Changelog

What's new in Black Magic AI

Every release, every fix, every feature. Synced directly from the desktop app's CHANGELOG.md. Latest: v0.4.18 · 2026-04-21

v0.4.18

2026-04-21

Changed · 1

Changed

Agent gallery cards now look like agents, not clones. 0.4.16's cards all rendered the same generic Bot icon next to the name and showed the bare slug underneath — the grid was a wall of identical briefcase-ish tiles with technical kebab-case strings that told the user nothing. Each card now pulls its frontmatter icon: (Radar for GEO Analyst, Linkedin for LinkedIn Outreach, Calendar for Meeting Prep, etc) and renders it inside a rounded icon tile with a per-agent accent color. The slug is gone, replaced by a tagline pulled from the agent's own body — the first real prose line, so "Research Agent" reads "You are the Research + Chat agent — the default generalist." instead of "researcher". Hover adds a subtle lift; the picked agent keeps its flame ring. Cards feel like distinct agents now.

v0.4.17

2026-04-21

Fixed · 1

Fixed

Upgrade now opens a real progress window. Previous versions silently exited the app and spawned brew in the background — from the user's perspective the screen just went blank for 60 seconds, which looked broken and led to frantic re-clicking. Clicking "Upgrade and relaunch" now opens a Terminal.app window titled "BlackMagic AI — Upgrading" that tails the real-time brew log, so you watch download percentages and install steps scroll live. The Terminal auto-closes when brew finishes (flag-file handshake), and the app reopens on top. No more silent upgrade loop.

v0.4.16

2026-04-21

Changed · 1

Changed

Agent gallery is now always laid flat on the empty chat page. 0.4.15 put the gallery in the empty state but hid it the moment an agent was picked (starter prompts replaced it). The gallery now stays visible above the starter prompts in a 4-column grid, and the currently-picked agent gets a flame-colored ring so there's no confusion about which agent will handle the next message. Starter-prompts heading also names the agent ("Starter prompts for GEO Analyst") so the two sections don't feel disconnected.

v0.4.15

2026-04-21

Fixed · 1Changed · 2

Fixed

Upgrade loop + zombie daemons. Three linked bugs that left users stuck on an "out of date" prompt with 10+ dead daemon processes and a broken install. First, daemonProcess.kill() only ran on window-all-closed, which on macOS doesn't fire on app relaunches — so every auto-upgrade cycle piled another daemon onto the stack. Cleanup now runs on before-quit + will-quit with a SIGTERM → SIGKILL escalation. Second, the upgrade dialog closed instantly when clicked and showed no feedback, so users clicked again and spawned a second brew process that collided on the download lock; there's now an upgradeInProgress guard plus a native macOS notification ("Downloading the latest version. The app will reopen in about a minute.") so the click clearly did something. Third, the upgrade shell script now force-kills any lingering /Applications/BlackMagic AI.app processes and clears stale brew .incomplete download locks before running brew upgrade, so a killed prior attempt can't brick the next one.

Changed

Sidebar is less cluttered. Dropped the per-agent Team section entirely — ten near-identical rows bloated the nav and buried the primary Chat entry point. Agents are now picked from the chat surface itself: the Agent dropdown in the header and a gallery of agent cards in the chat empty state. Renamed the "Work" section to "Automations". Pruned "Agent roles", "Org tree", and "Knowledge graph" from the main sidebar (still reachable by URL if anyone relied on them).
Chat empty state now routes through agents. With no messages in the thread, the chat shows either a grid of agent cards ("Pick an agent") or — once one is selected — that agent's own starter prompts pulled live from its vault frontmatter. One click on a card swaps the routing agent; one click on a starter prompt fills the composer.

v0.4.14

2026-04-21

Changed · 2

Changed

Agent activity stream redesigned. The old "thinking…" bubble with its truncated 6-line tool log was replaced with a structured activity stream inspired by Tukwork's CoWork surface. Each tool call renders as its own row with a status icon (pending spinner / ✓ done / ⚠ error), the tool name in title case, a bordered chip for the load-bearing argument (filename, domain, prompt id), and a muted tail with the full path / URL. Reasoning-summary deltas render as italic … <thought> rows with a spark icon, so the user sees the model's rationale for the next tool instead of a parade of opaque names. A live Processing… Ns timer at the bottom keeps long autonomous runs from feeling stalled.
Assistant messages now have a Copy · View as Markdown footer. Renders below every finished assistant bubble: one-click copy of the full markdown source, plus "View as Markdown" which opens the raw text in a new tab (easier than trying to select pretty- rendered prose). Same footer pattern as the CoWork reference so shipping a long plan or report no longer requires a trip to the runs page.

v0.4.13

2026-04-21

Added · 1Fixed · 2

Added

Agent cockpit now shows reasoning between tool calls. Long autonomous runs (GEO Analyst, Pipeline Ops, Closed-Lost Revival) used to render as a silent parade of tool names — you could see which tools fired but not why. The daemon now forwards the model's reasoning-summary deltas through the chat stream, and the cockpit renders them as dimmed … <thought> lines interleaved with the → tool / ✓ tool markers. Gives you a live window on what the agent is deciding, not just what it's doing.

Fixed

Agent no longer exits silently after hitting turn budget. runAgent used to cap at 20 turns and, if the model was still mid-tool-chain, write an empty final.md and leave the chat showing (empty). The default budget is now 50 turns, agents can raise it via max_turns: frontmatter (GEO Analyst defaults to 100), and when the budget does run out the agent writes an explicit closing message listing the last few tools called and telling the user to re-run to pick up from the files already written. No more mystery silences — the 2026-04-21T09-17-21-931Z-geo-analyst case.
Real brand logos on integration cards. 0.4.11 shipped colored tiles with initials ("H", "A", "SF", "飞", "MB", …) as a placeholder. Swapped for actual brand SVGs (Simple Icons paths, MIT licensed) — HubSpot sprocket, Salesforce cloud, Slack hash, Gmail envelope, Feishu/Lark bubble, Metabase dot-chart, Supabase lightning. Each path renders in white on the brand's colored background tile.

v0.4.12

2026-04-21

Fixed · 2

Fixed

Manual edits to company/contact/deal profiles are now discoverable and restorable. Saving a profile from the vault editor stamps human_edited_at: into the frontmatter; the Companies list surfaces a flame-colored edited badge on those rows with a tooltip explaining that re-enrichment stashes a backup under .bm/backups/ before overwriting. The company detail drawer now also lists the last 20 backups for that file with links into the vault viewer, so recovering a clobbered manual note is one click instead of a terminal dive (QA BUG-03).
New `GET /api/vault/backups?path=...` endpoint that returns the list of timestamped backups the daemon has already been writing under .bm/backups/ for profile files. Used by the Companies detail drawer; also handy for external tooling.

v0.4.11

2026-04-21

Added · 2Changed · 3

Added

Feishu, Metabase, Supabase now show up as cards on the Integrations page. The daemon already accepted keys for all three (Settings → Integration keys), but the card UI was missing so users couldn't see that Feishu/Metabase/Supabase were wired. Feishu joins the Messaging group; Metabase + Supabase get their own new "Data" group.
Brand tile for every integration card. Each provider now shows a colored square with its short glyph (H / A / SF / G / U / S / M / 飞 / MB) next to the name, so cards are visually distinguishable at a glance instead of all looking the same.

