…tion branching (Issue #79)

This commit implements three interconnected specifications:
- P1-A006: LLM Stats Collection + Session Summary for Normal Mode
- ai-block-fields-extension: Token usage, cost estimate, model version fields
- tree-conversation-branching: Fork/switch/delete conversation branches

Backend changes:
- Add LLMCallStats struct with token usage and timing metrics
- Refactor LLMService.Chat/ChatStream to return stats
- Implement BaseParrot with stats accumulation for all Parrot agents
- Add TokenUsage, CostEstimate fields to Block proto/message
- Add parent_block_id, branch_path for tree branching support
- Add ForkBlock, ListBlockBranches, SwitchBranch, DeleteBranch RPCs
- Implement postgres migration for new Block fields
- Add comprehensive tests for stats, ForkBlock, and cost calculation

Frontend changes:
- Add TokenUsageBadge and BlockCostBadge components for displaying stats
- Add BranchIndicator and BranchSelector for conversation branching
- Implement useBranchTree hook for branch operations
- Update SessionSummaryPanel to display token usage and timing
- Add frontend integration tests with Vitest
- Update i18n keys for new features

Documentation:
- Add AI_CHAT_INTERFACE.md with complete UI architecture diagrams
- Update ARCHITECTURE.md with Unified Block Model references

All 51 Loki Mode tasks completed. make check-all passes.

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Features:
- Add SessionBar component displaying aggregated session stats (cost, tokens, duration)
- Add BlockEditDialog for editing user inputs and creating branches
- Add type guards isStreamingStatus(), isErrorStatus() for BlockStatus checks
- Add isCompletedStatus() helper function

Code quality fixes (P0/P1/P2/P3):
- Fix JSX syntax errors: rename test files from .test.ts to .test.tsx
- Fix BlockStatus comparison bug (String() was always false)
- Reduce 'any' type usage in useBlockQueries.ts
- Extract magic number MILLI_CENTS_TO_USD constant
- Remove unused imports and variables

i18n:
- Add translations for session bar, edit dialog, and status labels

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

… audit fixes (Issue #79)

This commit completes the P0-P3 phase tasks and addresses all code audit findings:

**P2 Tasks:**
- Add BlockSummary restoration from persisted Block.sessionStats after page refresh
- Implement friendly branch path numbering (0/1/2 → A.2.3) with Base26 conversion
- Integrate SessionBar into ChatHeader for PC, mobile-only standalone display

**P3 Tasks:**
- Update AI_CHAT_INTERFACE.md to v0.93.1 with new UI patterns

**Code Audit Fixes:**
- P1-#1: Add safeBigIntToNumber with MAX_SAFE_INTEGER boundary check
- P2-#1: Add MAX_ITERATIONS limit to toBase26 to prevent potential infinite loops
- P2-#2: Optimize calculateSessionStats to process only last 100 blocks
- P3-#1: Replace hardcoded Chinese text with i18n t() calls
- P3-#2: Add detailed comment for MILLI_CENTS_TO_USD constant

**New Components:**
- HeaderSessionStats: Mode-differentiated session stats (Normal/Geek/Evolution)

**Bug Fixes:**
- Fix ForkBlock protobuf serialization using create(UserInputSchema, ...)
- Fix ToolCalls duplication with occurrence-based deduplication
- Fix import ordering per Biome lint rules

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Remove unused BlockListCacheData interface (re-added for internal use)
- Fix protobuf mock creation using proper types
- Change \`as any\` to \`as unknown as Type\` for observer mocks
- Use EmptySchema for delete/append operations
- Use ListBlocksResponseSchema for list operations
- Import EmptySchema from @bufbuild/protobuf/wkt

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Update @testing-library/jest-dom from 6.6.0 to 6.9.1
- Update @testing-library/react from 16.2.0 to 16.3.2

Rebase caused duplicate TrackToolCall call. Remove one instance.

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Fix cold-start latency issue in ScheduleParrot by sending thinking event
  before first LLM call (reduces perceived latency from 4s to immediate feedback)
- Add episodic_memory table to LATEST.sql for new installations
- Unify all Parrot agents to send thinking event before LLM call

Related to #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Add 6 missing tables to LATEST.sql for new installations:
- user_preferences: AI personalization preferences
- conversation_context: AI session persistence and recovery
- agent_metrics: A/B testing metrics
- tool_metrics: Tool call performance metrics
- chat_app_credential: Multi-platform chat app integrations
- ai_block: Unified Block Model (core table!)

