OpenClaw· 2026-07-28TODAY@openclaw ↗ OpenClaw 2026.7.2-beta.5: database crash recovery, cross-platform message durability, conversation rewind/branch, MCP App hosting, approval push notifications, meeting integration (Teams/Zoom/Meet), Wear OS companion, and 3,908 merged PRs
v2026.7.2-beta.5, published July 28, 2026, ships 3,908 merged pull requests from 200+ contributors across 8 main areas: ① database safety and crash recovery — quarantine storage, crash-recoverable SQLite snapshots, and schema-upgrade rollback mechanisms; live WAL verification, maintenance-race prevention, and backup integrity validation eliminate data loss paths; ② multi-platform message durability — accepted messages survive gateway restarts across Telegram, Signal, Slack, and other platforms via shared ingress drain and recovery; ③ conversation management — users can rewind or branch conversations from any individual message, with transcript switching in web and native apps; ④ interactive MCP App hosting — ticketed MCP App hosting with bounded tools, resources, and context updates; apps pin to persistent dashboards or launch from channel replies; ⑤ structured questions and approvals — agents post formatted question cards with option selections across all platforms; approvals gain push notifications, history tracking, and fair queuing; ⑥ meeting integration — native Teams, Zoom, and Google Meet support with automatic transcript collection; realtime Talk adds video and GPT Live functionality; ⑦ Wear OS companion app — home-screen controls, realtime Talk management, and audio-reactive playback; ⑧ local inference onboarding — guided setup for local model providers with downloadable models and in-process llama.cpp inference. Also included: Claude Opus 5 model integration, per-tab browser copilot with security isolation, fast active-memory recall with cross-conversation defaults, dynamic job scheduling with gated script payloads, Linux desktop integration with signed updates, and native Telegram Markdown list rendering. Security fixes address channel allowlist authorization boundaries, session export boundaries, secret redaction, and unsafe approval handling.
Claude Code· 2026-07-25@AnthropicAI ↗ Claude Code v2.1.220: bug fixes and reliability improvements (maintenance patch)
v2.1.220, published July 25, 2026, is a maintenance-only patch. The official release notes state only "bug fixes and reliability improvements," with no specific items, new features, or behavioral changes documented.
Claude Code· 2026-07-24@AnthropicAI ↗ Claude Opus 5 GA: 1M token context window, thinking on by default, 512-token prompt cache minimum, full effort ladder including max — same $5/$25 per MTok pricing as Opus 4.8
On July 24, 2026, Anthropic launched Claude Opus 5 (model ID claude-opus-5) as a step-change improvement over Claude Opus 4.8, with the largest gains in deep reasoning, agentic and long-horizon tasks, and test-time compute scaling. Core specs: 1M token context window (1M is both the default and the maximum — there is no smaller context variant), 128k max output tokens, thinking on by default (adaptive; no thinking field needed), full effort ladder (low, medium, high, xhigh, max) including the max level for deepest reasoning; disabling thinking with effort xhigh or max returns a 400 error, a breaking change from Opus 4.8. Prompt cache minimum drops to 512 tokens (down from 1,024 on Opus 4.8). Pricing: $5 / $25 per MTok, same as Opus 4.8; fast mode research preview (Claude API only) at $10 / $50 per MTok. Available on Claude API (claude-opus-5), Amazon Bedrock (anthropic.claude-opus-5), Google Cloud Vertex AI (claude-opus-5), and Microsoft Foundry. Beta features launching alongside: ① mid-conversation tool changes (beta header mid-conversation-tool-changes-2026-07-01) — add or remove tools between turns while preserving prompt cache hits; ② fallbacks "default" mode (beta header server-side-fallback-2026-07-01) — Anthropic automatically selects recommended fallback models by refusal category. Migration notes: thinking is on by default so revisit max_tokens for workloads that ran without thinking on Opus 4.8; remove any "include a final verification step" instructions carried from earlier models — Claude Opus 5 verifies its own work without prompting and those instructions cause over-verification.
