v0.2.0-beta.1 — Workspace, Research, and Desktop Agent

[ihabkhaled]

Claw v0.1.0-beta3 — Workspace, Research, and Desktop Agent

Release tag: v0.1.0-beta3
Range: 5b38dd5..3a3f9d1 (exclusive of 5b38dd5, inclusive of 3a3f9d1)
Previous release: v0.1.0-beta2 (ebdbedc)
Window: 2026-04-19 → 2026-04-26
Scope: 128 commits · 2,026 file changes · +105,138 / −17,909

---

Highlights

This release pushes Claw from a routing-and-chat platform into a full AI control plane: a complete workspace UI for twelve external providers, a new claw-research-service for live web research with multiple search adapters, a brand-new claw-agent-service with a device-paired desktop runtime, and per-module advanced model selection across every advanced chat module. Routing learns from feedback, the catalog classifies models dynamically, security debt from Dependabot and CodeQL is cleared, and the architecture stabilizes at sixteen independent backend services.

The headline shifts:

• Workspaces ship end-to-end UI for 12 providers — Gmail, Jira, GitHub/GitLab/Bitbucket, Slack, Google Docs, Confluence, Figma, SharePoint, OneDrive, ClickUp — each with a list page, detail dialog, and AI actions bar, plus cross-workspace multi-provider workflows.
• Research becomes a first-class module — a new microservice orchestrates Brave, Exa, Firecrawl, SerpApi, and Bing-RSS with an evidence bundle, content-safety guard, and a chat-side research toggle.
• Desktop agent lands phases A–D — device pairing, scope-aware policy engine, streaming command execution with cancel and timeout, and scheduled commands, all driven through a rewritten agent-cli.
• Advanced model selection per module — every advanced chat module (parallel, judge, compare, research, etc.) now exposes its own selector, with a compare-judge state machine and visible AI progress streamed into the UI.
• Approval Center v3 with cross-service action events flowing into audit, plus a new approval UI and AI actions dialog with a full i18n refactor (Hindi added — now 9 locales).
• Security cleanup: all open Dependabot and CodeQL alerts resolved, including the CodeQL #15 fix swapping cmd.exe for explorer.exe in openBrowser.

---

What's new

Workspace service — full UI rollout and 12-provider coverage

Workspaces went from a backend with partial scaffolding to a complete product surface.

• Provider UI streams 04 → 12 delivered list pages, detail dialogs, and AI action bars for: Gmail (Stream 04), Jira (05), Source Control / PRs (06), Slack (07), Google Docs (08), Confluence (09), Figma (10). Stream 11 added cross-workspace multi-provider AI actions; Stream 12 wired every provider page into sidebar nav and routes for the QA pass.
• Adapter coverage reached 12 providers: GitLab and Bitbucket (prompts 10/11), Confluence/Gmail/SharePoint/OneDrive (14/18/19/20), Figma and ClickUp (15/16). GitHub gained issue + PR sync and refresh across all three object types (prompt 09).
• Phased rollout 0 → 5 introduced platform-wide guardrails first (hide manual flow, error taxonomy), then connector detail page + IA cleanup (1), auto-sync on connect + activation UX (2), connector → chat handoff (3), action-lifecycle clarity + payload preview (4), and a deprecation-warning + reconnect UX with documentation (5).
• Approval Center v3 is in: a workspace-service module emits shared action.* events that the audit-service now consumes, with the frontend showing a unified approval queue and AI actions dialog.
• Scheduled sync automation (Stream 01 Phase 5) runs unattended syncs against connected providers.
• Cloud generation via internal chat endpoint plus a matrix CI lane for the workspace stack.
• Operational visibility: an operations center with live object refresh (prompt 06), a frontend admin console for providers + app configs (prompt 04), and a health-events endpoint for connector diagnostics (prompt 23).
• Security hardening (prompt 07): SSRF protections on outbound fetches, webhook signature verification, and idempotency keys on action execution.
• Auth and runtime: DB-backed provider registry with encrypted app configs (prompt 05 wires runtime auth flows to those configs), real OAuth probe + connect flow + nested sidebar, GitHub PKCE code_verifier, Bitbucket redirect_uri, automatic token refresh.
• UAT / manual test / regression pack (prompt 24) shipped alongside the UI work.
• Bug fixes: GitHub OAuth 500 (missing User-Agent, real error now surfaced), Gmail sync 403 with error_message shown in the UI, AI-actions run flow, stuck syncs unstuck, OAuth error states, and app-config edit.

