github.com/ktsoator/or

Added

• Added durable plan mode to coding sessions. Plan mode is persisted in session
  transcripts, injected into model requests as a planning policy, and requires
  explicit review through exit_plan_mode before execution. Plans are presented
  on a dedicated client review surface with plan mode controls, slash
  commands, and composer integration.
• Added durable todo checklists. The todo_write tool records whole-list
  snapshots as typed tool outcomes, checklist state survives session
  restoration, and current-turn progress is shown above the composer and kept
  in sync across conversation views.
• Added durable provider request checkpoints. Versioned request headers are
  committed at each provider checkpoint and model input is rebuilt from
  committed projections before dispatch, while credentials and other
  secret-bearing fields are excluded from durable reconstruction. Interrupted
  JSONL appends are rolled back to their previous byte boundary and repaired
  before retry.
• Added progress updates to the coding agent prompt. The agent now reports
  meaningful milestones during tool-driven work instead of narrating each tool
  call, while preserving the update protocol for custom base instructions.
• Isolated workspace previews and browser popups. Workspace files are served
  from a preview-only loopback origin, preview and web guests run in dedicated
  Electron partitions, and external navigation plus window.open requests are
  routed into the existing browser workspace.

Changed

• Session transcripts now require version 8. Transcripts from earlier releases
  are rejected and are not migrated automatically.
• Refreshed the built-in model catalog from models.dev to 425 models, with 382
  runnable models, and updated the English and Chinese support matrices.

Fixed

• Preserved browser tabs and conversation drafts across workbench
  conversations. User-opened browser tabs stay mounted when opening right-panel
  conversations or tasks, multiple conversation drafts coexist and retain
  input across tab switches, and background conversation updates no longer
  steal focus from an open browser tab.

Full Changelog

Compare v0.6.14...v0.6.15

Added

• Added end-to-end local observability for coding sessions. Run lifecycle,
  provider requests, retries, compaction, checkpoints, tool calls, approvals,
  token usage, cost, and latency are recorded as correlated structured events
  and assembled into session-scoped trace bundles.
• Added conversation diagnostics to the desktop client. Each conversation now
  exposes a trace view with an overview, execution trajectory, request content,
  raw data, tool definitions, and a resizable inspector, with live updates
  while a run is in progress.
• Added a composer context ring that surfaces the current context usage and a
  compact action in a dedicated popover, replacing the meter hidden in the
  model settings menu.
• Added durable session event checkpoints. Tool intent is checkpointed before
  dispatch and interrupted executions are repaired with explicit outcomes, so
  conversations can be reconstructed from committed events instead of inferred
  runtime state.

Changed

• Session transcripts now require version 6. Legacy version 2-5 transcripts are
  rejected and are not migrated automatically.
• Diagnostics are presented as session-scoped Turns and Steps with request and
  timing details, and activated skills render as Markdown in the diagnostics
  inspector.

Full Changelog

Compare v0.6.13...v0.6.14

Changed

• Refreshed the embedded model catalog from models.dev (394 to 404 models) and
  compiled explicit thinking levels for OpenAI Responses routes. Unverified
  OpenCode and Copilot reasoning controls remain fixed to provider defaults.

Fixed

• Kept the Settings default-model picker in sync with newly configured
  providers. The selector now reloads /api/models after provider saves or
  activation, so newly added models appear without leaving Settings.
• Refined Composer skill catalog navigation. Project Skills load from the
  active session workspace and are listed before system Skills, arrow-key
  navigation moves exactly one option with wraparound, and boundary scrolling
  stays in sync without highlight or cursor flicker.

Full Changelog

Compare v0.6.12...v0.6.13

Fixed

• Restored the user's login shell environment when Or is launched from Finder
  or Dock on macOS. Agent commands and stdio MCP servers can now resolve tools
  installed through Homebrew, nvm, mise, asdf, Volta, and user-local binary
  directories just as they do when Or is launched from a terminal.
• Kept desktop startup resilient when the login shell environment cannot be
  captured. Or now falls back to the inherited application environment after a
  bounded attempt instead of delaying or preventing startup.

