Release Notes

Last updated: September 2, 2026

v0.3.11 — September 2, 2026

The current Mac download at /download.

Gmail and Google Calendar connect out of the box. The app now ships its Google connection config, so a fresh install can connect Gmail from the System tab — no setup files. Google shows an "unverified app" notice during sign-in while our verification is in progress; Advanced → continue proceeds.

v0.3.10 — August 30, 2026

Connect sources on a fresh install. The System tab's sources table now lists every available source with its state — connect Gmail, grant access to Apple stores — even before anything has been ingested.
Backups: fingerprint everywhere, and exports. The System tab's back-up button now asks for Touch ID like the CLI does. Downloading your backup is sync pull --export-to — it writes a decrypted copy only where you point it and refuses your live data outright.

v0.3.9 — August 27, 2026

Sources state the real reason. A macOS store (Reminders, Notes, Calendar, calls, Voice Memos) the app can't read now says "present but unreadable — grant commissure Full Disk Access" instead of wrongly claiming no store exists.
Running views state their work. While a generation runs you see the model, route, round, and context size in flight — a long first call reads as what it is, not a hang.
Timeline and navigation fixes. Timeline day-drill and filters behave; the entities tab is addressable; assets refresh cleanly after updates.

v0.3.8 — August 21, 2026

Fix. Generations with web search enabled failed whenever the model chose not to search — the relay treated "no searches" as unreportable billing and withheld the answer. Zero searches now settles at token cost.
Failed refreshes stay visible. A view whose refresh failed used to show the previous run as latest with no trace of the attempt; the failure now stays on the surface until a refresh succeeds.
Send feedback from the app. Help → Send Feedback… It sends only what you write and the app version.

v0.3.7 — August 21, 2026

Fix. In the installed app, asks routed through Commissure credits failed with a model-catalog error before anything was sent — the credits model list didn't ship inside the app. It does now, and commissure sync keygen works from the installed app too.

v0.3.6 — August 21, 2026

Touch ID for backups. Enroll a backup key once (commissure sync keygen) and vault backups are approved with your fingerprint — no typed passphrase. Backup uploads also retry through flaky connections.
Site. Download counts are measured where downloads actually happen, and the euhedron.com page fits laptop screens.

v0.3.5 — August 20, 2026

Site. The ask assistant can record your feedback for the maker, knows the story behind the name, and each answer states how many lookups it made. Clearer release-state wording on the account page. No app changes.

v0.3.4 — August 19, 2026

A stable local address. The app's local service now serves on one well-known port, so command-line tools and agent integrations reach the same running service as the app window.

v0.3.3 — August 19, 2026

Reliability fixes. The local service's secure connections now verify correctly (sync status, announcements), the System tab loads faster on large libraries, and controls that can't apply on this install are no longer shown.

v0.3.2 — August 18, 2026

Smaller updates. The app now downloads only what changed between versions (differential updates), and interrupted downloads resume.

v0.3.1 — August 17, 2026

commissure is now a standard Mac app. Drag-install DMG; the app starts its local service, keeps it running, and opens signed in. No terminal steps anywhere.
Auto-updates. The app checks in the background, downloads updates, and prompts to restart. Check for Updates is in the app menu.
Apple Silicon. Requires an Apple Silicon Mac on macOS 13.4 or later; v0.2.15 was the last Intel-capable release.

v0.2.15 — July 23, 2026

One download for Intel and Apple silicon. The Mac app and every bundled native dependency are verified for both architectures, and the release launches under each before publication.
Route status distinguishes ready, depleted, rate-limited, signed-out, and policy-disabled, names the exact model and billing arm, and never reports an installed CLI or present key as able to run until a real call has succeeded.
Provenance at the write boundary. Captured items record the server-known session lineage that created them; a client cannot label its own write as coming from a more trusted device.
Passkeys verified end to end. Registration and sign-in use server-verified WebAuthn ceremonies with one-use challenges, origin and relying-party checks, counters, and backup-state tracking. New credentials require user verification; a pre-existing credential that has never demonstrated it can still sign in presence-only, and financial and operator boundaries then require the user-verified check.
The app, update API, download key, and deployment metadata read one publication record and expose the exact Worker version serving a request; when that record is absent or inconsistent they refuse rather than guess.