Changed

Company Profiler is now a first-class Team agent. Moved out of Sidebar → System → "Profile company" (where it was buried under Integrations / Agent roles / Settings) into the Team section at the top of the list. Seeded as agents/company-profiler.md with pin: first frontmatter so it always sorts above the other nine agents. Icon is Sparkles.
Sidebar agent rows respect `href:` frontmatter. Agents with an explicit href in frontmatter route there instead of the generic /team?slug=X cockpit. Company Profiler uses this to jump straight to /onboarding/bootstrap (the bootstrap-self playbook that crawls your domain + docs and populates the us/ tree) — clicking it in Team takes you to the form, not a chat. Other agents still land on the chat cockpit.
Sort by `pin` then name. Team list sort now honors a pin: first frontmatter field before alpha-sorting, so pinned agents never get bumped down as new ones are added.

v0.4.10

2026-04-21

Changed · 3

Changed

All 10 agent system prompts rewritten autonomous-first. Every agent now executes a full READ → PLAN → ACT → SUMMARIZE cycle in one run instead of stopping midway to ask the user what to do. The three classic "I can't continue without X" halt points — missing ICP, empty signal file, missing personas — now auto-bootstrap with a best-effort default marked draft: true in frontmatter so a human can review later, and the agent keeps going. Halts are reserved for genuine hard blockers (missing API credential a tool can't run without, destructive action requiring confirmation, persistent upstream 5xx) and come with a one-line statement of the exact resolution needed.
GEO Analyst auto-bootstraps personas + brand config. Step 1 (personas < 2) now derives 2 draft personas from us/company.md + us/icp.md + us/customers/top.md. Step 2 (no brands or no is_us: true) infers from us/company.md + us/competitors/* and calls geo_set_brands. Step 3 (thin prompt pool) generates candidates per persona × 6 query types. The full 10-step loop now runs to the weekly report without interruption.
Every seeded agent gains `revision:` frontmatter. ensureVault now compares the template's revision vs the user's on-disk file and overwrites when newer. This retires stale peec_* tool lists on vaults seeded in 0.4.3 and ships prompt rewrites without requiring manual vault cleanup. Future edits only need a revision bump to propagate.

v0.4.9

2026-04-21

Fixed · 1Changed · 1

Fixed

Sidebar breathing light is now per-agent, not global. Earlier versions lit every nav item (Skills, Companies, Integrations, Settings…) whenever any run was live, which made the whole sidebar look busy for no reason. Breathing is now scoped to the specific Team row whose agent has a live run. Non-Team rows no longer pulse — the Runs row's liveCount pill covers "how many agents are running total".

Changed

Cockpit starter prompt asks agents to execute, not narrate. The default "Kick off <Agent>" starter used to say "walk me through the first thing you would do … reply with: (1) concrete first step, (2) tools you would call, (3) expected output". That reliably produced essays instead of work. New starter "Run <Agent> end-to-end" tells the agent to actually execute, only stopping for genuinely ambiguous human decisions, and to summarize what it wrote when done.

v0.4.8

2026-04-21

Fixed · 1

Fixed

Auto-upgrade now reads the Homebrew cask directly. Both the launch-time hard gate and the renderer's soft upgrade banner used to compare against the R2 version.json manifest. That manifest is only refreshed by ./scripts/release.sh, which means any 0.4.x version bumped in the repo but not packaged looked "up to date" to the installed app — exactly why 0.4.6 on disk never saw 0.4.7. Source of truth is now the tap's cask file at https://raw.githubusercontent.com/blackmagic-ai/homebrew-tap/main/Casks/blackmagic-ai.rb, parsed with a regex on version "X.Y.Z". No R2 dependency, cache- busted per check.

v0.4.7

2026-04-21

Added · 4Removed · 1

Added

GEO is now native. Drops the Peec AI integration and replaces it with a daily sweep that runs your seed-prompt pool through ChatGPT (gpt-5.2 + web_search), Perplexity Sonar, and Google AI Overview (via SerpAPI) — all proxied through blackmagic.run, so you never manage upstream keys. Credits are charged at OpenAI / Perplexity / SerpAPI list price + 10% markup (8¢ / 2¢ / 2¢ per call). Results land in signals/geo/runs/<date>/<model>/<prompt>.json.
`/geo` dashboard. Four "biggest mover" cards at the top (SoV up/down, new/lost domain). Brand SoV bars render the prior-period ghost behind the current bar + a signed Δ column with up/down/flat arrows. SoV line chart overlays the prior period as a dashed line so you see the week-over-week shape, not just the absolute. Domain data split into four tables: biggest gains, biggest losses, new this period, lost this period. Gap sources (domains cited when competitors are mentioned but not when you are) keep their own panel.
**Twelve new geo_* tools** wired into the rewritten GEO Analyst agent: geo_list_prompts, geo_add_prompt, geo_remove_prompt, geo_list_brands, geo_set_brands, geo_run_prompt, geo_run_daily, geo_report_brands, geo_report_domains, geo_gap_sources, geo_sov_trend, geo_list_runs.
Two triggers. geo-daily fires POST /api/geo/run at 07:00 every day to refresh the pool; geo-weekly runs the GEO Analyst agent every Monday 09:00 against the week's stored snapshots.

Removed

Peec AI integration. The nine peec_* tools are gone. The peec_api_key setting is retired — proxying through blackmagic.run means no per-user third-party key to manage.

v0.4.6

2026-04-21

Fixed · 1

Fixed

Agent cockpit client-side exception. The scenarios useMemo lived after the early-return guards in AgentCockpit, so hook order changed between "loading" and "loaded" renders and React crashed the page with "a client-side exception has occurred". Moved the memo above the guards so hook order stays stable.

v0.4.5

2026-04-21

Changed · 3

Changed

Agent cockpit is now an interactive workspace. Clicking an agent used to land on a read-only "what it can do / tools / runs" page with a separate "Chat with X" button that page-jumped you away. Now the cockpit embeds ChatSurface inline: chat on the left (scoped to that agent, its own thread), skills + recent runs + tools in the right rail. No more hop to /Chat, no more "what do I do" confusion — the chat input is the primary action.
Empty state points at chat. When an agent has zero skills (like the new GEO Analyst), the "Skills" rail says "just chat on the left" instead of the old confusing "Add a playbook with agent: <slug>" message. A default starter prompt ("Kick off <Agent>…") is seeded into the empty chat so there's always one obvious next move.
Chat agent picker switches threads. Selecting a different agent now swaps the thread to that agent's own history (bm-team-thread-<slug>). The picker label reads "Agent:" with tooltip "Switch agents — each one has its own thread so you can run many in parallel". Swapping back to Default (Research Agent) restores the global chat thread.

