Skip to content

Latest commit

 

History

History
752 lines (647 loc) · 44.9 KB

guide.md

File metadata and controls

752 lines (647 loc) · 44.9 KB

Intercom project guide

Visual home · Prototype details

Current status, 13 November 2026: the owner has paused Intercom after rejecting the delivered experience. The live .18 work path stopped on concurrent tool-permission requests; its outcome is unknown. The capability inventory below is not an end-to-end acceptance claim. See the README accountability statement and development journey.

For the owner's clarified voice-overlay vision, what currently works and the missing integrated experience, read the illustrated vision/reality report. It records the pre-.16 gaps; the integrated local candidate below supersedes its source-versus-executable inventory, not its human-acceptance caveats.

Intercom microphone and context-mesh logo with frosted sea-glass nodes and connections

The microphone represents voice; varied-size connected nodes, rounded outer orbits and straight inner links represent the context mesh. Foreground links and a node overlap the microphone to give the mark depth.

A voice-first desktop assistant for working with existing Agency/Copilot sessions. Speak a request, choose terminal handoff or background execution, and hear the result when the requested work finishes.

Its proposed visual surface is a 3D Context Mesh sphere: active context and session paths light up while other relationships remain dimly visible.

Top priority: speed/performance by default, with selectable processing priorities. The prototype saves the selection and bounds graph rendering; background requests apply supported reasoning effort (low/medium/high) without changing the existing model. Speaking-voice settings are separate from processing priorities; measured speech quality and speed budgets remain unqualified.

Stage: Windows push-to-talk prototype with a local .18 candidate and included on-device recognition, system microphone and local voice defaults. Click Speak / Stop & send, or use the no-send speech check first. The restricted Intercom controller interprets text and clear plans target existing idle sessions; ambiguity still needs clarification, real tools retain permissions and work returns durable receipts. Actual multi-session and human multilingual qualification remain open. Every launch reads the actual local Context Mesh snapshot. Terminal mode opens the existing session for manual task entry, not automatic handoff. No wake word, active-session attach or new sessions. Electron is approved for the prototype, not selected as the final product host.

Engineering recommendation: Tauri 2 + Rust host, React/TypeScript + Three.js UI, native local-speech worker and SQLite jobs. This revises the initial Electron recommendation; hardware measurements and Agency/speech proof still gate selection.

Run the prototype

Included in local .18: recordings support up to five minutes. Audio stays local and in memory. At the limit, capture stops automatically and the final transcript waits for review, never automatic dispatch. Stop earlier to retain the selected automatic-send behavior. Long transcripts stay intact, but requests still need to fit 10,000 characters for Intercom or 1,000 for legacy/manual delivery. No automatic truncation or splitting. Published .12 still has its historical one-minute capture limit. See qualification and limits.

The latest local-only, unsigned Windows x64 candidate is .18: C:\devdesk\homedesk\Intercom\release\prototype-18-candidate\Intercom-Prototype-0.1.0-prototype.18-win-x64.exe. It includes warm native speech, five-minute capture, observed source-read activity, cited research synthesis, approved glass/water visuals, isolated portable extraction, persistent privacy-safe failure diagnostics and bounded automatic clarification replies. It is not a published GitHub release or a claim of human speech accuracy.

Open it and click the microphone: fresh profiles automatically use the included on-device Whisper model/Vulkan engine, system-default microphone, automatic language detection, Standard local Windows voice, speed-first processing and Intercom routing. No recognition-folder selection, microphone discovery or session-ID entry is required. Audio stays local. Initial capture starts on click; a structured Intercom clarification can automatically open the next bounded capture after the question finishes speaking. Tool permissions remain manual, and installed/authenticated Copilot plus its existing idle sessions remain environmental prerequisites for actual agent work.

Click the mic to finish your clarification reply; silence does not stop it automatically. Muted questions remain written and can also start reply capture. Advanced > General > Delivery behavior > Automatically listen for clarification replies disables this default-on behavior. Stop playback, Cancel, opt-out and hidden/minimized windows suppress pending opening. Ordinary replies and tool-permission prompts never trigger it. The previous unresolved request/question accompanies only the next accepted controller turn; it does not approve tools. See exact evidence and limits.

