed bedrock-mantle path had accepted it, so the profile carried temperature=0.7 until the migration.

• Bedrock invocation-logging custom resource is now idempotent — it previously overwrote an existing account-level logging configuration on deploy; it now checks first and skips if logging is already configured to a different destination (matching the documented "skipped if already configured" behavior). (#326)

• The deck-translation workflow is reachable again — translate-pptx existed as a workflow document but nothing routed to it: the MCP server instructions menu stopped at D and no persona mentioned it, so an agent asked to translate a deck could only find it by spontaneously calling list_workflows. It is now Workflow E in the server instructions (pointing at the new translate mode), with routing lines in the vibe and single personas. The translate scripts also moved from the repo-root scripts/ (deploy helpers, not shipped with the skill) to sdpm/scripts/ next to pptx_builder.py, so the workflow's scripts/translate_extract.py commands resolve relative to the skill root like every other workflow — and Layer 1 skill installs actually get them.

• make install-kiro no longer hijacks another checkout's Kiro setup — the MCP registration was re-registered with --force whenever the existing sdpm entry pointed elsewhere, and the composer-agent cleanup deliberately deleted configs generated for any checkout. Running the installer from a second clone therefore repointed a working install. Every artifact the installer can write (MCP registration, skill symlinks, legacy composer agent) is now classified as own / stale / foreign / unknown against the current checkout; only own, stale and absent are written, foreign requires --replace-existing, and unknown is never touched. Conflicts exit non-zero with a report instead of succeeding quietly. Behavior change: a leftover pointing at a checkout that still exists is now left alone. One pointing at a moved checkout is still repaired, which was the staleness that cleanup existed for.

• clients/kiro/install.py honours KIRO_HOME — it hardcoded ~/.kiro, so installs into a non-default Kiro profile silently wrote to the wrong one. KIRO_HOME is now also exported to the kiro-cli child process, since mcp add --scope global resolves "global" itself. Powers are global-scope only, so a separate profile is the only way to keep a Kiro CLI install and a Power install from colliding.

• Importer: white text baked onto light cards (white-on-white) — the converter adopted the shape style's fontRef color even when the shape's own lstStyle carried an explicit text color, which outranks fontRef in OOXML. Decks whose light-fill cards define dark text via lstStyle came back with fontColor: #FFFFFF. (#316)

• Importer/builder: arrowheads lost their size — tailEnd/headEnd w/len attributes are now round-tripped (arrowEndWidth, arrowEndLength, arrowStartWidth, arrowStartLength), so a large arrowhead no longer comes back as the small default.

• Importer/builder: gradient direction distorted on non-square shapes — the scaled attribute of linear gradients was dropped on import and forced to 1 on build. A scaled="0" gradient (true geometric angle) on a wide bar now renders at its original angle.

• Importer/builder: partially-rounded rectangles on pictures — image masks were limited to a 9-name allowlist and dropped their adjustment values. Any OOXML preset now passes through (round1Rect, round2SameRect, …) with its avLst preserved, so e.g. a picture panel with one rounded corner survives the round trip.

• measure: white-on-white false positives over gradient shapes — the contrast judge only recognized gradients written as style="fill:url(…)"; LibreOffice sometimes emits fill="url(…)" as an attribute, making the background "unknown" and falling back to the white slide background. Both forms are now treated as unknowable background (contrast check skipped). (#316)

• lint: shape-unknown-name noise on imported decks — camelCase raw OOXML presets (which the builder renders verbatim) are no longer flagged; pure-lowercase typos still are. (#316)

0.7.1 - 2026-08-05

Fixed

• Nested styled-text notation no longer leaks raw tags onto slides — agents sometimes emit {{bold:{{#FF0000:X}}}} instead of the canonical {{bold,#FF0000:X}}; the parser could not see through nesting and rendered the inner tag as literal text. A flatten pre-pass now normalizes nesting (any depth, partial nesting included) to the comma form, with inner attributes taking priority. Non-nested input is untouched. (#123)

• Builtin template download works from the Web UI — downloading blank-dark / blank-light did nothing: the download used a fetch()+blob path that requires CORS, and the builtin resource bucket has no CORS configuration (user templates live in a different bucket that has one). Downloads now use the same direct-link navigation as deck PPTX downloads (no CORS involved), the presigned URL carries Content-Disposition: attachment, and failures show an error toast instead of silently doing nothing. (#281)

