Changelog

Highlights

• Bring Team Context into Claude Code and Codex with provider-project and pathless activation, self-verifying setup, immediate in-session adoption, and a dedicated Context Tree settings page.
• Delegate GitHub Issue and pull-request tasks to a repository-scoped Team Agent, with terminal replies published by the First Tree GitHub App.
• Distribute complete Team Skill bundles—including nested references, scripts, and binary assets—across supported agent runtimes.
• Review open agent questions through the desktop and mobile Need you queue, ask for clarification without resolving the request, and process requests sequentially.
• Select Codex Fast mode and inject First Tree-managed MCP servers into Cursor and Kimi Code.
• Give invited members a progressive, personal-agent-first onboarding path and make the mobile PWA easier to install from desktop or phone.

Notable fixes

• Improved GitLab project-hook routing and clarified Context Reviewer and private-project setup states.
• Kept chat navigation, the composer, pinned selection, and Mobile Work responsive under background activity.
• Strengthened daemon ownership, runtime token recovery, resumed-message custody, tracked-request images, and attachment persistence.
• Run the production container as a non-root user.

Upgrade note

Existing First Tree GitHub App installations must re-approve the upgraded Issues: write permission before App-authored terminal replies can be published. Other GitHub automation continues to work while this permission is pending.

Highlights

• Mobile now has one unified Work surface for attention, unread, pinned, active, archived, and watched conversations, with fuller summaries and inline current state.
• Scheduled jobs are now first-class: agents can create and manage recurring work through the CLI, while Desktop shows schedule status, upcoming runs, and owner lifecycle controls.
• Settings → Setup now centralizes Team readiness and owner actions for repository automation, Context Tree, and Automatic Review, including guided GitLab System Hook setup with real merge-request routing verification.
• Context Tree workflows now support GitLab repositories, while review flows gain repair-first handling, stronger final-tree checks, and exact decision-influence receipts.
• The composer now exposes a focused live activity inspector so users can understand active agent work and jump to its timeline evidence.

Notable fixes

• New human input can retry Codex steering after an earlier append rejection instead of waiting for a long-running turn to finish.
• Failed managed-agent chats remain in Attention while offline and through reconnects until they genuinely recover.
• Mobile selection, sent mention rendering, GitLab entity links, and Context Review topic/merge behavior are more reliable.

Highlights

• Add Kimi Code as a first-class agent runtime provider, including capability detection, configuration surfaces, and runtime handling.
• Bring GitLab entity attention alongside GitHub with follow, following, and unfollow CLI commands, connection and identity settings, webhook routing, and native chat cards.
• Ship dedicated Context Tree review and audit skills together with the reusable First Tree QA workflow and cases.
• Add connected Google and GitHub account management plus an install-free quick start for invited team-agent users.

Notable fixes

• Retry transient provider failures even after visible output and report Claude token usage as accurate deltas.
• Align SCM attention semantics, make repository settings provider-neutral, keep renamed chat titles in sync, and improve participant and people-picker interactions.
• Harden OAuth cookie and credential handling while completing landing-campaign conversion measurement and restoring onboarding and growth analytics delivery.

Highlights

• Keep the full agent briefing checked in and session-stable across runtimes, with compact GitHub and GitLab guidance available without per-session capability drift.

Notable fixes

• Preserve correct GitHub issue and pull-request routing while simplifying GitLab operations around native glab guidance instead of a bundled GitLab skill.
• Restore mobile Chat to compact conversation rows and remove unintended management gestures and controls from mobile Now and Chat cards that could stretch rows or expose unsupported actions.

Highlights

• Bring the mobile workspace to production, including add-to-home-screen guidance and touch-first chat, triage, and picker improvements.
• Add the external Cursor Agent CLI as a supported runtime provider, with in-product setup, session resume, authentication handling, and replay-safe failure recovery.
• Prioritize conversations with attention and private pinned groups, activity-based ordering, richer filters, and cross-device pin synchronization.
• Let agents send workspace images through chat send, with improved image sizing, galleries, and a full-screen lightbox.
• Add agent-scoped Context Tree bindings, generated governance guidance, and stronger GitHub/GitLab forge setup workflows.