Claude Code· 2026-07-24@AnthropicAI ↗ Claude Code v2.1.219: Opus 5 as default model, sandbox.network.strictAllowlist, DirectoryAdded hook, subagent nesting depth 3, dynamic workflow medium default, and 10 bug fixes
v2.1.219 is built around the Claude Opus 5 launch. New features: ① claude-opus-5 is the new default Opus model (1M context, Fast mode at $10/$50 per MTok); Opus 4.7 removed from Fast mode — /fast now covers Opus 5 and Opus 4.8 only; ② sandbox.network.strictAllowlist setting denies non-allowlisted hosts for sandboxed commands without prompting; ③ DirectoryAdded hook fires after /add-dir or SDK register_repo_root registers a new working directory mid-session; ④ headless stream-json init event gains mcp_server_errors listing --mcp-config entries skipped by config validation; terminal runs print a startup warning; ⑤ workflowSizeGuideline settings key sets the Dynamic workflow size guideline from any settings file (/config row hidden when set); ⑥ depth-2+ subagents appear in stream-json when --forward-subagent-text is set, keyed by the spawning Agent tool_use id. Behavioral changes: dynamic workflow default changed to medium (aim for fewer than 15 agents); subagent nesting depth raised from 1 to 3 — restore with CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1; Managed MCP ${VAR} entries now resolve from startup environment and managed-settings env; claude-api skill updated to default to Opus 5 with migration path from Opus 4.8. Bug fixes (10): claude -p drops text already produced when a turn dies on a mid-stream API error; permission approved during self-hosted runner restart dropped on session resume; Fable model row shows "Requires usage credits" for plans that include it due to stale cache; SIGTERM during runner startup leaves stale active row until lease expires; /model picker shows merged Opus row as "Opus" instead of "Opus (1M context)"; copy-on-select inside GNU screen prints base64 into terminal; Remote Control clients keep stale fast-mode status after model switch, reconnect, or failed org check; CLAUDE_CODE_GIT_BASH_PATH on Windows used as bash when path is not a bash/sh binary — now ignored with warning; Vim mode pressing ← on empty prompt from NORMAL mode now correctly returns to agent view; screen-reader mode rewrote entire input line on every keystroke instead of echoing only the typed character.
MiMo Code· 2026-07-24@XiaomiMiMo ↗ MiMo Code v0.1.9: planning engine improvements, MiMo free API sunset enforced, enhanced slash command discovery, and workflow tools now opt-in
v0.1.9 ships 8 merged PRs (#1877, #1883, #1884, #1886, #1887, #1893, #1900, #1901) with these changes: ① planning engine improvement — non-Yes answers to plan_exit/plan_enter now guide the model to continue planning rather than exit early; ② provider fix — OpenAI response header wait times are now bounded, preventing hangs from slow or missing response headers; ③ TUI enforces MiMo free API sunset, prompting users to migrate to the paid API; ④ enhanced slash command discovery in the TUI; ⑤ fix for GPT subagent tool catalog resolution by catalog model; ⑥ workflow tools changed to opt-in (no longer enabled by default); ⑦ Python toolchain ruff rules aligned with upstream defaults, adding UP (pyupgrade) and I (isort) rule sets. Three contributors: yanyihan-xiaomi, qiaozongming, and MiMoHardFather.
MiMo Code· 2026-07-23@XiaomiMiMo ↗ MiMo Code v0.1.8: built-in /compose-next command, Codex/GPT login support, Office skill MIMO_* bundled runtimes, and bug fixes
v0.1.8 ships four additions: ① a built-in /compose-next command lands alongside the now-deprecated Compose Agent, with migration recommended; ② Codex (GPT) login option added with model-filtering removed — full GPT model series supported, GPT tool guidance and skill search aligned; ③ Office skills (docx/pptx/xlsx) now resolve bundled runtimes via MIMO_* env vars, fixing the Windows soffice profile URI bug and adding per-OS CJK font guidance; ④ mimocode-docs skill updated, provider configuration guidance hardened, and missed v0.1.6/v0.1.7 feature coverage synced. Bug fixes: dream/distill writes now confined to memory and .mimocode directories; empty-content user message creation guarded at createUserMessage call site and schema level; session time_updated preserved during global re-parent operations; checkpoint writer recovery contained. Maintenance: unused sst and @actions/artifact packages removed, eliminating an unlicensed transitive dependency; reminder logic simplified with a new enable flag.
