blumi v0.7.0 — code intelligence + autonomy

blumi v0.7.0 — the code-intelligence + autonomy release. Local-first as always: every new capability is opt-in and the curl … | sh install + blumi tui / serve behave exactly as before.

Highlights

Code intelligence (F1–F7)

• code_explore — one tool returns a symbol's source + callers/callees + related symbols + change blast radius (no more stringing four lookups together).
• Semantic knowledge layer — opt-in cached LLM per-file summary + architecture layer + business domain, with a dependency-ordered guided tour surfaced in the TUI /knowledge overlay, the web Control Center, and the blugo Code tab. (knowledge.semantic, blumi knowledge summarize|explain.)
• Live code-graph auto-sync — opt-in watcher keeps the graph in step with edits (knowledge.autosync).
• Framework-aware routes — GET /users/:id indexed as graph nodes for Axum/Actix/Rocket, FastAPI/Flask/Django, Express/Fastify/NestJS, gin/echo/chi, Rails, Spring.
• Hash-anchored edits — FileEdit can replace by line range + a content hash FileRead surfaces (big edit-token savings, fails safe on a stale view).
• debug tool — batch lldb/gdb/pdb/dlv (breakpoints → backtrace → inspect).

Autonomy, routing & efficiency

• Hermes goal-oriented autonomy (P2) — durable goal + step-ledger, completion detection + stall backoff, RPL phase-2 worktree-sandbox branch simulation, goal-aware sub-agents.
• Effort tiers — --effort low|standard|high|xhigh scales reasoning depth, ceilings, RPL, fan-out, thinking, and fallback-chain length (standard = no-op).
• Provider fallback chains — a routing tier can be an ordered provider list; falls through on connect error/timeout.
• UltraCode-Shim — built-in ultracode provider preset for an effort=xhigh + adaptive-thinking envelope.

Stability & quality (P0 + audit)

• Per-tool + LLM-stream timeouts, stuck-turn watchdog, cancellable approvals + panic-safe TurnDone, broadcast-lag self-heal — a hung tool or stalled provider recovers instead of freezing the UI.
• Grid "no live peers" flapping fixed; /api/retrospect 401 fixed; child-process output capture capped (OOM guard) + LSP/persist/grep fixes + three races.
• Crush-grade TUI + web UI: styled approve / session / deny buttons.

Skills

• ~146 bundled SKILL.md skills, now including Addy Osmani's agent-skills SDLC pack (addy-*, 24 skills) and a refreshed taste-skill pack (taste-*). blumi skills list / blumi skills sync.

Install / upgrade

curl -fsSL https://raw.githubusercontent.com/ankurCES/blumi-cli/main/install.sh | sh

Existing installs pick up the new bundled skills on next launch (the skill bundle re-materializes to ~/.blumi/skills and never clobbers your own).

Full details in CHANGELOG.md.

blumi v0.6.0 🌸 — structural code graph + retrospective memory

Highlights

🕸 Structural code graph (opt-in)

A typed, scope-resolved code graph via tree-sitter for Rust, Python, Go, JavaScript, and TypeScript (--features code-graph + knowledge.graph.mode = "structural"). A unified code_graph tool — callers / callees / impact (the change blast radius) / implementers — surfaced via the CLI (blumi knowledge impact|callers|callees|implementers), the gateway (POST /api/knowledge/graph), the TUI /knowledge hot-spots, and blugo's Code-tab Impact sheet. Graph-aware code_search fills spare result slots with typed neighbors, and the RPL blast radius now folds in code-graph fan-in so editing a heavily-referenced file faces tighter review.

🔁 Retrospection — daily memory consolidation from chat

A background pass replays each session's new transcript only (differential, from a watermark — never re-reading the whole history) and distills durable learnings into long-term memory: dedup-merged, provenance-tagged, and diffused across the grid. Full observability + manual control — a per-node run-log, Run now (differential) / Rebuild (full re-process), and a per-node summary across the grid — in blugo's new Retro tab. (memory.retrospect, default on, every retrospect_hours.)

🧠 Memory fixes

• Pinned /goal — a standing objective re-injected each turn, surviving context rollovers.
• Fitness guards — reward / note_used only touch active memories.
• Opt-in conflict resolver — the sweep classifies same-topic memory pairs and supersedes the outdated side (memory.resolve_conflicts).