Notable fixes

• Install Linux daemon services in system scope when the CLI runs as root, while preserving user-scoped systemd behavior for normal users.
• Route user-requested issues and pull requests to the repository the work belongs to instead of accidentally defaulting to First Tree's tracker.
• Improve iOS standalone sizing, mobile card previews, install-guide copy, and narrow-screen timeline layouts.

Compatibility

• Context Tree validation is now strict and always on. Trees with malformed metadata, unsafe symlinks, invalid soft_links, or cross-content-class links must be repaired before their next tree PR can pass.

Highlights

• Run First Tree agents on Windows through a per-user Task Scheduler daemon supervisor, with service restart/recovery hardening and reliable npm/Codex executable resolution.
• Attach documents and other files from existing chats, new-chat drafts, and request answers; agent runtimes can download non-image attachments for Claude and Codex handoff.
• Start or recover Context Tree setup in a private chat, resolving sources from declared workspace repositories, a local Git repository, or a GitHub URL.
• Use GPT-5.6 Sol, Terra, and Luna Codex models with max and ultra reasoning, and discover the Codex CLI bundled with the ChatGPT desktop app.
• Hosted macOS/Linux onboarding now uses the channel-aware portable installer with simpler copyable install and login commands.

Notable fixes

• Surface terminal provider failures as durable, actionable chat notices, including clearer Claude proxy and connectivity guidance.
• Resolve runtime-session tokens fresh from per-agent files and tolerate token-rotation races during soft enforcement.
• Make first-tree upgrade fall back to the current channel's latest release when no server URL is configured.
• Let self-directed GitHub assignments and mentions create a tracking chat when none exists.
• Reject malformed double-escaped agent messages while preserving intentional stdin and message-file bodies.
• Correct CommonMark fence handling in document-link previews to avoid missed or spurious links.
• Reduce generated agent briefing size while preserving the hard collaboration, worktree, Context Tree, and skill rules.

Compatibility notes

• Windows daemon support is per-user and starts after sign-in through Task Scheduler; it is not a Windows Service or pre-login daemon.
• The new mobile /m/* shell remains channel-gated and disabled in production.
• Existing npm-installed clients keep upgrade compatibility; hosted fresh installs on macOS/Linux use the portable installer.

Highlights

• A new one-command installer is available for machines where installing Node.js or npm is inconvenient. It downloads a self-contained First Tree CLI and can keep itself up to date.
• CLI updates are safer. Existing installs are handled more carefully, and release downloads are checked before users are pointed at them.
• GitHub setup is easier to understand and fix. The settings page now shows clearer connection status, more team members can inspect setup, and missing repository access gets a direct recovery path.
• Signing in from the CLI is faster with shorter codes, and first-time setup gives clearer guidance about where team data will live.

Fixes

• Fixed a release issue that could block publishing the CLI even when the build was already correct.
• Improved recovery after reconnecting or resuming an agent, so fewer sessions get stuck after local state changes.
• Made review and setup messages clearer when a human decision is needed or when GitHub permissions are missing.
• Made chat input, image uploads, and status messages behave more consistently.

Upgrade

``bash npm install -g first-tree@0.5.12 ``

New installer:

``bash curl -fsSL https://download.first-tree.ai/releases/prod/install.sh | sh ``

Highlights

• Onboarding redesigned around value-first GitHub access, and the create-agent step is skipped when a member already has a personal agent (#1311, #1337).
• Self-service team leave from the workspace UI (#1290).
• Proxy compatibility without managing your proxy: the daemon now reads a user-owned daemon.env instead of baking a point-in-time snapshot into the service unit, ending the recurring 403 Request not allowed egress freeze (#1335).
• first-tree-welcome skill is now demand-driven with first-task verification (#1341, #1344).

