Changelog

Home keeps a living briefing with open loops, questions, time allocation, and connector context so you can pick up where work left off.

Learning supports Low and High capture modes, and the Learning HUD Ask can use recent on-screen context while you stay in flow.

Automations can keep running in the background, and Stop lives on the Automations page instead of interrupting Chat.

Sensitive details such as passwords, emails, auth tokens, API keys, private keys, and connection strings are cleaned before the Agent reads your captures.

Pacific Machines now requires version 0.1.61 or newer so every install gets Home, capture modes, and the privacy cleanup updates.

The Learning HUD expands into a clean capsule without a faint rectangular plate or shadow behind it, so it stays readable over your desktop.

While Learning is on, the header Learning control opens a status popover with capture details and an explicit pause action.

Schedules is renamed Automations in navigation and the account menu. The standalone Skills page is removed; Chat still runs skills through slash commands.

Pacific Machines now requires version 0.1.60 or newer so every install gets the Learning HUD and Automations updates.

The local agent runtime is locked at build time and shipped inside the signed app, so Pacific Machines no longer downloads those packages from the internet on your computer.

Staff remote support is off by default. You must enable Settings → Support → Staff remote support before an authorized support session can run on this device.

On Mac, capture folders and database files are restricted so only your user account can read them. Older installs are corrected when the app starts and when Learning starts.

Pacific Machines now requires version 0.1.59 or newer so every install gets these security and support controls.

Windows Learning no longer fails to start on PCs missing a common Microsoft runtime, which previously looked like a permissions problem.

Pacific Machines now includes that Windows runtime with the app, so Learning can start on clean machines without a separate install.

Pacific Machines now requires version 0.1.58 or newer so every install gets the Windows Learning startup fix.

Low-mode Learning keeps the app, window, and text context needed for daily recaps without saving continuous screenshots, OCR frames, or video.

The Learning HUD sits higher at the top center of the screen so it stays visible without blocking your work.

Learning capture is more stable on Mac and Windows, including a fix for Mac installs where the capture bridge could crash and restart in a loop.

Pacific Machines now requires version 0.1.57 or newer so every install gets the lighter Low-mode Learning path and the matching capture fixes.

Chatting about your day no longer pauses Learning. Capture keeps running while Chat looks up what you were doing.

Chat answers about recent work are based on your capture history, so they come back faster and stay tied to what was actually on screen.

Pacific Machines now requires version 0.1.56 or newer so every install gets uninterrupted Learning and the matching Chat answers.

Find Opportunities is Brain-first: it uses what Update Brain already learned, so you no longer wait on a separate evidence pack or multi-stage discovery pipeline.

Update Brain catches your organization up through yesterday and keeps Chat able to write Brain facts when you ask it to remember something.

Check Brain is non-blocking and opens a result dialog, so you can keep working while Pacific Machines confirms Brain health.

Pacific Machines now requires version 0.1.55 or newer so every install gets the Brain-first Find Opportunities flow and the matching cloud guards.

When your build is below the minimum required version, Pacific Machines blocks with a clear Update action that downloads and installs through the signed updater on macOS and Windows.

The required-update screen no longer sends you to a downloads folder — Update is the only path, so you stay inside the verified release channel.

macOS installer and updater artifacts are branded Pacific Machines, matching the app name you see after install.

The minimum-version gate keeps a clear Update action available whenever your build is below the required version.

Required updates download and install through the signed updater on macOS and Windows instead of sending you to a separate downloads step.

On Windows, Pacific Machines could stay stuck waiting for Screen capture even when Screenshots and screen recording was already allowed. Learning now confirms access with a real capture check so permission can clear and recording can start.

Learning uses a lighter capture mode by default, so scrolling and same-app window switches no longer generate unnecessary capture work. That keeps more of your computer free during meetings and everyday multitasking.

After you open Windows Privacy settings for Screen capture and come back, Pacific Machines restarts its capture check so the new permission state shows up without a full app restart.

Saving an ignore list no longer deletes matching past captures from your local database. From that point on, Learning simply skips the ignored app or website.

The Learning pill could stay paused because an ignored website host was also matched against unrelated window titles. Desktop ignore rules are scoped to app names, and URL rules stay URL-only.

After you leave an ignored app, capture resumes cleanly instead of getting stuck on a stale paused state.

