[Lazarus-931]

this shortens the built-in system prompt so nativ sends fewer prompt tokens during prefill while keeping the same profile, style, tool, browser, terminal, and safety instructions.

i tested the original and shorter prompts on five model sizes: qwen3-4b, llama-3.1-8b, glm-4.7-flash, qwen3.5-122b-a10b, and deepseek-v4-flash optiq. the shorter prompt matched the original semantic score on every model in this test suite, used around 46% fewer prompt tokens, and reduced median uncached prefill by 42.5% in this benchmark.

i also compared the original pr 635 chat_work prompt with the rewritten routing-focused version using exact tool-name, action, argument, generated-content, and plain-text assertions. the original and rewritten scores were:

model	original	rewritten
llama-3.2-1b	0/7	0/7
qwen2.5-1.5b	1/7	2/7
llama-3.2-3b	4/7	4/7
qwen3-4b	6/7	6/7
mistral-7b-v0.3	2/7	2/7
llama-3.1-8b	4/7	5/7
gemma-3-12b	1/7	5/7
gemma-3-27b	1/7	4/7
glm-4.7-flash	7/7	7/7
llama-3.3-70b	5/7	5/7
qwen3.5-122b-a10b	6/7	6/7
deepseek-v4-flash optiq	1/7	1/7

deepseek loaded and passed the plain-text case, while its six tool cases hit the same engine keyerror in both arms.

across that matrix, the rewritten chat_work prompt scored 47/84 versus 38/84 for the original, used 13.7% fewer prompt tokens on average, and reduced estimated prefill by 15.1%. it matched or improved the original score on every model in the matrix.

llama 8b scored 26/27 with both prompts because it labeled one meta-question incorrectly, but its response still followed the requested terse, no-emoji, unformatted style. these numbers are from these model benchmarks and are not meant as a universal speed claim.

the nativ app builds and 78 focused NativSettingsTests and ChatWorkTests pass.

[lucasnewman]

@Lazarus-931 Can you file an issue for the DeepSeek model failure? Otherwise this looks fine to me.

[Lazarus-931]

@Lazarus-931 Can you file an issue for the DeepSeek model failure? Otherwise this looks fine to me.

I'll follow up with that!