Notable fixes

• Channel-correct daemon.env path, with inline-comment stripping, empty-value skipping, and tighter egress-vs-auth detection (#1339).
• Stop treating a flaky codex --version check as a missing binary (#1329).

Internal

• Skill-eval tooling: deterministic grading, quality judge, result operations, and seed/read eval gates (#1327, #1331, #1332, #1333, #1338, #1343); removed the unmaintained packages/e2e harness (#1336).

Upgrade

``bash npm install -g first-tree@0.5.11 ``

Highlights

• Native engines are now externalized. The Claude and Codex native engines resolve from your PATH (with a one-click install) instead of shipping inside the package — global installs are smaller (~15MB saved by pruning unused Claude SDK JS peer deps), capability detection is install-aware, and missing runtime credentials can be logged in directly from chat.
• Shell-safe chat bodies. chat send / chat ask accept --message-file / -F / stdin for shell-safe message bodies; inline bodies are now guarded against shell residue, and chat send/ask/update are framed as real CLI tool calls so replies actually get delivered.
• Long tasks survive idle-suspend. The client working_grace default is widened 1h → 12h.
• Inline Context build entry. Connect GitHub and pick a repo inline when starting a Context build.

Notable fixes

• Login-shell PATH probe is now shell-agnostic — fish users are no longer reported as missing — and canonicalizes PATH dirs; CLAUDE_CODE_EXECUTABLE is gated on executability.
• Stop iOS Safari auto-zoom on mobile inputs (16px floor).
• Context Tree IO diagnostics performance: prefilter/index/reduce candidate scans, time decisions synchronously, reuse the tree binding.
• Onboarding: require a personal agent for readiness, default new chats to a cached starter agent, org-level tree setup recovery.
• Added skill eval gates (welcome, write, floor coverage matrix) and refactored the read suite into the core framework.

Highlights

• Redesigned multi-team switching & management; team rename moved into the switcher (#1248, #1258)
• In-product browser-OAuth Connect for the Claude/Codex SDK — no separate CLI login; surfaces login failures and fixes the fallback sign-in URL (#1194, #1225)
• Context Reviewer PR automation, with an immediate-save Switch setting plus PR-update wake/dedupe (#1233, #1260, #1236, #1254)
• Build a Context Tree on a personal GitHub account, and pick which repos seed tree init (#1263, #1213)
• Richer ask answer card: @mention + image attachments, plus Enter/Esc shortcuts (#1256, #1179)
• Chat description surfaced as a pinned, collapsible task summary (ChatSummary) (#1195, #1223)
• Offline-agent + leave-a-task affordance in the chat timeline (#1246)
• Adopt a member's first agent as their delegate (#1240)
• Onboarding revamp: agent-first kickoff, work chat starts before tree setup, skill renamed to first-tree-welcome (#1218, #1231, #1230)
• Restored chat send <human> and retired the agent-final-text mirror; rebind the harness "user" to the runtime so agents reliably chat send (#1190, #1232, #1226)
• GA4 instrumentation + cross-domain linking on the cloud SPA (#1201, #1203)
• Instant-save profile & add-repo forms; agent-detail IA cleanup (#1212, #1211, #1180, #1182)

Notable fixes

• Claude SDK provider error handling; session limit surfaced as a provider error; resolve the bundled Claude native binary and drop codex device-code (#1242, #1228, #1219)
• Codex provider: app-server token replay/usage deltas, stop retrying deterministic compact failures, suppress expected shutdown noise (#1243, #1241, #1185, #1183, #1222)
• Reject empty / placeholder message bodies at the write boundary (#1257)
• Defer idle-suspend while a provider has a live subprocess (#1253)
• Session provider retry policy + transient agent-config fetch retry during bring-up (#1192, #1239)
• Reconnect after inbox recovery timeout; resume foreground polls on focus/pageshow (#1237, #1244)
• Restore visible text in the chat ask answer input (#1266)