• grid tool returns actionable errors and supports % / repeat() — unsupported CSS track syntax (auto, minmax(), previously also % and repeat()) crashed with an uncaught ValueError instead of an error message the agent can react to. % and repeat(n, X) are now supported, unsupported syntax returns {"error": ...} naming the offending token, and the supported subset is documented in the tool docstring and grid guide. (#282)

• Live preview: icons relying on even-odd fill no longer render as solid boxes — LibreOffice's SVG export declares fill-rule="evenodd" only on the root <svg>, which the per-component fragment split dropped, so multi-subpath line-art icons (many AWS resource icons) lost their cutouts and appeared as filled rectangles in the live preview (final PNGs were unaffected). Root inheritable attributes are now propagated onto each fragment. (#288)

• Architecture diagram box auto-height now works on non-16:9 templates and with CJK text — the engine used a fixed 16:9 pt-to-px ratio for text measurement, causing boxes to undersize on 4:3 and other aspect ratios, and did not account for fullwidth (CJK) character width. analyze_template now reports ptPerPx in slide_size, the agent records it in deck.json slideSize, and arch_diagram accepts a pt_per_px parameter for accurate calibration. (#285)

• Remote MCP: analyze_template now includes slide_size in cached results — the cached template analysis omitted the slide_size field, so subsequent calls returned an incomplete response and the agent could not populate deck.json slideSize. (#285)

• Custom templates with non-16:9 slide sizes (4:3 etc.) now lay out correctly — the engine's px coordinate system followed the template width but assumed a fixed height of 1080, so 4:3 decks left the bottom quarter of every slide empty, reported false out-of-bounds warnings, mismatched placeholder coordinates, mis-measured text overflow, and rendered cropped previews in the Web UI. The canvas is now derived from the template's real dimensions (1920 px wide, height following the aspect ratio — 4:3 becomes 1920×1440), and analyze_template / deck.json carry the canvas size so slide composition uses the full slide. Behaviour for 16:9 templates is unchanged. Known limitation: architecture diagram boxes with an omitted box.height can still under-estimate text height on non-16:9 templates — specify box.height explicitly. (#208)

• Live slide preview: the first few slides now animate — animation was suppressed for 3 seconds after the slides tab appeared, and because a new deck switches to that tab as soon as slides arrive, the first slides were always shown instantly. Suppression is now based on whether the slides already existed when the view mounted, instead of a timer.

• Live slide preview: the PNG fallback now actually appears — the error state was reset on every 1-second poll, so the fallback was unmounted before it could be seen; a failure to find the render container also marked the slide as permanently processed, leaving an empty black box. Slides with nothing to draw now fall back to the rendered PNG, and the fallback image retries expired signed URLs.

• AWS: uploaded custom templates now apply to PPTX generation — the remote server's template resolution only searched builtin templates, so a deck referencing an uploaded user template silently fell back to blank-dark.pptx. Generation now resolves user templates first (same order as analyze_template), and an unresolvable template name raises an explicit error listing available templates instead of silently using the wrong design. (#206)

0.7.0 - 2026-08-03

Added

• Web UI: live design studio — full visual redesign. Light/Dark/System themes (default dark), 90–125% text scaling, studio color tokens (5 agent work colors, ink for deliverables, red reserved for errors), and a two-voice typography system (Bricolage Grotesque for UI chrome, Fraunces for document surfaces).
• Web UI: artifact-first review surfaces — chat tool activity as a compact agent work ledger; brief as a reviewable contract document with approval state; outline as single-column narrative slides (number rail, slim skeleton cards, enriched slide faces with an evidence/visual/notes spec sheet); art direction as per-slide style cards sharing the same rail grammar, with template and style sections on one alignment axis.
• Outline workflow now asks for ## section headings when a deck has distinct parts, so review surfaces can render the story arc as chapters.

Changed

• Style cover extraction is unified client-side: both /styles APIs (cloud and local) now serve raw style HTML (html field replaces coverHtml), and the style gallery opens previews with zero additional round-trips.
• Style previews and thumbnails keep the style author's own canvas background instead of forcing transparency.