For a no-send trial, use Input > Speech check > Test speech locally; stop with the bubble. This check does not replace your draft. Existing explicit legacy/custom recognition and microphone choices remain unchanged. The included native-optimized Vulkan build is qualified on this development machine only, not as a generic Windows/GPU distribution.

For optional previews while recording, enable Talk > Live transcription (local draft). It starts off on each mount to preserve the speed-first default. These are serialized local batches using the same resident model, not word-by-word streaming; each completed preview is followed by a three-second wait. Previews never send requests. Stopping cancels/settles preview work before final recognition; cancelling discards previews and preserves your typed draft. This option uses more CPU/GPU.

The latest published prototype remains the unsigned .12: Download the .12 EXE from the prototype prerelease. Only the local .18 executable and its checksum remain in release, following owner-requested cleanup after exact portable verification on 2026-11-13. Superseded .17 and candidate unpacking/build intermediates were removed only after a fresh check found no running Intercom process; published GitHub releases were not changed.

The duplicate-launch regression verifies that the primary model and resident helper survive the rejected second launcher, then transcribes again. Missing resources still refuse capture explicitly. Open Advanced > Display > Local diagnostics for the persistent directory and Open diagnostics folder. The default is C:\Users\mishrad\AppData\Roaming\IntercomVoicePrototype\diagnostics. Logs omit speech/request content; automatic failure snapshots contain process metadata and resource presence, not memory contents. Storage failures appear visibly. See retention and recovery. Close an older build before launching .18; no running app is switched for you.

Research now shows actual source checks, successful reads and failures in a bounded activity trail; successful reads trigger finite mesh pulses. This is observed host activity, not hidden model reasoning. Ordinary collection is read-only and opens an extractive report. For a model-written synthesis, explicitly select an existing idle session in the report controls, or clearly name that exact session in the request. The normal durable queue, workspace exclusions and manual tool approvals apply. Validated citations link retained evidence; they do not prove semantic correctness. See the current .17 receipt.

Prototype .12 includes the microphone fixes since .7: session-independent local capture/transcription, a stationary microphone in both views, a collapsed single-row floating dock, accessible settings/transcripts and explicit review of uncertain/long recognition. Hover or pin mesh nodes to identify them without permanent labels. It also packages the reorganized repository and retains named installed-voice profiles, pitch/speed/volume, deep graph zoom, maximize/restore and confirmed existing-session background work.

The live source acceptance path captured synthetic speech, executed the exact approved read-only Get-Location in the existing session, returned "Hello world." and completed local playback. Existing saved grants meant no native permission prompt appeared; approval mediation is tested with fixtures, not live-qualified. The packaged candidate is qualified separately from that historical live task. Physical FDUCE and human voice-quality acceptance remain unverified. Checksums, the feature matrix and exact limits are in the prototype guide.

Windows may show an unknown-publisher/SmartScreen warning; verify the release checksum and follow your device's security policy. No Node.js installation is needed to run the EXE. First launch extracts the bundled Electron runtime; startup speed has not been benchmarked.

The earlier fixture-only .1 prerelease is preserved unchanged. The intermediate .2-.6 builds are historical qualification records; their features are included in .12.

From C:\devdesk\homedesk\Intercom in PowerShell, with Node.js/npm installed:

npm ci
npm start

Current source: focused glass controls

Current source defaults to the Intercom node, not a designated work session. Its restricted controller receives transcript text and minimal catalog metadata; the host validates and durably queues clear multi-session plans without a second target/task confirmation. Ambiguity still needs clarification and real tools retain exact one-shot approvals. Background work does not occupy the mic. Work > Plans offers request/action selectors and Activity / Permission / Result tabs. Full details remain accessible inside bounded text areas. An ordinary agent response is not independently verified task success.

The host enforces dependencies, excluded status targets and at most two non-overlapping workspaces; recovery never automatically retries an uncertain dispatch. Current checks pass 62 host cases, 19 renderer/recording cases and 11 controller cases. The full desktop suite passes 54 cases; the .17 unpacked package also passes nine focused cases. Actual portable synthetic capture passes native local live-draft and final-recognition checks, including X/Y exclusion preservation, same native PID reuse, shutdown and no automatic send. The voice suite is 51/52: legacy Windows confuses synthetic "please" with "police"; the strict assertion is retained. These include exact permission routing and cancelled interpretation. Fixtures are not real multi-session execution evidence. Local Whisper recognition is now connected as described below. The older portable dependency candidate predates this wiring; local .17 includes it, and published .12 remains unchanged. See runtime contracts for the exact boundary.

