Changelog

When AI is enabled, "Chat with Duncan", "Clear Duncan Chat", and

"Close Duncan" now appear at the very top of the command palette

results. Previously they were buried after navigation and sync commands.

"Clear Chat" calls ClearAll on the AI tray, clearing all messages,

intent suggestions, and content suggestion history in one action.

AI commands are conditionally included — they don't appear when AI

is disabled, keeping the palette clean.

Version: 1.50.7 → 1.50.8

Local models (Phi-3, Mistral) ignore multi-rule system prompts and parrot

examples verbatim. Three-pronged fix:

1. System prompt collapsed to a single line with minimal directives. No

numbered rules, no examples. Small models handle one dense instruction

paragraph better than structured rule lists.

2. Sanitizer expanded to strip markdown bold/italic, markdown headers,

bullet prefixes, "Note:" disclaimer lines, and all known chat template

tokens (including Llama 3.x eot_id/start_header_id variants).

3. Hard sentence-count truncation per tier: Short caps at 2 sentences,

Medium at 5, Long at 30. Applied after all other sanitization so the

model physically cannot produce a wall of text for a greeting.

Token budgets also tightened: Short 100->60, Medium 300->250.

Three fixes for the Duncan chat tray:

1. Removed example conversations from system prompt — local models were

parroting them verbatim instead of following the behavioral guidance.

Replaced with terse directive-only prompt.

2. Added response sanitizer that strips raw chat template tokens

(<|assistant|>, <|end|>, <|im_start|>, etc.) and role prefixes

("- User:", "Assistant:", "Duncan:") that local models leak into output.

3. Injected user's display name (from app settings) into the system prompt

so Duncan addresses the user by name naturally.

Introduces a three-tier response budget system (Short/Medium/Long) that

classifies the user's message intent before sending the AI request. Short

tier (greetings, yes/no, can't-do topics) caps at 100 tokens with 1-2

sentence guidance. Medium tier (explanations, how-to, brainstorming) allows

300 tokens and 3-5 sentences. Long tier (summarize a page, draft an email,

long-form writing) allows 800 tokens with paragraph-level output. The

classifier uses keyword matching with message length heuristics, and the

system prompt length rule adapts per tier. All configuration remains

centralized in AiPersona.cs.

The previous system prompt was a single generic sentence that the model

ignored, producing wall-of-text responses with repetitive AI disclaimers.

Replaced with a strict rule-based prompt enforcing 1-2 sentence replies,

no self-identification, no filler, and clear boundaries around what Duncan

can and cannot do (no internet, local-only). Dropped MaxTokens from 1024

to 200 and Temperature from 0.7 to 0.4 to further constrain verbosity.

Moved the system prompt into AiPersona.cs alongside the Name constant so

all persona configuration lives in one file.