⏱ Reliability

• Auto-wake on context rollover (llm.wake_on_rollover) — long / autonomous turns no longer stall at the token ceiling after a compaction.

📱 blugo

• Control Center is now its own full screen (12 tabs incl. Code · Graph · Retro).
• Code Impact sheet + the Retro tab. App → 1.1.0+13.

Full details: CHANGELOG.md

blumi v0.5.0 🌸 — self-improving memory + RPL-Judgement

Bring your own everything. What you configure is what you get.

This release makes blumi's agent learn from outcomes and think before it acts.

🧠 Self-improving agent memory

• Probationary failure→fix learning — a guided recovery is a pending hypothesis until the same tool is observed to succeed (cross-step verify, now on by default); only verified fixes become recallable. No more optimistic guesses masquerading as fixes.
• Value-based fitness — memories carry a learned value (separate from retrieval utility): rewarded for being in context on productive steps, decayed on failures, corroborated across grid nodes. Eviction now keeps what's useful, not what's merely popular.
• Structure-aware recall — hub-suppression over the memory graph (generic "matches-everything" memories stop dominating) + recall seeded from recent turns.
• Curation everywhere — the SEDM consolidation/eviction/graph/evolution sweep now runs for standalone blumi run / blumi tui, not just the gateway.
• Conflict + diffusion groundwork — reversible supersede + conflict detection; diffusion shares highest-value memories and raises value on cross-node agreement.

⚖️ RPL-Judgement (opt-in) — "Raskolnikov's Psychological Loop"

An adversarial, regret-minimizing pre-execution loop. Before a high-blast tool batch runs, blumi maps its blast radius, an adversarial "Porfiry" judge must approve the plan (reject → re-plan), and after execution the Error Delta ("regret") is written back to memory to feed fitness. Off by default (rpl.enabled) — it trades extra LLM calls for fewer catastrophic actions. A standard agent maximizes success; an RPL agent minimizes regret.

📱 blugo (phone companion)

• New mascot logo — the armed-hornet "beekeeper" bursting from the blumi bloom (app icon + network-diagram hub).
• "Bloom" live wallpaper — the flower blooms into the logo and the wasp fades in; replays on fold-open (hinge sensor).

---

Local-first and source-built — install.sh builds from this tag. See the CHANGELOG and the wiki.

blumi v0.4.0 🌸

Added

• blugo Dispatch + FCM push — a Telegram-style way to chat with each node
  and get pinged when it replies, even with the app backgrounded. On by default
  on the LAN; zero config beyond dropping the Firebase files in place.
  • Dispatch surface — a Dispatch entry on the welcome diagram (per-node
    action) and a dedicated inbox screen (a row per saved gateway + a
    Broadcast channel). Each per-node thread is a dedicated, isolated
    session (separate from the workbench chat), reusing the full chat UI
    (markdown, tool/approval/plan cards, voice). Broadcast fans one message
    to every saved gateway and shows each node's reply card.
  • FCM (Firebase Cloud Messaging), HTTP v1 — the gateway pushes a reply
    preview to the phone on turn completion (dispatch and the main chat), so a
    backgrounded/killed app still gets notified. Gateway: a device-token
    registry (~/.blumi/fcm.json), POST /api/push/fcm/register|unregister, and
    an FCM v1 sender that mints a Google OAuth2 token from a service account
    (RS256 JWT, cached) — enabled automatically when
    ~/.blumi/fcm-service-account.json is present, a silent no-op otherwise.
    blugo: firebase_messaging integration that registers its token with
    every gateway and routes a notification tap to the right dispatch thread
    (graceful no-op without Firebase config, falling back to local notifications).
  • Notification status-bar icon — a purpose-built white, alpha-only
    flower-outline small icon (ic_stat_blumi), wired into both the OS-drawn
    FCM path (default_notification_icon + a rose default_notification_color)
    and the local-notifications plugin. Android masks small icons to their alpha
    channel, so the full-colour launcher icon previously collapsed to a white
    blob; it now shows the blumi flower.
  • Launch behaviour — a cold start (after the app is killed) now lands on
    the network-diagram menu instead of silently reconnecting to the last
    gateway; a warm resume (background→foreground) keeps the live state and
    returns you to the last screen. Tapping a saved node still reconnects
    instantly via its stored token.
  • Welcome hub = the blumi flower → Dispatch. Every node on the
    network-diagram menu now wears the eight-petal logo flower (a faithful
    port of blumi-logo.svg) instead of the old five-petal glyph — a large,
    prominent centre hub flanked by smaller satellite gateways. The centre hub
    opens Dispatch directly (adding a gateway has its own ＋ button; the now
    redundant top-right Dispatch button was removed). The open plays a short
    "Entering bluuuum mode…" splash — the same logo bloom spins on the
    current theme background while a brand glow blooms out from behind it and
    washes the screen in the Living-Rose gradient, which then recedes back to
    the dark background as the text fades in. Tap to skip; bypassed entirely
    under reduce-motion.
  • Concurrent dedicated sessions (gateway) — /api/chat/send,
    /api/messages, and /api/chat/stream accept an optional session_id so a
    client can drive a specific session concurrently with the active one
    (a small dispatch-session registry); the dispatch SSE is pinned and never
    follows workbench swaps. The session_id-less paths are unchanged.

