On 13 November 2026, the owner stopped using Intercom because it had not delivered the requested experience: speak naturally, have existing Copilot sessions do useful work, and see their observed context activity in the mesh. The owner's assessment was explicit: "the things that I wanted to achieve I have not been able to achieve" and "you have generally disappointed me".
The latest live .18 run reached an existing session, but stopped when the connector rejected concurrent tool-permission requests. Its outcome is unknown, not completed; any partial work needs inspection before a retry. Successful transcription, clarification capture, isolated tests and attractive visuals did not establish a reliable end-to-end product. The coding agent overstated readiness by treating those milestones as sufficient delivery.
Implementation is paused at the owner's request. This source push records the work and the failure; it is not a fix, a new binary release or human acceptance. Any future resumption must prove the complete delivered workflow, including real permission handling and truthful failure recovery, before claiming success. See the development journey and accountability record.
Sea-glass lettering, frosted mesh and softly flowing multicolor waves. View the still logo.
Inspired by mid-century switchboard operators and the conductor of an opera orchestra, Intercom aims to turn one spoken request into coordinated work across existing Agency/Copilot sessions: connect the right specialists, cue their tasks and bring their replies back. Its sea-glass Context Mesh visualizes those connections, lighting paths only for observed activity.
KDD exchange, 1967. Asahi Shimbun. Source / rights. Used with owner-confirmed permission.
GUI demos below show the current source, not published .12 or speech-quality evidence. Contexts are fictional; audio input is synthetic. Each has a reduced-motion still.
Smooth glass in motion. Your input ripples through its reflections, not its rotation. Reduced-motion still · Synthetic local speech; silent GIF.
Floating Mesh still · Owner-supplied wallpaper; actual transparent GUI capture, staged composite.
Integrated local .18 (not published): local development connects the restricted controller to a durable multi-session scheduler, with Intercom-node default focus, conditional checks, exclusions and manual tool approvals. Clear requests need no second target/task confirmation. Disposable fixtures verify this flow. Recorded English input and one strictly bounded existing-session reply also pass in separate protected test lanes; real multi-session work and human speech accuracy remain unqualified. See the recorded-input qualification. Local candidate .18 includes integrity-checked multilingual Whisper, automatically selected on fresh profiles with the system microphone, auto language, local voice and speed-first routing. Existing explicit engine choices stay unchanged; permissions remain manual. It also tolerates unnamed catalog sessions without guessing targets and offers optional Talk > Live transcription (local draft): periodic on-device previews, off by default, never automatic partial sends. Recognition now keeps one native model loaded across drafts and final transcription through private pipes, without a network listener. There is no silent Windows/CPU fallback. An 8 kHz synthetic example confused session "Y" with "why": speech accuracy is not accepted. The exact .18 portable passes synthetic local speech, warm-worker reuse and shutdown checks on this machine; broader GPU and human speech qualification remain open. The published .12 build is unchanged.
Clarification replies (local .18): after Intercom asks a structured clarification, the mic opens automatically when local playback finishes, or after written display when muted. Click the mic to finish your reply. Stop playback, Cancel, opt-out and hidden/minimized windows suppress pending listening. Disable Automatically listen for clarification replies under Advanced delivery settings. This is bounded capture, not continuous listening or automatic detection of arbitrary downstream questions.
Five-minute recordings (local .18): capture and transcribe up to five minutes locally. Automatic cutoff preserves the transcript for review and never sends an unfinished request; ordinary manual stopping retains existing send settings. This is not included in published .12. See long-recording behavior and evidence.
Observed context research (local .18): ask Intercom to find a project's contexts and session history. A separate read-only lane inspects registered notes and requested, exactly attributed local histories, then saves and opens a private illustrated HTML report with citations and a coverage audit. No existing session is resumed for ordinary collection. Actual source reads now drive finite mesh pulses and an attributable activity trail. Optionally select an exact existing idle session for cited Copilot synthesis through the normal permission-controlled queue. Citation locations are validated, not the truth of every conclusion. Topic references are not proof of ownership, completion or live access. These features are not included in published .12. See current candidate and limits.
Latest unsigned executable:
C:\devdesk\homedesk\Intercom\release\prototype-18-candidate\Intercom-Prototype-0.1.0-prototype.18-win-x64.exe.
Only this executable and its checksum remain in the local release folder;
existing published GitHub releases have not been changed.
Portable recovery and diagnostics: .17 isolates each launch's extracted files so a rejected duplicate cannot remove the first instance's speech model. Persistent metadata-only logs and automatic process/resource failure snapshots are available under Advanced > Display > Local diagnostics. No audio, transcripts, prompts, responses or full memory dumps are logged. See diagnostic location and limits.