v0.2.14 — July 19, 2026

Apple Notes, indexed. Your notes join the corpus as a first-class source — searchable and available to every agent, alongside your messages and mail.
Finish signup inside the app. New to commissure? If you don't have a passkey yet, the sign-in screen emails you a link that creates your account and connects the app in a single step.
Source freshness tells the truth. ChatGPT and Claude exports age by the export's own date — the System tab flags them when the archive is old, not when it was last imported, so you know when a re-export is due.
Launch does not depend on Apple being reachable. The app inside the DMG carries its own notarization ticket, so supported Macs can verify it even on locked-down or offline networks.

v0.2.13 — July 18, 2026

Features you can switch off. System → features lists every view with an on/off toggle — flip one and its tab appears or leaves the sidebar immediately. Research views aimed at a job search ship off; turn them on if that's your use.
Results read like documents. Generated results render markdown — inline code, bold, lists — with short fields (kind, severity, basis) as compact property chips instead of label rows.
Read any feature's prompt. commissure prompts lists every AI feature; prompts <name> prints the exact composed prompt it receives, its tools, and its output schema — the same text each view's ⚙ shows.
Models, grouped honestly. The per-feature model map splits core (rides your default) from utility (titles, enrichment — shipped pinned to a cheap model), so nothing runs on a model you can't see.
Activity leads with runs. The activity tab moves next to agents and system, and opens on the run ledger — every generation, newest first — with the full event feed one dropdown away.
Drafter. A new view that shapes any tracked project's draft or spec: it reads your reactions and proposes anchored edits you apply or decline — nothing changes the text but you.

v0.2.12 — July 18, 2026

A calmer search bar. Search modes and sort stay up front; the occasional filters — sources, a name to scope by, a date range — tuck behind one "filters" control, and the empty search centers the keyboard shortcut.

v0.2.11 — July 18, 2026

Setup that states, not sells. The first-run wizard reads plain now — pick how AI runs, connect your sources, done — with the loud "ready" highlights quieted to match.
Search names what it finds. Results group under "entities" — people and companies alike, each row showing its own kind — instead of labeling everything "people"; the empty search shows the keyboard shortcut, not a hint.

v0.2.10 — July 17, 2026

Ask your corpus from anywhere. Your question travels to your device, which answers from your local corpus and sends the answer back — from the web, your phone, or your wrist. The corpus is never shipped to a server; a local model sends nothing out, a cloud one sees only the snippets it needs, like any ask. Off by default; turn it on in the syncing table.
Sharper research, with its sources shown. The scout and shortlist views now verify company facts against the live web and cite them — sources render as clickable links, corpus references as chips that open the original. Web searches are recorded in the trace, so you can see exactly what a run checked.
Your mail heals itself. Incoming Gmail now reconciles a rolling window each pull, so a message a transient hiccup dropped gets picked up on the next sync instead of vanishing for good.
Per-feature model control. The account panel lets you pin any feature to its own model — fast models for cheap work, your best one for judgment — with the whole map visible.
A calmer System tab. Sectioned like a settings screen, with the developer-only "newer code on disk" note demoted to a quiet chip instead of a wall of red.

v0.2.9 — July 17, 2026

Syncing, visible and yours to flip. The System tab gains a syncing table: every channel that can leave your device, one row each with live state and an on/off toggle — your rulebook, open items, the coordination reading, the backup vault. The corpus row says never, and means it.
Rules and items on your account. commissure sync rules and sync items (or the toggles) publish read-only copies to your account page — glanceable from any device, stamped with their age, clearable any time. Items travel as stubs: titles, never bodies.
Share your rulebook as a pack. commissure rule share <slug> mints a versioned pack at a link you can hand anyone; rule import pulls one in as its own layer — your rules and the defaults stay untouched, every overlap named.
One-click encrypted backup. The vault row's back-up button encrypts corpus.db + kernel.db on your device and uploads — the passphrase never leaves the machine, and the row turns amber when the snapshot is over a week old: its age is your loss window.
Capture from anywhere. A mailbox for quick input: commissure capture send drops a one-liner from any signed-in device; the Mac pulls it into your signals and items with recorded channel provenance. The watch capture screen is built and headed for TestFlight.
Sources say when there's more. Agent-history and export sources (cursor, codex, claude code, chat exports) now show "new since …" when the source holds newer data than your last pull.
Where AI calls go, always visible. A route chip in the status bar and a line under ask name the active provider and model — per call, not per guess.

