Change Notes

Settings → Backend & Safety now offers two text backends: Google AI (cloud — Google's safety filters and Prohibited Use Policy apply) and Local model (any OpenAI-compatible endpoint: Ollama, LM Studio). On the local tier Synthograsizer applies no content filters of its own — your hardware, your discretion, within the Terms §6 floor. Mixed-mode v1: images, video, and music always generate via Google. Includes endpoint test-connection + model picker (GET /api/backend/local/models), JSON-mode mapping with graceful fallback for servers that reject response_format, and SSE streaming. Hosted instances are pinned to Google and reject config changes.

Per-category selectors (Harassment / Hate speech / Sexually explicit / Dangerous content / Civic integrity) adjust Google's safety_settings within what its API permits — values pass through verbatim and Google's rejections surface honestly at generation time. Saved server-side; precedence: per-request > saved defaults > permissive baseline. Hosted instances clamp per-request settings to the operator's defaults — closing a pre-existing hole where any client could send BLOCK_NONE.

Safety blocks are now a typed error end-to-end (SafetyBlockedError → structured 422), so the error panel shows what was blocked and why, with a Report wrongly blocked button: prefilled GitHub issue form (wrongly_blocked.yml) and/or a local report to POST /api/feedback (JSONL on the operator's disk). Prompts are never auto-included — sharing one is an explicit checkbox. General "Send feedback" lives in the settings modal.

synth_fetch now validates URLs and every redirect hop against private/loopback/metadata addresses with a 10s timeout and 2 MB cap (workflow-engine/urlGuard.js); ANALYZE_URL gets a scheme/private-literal pre-check. First image upload shows a one-time consent notice (images go to Google's APIs — see Terms §7). Hosted mode (SYNTH_HOSTED=1): env-only API key, per-IP rate limiting (429 + Retry-After), and an hourly retention purge (default 30 days). None of this affects a normal local install.

The AI Studio Tools grid switched from a hard repeat(3, 1fr) to repeat(auto-fit, minmax(170px, 1fr)), so tiles reflow 3-up → 2-up → 1-up instead of squeezing. The app bar's controls are now grouped into wrap units (Mode / Links / Theme) that drop to a second row as whole groups rather than splitting mid-group, with a stacked-bar breakpoint at ≤1080px and tightened chrome below 600px. Verified clean at 1577 / 1024 / 768 / 375.

The app-bar mode container had flex-shrink: 0, pushing the Cel-Pastel button past the viewport (partially unclickable) on screens narrower than ~1620px. It now shrinks and wraps to a tidy right-aligned row; dividers that would dangle at a row edge are dropped once wrapping begins.

MIDI / Scope / Display / API were four full-width rows each hugging the right edge under a left-aligned header — lopsided and disorganized. They now pair into a wrapping grid (Display spans its row for the BG swatch), matching the treatment the Cel-Pastel theme already had.

The Hardware theme pins the center value to a fixed 38px with nowrap + ellipsis, so longer values like "razor-edge stillness" rendered as "razor-edge st…". The dormant resizeControlValueText() hook is reimplemented as a shrink-to-fit (down to a 14px floor, transition suspended during measurement) and now runs on every value change, not just window resize.

.btn-glitcher-large and .btn-workflow-large were missing from the Hardware theme's accent block (and Glitcher from Cel-Pastel's icon group), so they fell back to an un-themed grey. Glitcher gets a teal tint matching its brand accent; Workflows joins its Flow-category green. Output action buttons (Copy / Like / Code / Send to Chat) were re-colored from the label tint (~1.9:1) to full text contrast (~9.7:1) so they no longer read as disabled.

Every Gemini-facing prompt in backend/services/ gained an explicit "interpret the request" layer: terse user input is expanded into a specific, opinionated reading rather than a generic one, and explicit user constraints (named variables, counts, fixed phrases) are honored over defaults. Narrative generation now requires exactly one self-contained prompt per image in input order; image-variation requires exactly five; smart-transform handles the EMPTY-input and no-reference edge cases. The external-LLM authoring guides (SYSTEM_PROMPT.txt and friends) were de-duplicated against the canonical prompt and corrected — one shipped a stale flat-array schema.

The accreted styles/fixes/*.css patch cascade (where .effect-module was redefined 11× across files) was inlined into effect-studio.css in exact load order — byte-identical cascade, verified in-browser. The orphaned quick-fixes.css (referenced nowhere) was deleted.