Saved-context research

In local .17, ask for a project/context research report from Talk. The controller receives the request and minimal mesh/catalog metadata, not collected file contents. A distinct read-only host lane reads the four fixed registered context notes and, when requested, exact indexed session histories. It saves an illustrated, cited HTML report and JSON evidence under the app profile's research directory and asks the OS to open the HTML default handler. The receipt shows its full path; Open context report retries an opening failure. Catalog notices stay collapsed and do not imply a context-read failure.

The report distinguishes project-name matches, topic references, recorded requests, assistant statements, saved topology and unavailable sources. Extractive summaries are not model-written conclusions. This is not a census of all sessions ever or proof of live access. Ordinary collection sends no work. Actual successful source reads produce finite mesh pulses and an attributable trail. Explicit exact idle-session selection can request cited Copilot synthesis through the normal durable work queue and manual permissions; source excerpts then enter that session's provider context. Location validation is not proof that a model's conclusion follows from the source. Large histories stream within explicit limits; omissions and parser gaps are visible. Generated reports remain private, outside publication artifacts. This feature is included in local .17, not in published .12. See the research qualification and runtime contract.

Included speech defaults and optional custom engines

Packaged .17 resolves its included model and seven pinned native binaries from Electron's external resources\speech directory. Packaging verifies the exact hashes and retains whisper.cpp/OpenAI/miniaudio license notices. A private-pipe helper loads the model before capture and reuses it across drafts and final recognition. Cancellation shuts it down; the next operation reloads. Idle model RAM is not yet measured. Fresh profiles need no settings file. Saving a language hint records a symbolic bundle source, not a temporary portable extraction path; the current resource path resolves on each launch. Missing/corrupt resources refuse recording before opening a microphone. No Windows/CPU fallback, download service or automatic permission grant exists.

For a custom engine, open Input > Recognition > Choose local Whisper engine and select a prepared directory containing ggml-large-v3-turbo-q5_0.bin and a runtime subdirectory. Only the exact pinned CPU/BLAS or development Vulkan file set is accepted. The host checks hashes before saving the selection and again before recording, with an unchanged-artifact cache. Configuration survives restart. This optional override is a prepared-folder picker, not an automatic downloader or installer. Source runs and historical .13 do not have the packaged bundle. See the qualification record for the current artifacts and limitations.

Source profiles without a bundle require setup. Use legacy English recognition explicitly selects the installed Windows engine instead; there is no silent fallback. Missing or corrupt local artifacts refuse capture before opening a microphone and preserve any unsent draft. Changing/cancelling setup never sends a request. Selected Vulkan must report actual Vulkan execution or transcription fails.

Automatic detection is the included default; optionally choose English or Hindi in Recognition. The hint is saved across restarts; it is not translation or a promise of recognition accuracy.

Default Intercom requests speak checking progress and conversational replies using the local voice. Work announcements include newly settled live plans, defer while recording and retain truthful per-action outcomes. Restored receipts are not replayed; a normal Copilot response is not proof of task success.

Input > Microphone > Test selected measures two seconds of the selected input, reports peak/RMS and browser-reported source rate separately from the recognition PCM rate, and rejects silence or substantial clipping. These are browser/media measurements, not independently measured hardware resolution. Input > Speech check > Test speech locally records and transcribes without sending any request or replacing your working draft. Stop with the bubble; cancel discards late results. The separate result shows local text, capture format, duration, recognition time and detected language. Audio is not saved. Hindi may appear in its original script here; Roman cleanup happens during Intercom interpretation, not this raw diagnostic.

Synthetic native execution, rate conversion, settings persistence and exact owned-process cancellation work. That does not establish speech quality: the 8 kHz test confused "Y" with "why" despite high token probability. Current source forces review when either WAV or browser-reported source rate is below 16 kHz, even if capture upsamples it. That blocks automatic sending; upsampling does not restore lost detail or fix the name-recognition failure. Human English/Hindi/Hinglish and FDUCE acceptance, speed budgets and a portable GPU distribution remain open. Roman normalization/filler cleanup occurs in the restricted Intercom interpreter; legacy destinations do not use that cleanup.