v0.2.8 — July 16, 2026

Track companies, not just people. The entities tab now holds tracked entities — companies, orgs, products you add yourself or elevate from a research run. Give one a website and your own mail and browsing history resolves onto it, so "who do I know there" is a lookup, not research.
Scout: research the outside world. A new view that grounds in a project's spec, searches your corpus for warm paths, and — first in the app — searches the live web, emitting candidates with cited why-fits. Every candidate carries the corpus check's verdict; keepers elevate to tracked entities in one click.
Shortlist: the shelf, ranked. Every tracked entity scored against your criterion relative to the rest — each score carries its evidence and the single best next move. It re-verifies warm paths instead of inheriting them: in its first live run it caught a departed contact and a bounced address the discovery pass had trusted.
Web search is a tool you grant. Provider-side live web search joined the tool registry — grant or revoke it per feature like any tool (commissure tools grant scout web_search), visible in the grants table.
Configure any feature's toolkit. Tool grants are editable everywhere: commissure tools grant/revoke/set/reset in the CLI, checkbox editing in the system panel, same validation both ways — a typo returns the catalog instead of silently stripping a tool.
References: pinned exemplar repos. Register codebases (openai/codex, the MCP SDKs, Google's gemini-cli) as evidence audits and asks can search and quote at an exact pinned commit — citations open the cited file at the cited line, and stay stable when the reference re-pins.
Runs scope to a project. Any view can run against one tracked project — its spec rides as the stated intent (--project in the CLI, a scope picker on every view toolbar). Audits read the project's actual code through project-scoped tools.
Transient failures don't kill runs. Model calls retry rate limits, overloads, and dropped streams with a bounded backoff ladder — auth and quota failures still surface immediately, and every retry is recorded on the run, never silent.
A failing key can't hide. Key failures leave a visible stamp — commissure keys list shows "failing since" with the error next to a present-but-rejected key, and the routing table says "key present" instead of claiming "ready" for keys it never probed.
Every run explains itself. The run's self-assessment note is schema-required and rendered under its results — a clean pass reads as a verified verdict, never a blank list — and traces always retain what the run did, even when large outputs are trimmed.

v0.2.7 — July 16, 2026

Projects. A new tab tracks your projects — each card shows the repo's live state (branch, uncommitted work, last commit), which agents are working in it right now, and its open items. Items can be scoped to a project; everything is also available from the CLI (commissure project).
Specs. Every project can carry a spec — the written intent of what it should be. Revisions are kept whole with their history and reason; runs that rank or audit a project ground in its spec.
Rank your portfolio. A portfolio view scores every tracked project against any criterion you give it ("momentum", "closest to revenue"), each score carrying its grounded why.
Several accounts per provider. API keys now hold named slots — keep two Anthropic or OpenAI accounts stored, see each key's masked identity, and switch the active one in a click (account tab) or one command (commissure keys use).
Newest messages arrive on time. iMessage ingest now reads Apple's write-ahead log, so messages from minutes ago are searchable immediately instead of after macOS checkpoints — sometimes days later.
Long runs aren't cut down mid-flight. The run reaper now checks a run's live heartbeat before declaring it dead, so a 20-minute multi-agent sweep completes instead of being marked "timed out" at ten.
Runs that produce nothing say so. A run whose output has no readable content is marked failed with the reason and cost — the panel keeps its last good generation instead of banking husks as success.
Honest chrome. One floating popover at a time; status chips are click-to-open with an "open ↗" into their full section; the machine-resources chip shows load and commissure's memory; unknown thread ids read as not-found instead of an empty conversation.