Full Changelog

Compare v0.6.11...v0.6.12

Added

• Added conversation branching from assistant responses. A branch opens as a
  separate session with the conversation history through the selected response,
  inherited session settings, a distinct title, and direct navigation between
  the source conversation and its branches. Workspace files remain at their
  current state and are not rolled back.
• Added in-place editing for earlier user messages. Or now removes the later
  conversation turns, keeps the same session identity, settings, and workspace,
  and generates a replacement response from the edited message.
• Added branch-point navigation that returns to and briefly marks the complete
  source response turn. If editing the source conversation removes that point,
  the branch keeps its source-session link and reports that the original branch
  point was edited instead of locating unrelated content.
• Added compact project and session details on sidebar hover, plus a native
  Reveal in Finder action for project directories.

Changed

• Moved MCP connection ownership to the application so compatible connections
  can be warmed and reused across conversation runtimes. Configuration reloads
  use isolated generations, allowing active sessions to finish with their
  existing connections while new sessions receive updated tools.
• Consolidated the repository's curated Skills, MCP server recommendations, and
  CLI tool recommendations under catalog/. The collection includes setup and
  security guidance for GitHub MCP and practical usage guidance for ripgrep.

Fixed

• Kept direct child branches as independent sessions when their source session
  is deleted, while preserving valid relationships between nested branches.

Full Changelog

Compare v0.6.10...v0.6.11

Added

• Added MCP tool integration over local stdio commands and Streamable HTTP
  endpoints. MCP servers can now be added, edited, enabled, tested, and removed
  from the application sidebar, with optional project scoping and discovered
  tool inspection.
• Made connected MCP tools available to the agent through Or's existing Ask,
  Auto edit, and Full access permission modes. Text, structured data, and image
  results remain available in live and restored conversations, including a
  full-size preview for returned images.

Fixed

• Preserved complete tool-result output instead of silently limiting it to 12
  lines. Large results are now bounded to 1,000 lines or 64 KiB and include an
  explicit truncation notice when either limit is reached.

Full Changelog

Compare v0.6.9...v0.6.10

Added

• Added an expandable context-window breakdown to the model menu. Or now shows
  estimated allocations for messages, system tools, the system prompt, Skills,
  project context, and free space while keeping the provider-reported token
  total authoritative.

Changed

• Simplified permissions to three session modes: Ask, Auto edit, and Full
  access. Browser navigation and page inspection are now explicit network
  operations, so Ask and Auto edit request approval before using them while
  Full access continues without prompts.
• Separated context occupancy from response billing usage. The model menu now
  reports the latest measured context window, while the response footer reports
  usage for the visible run.
• Made the authenticated coding-desktop sidecar the only production backend
  for Or. Built-in filesystem and shell tools now use their local production
  implementation directly, removing unused hosting and execution layers while
  preserving their model-facing behavior.

Fixed

• Restored sidebar, Settings, and workbench resizing after pointer capture is
  interrupted by a window blur, hidden document, missed button release, or
  another lost-pointer event. The resize cursor and drag state now reliably
  return to normal.

Upgrade Notes

• The legacy Read only permission mode has been removed. Existing sessions
  saved with read_only open in the conservative Ask mode.
• Ask and Auto edit now require one-time approval before opening HTTP(S) or
  localhost pages and before reading visible browser-page content. Listing
  session tab metadata and opening workspace HTML previews keep their existing
  internal and workspace-read behavior.
• The standalone coding Web/API server, fixed localhost:8787 hosting,
  cross-origin API access, OR_CLIENT_ORIGIN, and VITE_API_ORIGIN are no
  longer supported. Or is a Desktop-only product and continues to use its
  authenticated random-port sidecar internally.