v0.4.4

2026-04-21

Fixed · 1

Fixed

"Upgrade and relaunch" relaunch loop. The auto-upgrade script ran brew upgrade --cask blackmagic-ai without first refreshing brew's tap cache, so brew saw no newer version, exited 0, and the app relaunched into the same "out of date" dialog forever. The script now runs brew update --quiet first, passes --greedy (bypass brew's cask-auto-update heuristics), and if the installed version still doesn't match the latest after the upgrade, force a brew reinstall --cask blackmagic-ai. Log at ~/Library/Logs/BlackMagic AI/auto-upgrade-*.log keeps a diff of installed vs latest so the next stuck upgrade is visible at a glance.

v0.4.3

2026-04-21

Added · 4Changed · 1

Added

GEO Analyst agent (agents/geo-analyst.md), seeded to every vault. Executes the 12-step Generative Engine Optimization loop from the internal GEO PRD — ICP lock → Seed Query audit → prompt expansion → (Peec fan-out) → 6-field response parsing → 6-metric scoring → Gap Source analysis → Owned/Earned/Paid recommendations → draft distribution → 48h source-drop alerts → pipeline attribution → weekly report. English-only scope (ChatGPT, Google AI Overviews, Perplexity, Gemini, Claude, Copilot, Grok). Model: gpt-5.4.
Peec AI integration (BYOK, X-API-Key, base https://api.peec.ai/customer/v1). Nine tools wired into the daemon: peec_list_brands, peec_list_prompts, peec_create_prompt, peec_prompt_suggestions, peec_accept_prompt_suggestion, peec_report_brands (SoV / Citation Rank / sentiment / visibility), peec_report_domains (cited sources — the highest-leverage GEO data, drives Gap analysis), peec_report_urls, peec_source_content. Key goes in Settings → Integrations → Peec AI. Peec's API is Enterprise-tier + beta per their docs; errors surface verbatim.
`geo-weekly` trigger (Monday 07:00 cron). Fires the geo-analyst, pulls the week's Peec snapshot, diffs against last week's run, writes the weekly report to signals/geo/weekly/<iso-week>.md, and drops 48h source-drop alerts into signals/geo/alerts/.
`signals/geo/{runs,weekly,actions,alerts}` directories in the vault skeleton so weekly snapshots, reports, action lists, and source-drop alerts each have their own home.

Changed

Triggers now accept an `agent:` frontmatter field. Previously triggers either ran a shell command, a playbook, or fell back to the hardcoded researcher agent. With agent: geo-analyst (or any agent slug) a cron can now fire any agent directly.

v0.4.2

2026-04-21

Removed · 1Changed · 1

Removed

Demo brand names in Chat default scenarios. DEFAULT_SCENARIOS still held six GTM prompt templates referencing acme.com, beta-corp, jane@acme.com even though the rendered scenario cards were dropped in 0.3.4 — the array still shipped in the bundle and its length > 0 was what gated the "What do you want to do? / Reference files in your vault with [[wikilinks]]" copy. Array is now empty; Chat's empty state is blank unless a caller passes its own scenarios. The Team cockpit's agent-scoped scenarios are unaffected.

Changed

Empty-state renders scenario cards when callers provide them. For callers that do pass scenarios (e.g. agent-specific chats), the cards come back as a tidy grid, without the generic "What do you want to do?" title on top.

v0.4.1

2026-04-21

Fixed · 1

Fixed

Chat 呼吸灯 now persists through the whole run. The thinking bubble previously had a lone animate-pulse dot that vanished the moment streaming text started, so during long replies the UI looked idle. Upgraded to the same animate-ping halo the sidebar and cockpit use, and pinned a second breathing dot next to the Chat title that stays on for the full duration of sendMut.isPending — streaming phase included.

v0.4.0

2026-04-21

Added · 2Changed · 1

Added

Agent cockpit at /team?slug=<agent>. Clicking an agent in the Team sidebar now lands on a purpose-built page instead of opening the raw .md file or a duplicate chat. Shows identity + tools, status strip (last run · skill count · live indicator), the playbooks that agent owns as runnable skill cards (with inline inputs), and the 8 most recent runs by this agent. Header CTA deep-links into the main Chat with the agent preselected via /?agent=<slug>. Works for any agent file in the vault — no more 404 on dynamic slugs under static export.
Chat agent picker. The main Chat header now has a small dropdown populated from agents/*.md. Pick "auto (researcher)" or route the next message to any specific agent (Deal Manager, LinkedIn Outreach Agent, etc.). Setting it from the URL (/?agent=<slug>) preselects for deep links.

Changed

Friendlier display names for the three base agents. ae → "Deal Manager", researcher → "Research Agent", sdr → "Outreach Agent". Slugs stay stable so existing playbooks (agent: researcher, etc.) keep working. A migration rewrites existing vaults' frontmatter only when the name: field still matches the bare slug, so user renames are preserved.

v0.3.12

2026-04-21

Added · 1

Added

Seed six GTM persona agents into every vault. The vault-backed sidebar in 0.3.11 correctly showed whatever was in agents/*.md, which for most projects meant just ae, researcher, sdr. Now every vault also seeds website-visitor.md, linkedin-outreach.md, meeting-prep.md, lookalike-discovery.md, closed-lost-revival.md, and pipeline-ops.md with real system prompts, tool allowlists, and Lucide-icon frontmatter. Existing vaults pick them up on next daemon start (missing-files-only write). Users can edit or delete any of them — sidebar reflects whatever is on disk.

v0.3.11

2026-04-21

Changed · 1

Changed

Sidebar Team is now vault-backed. The Team section used to be a hardcoded Swan-style GTM list (Website Visitor Agent, LinkedIn Outreach Agent, …). It now reads agents/*.md from the active project's vault and lists whatever role files exist there. Each entry links to the agent's definition in the vault editor so you can inspect/edit the system prompt directly. Falls back to the old hardcoded list only if the vault has no agents at all.

v0.3.10

2026-04-21

Added · 1

Added

Sidebar breathing light while a run is live. Every nav item (Skills, Sequences, Triggers, Companies, Contacts, Deals, Org, Knowledge graph, Files, Integrations, Agent roles, Settings, Team agents) now shows the same animate-ping flame dot that the Runs row uses whenever liveCount > 0. Gives the whole sidebar a shared "呼吸灯" pulse so it's obvious at a glance that an agent is working, even when you've navigated away from /runs.