Fixed

• Style thumbnails rendered the first slide at 70% (standalone-viewing body zoom was not reset in the cloud cover path), leaving gutters around the cover; slides now fill thumbnails edge-to-edge.
• Unpainted regions of style slides (rounded corners, frame decorations) showed as opaque white inside the dark UI.
• One deck slide always fits the viewport in the full-size carousel view.
• Prose art direction (art-direction.md, no style selected) rendered as raw markdown inside an iframe; it is now typeset as a document.

0.6.0 - 2026-08-02

Breaking

• Stateless attachment pipeline — upload_file, read_uploaded_file, and public MCP pptx_to_json tools are removed. Use read_attachment(source) and import_attachment(source, deck_id) instead. The new tools are stateless (no session storage, no uploadId) and accept local paths, S3 keys, or URLs directly.
• measure_slides standalone tool removed — measurement is now triggered exclusively via run_python(measure_slides=[...]).
• list_asset_sources tool removed — call search_assets(query="") for the same discovery listing (sources with counts).
• run_python / run_style_python save parameter removed — writes always persist; the deprecated flag is no longer accepted.
• run_python files parameter removed — use read_attachment to access uploaded file content, then reference by path in code.
• Session restart required — after git pull, restart all Local Web UI / ACP sessions. The next spawn will pick up the updated agent definitions (agents-sync.ts re-derives from acp-agents/). make install-kiro is only needed for global MCP config cleanup, not for allowlist updates.

Changed

• search_assets now supports discovery mode: calling with an empty query (query="") returns all asset sources with counts, replacing the removed list_asset_sources tool.
• diff_pptx now accepts committed import bundle directories as input, enabling the hand-edit sync workflow without the public pptx_to_json tool.
• Cloud file attachments use POST /attachments/presign + direct S3 PUT; Local Web UI uses POST /api/attachments. The [Attached:...] marker format is now [Attached:{"v":1,"name":"...","source":"..."}].

Fixed

• Cloud and local ACP deck agents now expose arch_diagram, as required by the composition workflow, so architecture, system, and flow diagrams use automatic routing and crossing minimization instead of silently falling back to manual placement.
• Attachment presigned PUTs now use Signature Version 4 — S3 rejects the conditional write (If-None-Match) with legacy SigV2 URLs, which broke all browser uploads in some regions (e.g. ap-northeast-1).
• The Web UI no longer sends a message when an attachment upload fails: input and attachments are kept for retry and the actual error (e.g. quota exceeded) is shown instead of a generic failure.
• Per-user raw attachment caps recalibrated for internal/team deployments (1000 objects / 50GB, overridable via ATTACHMENT_MAX_OBJECTS / ATTACHMENT_MAX_BYTES), and S3 lifecycle rules are prefix-only so attachment objects from pre-0.6 releases (which carry no sdpm-class tag) also expire. Deployments upgrading from the old upload pipeline should purge leftover uploads/ objects — they otherwise count toward the quota.
• servers/remote/constraints.txt regenerated via make lock; cachetools / protobuf are pinned by aws-opentelemetry-distro and are now excluded from Dependabot bumps until the distro itself is upgraded.

0.5.3 - 2026-08-01

Changed

• Kiro CLI: composer sub-agents are now self-spawned — make install-kiro no longer generates ~/.kiro/agents/sdpm-composer.json; the orchestrating agent spawns composer workers itself and pulls the composer behavior through start_presentation(mode="composer"). Upgrading from v0.5.2 or earlier, re-run make install-kiro once: it removes the legacy generated agent file (only if unmodified; a customized file is left in place with a warning).
• start_presentation now accepts mode="single" (one agent handles dialogue and composition end-to-end), making every persona reachable through the port.