Install / upgrade

`` npm install -g first-tree@0.5.9 ``

Highlights

• Composer drafts are now cached per chat in local storage, so an unsent message survives switching chats and reloading the page (with rollback / stale-ref hardening). (#1165, #1167, #1168, #1169)
• Returning sign-in lands you back in your last-used org, plus onboarding copy polish. (#1148)

Notable fixes

• The chat-ask answer card is phone-friendly: Reply / Skip stay reachable on short viewports, the card lifts above the on-screen keyboard, and long option previews wrap instead of overflowing. (#1177)
• Agent runtime: prepend the channel CLI bin for agents (#1164), batch-settle Claude Code coalesced inputs (#1159), and inject chat context per provider session (#1157).
• Route the Codex handler through the app-server. (#1155)
• agent-detail edits now save immediately. (#1171)
• Right-rail Summary uses a single, tighter type scale. (#1150)

Internal

• Staging-only toggle to hide agent final-text mirrors (#1149); docs + CI upkeep (#1166, #1163, #1050).

Highlights

• Onboarding revamp. Setup is reframed as "connect your coding agent": a ceremonial welcome for admins and invitees, restructured connect steps (Install First Tree → Create agent → Connect GitHub), an idempotent server-side kickoff endpoint, connected-account confirmation on the connect-code step, and a per-org draft for onboarding repo selection.
• Agent ↔ human comms split into chat send / chat ask / chat update. chat ask raises a tracked, blocking open question with a docked answer surface in the web composer; chat update upgrades the chat description into a living status report. Open-question prompts render markdown with clamping and question caps.
• Explicit GitHub follow/unfollow. A github follow / unfollow CLI replaces session-event auto-binding for wiring PRs and issues to a chat, and GitHub entity titles now show in the chat right sidebar.
• Context Tree authorship & attribution. Ships the first-tree-write skill, detects Context Tree writes from git status with git-derived complete writes in the Context tab, attributes refs by git repo identity, and moves agent-managed source repos under <workspace>/source-repos/<name>.
• "New version available" refresh chip in the web topbar, with detection and a refresh entry on narrow layouts.
• Web inspection panel. The chat right sidebar is restructured as an inspection panel; a dark, brand-consistent login page; identicon/pixel-avatar fallbacks; and a reworked jump-to palette with better recents and search.

Notable fixes

• Server no longer mints a second chat for a PR's own creator.
• Inbox delivery hardening: terminal delivery ack enforcement, fair delivery windows, ack-recovery scheduling, and a delivery-status index.
• Claude/Codex TUI suspend/resume hardening, including fencing suspend during resume.
• Removed agent re-bind (provider/computer switching); login --override rotates the local client identity instead.
• Chat description guarded against literal \n escapes; long unbreakable runs wrap in chat markdown.
• Onboarding gate scoped to memberships; preview states synced.

Upgrade

``bash npm install -g first-tree@latest ``

Highlights

• Chat requests now have explicit open-question resolution, chat descriptions paired with topics, and Needs attention keeps open requests pinned.
• Agents now operate on the chat-send-only contract and include the new first-tree-read skill for scoped Context Tree reads.
• Onboarding and Context Tree setup recover better when admins skipped connecting code, with clearer invitee and team copy.

Notable fixes

• Require explicit production server secrets and remove the retired Feishu/Kael adapter subsystem.
• Lower the published Node engine floor to >=22.13.0 to match resolver-enforced dependencies.
• Degrade GitMirrorAuthError as a source-repo checkout condition, restore composer footer toolbar clicks, and guard prompt edits from pasted assembled briefings.

Notable fixes

• Fixes the onboarding flow so users are not bounced away when their organization gains a usable agent mid-flow; the leave-bounce decision is made at entry instead of on every render.

