Early DS4 layers use token-id hash routing instead of top-k routing. */

Dispatch for dense F32/F16/Q8_0 tensors used by auxiliary projections. */

Do an iteration step towards the next element. At the end of the step the * iterator key will represent the (new) current key. If it is not possible

Validate Responses tool outputs before rendering. * * A tool output with a call_id is meaningful only if either: * 1. DS4 still has the matching

Decode a UTF-8 sequence starting at 's' into a Unicode codepoint. * Returns the codepoint value. Assumes valid UTF-8 encoding. */

Return true when canonicalization would replace already-sampled tokens. * * A DS4 session checkpoint is more than a token vector: the backend state

Slow reference recognizer used by tests. */

Restore the live prompt after output. The primary path draws it in the * reserved bottom rows; the fallback path keeps the older one-row-below-outpu

Final language-model head: HC collapse, RMSNorm, and Q8_0 vocab projection. */

Pop an item from the stack, the function returns NULL if there are no * items to pop. */

Insert/remove the Think Max prefix inside the existing transcript. The * prefix lives after BOS, before any system/developer text, which mirrors the

Given a buffer, position and total length, return the byte length of the * grapheme cluster at the current position. */

Submit one user turn if the worker is idle. Busy submissions are rejected so * the UI can keep the typed text editable instead of silently queueing

Drain available output, notice process exit, and enforce timeout. This is * called opportunistically by status/wait/compaction instead of a backgrou

Split a text buffer into line spans. content_end excludes CR/LF so callers * can print or compare line content without newline spelling differences.

Validate Anthropic tool results before rendering. * * A tool_result.tool_use_id is valid if it is either still bound to the live * Anthropic assist