Research — new service and chat integration

A four-phase rollout introduces dynamic search orchestration as a real product capability.

• Phase 1: new claw-research-service for dynamic search orchestration.
• Phase 2: fetch layer, evidence bundle, and research runs.
• Phase 3: chat-side research-mode toggle that wires the service into prompt assembly.
• Phase 4: research admin pages, chat research toggle, and evidence viewer in the frontend.
• Search adapters: Brave, Exa, Firecrawl, and SerpApi.
• Hardening: content safety + scrape module + search-provider bootstrap; DuckDuckGo fallback switched to Bing-RSS for reliability; low-relevance result validation; tightened research context assembly in chat.
• Provider migration: provider kinds migrated, execution service tightened, evidence bundle finalized.

Desktop agent — phases A through D

A new service-and-CLI pair brings safe local command execution under user policy control.

• claw-agent-service lands with phase A + B services, DTOs, controllers, managers, repositories (device, refresh, pairing, device-code, policy), domain types, enums, constants, utilities, guards, decorators, and DB schema/migrations.
• agent-cli rewrite around device tokens with a scope-aware runtime.
• Frontend: connect-device flow, devices management page, risk-tier badge, plus i18n keys (agent.connect, settings.devices) across all 8 locales.
• Phase C: streaming command output + cancel + timeout + an SSE endpoint.
• Phase D: scheduled commands and an expanded scope enum.
• ADR-015 documents the auth model; ADR-016 the policy engine; ADR-017 the roadmap with phases E + F deferred against an unblocked foundation.
• Service guide plus phase A + B env vars, install scripts, CI, and a root CLAUDE.md.
• Shared types: device + policy event patterns added.
• Risk-aware command flow integrated into app config.

Advanced model selection (chat-service + frontend)

Per-module model choice replaces a single global selector.

• Chat-service: new ModelSelectionMode and CompareJudgeState enums; AdvancedModuleModelSelectionService plus extensions to LocalModelSelectionService; injection into all 7 advanced module managers; advancedModelSelectionFields schema wired into 7 module DTOs; advanced-model-selection types updating event/judge/parallel shapes; selection-path coverage across 6 advanced module specs (plus 2 new specs); execution + stream types extended with provenance and progress events.
• Frontend: AdvancedModuleModelSelector component with judge / research / compare-judge supporting components; rendered on all 7 advanced module pages and thread detail; wired into 7 module page hooks plus judge-review and in-thread-compare; updates to chat composer, message bubble, and parallel/research UI; new model-selection / judge / research / display utilities; types extended for advanced model selection plus judge/research/progress payloads; new enums for selection mode, judge lifecycle, and visible progress.
• Streaming AI progress is now visible in chat with a controlled model selector and a stream progress indicator.

Routing intelligence

• Router education and feedback learning.
• Learned priors and judge examples added.
• Enriched router prompt intelligence (twice — broader pass and a follow-up pass adding guards).
• Cloud preference for coding/reasoning, with policies gated on explicit user mode.
• Recognition for gpt-5 model prefixes in connector logic.
• Honest judge provider labels in chat output.

Ollama service and catalog

• Catalog classification: search/browser helper models classified dynamically; new search_browser_score column on the model catalog; expanded catalog with registry-generated entries and discovery improvements; URL-safety hardening.
• Connector sync: exact connector model keys synced; ollama catalog sync added in connectors.
• Runtime: empty-response handling fixed; assistant bubbles get a fallback when empty.
• Discovery admin UI for sources, candidates, runs, and hardware packs (frontend); DISCOVERY_* env vars added; engineering doc on the model discovery pipeline.

Approval center, AI actions, and i18n

• Frontend approval center UI and AI actions dialog with a full i18n refactor.
• Hindi locale added — translations cover 109 new UI strings across all 9 languages.
• Translations, module constants, and research provider form updates landed in a single sweep.
• i18n now supports arbitrary nesting depth in translation keys.

Audit, chat, and provenance

• audit-service consumes workspace action events.
• chat-service: fast-path execution budgets and lightweight routing controls; fast-path execution ternaries parenthesized with refreshed tests; response provenance persisted and rendered.

Docs, infra, and skills

• All root docs updated for the 16-service architecture.
• Backend/frontend reference documentation plus an audit-gap tracking doc.
• Non-negotiable rules, skill runbooks, and a testing mandate added.
• 20 engineering-mindsets doc.
• Improved local setup guidance and scripts (installer docs).
• Portable AI skills framework under work/skills/.
• ClamAV images switched to the Debian variant in Docker.
• Dependabot configuration added; cadence later changed from weekly to daily.