Highlights

• Ships the W1 workspace-rooted Context Tree flow: workspace.json, .first-tree-workspace state, per-agent source repo clones, first-tree-seed, GitHub repo initialization, and W1-safe tree verify.
• Improves agent-team collaboration with explicit group-chat recipients, tracked human requests, channel-aware command text, clearer async handoffs, and unconditional First Tree / Context Tree reading rules in agent briefings.
• Refreshes onboarding and admin workspace UX across repo connection, Settings Resources, agent Capabilities / Instructions / Usage, Team, chat identity hovercards, and narrow viewport behavior.
• Adds the Claude Code TUI runtime path and improves agent runtime reliability around final responses, usage accounting, session mappings, source checkout fallback, and workspace cleanup retries.
• Aligns CLI upgrade behavior with the server-advertised target version and refreshes the auto-update target before install.

Notable fixes

• Includes bundled skills and CLI README in the published npm tarball.
• Recovers inbox delivery and acknowledgement durability across suspend, redelivery, and delivered-prefix paths.
• Fixes WebSocket auth handshake classification, client presence close grace, and several onboarding access guards.
• Recovers document previews from message metadata when the snapshot cache is gone.
• Fixes web interaction issues around select options in modals, prompt editing, tab overflow, button text colors, and mention/request rendering.

Compatibility notes

• Legacy GitHub Scan and pre-W1 Context Tree initialization paths are retired.
• Existing clients migrate workspace state forward through one-shot cleanup paths; no manual migration is expected for normal npm users.

Highlights

• Need-Human-Attention (NHA) primitive — server + web + CLI + skill (#602, #615, #616, #617)
• In-flight message recovery via agent:bind reset + deferred ack (#612)
• Card-based IA for Settings → Computers, with 4-state status pill and connect dialog (#594, #586, #603, #607)
• Per-agent slash-command catalog + composer popover (#588)
• Channel-aware CLI binding in agent prompts + server hints (multi-env) (#627)
• Translate codex/claude SDK auth errors into actionable chat hints (#618)
• Responsive workspace layout for mobile viewports + responsive Settings shell (#577, #596)

Notable fixes

• Client resilience — error taxonomy, child process registry, auth paused, session/bind retry, stream sniff (#582)
• Retry SDK timeouts + tighten fetchAgentConfig budget (#631)
• Bundle skills/ in CLI tarball + auto-refresh on CLI drift (#605)
• Include node path in launchd / systemd units (#622, #625)
• Dedup concurrent Context Tree syncs sharing one clone dir (#606)
• Widen appendEvent seq retry budget + jitter to absorb concurrent bursts (#570)
• Stop ComposeStatusBar rendering the lead agent twice (#630)

Highlights

• New beginner-first onboarding flow at /onboarding (#529)
• Instant chat switching via IndexedDB caching (#286)
• Unified agent work-status across chat list, sidebar, and composer (#524)
• Agent runtime errors now surface as structured session events (#554)

Notable fixes

• Auto-retry transient git fetch / set-head failures (#548)
• Auto-recover worktree after orphaned dev-server leftovers (#527)
• Forward shell proxy env into launchd / systemd units (#526)
• Server: route GitHub events by human; UUID v7 message IDs (#544, #549)

Patch release. Fixes npm install -g first-tree failing on v0.5.0 with EUNSUPPORTEDPROTOCOL — the published tarball leaked a workspace:* reference into its dependencies. (#523)

Upgrade: npm install -g first-tree@latest

First release under the unified first-tree npm package, after merging the previously-separate first-tree-hub (Agent Hub client + daemon) and first-tree (Context Tree onboarding + GitHub Scan) into a single CLI.

What's in this release

• Unified CLI — first-tree (with ft short alias) now exposes:
• Agent Hub client: first-tree login, first-tree chat …, first-tree connect …
• Background daemon: first-tree daemon start|stop|status
• Context Tree workflows: first-tree tree … (onboarding, sync, write)
• GitHub notification scan: first-tree github scan …
• Single source of truth — one repo (agent-team-foundation/first-tree), one npm package, one binary.
• Drop-in for both prior tools — local data path ~/.first-tree/hub/ is unchanged; auth tokens, daemon state, and Context Tree bindings carry across.

Install / Upgrade

New users ```bash npm install -g first-tree first-tree login