0.5.2 - 2026-08-01

Changed

• run_python persistence semantics unified: file writes now always persist — the save flag is deprecated and ignored (silent data loss when omitting save=True on Cloud is no longer possible). The deck's PPTX artifact refreshes automatically whenever the deck changes; measure_slides remains the trigger for the expensive verification pass (render, text overflow measurement, previews). Cloud sandbox write-back is now diff-based (changed/new files only), preventing a stale sandbox copy from overwriting newer S3 writes

Fixed

• Cloud: superseded PPTX artifacts are now deleted after each refresh — the automatic artifact refresh no longer accumulates orphaned objects in S3 (update_deck returns previous values via UPDATED_OLD)

0.5.1 - 2026-07-31

Added

• make doctor — diagnoses local setup (uv / LibreOffice / poppler, checkout path anchors) with a moved-checkout hint
• make smoke — boots the local MCP server over real stdio and verifies template/persona resolution (also runs in CI)
• GitHub Releases are now created automatically on tag push (notes extracted from this changelog)

Fixed

• Cloud agent output-token limit: model profiles now set an explicit max_tokens (Claude 32768, others 8192) — Bedrock's small default truncated long single-call outputs (e.g. writing specs/brief.md from a long article) and killed the turn with a generic error. MaxTokensReachedException is also classified now (max_output) so the Web UI shows an actionable message instead of "something went wrong".

Changed

• L4 agent personas unified: the cloud agent (Strands) now fetches mode behavior from personas/*.md through the same start_presentation(mode=...) port as every other client, instead of carrying its own copies in agent/prompts/role/. Duplicated role/workflow prompt files were removed; only transport-specific wiring (attachment wire format, compose_slides report format) remains in agent/prompts/. Prompt changes now touch only personas/ for all layers.
• Internal API move: sdpm.engine.diff.diff_report / load_slides_json_or_pptx moved to sdpm.api (dependency-rule fix; engine.diff now exposes the pure diff_slides(base, edit)). These were internal APIs — update imports if you consumed them directly.

0.5.0 - 2026-07-31

Breaking architecture cleanup. No changes to the slide JSON schema — existing decks and cloud data keep working. See the migration guide for upgrade steps.

Changed (breaking)

• Directory layout: skill/ → sdpm/, mcp-local/ → servers/local/, mcp-server/ → servers/remote/, agents/ → clients/claude-code/agents/
• Engine split: the sdpm package is now organised as two peer subpackages — sdpm.engine (json ↔ pptx) and sdpm.knowledge (references / assets)
• Skill files removed: mode behavior now lives in personas/*.md and is served to any MCP client via the new start_presentation(mode=...) tool. Claude Code plugin no longer installs skills; Kiro installer no longer symlinks skill directories
• Single tool contract: every MCP tool is defined once in sdpm.tools; both servers bind the same functions (local: 24 tools, remote: 22 tools)
• Docs: English docs are canonical; Japanese docs reduced to README and getting-started

Added

• start_presentation(mode=...) MCP tool — serves vibe / spec / style / composer / single personas to any MCP client
• SDPM_SKILL_ROOT environment override for path anchors (used by the remote Docker image, gateway integrations)
• Migration guide: docs/en/migration-v0.5.md

Fixed

• Pinned mcp>=1.28.1,<2 across local server, remote server, and agent — mcp 2.0.0 removed mcp.server.fastmcp and crashed fresh container builds

0.4.0 - 2026-07-30

• Kiro CLI support: installer, composer agent, skill dispatch (#207)
• PPTX import: bring existing decks into the agent + edit flow, hand-edit sync via diff_pptx (#149, #215, #220)
• Template notes (built-in and local), template picker in Art Direction pane (#203, #204)
• User-local styles, assets, config, and template directories with cross-platform paths (#96, #99)
• Per-user model switching via Settings (#100)

0.3.x - 2026-05-12 .. 2026-06-02

• 0.3.0: composer web fetch, image aspect-ratio fit, SVG color fixes (#139, #146)
• 0.3.1–0.3.8: stability fixes (template upload in local mode, template analysis via uv run), workshop content, one-click deploy buttons (#167, #168, #170, #171)

0.2.x - 2026-05-01 .. 2026-05-11

• 0.2.0: agent separation, parallel slide generation, model config refactor (#71)
• 0.2.1: fontSize token discipline check, Python 3.14 compatibility fixes (#133, #136)

0.1.0 - 2026-05-01

• Initial release: spec-driven slide generation (Engine json ↔ pptx, CLI, local/remote MCP servers, Strands Agent, React Web UI, CDK stacks)