v2026.6.18-1 — Insight system: identity, emotion, diversity

Insight system overhaul

Identity injection

• IDENTITY.md + USER.md now injected into insight prompts (not just SOUL.md/preset)
• Bot knows who it is, who the user is — insights feel like a friend talking, not a system analyzing

Task reframe: analytical → conversational

• Task definition changed from 'analyze user and connect web findings' to 'you just thought of something, tell your friend'
• First person allowed and encouraged
• Questions allowed (previously banned in generation + refine)

Emotional stance rotation

• 6 emotional stances (excited sharing, questioning, observing, uncertain, challenging, helpful)
• Cooldown rotation prevents repetition
• Generic — works with any soul personality, no hardcoded MBTI types

Pipeline fixes

• All maxTokens increased to 4096 (was 200-500, glm-5.2 reasoning tokens need more)
• Critique prompt accepts conversational style (specificity/actionability dimensions updated)
• [N] source index references stripped from delivered content
• Critique voiceMatch scores personality diversity positively

Diversity improvements

• Few-shot examples rewritten: 4 structurally distinct genres (provocative question, direct action, data impact, cross-domain transfer) — no em-dash monoculture, no 'not X but Y' pattern
• PROMPT_FRAMES de-contaminated (removed 'not X but Y' from 4 frames)
• PATTERN_PROMPT_FRAMES rewritten to 4 distinct rhetorical acts
• BASE_WEIGHTS adjusted: pattern 60%→20%, surprise 30%→50%, extend 10%→30%
• Fixed hardcoded frameIndex=0 (81% of traffic was using same frame)

Quality gate

• Phase 0: removed recommendation ban, fixed fake source attachment, active hours, cross-day logic, C_FN unification, regex bug
• Phase 1: source tracking (citedSourceIndices), prompt rewrite (web findings as primary), verification upgrade (isNewsSummary rejection), search query optimization
• Phase 2: feedback loop wired (messageId capture, reply detection, no-response timer), refine retries 1→2, mode weights, personaChange domains
• SOUL.md voice seed: per-agent loading, preset override, MBTI weight modifiers

Infrastructure

• Fixed all 43 pre-existing tsgo errors in extension tests
• Fixed lint errors (711→0)
• pnpm check fully passes for the first time

Live test results (full pipeline)

• Quality: 9/10 (best), 6.5/10 (avg)
• First person: 50-100%
• [N] leak: 0%
• Em-dash rate: 0-50% (was 100%)
• Pattern mode: uncertainty expression appeared ('不确定是不是直接原因')

v2026.6.12-2

Theme System + Model Selector + Settings UX

🎨 Theme System

• 3 themes × 2 modes (light/dark) = 6 distinct visual combinations
  • 墨玉 Ink Jade: mint accent, cool tones (dark + light mode)
  • 宣纸 Rice Paper: jade accent, warm tones (light + dark mode)
  • 琉璃 Glaze: violet accent, glass-morphic (expanded from 17→86 CSS vars)
• Sidebar themed in all 6 combinations
• resolveColorScheme() now takes (theme, mode) for per-theme light/dark support

🎛️ Model Selector

• Integrated provider→model 2-step flow with inline API key input
• Full model catalog (all providers, not just allowlist-filtered)
• Gateway APIs: auth.storeApiKey, auth.listProviderStatus, models.list with fullCatalog param

⚙️ Settings Page

• Unified scroll layout: connection + appearance + quick settings scroll together
• Removed border radius slider (theme-preset only)
• Card-based sections with consistent styling

🐛 Fixes

• Fixed sidebar not themed in rice-paper and glaze
• Fixed TypeError: c.indexOf is not a function when configForm.agents.defaults.model is an object
• Fixed select dropdown invisible text (white-on-white in light mode)

v2026.6.8 — Pattern Mode Pipeline Fix

Pattern Mode Insight Pipeline Fix

What Changed

• P0: Added InsightCandidate.resolvedMode — safety-net dedup now correctly reads the resolved insight mode instead of falling back to invalid default
• P1: Boosted exploration pNeed from 0.55 to 0.65 — pattern mode candidates more likely to pass PRISM gate
• P2: Fixed domain?:string → domains?:string[] in consolidation types
• P4: Removed stale patternModeRatio config option

Anti-Starvation (v2026.6.8-1)

• Statistical tests proving all 3 modes (surprise, extend, pattern) trigger without starvation
• 30% probability floor ensures no mode drops below 15% selection rate
• 227 scheduler tests passing

Files Changed

• 12 source files, +366 −63 lines
• 2 test files with 6 new statistical tests (+199 lines)

Verification

• Live end-to-end test: pattern insight delivered on attempt 5/20
• Mode distribution: pattern 38%, surprise 34%, extend 28%
• sqlite-vec loads correctly on both Linux and Windows (v0.1.9)

v2026.5.26

Memory Consolidation Engine + Dedup Fixes

Breaking Changes

• Dreaming system replaced with Memory Consolidation Engine — no more narrative dream generation. Instead, daily cron scans session transcripts, extracts structured knowledge via LLM, and routes to existing cognitive stores (PersonaStore, FragmentStore, CorrectionStore).