v0.2.6 — July 15, 2026

Reads resolve people. In search results, item reads, and thread reads (dashboard, CLI, and MCP), message senders and thread titles show the person's name where the handle maps to a known entity; unmapped handles stay as-is.
GPT-5.6 models selectable. GPT-5.6 Sol, Terra, and Luna can be chosen in routing, view model panels, and metered credits, priced per OpenAI's published launch pricing.
Runs show from any surface. A generation started from the CLI or an agent marks its view tab and run button while in flight, and the open panel reloads when the run lands.
Status chips state what changed. The embeddings chip distinguishes an existing index with an unreachable query embedder ("semantic offline" — hybrid search continues with exact matching) from an index still building; the agents chip lists each active surface·project pair; the embeddings, activity, agents, and processes chips all peek on hover and toggle on click.
Result cards open into the drawer. Reactions, evidence, and provenance live in the detail drawer each card opens; card actions are named for what they do (file as item · delve); entries from one known-degenerate earlier run are labeled as such.
Site ask keeps its thread. On commissure.ai, the ask panel's conversation persists across page reloads in the same tab, and follow-up questions carry the earlier turns.
Higher run ceilings, labeled when hit. View runs default to a 150-tool-round ceiling, and a run that reaches its ceiling is marked "coverage may be incomplete" instead of finishing silently.

v0.2.5 — July 14, 2026

Semantic search works out of the box. Meaning-based search now ships inside the app — nothing to install, nothing to run alongside it. If it's ever unavailable, search falls back to exact matching and says so honestly.
The menu-bar widget ships with the app. Turn it on or off from the dashboard — no separate install or hand-launch.
Tune how views run, in the app. Each AI view's dispatch — how many agents, which models, each agent's focus — is editable in its config panel, with a reset back to the shipped defaults. Multi-agent results say which model produced each entry.
Star what's good. Any view result can be favorited; stars are kept as taste signals that future runs read.
Attach an image. Reactions and item capture accept a picture alongside your note.
Long results get a contents rail. Multi-section results show a section index, and past generations sit in one snapshots dropdown everywhere.
Drafting and sending repaired. The draft → approve → send flow works again after an internal refactor orphaned it, and drafts regenerate in place.
Local settings don't ask for a login. Appearance, routing, and your own API key work signed-out — and even when commissure.ai is unreachable. Only account things (credits, tier, usage) ask you to sign in.
Fixes & hardening. Restores stage to a temp file so a failed restore can't half-overwrite your data; dashboard output escaping tightened; a crafted entity id can't escape its cache; CLI get/thread crash fixed.

v0.2.4 — July 12, 2026

Credits cover both providers. Claude and GPT models both run on commissure credits — pick any model, pay per call. The subscription is no longer sold; existing subscribers keep their plan.
Run a view as several agents. Any AI view can run as N concurrent agents — each on its own angle if you set one — merged by diversity or agreement. Configurable per view; the launch dialog takes a focus for any single run.
Past runs are kept. Every view keeps its last 12 generations — pick any from the toolbar, or download a run (or a single result with its evidence and trace) as a timestamped file.
Ask is in chat order. The conversation sits above the composer; past asks collapse, filter, and sort by cost.
Search shows people. Message results name the resolved sender, not a bare phone number.
Dictate anywhere you capture. A mic on reactions, item capture, notes, and ask — where the browser supports speech recognition.
Check an item against reality. Every tracked item has an analyze action that verifies it against the current state and notes the assessment, with the full trace.
Longer results, fewer failures. Output limits are safety ceilings now, so long structured results no longer truncate mid-run.

v0.2.3 — July 1, 2026

Photo text search. Text inside your iMessage photos is now read on-device (Apple Vision) and made searchable — the image never leaves your device.
Choose your model. Pick the model your AI runs on in account → routing (Opus, Sonnet, Haiku, and Fable 5 where your access allows). If a model isn't reachable you get a clear reason, not a silent failure.
Sign-in emails reach the inbox. Magic links now pass SPF, DKIM, and DMARC, so they stop landing in spam.
Local store locked down. ~/.commissure is readable by your macOS account only.
Fixes. Crashed runs no longer show as live; the mobile site no longer zooms the sign-in box or offers the Mac download on a phone.
Update notices. The app tells you when a newer version is available.

Earlier

Earlier builds were private/internal without public notes. Public release notes start at v0.2.3.

← back to commissure