Full Changelog

Compare v0.6.8...v0.6.9

Changed

• Unload idle conversation runtimes after their last viewer disconnects, while
  keeping active runs, approvals, questions, title generation, and background
  tasks pinned. The client now evicts inactive thread state under the same
  rules, reducing memory growth during long desktop sessions.
• Make Agent Skills the single reusable prompt and workflow mechanism. The
  composer slash catalog now contains built-in commands and Skills without the
  overlapping Prompt Template subsystem.
• Keep AI-generated session titles without a separate utility-model setting.
  Title generation now uses the session model in the background while the first
  prompt remains available as the immediate fallback.
• Store product-facing tool outcomes directly in the append-only session log
  instead of maintaining a parallel transcript-details file.
• Simplify instruction loading and prompt construction. Or now reads only
  AGENTS.md and AGENTS.local.md, keeps its stable identity separate from the
  dynamic session environment, and reports the actual Agent tool directory as
  session cwd context.
• Use the Electron native directory picker exclusively when adding a project.
  The browser-based filesystem picker and its backend directory API have been
  removed.
• Split the engine session implementation into focused assembly, lifecycle,
  run, journal, and context-refresh owners. Unused plugin, capability, legacy
  compatibility, and internal API layers were removed, reducing the codebase by
  nearly 6,000 net lines since v0.6.7.

Fixed

• Tightened existing and newly created session storage to 0700 directories
  and 0600 transcript files so prompts, code, and tool output are not readable
  by other local users.
• Prevented inactive server sessions, replay frames, and client thread state
  from accumulating for the lifetime of the app. Reopening a session restores
  it lazily from durable storage.

Upgrade Notes

• Prompt Templates and their argument expansion are no longer supported.
  Convert reusable workflows from ~/.or/prompts/ or
  <workspace>/.or/prompts/ into Agent Skills under
  ~/.agents/skills/<name>/SKILL.md or
  <workspace>/.agents/skills/<name>/SKILL.md.
• The Utility Model setting has been removed. AI titles continue to work and
  use the model selected for the session.
• Rename CLAUDE.md to AGENTS.md and CLAUDE.local.md to
  AGENTS.local.md; the Claude-specific fallback names are no longer read.
• Provider settings must use schema version 4. Versions 1-3 are no longer
  migrated; reconfigure providers in Settings if an older file is rejected.
• Legacy usage JSONL files are no longer imported into SQLite. Users already
  running v0.6.7 have completed this migration; older installations should
  upgrade through v0.6.7 first if historical usage totals must be retained.
• Existing conversation text remains readable, but legacy tool details and
  preview metadata stored only in the old transcript-details sidecar are not
  imported into the consolidated session log.
• The -cwd flag and OR_CWD environment variable have been removed. Project
  directories continue to be selected per session through the desktop UI.
• Selecting a new project now requires the Electron desktop bridge. The local
  HTTP API remains part of the desktop application, but the browser filesystem
  picker and /api/directories endpoint are no longer available.

Full Changelog

Compare v0.6.7...v0.6.8

Added

• Added support for the open Agent Skills specification. Or now discovers user
  Skills from ~/.agents/skills/<name>/SKILL.md and workspace Skills from
  <workspace>/.agents/skills/<name>/SKILL.md, with workspace Skills taking
  precedence when names collide.
• Added a curated, redistributable Skill collection under skills/. The first
  packages are Vercel Labs' MIT-licensed find-skills and iOfficeAI's
  Apache-2.0-licensed officecli, with pinned upstream revisions and licenses.
  They can be installed from GitHub with commands such as
  npx skills add ktsoator/or@officecli -g -y.

Changed

• Unified built-in commands, Prompt Templates, and Skills under the composer's
  / menu. Opening the menu refreshes both file-backed catalogs, and a Skill
  remains distinct from a same-named Prompt Template.
• Preserved activated Skill instructions as protected session context across
  compaction and session restoration. Repeated activation reuses the original
  instruction snapshot instead of loading a changed file mid-conversation.
• Omitted the Skill catalog, protocol text, and tool entirely when no valid
  Skills are available.
• Tightened SKILL.md validation to the standard fields and naming limits.
  Skill bodies are loaded unchanged, relative paths resolve from the Skill
  directory, and allowed-tools never bypasses Or's permission policy.

Upgrade Notes

• Move existing Skills from ~/.or/skills/ to ~/.agents/skills/, and move
  workspace Skills from <workspace>/.or/skills/ to
  <workspace>/.agents/skills/.
• The composer no longer uses $ to open Skills. Use / for commands, Prompt
  Templates, and Skills.
• Legacy prompt-template substitution and nonstandard top-level Skill fields are
  no longer supported. Put version information under metadata.version.
• Bundled resources in user-level Skills remain outside the workspace and may
  require normal read approval. Executing a bundled script continues to use the
  normal command permission flow.

Full Changelog

Compare v0.6.6...v0.6.7