The Learning pill could stay paused because an ignored website host was also matched against unrelated window titles. Desktop ignore rules are scoped to app names, and URL rules stay URL-only.

After you leave an ignored app, capture resumes instead of staying stuck on a stale paused HUD state.

On Windows, Chat could fail immediately with “we're having problems” because assistant packages were still installing when the first message started. Those packages are prepared before the assistant process starts.

Chat no longer shows Ready while setup is still incomplete; send stays gated until the runtime finishes, with wait and retry timeouts for slow machines.

If the assistant process fails to start, Ask retries once before surfacing the error.

Each day is measured into non-overlapping capture blocks first; work episodes only assign purpose to those blocks, and every start time, end time, duration, and day metric is derived from them.

A recap can no longer keep overlapping episodes or totals that do not match the minutes that were actually captured.

While an analysis runs, the progress card shows the current step and elapsed time, including what each subagent is doing.

Analysis runs that stop making progress are cut off with a clear failure instead of appearing to run forever.

Analysis agents are scoped to the Pacific Machines data folder and cannot recursively scan your home folder or screenshot store.

Every analysis kind inherits the same stall detection and heartbeat reporting, not only daily recap reconstruction.

Workflow and opportunity builders only read the files named in their task, which keeps discovery faster and more predictable.

Work-episode reconstruction no longer searches the whole Pacific Machines data folder or the filesystem root while building a day.

Capture database reads for a day are bounded with a timeout and output size limit.

Subagent progress heartbeats and a no-progress cutoff make long recap runs easier to follow and less likely to hang.

Settings opens Changelog, Documentation, Security, and Terms on pacificmachines.ai in your default browser instead of embedding separate copies in the app.

The Privacy dashboard keeps local capture, cleanup, ignore-list, and deletion controls, and links out to the public Privacy Policy.

Pressing Learning while capture was already starting in the background showed “recorder is already running”. Both paths now share one start, so the session simply continues.

Automatic privacy redaction after ending a session could fail with a database lock while capture was still writing. Redaction now writes in short batches and waits its turn.

Workspace membership sync failed when the connection dropped mid-request. It now retries with backoff and finishes on its own.

Find opportunities shows which days still need a daily recap instead of “Something went wrong”.

Internal diagnostics for Chat startup, capture, and sync failures are far more detailed so issues get fixed faster — nothing extra is shown in the app.

The Windows app is named Pacific Machines / Pacific Machines Preview in Task Manager, the installer, Start Menu, and the install folder — not Jarbas or JarbasPreview.

Updating from an older Windows install renames the legacy Jarbas or JarbasPreview folder in place so you keep a single app.

Daily recaps lead with an episode overview and a full work-episode timeline instead of jumping straight into other scenes.

Calendar Day view cards show observed actions and full episode descriptions for every timeline layout.

Days ready for capture show generate cues, and active time is always visible once a recap exists.

Chat no longer fails after large recaps because history messages exceeded size limits.

Work episodes are reconstructed more consistently so timelines stay complete and readable.

Workflow discovery agents now prefer canonical files under ~/.jarbas and avoid broad home-folder searches unless an external path is explicitly provided.

Opportunity publishing now reads the current UI language for every turn instead of reusing the language from an older Chat process.

Publish receipts and cloud payloads no longer duplicate the full workflow audit, preventing oversized messages and save requests.

Following up in Chat after publishing no longer fails because an opportunity JSON message exceeds the history limit.

In dark mode, the company description text area on create organization uses the same dark field styling as the other inputs.

After entering your job title, you can describe your role in a text area; that description is stored on your account.

After sign-up, the app shows create/join company instead of staying on Loading while the session is still pending.

A finished Find opportunities run can no longer save a hand-built stub map with missing steps, apps, and evidence.

Opportunity titles and details follow the language selected in the app, instead of mixing English into other languages.

When a workflow step has more than one next step, the diagram draws separate arrows instead of one blended fork.

Finding opportunities needs fewer repair rounds, so a full discovery run finishes sooner.

The Learning pill could stay hidden on Windows after starting Learning, because two requests to open it raced each other.

Starting the app no longer waits for a check of the bundled shell, which could stall the first launch after installing while Windows scanned the new files.