Changed

• blugo — full UI/UX redesign. The phone app is rebuilt on a proper design
  system and gets a brand-new welcome experience.
  • Welcome = an interactive grid network diagram. Instead of a list+form,
    the connect screen now draws a radial hub-and-spoke map (native
    CustomPainter): this device at the hub, saved gateways orbiting it on
    gradient spokes, and auto-discovered (mDNS) gateways as dashed/dotted
    nodes with a ＋ badge. Each node carries a canvas-drawn Living-Rose flower
    glyph; a radar sweep animates while scanning. Tap a saved node →
    Connect · Edit · Delete-on-the-side; tap a discovered node → a connect
    sheet with the password auto-focused and name/host/port pre-filled from
    discovery; ＋ adds a gateway by IP. New AppController.editServer
    supports renaming and re-pointing a saved gateway.
  • Design-system foundation. A new lib/ui/kit/ — BlumiTokens
    (ThemeExtension: the Living-Rose ramp, status colors, AA-safe muted text,
    radii/spacing, brand gradient), a shared motion language (durations,
    PressableScale, staggered entrances, cross-fades, reduced-motion aware),
    and a reusable widget vocabulary (cards, section headers, status dots/pills,
    badges, gradient buttons, empty states, sheets/dialogs). theme.dart now
    attaches the tokens and full component themes (cards, inputs with a focus
    ring, buttons, chips, tabs, dialogs, sheets, snackbars, …) to all six
    palettes.
  • Every screen restyled. Home gets a gradient wordmark header, a
    ListView.builder transcript with RepaintBoundarys, accent-bordered chat
    bubbles, kit tool/approval/clarify/plan cards, an animated context meter, and
    a themed composer; connect↔home cross-fades. The control center gets a
    gradient header, leading-icon tabs grouped into Agent · Work · Grid ·
    Knowledge, uppercase section headers, and de-duplicated grid metrics
    (removed from Settings — the Grid tab owns them). The command palette moves
    onto the shared bottom-sheet with pressable rows; markdown code/inline blocks
    are token-ized (atom-one-dark kept for syntax).
  • Accessibility & perf: Semantics on icon nodes, AA-contrast muted text,
    reduced-motion short-circuits, RepaintBoundary on animated leaves.
  • Verified on the Pixel 9 Pro Fold; blugo build bumped to 1.0.0+3.

---

Full changelog: https://github.com/ankurCES/blumi-cli/blob/v0.4.0/CHANGELOG.md · Compare: v0.3.0...v0.4.0

blumi v0.3.0 🌸

Added

• TUI /knowledge + /memories overlays — the terminal UI reaches parity with
  the web Control Center's knowledge/memory views. /knowledge shows the code
  knowledge base (indexed files / symbols / vectors + per-source counts);
  /memories browses semantic long-term memory entries (namespace / kind /
  utility / hit-count, pinned marked ★). Both are read-only, any-key-to-close popups
  (mirroring /heal); they degrade to a notice when knowledge/memory is disabled.
  (The existing /memory still views the MEMORY.md / USER.md files.)

