v0.3.9

What's changed

• AI: persist generated images as workspace assets
• AI: apply preset image model, and accept every preset in SetConfig
• AI: generate a design as editable SVG (backend, DeepSeek/text LLM)
• AI: turn a generated SVG into an editable design (frontend)
• Backend: retry the initial DB connect on boot instead of failing once (#16)
• Docs: add the F28 leverage map and reference pointer
• AI: restore server-side assistant tool parity via a shared manifest
• AI: generated decks carry speaker notes
• AI: append generated pages by default instead of replacing the document
• AI: close the provider-config gaps
• Docs: sync F28/F39 feature docs to the shipped code
• Style: rename exported UPPER_SNAKE constants to camelCase
• Fix review findings: frontend test gaps, note-cap parity, azure edit capability
• Apply the full review's findings across the AI config and prompt seams
• AI: point the save error at the missing base URL
• Close the review's remaining AI-config and parity gaps
• AI config: PATCH semantics at the server and true capability gating
• AI: request natively schema-constrained output from the provider
• AI: repair-loop feedback with concrete errors in the orchestrator
• AI: port the prompt rule corpus into every generation prompt
• AI: review the outline before generating, with generation dials
• AI: placeholder-first generation imagery with reuse and stock routing
• Fix the Phase 1 review findings across the generation pipeline
• Schema v21: optional capacity hints on Placeholder
• AI: layout-grounded generation with derived content schemas
• AI: narrative ops - insertAgenda, splitSlide, insertComparison
• AI: per-slide regeneration with layout re-selection and image diffing
• Fix the Phase 2 review findings across layout-grounded generation
• Fix the follow-up review findings on the layout-generation seams
• AI: ingestion depth - multi-file grounding, office formats, scan detection
• AI: web search grounding with untrusted-source framing
• AI: chart values computed from attached data, never estimated
• AI: streaming generation - SSE progress and instant decks
• Fix the Phase 3 review findings across ingestion, search, and streaming
• AI: theme UI, brand-to-theme bridge, and AI theme generation (F28 T19)
• Editor: extract a reusable layout set from a deck (F28 T20 stage 1)
• Editor: vision correction and capacity verification for extracted layouts (F28 T20 stages 2+3)
• Docs: sync the F28 spec with whole-deck layout extraction (T20)
• Brand: draft a brand kit from a company domain (F28 T21)
• Docs: record the Phase 4 design-system features (T19, T20, T21)
• Fix the Phase 4 review findings across themes, extraction, and brand drafts
• Export: PPTX fidelity golden set + camera bubble in present-and-record (F28 T22)
• Fix the Phase 5 review findings across PPTX fidelity and recording
• Editor: make the brand re-skin identity-safe for neighboring undo entries
• Editor: extend restyle undo identity safety to the node level
• Docs: catch the F28 capability tables up to the shipped product
• Docs: add the F28 completion task board (7 phases, 38 tasks)
• Import: archive-bomb guards on the client zip reader (F28 completion C01)
• Present: per-transition easing and exit transitions (schema v22, F28 completion C02+C03)
• Editor: transition gallery picker and animate-all-slides (F28 completion C04+C05)
• Fix the Phase 1 review findings: exit-only transitions, key drops, swatch churn
• Engine: Magic Move depth - nested matching, appearance tween, forced match, word morph (F28 completion C06-C10)
• Fix the Phase 2 review findings: word-morph measurement parity and plan caching
• Engine: animation depth - motion paths, keyframe channels, spring physics, painter, media triggers (schema v23, F28 completion C11-C15)
• Fix the Phase 3 review findings: channel boundary holds and trigger fall-through
• Present: interactivity and delivery - interactions v2, captions, coach, kiosk, phone remote, control hand-off (schema v24, F28 completion C16-C22)
• Fix the Phase 4 review findings: pairing-code leak, speech loop, CRDT key hygiene
• Scale and interop: present prefetch, lazy thumbnails, ODP import, markdown outlines (F28 completion C23-C26)
• Fix the Phase 5 review findings: prefetch memory pin and honest scope notes
• Assist: accessibility fixes and AI completion - one-click a11y fixes, title sync, chart alt text, Magic Switch, brief interview, critique loop, citations, image themes (F28 completion C27-C34)
• Fix the Phase 6 review findings: citation integrity, review-outline clobber, CRDT name key, nested-node fixes
• Collab and sharing: slide status board, named link analytics, iframe embed, reuse slides (F28 completion C35-C38)
• Fix the Phase 7 review findings: background-image asset carry, picker thumbnail gating, board drag hygiene
• Docs: add the F40 AI platform task board (generation API, MCP, catalog, template-from-PPTX, adaptive reflow)
• API: public generation API - scoped keys, headless goja composer, prompt-to-deck endpoint, embedded docs (F40 E01-E06)
• Fix the Phase 1 (F40) review findings: header-only keys, job-poll tenancy, rune-safe truncation
• API: MCP server over the generation API + key-activity audit trail (F40 E07-E08)
• Polish the F40 Phase 2 review findings: share-refusal clarity and the Origin-validation stance
• Catalog: 16 slide layouts, 33 themes as shared data, 31 deck template seeds, generate-with-theme (F40 E09-E12)
• Fix the F40 Phase 3 review findings: hue-ladder duplicate, theme choice lost on regenerate, themed-compose coverage
• Templates: one-click AI template-from-PPTX + templates as first-class generation targets (F40 E13-E14)