The Windows installer is roughly half its previous size, so downloads finish and Windows is less likely to block them.

Windows Chat no longer fails activity and capture questions because Bash was missing on machines without Git for Windows.

Pacific Machines includes a real Bash runtime for Windows Chat, so looking up your local activity database works on a stock install.

The app bundle is named Pacific Machines / Pacific Machines Preview on macOS, so the Dock and the Screen Recording, Accessibility and Input Monitoring lists no longer show Jarbas.

Installing an in-app update on an older Jarbas.app or JarbasPreview.app renames it to the new name before relaunching.

Settings and onboarding now include Input Monitoring on macOS, with Grant and Open Settings.

Learning start and auto-start now treat missing Input Monitoring like missing Screen Recording or Accessibility, instead of starting a session that stores nothing.

Permission warnings name Input Monitoring when it is the missing grant.

Learning no longer reports recording when macOS Accessibility access is missing, which left capture running without storing anything.

Start recording on launch no longer fails silently — a missing permission now says so instead of leaving Learning quietly off.

A live Learning session is checked for stored activity and warns you if nothing is being captured, including when Input Monitoring is turned off.

Assistant tools are loaded only once on Windows, preventing startup conflicts caused by drive-letter casing.

Windows x64 preview installers now bundle matching x64 runtime components.

Windows now loads from the correct localhost origin so Clerk sign-in no longer fails or leaves a blank screen.

Windows installs a real agent-browser executable and prepares Chrome without blocking Pi agent setup.

Windows and Mac ship under the same app version with platform-explicit installer names.

Mac no longer crashes when Learning starts or auto-starts on launch.

Starting Learning on Windows no longer freezes Chat, History, Recaps, and Settings.

Chat no longer hangs when Windows Git Bash was set as the shell environment.

Windows sidebar spacing above the organization switcher matches the content chrome.

Share on Team Dashboard creates a short live link anyone can open in a browser.

Shared dashboards support scrolling and person opportunities without signing in.

Share links no longer depend on this computer’s local address.

Alternate names for the same org, product, team, or app no longer become separate Opportunity cards when the recurring trigger and outcome match.

Workflow inventory validation rejects label-only duplicates before investigators fan out.

Find Opportunities identifies specific recurring workflows from your measured time—not broad umbrellas—and turns each into its own Opportunity card.

Opportunity detail is easier to review, with clearer feedback and regeneration when a card needs another pass.

Chat shows which Plugins are connected for the conversation.

Chat activity transcripts stay collapsed by default so the reply stays easy to read.

Current workflow diagrams show distinct run shapes as modes with visible lanes instead of one forced left-to-right path.

Find Opportunities keeps major responsibilities as separate cards when evidence shows independently useful outcomes—not sibling lanes of one map.

Merge vs split for major workflows follows your observed role outcomes and supporting stages, not hardcoded category pairs.

Workflow graphs are easier to inspect when modes, frequency, and bottlenecks matter.

Find opportunities builds problem-first cards from observed work, with Evidence, Workflow, Impact, and Build tabs.

Workflow maps show today and the automation path as separate diagrams, with readable nodes and clearer loops.

The Evidence tab presents observed and inferred facts in a scannable layout instead of dense all-caps blocks.

Saving the same discovery twice no longer creates duplicate opportunities.

Fullscreen workflow diagrams hide the window traffic lights so the title stays readable.

Find opportunities personalizes from your Brain and recap time sinks — not from which Plugins happen to be connected.

Plugin feasibility checks whether a toolkit exists in Composio; connected-today does not change discovery or ranking.

Opportunity titles, New/Viewed badges, and observed-episode labels follow the app language.

Workflow diagrams no longer show empty Action nodes that only repeat an app name.

Pacific Machines Preview no longer opens an empty window — the main UI loads on first launch.

Daily Recaps rebuild your day as adaptive work episodes instead of fixed ten-minute buckets, so stories follow how you actually switched tasks.

Team Reports synthesize only from saved member recaps — they never read personal capture or brain files.

Organization context syncs into the personal brain without wiping existing notes when you create or join a company.

New routine opens Chat to set up a managed schedule instead of a create dialog.

Theme (System / Light / Dark) lives in Settings; Schedules cards use the same surface styling as Settings.

