The prompt renderer escapes DSML text so a tool argument can safely contain * shell operators or closing tags. The generated-DSML parser must undo e

The input embedding starts all HC streams with the same token vector. */

Sink-aware attention over a set of KV rows. The learned sink logit is part * of the softmax denominator but contributes no value vector. */

Attention output projection is grouped: each group first maps its heads to * a 1024-rank low vector, then all groups are projected back to 4096. */

KV projection has one KV head of width 512, followed by a learned RMSNorm. */

Decode version of routed MoE: same math as layer_routed_moe_one(), but all * large temporaries come from the persistent scratch arena. */

Beep, used for completion when there is nothing to complete or when all * the choices were already shown. */

Return the number of bytes to move left from pos. If pos is at the end of a * folded range, the whole hidden range is skipped by one cursor movement.

Clear all remembered fold ranges. */

This is just a wrapper the user may want to call in order to make sure * the linenoise returned buffer is freed with the same allocator it was * cre

This is the API call to add a new entry in the linenoise history. * It uses a fixed array of char pointers that are shifted (memmoved) * when the hi

Single live protocol-tool state. * * This is not an implementation of durable remote conversation storage. It is * only an in-memory binding from

Batched Q8_0 matmul for prefill: quantize all token activations, then scan * weight rows once per output channel. */

Allocate the Metal graph state for a chosen raw-cache capacity. The model * weights are not copied here; tensors reference the mapped GGUF. */

Encode one complete layer for prefill by chaining attention and FFN batches. */

Pick a raw SWA cache size for Metal. During batched prefill it must cover * the previous window plus the current ubatch. */

Like raxIteratorNextStep() but implements an iteration step moving * to the lexicographically previous element. The 'noup' option has a similar * ef

Allocate a new non compressed node with the specified number of children. * If datafield is true, the allocation is made large enough to hold the *

Calls the two low level functions refreshSingleLine() or * refreshMultiLine() according to the selected mode. */

Agent session IDs are intentionally independent from the rendered transcript: * once a session has a title and creation time, resaving it keeps the s

Shut down the worker and release owned resources, including any live bash * process groups. */

Synchronize the live DS4 session to a transcript. This is the agent's main * cache-saving operation: if the requested transcript extends the live se