Release History

What Comes Next

dreamcontext ships first-class support for Claude Code — seven hooks; core sub-agents (the initializer and curator skill families, the dreamcontext-deep-research synthesis skill, explore, the three primary RemSleep specialists, plus conditional sleep-federation and sleep-migration specialists); optional pack agents. The core idea — structured files, pre-loaded context, consolidation cycles — is not tied to any one agent, and the architecture is designed so that others (Gemini CLI, Copilot, custom agents) can plug into the same _dream_context/ directory with their own integration layer.

How it got here. Each release widened the system along one axis:

• v0.1.x — the neuroscience-inspired memory core (bookmarks, decay tracking, warm knowledge, triggers, transcript distillation), the web dashboard, and the first skill packs.

• v0.2.0 — Council debates, the Brain graph, Obsidian integration, lightweight root-instruction installs, and the catalog-driven skill-pack CLI.

• v0.3.0 — the sleep-cycle fan-out: the single consolidation agent became three parallel specialists (sleep-tasks and sleep-state always, sleep-product conditionally), with the main agent orchestrating directly and no separate orchestrator (see The Sleep Cycle).

• v0.4.0 — the manifest-aware install/update system (the setup front door, .install-manifest.json, stale-file pruning), multi-product/monorepo support, BM25 memory recall (deterministic, zero new deps, hook-injected by default), the tightened 150-line anti-bloat ceiling, and a 3D Brain graph.

• v0.5.0 / v0.5.1 — the "persistent brain" positioning, the one-command install and in-session update nudge (see Install & Update), the goal-skill and multi-review orchestration packs (joining Council under one decompose-then-synthesize pattern), optional Haiku-assisted recall, a context gate redesigned to surface the full skill catalog instead of a pre-picked subset, and the dashboard security hardening — loopback bind, CSRF guard, CORS lockdown, path-traversal guard — shipped as the v0.5.1 fast-follow before the first public release (see Security model).

• v0.6.0 — the recall engine overhaul detailed in Memory Recall (BM25F field weighting, EN/TR stemming, a synonym table, score/rank decoupling — recall@1 68→85%, @3 82→95%), continuous capture (auto-digest + salience + a capture guard so machine-captured notes never outrank curated ones), and the backend control-plane (config/packs/version-check/vaults routes) that the desktop app is built on.

• v0.7.0 — recall v3 (link-aware ranking, a snapshot token budget, an agent-feedback loop, and the graphify spec), multi-people awareness (attribute work per person), PRD freshness tracking with features doctor, the Eisenhower matrix view, and Excalidraw boards as first-class knowledge (their extracted text indexed, never the scene JSON).

• v0.8.x — the part of the vision that used to live in this section: the desktop app (Tauri 2 multi-vault launcher, in-app onboarding, the Sleepy notch companion, and continuous no-notarization updates), federation (read-only live cross-vault recall; the copy-based sync inbox shipped then was pulled back to the roadmap), a pluggable task backend (local markdown or a ClickUp list behind one adapter), a faceted tag taxonomy with sleep-time maintenance, and a versioned migration system (registry + ledger + a sleep-migration specialist) that keeps installed projects' assets in step across releases.

• v0.9.0 — GitHub Issues as a second remote task backend (issue-body-as-task, dc:* status labels, soft-close on delete — the same pluggable adapter and sync engine as ClickUp), and the brain's own lifecycle skills: the interactive initializer bootstrap (now with hook auto-detection of a missing/sparse brain, a migration, or a large new docs source) and the curator refactor pass (the MOVE/MERGE/SPLIT/RENAME/RE-TYPE/RETIRE re-organization sleep won't do, plus an atomic knowledge merge) — see Core skills: bootstrap, refactor, deep-research, and maintain the brain.

• v0.9.1 — the first half of a richer task model, working across both cloud backends: task date ranges (a planned start alongside the due/end, start≤due, backlog mutual-exclusion, and an auto-stamped start on first in_progress), assignee resolution against the live member roster (fuzzy match, ambiguous-aborts, unmatched-warns via SyncReport.warnings[] — never the old silent fallback to the token owner), and three new time-axis dashboard views for the same tasks: a Timeline (Gantt), a Calendar, and an Activity heatmap.