v0.3.9

2026-04-21

Changed · 1

Changed

Auto-upgrade on out-of-date launch. The launch-time version gate used to offer only "Copy command and quit". It now offers Upgrade and relaunch when Homebrew is on disk — the app spawns a detached shell that waits for the current process to exit, runs brew upgrade --cask blackmagic-ai, then open -a "BlackMagic AI" to reopen automatically. Output goes to ~/Library/Logs/BlackMagic AI/auto-upgrade-*.log so a failed upgrade leaves a breadcrumb. Homebrew path is probed at /opt/homebrew/bin/brew and /usr/local/bin/brew; on systems without brew we fall back to the previous copy-command dialog.

v0.3.8

2026-04-20

Added · 1Fixed · 1

Added

End-to-end LinkedIn outreach loop — new universal skill li-campaign-loop + sequence linkedin-post-signal + preset trigger linkedin-daily-outreach (weekday 09:00). The loop gathers LinkedIn engagement signal, picks the top prospects, enriches each contact, calls draft_create for both linkedin_connect and linkedin_dm, enrolls the contact into the LinkedIn sequence, and writes a summary note to signals/linkedin/<date>-loop.md. Replaces the Apidog-era LinkedIn trigger that only printed usage output. Available in every project — seeded on vault creation and installable into existing vaults via Triggers → Install presets.

Fixed

Chat tool stream — → list_dir / ✓ read_file lines now include the load-bearing argument (path, url, domain, query, …) so users can see what the agent is actually touching instead of staring at an opaque list of tool names. The daemon forwards the call arguments on both pending + completed events.

v0.3.7

2026-04-20

Fixed · 1

Fixed

Project picker first-launch dismissal (BUG-005 follow-up) — the picker now always renders as a modal with Escape / outside-click / ✕ dismissal, including the very first app launch. Previously 0.3.0 only made the picker dismissable when re-opened from the sidebar; the first-run page-mode fallback was non-dismissable.

v0.3.5

2026-04-20

Fixed · 2

Fixed

Chat — "(empty)" bubble no longer renders next to "thinking…". The placeholder assistant message is hidden while the request is in flight; the thinking bubble owns that slot.
Codex progress — file_read, web_fetch, patch, and every other codex step now surface in the "working…" list so the UI no longer looks stuck during multi-step runs.

v0.3.4

2026-04-20

Added · 1Changed · 1Fixed · 1

Added

Dark mode by default. Light mode is still available via the sidebar toggle.

Changed

Chat empty state — dropped the hardcoded GTM scenario cards. Non-GTM vaults saw them as noise. Keeps a minimal header + wikilinks hint.

Fixed

Sidebar version label — daemon reports the real BM_APP_VERSION instead of a hardcoded 0.3.0.

v0.3.3

2026-04-20

Fixed · 1

Fixed

Gatekeeper launch crash — afterPack signs every Electron Framework helper (chrome_crashpad_handler, etc.) explicitly before sealing the framework bundle. Resolves "BlackMagic AI cannot be opened because of a problem" on 0.3.0–0.3.2.

v0.3.0

2026-04-20

Fixed · 13

Fixed

Companies — strict domain regex gate; malformed domains can no longer start a run. (BUG-001)
Daemon `writeVaultFile` — timestamped backup under .bm/backups/ before overwriting any companies/, contacts/, or deals/ profile, so retries can't silently clobber manual edits. (BUG-002)
Sidebar live-runs badge — tighter filter (done flag, turns, tokensOut, 2-minute ceiling) plus refetch on window focus. (BUG-003)
Chat — reconcile the last assistant message to the server's final payload on done. (BUG-004)
Project picker — opens as a modal with Escape and outside-click dismissal. (BUG-005)
Vault + Settings — show the active vault path. (BUG-006)
Playbooks — Run Playbook disabled while required inputs are empty or malformed. (BUG-007)
Agents — inline "New agent" form replaces window.prompt(). (BUG-008)
Versioning unified at 0.3.0 across workspaces. (BUG-009)
Contacts empty state with CTAs to Companies + Chat. (BUG-010)
Org tree counts only files. (BUG-011)
Knowledge graph shows match count + "no match" feedback. (BUG-012)
Sequences Run walker now reports enrollments · fired · failed. (BUG-013)

v0.2.10

2026-04-20

Fixed · 1

Fixed

Chat textarea now grows based on content length on every input change (typing, scenario clicks, post-send clear).

v0.2.9

2026-04-20

Added · 2Fixed · 1

Added

Force upgrade on launch — every launch checks the R2 manifest and blocks with a brew-upgrade dialog if the app isn't the latest.
Integrations: Feishu / Metabase / Supabase. - Feishu: feishu_notify (custom bot webhook), feishu_send_message (tenant token DM / group), feishu_bitable_list_records. - Metabase: metabase_run_card, metabase_query_sql, metabase_search. - Supabase: supabase_select, supabase_insert, supabase_update, supabase_rpc.

Fixed

Sidebar live-runs badge — honours the new done flag from /api/agent/runs so finished runs stop pulsing live.

v0.2.8

2026-04-20

v0.2.7

2026-04-20

Added · 4Fixed · 3

Added

Company Profiling Agent (onboarding wizard) at /onboarding/bootstrap. Takes a domain (+ optional docs URL + extras), runs the bundled bootstrap-self playbook, populates the us/ tree.
Apollo (direct API) + Attio (full toolkit). - Apollo: apollo_search_people, apollo_enrich_person, apollo_organization_search. Replaces the Apify actor scrape. - Attio: attio_search_records, attio_create_record (with upsert), attio_update_record, attio_create_note, attio_add_to_list.
Upgrade banner — R2 version.json check on launch; if a newer version exists, renderer shows a top banner with a one-click-copy brew upgrade command.
Sidebar → System → Profile company entry.

Fixed

login-gate hydration — SSR and first client render no longer diverge on bridge presence.
/api/chat streaming — the builtin Responses fallback streams token-by-token via SSE instead of returning one JSON payload. Event shapes normalised ({ delta }, { message }).
Distribution — dropped electron-updater and Windows builds; updates ship exclusively through brew cask.

v0.2.6

2026-04-19

Fixed · 3Added · 1

Fixed

Chat 404 "endpoint not available on API subdomain" — middleware accepts both /v1/* (canonical) and /api/v1/* (legacy) on api.blackmagic.run.
Gatekeeper "app is damaged" — Homebrew cask under blackmagic-ai/tap; ad-hoc signing in afterPack; postflight runs xattr -cr as a backstop.
Billing — gpt-5.3-codex aliased to gpt-5.4 upstream so older daemons keep working. Credits no longer expire on monthly refresh.

Added

Dashboard "Install and launch" panel with copyable install + launch + upgrade brew commands.

Stay in the loop

Black Magic AI ships daily. Install the desktop app to get these updates automatically — or check the roadmap to see what's next.

