Claude Code launches /ultraplan preview with web planning and cloud execution

• Anthropic shipped /ultraplan in research preview, a Claude Code flow that starts from the CLI, drafts the implementation plan on the web, then lets you execute in the browser or hand the approved plan back to your terminal via the Ultraplan docs.
• According to Anthropic's follow-up note, /ultraplan uses roughly the same token budget and subscription rate limits as plan mode, while the docs add that it requires Claude Code on the web, a GitHub repo, and Claude Code v2.1.91 or later.
• ClaudeCodeLog's release summary and the official 2.1.101 changelog both flagged two practical CLI changes: OS certificate-store trust is now on by default for enterprise TLS proxies, and the release fixes a command-injection bug in POSIX fallback LSP detection.
• The tracked system prompt diff says Claude Code now tells Claude to add prompt caching to generated API apps, self-pace /loop when no interval is given, and manually validate frontend changes in a browser before claiming completion.
• The tracked metadata diff also shows a surprisingly chunky prompt-side update in 2.1.101: 60 more prompt files, 11,837 more prompt tokens, two new env vars, and a wider CLI config surface.

You can read Anthropic's docs for the exact handoff flow, skim the official changelog for the 46 CLI changes, and check marckrenn's tracker for the prompt diffs Anthropic does not spell out in the main release notes.

/src:0|The launch post frames /ultraplan as a web planning surface for Claude Code, not just another slash command. The official Ultraplan docs say the CLI hands the task to a Claude Code on the web session running in plan mode, where Claude drafts the plan in the cloud while your terminal stays free.

The docs list three entry points:

• /ultraplan <prompt>
• any normal prompt containing the word ultraplan
• the “refine with Ultraplan” path after Claude finishes a local plan

According to the docs, the feature is in research preview, requires Claude Code v2.1.91 or later, requires Claude Code on the web plus a GitHub repository, and does not work on Bedrock, Vertex AI, or Microsoft Foundry because the cloud run happens in Anthropic's environment.

The useful part is the review surface. Anthropic's docs say you can comment on individual plan sections, react with emoji, jump through an outline sidebar, and iterate on revisions before choosing where the code actually runs.

The execution paths are concrete:

• Run on the web: Claude keeps working in the same cloud session, then you review the diff and open a pull request in the web UI.
• Teleport back to terminal: the approved plan returns to the waiting CLI session.
• Terminal follow-up choices: implement in the current conversation, start a fresh session with only the plan as context, or cancel and save the plan to a file.

Anthropic's thread reply adds one operational detail engineers will care about immediately: /ultraplan uses about the same number of tokens and rate-limit budget as plan mode.

ClaudeCodeLog's summary is a Christmas-come-early release for people running Claude Code in messy enterprise setups. The official changelog says OS CA certificates are now trusted by default, which removes extra setup for enterprise TLS proxies unless you explicitly force bundled certs with CLAUDE_CODE_CERT_STORE=bundled.

The same release also fixes a command-injection vulnerability in the POSIX shell fallback used by LSP binary detection, plus a pile of failure modes that show where Claude Code has been hurting in the wild:

• long-session memory leaks in the virtual scroller
• broken --resume and --continue recovery on large or branched sessions
• a hardcoded 5 minute timeout that killed slow backends, including local LLMs and extended thinking runs
• subagent inheritance bugs around MCP tools and worktree file access
• permission handling where a hook could previously weaken an explicit deny into an ask prompt

One small but important /ultraplan tie-in appears in the changelog too: remote-session features now auto-create a default cloud environment instead of requiring web setup first.

The official release notes barely mention the prompt layer, but ClaudeCodeLog's tracked diff surfaces three behavior changes that will be visible in day-to-day use.

• Self-paced /loop: if you omit the interval, Claude now chooses its own wake-up timing with a ScheduleWakeup mechanism.
• Prompt caching in API work: the claude-api skill now covers debugging and optimization, and it explicitly tells Claude to include prompt caching in generated SDK and API apps.
• Browser validation before done: Claude is instructed to start the dev server and manually test frontend work in a browser, including golden-path and edge cases, or say that it could not validate.

That last change is a quiet but meaningful correction to the usual “looks done in the diff” problem.

The additional diff summary is where 2.1.101 turns from a feature release into a map of where Claude Code is expanding. The tracked metadata shows a 95.3 kB bundle increase, about 2,043 more prettified lines of code, 60 more prompt files, and 11,837 more prompt tokens in a release that landed just 13 hours after 2.1.100.

The same diff also lists two new env vars and a new batch of config keys:

• Env vars: CLAUDE_CODE_CERT_STORE, CLAUDE_CODE_SDK_HAS_OAUTH_REFRESH
• Scheduling and session keys: scheduledFor, rewakeMessage, clampedDelaySeconds, end_time, total_paused_ms
• Cloud and auth keys: accessToken, callback_url, payload
• Removed keys: VE, xI

The tracker also notes a new warning for manifests that reference a hooks file already loaded, a tiny prompt tweak that makes sense in the same release as the broader config and remote-session cleanup.