OPEX v0.33.0

OPEX v0.33.0

Builds on v0.32.0's model catalog with capability-aware model UX, adds shareable read-only sessions and OpenAPI tool import, and refines the approval workflow.

Model catalog & capabilities

• Phase 3c — capability data + a temperature gate, reasoning_content detection from models.dev (interleaved field), and a vendor-hinted resolution tier from discovery owned_by.
• Chat model UX — model capability badges in the chat model dropdown (context window + vision / reasoning / tools) and reasoning_content exposed in the models endpoint with a badge.
• Catalog-priced usage $ and a max_tokens cap driven by the catalog.

Sessions & tools

• Read-only shareable links — share a session view without granting write access (Tier-3 #6).
• OpenAPI import — an operator-facing importer that turns an OpenAPI spec into tools (Tier-3 #7).

Approvals & agent behavior

• "Always allow" button on the approval card, with wildcard tool-policy / approval matching.
• Structured compaction summary for cleaner long-session compaction.
• Hermes parity — persist the per-turn model override across restart; per-category context breakdown in usage.

Provider UX

• Streamlined provider creation (auto type / base_url, type moved to Advanced) and a tabbed, mobile-friendly provider edit dialog.

Tests & internals

• HTTP-recorder cassette tests for providers and codemode "tools-as-code".
• Brand mark refresh (woodcut walnut engraving) across the sidebar and chat welcome.

---

Full changelog: git log v0.32.0..v0.33.0

OPEX v0.32.0

OPEX v0.32.0

A large release dominated by the live model catalog, a full-scale design-system migration of the web UI, and the ITS (its.1c.ru) browser tool.

Live model catalog

• Phase 1–3 — model metadata sourced from models.dev (context windows) and an OpenRouter loader, replacing the hand-maintained context-window table.
• Provider preset picker — add 150+ providers OPEX doesn't ship out of the box.
• Capability gating — capability data drives a max_tokens cap, a temperature gate, and catalog-priced usage ($).
• Operator-configurable context_window and a per-model context-window table (supersedes the single override) — fixes the MiMo 128k fallback.

Design-system migration

• Every major page migrated onto shared design-system primitives — Secrets, Backups, Channels, Integrations, Tasks, Providers, Agents, Monitor, Config, Skills, Memory, Access, Tools, Login, Setup wizard, Webhooks.
• New primitives: Card, IconTile, StatCard, Table, DataRow, CopyableCode, StatusBadge/StatusDot, Badge sizes, Chip, Pagination, SegmentedControl, SearchInput, SectionHeader, ScrollableTabsList, Stepper, Alert, AuthShell, ErrorState, DialogContent size prop.
• Non-color design tokens (type / layout / z / elevation / tap-target) and a no-raw-design-values lint guard to keep pages on the system.

ITS (its.1c.ru) tool

• Agent-facing its YAML tool (search / read via toolgate), backed by real its.1c.ru selectors and a multi-step SSO login with conservative relogin.
• toolgate /its/search + /its/read endpoints (serialized, hard timeout), HTML→markdown extraction, and a search-results parser (TDD).
• browser-renderer gains a generic ProfileManager — persistent stealth context per profile — so the ITS session survives across calls.
• ITS credentials resolved from the vault via GET /api/internal/its-credentials (cached, TTL).

File Handlers UI

• File Handlers tab with a HandlerEditor (CodeMirror Python + descriptor form), split into Settings/Code tabs, and edit / create / delete / reset actions with status badges.
• Backend: GET/POST/PUT/DELETE /api/handlers with toolgate validation (POST /handlers/validate) and audit.

Other

• Server-side search for /api/audit + a global audit search in the UI.
• Per-(agent, tool) tool-quality penalty (migration 070) and a per-tool TTL/threshold for the semantic-search cache.
• Total time budget for the decision-webhook chain (on_chain_timeout, default open).
• Imported 8 server-authored skills (incl. its-search-optimized).
• Complete i18n pass and an app-icon rebrand (Bot → Nut) with a matching favicon.

---

Full changelog: git log v0.31.0..v0.32.0

OPEX v0.31.0

OPEX v0.31.0

The first release under the new name OPEX (formerly HydeClaw). It rolls up two development cycles (0.30.x → 0.31.0): the completed Hermes-parity MEDIUM cycle and the move to zero-config media tools.