---

Bug fixes

• feat(security) / fixes resolving all open Dependabot and CodeQL alerts.
• CodeQL #15 — cmd.exe replaced with explorer.exe in openBrowser.
• GitHub OAuth 500 → User-Agent header added, real error surfaced.
• Gmail sync 403 → error_message exposed in UI.
• Workspace AI actions run, stuck syncs, OAuth error states, app-config edit.
• GitHub PKCE code_verifier, Bitbucket redirect_uri, automatic token refresh.
• Empty assistant bubbles render fallback content.
• Research: low-relevance search results validated; chat research context assembly tightened.
• Ollama runtime empty-response handling.
• Chat fast-path ternary parenthesization with test refresh.
• CI: forceExit + 30s timeout across all Jest configs; multiple CI fix passes.

---

Breaking changes and migration notes

• Workspace manual flow is hidden behind the new connector detail experience; integrations relying on the old manual path should reconnect via the new flow (Phase 0 guardrails + Phase 5 reconnect UX).
• Provider registry is now DB-backed with encrypted app configs; runtime auth flows read from the database, so existing provider configurations should be re-seeded if you were carrying them inline.
• agent-cli has been rewritten around device tokens with a scope-aware runtime — older agent-cli installs need to re-pair through the new connect-device flow before they can issue commands.
• Research is opt-in via the chat research-mode toggle; admin pages let you configure providers but the toggle must be on per-thread to invoke the service.
• Routing now prefers cloud for coding and reasoning unless a user explicitly sets a mode; review your routing policies if you depended on local-first defaults for those task classes.
• 16-service topology: claw-research-service and claw-agent-service are added to the deployment; ensure your compose / orchestration includes them and that new env vars (DISCOVERY_*, agent phase A+B) are set.

---

Service inventory

After this release the backend stabilizes at 17 application packages: claw-frontend plus 16 backend services — auth, chat, connector, routing, memory, file, audit, ollama, health, workspace, research, agent, image, file-generation, client-logs, and server-logs.

---

Acknowledgements

128 commits across the workspace UI streams, advanced model selection rollout, research-service phases 1–4, agent phases A–D, routing improvements, ollama catalog work, security hardening, and a 9-language i18n pass. Thanks to everyone who reviewed, QA'd the streams, and stress-tested the connector flows.

---

Full diff: 5b38dd5..3a3f9d1
Last commit included: 3a3f9d1 — fix(security): replace cmd.exe with explorer.exe in openBrowser (CodeQL #15)

---

This discussion was created from the release v0.2.0-beta.1 — Workspace, Research, and Desktop Agent.

[jingchang0623-crypto]

16 backend services in one release — respect. That is either a very brave or a very caffeinated team 😄

The Desktop Agent caught my eye. We run agent tasks on home servers and the scope-aware policy engine is exactly what we needed. Currently using a janky bash script wrapper that checks every command against an allowlist... your Phase A-D approach is way more elegant.

One concern from the trenches though: the scheduled commands feature. We learned the hard way that giving agents cron-like scheduling power without strict budgets leads to the "infinite meeting loop" — agents scheduling tasks that trigger other agents that schedule more tasks in an O(n!) escalation.

Two things that saved us:

1. Communication budgets — each agent gets a hard message limit per task cycle
2. Cross-agent health checks — if Agent A does not hear from Agent B within N hours, it alerts the human

Curious if your Phase E roadmap includes anything like fleet-level orchestration policies. The per-device scope model is great for a single desktop, but once you have 5+ agents talking to each other through shared services, you need fleet-level governance.

The research service with multiple search adapters is 🔥 though. Currently paying for 3 different search APIs separately — having a unified orchestrator would cut our costs significantly.

Impressive release. Will be watching the desktop agent evolution closely.

(Full disclosure: we documented our multi-agent production nightmares at miaoquai.com/stories — might be useful reading for the agent-service team)

[ihabkhaled]

Thank you so much, all your concerns and comments are noted and will be revised and worked on.
The thing that the product/tool is not that mature yet, so still lots of mistakes, lots of discrepancy.
But I am proud of the most is dedicating the tool to be constructed well in a very strict architecture plan, at least to help me in the future to maintain easier.

Thank you ❤️