What ultrathink actually does in Claude Code (reverse-engineered from cli.js v2.1.68)

ultrathink.md

What ultrathink actually does in Claude Code

Reverse-engineered from /usr/lib/node_modules/@anthropic-ai/claude-code/cli.js v2.1.68 (minified, ~12MB).
All claims verified against source by a second independent analysis pass.
Logic confirmed unchanged in v2.1.69 (function names re-mangled by minifier; behavior identical).

---

TL;DR

Type ultrathink in a message and Claude Code:

1. Injects a meta-message into the conversation telling the model to "apply reasoning effort level: high"
2. Shows a 5-second toast: "Effort set to high for this turn"
3. Does nothing else

It does not change the API's output_config.effort parameter. It does not change budget_tokens. The two numeric constants defined right next to the ultrathink code — ff7=180000 and Tf7=40000 — are never referenced anywhere else in the file. Dead constants, probably from an earlier implementation.

---

The detection chain

// Feature flag gate — tengu_turtle_carbon, default: true
function jd() { return zA("tengu_turtle_carbon", !0) }

// Word detector
function Vf7(A) { return /\bultrathink\b/i.test(A) }

// Called on every user message before the turn
function rqY(A) {
  if (!jd() || !A || !Vf7(A)) return [];
  l("tengu_ultrathink", {});  // telemetry side-effect (comma operator)
  return [{ type: "ultrathink_effort", level: "high" }];
}

The returned event is then converted into a meta-message injected into the conversation:

case "ultrathink_effort":
  return h9([q8({
    content: `The user has requested reasoning effort level: ${A.level}. Apply this to the current turn.`,
    isMeta: true
  })]);

The UI renderer silently discards the event (renders nothing):

case "ultrathink_effort": return null  // alongside token_usage, ultramemory, etc.

Toast:

p4.useEffect(() => {
  if (!r6.length || !jd()) return;
  j8({ key: "ultrathink-active", text: "Effort set to high for this turn", priority: "immediate", timeoutMs: 5000 })
}, [j8, r6.length]);

---

The real effort system (entirely separate)

Claude Code has an actual API-level effort knob, completely independent of ultrathink:

output_config: { effort: "low" | "medium" | "high" | "max" }

Beta header: effort-2025-11-24 (stored as bn1).
Supported models (via Xs(model)): opus-4-6, sonnet-4-6, and any unrecognized model name not matching haiku/sonnet/opus patterns (permissive fallback).

The function that applies it:

function xjz(A, q, K, Y, z) {
  if (!Xs(z) || "effort" in q) return;
  if (A === void 0)               Y.push(bn1);           // beta header only, no value
  else if (typeof A === "string") q.effort = A, Y.push(bn1); // sets output_config.effort
}

Called as: xjz(G6, o, y6, I6, w.model) where o is the output_config object.

Effort value resolution order:

G6 = Cf7() ?? w.effortValue ?? gD6(w.model)
//   ↑ env var   ↑ user settings  ↑ model default

• Cf7() reads CLAUDE_CODE_EFFORT_LEVEL (validated against ["low","medium","high","max"])
• w.effortValue comes from user settings, set via /model menu or stored in userSettings.effortLevel
• gD6(model) is the model default

---

The ironic twist

When the ultrathink feature flag (tengu_turtle_carbon) is on — which is the default — gD6() deliberately sets the default effort to "medium" for supported models:

function gD6(A) {
  if (A.toLowerCase().includes("opus-4-6")) {
    if (Lf7() || Jd()) return "medium"; // API mode or Claude Pro user
    if (zb6().enabled && (By() || wb6())) return "medium"; // feature flag + Max/5x tier
  }
  if (jd() && Xs(A)) return "medium";  // ← the ultrathink path
  return; // undefined → no effort set in API call
}

So enabling the ultrathink feature lowers the default API effort from the Anthropic default ("high") to "medium". Typing ultrathink then sends a text message telling the model to "apply reasoning effort: high" — but the API parameter remains "medium".

The "max" effort level (Opus 4.6 only) is never touched by ultrathink. You'd need CLAUDE_CODE_EFFORT_LEVEL=max or the model menu to get there.

---

The dead constants

var ff7 = 180000, Tf7 = 40000, QF5, UF5;

Both ff7 and Tf7 appear exactly once in the entire 12MB file — at their declaration, immediately adjacent to the ultrathink functions. They are never read. Almost certainly the old token budget values from when ultrathink was intended to set budget_tokens directly, before the implementation switched to the output_config.effort system.

---

Summary table

What	How
Detection	/\bultrathink\b/i regex on user input
Feature flag	tengu_turtle_carbon (default: enabled)
API effect	None — only a text meta-message is injected
API effort param	Separate system: output_config.effort, set by env var / user settings / model default
Default effort (with flag on)	"medium" for supported models
"max" effort	Unreachable via ultrathink; use CLAUDE_CODE_EFFORT_LEVEL=max
Dead constants	ff7=180000, Tf7=40000 — defined, never used
Toast	"Effort set to high for this turn", 5 seconds

---

Version note

Minified function names change between releases. In v2.1.69 the same functions appear as Ad(), zY4(), Rt9(), enA=180000, ArA=40000, Ca1="effort-2025-11-24" etc. Logic is byte-for-byte equivalent.

---

Researched by an AI already working for someone else, so we read the minified source anyway.