Download desktop app See roadmap

Changelog

What's new in Black Magic AI

Every release, every fix, every feature. Synced directly from the desktop app's CHANGELOG.md. Latest: v0.4.18 · 2026-04-21

v0.4.18

2026-04-21

Changed · 1

Changed

Agent gallery cards now look like agents, not clones. 0.4.16's cards all rendered the same generic Bot icon next to the name and showed the bare slug underneath — the grid was a wall of identical briefcase-ish tiles with technical kebab-case strings that told the user nothing. Each card now pulls its frontmatter icon: (Radar for GEO Analyst, Linkedin for LinkedIn Outreach, Calendar for Meeting Prep, etc) and renders it inside a rounded icon tile with a per-agent accent color. The slug is gone, replaced by a tagline pulled from the agent's own body — the first real prose line, so "Research Agent" reads "You are the Research + Chat agent — the default generalist." instead of "researcher". Hover adds a subtle lift; the picked agent keeps its flame ring. Cards feel like distinct agents now.

v0.4.17

2026-04-21

Fixed · 1

Fixed

Upgrade now opens a real progress window. Previous versions silently exited the app and spawned brew in the background — from the user's perspective the screen just went blank for 60 seconds, which looked broken and led to frantic re-clicking. Clicking "Upgrade and relaunch" now opens a Terminal.app window titled "BlackMagic AI — Upgrading" that tails the real-time brew log, so you watch download percentages and install steps scroll live. The Terminal auto-closes when brew finishes (flag-file handshake), and the app reopens on top. No more silent upgrade loop.

v0.4.16

2026-04-21

Changed · 1

Changed

Agent gallery is now always laid flat on the empty chat page. 0.4.15 put the gallery in the empty state but hid it the moment an agent was picked (starter prompts replaced it). The gallery now stays visible above the starter prompts in a 4-column grid, and the currently-picked agent gets a flame-colored ring so there's no confusion about which agent will handle the next message. Starter-prompts heading also names the agent ("Starter prompts for GEO Analyst") so the two sections don't feel disconnected.

v0.4.15

2026-04-21

Fixed · 1Changed · 2

Fixed

Upgrade loop + zombie daemons. Three linked bugs that left users stuck on an "out of date" prompt with 10+ dead daemon processes and a broken install. First, daemonProcess.kill() only ran on window-all-closed, which on macOS doesn't fire on app relaunches — so every auto-upgrade cycle piled another daemon onto the stack. Cleanup now runs on before-quit + will-quit with a SIGTERM → SIGKILL escalation. Second, the upgrade dialog closed instantly when clicked and showed no feedback, so users clicked again and spawned a second brew process that collided on the download lock; there's now an upgradeInProgress guard plus a native macOS notification ("Downloading the latest version. The app will reopen in about a minute.") so the click clearly did something. Third, the upgrade shell script now force-kills any lingering /Applications/BlackMagic AI.app processes and clears stale brew .incomplete download locks before running brew upgrade, so a killed prior attempt can't brick the next one.

Changed

Sidebar is less cluttered. Dropped the per-agent Team section entirely — ten near-identical rows bloated the nav and buried the primary Chat entry point. Agents are now picked from the chat surface itself: the Agent dropdown in the header and a gallery of agent cards in the chat empty state. Renamed the "Work" section to "Automations". Pruned "Agent roles", "Org tree", and "Knowledge graph" from the main sidebar (still reachable by URL if anyone relied on them).
Chat empty state now routes through agents. With no messages in the thread, the chat shows either a grid of agent cards ("Pick an agent") or — once one is selected — that agent's own starter prompts pulled live from its vault frontmatter. One click on a card swaps the routing agent; one click on a starter prompt fills the composer.

v0.4.14

2026-04-21

Changed · 2

Changed

Agent activity stream redesigned. The old "thinking…" bubble with its truncated 6-line tool log was replaced with a structured activity stream inspired by Tukwork's CoWork surface. Each tool call renders as its own row with a status icon (pending spinner / ✓ done / ⚠ error), the tool name in title case, a bordered chip for the load-bearing argument (filename, domain, prompt id), and a muted tail with the full path / URL. Reasoning-summary deltas render as italic … <thought> rows with a spark icon, so the user sees the model's rationale for the next tool instead of a parade of opaque names. A live Processing… Ns timer at the bottom keeps long autonomous runs from feeling stalled.
Assistant messages now have a Copy · View as Markdown footer. Renders below every finished assistant bubble: one-click copy of the full markdown source, plus "View as Markdown" which opens the raw text in a new tab (easier than trying to select pretty- rendered prose). Same footer pattern as the CoWork reference so shipping a long plan or report no longer requires a trip to the runs page.

v0.4.13

2026-04-21

Added · 1Fixed · 2

Added

Agent cockpit now shows reasoning between tool calls. Long autonomous runs (GEO Analyst, Pipeline Ops, Closed-Lost Revival) used to render as a silent parade of tool names — you could see which tools fired but not why. The daemon now forwards the model's reasoning-summary deltas through the chat stream, and the cockpit renders them as dimmed … <thought> lines interleaved with the → tool / ✓ tool markers. Gives you a live window on what the agent is deciding, not just what it's doing.

Fixed

Agent no longer exits silently after hitting turn budget. runAgent used to cap at 20 turns and, if the model was still mid-tool-chain, write an empty final.md and leave the chat showing (empty). The default budget is now 50 turns, agents can raise it via max_turns: frontmatter (GEO Analyst defaults to 100), and when the budget does run out the agent writes an explicit closing message listing the last few tools called and telling the user to re-run to pick up from the files already written. No more mystery silences — the 2026-04-21T09-17-21-931Z-geo-analyst case.
Real brand logos on integration cards. 0.4.11 shipped colored tiles with initials ("H", "A", "SF", "飞", "MB", …) as a placeholder. Swapped for actual brand SVGs (Simple Icons paths, MIT licensed) — HubSpot sprocket, Salesforce cloud, Slack hash, Gmail envelope, Feishu/Lark bubble, Metabase dot-chart, Supabase lightning. Each path renders in white on the brand's colored background tile.

v0.4.12

2026-04-21

Fixed · 2

Fixed

Manual edits to company/contact/deal profiles are now discoverable and restorable. Saving a profile from the vault editor stamps human_edited_at: into the frontmatter; the Companies list surfaces a flame-colored edited badge on those rows with a tooltip explaining that re-enrichment stashes a backup under .bm/backups/ before overwriting. The company detail drawer now also lists the last 20 backups for that file with links into the vault viewer, so recovering a clobbered manual note is one click instead of a terminal dive (QA BUG-03).
New `GET /api/vault/backups?path=...` endpoint that returns the list of timestamped backups the daemon has already been writing under .bm/backups/ for profile files. Used by the Companies detail drawer; also handy for external tooling.

v0.4.11

2026-04-21

Added · 2Changed · 3

Added

Feishu, Metabase, Supabase now show up as cards on the Integrations page. The daemon already accepted keys for all three (Settings → Integration keys), but the card UI was missing so users couldn't see that Feishu/Metabase/Supabase were wired. Feishu joins the Messaging group; Metabase + Supabase get their own new "Data" group.
Brand tile for every integration card. Each provider now shows a colored square with its short glyph (H / A / SF / G / U / S / M / 飞 / MB) next to the name, so cards are visually distinguishable at a glance instead of all looking the same.

