v0.2.37

Highlights

OpenCode is now a native Hlid provider

• Added native OpenCode session creation, observation, history loading, resuming, and reconciliation in Raven.
• Unified managed and external OpenCode server handling, including Windows and WSL routing.
• Exposed the approved Hlid tool surface to OpenCode sessions and improved provider-owned status and error reporting.

Easier agent and provider setup

• Added a provider-neutral Agent setup directory that discovers available agents and clearly shows what is selected, detected, or still needs setup.
• Added reviewed managed ACP installation flows with explicit Windows or WSL targets.
• Updated first-run setup so people can select an available agent, configure local AI, or safely leave agent setup for later.
• Surfaced provider capability health so stale, unavailable, and working provider states are easier to understand.

More capable Routines

• Added clearer fixed-interval scheduling in minutes, hours, days, or weeks.
• Added monthly schedules, including predictable handling for shorter months.
• Added archived Routine browsing and recovery; restored Routines remain disabled until intentionally re-enabled.
• Improved schedule activation, attention state, and recovery behavior.

Raven and presentation improvements

• Improved live-session switching and restoration across providers.
• Expanded agent and tool activity rendering, including subagent and task activity details.
• Stabilized the composer and balanced presentation behavior across performance profiles.
• Improved Automatic performance recommendations so capable systems are less likely to be placed in an overly constrained mode.

Provider and protocol maintenance

• Updated Codex protocol support, Claude Agent SDK, ACP SDK, and supporting build/test dependencies.
• Corrected OpenCode GPT-5.6 Sol pricing estimates and the associated stored Ledger data.
• Updated Hlid operating brief v2 so exact-reference and cross-provider ownership boundaries are delivered consistently.

Documentation

• Reorganized the user guide into focused guides for getting started, chat and sessions, providers and models, Vault and Relics, Routines, access, voice, and maintenance.
• Refined the README around agent discovery, local AI, onboarding, and Routine scheduling.
• Refreshed the tan-theme screenshots for first run, Watch, Raven, Forge, agent setup, local AI, Routines, Vault, and Relics.

Full Changelog: v0.2.23...v0.2.37

v0.2.23

What changed

• Reorganized the README into a clearer feature overview with linked sections for sessions, providers, voice, mobile access, security, development tools, automation, history, references, Obsidian, and Relics.
• Added constrained-device performance profiles and reduced Raven rendering work for lower-powered systems.
• Kept mobile composer context visible when pickers open.
• Hardened OpenCode server controls and Forge notification and navigation behavior.

Validation

• bun install --frozen-lockfile
• bun audit
• bun audit --audit-level=high --prod
• bun run validate
• bun run test:db
• bun run build:win

Full Changelog: v0.2.18...v0.2.23

v0.2.18

Highlights

• Import ChatGPT exports through Hlid's authenticated data API, including from Tailscale UI origins.

• Keep materially different OpenCode Bash commands visible as separate activity rows when OpenCode reuses one ACP tool ID.

Improvements and fixes

• Add History to Simple navigation between Library and Settings so imported sessions are directly accessible.

• Preserve genuine streamed Bash command construction as one activity row and leave non-OpenCode ACP providers unchanged.

Full Changelog: v0.2.17...v0.2.18

v0.2.17

What changed

• OpenCode can now run as a Hlid-managed server, share native sessions with Desktop, and hand resumable sessions into Raven.
• OpenCode plan files join Hlid's normal review flow, while shared settings and onboarding survive restarts more reliably.
• Ledger can import ChatGPT data exports as immutable, resumable conversation snapshots and now prices supported OpenCode usage.
• Local Ollama models can use their real context and tool capabilities instead of being forced into one shape.
• Simple Forge now surfaces updates, and the Claude 2.1.240 and Codex 0.149 integrations are current.
• WSL delegation, TLS browser routing, reasoning-only ACP replies, onboarding, and Sol promotional pricing all got cleanup along the way.

The README, user guide, and Forge screenshot are current for v0.2.17.

Full Changelog: v0.2.2...v0.2.17

v0.2.2

Hlid 0.2.2

Simple view keeps Home, Chat, Library, and Settings up front while the full Hlid toolset stays one click away.

• Add Guided and Custom first-start paths, including a safe starter workspace option.
• Add Simple and Full navigation with a compact More menu.
• Give Simple Forge a readiness page and shortcuts for common setup.
• Add a reviewed local AI setup checklist without bypassing existing confirmations.
• Refresh the README, user guide, and Forge screenshot for the new experience.
• Make transient UI tests use their real async flow and controlled timers so release CI is deterministic.

Full Changelog: v0.2.1...v0.2.2

v0.2.1

Fixes

• Read and extract DirectML runtime archives with Hlid's bundled ZIP implementation instead of requiring external tar or unzip executables.
• Restore verified DirectML runtime installation from Hlid release assets on packaged Windows builds.

Validation

• Verified the qualified v0.2.0 DirectML release asset end to end.
• Passed bun run validate and bun run build:win.

Full Changelog: v0.2.0...v0.2.1

v0.2.0

Highlights

• Add managed ACP installs, Windows-hosted Ollama with a scoped WSL relay, and release-delivered DirectML speech runtime downloads.
• Add managed Git worktrees for durable Hlid delegation and session-scoped public browser targets beside Project Preview.
• Refresh the README, user guide, and current Watch, Vault, Relics, Forge, and Raven mobile screenshots for 0.2.0.
• Align the production duplication gate with the existing 0.2% repository baseline, so further increases continue to fail CI.