Claude Code· 2026-07-22@AnthropicAI ↗ Claude Code v2.1.218: /code-review as background subagent, /deep-research manual-only, Plan+Auto Bash via classifier, skill fork background by default, and 20+ bug fixes
v2.1.218 ships ten feature and behavioral changes: ① /code-review now runs as a background subagent so review work no longer fills the conversation context; stacked slash commands keep the stack as the review target; ② /deep-research no longer launches automatically — it must be invoked explicitly; ③ in Plan mode with Auto, Bash commands the static analyzer cannot prove read-only are now judged by the auto-mode classifier instead of opening a permission dialog; ④ fast mode changes via /config model=<x> or Remote Control now show an announcement; ⑤ server-managed settings: benign feature and cost toggles no longer trigger approval prompts; ⑥ agent names in Agent markdown files may no longer contain ":" (reserved for plugin namespacing); ⑦ skills with context: fork now run in the background by default — opt a skill out with background: false; ⑧ frontmatter boolean fields now accept yes/no/on/off/1/0 (case-insensitive) alongside true/false; ⑨ claude mcp list and /mcp now display HTTP status and error text when an MCP server fails to connect, with warnings for MCP config values that have hidden leading or trailing whitespace; ⑩ --ax-screen-reader mode adds screen-reader announcements for Option+Delete, Ctrl+W, Cmd+Backspace, Ctrl+U, and Ctrl+K deletions. 20+ bug fixes: Windows paths with \u-prefixed segments (e.g. C:\Users\unicorn) were corrupted into CJK characters in tool inputs; left arrow key silently discarded conversations with no undo — now prompts for confirmation; multi-line paste collapsed to a single line with j where newlines were encoded as Ctrl+J; /context reported stale pre-compact token usage after compacting from the message picker; /ultrareview rejected descriptive arguments like "review my auth changes" — now runs a review of the current branch with the text as a note; /code-review ultra silently ran a local review in non-interactive sessions — now launches cloud review; Bedrock application-inference-profile ARNs priced at incorrect model rates; long IDE selections truncated mid-emoji caused mojibake; silent tool executor errors could be dropped; engine teardown race started and abandoned phantom turns; spurious "[Request interrupted by user]" messages appeared after interrupted tool calls; VoiceOver read "new line" instead of echoing a typed space at input end in screen-reader mode; plugin and settings panels did not move the terminal cursor to the focused row for screen readers or magnifiers; deeply nested watched directory deletions or moves triggered maximum-call-stack crashes; PR events occasionally lost when a session exited immediately after creating or linking a PR; Bedrock setup wizard failed profile verification for assume-role profiles in partitioned AWS regions and proxy-only networks; monotonic clock adopted to fix rare negative or incorrect turn-duration measurements; MCP authentication startup notice over-counted claude.ai connectors; prompt history entries dropped or duplicated under concurrent or failed writes; context-overflow retry loop re-sent identical requests with a large thinking budget — Ctrl+B backgrounding now applies the same caps as other paths; agent frontmatter hooks ran from untrusted folders — now require the agent file's folder to have accepted workspace trust; fork-session lineage lost after compaction in headless and SDK sessions; sessions with malformed delta attachments crashed on resume or failed every turn; remote sessions continued sending heartbeats after worker replacement causing long-lived processes to endlessly retry rejected requests. UX improvements: dangerous-rm, background-&, and suspicious-Windows-path checks no longer open permission dialogs and are adjudicated by the auto-mode classifier; sandbox command restrictions improved for IDE interactions; trust dialogs now name the repository root the grant covers; /ultrareview error feedback improved so Claude can correct invalid arguments rather than retrying unchanged.
Claude Code· 2026-07-22@AnthropicAI ↗ Claude Managed Agents July 22: model effort level, environment + memory-store lifecycle webhooks (7 event types), initial_events session pre-launch, thread event stream deltas, and optional agent version field
July 22, 2026: Claude Managed Agents ships five API additions in one update. ① model effort level — pass effort inside the model object when creating or updating an agent to set the reasoning effort (low, medium, high, or max); see Effort levels. ② Environment and memory-store lifecycle webhooks — Managed Agents webhooks now cover four environment.* event types and three memory_store.* event types (seven total), letting you react to environment and memory store lifecycle changes without polling; see Subscribe to webhooks. ③ initial_events session pre-launch — pass initial_events on POST /v1/sessions with up to 50 user.message or user.define_outcome events; a non-empty list starts the agent loop in the same call, eliminating a separate send-events request to start work. ④ Thread event stream deltas — GET /v1/sessions/{id}/threads/{thread_id}/stream now accepts the same event_deltas[] query parameter as the session-level stream, allowing real-time preview of a subagent's text as it generates; each connection previews only the thread it is reading. ⑤ Optional version field on agent updates — version is now optional when updating a Managed Agents agent: omit it for an unconditional update, or supply it for optimistic concurrency (a version mismatch returns 409); see Update semantics.