The current source UI differs from the published .12 download below. The bottom of the workspace contains only a glass microphone icon. Its tooltip and accessible name describe Speak, Stop & transcribe / send, or Cancel request; recording level fills the button. Recording does not require a connected session. The earlier subtle pearlescent glass shell is restored around the smooth lavender/mint microphone; the opaque rainbow texture is no longer used. Six soft reflection patches turn in CSS 3D beneath fixed edge shading and lighting. A seeded, 32-second periodic trajectory changes direction on X, Y and Z without random-target snaps. One Web Animation interpolates dense 50 ms orientation samples; the microphone stays upright. The internal golden loop and narrow upper reflection stroke remain removed. This is a lightweight optical approximation, not physical ray tracing or live environment capture. Motion pauses when the document is hidden or the native window reports paused, and reduced motion removes it. Gradients and shadows are not animated. The owner rejected both whole-shell expansion and drawn ring outlines. During capture, existing measured input instead disturbs a local 64x64 water-height/velocity field at the centre. A finite-difference wave update propagates the disturbance into a damped rim; stronger input injects stronger disturbances, no faster than once per 240 ms. Surface slopes become red/green sampling offsets in a local SVG displacement map that refracts the existing glass reflections. The same normals produce soft directional light and shade on a clipped transparent canvas. There are no stroked circles. The microphone glyph, rim, glass size and hit area remain undistorted; the capsule still fills vertically. The owner's jQuery-ripples reference uses height updates and normal-based image sampling too. Its source was inspected, not imported: this implementation adds no jQuery, dependency, WebGL scene, external request or audio stream. Water painting is capped at 30 fps, with fixed 90 Hz numerical steps and at most 100 ms of catch-up. Silence stops disturbances, fades the residual waves and removes the filter/loop after 1.6 seconds. Hidden/native pause, reduced motion, cancellation and unmount clear both water buffers and the filter. The separate glass Web Animation is never restarted, steered or paused by input levels. The button has no backdrop blur; 60 px workspace and 34 px floating targets are unchanged. These are rendering-cost safeguards, not a measured hardware performance claim. During capture, original mint/cyan ribbons flow outward symmetrically from the microphone. Their height and the microphone capsule's vertical fill follow the same measured input level, not random animation. Floating mode keeps smaller wings around its microphone. Painting is capped at 30 fps, stops when hidden/ minimized, and is removed when capture stops; reduced motion uses a stationary shape that still reflects the level. No extra microphone stream is opened.

The right panel opens on Talk. Each category keeps its controls in view:

Tab Controls
Talk Transcript, explicit send, response and recording cancellation
Input Microphone: selection/load/measured test; Recognition: engine/language; Speech check: local-only transcription
Session Intercom orchestration destination; legacy chat/session connection choices
Voice Local playback; Voice / Tune / Profiles / Preview sub-tabs
Work Plans with Activity / Permission / Result; legacy Target / Task / Review / Receipt
Advanced General: automatic sending and processing priority; Display: drift; About: limits

Use Left/Right or Home/End on a tab to switch by keyboard. Draft text, session entry and unsaved voice edits survive category changes. The single recorder stays mounted across tabs and workspace/floating transitions.

In Floating Mesh, a small microphone sits immediately beside the native drag grip, outside the default-collapsed controls. Expand the arrow to show the fitted grid of graph/window controls, settings, conversation and work/receipts. The grid does not scroll, including at the 360 px minimum size. Settings and work return to the workspace; conversation toggles the floating transcript. Recognition review and errors open that transcript automatically. A pending permission highlights the dock arrow and work control; it never grants approval.

Settings categories fit the 960 x 680 minimum workspace without scrolling. Long transcripts, work confirmations and receipts may still scroll inside their content panels; no setting is hidden to fake a fit. The published .12 binary, release assets and any already-running instance retain their earlier layout.

Shared graph and window controls

Use the top-right Maximize window square to fill the screen; it becomes Restore window. Scroll over the graph to zoom toward the pointer, use + / - for centered zoom, or right-drag to pan. Zoom can enter the sphere; Reset view restores the starting camera and graph orientation.

Node identification in .12: hover a colored node for its name, or click/tap to pin its details and base path. Names stay hidden otherwise. Close with Escape, the close button, a same-node click or blank-space click. Keyboard: focus the graph, arrow through nodes, Enter/Space to pin. Inspecting a node does not change the work target.