• v0.9.2 — the second half: project task overrides (an overrides/task.md that redefines the task template and declares custom fields — with optional required hard-fail and ask-the-user flags — which sync to ClickUp native fields / GitHub labels-and-body-blocks and brief the agent on how to fill them), and a sprint-aware version filter in the dashboard (a current/planning/released sprint model backed by state/.active-version.json, with set-current and mark-complete actions, plus dashboard-side cloud-token entry and a dry-run provision preview).

• v0.10.0 — the violet rebrand (new logo, regenerated Tauri icon, a four-group sidebar: Workspace / Memory / Brain / Control Panel); Sleepy "Search & Ask" (instant BM25 local search with a zero-token extractive answer, plus a real multi-turn, read-only Claude Code conversation streamed over SSE — see The Dashboard and Security model); the tasks board rebuild (independently persisted saved views, two-pane include/exclude filtering, a Versions popover, card Properties, and an at-risk alert, backed by a shared/local persistence split — overrides/board.json vs state/board.local.json); the dreamcontext-deep-research core skill (corpus-native iterative research that fans out dreamcontext-explore); a one-shot taxonomy audit --fix bulk tag normalizer; and a salience/bookmark pipeline fix.

• v0.10.1 / v0.10.2 — the desktop-app release train and the in-app Agent surface that superseded the Sleepy dashboard page: a multi-session, split-pane terminal (per-pane tabs, ⌘D drag-to-split, minimize-to-corner dock, detached-DOM persistence so the PTY never remounts), a ⌘K command palette over live recall, image-drop → in-vault injection, and a first-run prerequisite installer for the Claude CLI / node-pty. Also: board version smart-buckets (Current / Backlog / Completed), the multi-assignee AvatarStack (per-person hue), an atomic sleep consolidation mutex with rescaled debt thresholds, timeline-gantt polish, and a 404-on-missing-asset fix that ends the stale-chunk MIME error. (v0.10.2 is a pure republish — no code delta over v0.10.1.)

• v0.10.5 — the GitHub task-image bridge (a local screenshot uploads to a dedicated dreamcontext-assets branch, content-sniffed by magic bytes and content-addressed, so it renders inline on the synced issue while the local task keeps its canonical path), the root-cause orphaned-dashboard-server fix (a parent-death watchdog that reaps the Node server on every app-exit path the Rust handler misses), the project's first CI workflow (ci.yml runs vitest on every push and PR — closing the gap that let a recall regression ship at v0.10.1), overflow-proof Excalidraw card rendering, and a knowledge board-folder grouping / embedded-asset resolution fix. It also lands the PO-authored OKR roadmap — an objective board (core/objectives/, dreamcontext roadmap objective) with many-to-many task links, a full-DAG forecast cascade, and target-vs-forecast slip detection — surfaced as an interactive dashboard Roadmap page: a forecast timeline with drag-to-reschedule live cascade, drag-to-link dependencies, a status-column board view, and an inline-editable detail panel.

• v0.18.0 — the Task Manager: a task's own Claude Code session pinned inside the task view, maintaining the document (revise / summarize / split / status-reconcile) via a new core task-manager skill, with the terminal DOM re-parented from the one global Agent surface so it survives navigation. Around it: anchored doc comments (select a span, batch comments, one Send delivers them to the session — quote-anchored so they survive a live rewrite), a live doc refresh with a git-style "changes this session" diff, and delegate-to-Claude from the board. The Agent dock went living — screen-true status classification, an asking/"needs you" state that shakes, chimes, and jumps the queue. The embedding index gained a second consumer: dreamcontext embed dedup, a semantic near-duplicate gate that makes sleep merge instead of duplicate. The excalidraw kit became atomic — 44 builders (charts, wireframes, real-proportion devices) drivable from pure JSON, with a 3-check render audit. Plus dreamcontext tasks version, a doctor warning when two projects share one task container (#184), and a sync fix so a push can no longer advance the pull watermark (#185).