Changed

Company Profiler is now a first-class Team agent. Moved out of Sidebar → System → "Profile company" (where it was buried under Integrations / Agent roles / Settings) into the Team section at the top of the list. Seeded as agents/company-profiler.md with pin: first frontmatter so it always sorts above the other nine agents. Icon is Sparkles.
Sidebar agent rows respect `href:` frontmatter. Agents with an explicit href in frontmatter route there instead of the generic /team?slug=X cockpit. Company Profiler uses this to jump straight to /onboarding/bootstrap (the bootstrap-self playbook that crawls your domain + docs and populates the us/ tree) — clicking it in Team takes you to the form, not a chat. Other agents still land on the chat cockpit.
Sort by `pin` then name. Team list sort now honors a pin: first frontmatter field before alpha-sorting, so pinned agents never get bumped down as new ones are added.

v0.4.10

2026-04-21

Changed · 3

Changed

All 10 agent system prompts rewritten autonomous-first. Every agent now executes a full READ → PLAN → ACT → SUMMARIZE cycle in one run instead of stopping midway to ask the user what to do. The three classic "I can't continue without X" halt points — missing ICP, empty signal file, missing personas — now auto-bootstrap with a best-effort default marked draft: true in frontmatter so a human can review later, and the agent keeps going. Halts are reserved for genuine hard blockers (missing API credential a tool can't run without, destructive action requiring confirmation, persistent upstream 5xx) and come with a one-line statement of the exact resolution needed.
GEO Analyst auto-bootstraps personas + brand config. Step 1 (personas < 2) now derives 2 draft personas from us/company.md + us/icp.md + us/customers/top.md. Step 2 (no brands or no is_us: true) infers from us/company.md + us/competitors/* and calls geo_set_brands. Step 3 (thin prompt pool) generates candidates per persona × 6 query types. The full 10-step loop now runs to the weekly report without interruption.
Every seeded agent gains `revision:` frontmatter. ensureVault now compares the template's revision vs the user's on-disk file and overwrites when newer. This retires stale peec_* tool lists on vaults seeded in 0.4.3 and ships prompt rewrites without requiring manual vault cleanup. Future edits only need a revision bump to propagate.

v0.4.9

2026-04-21

Fixed · 1Changed · 1

Fixed

Sidebar breathing light is now per-agent, not global. Earlier versions lit every nav item (Skills, Companies, Integrations, Settings…) whenever any run was live, which made the whole sidebar look busy for no reason. Breathing is now scoped to the specific Team row whose agent has a live run. Non-Team rows no longer pulse — the Runs row's liveCount pill covers "how many agents are running total".

Changed

Cockpit starter prompt asks agents to execute, not narrate. The default "Kick off <Agent>" starter used to say "walk me through the first thing you would do … reply with: (1) concrete first step, (2) tools you would call, (3) expected output". That reliably produced essays instead of work. New starter "Run <Agent> end-to-end" tells the agent to actually execute, only stopping for genuinely ambiguous human decisions, and to summarize what it wrote when done.

v0.4.8

2026-04-21

Fixed · 1

Fixed

Auto-upgrade now reads the Homebrew cask directly. Both the launch-time hard gate and the renderer's soft upgrade banner used to compare against the R2 version.json manifest. That manifest is only refreshed by ./scripts/release.sh, which means any 0.4.x version bumped in the repo but not packaged looked "up to date" to the installed app — exactly why 0.4.6 on disk never saw 0.4.7. Source of truth is now the tap's cask file at https://raw.githubusercontent.com/blackmagic-ai/homebrew-tap/main/Casks/blackmagic-ai.rb, parsed with a regex on version "X.Y.Z". No R2 dependency, cache- busted per check.

v0.4.7

2026-04-21

Added · 4Removed · 1

Added

GEO is now native. Drops the Peec AI integration and replaces it with a daily sweep that runs your seed-prompt pool through ChatGPT (gpt-5.2 + web_search), Perplexity Sonar, and Google AI Overview (via SerpAPI) — all proxied through blackmagic.run, so you never manage upstream keys. Credits are charged at OpenAI / Perplexity / SerpAPI list price + 10% markup (8¢ / 2¢ / 2¢ per call). Results land in signals/geo/runs/<date>/<model>/<prompt>.json.
`/geo` dashboard. Four "biggest mover" cards at the top (SoV up/down, new/lost domain). Brand SoV bars render the prior-period ghost behind the current bar + a signed Δ column with up/down/flat arrows. SoV line chart overlays the prior period as a dashed line so you see the week-over-week shape, not just the absolute. Domain data split into four tables: biggest gains, biggest losses, new this period, lost this period. Gap sources (domains cited when competitors are mentioned but not when you are) keep their own panel.
**Twelve new geo_* tools** wired into the rewritten GEO Analyst agent: geo_list_prompts, geo_add_prompt, geo_remove_prompt, geo_list_brands, geo_set_brands, geo_run_prompt, geo_run_daily, geo_report_brands, geo_report_domains, geo_gap_sources, geo_sov_trend, geo_list_runs.
Two triggers. geo-daily fires POST /api/geo/run at 07:00 every day to refresh the pool; geo-weekly runs the GEO Analyst agent every Monday 09:00 against the week's stored snapshots.

Removed

Peec AI integration. The nine peec_* tools are gone. The peec_api_key setting is retired — proxying through blackmagic.run means no per-user third-party key to manage.

v0.4.6

2026-04-21

Fixed · 1

Fixed

Agent cockpit client-side exception. The scenarios useMemo lived after the early-return guards in AgentCockpit, so hook order changed between "loading" and "loaded" renders and React crashed the page with "a client-side exception has occurred". Moved the memo above the guards so hook order stays stable.

v0.4.5

2026-04-21

Changed · 3

Changed

Agent cockpit is now an interactive workspace. Clicking an agent used to land on a read-only "what it can do / tools / runs" page with a separate "Chat with X" button that page-jumped you away. Now the cockpit embeds ChatSurface inline: chat on the left (scoped to that agent, its own thread), skills + recent runs + tools in the right rail. No more hop to /Chat, no more "what do I do" confusion — the chat input is the primary action.
Empty state points at chat. When an agent has zero skills (like the new GEO Analyst), the "Skills" rail says "just chat on the left" instead of the old confusing "Add a playbook with agent: <slug>" message. A default starter prompt ("Kick off <Agent>…") is seeded into the empty chat so there's always one obvious next move.
Chat agent picker switches threads. Selecting a different agent now swaps the thread to that agent's own history (bm-team-thread-<slug>). The picker label reads "Agent:" with tooltip "Switch agents — each one has its own thread so you can run many in parallel". Swapping back to Default (Research Agent) restores the global chat thread.