• Messaging gateway as a managed service — blumi gateway now mirrors
  blumi serve's service layer. blumi gateway run launches every configured
  transport (Telegram / Discord / Slack / WhatsApp) concurrently in one process; and
  install / uninstall / start / stop / status register it as a launchd
  (com.blumi.gateway) or systemd-user (blumi-gateway) service with auto-start +
  crash/reboot restart (KeepAlive · Restart=always). status reports the service
  state + which transports are configured; logs at ~/.blumi/gateway.log. The
  existing single-transport commands (blumi gateway telegram, …) still work for
  foreground use.

• Duplicate-bot guard (Telegram) — the Telegram gateway now detects a 409
  Conflict (another consumer polling the same token — e.g. a stray bot on a grid
  peer) and logs a loud warning to gateway.log, then backs off and keeps
  retrying. Previously a 409 parsed as an empty update batch and was invisible —
  the cause of silent double-replies when two nodes ran the same token.

• Telegram voice toggle — gateway.telegram.voice (off by default) gates the
  Telegram bot's voice handling: inbound voice-note transcription and spoken
  (TTS) replies. With it off, a voice note gets a short "voice is off" reply and the
  bot answers in text only. Set "gateway": { "telegram": { "voice": true } } to
  re-enable (still needs global voice.* configured). Note: this flips the prior
  always-on behavior — Telegram voice is now opt-in.

• Lifecycle hooks (Claude-Code-style) — two events, off by default:

  • user_prompt_submit runs shell commands when you submit a prompt; each
    command's stdout is injected as cache-safe background context for the turn (the
    prompt is piped to stdin, cwd = workspace, per-hook timeout).
  • pre_tool_use runs ahead of permission policy before a tool executes: the
    {tool, input} payload is piped to stdin and a non-zero exit blocks the call
    (stderr/stdout becomes the denial reason), while a spawn error or timeout fails
    open (allows) so a broken guardrail can't brick the agent. A matcher scopes a
    hook by tool name (substring; empty = all). Hooks are read on session build, so a
    newly added one applies on the next reload_self/restart.

• Completion notifications (notify, off by default) — when an autonomous
  run finishes (blumi loop or always-on discovery), blumi fans out a short alert
  to the channels you enable. First wave (server-side push — reaches you with no app
  open): an OS desktop notification (macOS osascript / Linux notify-send) and
  a proactive gateway-bot message (Telegram / Discord / Slack / WhatsApp) to a
  configured chat/channel, reusing the gateway.* credentials. Config:
  notify { enabled, on:[loop,discovery,turn], desktop, bot:{transport,target}, web_push }. blumi loop --notify still fires a one-off desktop notification even
  when notify is off. Second wave (live-stream surface, needs no config): a
  browser in-tab alert — when a turn you started finishes while the web tab is
  backgrounded, blumi flashes the title, badges the favicon, plays a short ping,
  and drops a click-to-focus toast (silent while the tab is focused); and a blugo
  phone notification (flutter_local_notifications) — a heads-up local
  notification when a turn completes while the app is backgrounded (Android 13+
  runtime permission requested on first launch). Final wave (notify.web_push):
  browser Web Push (VAPID + RFC 8291) — a keypair + subscription store under
  ~/.blumi/push.json, GET /api/push/key + POST /api/push/{subscribe,unsubscribe},
  a service-worker push handler, and an Enable button in the Control Center.
  Pushes fire on loop / discovery / turn completion to every subscribed
  browser. ⚠️ Browser Web Push requires a secure context (HTTPS or
  http://localhost), so over a plain-HTTP LAN it stays dormant until you add TLS.

• Web Control Center panels — the browser/phone Control Center gains four
  tabs over the new backends: routing (tiers + $ saved), entries
  (white-box memory: pin / edit / delete), discovery (always-on status +
  reports), and git (read-only status / diff / log). One dist rebuild.

• Workspace create/clone wizard (TUI) — /new-workspace <path> creates a
  folder (+ git init) and opens it; /clone-workspace <url> [dir] git-clones a
  repo and opens it. Both append to the workspace pane (extends /open-workspace).

• Web git panel (read-only) — GET /api/git/{status,diff,log} expose the
  workspace's git status / diff --stat / recent log (behind the gateway password)
  so the browser/phone can review what the agent changed. (Staging/commit + the
  React panel are follow-ups.)