• v0.20.x — the proactive learning layer (falsifiable theses with pre-registered predictions, a cited evidence ledger, arithmetic-derived confidence, re-tested every sleep cycle by a sleep-learn specialist, opt-in), funnel analytics in Lab (a routed multi-page view over funnel-set/v1: all-funnels comparison, per-funnel step lanes with drop badges, click-two-steps A→B conversion, and a significant-change collapse for long funnels), and the task-sync integrity closeout (tasks dedup for duplicate mirror families, deterministic map merging, strict corrupt-ledger reads that refuse rather than clobber, and a snapshot-guarded pull that re-links orphans instead of duplicating them). Plus Announcements (a What's New page and on-load popup), the patterns core skill, and a v0.20.1 patch for task sync at scale (bulk first-sync, live progress, lock-ownership hardening).

• v0.21.0 — the agent surface becomes a first-class app view. Chat left preview as a 12-state native renderer (transcript hover actions, tool cards with real diffs, permission and survey cards, sub-agent drill-in) on unified 44px chrome, and the agent now knows it is in Chat: boards render on a live canvas, media plays inline, backticked paths become chips, and a dream-actions fence becomes real buttons. Its answers can be things you use rather than read — a dream-view fence renders an interactive chart, a comparison page, or a pinned always-on-top checklist window that sends your filled-in list back as one message, under its own narrow Tauri capability. Chat became the standard agent screen; the raw terminal stayed as Terminal (legacy).

• v0.22.0 — Automations: scheduled headless runs the brain owns, disabled until you install the dispatcher and approve each job on this machine, private by default, and gated by a hash of exactly what will happen. Three things make them more than cron — they learn (a bounded playbook plus a twenty-entry lesson ledger, read before each run as untrusted notes that lose to the approved prompt), they are visible (automations session replays the headless run turn by turn, tool calls and failures included, with the same drill-in in the dashboard), and their notification carries the result rather than announcing that a file was written. Alongside: recall widened to all nine channels with a --level importance filter, and Lab custom scripts moved into a short-lived child process so an edited shared library is never stale behind an ESM cache.

• v0.23.1 — the brain learns who it is talking to, and travels a third as heavy. 1.user.md becomes a people/ roster with one constitution per person; the soul's accumulated rules move out into first-class patterns that can no longer be demoted into silence; and the session snapshot was rebuilt to load everything the agent needs in roughly a third of the characters. It ships as .1 because the .0 cut never reached anyone: review found six defects in the people-first migration — the one step that DELETES its own source file — all of them in places where a wrong answer is indistinguishable from a right one once 1.user.md is gone. The rule that came out of it is now enforced in the code: a migration may move prose, but it may never CHOOSE whose prose it is; anything ambiguous is handed to an agent as a task instead of guessed alphabetically. v0.23.1 also carries v0.22.0 (Automations), which was built but likewise never published, so a single upgrade from v0.21.0 delivers both. Alongside: Chat's answers got readable (a body-column leading, lists that no longer sit tighter than the prose around them, and a highlighter the agent can reach for when bold is already carrying structure), tool-call runs collapse again, and the skill classifies a PDF's text layer before handing pages to the vision model.

What's next. With the desktop app, its in-app Agent surface, federation, a two-backend task model, and the Sleepy/board/rebrand pass all shipped, the near-term road is depth: a read-only public web surface (safely share a single knowledge file or Excalidraw board over a tunnel), an Overview surface in the app, the accessibility/responsive/i18n polish still deferred from the beta, Windows/Linux desktop builds, and widening federation from live read toward a redesigned, opt-in offline-mirror mode (the one thing live read can't do) and richer cross-project awareness. The agent-integration surface keeps broadening too — the architecture is deliberately not tied to Claude Code, so Gemini CLI, Copilot, and custom agents can plug into the same _dream_context/ directory with their own integration layer.

The long-term vision: your whole project — knowledge, PRDs, tasks, roadmap, metrics — lives in your repo, structured and version-controlled, and any agent you choose to work with can pick it up. You, your team, and your agents all see the same big picture. The human stays in the loop. The context stays portable. The system gets better every session because consolidation enforces the discipline that makes that possible.

For setup, commands, and getting started, see the README.

---

Part of the dreamcontext deep dive — Home · README