Published .12 control layout

Click Floating mesh to hide the panels and window background. Prototype .10 and newer, including .12: the bottom dock starts collapsed. Its Drag to move grip moves the native window; the adjacent arrow expands one horizontal row containing graph zoom/reset, restore, pin, close and window-size controls. At small sizes the row scrolls horizontally rather than wrapping. The size buttons resize the native window without changing graph zoom. There is no separate top grip or hovering zoom toolbar in this view.

Prototype .11 and newer, including .12: the large Speak button is fixed at the bottom center of the whole window, in its own operational bar below the graph and scrolling panels. It never shifts when recording or cancellation controls appear. In floating mode it stays centered above, and outside, the collapsed dock. The input meter appears while recording; x cancels the request. It uses the same selected input and recording in workspace/floating mode: switching views does not stop or reopen the stream. Settings beside the mic opens/focuses microphone selection in the workspace, or toggles the scrollable floating voice panel. The transcript is first; microphone/session setup and playback profiles start collapsed. Stopping opens recognition review/results or an explicit error in floating mode. Preparing a work request returns to the workspace for the existing separate target/task confirmation; it does not dispatch work. The old inert live-mode footer is not used; the new bar contains the actual recorder. Explicit fixture demo remains synthetic and separate from live capture.

Escape dismisses an open voice panel or node label before restoring the workspace; Ctrl+Shift+M toggles views while Intercom has focus. Returning from a maximized workspace retains its previous normal-window bounds/maximization. Older .7-.9 builds retain the top drag handle and two-row dock; they do not update automatically. See the current prototype details.

Prototype .8 and newer, including .12: click Settings beside Speak first, then Load microphones, select FDUCE (or your preferred input), optionally Test selected, then Speak. The meter shows the selected input while recording. No Copilot connection is needed for recording or local transcription. Click Stop & transcribe to retain an editable transcript; without a connection, nothing is sent, even if automatic sending is checked. Load/Test only checks access, not capture or recognition. The prototype guide identifies the fixed .12 EXE; the older .7 download is unchanged and still requires connection before Speak.

To send, paste your existing dedicated session UUID and click Connect existing session, then Send text to Copilot. Connecting alone never sends an existing transcript. Its terminal must be closed/idle; Intercom refuses a second writer. Agency and Copilot must be installed, on PATH and authenticated. No session UUID or credentials ship in the EXE. Subsequent recordings with a connected session and automatic sending enabled use Stop & send and read the returned answer. Disable automatic sending to review/edit before sending. Typed fallback remains available.

Prototype .9 and newer, including .12: low-confidence recognition preserves the transcript for review instead of discarding it. Nothing is automatically sent or proposed in any destination; review/edit, then explicitly send or prepare work. Transcripts over 1,000 characters remain intact but must be shortened before sending. There is no automatic truncation or splitting. Empty recognition still gives a distinct error. This improves recovery, not recognition accuracy; the engine remains Windows English (US), regardless of the speaking-voice profile. See the current prototype details.

English (US) Windows speech recognition and an installed English voice are required; nothing is downloaded. Audio stays in memory and is not uploaded. Text goes to the existing Copilot account/session and may remain in its history. Use Show fixture demo for the older synthetic routing controls.

Expand Voice profiles and playback, then Voice profiles / customize speech, without connecting a session. Choose an installed voice, adjust pitch/speed/volume, and Preview voice. Preview does not save edits. Save & use profile keeps and activates them; Duplicate profile creates another named configuration. Profiles survive relaunch and apply to progress, replies and work updates. Gender is installed voice metadata, not a transformation; pitch depends on the voice, and speed uses Windows' scale. The required future arbitrary identity/timbre/emotional style engine has a separate provider boundary but is not implemented here. See the speech contract.

Installed voice profiles in the packaged desktop

Packaged .7: saved speaking profile with installed voice, pitch, speed and volume. The sidebar scrolls to preview/save actions below. Contexts are a synthetic test snapshot loaded through the native mesh reader, not the owner's private mesh.

For real work, choose Background task as the voice destination, then review the proposed text below. Select the target context and an existing idle session, click Review target and task, then confirm the exact immutable proposal. Approve or reject each surfaced operation once. Review the returned response and receipt; a response is not independent verification of arbitrary work. One background job is supported; voice proposals and floating controls remain available while it runs. The separate terminal destination needs manual task entry and has no monitored completion.