Hermes-parity

• Mid-run clarify tool. An agent can ask the user a clarifying question in the middle of a turn (ClarifyManager + waiter/heartbeat) — delivered on the web (ClarifyCard) and in channels (inline buttons + text intercept, taking priority over approval). The default timeout is aligned with request_timeout.
• Checkpoint / rollback. A shadow git repo takes per-turn snapshots of workspace/agents/{agent}/ before every write / edit / delete / rename. Restore via the /rollback slash command (list / N / diff / file) and via REST (list / diff / restore), plus a CheckpointPanel in chat. Prune by keep+ttl, with anti-traversal and a kill-switch.
• Runtime decision-webhooks (user hooks). A synchronous HTTP gate at three lifecycle points: BeforeMessage (block / inject-context), BeforeToolCall (block-veto / modified-args), and AfterToolResult (transform-result + provenance). SSRF protection, provenance sanitization, audit (AuditEvent::HookDecision), and an editor in AgentEditDialog.
• VAD (voice). Client-side RMS VAD: auto-stop recording on silence, a level indicator, skip-empty, and a continuous hands-free mode. Server-side silence trimming via ffmpeg in toolgate.
• apply_patch (V4A). A system tool that applies contextual patches in the V4A format (Update + Add, multi-file, atomic: validate-all → write-all) with a path guard, read-only checks, and an automatic checkpoint.

Capability tools

• Media tools (STT / TTS / Vision / ImageGen / web-search) are now zero-config — built in and resolved directly from the active providers (5 YAML definitions removed). They appear in the UI as builtin (read-only) and are added to SUBAGENT_DENIED_TOOLS.

Infrastructure

• Infra-monitor. OPEX watches backup and curator: heartbeat self-heal (enable + create if missing) and a watchdog backstop alert (infra_jobs, throttled transitions).
• Removed the GSD plugin (.planning + CLAUDE.md sections + local agents/hooks).

UI & fixes

• Responsive mobile layout for the Integrations page (stacked account/binding rows).
• Radix ScrollArea replaced with native scrolling where touch scrolling broke on mobile.
• Many fixes: gemini-cloudcode test compilation, regenerated SSE wire fixtures, clippy, and preserving webhooks on PUT (data-loss fix).

---

Full changelog: git log v0.29.0..v0.31.0

HydeClaw v0.30.0

HydeClaw v0.30.0

A large feature release centered on autonomous goals, durable resumption, and a broad security-hardening pass. (Released under the old name HydeClaw; the project was renamed to OPEX in v0.31.0.)

Autonomous goals (/goal)

• Goal driver loop — /goal and /subgoal slash commands start a background driver that runs a turn, delivers the result, asks an LLM judge whether to continue, and decides the next step (GoalDriverPool, session_goals table, verdict/command parsers).
• Per-session serialization — user turns are serialized against the goal driver via a per-session lock, so a live message never races the autonomous loop.
• Cron-driven goals — autonomous_goal is reachable through the agent cron tool (create/update/run_once) with durable channel delivery.

Durable resumption (Phase 1)

• Re-drive of crashed goals — a startup resumer re-drives autonomous cron goals that crashed between turns, backed by an atomic claim_redrive primitive and a redrivable-goal query.
• Interrupted-verify guard — blocks blind re-runs of non-idempotent work after an interruption.
• Owner notifications — owners are notified (via channel push) of interrupted interactive /goal runs.

Security hardening

• Domain blocklist — [security] blocked_domains config + a glob url_policy matcher, enforced on browser_action, web_fetch, and URL args of YAML tools (browser / screenshot_web).
• Code-execution scanning — pre-exec scan of code_exec (block on host, warn in sandbox) plus non-blocking dangerous-code warnings on workspace_write/workspace_edit.
• Prompt-injection blocking — high-severity injection patterns in SOUL.md/IDENTITY.md system-prompt files are now blocked (not just logged); patterns tagged with severity and expanded to cover C2 / exfiltration / persistence.

Channels & voice

• Per-chat voice mode — /voice on|off|status (DB-backed, channel_voice_modes); when on, every reply is also sent as a voice message.
• Richer browser automation — browser_action gains scroll / hover / drag / back / press and JS-dialog handling, with a testable dispatch extraction.
• In-session todo tool — session_todos table + an Active TODO block injected into the system prompt each turn (survives compaction).
• /compact surfaced in the chat slash-menu autocomplete.

Providers & media

• Web search — search_web via toolgate /v1/search with a live active-provider list; new websearch capability with searxng / ollama / brave plugins; SearXNG replaced by Ollama Cloud web search.
• New media providers — vendored Silero TTS (with EN→Cyrillic transliteration), Xiaomi MiMo (TTS/STT/Vision), and OpenRouter STT (Whisper / GPT-4o-transcribe).
• gemini-cloudcode provider (feature-gated) — full LlmProvider impl with Google OAuth routes, streaming delta synthesis, and integration tests.

UI & infrastructure

• Provider capability groups + web-search + priority editor; adoption of the Field/EmptyState/neu-card design system with a11y badges; an i18n pass and mobile UX polish (shortcuts, swipe, lightbox, shiki tool output).
• Startup sweep that deletes orphaned tool-result rows; MetricsRegistry counters for re-drive / interactive-notify events.

---

Full changelog: git log v0.29.0..v0.30.0