Removed tmp/, scratch/, an empty presentation/ shell, a stray root node_modules/, two superseded ChatRoom build bundles, an orphan tweaks-panel.jsx, a dead generateAgentResponseSync export, and the untracked .design-handoff/ raw export (the curated design-handoff/ remains). REPO_MAP.md updated with each pass.

When you close the Agent Studio modal with at least one message in the transcript, the session is persisted to JSON/Agent Studio Logs/<agent-name>-<hash>.json with the same payload shape as the manual ⤓ Export button: goal, mode, agents, messages, workflow chips, artifacts, and composer context if launched from the Composer. Empty sessions (modal opened but no chat) are skipped to keep the folder clean. Content-addressed dedup means closing/reopening the same conversation lands on the same file.

When the Workflow Runner fires workflow_complete (success OR failure), the run is logged to JSON/Workflow Outputs/<workflow-name>-<hash>.json with the template name, run id, params used (image params stripped — they're huge base64 blobs), step results, status, and duration. Failed runs are saved too — useful for diagnostics. The run metadata is captured at _runWorkflow start time so the complete handler has everything it needs.

POST {kind, content, filename_hint?}. Files land at Synthograsizer_Output/JSON/<Kind>/<slug>-<6-char sha256>.json. Server-side dedup is content-addressed (id + timestamps stripped before hashing) so repeated saves of the same logical content land on the same file. Supported kinds: agent_profile, taste_profile, agent_log, story_output, workflow_output (last three reserved for follow-ups). Sibling endpoints: GET /api/list-outputs/{kind} for picker UIs (returns lightweight per-kind summary), GET /api/get-output/{kind}/{filename} for full content, DELETE /api/delete-output/{kind}/{filename}.

Every .save() now fires a best-effort POST to /api/save-output alongside the local localStorage / IndexedDB write. Failures are silent (backend may be offline). Net effect: on first load of the Composer, every built-in agent gets re-saved through the store and back-filled to disk — instant historical backup. The user no longer loses agents to a wiped browser cache.

Composer Library overflow menu (⋯) gets a "📁 Restore from disk…" item that lists every Agent Profiles/*.json with a one-click Restore button per row. The active-profile pill on the Synthograsizer app bar gets the same affordance for Taste Profiles/*.json. The pill now stays visible even when no profile is active (shows "No active profile") so Restore is always reachable.

All 4 tuned agents auto-placed in channels 1–4, session named "<profile> · First Session", goal pre-filled, three shared anchors (session_context, tone, output_rules) derived from the profile's tagline + axes + narrative inclinations. One click "Launch Session" and the crew is collaborating.

Mirrors AgentProfileStore's shape; full CRUD + subscribe. Lives in localStorage (lightweight at this scale). An artist can keep "Neon Sirens" and "Quiet Erosion" side-by-side and switch which one their agents read from.

resolveProfileBio() now injects synthetic placeholders — {{taste_tendency}}, {{taste_narrative}}, {{taste_subject}}, {{taste_profile_name}}, {{taste_tagline}}, {{taste_palette_primary/secondary/accent}}, {{taste_axes_summary}} — resolved at render time against whichever profile is active. Switch profiles, every agent re-skins instantly. Profile-defined variables with the same name still win (no surprise overrides).

Like openSessionFromRecipe but addresses agents by profileId (deterministic) and accepts sharedAnchors + sessionName in addition to goal. Fired from a one-shot sessionStorage handoff so the redirect from /taste-profile/ is seamless.

New mode: "p5_edit" endpoint takes the loaded sketch + an instruction and returns RAW JavaScript (no JSON envelope, no schema overhead). The frontend merges the result back, leaving promptTemplate and variables untouched. Surfaced as an inline "Edit scope" toggle in the Remix panel — appears only when the loaded template has a p5.js sketch, defaults to "Sketch only" (the common AV case). Should dramatically reduce timeouts on Pro for visual edits since the model no longer has to re-emit the entire template every time. Includes a REQUIRES_FULL_REMIX escape hatch: if the model decides the edit needs variable / prompt changes it prepends a comment and the frontend prompts the user to switch scope.

Persistent indicator showing which Taste Profile is currently active (with a green / cherry status dot). Clicking it opens a small menu listing all saved profiles, with switch / clear / "Manage on Taste Profile" actions. Subscribes to TasteProfileStore changes so it stays in sync if the active profile is updated in another tab. Hidden in Composer mode (where the Taste-Profile shortcut already gets hidden). Styled for both Hardware and Cel-Pastel themes.