Switching organizations no longer hits auth gaps that broke item-view loading.

Minimum compatible version is 0.1.19. Opening an older build shows a required update screen, and cloud features refuse outdated clients until you install the latest.

Chat identifies as the Pacific Machines assistant — never as Gemini or another model name.

Find automation opportunities presents scannable comparison and impact tables instead of long bold prose.

Activity reports call find-automation-opportunities for their Automation Opportunities section instead of inventing cards.

Chat can create and manage Schedules via /manage-schedules (create, list, pause, resume, delete). Same on-disk format as the Schedules UI.

Scheduled runs open in Chat (not a dialog). Each run is a real conversation you can continue messaging in.

Chat keeps the personal brain up to date on its own: consolidates misplaced facts, fills empty canonical files (e.g. identity), and deduplicates — no need to ask “update the brain.”

Recaps and brain updates may use already-connected Plugins to clarify meetings and app threads when capture alone is opaque (read-only; never mid-recap connect flows).

Scheduled automation runs now send the prompt correctly to the agent (was finishing with no reply).

Merged daily, weekly, monthly, and yearly recap automations into one /create-recap automation that takes a date or range. Day mode includes work-episode reconstruction as internal docs (WORK_EPISODES.md).

Creating a Daily Recap (or week/month/year rollup) opens Chat and sends the request automatically, instead of the progress checklist on the Recaps page.

Capture database rules now live in the always-on system prompt instead of a separate /local-data automation.

Renamed the connected-apps automation from /composio to /plugins. Tool names stay internal; Chat never says Composio to you.

Chat markdown uses real prose spacing (paragraphs, lists, headings) instead of uneven hardcoded gaps.

Brain Preview follows in-brain Markdown links without a full-page reload, and hides HTML authoring comments.

Stopped installing the Pacific Machines automation into Cursor, Claude, and Codex. Leftover copies are removed on launch.

Daily, weekly, monthly, and yearly recaps keep the analysis process alive after the JSON write so /update-brain can finish updating your brain.

Retired analysis kinds (opportunities, team reports, background activity reports) are blocked at start with a clear message — use Chat /generate-report and the Team page instead.

Recaps calendar loads lightweight list metadata; full stories load when you open a recap. Period-recap prompts no longer embed the full capture reading protocol.

Schedules automation picker hides analysis-only automations that cannot run as scheduled agent jobs.

Brain graph pauses its animation when the window is hidden to save CPU.

Settings → Automatic capture: optionally start recording when Pacific Machines opens, and optionally open the app at login so capture can begin when you turn on this computer.

Closing the main window on macOS hides it instead of destroying it, so clicking the Dock icon brings Pacific Machines back. Quit still fully exits.

Settings notes that macOS hides notification banners while Screen Recording is on, and how to allow them in System Settings → Notifications.

App language setting (English, German, Spanish, Italian, Portuguese, French). Menus and labels follow your choice; AI instructions and brain stay in English.

Brain sidebar page: browse your personal brain with an Obsidian-style file tree, rendered Preview, and editable Files source view.

Seeded brain structure: identity, voice, beliefs, preferences, relationships, goals, projects, decisions, playbooks, journal, lessons, open loops, glossary, working memory, and inbox.

New /update-brain automation runs after every Daily Recap to grow working memory, projects, lessons, and inbox proposals without inventing identity.

Team sidebar: browse everyone in your organization and open each person’s saved activity reports.

Chat `/` and Automations now show built-in automations (generate-report, skill-creator, update-brain, and more) alongside yours in one library; built-ins can’t be deleted or edited.

Brain Graph view: an Obsidian-style interactive map of your notes — drag, pan, zoom, click to focus, double-click to open a file.

Brain reorganized for scale: core/, now/, work/, topics/, archive/, and time-scaled memory/ (daily → weekly → monthly → yearly), each with README guidance. Agents may add folders, files, and READMEs as life expands.

Brain has three views: Graph (interactive map), Preview (rendered note), and Files (source editor).

Onboarding asks for your job title after you create or join an organization, saves it to your account, and writes it into your brain identity.

Weekly, monthly, and yearly recaps: from Week, Month, or Year calendar view, generate a story that rolls up lower-level recaps into ~/.jarbas/weekly-recaps, monthly-recaps, or yearly-recaps and updates brain memory/weekly, memory/monthly, or memory/yearly.