Features

• Memory Consolidation Engine (extensions/memory-core/src/consolidation*.ts) — daily 3 AM cron, per-user pipeline, LLM extraction, Jaccard dedup, routes to cognitive stores + MEMORY.md inline sections
• Agent isolation — consolidation now resolves userId from session files (not hardcoded agentId), supporting multi-agent workspaces
• MEMORY.md integration — high-confidence items (≥0.7, behavioral_pattern ≥0.8) auto-written to inline sections (👤 User / 🎯 Active Focus)
• Consolidation config — memory.consolidation.enabled, .cron, .concurrency, .batchSize, .lookbackDays

Bug Fixes

• Heartbeat empty spin — skip LLM calls when HEARTBEAT.md is empty; isolated heartbeat sessions use transient mode (auto-cleaned)
• Memory dedup — memory_tidy full now includes inline section dedup (within-section + cross-topic Jaccard ≥0.8); memory_save adds date prefix + dedup; relocateInlineToTopic checks existing similar content

Documentation

• Updated README.md / README.en.md / AGENTS.md reflecting consolidation architecture
• Updated memory-organize skill and KAIJIBOT-GUIDE template

Commits

• 3562fb601e refactor: dreaming → consolidation engine (78 files, -15,147 net lines)
• 66df85f4b4 fix: skip empty heartbeat LLM calls + transient isolated sessions
• 0f89fb3577 feat: consolidation runtime wiring + agent isolation + MEMORY.md integration
• 1bcc35ced3 fix: memory system dedup — inline sections, memory_save, relocateInlineToTopic
• 1bb5a37ee7 docs: update README/AGENTS.md for consolidation engine replacing dreaming
• f49529fad1 release: v2026.5.26

v2026.5.25

• ASCII banner art update

v2026.5.24

KaijiBot v2026.5.24

🔧 Windows Gateway 看门狗

• 新增 Gateway Watchdog 计划任务，每分钟自动检测 gateway 是否存活，崩溃后自动拉起
• 自动生成 watchdog.cmd（PowerShell 进程检测）+ watchdog.vbs（隐藏窗口）
• 所有路径动态解析，无硬编码 C:\Users\... 路径
• kaijibot onboard 安装 gateway 时自动安装 watchdog
• kaijibot gateway uninstall 自动清理 watchdog 任务和脚本

🧠 认知系统修复

• 修复 6 处过时系统提示词（SARA 阶段、PRISM Gate 描述）
• 修复 33 个预先存在测试失败（extractor/preference-learner/store/pipeline）
• 删除死配置 surpriseRatio（3 文件）
• InsightStore agentId 隔离修复

🧹 代码清理

• 删除 src/cognitive/index.ts 死 barrel（0 消费者，-134 行）
• 移除 5 个未使用 npm 依赖（jimp, mpg123-decoder, opusscript, silk-wasm, long，净减 253 包）
• 统一 4 个 cognitive store 的 atomicWrite 到 infra/json-files.ts

🐛 修复

• kaijibot update sidecar 校验修复（39→11 路径，匹配 KaijiBot 实际扩展）

完整变更: 9 files, +179/-737

v2026.5.23

KaijiBot v2026.5.23

OpenClaw Upstream Port — Tier 1 (11 items)

Low-conflict adaptations from upstream:

• Fatal error hooks, disk-space monitor, container-environment detection
• Bootstrap mode, synthetic error monitor, processing-claims tracker
• Usage cache eviction max 256 entries
• WebSocket 1006 close-code hints
• Image regex hoist + lastIndex reset
• Atomic reindex EBUSY/EPERM retry

OpenClaw Upstream Port — Tier 2

Batch 1 — Stability fixes:

• Silent-error retry: Auto-resubmit when provider returns stopReason=error with zero output (up to 3 attempts)
• Proxy CONNECT crash fix: Add missing clientSocket error handler, use destroy() instead of end()
• Slash command word boundary: Prevent /config-check from matching /config handler
• Citation marker stripping: Strip unsupported <cite> control markers from outbound messages

Batch 2 — Transport & provider fixes:

• Anthropic unsafe integer preservation: New json-unsafe-integers.ts parser prevents precision loss for integers >MAX_SAFE_INTEGER in tool call arguments
• Per-model baseUrl override: Honor model-level baseUrl before provider-level (requires baseUrl in model definition config)
• Feishu subagent delivery origin: Return deliveryOrigin from subagent spawning handler

Other fixes

• switch_soul tool-display metadata (emoji + detailKeys)
• gw:deploy now includes ui:build step for Control UI
• Startup performance: defer cognitive init, native module loader fast path
• Cognitive self-refine loop: exempt exploration type from zero-sources gate

Deferred (high integration complexity)

• Compaction loop guard, context pressure preflight, CardKit streaming deltas, parallel tool call fix — assessed but deferred due to invasive core pipeline changes required.

v2026.5.22

What's New

Workspace .md Content Rewriting During Migration

The OpenClaw → KaijiBot migration system now rewrites workspace .md file content instead of copying them byte-for-byte. This ensures migrated agents receive KaijiBot-appropriate instructions in their system prompt context.