Layered on top of the Hardware theme via body.theme-cel-pastel — Hardware remains the default and is unaffected when the picker is set to Hardware

Persists to localStorage['synthograsizerTheme']; each page gracefully normalizes the other's vocabulary ('dark' ↔ 'hardware') so navigation stays consistent

Covers the full Taste-Profile surface (steps 1–4): dropzone, pair cards, synth spinner, stat cards, aesthetic axes, palette swatches, tag chips, agent cards, JSON preview

Standard Synthograsizer page is unaffected at runtime; the AV hooks in app.js stay no-op unless window.SYNTH_AV_MODE is set (only by av.html)

A variable set to source: 'code' drives OSC only and is masked from the LLM prompt; set to 'ai' fills the prompt without firing OSC; 'both' does both

Send types: float, int, trigger. Curves: linear, exponential, logarithmic, custom. Sends only fire when the normalized value actually changes

Profiles persist via the new AVProfileStore; override edits apply to the running surface immediately without reload

Same template can have a performance-tuned variant (smaller knob set, OSC-wired) without forking the LLM template. Ships with AV variants for svg-flow-particles, lissajous-lab, synthograsizer-prompt

Pre-built OSC address templates targeting Resolume's standard composition routing

Lets the artist leave an axis unconstrained instead of forcing a false choice; the marker biases to 0.5 (range 0.4–0.6) in buildProfile() and the backend prompt was hardened to honor neutral answers

Reads tEXt, iTXt, and zTXt chunks. Knows the formats used by AUTOMATIC1111 (parameters), ComfyUI (workflow / prompt), NovelAI (Description), and generic Comment

Frame is rendered onto a canvas and resized like any still image; lets you train on a GIF without uploading the full animation

Clean up mistakes without restarting the upload step

Every agent now reliably ships bioTemplate + variables + anchors; the Send-to-Composer flow has a client-side fallback that synthesizes a minimum-viable template if any field is still missing

Saves all 4 agents, verifies each landed in the cache, then waits ~600 ms before redirecting so the Composer page's loadAllAsync() doesn't race the writes

Auto-save now derives a 1–3-word slug from the template name (or first source tag) + 6-character SHA-256 of the canonical content. No more timestamp spam in the output directory

If any existing file in the output directory has the same content fingerprint (even under a different name), the save returns status: exists pointing at the original instead of creating a copy

Replaces synchronous 5 MB-limited localStorage; profile saves/loads are now fully async, eliminating UI jank on large agent rosters with embedded assets

Existing profiles are transparently moved on first run; no manual export/import step required

Legacy call-sites using AgentProfileStore.loadAll() continue to work until they are updated to loadAllAsync()

resolveProfileBio() now re-rolls weighted-random attribute picks up to 3 times to avoid choosing the same value as the previous interaction, producing more varied and dynamic agent behaviour

Replaces fragile regex; gracefully handles markdown code fences, preamble/postamble prose, trailing commas, and single-quoted strings hallucinated by the model

The kill-switch would silently cancel complex long-running tasks (e.g. video generation) at exactly 5 minutes regardless of activity

Timer resets automatically on every workflow_step_start / workflow_step_chunk event; connections only close if the backend goes genuinely silent for 10 minutes

Pro model generating a full self-contained p5 sketch (lookup maps, 4–7 variables, draw loop) was being cut off by the shared 150s frontend abort; now 300s frontend, 300s backend, 300s workflow-engine path — all three layers aligned

Structural rules (JSON format, variables, anchors) always enforced; tone and style are now customizable per-user

Current profile JSON sent alongside the generation prompt so the model can make targeted changes

State persisted to localStorage; factory-default reset button included

Identity row and Bio template highlighted with gen-context-active state; taste vector bias automatically included when a vector is set

Attach up to 8 reference images to visually guide template evolution via multimodal prompt assembly

Single persistent top bar across all modes; layout-switcher promoted to app-bar-modes; hardware theme fully applied to Agent Studio

Each room is a discrete workflow step; knobs in Run room control live generation parameters in real time

Previously, muted agents still counted toward quorum threshold, producing incorrect vote outcomes

Full interactive demo with no API key required; rate-limited to lightweight models

Bidirectional bridge between Synthograsizer prompt output and Glitcher effect parameters

Cream chassis, warm brown borders, NeoBrutalism shadows, monospace labels throughout

Improved output coherence, longer context window, faster multimodal responses across all generation endpoints

Tokens appear as they arrive — no more waiting for full response before text renders

Multi-agent Codenames with roles (spymaster / operative); agents maintain secret layer inaccessible to opponents