Validation

• bun run validate
• bun run build:win
• git diff --check

Full Changelog: v0.1.92...v0.2.0

v0.1.92

Highlights

• Added provider-native approval and permission controls for Codex and Claude, including Codex auto-review, permission profiles, Claude automatic modes, spend controls, peer inbox review, and clearer subagent progress.
• Made OpenCode a first-class ACP provider with managed Windows and WSL installs, model visibility controls, session import, updates, and improved runtime continuity.
• Added optional, private push notifications for sessions and routines, with per-device categories, batching, pause controls, exact navigation, and stronger PWA recovery.
• Added Local Conversation and faster local neural speech with pronunciation testing, a release-built qualified DirectML runtime, and CPU fallback.
• Reworked Forge and navigation with plain-language defaults, configurable names, better capability discovery, and more dependable provider readiness.

Improvements and fixes

• Improved Raven history, mobile context, active subagent grouping, background work reconciliation, and provider switching.
• Hardened MCP status and structured questions across Codex, Claude, and ACP providers.
• Reduced notification and diagnostics overhead, bounded long histories, and expanded PWA, offline, and performance coverage.
• Made speech and notification CI fixtures deterministic and pinned the reproducible, GPU-qualified DirectML r2 runtime.
• Updated the user guide, example configuration, and every documentation screenshot for 0.1.92.

Full Changelog: v0.1.69...v0.1.92

v0.1.69

Highlights

• Added Codex Apps and connectors to Raven and Forge, with installed and available inventories, authentication and usability status, connector health, and supported provider-owned sign-in flows.
• Added realtime Codex dictation alongside recorded audio and Raven Live. Provider-generated images now stay in conversations and Relics, and supported Windows hosts can return interactive visualizations to Raven.
• Made sleeping Raven turns durable across normal Hlið restarts and surfaced them in Attention until their provider usage window resets.
• Expanded live Claude sessions with file checkpoint rewind previews, native skill refresh, MCP server configuration and runtime controls, per-server approval overrides, and MCP elicitation.
• Reworked Raven activity into bounded response cards with first-class task progress and separate live background activity for supported Claude and Codex work.
• Added live provider capability discovery, including exact hlid_help lookup, search, filtering, and revision-safe pagination.

Improvements and fixes

• Kept long Raven sessions responsive with paged history and activity, lazy tool payload reads, batched reconnect replay, faster mobile resume, and stable pinned scrolling after refresh.
• Improved Project Preview with retained-capture navigation and safer stop handling. File-change views now show per-file addition and deletion counts, and modal backdrops close consistently when the current action allows it.
• Fixed Claude cumulative cost accounting, historical cost rows, steering continuation, streamed reply spacing, and native context resets. Updated Terra and Luna pricing with an effective-dated repair.
• Added storage cleanup previews, lossless managed-image optimization, compacted historical payloads, lightweight database optimization, and a guarded full reclaim action.
• Scoped instruction files, provider controls, and capability reporting to the configured provider and actual host environment.
• Hardened child-process shutdown, realtime voice lifecycles, Claude session safeguards, Codex MCP paging, and compatibility with current Claude, Codex, ACP, Vite, and security dependency releases.

Full Changelog: v0.1.40...v0.1.69

v0.1.40

Highlights

• Added session-scoped Project Preview to Raven. Agents can start and manage a preview, inspect it at desktop, tablet, or mobile sizes, catch console and request failures, work from annotated feedback, and save approved captures.
• Added durable Hlið child sessions across Claude, Codex, and installed ACP providers. Children keep their own workspace, provider controls, permissions, transcript, usage, and Ledger history, with nested provenance, attention, steering, cancellation, and explicit restart continuation.
• Expanded Raven with cross-provider steering, native Claude workflow controls, live session switching, attention-aware navigation, and turn-linked context receipts with history and inspection.
• Added optional local neural read aloud with Kitten and Piper, alongside device and Microsoft voices. Packaged Whisper speech-to-text now includes reviewed CPU and Vulkan backends.
• Added exact workspace references to the shared @ picker and Forge controls for adding or removing Hlið's agent and vault MCP servers from Claude Desktop.

Improvements and fixes

• Kept Raven provider, model, effort, and permission controls stable across reloads, reconnects, and provider changes. Accepted steering stays visible, incomplete turns keep their usage, and model costs remain attached to the turn that produced them.
• Updated the Codex integration for CLI 0.146 and stopped optional account or rate-limit reads from disconnecting an active chat when they time out.
• Kept long sessions responsive, moved packaged WebSockets onto the UI origin, and moved slow WSL and metadata discovery work off the active connection path.
• Hardened Project Preview restoration, relay hydration, hot reloads, browser recovery, capture behavior, and process cleanup. Only previews still owned by the live manager return after a reconnect or session resume.
• Fixed both Hlið MCP attachment and Obsidian CLI launching when Claude Desktop starts the server with its rebuilt Windows environment.
• Made provider extension and managed-skill changes environment-scoped and recoverable without disturbing an active turn. Updated Umbod core to 0.0.12 and kept its analytics refresh work out of the main UI path.
• Reworked Forge around its real content width so descriptions, statuses, actions, tabs, and technical values stay readable on desktop, tablet, and mobile. The API Reference now loads through the server boundary and documents the current routes without pulling Node modules into the browser.
• Made release CI build and audit the reviewed Whisper runtime automatically on Windows 2022, preserve the pinned upstream license bytes, then pass that same-run archive and manifest directly into the Windows executable build.
• Refreshed the README, user guide, third-party notices, and tan screenshots for the current app.

Full Changelog: v0.0.223...v0.1.40