• Smart (cost-aware) model routing (PilotDeck-inspired) — per turn, a fast
  heuristic (and, on ambiguous turns, a local "judge" model) picks a difficulty
  tier and routes to a light vs flagship model; delegated sub-agents default
  to the cheap tier. Config router (mode = off|heuristic|hybrid|judge,
  light/heavy/judge provider+model, heuristics, subagent_tier,
  prefer_grid_light); default off. Live toggle + savings view: TUI /route
  (per-tier counts + $ saved vs all-heavy), GET /api/route,
  Command::SetRouterMode. Model swaps are gated to tier changes (prompt-cache
  safe); the judge fails safe to the light tier.

• White-box memory editor — list / view / pin / delete / edit individual
  semantic-memory entries (not just the MEMORY.md/USER.md files). Pinned entries
  are exempt from SEDM eviction + consolidation; editing re-embeds + resyncs FTS5.
  POST /api/memory/{list,pin,delete,update} (migration 0007 adds a pinned
  column). Editing/pinning a user-namespace entry stays local (never diffuses).

• Always-on proactive discovery (PilotDeck-inspired; off by default) — the
  gateway periodically (gated by cadence / rate-limit / board-busy / open-cap)
  runs one read-only turn to surface candidate tasks, adds them to the board as
  Discovered: todos, and lands a redacted markdown report (~/.blumi/reports/) +
  an agent-namespace discovery memory. Config always_on
  (enabled/autonomy/cadence_secs/…). Surfaced via GET /api/always-on, a
  blumi serve status line, and the TUI /discoveries overlay. (Autonomous
  low-risk execution in a worktree/snapshot is a planned follow-up.)

• Per-task cost telemetry — each board Task now accumulates input_tokens /
  output_tokens / cost_usd (priced from the model's list price); blumi loop
  records the per-task token delta, surfaced in the TUI /board ($/task + total)
  and /api/tasks. The model price map moved to blumi_config::pricing so
  routing, per-task cost, and the TUI meter share one source of truth.

• Unified blumi serve + web UI — blumi serve already serves the embedded
  React UI; blumi web is now framed as a localhost dev shortcut (+ a --port
  flag), and the Web-UI URL is printed by serve pair / install / status.

• TUI /open-workspace — a file-browser popup to open any folder as a
  workspace: ↑/↓ move, → enter a folder, ←/backspace go up, space opens
  the highlighted folder as a workspace (keep browsing), enter opens + closes,
  esc cancels. Git repos are flagged; opened folders appear in the left
  workspace pane immediately and are persisted to recents.

• Grid-embed offload transport — embeddings.backend = "grid" now routes
  embedding to the strongest GPU peer via a GridEmbed hook + secret-authed
  POST /api/grid/embed, with a TTL-cached peer choice and a local fallback
  (a lean node degrades to FTS5 when no peer is up). Closes the v0.2.0 follow-up.

• Cross-step recovery confirmation — a guided recovery is marked verified
  only when the retried tool actually succeeds on a later step (ground truth, not
  just "a fix was suggested"); the confirmed fix's utility is reinforced. Toggle
  with heal.verify (the field's meaning is now cross-step confirmation, no LLM).

• TUI /heal overlay — a self-healing summary (recovery / evolution / proposal
  counts + recent items) via a new Store::heal_summary, alongside the existing
  inline ⚕ self-heal traces and the blugo Heal tab / /api/heal.

Fixed

• NVIDIA CUDA build on Linux (BLUMI_CUDA=1) — two issues:
  • Build: pin ort-sys to =2.0.0-rc.9 and restore --locked on the
    installer's CUDA path. ort's range dependency on ort-sys floated to rc.12
    on a non-locked resolve, whose download-binaries build is broken
    (TLS-feature / ureq mismatch).
  • Runtime: CUDA's ONNX Runtime is a shared lib, so cargo install (binary
    only) left libonnxruntime.so unresolvable → every blumi invocation failed
    with "error while loading shared libraries". The installer now ships the .so
    next to the binary (copy-dylibs + $ORIGIN rpath) and verifies the binary
    loads, auto-falling back to a lean (CPU) build otherwise — so a reinstall can
    never leave a binary that won't start.
    Apple CoreML builds were unaffected (statically linked, already --locked). For
    Linux GPU the reliable path remains a local server (Ollama) for LLM + embeddings.

---...