The mesh reads the published local snapshot on every launch; it does not rebuild Context Mesh itself. Source/load timestamps and stale warnings are visible. Missing or malformed data produces an error, never a silent fixture substitute. Bright spokes indicate observed job/tool locations, not inferred reads of links.

See the prototype guide for controls, architecture, performance behavior, checks and limitations.

Intercom desktop workspace with graph zoom and window controls

Actual packaged v0.1.0-prototype.5 capture: maximize/restore is beside minimize and close; graph zoom/reset controls sit beside the sphere. This image uses explicit synthetic demo contexts; normal startup reads the actual local mesh.

Floating Mesh with visible movement handle and window-size controls

Actual packaged .6 fixture capture: grab Drag to move above the sphere; use Smaller / Larger in the always-visible bottom dock. The window is 540 px wide while the separate graph zoom is 125%. Surrounding pixels are transparent; synthetic contexts do not represent actual context reads.

Start here

The repository uses numbered, evidence-backed workstreams. The owner authorized the initial fixture sub-slice, local-English conversation and subsequently confirmed real work with manual approvals and a fresh actual mesh each launch. Multilingual/hardware/performance and broader orchestration remain open.

Development boundaries

Repository layout

The root keeps the package/build configuration, browser entry, contributor instructions and numbered workstreams. Implementation and evidence are grouped by responsibility rather than mixed into those entry points:

Location Contents
src/app Application shell and shared styles
src/components Mesh, microphone, conversation, voice-profile and work UI
src/domain Context, processing-profile and routing model
src/voice Selected-input recording and WAV encoding
desktop/main.cjs, desktop/preload.cjs Electron entry and isolated renderer bridge
desktop/services Mesh worker, voice, synthesis, ACP and work hosts
desktop/speech Local Windows speech script and companion text
public/audio, public/branding Runtime worklet and application icons, copied into dist
tests/unit, tests/e2e, tests/smoke, tests/helpers Unit suites, desktop fixtures, opt-in smoke scripts and synthetic audio
tools/docs README artwork generator
docs/product, docs/architecture, docs/research Requirements, design/contracts and qualification research
docs/prototype, docs/development Prototype evidence, contributor workflow and tracking policy
docs/assets/branding, docs/assets/readme, docs/assets/screenshots Original identity, generated showcase and source captures
TODO Stable numbered workstreams, index and development ledger

dist, .local, release and .agent-context remain ignored generated, qualification, delivery and private-continuity locations respectively. The folder reorganization does not migrate profiles or replace a running release.

Run the existing npm commands from the repository root. npm test discovers the renderer/domain unit suites; npm run test:voice and npm run test:host cover native services. npm run test:desktop uses disposable profiles and synthetic audio. Live smoke scripts are not part of those suites and still require fresh explicit consent before contacting a real Copilot session.

The Test workflow runs the same lint, build, renderer, native voice/host and Electron desktop commands on Windows Server 2025 with Node 26 for main pushes, pull requests and manual dispatches. It uses npm ci, one desktop worker and the existing synthetic/disposable fixtures. Only synthetic desktop failure outputs are uploaded, retained for three days. It does not run live smoke scripts, real sessions, physical microphones, packaging, releases or deployment. Hosted results are separate from human speech quality, hardware acceptance and the unfinished local rebuild.

After a successful build, independent native/host/desktop suites still run if an earlier test suite fails; cancellation stops them. Failures are not ignored. The first hosted run found a real recognition mismatch (Please became police), despite passing local synthetic checks; its speech assertion is retained rather than weakening the suite to obtain a green badge.

For isolated packaging while a retained release is in use, choose a separate ignored output rather than replacing it:

npm run package:win -- --config.directories.output=.local\reorganization-package

Runtime boundaries

Local agent execution does not guarantee offline model processing. Speech privacy, permissions, retention, cancellation, and confirmation retain explicit boundaries. Only own-main-frame audio is permitted; camera/other browser permissions and renderer network requests remain denied. Narrow host IPC controls speech, sanitized mesh loading and confirmed existing-session work; no generic renderer shell/filesystem bridge exists. The single atomic job receipt is not the researched SQLite queue/outbox. Continuous listening, cloud audio, active-session attachment and new services remain outside this prototype.

Private session continuity lives in the Git-ignored .agent-context directory; it is not part of public documentation or release packages.

