Cap persona + diary at 100 tokens; shorten prompts; stream the diary

- maxTokens 100 for both persona and diary (was 220/768 via thinkMaxTokens) so models
can't ramble past the answer. Prompts tightened to match: persona about → "1-2 short
sentences (~25 words)"; diary → "1-2 vivid sentences (~60 words), be brief".
- War diary now streams visibly: it used the aggressive stripThink live, which an
unclosed <think> emptied — so it showed blank until the end. Use stripThinkFinal in
the live onToken so the prose (and narration) appear token-by-token.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

web/diaryPanel.js

@@ -5,7 +5,9 @@
 import { streamChat, ensureModel, currentModel, currentModelId, getEngineId, backendLabel } from '/web/runtime.js'
 import { mountTtsBar } from '/web/ttsBar.js'
 import { makeNarrator, ensureTts } from '/web/tts.js'
-import { DIARY_SYSTEM, diaryUserPrompt, stripThink, stripThinkFinal, noThink, thinkMaxTokens } from '/web/personaPrompts.js'
+import { DIARY_SYSTEM, diaryUserPrompt, stripThinkFinal, noThink } from '/web/personaPrompts.js'
+
+const MAX_TOKENS = 100 // short diary entries — cap matches the "~60 words" prompt
 
 function el(tag, props = {}, kids = []) {
   const n = document.createElement(tag)
@@ -54,7 +56,7 @@ export function mountDiaryPanel(host) {
       '=== TINY ARMY · DIARY DEBUG ===',
       `engine:   ${getEngineId()} · ${backendLabel()}`,
       `model:    ${currentModelId()} (${currentModel().label})`,
-      `input:    unit=${unit.value} traits=${traits.value} maxTokens=${thinkMaxTokens(currentModelId(), 220)}`,
+      `input:    unit=${unit.value} traits=${traits.value} maxTokens=${MAX_TOKENS}`,
       `outcome:  ${outcome}`,
       `--- raw output (${(raw || '').length} chars) ---`,
       raw || '(empty)',
@@ -139,10 +141,10 @@ export function mountDiaryPanel(host) {
       await ensureModel((frac, label) => { status.textContent = label || `downloading ${currentModel().label}… ${Math.round(frac * 100)}% (one-time)` })
       status.textContent = `writing on your device with ${currentModel().label}…`
       await streamChat(DIARY_SYSTEM, diaryUserPrompt(unit.value, traits.value) + noThink(currentModelId()), {
-        maxTokens: thinkMaxTokens(currentModelId(), 220), temperature: 0.9,
+        maxTokens: MAX_TOKENS, temperature: 0.9,
         onToken: (piece) => {
           raw += piece
-          lastBody = stripThink(raw)
+          lastBody = stripThinkFinal(raw)
           out.textContent = header + lastBody
           if (live) { const delta = lastBody.slice(spokenLen); if (delta) { live.push(delta); spokenLen = lastBody.length } }
         },

web/personaPanel.js

@@ -5,9 +5,10 @@
 import { streamChat, ensureModel, currentModel, currentModelId, getEngineId, backendLabel } from '/web/runtime.js'
 import { extractLivePersona } from '/web/personaStream.js'
 import { parsePersonaJson } from '/web/personaParse.js'
-import { PERSONA_SYSTEM, personaUserPrompt, stripThink, stripThinkFinal, noThink, thinkMaxTokens } from '/web/personaPrompts.js'
+import { PERSONA_SYSTEM, personaUserPrompt, stripThink, stripThinkFinal, noThink } from '/web/personaPrompts.js'
 
 const CLASSES = ['Warrior', 'Ranger', 'Monk', 'Assassin', 'Mage', 'Paladin', 'Cleric', 'Knight']
+const MAX_TOKENS = 100 // short personas — keep models from rambling past the JSON
 
 function el(tag, props = {}, kids = []) {
   const n = document.createElement(tag)
@@ -59,7 +60,7 @@ export function mountPersonaPanel(host) {
       '=== TINY ARMY · PERSONA DEBUG ===',
       `engine:   ${getEngineId()} · ${backendLabel()}`,
       `model:    ${currentModelId()} (${currentModel().label})`,
-      `input:    class=${sel.value} seed=${seed.value || '(none)'} maxTokens=${thinkMaxTokens(currentModelId(), 220)}`,
+      `input:    class=${sel.value} seed=${seed.value || '(none)'} maxTokens=${MAX_TOKENS}`,
       `outcome:  ${outcome}`,
       `--- raw output (${(acc || '').length} chars) ---`,
       acc || '(empty)',
@@ -95,7 +96,7 @@ export function mountPersonaPanel(host) {
       await ensureModel((frac, label) => { status.textContent = label || `downloading ${currentModel().label}… ${Math.round(frac * 100)}% (one-time)` })
       status.textContent = `writing on your device with ${currentModel().label}…`
       await streamChat(PERSONA_SYSTEM, personaUserPrompt(sel.value, seed.value) + noThink(currentModelId()), {
-        maxTokens: thinkMaxTokens(currentModelId(), 220),
+        maxTokens: MAX_TOKENS,
         onToken: (piece) => {
           acc += piece
           thinkEl.textContent = acc  // raw view shows the model's <think> reasoning too

web/personaPrompts.js

@@ -6,7 +6,7 @@ export const PERSONA_SYSTEM =
   'fighter writes its own legend. Given a class and an optional seed, return ONE JSON ' +
   'object and NOTHING else, with exactly these keys:\n' +
   '  "name": a short evocative soldier name (2-4 words),\n' +
-  '  "about": 1-3 sentences of backstory in a heroic, slightly wry war-legend tone,\n' +
+  '  "about": 1-2 short sentences of backstory (about 25 words) in a heroic, wry war-legend tone,\n' +
   '  "specialty": a 1-3 word combat specialty,\n' +
   '  "personality": a 1-3 word personality tag,\n' +
   '  "vibe": a 1-3 word vibe.\n' +
@@ -14,9 +14,9 @@ export const PERSONA_SYSTEM =
 
 export const DIARY_SYSTEM =
   'You are a tiny soldier in the auto-battler Tiny Army, writing a short first-person ' +
-  'war-diary entry. Given your name and traits, write 2-4 vivid sentences in first ' +
-  'person about a day on the battlefield — heroic, grounded, a touch of dark humor. ' +
-  'Prose only: no headings, no lists, no preamble.'
+  'war-diary entry. Given your name and traits, write just 1-2 vivid sentences (about ' +
+  '60 words, no more) in first person about a day on the battlefield — heroic, grounded, ' +
+  'a touch of dark humor. Prose only: no headings, no lists, no preamble. Be brief.'
 
 export function personaUserPrompt(unitClass = '', seed = '') {
   const s = seed && seed.trim() ? ` Seed inspiration: "${seed.trim()}".` : ''