effortless Claude Code mod: set Claude Code reasoning effort automatically

effortless is a community mod by HeyCubit (HeyCubit/effortless). It sets the Claude Code reasoning effort for every prompt you send. Before the request goes out, a small judge reads your message and the recent conversation and picks an effort from low to max, so quick questions run at Low and hard jobs at High. You do not touch the Effort control. Short follow-ups such as "go" or "yes" keep the effort already chosen, and a message that names a level ("max effort") runs at it with no judge.

Above the prompt it draws a band. According to the author's README, the band shows the effort and model in use, how full the context is, a countdown until the prompt cache goes cold, an Auto switch, a settings gear, and a Compact button. When Haiku judges that a fresh chat would pay off, Compact becomes a Handoff button that writes a summary, clears the chat and carries on. Further down it can warn about a swamped context, running-low plan limits and a cold cache, and it adds an agent panel for subagents.

The mod is large: one 5,400-line module plus a bundled second plugin (textless) in the same repository, which we did not onboard. The author's README claims 21% lower cost than Opus on medium in its own benchmark; we did not reproduce that.

There is no recording yet. The mod needs its own setup guide, a judge call before every prompt and a few turns before the band has anything to show, and the author's test suite fails 52 of 274 tests on Claude Code 2.1.303, including its update tests (see below). We kept this page as a draft rather than record a mod we could not verify end to end.

The install block above adds the author's marketplace (HeyCubit/effortless, which lists effortless and textless) and installs the plugin from it. We tested commit df5602e on the default branch; the marketplace entry itself points at the author's stable branch, which may be a different commit. Run /reload-plugins or restart Claude Code. A short setup guide opens above the prompt the first time: keep Haiku alone or add Jev, lean cheaper or smarter, and choose how handoffs are written. /effortless setup opens it again.

The plugin is one function-hooks module, hooks/register.tsx. On prompt.submit it asks the judge for a verdict, and on turn.step it rewrites the outgoing request's effort. Compaction is also taken over by a session.compact hook:

on('session.compact', async ($, e, next) => {
  // A subagent's own compaction, the session's model by choice: Claude Code's own.
  if (e.agentId !== undefined || config.compactWith !== 'haiku') return next(e)
  // Ahead-of-time summaries would be written by the chat's model and then not used: none.
  if (e.trigger === 'precompute') return { skip: 'effortless compacts with Haiku when it is time' }
  const messages = Array.isArray(e.messages) && e.messages.length ? e.messages : await $.session.messages().catch(() => [])
  const got = await haikuCompaction($, messages, e.instructions).catch((error: unknown) => ({ fail: String(error).slice(0, 80) }))
  ...

The judge calls $.model.complete with Haiku 5.5, falling back to the haiku alias, with a 6 second timeout. With a TypeSafe key it asks Jev first over HTTP and falls back to Haiku when Jev is unsure, slow or down. The cache countdown comes from the cache lifetime in each response's usage, and Compact appears once it reaches zero.

The plugin declares 16 options (/plugin configure effortless@effortless or the gear in the band): the judge (auto, haiku, jev), an effort slider from -2 to 2, a floor and ceiling for Auto, which alerts to hide, the handoff skill and first choice, "Compact with" (haiku or session), "Cheaper model when it can" (off by default), "Effort for helpers" (off by default), a swamp threshold, and appearance, theme and layout (default band or minimal footer buttons). The /effortless command also takes auto, stats, handoff, settings, setup and debug.

We read the manifest, hooks/hooks.json, commands/effortless.md and the parts of hooks/register.tsx that touch the network, processes, files, environment variables, the model and prompts. Of the other modules (agents.tsx, progress.tsx, theme.ts, art.ts, brand-mark.ts, about 100 KB) we only searched for calls and encoded or hidden code; we did not read them line by line, and we did not read the textless plugin or the site, bench and tools folders, which the plugin does not load. There is no package.json, no bundled or minified code, and no binary. Findings:

• Network ($.http.fetch): https://api.typesafe.ai/v1/systemone (or the JEV_URL environment variable) only when a TypeSafe key is found; it receives the next message (up to 4,000 characters), the recent conversation and the current model and effort, with the key as a Bearer token (register.tsx:724). The release check reads api.github.com/repos/HeyCubit/effortless/contents/public.json (register.tsx:2685).
• Processes ($.process.run): git for the release check, creating a small repo under the Claude data folder (.../data/effortless-release-check) and running a depth-1 fetch from github.com/HeyCubit/effortless (register.tsx:2668-2678), with no prompt; claude plugin marketplace update effortless when git fails (:2694); on Update, claude plugin marketplace update and claude plugin update, then cp -R of the new version over the running plugin cache folder (:2840-2891); on Uninstall, claude plugin uninstall after two presses (:2529). All are argv arrays, not shell strings.
• File writes ($.fs.write): a key pasted in the setup guide or Judge card is saved in plain text to ~/.config/jev/.env (:1893-1917); the README discloses this and points to /plugin configure, which uses Claude Code's secret storage. A debug log ~/.claude/effortless-render-<chat>.log is rewritten every second with draw timings (:2262-2272). /effortless bench writes two report files in the current directory (:4024). A proof log is written only if EFFORTLESS_LOG is set or the mod runs from a dev-mods folder. It writes nothing to settings.json, shell files, LaunchAgents, cron or git hooks. Settings go through $.config.set and its own store.
• Reads: ~/.claude/effortless-preview.json (a preview aid), ~/.config/jev/.env (only if the jev judge is picked outright, or after you saved a key there), and the plugin's own plugin.json.
• Daemons: none. It uses $.clock.every timers that stop with the session.
• Environment: EFFORTLESS_LOG, EFFORTLESS_MODEL_UI, HOME, JEV_URL, TEMP, TMPDIR, TYPESAFE_API_KEY, USERPROFILE. The key leaves the machine only to the Jev endpoint.
• Telemetry: none found. The release check sends no data beyond a normal git or HTTPS request to GitHub.
• Model influence: it rewrites each request's effort, and with "Cheaper model when it can" on (default off) it may send a prompt to a cheaper model. On session.compact it replaces Claude Code's summary with one written by Haiku 5.5 (default on): your transcript goes to Haiku through your own login, and the summary is injected as the first message with a fixed "continue the conversation" text (register.tsx:3800). The handoff submits its own prompts and runs /clear. We found no hidden or zero-width text and no instructions aimed at the model beyond these prompts. commands/effortless.md is only used when the mod did not load.
• Gating hooks: session.compact, prompt.submit, agent.spawn, tool.call, command.run and others have no .catch, as validate notes. tool.call only observes. session.compact, classic.SessionStart and classic.PostModelSwitch do not run for a plugin you install where an organization manages Claude Code.
• External requirements: none required. The claude CLI is needed for Update and Uninstall; a TypeSafe account only for Jev.
• License: MIT, LICENSE file present.

claude plugin validate reported 1 error and 1 warning: effortless.updateCheck is not declared: the manifest's types contract must name it (the update button's state is missing from types/index.d.ts), and no author in plugin.json. claude plugin test ran 274 tests across 15 files: 222 passed and 52 failed. Many failing test names look like textless features (scenes, cards, chess), but the failures also include effortless's update tests (Update, Uninstall, the .local/bin fallback, the HTTPS rewrite) and the key-from-~/.config/jev/.env test, whose expected Windows-style path was joined to the working directory on macOS. We did not investigate each one. Because of the validate error and these failures, we have not marked the page live.

Checked with Claude Code 2.1.303 on macOS (validate and tests only; the mod was not loaded in a recording). The mod needs function hooks and is written for the current Claude 5 models: Auto changes the effort only on Opus 5.5 and Sonnet 5.5, and pauses on other models. If the band does not show up:

• Run /reload-plugins or restart Claude Code, then /effortless debug for a status line.
• Set Look to default; minimal draws only small footer buttons.
• If Update fails, run claude plugin update effortless@effortless yourself.
• If a key does not seem to work, check the judge setting, which is auto by default and ignores ~/.config/jev/.env unless you pick jev.