Activity reports auto-save to your account when Chat finishes generating them, and the Reports sidebar lists every saved document for full-screen reading.

Generate Report automation rebuilt: it always confirms the timeframe (or accepts all time), reads your daily recaps, and produces a 17-section activity-pattern analysis with a source audit, executive dashboard, repetition map, leverage matrix, and action plan.

Reports render as styled HTML documents in chat — KPI cards, confidence badges, bar charts, flow diagrams, loop cards, and a 2x2 leverage matrix — with live rendering while the report streams and an expand-to-fullscreen view.

Report visuals follow the app design system in light and dark mode, with clean tables capped at four columns and evidence bullets instead of cramped cells.

Retired the find-opportunities, list-opportunities, list-reports, team-report, and record-feedback automations, and removed cloud-stored opportunities; personal activity reports now generate from daily recaps and save to your account.

Daily Recap builds a timeline of work episodes — one coherent objective each — instead of app-by-app time slices.

Recap story timeline cards show objective, actions, outcome, and apps for each episode, with a clear episode count.

Daily Recap calendar switches between Month, Week, and Day views, with navigation and a day detail panel for open or create.

Find Opportunities reads saved daily recaps first, clusters work episodes into workflows and responsibilities, then qualifies automation — it no longer rebuilds the day from raw capture.

Find Opportunities will not run until every captured day in the selected range has a daily recap, so opportunities stay grounded in the same episode story you already read.

Opportunity detail is a decision page: why automate this, current process, what the agent would own, when it is done, and what you still decide.

Thin evidence can mark an opportunity Keep observing, with an explicit build decision instead of forcing Ready or Blocked.

Completion notifications use the Pacific Machines app icon instead of the Script Editor glyph in development.

Date calendars inside Find Opportunities and other range dialogs stay open when you pick a day — the dialog no longer closes on the first click.

Analysis AI tabs load the full saved transcript (thinking text and tools in order), not only the slim checklist that said nothing was saved.

Connect / link / authorize requests no longer search local capture. Chat looks up the Plugin for that app first, even when the name is unknown or misspelled.

If Chat finishes a turn with no visible answer, it now says so instead of leaving a blank bubble after minutes of work.

Find Opportunities only keeps workflows a automation can run through Plugins. Desktop control or browser-only plans are rejected.

Opportunity cards show when they were generated.

Learning pill and Learning button timers stay in sync after stop and start again.

On launch, Pacific Machines installs a automation into Cursor, Claude, and Codex (when those apps are present) so they can recap your local work history.

If that automation is missing or out of date, the next launch replaces it so it matches this build exactly.

Changelog now includes 0.1.14 (overlay window chrome), which was missing from the in-app history.

The main window uses overlay chrome: no title bar, and the close / minimize / zoom buttons sit on the sidebar like a hiddenInset Mac window.

Daily Recap hides those traffic lights while a story is open, then restores them when you leave.

Sidebar account menu and shell spacing are tighter so the frame feels like one piece of glass, not a titled browser window.

Staff Admin Panel lists people and organizations, with plan, trial, and spend.

Open a person to see every org they belong to, with that org’s opportunities and reports in the same card.

Person detail can edit account fields, org role, and trial, and add or remove membership. Deleting an organization stays on the organizations table.

The Chat button on opportunity and report headers shows the Chat label and the same icon as the sidebar.

The Learning pill only stays on screen while capture is actually recording. Find Opportunities and other runs no longer leave it up.

Staff Admin Panel lists every organization next to people, with members, trial, and spend.

Click a person in the Admin Panel to open their detail page: every org, with opportunities and reports in each.

Staff person detail is editable: account fields, org role, trial, add or remove membership, and delete an organization.

Opportunity cards use a red New badge, and a green Ready badge when the workflow is ready to build.

How often on opportunity cards is the number of times we saw the workflow, not Weekly or Daily.

Clicking a process step opens one tight Markdown writeup of what happened and the capture evidence, instead of Takes / Makes / Seen blocks.

Opportunity Summary, confidence, and ready writeups are full readable Markdown: short paragraphs, clear subheads, and bullets only where they help.

Find Opportunities can save Thin items. A 50% confidence gate was rejecting every insert, so the board stayed empty even when work was found.