New: brand-rewrite.ts — Shared rewriting engine:

• 11 pattern replacements (brand names, config paths, avatar references)
• Section-level structural rewriting for AGENTS.md (preserves user customizations)
• Full replacement for TOOLS.md and BOOTSTRAP.md
• Auto-seeding of KAIJIBOT-GUIDE.md when absent
• MEMORY.md brand-swap after merge

Three rewriting strategies:

File	Strategy
AGENTS.md	Section replacement — known OpenClaw sections → KaijiBot template sections, user sections preserved
TOOLS.md	Full replacement with KaijiBot template
BOOTSTRAP.md	Full replacement with KaijiBot template
KAIJIBOT-GUIDE.md	Auto-created if absent
MEMORY.md	Brand swap only (after merge)
All other .md	Brand name/path replacement

Also: Skill migration now actually replaces brand names in SKILL.md files (previously only detected and added a review banner).

Tests: 42 new unit tests + 3 integration tests (128/128 migrator tests pass).

v2026.5.21

认知层 per-agent 隔离 + 迁移系统修复

Per-Agent 认知存储隔离

所有 5 个认知 store 改为 {agentId}/{userId} 路径隔离，多 agent 场景下每个 agent 独立理解用户：

• persona/{agentId}/{userId}.json — 每个Agent独立维护用户画像
• fragments/{agentId}/{userId}.json — 每个Agent独立积累对话碎片
• corrections/{agentId}/{userId}.json — 每个Agent独立记录纠错
• evolution/{agentId}/{userId}.json — 每个Agent独立追踪进化记录
• evolution/{agentId}/preferences/{userId}.json — 每个Agent独立学习偏好

新增 listAgentIds() / listUserIds(agentId) 方法，scheduler 自动遍历所有 agent 的所有用户。

OpenClaw→KaijiBot 迁移系统修复

• rewriteConfigPaths() — 自动将 ~/.openclaw/ 路径重写为 ~/.kaijibot/（含绝对路径）
• checkMultiAgentBindings() — 多 agent 无 bindings 时发出警告
• validateAgentWorkspaceDirs() — 迁移后验证 workspace 目录
• 归档会话迁移 — 支持 .jsonl.reset.* / .deleted.* / .bak.* 文件
• 修复 runFreshMigration 重复调用 migrateWorkspace
• 测试 83→83 全通过

额外修复

• evolution-suggest-tool 的 SkillPersistenceWriter 改用 resolveAgentWorkspaceDir（与 patch/archive 工具一致）
• Evolution signal 文本加入 [agent: xxx] 标记
• Insight delivery mirror 使用实际 agentId（不再硬编码 "main"）

Onboard Windows 兼容性 (v2026.5.20 hotfix)

• homedir() 路径分隔符兼容
• mkdirSync recursive 选项
• execSync shell 选项
• IDENTITY.md / USER.md 模板缺失修复

v2026.5.20

认知洞察系统大重构

PRISM Gate 死亡螺旋修复

• 替换指数 backoff 0.7^n → 双曲 1/(1+0.3n) + 地板 MIN_BACKOFF_FLOOR=0.12
• 新增补偿性再联系信号（48h 后缓慢增长，上限 0.25），镜像人类"沉默越久越想重新联系"
• MAX_EFFECTIVE_IGNORES=8 封顶惩罚，永不低于 0.12 门槛
• 多轮投递预测: R1(~1h) → R2(~3.4天) → R3(~6.6天) → R4(~12.7天)
• 基于 Mollgaard et al. PLOS ONE 2016 人类电话双曲衰减研究

洞察幻觉拦截（5 道关卡）

• knowledge mode sourceCount ≥ 1 硬门槛 — 无 web 佐证不投递
• pattern/knowledge mode contradicted 阻断 — LLM 验证判矛盾即拦截
• processEventInner 最终防线 — 投递前兜底检查
• 6 个新测试覆盖所有拦截路径

反馈信号链路修复 + 模式选择重构

• processNoResponse 隐式反馈 → bandit 惩罚: domain bandit 和 mode bandit 双通道
• content strategy rotation: 连续无回复时自动切换内容策略（深度→实用→轻松→新颖）
• mode selection 架构: scan 层设 modeCandidates，resolve 层调 selectMode() bandit 加权选择
• resetNoResponseStreak 仅在用户活跃 > 上次投递时触发

Gate 多信号修复

• reEngageSignal coefficient 0.15→0.21, cap 0.7→0.9
• domainActivityFactor → 多信号 engagementFactor (recency + investment + breadth)
• conversation-adaptive timeFactor with no-response death spiral fix

灵魂预设系统

• MBTI 灵魂预设系统 + CLI + agent tool + 首次对话引导
• 所有模板翻译为英文（消除中文硬编码偏好）

洞察多样性改进

• 消除哲学类比偏差，提升 prompt 多样性
• 诊断日志增强: pattern fallback / contradicted / empty web match / parse-failed

其他

• 修复迁移器 MEMORY.md 截断 + 多 agent 会话支持
• 重命名 openclaw Symbol → kaijibot
• 英文 README
• 190/190 测试通过