README artwork

The README's single inspiration paragraph pairs a mid-century switchboard operator (listening and connecting existing sessions) with an opera-orchestra conductor (coordinating independent work and returning results). It describes product intent, not completed speech/orchestration qualification; no UI redesign is part of this documentation change. The photo floats right at 52% width using GitHub-compatible HTML, without a bordered layout table. Text wraps beside it; the right-aligned credit follows a clearing break so the gallery starts below the photograph. The accompanying unchanged photograph is docs/assets/readme/kdd-telephone-exchange-1967.jpg, supplied by the owner. Wikimedia Commons provenance credits Asahi Shimbun and a 31 March 1967 source. It is not a 1950s photograph. Commons notes Japanese public-domain status and a possible U.S. copyright restriction; the owner explicitly confirmed permission to publish this photo. That permission is not a blanket public-domain or repository-license claim.

The animated hero, reduced-motion still and static documentation mark use frosted sea glass: muted seafoam, milky diffuse shading, fine fixed grain and soft rim highlights rather than polished neon surfaces. The title, raised link buttons and gallery framing share that palette with dark readable text. The lowercase intercom. wordmark retains its original Segoe UI Bold typeface with grainy seafoam ink and soft violet pigment washes. The subtitle and navigation remain in legible Segoe UI. The animated and static documentation microphones use one smooth vertical violet-to-seafoam gradient, without the wordmark's streaked pigment washes. Both logo microphones use a brighter lavender-to-mint gradient. The animated fill rises and falls with the illustrative voice level. Its unfilled capsule is transparent, not painted with a second material; the visible fill retains one smooth gradient and the rim also breathes. The static mark retains its original silhouette, foreground node and crossing links. Runtime/EXE icons are not changed. GitHub controls the page itself; these are rendered images, not a custom CSS theme or a backdrop filter. Original screenshots are not recolored or retouched.

The visual README uses an original, looping 3D-projected mesh around a microphone with a smoothly blended violet/seafoam body. Its 432 x 432 px central sphere is three times the previous 144 px logo. The transparent 936 x 432 px canvas adds mirrored cyan, mint, lavender, rose and amber wave ribbons behind the center. Layered rolling harmonics, breathing strand spacing and a broad, softly blurred colored glow give the wings a flowing, ethereal feel. Colors drift through the ribbons in a seamless cycle; their energy follows the same illustrative voice envelope without overpowering the frosted mesh. The voice level is simulated: a GitHub README cannot access your microphone, read your context activity, or run the desktop application. The animation does not change Intercom's runtime logo or microphone behavior. The logo has no embedded caption or rectangular backdrop: only the circular center glow remains, with transparent space between and around the wave lines. There is no background panel behind the wings. The GIF uses palette transparency and clears each frame to avoid trails; the still PNG keeps its soft alpha edge. Linked button artwork uses raised faces and soft shadows, not executable controls or CSS hover effects.

The behavioral reference is Microsoft's New microphone volume indicator in Teams (Anqi Chen, August 14, 2025): mic levels move up/down with incoming speech. The Intercom artwork is original, not a copied Teams icon.

The README offers a still image when the renderer supports the reduced-motion media query. You can also open the still directly. All four README GUI panels are six-second recordings of an earlier local source renderer: voice/tuning tabs, mesh orbit/zoom, the floating dock, and input-reactive microphone waves. They still show the rejected rotating rainbow film, not the latest restored glass/3D motion; their next refresh is held for visual acceptance of the local preview. Each has a reduced-motion PNG and a direct still link. Historical screenshots remain evidence in this guide; they are not substituted into the new recordings.

The standalone enlarged bubble showcase has been removed from the README. A separate 560 x 560 px GIF remains at assets/readme/readme-bubble.gif, with its matching PNG still, for inspecting the actual idle control. That documentation candidate still shows the superseded rainbow shell. The latest local preview is .local/readme-capture/readme-bubble.gif: 560 x 560 px, 640 frames, 32 seconds at 20 fps, sampled deterministically from the actual control's Web Animation timeline. It is not live audio or a performance measurement. A 3x screenshot of the 84 px crop provides a 252 px source before enlargement. It has its own reduced-motion still and caption. No reference photograph is copied into the artwork. The original logo GIF, wordmark and button artwork are unchanged by this refresh.