Find Opportunities walks every captured day and emits every automatable work workflow, not a shortlist of the best two. Thin items stay on the board and now explain why they need more proof, or why they are ready.

Opportunity cards show how many times we saw the workflow, impact, and a confidence bar. Detail scores use percentages only when they mean something (confidence), and frequency, importance, time, and evidence stay as plain labels.

Opportunity confidence is a real per-item score. Thin stays a low percent, but repeated work with a complete process can show Solid or Strong instead of every card reading 44%.

Opportunity details show the captured process as an n8n-style map. Click a step to open that step only in a side inspector. A second map, titled If this gets automated, appears only when automation would change the workflow. Find Opportunities writes replayable procedures instead of 8-word captions.

Find Opportunities uses the account preferred model instead of always forcing Opus 5.

The default preferred model is Gemini 3.5 Flash, not Opus 5.

Opportunity details put the date above the title, list used apps in their own section, and explain the automation in a short paragraph instead of bullets.

Feedback on opportunities and reports opens in a dialog instead of a popover.

Convex feedback pipeline: structured answers (usefulness, accuracy, missing steps) plus freeform notes, keyed per user, artifact, and target version.

record-feedback automation so Ask can save critiques when you correct, dismiss, or praise a report or opportunity in chat.

Feedback popover in report and opportunity headers, plus inline panel at the bottom of each document.

Opportunity story player with cover, workflow map, per-run evidence scenes, score dashboard, and feedback scene.

Opportunity notes drawer and chat controls that stay visible while the Ask dock is open.

Evidence gap callouts when a workflow needs more capture runs, with concrete requirements before confidence can rise.

Opportunity scoring, evidence timelines, visual analytics, and insight helpers on the detail page.

In app Privacy Policy page (capture paths, retention, cloud use) and live PII redaction preview on sample text.

Modular onboarding: permissions, job role, company tool stack, and activation loop reveal; progress survives macOS screen-recording restarts.

Changelog page in the account menu with per-version history and your installed build.

Generate report dialog restored with last-7-days default, period comparison hints, and clearer errors while LLM settings load.

Redesigned opportunity list cards with clearer status, evidence hints, and story entry points.

Daily recap story player refreshed to match the opportunity story visual system.

Ask waits for macOS desktop-control grants before loading computer-use extensions, so chat no longer blocks on permission prompts.

Per evidence run yes/no confirmations on opportunity feedback for repeated workflow accuracy.

New and Viewed badges on reports, opportunities, and daily recaps after first open.

Feedback button shows a label, sits left of Chat, scrolls inside the popover, and places Send below the note field.

Report recommendation bullets render markdown bold instead of raw asterisks.

generate-report and team-report automations expanded with capture-grounded prompts and feedback hooks.

Report generation no longer silently fails when the generate-report dialog was missing from the build.

Dialog viewport wiring so report and analyze-range modals render and focus correctly.

Onboarding emits permission grant mix events (screen recording, accessibility, notifications, etc.) to PostHog.

Analytics helper fires grant-mix updates only when the mix changes, not on every permission poll.

Staff Users table shows each person's last reported desktop app version and when it was updated.

Convex user sync reports Tauri getVersion on sign-in so staff can audit builds without opening each machine.

Drive DMG ship automation records the 0.1.9 file id so the next public build adds instead of overwriting.

PostHog project token and host fetched from Convex instead of being baked into the desktop bundle.

Staff Users page App version column synced from the desktop client on sign-in.

Convex user row stores appVersion and appVersionUpdatedAt whenever the desktop app reports its build.

Organizations Convex module extended to support staff directory and version metadata.

Drive DMG ship automation records the 0.1.8 file id for versioned archive uploads.

Team report generation accepts one member with a personal report in the selected date range.

Team synthesis prompt and validation updated so solo teams get a full narrative, not a validation error.

team-report automation documents the single-member path and expected output shape.

Drive DMG ship automation records the 0.1.7 file id for the versioned archive workflow.

Daily recap and report agents query ~/.jarbas capture SQLite directly and cite frame timestamps and on-screen text.

generate-report and daily-recap automations tightened to reflect measured activity, not assumed schedules.

Pi analysis pipeline passes capture context into report and recap generation runs.

