Changelog — October 9, 2026

Nine releases in nine days; Claude Mods and Haiku 5.5 are the headlines. 2.1.287 adds Claude Mods: plugins may now modify deeper behavior, and a built-in mod called "You should know" can be turned on with /plugin enable cc-plugin-you-should-know@builtin (first-party sessions with telemetry on). 2.1.293 adds Claude Haiku 5.5 (claude-haiku-5-5), now the default Haiku model on the Anthropic API. If an MCP server stops connecting after 2.1.287, add "bareElicitationCapability": true to its MCP config entry.

Commands and hooks. claude project purge is now claude purge; the old name still works and prints a notice (2.1.288). /code-review takes --max-findings <n>|all (2.1.288). claude attach and claude logs accept part of a session name instead of the ID (2.1.290). claude plugin install gains --marketplace <source> (2.1.292), and the Agent tool gains an effort parameter. Command and HTTP hooks accept onFailure: "block" (2.1.295): a hook that cannot start, times out or exits with an unexpected code now blocks the action instead of letting it through.

Environment variables and defaults. New: CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS caps how long unattended retry mode waits out 429 and 529 errors (2.1.295); CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS sets a longer base delay for 529 retries (2.1.292); CLAUDE_CODE_DISABLE_STRUCTURED_OUTPUTS turns structured outputs off for gateways that reject them (2.1.288); CLAUDE_CODE_WEB_SEARCH_REFILLS_PER_HOUR sets the rate at which an interactive session's WebSearch budget refills, which is now 100 calls an hour instead of ending after 200 (2.1.290). Changed: a repository's .claude/settings.json can no longer set CLAUDE_CODE_DISABLE_ATTACHMENTS (2.1.290), and local stdio MCP servers and claude.ai connectors now negotiate MCP protocol version 2026-07-28 by default; MCP_PROTOCOL_NEGOTIATION=legacy opts out (2.1.292, 2.1.295).

Releases → Affects: /claude-cowork/setup/, /claude-cowork/pricing/, /claude-cowork/faq/effort-levels/, /claude-cowork/faq/claude-code-undocumented-configuration/, /commands/

Claude Haiku 5.5 launched on 2026-10-07 as claude-haiku-5-5, with a 1M-token context window, 128K max output, and adaptive thinking with the effort parameter. It is priced by prompt length: $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, and $0.50 and $2.50 above that (cache reads $0.01 and $0.05). A request's prompt length counts all input tokens including cache reads and writes, and each request is priced on its own. Haiku 4.5 listed at $1 and $5.

Code written for Haiku 4.5 can break. Manual extended thinking (budget_tokens) now returns a 400 error, adaptive thinking is on by default so a response can begin with thinking blocks, and the same text counts as more tokens. Anthropic publishes a migration guide and a Haiku 5.5 prompting page.

Also announced. Prompt-cache reads on Claude Sonnet 5.5 dropped from $0.20 to $0.10 per million tokens; cache writes and other prices are unchanged. Claude Max and Team plans now include monthly Claude API credits: $100 (Max 5x), $200 (Max 20x), and $20 per Standard or $100 per Premium Team seat, pooled and capped at $500; they cover the Claude API, Managed Agents and the Agent SDK but not Claude Code, and do not roll over. In Claude Managed Agents, web_search and web_fetch now follow allowed_hosts under limited networking, and web_fetch only fetches URLs that have already appeared in the session (otherwise it returns url_not_in_prior_context).

API release notes → Affects: /claude-cowork/, /claude-cowork/pricing/, /claude-cowork/vs-api/, /claude-cowork/faq/effort-levels/, /tools/cost-calculator/

2026.9.9 adds model support and better update recovery. It adds GPT-6.1 Sol to the Codex model list and supports Claude Haiku 5.5 through Anthropic and Claude CLI, improves recovery after failed updates, fixes missing iMessage replies, and stops an old scheduled job from interrupting a newer conversation. 2026.9.8 (2026-10-03) fixed missing replies between agents, reduced memory use when running many Codex agents, and made update repair keep the plugins you had allowed and enabled.