The captures use an isolated build, disposable profile, 48 fictional contexts and 144 structural connections for a denser globe. This density applies only to the application demos; the logo keeps its original nine-node geometry. The captures use no real catalog, only an enveloped synthetic tone through Chromium's fake audio device. Recognition readiness is stubbed for recording the UI; no recognition, speech playback or Agency dispatch runs. The copied documentation host stays hidden/offscreen, disables background throttling and keeps rendering unpaused; the renderer/preload are unchanged. No actual user instance, private mesh, physical microphone, desktop or other application is recorded. Floating frames retain genuine native-window transparency and are enlarged over the owner-supplied sea-glass wallpaper (assets/readme/sea-glass-wallpaper.png), not presented as an on-desktop recording. Fictional node colors use saturated coral, amber, plum and indigo to contrast with the cool sea-glass palette. The caption identifies the wallpaper and compositing. These are source-build demonstrations, not a new release, proof of native overlay behavior on a human desktop, or speech-quality acceptance.

Reproduce the documentation assets with Python and Pillow (rendered with Pillow 12.2.0 and Windows Segoe UI; these are artwork tools, not application dependencies or bundled fonts):

python .\tools\docs\render_readme.py

Run from the repository root. This command regenerates branding and link buttons, leaving the separately captured GUI assets untouched. The branding generator uses no audio, network, session connector or live app. Its fixed cyclic rotation and illustrative speech envelope produce a 144-frame, 20 fps, 7.2-second loop with a shared sea-glass palette with dedicated multicolor wave ramps. Fixed grain seeds keep the material reproducible without random frame-to-frame flicker. The original high-resolution static mark is retained in tools/docs/sources/intercom-logo.png; the generator changes its material, not its geometry or alpha silhouette. Runtime/EXE icons remain unchanged.

Re-record the GUI gallery using the installed development dependencies:

npm run build -- --outDir .local\readme-capture\app\dist
node .\tools\docs\capture_readme.cjs

Refresh the approved speaking showcase directly in the README assets:

npm run build -- --outDir .local\water-refraction-capture\app\dist
node .\tools\docs\capture_readme.cjs --readme-speaking

This writes the speaking GIF and matching reduced-motion still under docs\assets\readme. The private modes below leave those assets unchanged. The full GUI command also creates a private idle bubble pair; only the selected speaking pair is embedded in the README.

Capture only the current bubble preview, without replacing README media:

npm run build -- --outDir .local\readme-capture\app\dist
node .\tools\docs\capture_readme.cjs --bubble-only

This writes the preview GIF and still under .local\readme-capture. Capture the level-reactive bubble using locally synthesized Windows speech:

npm run build -- --outDir .local\water-refraction-capture\app\dist
node .\tools\docs\capture_readme.cjs --bubble-speaking

This writes intercom-speaking-bubble.gif and its PNG under .local\water-refraction-capture, preserving the rejected outlined-ring and expansion previews, the idle preview and README media. The 14-second, 280-frame example captures the actual recording/RMS pipeline with synthetic speech and pauses, not a physical microphone or injected UI level. The GIF is silent. The hidden disposable app checks measured input, active surface displacement, fixed shell size, final silence and released tracks; transcription, playback and work dispatch stay blocked. Desktop checks separately verify actual filtered/unfiltered pixel differences, sustained input, unchanged rotation identity/advancing time and lifecycle cleanup. Numerical tests cover outward propagation, signed slopes, input strength, damping and transparent lighting at rest. Six decoded GIF phases were inspected, not complete playback. The rejected film artwork generator has been removed; the current glass does not load a texture asset.

The capture tool launches only its copied, hidden documentation host, refuses changed host-patching assumptions, blocks transcription/playback/session dispatch, verifies input levels, floating alpha and reflection motion, cancels capture, closes its own app and removes its exact disposable profile. Intermediate frames remain ignored under .local\readme-capture; Pillow creates bounded 960px, 10 fps looping GIFs with a shared palette per scene and delta frames. The full command also captures 640 idle Web Animation samples and creates readme-bubble.gif plus its PNG still. Explicit screenshot scaling and PNG dimension assertions guard the close-up resolution; device-scale flags alone did not enlarge the hidden host's Playwright crop. Every decoded GIF frame is compared with its quantized source, including the complete loop duration. These opaque gallery composites keep prior frame contents; the transparent hero still uses clear-frame disposal.