Report and recap pipelines reject placeholder timelines when capture rows are missing for the selected range.

Agents no longer fabricate end-to-end weekly narratives when capture density is thin.

Drive DMG ship automation records the 0.1.6 file id for versioned public builds.

Daily recap output is a structured written summary (title, narrative, apps, timeline) saved locally per calendar day.

Recap story player for stepping through recap scenes inside the Daily Recap tab.

daily-recap automation prompts for end to end readable copy instead of bullet-only output.

Daily Recaps page simplified around the new written recap model and story entry.

Windows CI publishes versioned NSIS installers to the correct artifacts path.

Windows CI tolerates GitHub Actions artifact quota failures without failing the whole release job.

Drive DMG ship automation records the 0.1.5 file id for the versioned archive workflow.

Cloud-synced opportunities in Convex with local migration from legacy on-disk files.

Staff Users pages and org admin directory for Pacific Machines operators.

PostHog analytics identity sync, org-aware events, and deduped team-report and automation-creation tracking.

Opportunity evidenceMoments with frame timestamps, OCR snippets, and multi-image screenshot paths.

Report opportunity refs: thin links from personal reports back to matching opportunities in the same period.

Report and recap UI leads with timeline chapters before aggregate KPIs and recommendations.

Opportunity list downgrades or hides high-confidence labels until evidence score and consistency thresholds pass.

Opportunity reading layout tightened; opening a report scrolls from the top instead of mid-document.

Report list polish, everyday wording in disclosures, and clearer With Agent copy.

Executive brief uses lead sentence plus bullets; app logos render consistently in report headers.

Reports table refactor with pagination and cleaner list-mine queries for org members.

Ask account context sync and staged snapshot prefix so cloud opportunity lists recover mid-chat.

Clerk org membership re-verified hourly; stored memberships expire after 24h without Clerk confirmation.

Membership revoked when Clerk says the user left the org; JWT org-claim lag tolerated via stored fallback.

Opportunity and report create, update, and delete require proven org membership, not just file ownership.

listMine and staff directory gated on membership; self-read checks cannot bypass missing org claims.

Opportunity dedupe trusts localId first so two items sharing jobId and title are not collapsed silently.

Legacy migration skips .DS_Store and system files instead of failing forever.

Opportunities run view clears on save failure or empty results instead of spinning forever.

Cross-org writes blocked; migration data loss and PII leak paths from pre-prod review closed.

Range data reset covers unmigrated legacy opportunities; upsert_analysis_item permission added.

Drive DMG ship automation records the corrected 0.1.4 installer id for versioned uploads.

macOS updater .tar.gz excludes AppleDouble (._*) metadata that prevented Tauri from unpacking signed bundles.

In-app Update can install notarized builds without manual DMG reinstall after this repair.

Drive DMG ship automation records the 0.1.3 file id so the next public DMG is archived, not overwritten.

Drive upload workflow writes PacificMachines_0.1.x_aarch64.dmg without overwriting the previous DMG.

Notarized DMG filename derives from package.json version instead of a fixed 0.1.0 label.

DMG ship automation documents patch increment on every public build and records file ids for the archive.

Check for updates in Settings using the Tauri updater plugin against the public release manifest.

Download progress and automatic relaunch after a signed update installs.

Story surface reading experience for reports and opportunities with chapter navigation and serif document typography.

Daily Recaps tab on the same story visual system as reports and opportunities.

Resizable document chat rail on opportunity and report detail pages.

Learning HUD shows pause state for ignored apps and applies capture filter edits live.

Opportunity brief denser and easier to scan; process-map cards tightened; short one-off workflows skipped.

Opportunity stories use markdown bullets and everyday language instead of unlock jargon.

Report story chapters aligned with opportunity structure; agent stream panels stay readable during analysis.

App update button in the header; polish on opportunity and report story surfaces.

LLM provider API keys loaded from Convex env instead of persisting secrets in llm.json on disk.

Windows CI stages binaries on a GitHub release when Actions artifact quota is full.

Release workflow bumps patch version and publishes updater artifacts alongside the DMG.

Windows CI no longer depends solely on GitHub Actions artifacts, avoiding quota failures on publish.

DMG ship automation always increments patch so in-app Update can see a newer build on the next ship.