Two other tracks. 2026.8.35 (2026-10-02) is a gateway-only "extended-stable" release, which its notes call OpenClaw's current equivalent of LTS: the end-of-August build plus critical security fixes and GPT-6.1 Sol support. 2026.10.1-beta.1 and beta.2 (2026-10-08) are on the npm beta channel and add Code Mode values that persist across cells and Gateway restarts, click-to-play YouTube cards in the Control UI, and import of coding-agent transcripts into durable sessions; beta.2 is a 40-commit hotfix. Betas are not for production installs.

Release notes → Affects: /openclaw/, /openclaw/setup/, /openclaw/skills-guide/, /openclaw/email/

A patch tag that rolls up about 2,100 merged PRs since v0.21.5. It is the first release cut by Hermes's new stable release pipeline (one attempt ref, a tested Docker image, a receipt tag at publish) and ships the tag, the GitHub release and the Docker image. The desktop app, Termux packages and the Microsoft Store stay on their current build and move with the next bundled release; full curated notes arrive with v0.22.0.

Listed in the window but not documented in the notes: live transcription while you speak (stt.streaming), spoken voice turns routed to their own model (auxiliary.voice_chat), third-party plugins in a per-profile plugin host (plugins.isolation: host), and a summary-first hermes status with --short and --full. Treat these as leads until v0.22.0. Tags such as rc.5-v0.21.7 in the releases feed are release-candidate attempts, not a release: 0.21.7 has not shipped.

Releases → Affects: /hermes/, /hermes/setup/, /hermes/memory/, /hermes/dashboard/

Two stable releases since 7.8.1. The multi-project Agent Manager became the default in the 7.8.7 pre-release (the experimental toggle and the old single-project view are gone). 7.8.8 builds on it: you can now pin worktrees in the Agent Manager sidebar (Shift+click or the context menu) and opens chat web links in-app when the experimental Integrated Browser is on; choose "System browser" under "Open links in" to opt out. 7.8.7 also lets the Integrated Browser open in an editor tab for any session through the Open Integrated Browser command.

7.8.3 adds a manual Fetch Models button to the Custom Provider settings dialog, suggests files from every folder of a multi-root VS Code workspace when you type @, and adds a copy-content button to the Documents viewer toolbar.

Releases → Affects: /kilocode/, /kilocode/setup/, /kilocode/orchestrator/

No release tags, so nothing to bump; the commit stream is mostly fixes. The last 20 commits recover confirmed legacy sandboxes after failed upgrade probes, restore shared native NVIDIA sandbox requests, authenticate the Llama Guard service before loopback publication, keep forward ports occupied instead of failing onboarding when ownership can't be proven, explain WSL GPU readiness, and restore native startup and NVIDIA inference on Brev.

A command to re-check. NVIDIA's command reference (docs/reference/commands.mdx on main, read 2026-10-08) no longer lists nemoclaw <name> snapshot or its create, list and restore subcommands. The backup command it does list is nemoclaw backup-all. If older instructions tell you to run snapshot create before maintenance, confirm it against your installed version first.

Commits → Affects: /nemoclaw/, /nemoclaw/setup/

OpenAI added Ultrafast mode for GPT-6.1 Sol on 2026-10-08. Set service_tier: "ultrafast" on a gpt-6.1-sol Responses API request to cut the time between generated output tokens. It is available to all API users subject to rate limits, with global processing and US and EU data residency. It is a premium tier: $12 input, $0.60 cached input, $15 cache write and $60 output per million tokens for prompts up to 272K tokens ($24, $1.20, $30 and $90 above that), six times the standard $2 and $10. Standard GPT-6.1 Sol prices are unchanged.

Also from OpenAI. The Decisions API is in beta with gpt-6-luna (2026-10-06): it turns text and images into typed answers, which OpenAI says is 10 times faster than the Responses API. API usage tiers were simplified from five to three: Build, Launch and Grow.

Pricing → Affects: /chatgpt/pricing/, /chatgpt/api-vs-chat/, /tools/cost-calculator/