v0.4.4

2026-04-21

Fixed · 1

Fixed

"Upgrade and relaunch" relaunch loop. The auto-upgrade script ran brew upgrade --cask blackmagic-ai without first refreshing brew's tap cache, so brew saw no newer version, exited 0, and the app relaunched into the same "out of date" dialog forever. The script now runs brew update --quiet first, passes --greedy (bypass brew's cask-auto-update heuristics), and if the installed version still doesn't match the latest after the upgrade, force a brew reinstall --cask blackmagic-ai. Log at ~/Library/Logs/BlackMagic AI/auto-upgrade-*.log keeps a diff of installed vs latest so the next stuck upgrade is visible at a glance.

v0.4.3

2026-04-21

Added · 4Changed · 1

Added

GEO Analyst agent (agents/geo-analyst.md), seeded to every vault. Executes the 12-step Generative Engine Optimization loop from the internal GEO PRD — ICP lock → Seed Query audit → prompt expansion → (Peec fan-out) → 6-field response parsing → 6-metric scoring → Gap Source analysis → Owned/Earned/Paid recommendations → draft distribution → 48h source-drop alerts → pipeline attribution → weekly report. English-only scope (ChatGPT, Google AI Overviews, Perplexity, Gemini, Claude, Copilot, Grok). Model: gpt-5.4.
Peec AI integration (BYOK, X-API-Key, base https://api.peec.ai/customer/v1). Nine tools wired into the daemon: peec_list_brands, peec_list_prompts, peec_create_prompt, peec_prompt_suggestions, peec_accept_prompt_suggestion, peec_report_brands (SoV / Citation Rank / sentiment / visibility), peec_report_domains (cited sources — the highest-leverage GEO data, drives Gap analysis), peec_report_urls, peec_source_content. Key goes in Settings → Integrations → Peec AI. Peec's API is Enterprise-tier + beta per their docs; errors surface verbatim.
`geo-weekly` trigger (Monday 07:00 cron). Fires the geo-analyst, pulls the week's Peec snapshot, diffs against last week's run, writes the weekly report to signals/geo/weekly/<iso-week>.md, and drops 48h source-drop alerts into signals/geo/alerts/.
`signals/geo/{runs,weekly,actions,alerts}` directories in the vault skeleton so weekly snapshots, reports, action lists, and source-drop alerts each have their own home.

Changed

Triggers now accept an `agent:` frontmatter field. Previously triggers either ran a shell command, a playbook, or fell back to the hardcoded researcher agent. With agent: geo-analyst (or any agent slug) a cron can now fire any agent directly.

v0.4.2

2026-04-21

Removed · 1Changed · 1

Removed

Demo brand names in Chat default scenarios. DEFAULT_SCENARIOS still held six GTM prompt templates referencing acme.com, beta-corp, jane@acme.com even though the rendered scenario cards were dropped in 0.3.4 — the array still shipped in the bundle and its length > 0 was what gated the "What do you want to do? / Reference files in your vault with [[wikilinks]]" copy. Array is now empty; Chat's empty state is blank unless a caller passes its own scenarios. The Team cockpit's agent-scoped scenarios are unaffected.

Changed

Empty-state renders scenario cards when callers provide them. For callers that do pass scenarios (e.g. agent-specific chats), the cards come back as a tidy grid, without the generic "What do you want to do?" title on top.

v0.4.1

2026-04-21

Fixed · 1

Fixed

Chat 呼吸灯 now persists through the whole run. The thinking bubble previously had a lone animate-pulse dot that vanished the moment streaming text started, so during long replies the UI looked idle. Upgraded to the same animate-ping halo the sidebar and cockpit use, and pinned a second breathing dot next to the Chat title that stays on for the full duration of sendMut.isPending — streaming phase included.

v0.4.0

2026-04-21

Added · 2Changed · 1

Added

Agent cockpit at /team?slug=<agent>. Clicking an agent in the Team sidebar now lands on a purpose-built page instead of opening the raw .md file or a duplicate chat. Shows identity + tools, status strip (last run · skill count · live indicator), the playbooks that agent owns as runnable skill cards (with inline inputs), and the 8 most recent runs by this agent. Header CTA deep-links into the main Chat with the agent preselected via /?agent=<slug>. Works for any agent file in the vault — no more 404 on dynamic slugs under static export.
Chat agent picker. The main Chat header now has a small dropdown populated from agents/*.md. Pick "auto (researcher)" or route the next message to any specific agent (Deal Manager, LinkedIn Outreach Agent, etc.). Setting it from the URL (/?agent=<slug>) preselects for deep links.

Changed

Friendlier display names for the three base agents. ae → "Deal Manager", researcher → "Research Agent", sdr → "Outreach Agent". Slugs stay stable so existing playbooks (agent: researcher, etc.) keep working. A migration rewrites existing vaults' frontmatter only when the name: field still matches the bare slug, so user renames are preserved.

v0.3.12

2026-04-21

Added · 1

Added

Seed six GTM persona agents into every vault. The vault-backed sidebar in 0.3.11 correctly showed whatever was in agents/*.md, which for most projects meant just ae, researcher, sdr. Now every vault also seeds website-visitor.md, linkedin-outreach.md, meeting-prep.md, lookalike-discovery.md, closed-lost-revival.md, and pipeline-ops.md with real system prompts, tool allowlists, and Lucide-icon frontmatter. Existing vaults pick them up on next daemon start (missing-files-only write). Users can edit or delete any of them — sidebar reflects whatever is on disk.

v0.3.11

2026-04-21

Changed · 1

Changed

Sidebar Team is now vault-backed. The Team section used to be a hardcoded Swan-style GTM list (Website Visitor Agent, LinkedIn Outreach Agent, …). It now reads agents/*.md from the active project's vault and lists whatever role files exist there. Each entry links to the agent's definition in the vault editor so you can inspect/edit the system prompt directly. Falls back to the old hardcoded list only if the vault has no agents at all.

v0.3.10

2026-04-21

Added · 1

Added

Sidebar breathing light while a run is live. Every nav item (Skills, Sequences, Triggers, Companies, Contacts, Deals, Org, Knowledge graph, Files, Integrations, Agent roles, Settings, Team agents) now shows the same animate-ping flame dot that the Runs row uses whenever liveCount > 0. Gives the whole sidebar a shared "呼吸灯" pulse so it's obvious at a glance that an agent is working, even when you've navigated away from /runs.