• Fix the F40 Phase 4 review finding: template themes reach the freeform path
• AI Studio: adaptive reflow for layout-linked slides (F40 E15-E17)
• AI Studio: harden the reflow review findings (F40 E15-E17)
• Editor: one assistant bubble per intent in the AI panel
• Editor: keep the AI chat alive across tool switches and surface its state
• AI: tell the user WHY a provider call failed, and stop stale models
• AI Studio: give generated decks a real visual system
• Editor: let the AI panel collect several attachments in one sitting
• Editor: drag and drop files onto the AI panel to attach them
• Fix the review findings from the AI panel and deck-styling work
• Frontend: clear the standing lint warnings
• AI: add Moonshot (Kimi) as a first-class provider
• Docs: correct the provider counts for the Kimi preset
• Editor: make the AI base URL editable for every provider
• AI: carry the real failure reason through the generation stream
• Editor: stop the confirm banner from discarding the user's style choices
• Editor: give the generation job poll a realistic budget
• Editor: make one undo actually revert a generation, and never cross designs
• Fix two gaps in the AI trust fixes
• Editor: staged progress and a Stop button for AI generation
• Editor: surface failed slide copy, and stop image slots reading "Text"
• Editor: make the AI panel operable and announced
• Make Stop actually reach the long calls
• Editor: fold async AI results into the turn they continue, and keep focus
• Editor: give per-slide regeneration a button, and make the style dials stick
• Dashboard: start a design from a brief
• Fix three gaps in the discoverability work
• Dashboard: rebuild the brief field as a proper composer
• Dashboard: tune the brief composer's proportions
• Dashboard: centre the hero on the brief, and move the blank-start action
• Dashboard: attach documents to the brief
• Dashboard: drag documents onto the brief to attach them
• Editor: do not lose a dashboard brief when no AI provider is connected
• AI: give the provider config a home in workspace settings
• Fix two gaps in the AI settings extraction
• AI settings: label the fields and fit the settings page
• Dashboard: tell the two key sections apart, and lead with AI
• AI settings: drop the card and match the app's field styling
• AI settings: separate the web-search fields from the model fields
• AI settings: let the form use the full section width
• AI settings: move the save action to the end edge
• Fix the editor panel's copy of the AI form
• Editor: close the five remaining P1 findings from the AI audit
• Editor: make the follow-up chips deterministic, and survive a reload
• Editor: translate the assistant's skip reasons
• Editor: reach past AI conversations again
• Dashboard: expose the AI usage limits the server already enforces
• Fix three gaps in the history and policy work
• Editor: retry a failed generation with the review you already did
• Dashboard: say whether this workspace can actually generate
• AI settings: show that a key is stored, and lift the status to the heading
• AI: let a workspace disconnect its provider
• AI settings: rename Disconnect to Reset provider, beside Save
• UI: a real confirmation dialog, instead of the browser's
• AI: prove the provider works instead of assuming a saved key does
• Dashboard: present monthly AI token usage as the estimate it is
• AI settings: report the right cause, and stop claiming a key that is not there
• AI: let a workspace pair a text provider with a separate image provider
• AI settings: keep the image provider's stored model and host across a switch
• AI settings: say when an optional section could not be read
• AI: give the image provider's save rejections their own reasons
• AI settings: check the image provider's key, and cover the form with tests
• AI: apply the SSRF gate to redirects, not just the configured URL
• AI settings: give the image provider the same status treatment as the main one
• AI settings: stop writing an image config that was never asked for
• Assistant: refuse a turn that neither answers nor acts
• AI: tell a workspace its providers cannot read images, not that it has none
• AI: remove the backtracking blowup when stripping a code fence
• AI search: cap the result list before growing it, not after
• Backend: say what was unreachable when the DB connect fails (#18)
• Report a failed sign-in database query as 500, not 401 (#20)

Platform notes

• Each archive contains just the self-contained hycanvas binary with the frontend embedded; see the README for setup (.env, PostgreSQL). The first-run wizard generates the configuration for you, and .env.example in the repository documents every setting.
• First run: ./hycanvas service start daemonizes the server (pidfile + logfile next to the binary; Linux, macOS, and Windows) and asks whether to set up in the browser or in the terminal. The browser wizard at /installation/step-1 is unlocked by the one-time access secret it prints; the CLI wizard asks the same questions right there. Other verbs: stop, restart, status, log -f. ./hycanvas start runs in the foreground.
• The macOS binaries are not signed or notarized. If macOS quarantines the download, run xattr -d com.apple.quarantine hycanvas or right-click the binary and choose Open.
• ffmpeg is only required for server-side video export; everything else works without it.

What's Changed

• Retry the initial DB connect on boot instead of failing once by @devadattas in #16
• Say what was unreachable when the DB connect fails by @devadattas in #18
• Report a failed sign-in database query as 500, not 401 by @devadattas in #20

Full Changelog: v0.3.8...v0.3.9