Claude Code adds /goal checks for long-running tasks

• Anthropic shipped /goal in Claude Code, a completion check that keeps a session running until a stated condition is met, according to ClaudeDevs' /goal announcement and ClaudeDevs' command example.
• The same thread framed /goal as one part of a longer-running toolkit that also includes /loop, /schedule, stop hooks, and auto mode, as ClaudeDevs' /loop post, ClaudeDevs' /schedule post, and ClaudeDevs' stop hook post laid out.
• Fast mode for Claude Opus 4.7 is now in research preview on the API and in Claude Code, with Claude Code users able to opt in now and getting it as the default fast mode on Thursday, per ClaudeDevs' fast mode post and ClaudeDevs' rollout timing post.
• Anthropic also said fast mode is already showing up in partner tools including Cursor, Warp, Windsurf, v0, Emergent, and Factory, according to ClaudeDevs' integration list and bentossell's FactoryAI repost.
• Outside Anthropic's own rollout thread, users were already treating /goal as a sticky workflow pattern in other coding agents too, while one viral complaint about Opus 4.7 claimed non-English prompts can spike usage much faster, per LLMJunky's Codex steps and om_patel5's tokenization complaint.

You can read Anthropic's fast mode docs, the goal docs, and the API waitlist page. Anthropic's own thread also bundled a tiny taxonomy of long-running controls, from /goal and /loop to /schedule and stop hooks. On the rollout side, the company named Cursor, Warp, Windsurf, v0, Emergent, and Factory on day one.

Anthropic's most concrete addition is /goal, which turns a vague "keep going" request into a named completion condition.

The example prompt is plain: /goal all tests in test/auth pass and the lint step is clean. In Anthropic's description, Claude checks that condition whenever it tries to stop, then either keeps working or returns a "Goal achieved" summary, as ClaudeDevs' Ralph loop explanation put it.

Anthropic bundled /goal with four other controls that make more sense as a set than as isolated commands.

• /goal: keep working until a completion condition is met, per ClaudeDevs' example.
• /loop: rerun Claude repeatedly for iterative refactors or backlog burn-down, per ClaudeDevs' /loop post.
• /schedule: trigger Claude on a cadence such as nightly test runs or weekly cleanup, per ClaudeDevs' /schedule post.
• Stop hooks: gate whether Claude is allowed to finish, including test-suite runs or CI endpoints, per ClaudeDevs' stop hook post.
• Auto mode: remove the human confirmation bottleneck for longer runs, enabled with Shift+Tab in the CLI or via the desktop mode selector, per ClaudeDevs' auto mode post.

The useful detail here is that Anthropic is naming several ways to keep the agent alive after the first draft. /goal handles end conditions, stop hooks add programmatic gates, and auto mode removes the need for constant manual approval.

The second ship is rollout plumbing for Opus 4.7 fast mode.

Anthropic said fast mode is in research preview on the API and in Claude Code, with Claude Code users opting in first and then getting Opus 4.7 as the default fast-mode model on Thursday. For API users, Anthropic pointed people to a waitlist rather than broad availability, via ClaudeDevs' API waitlist post.

The docs links in the thread point to Anthropic's fast mode documentation and goal documentation, which is where the company is putting the operational details rather than in a blog-style announcement.

Anthropic did not keep the fast-mode preview inside its own surfaces.

The named day-one integrations were:

• Cursor
• Emergent
• Factory
• v0
• Warp
• Windsurf

That is a pretty fast ecosystem spread for something still labeled research preview. Factory's repost also showed the model name directly in product UI, which suggests partners were ready with the surface-level wiring before Anthropic made the thread public.

/goal is already spilling out as a generic coding-agent pattern, not just a Claude Code command.

One user documented how to surface /goal inside the Codex app through a command palette flow, then joked about Codex "one shot"-ing the assigned goal. That does not prove shared internals, but it does show the same interface idea landing across coding agents at roughly the same moment: give the model a persistent end-state, then let it run until it hits it.

The last useful wrinkle came from outside Anthropic's announcement thread. om_patel5's tokenization complaint claimed Opus 4.7 can burn through a paid session much faster on German prompts than on equivalent English prompts, attributing the gap to tokenizer behavior rather than a transient bug.

That post cited a Reddit complaint and gave a specific comparison: the same user's English prompt allegedly consumed 37% of a session on Opus 4.7, while the German version consumed 100%. Anthropic did not mention any language-specific usage caveat in the rollout tweets, so for now this sits as an attributed user report around the same Opus 4.7 release cycle.