v0.3.9

2026-04-21

Changed · 1

Changed

Auto-upgrade on out-of-date launch. The launch-time version gate used to offer only "Copy command and quit". It now offers Upgrade and relaunch when Homebrew is on disk — the app spawns a detached shell that waits for the current process to exit, runs brew upgrade --cask blackmagic-ai, then open -a "BlackMagic AI" to reopen automatically. Output goes to ~/Library/Logs/BlackMagic AI/auto-upgrade-*.log so a failed upgrade leaves a breadcrumb. Homebrew path is probed at /opt/homebrew/bin/brew and /usr/local/bin/brew; on systems without brew we fall back to the previous copy-command dialog.

v0.3.8

2026-04-20

Added · 1Fixed · 1

Added

End-to-end LinkedIn outreach loop — new universal skill li-campaign-loop + sequence linkedin-post-signal + preset trigger linkedin-daily-outreach (weekday 09:00). The loop gathers LinkedIn engagement signal, picks the top prospects, enriches each contact, calls draft_create for both linkedin_connect and linkedin_dm, enrolls the contact into the LinkedIn sequence, and writes a summary note to signals/linkedin/<date>-loop.md. Replaces the Apidog-era LinkedIn trigger that only printed usage output. Available in every project — seeded on vault creation and installable into existing vaults via Triggers → Install presets.

Fixed

Chat tool stream — → list_dir / ✓ read_file lines now include the load-bearing argument (path, url, domain, query, …) so users can see what the agent is actually touching instead of staring at an opaque list of tool names. The daemon forwards the call arguments on both pending + completed events.

v0.3.7

2026-04-20

Fixed · 1

Fixed

Project picker first-launch dismissal (BUG-005 follow-up) — the picker now always renders as a modal with Escape / outside-click / ✕ dismissal, including the very first app launch. Previously 0.3.0 only made the picker dismissable when re-opened from the sidebar; the first-run page-mode fallback was non-dismissable.

v0.3.5

2026-04-20

Fixed · 2

Fixed

Chat — "(empty)" bubble no longer renders next to "thinking…". The placeholder assistant message is hidden while the request is in flight; the thinking bubble owns that slot.
Codex progress — file_read, web_fetch, patch, and every other codex step now surface in the "working…" list so the UI no longer looks stuck during multi-step runs.

v0.3.4

2026-04-20

Added · 1Changed · 1Fixed · 1

Added

Dark mode by default. Light mode is still available via the sidebar toggle.

Changed

Chat empty state — dropped the hardcoded GTM scenario cards. Non-GTM vaults saw them as noise. Keeps a minimal header + wikilinks hint.

Fixed

Sidebar version label — daemon reports the real BM_APP_VERSION instead of a hardcoded 0.3.0.

v0.3.3

2026-04-20

Fixed · 1

Fixed

Gatekeeper launch crash — afterPack signs every Electron Framework helper (chrome_crashpad_handler, etc.) explicitly before sealing the framework bundle. Resolves "BlackMagic AI cannot be opened because of a problem" on 0.3.0–0.3.2.

v0.3.0

2026-04-20

Fixed · 13

Fixed

Companies — strict domain regex gate; malformed domains can no longer start a run. (BUG-001)
Daemon `writeVaultFile` — timestamped backup under .bm/backups/ before overwriting any companies/, contacts/, or deals/ profile, so retries can't silently clobber manual edits. (BUG-002)
Sidebar live-runs badge — tighter filter (done flag, turns, tokensOut, 2-minute ceiling) plus refetch on window focus. (BUG-003)
Chat — reconcile the last assistant message to the server's final payload on done. (BUG-004)
Project picker — opens as a modal with Escape and outside-click dismissal. (BUG-005)
Vault + Settings — show the active vault path. (BUG-006)
Playbooks — Run Playbook disabled while required inputs are empty or malformed. (BUG-007)
Agents — inline "New agent" form replaces window.prompt(). (BUG-008)
Versioning unified at 0.3.0 across workspaces. (BUG-009)
Contacts empty state with CTAs to Companies + Chat. (BUG-010)
Org tree counts only files. (BUG-011)
Knowledge graph shows match count + "no match" feedback. (BUG-012)
Sequences Run walker now reports enrollments · fired · failed. (BUG-013)

v0.2.10

2026-04-20

Fixed · 1

Fixed

Chat textarea now grows based on content length on every input change (typing, scenario clicks, post-send clear).

v0.2.9

2026-04-20

Added · 2Fixed · 1

Added

Force upgrade on launch — every launch checks the R2 manifest and blocks with a brew-upgrade dialog if the app isn't the latest.
Integrations: Feishu / Metabase / Supabase. - Feishu: feishu_notify (custom bot webhook), feishu_send_message (tenant token DM / group), feishu_bitable_list_records. - Metabase: metabase_run_card, metabase_query_sql, metabase_search. - Supabase: supabase_select, supabase_insert, supabase_update, supabase_rpc.

Fixed

Sidebar live-runs badge — honours the new done flag from /api/agent/runs so finished runs stop pulsing live.

v0.2.8

2026-04-20

v0.2.7

2026-04-20

Added · 4Fixed · 3

Added

Company Profiling Agent (onboarding wizard) at /onboarding/bootstrap. Takes a domain (+ optional docs URL + extras), runs the bundled bootstrap-self playbook, populates the us/ tree.
Apollo (direct API) + Attio (full toolkit). - Apollo: apollo_search_people, apollo_enrich_person, apollo_organization_search. Replaces the Apify actor scrape. - Attio: attio_search_records, attio_create_record (with upsert), attio_update_record, attio_create_note, attio_add_to_list.
Upgrade banner — R2 version.json check on launch; if a newer version exists, renderer shows a top banner with a one-click-copy brew upgrade command.
Sidebar → System → Profile company entry.

Fixed

login-gate hydration — SSR and first client render no longer diverge on bridge presence.
/api/chat streaming — the builtin Responses fallback streams token-by-token via SSE instead of returning one JSON payload. Event shapes normalised ({ delta }, { message }).
Distribution — dropped electron-updater and Windows builds; updates ship exclusively through brew cask.

v0.2.6

2026-04-19

Fixed · 3Added · 1

Fixed

Chat 404 "endpoint not available on API subdomain" — middleware accepts both /v1/* (canonical) and /api/v1/* (legacy) on api.blackmagic.run.
Gatekeeper "app is damaged" — Homebrew cask under blackmagic-ai/tap; ad-hoc signing in afterPack; postflight runs xattr -cr as a backstop.
Billing — gpt-5.3-codex aliased to gpt-5.4 upstream so older daemons keep working. Credits no longer expire on monthly refresh.

Added

Dashboard "Install and launch" panel with copyable install + launch + upgrade brew commands.

Stay in the loop

Black Magic AI ships daily. Install the desktop app to get these updates automatically — or check the roadmap to see what's next.

Download desktop app See roadmap

Changelog - blackmagic.run