This ensures new installations have all required tables from the start,
preventing "relation does not exist" warnings.

Related to #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Add prevention mechanism for future migration sync issues:
- store/migration/postgres/CLAUDE.md: Complete migration development guide
- scripts/check-migration-sync.sh: Automated sync verification script
- Link to migration guide in root CLAUDE.md

Key principles:
1. Any DB change must update BOTH migrate/*.up.sql AND LATEST.sql
2. Run check-migration-sync.sh before committing migrations

Related to #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Remove 5 outdated migration files from root of postgres dir:
- 20260203_chat_apps_credential.* (superseded by 20260203070000_*)
- 20260203_conversation_channel_type.* (superseded by 20260203130000_*)
- V0.53.4__add_chat_app_app_secret.sql (superseded by 20260203080000_*)

These files:
- Use old naming format (YYYYMMDD_ and V*.*__)
- Are not read by the migration system (migrator.go only reads migrate/)
- Have newer counterparts in migrate/ directory

Directory structure now cleaner:
- CLAUDE.md (migration guide)
- migrate/ (incremental migrations)
- schema/ (LATEST.sql)

Related to #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Update freshness timestamp to 2026-02-07 (v0.93.1)
- Document four-layer routing system (Cache → Rule → History → LLM)
- Add AI model strategy overview table
- Update intent classification to use SiliconFlow + Qwen2.5-7B-Instruct
- Document fast path optimization (time + query keywords)
- Remove deprecated ChatRouterConfig from API documentation
- Add structured JSON Schema output for LLM classifier

Ref:
- ai/router/service.go: four-layer routing implementation
- server/router/api/v1/ai_service.go: SiliconFlow + Qwen integration
- ai/router/rule_matcher.go: fast path optimization

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Add asynchronous LLM connection warmup to reduce first-request latency.
The warmup sends a minimal ping request (max_tokens=1, temperature=0)
to pre-establish TCP/TLS connections before the first real user request.

Changes:
- Add Warmup(ctx) method to llmService with direct API call
  - Uses max_tokens=1 (minimum cost, ~85% cheaper than default)
  - Uses temperature=0 (fastest deterministic response)
  - 5s timeout, failures don't affect service startup
- Call Warmup asynchronously in NewAPIV1Service after LLM init
- Add provider() helper for logging

Expected impact:
- First request latency: ~5400ms → ~4900ms (save ~500ms)
- Warmup cost: ~¥0.0000006 per startup (vs ¥0.000005 before)

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Add guidance for using TODO LIST to track multiple tasks, preventing
"memory loss" or losing direction during multi-task sessions.

When to use TODO LIST:
- Multiple optimization points found during analysis
- User requests to tackle issues one by one
- Tasks take >1 hour or have multiple steps

Workflow:
1. Create tasks with TaskCreate after analysis
2. Track status with TaskList
3. Update to in_progress before starting
4. Mark completed when done

Example practice from this session:
- Analyzed logs → found 5 issues
- Created 4 TODOs for remaining work
- Completed #1 (connection warmup), tracked others

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Use LEFT JOIN + COUNT GROUP BY to fetch block counts in a single query,
eliminating the N+1 query problem where each conversation triggered
a separate ListAIBlocks call.

Changes:
- store.AIConversation: add BlockCount field
- postgres.ListAIConversations: use LEFT JOIN ai_block with COUNT
- Remove per-conversation ListAIBlocks calls from handler

Before: N conversations → N+1 database queries
After:  N conversations → 1 database query

Performance impact:
- 10 conversations: 11 queries → 1 query (91% reduction)
- 100 conversations: 101 queries → 1 query (99% reduction)

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Tool-based agents (MemoParrot/ScheduleParrot/AmazingParrot) don't have
direct LLM calls, so token stats are always zero. Instead of logging
misleading zeros, log meaningful tool metrics:

Changes:
- When PromptTokens=0 and CompletionTokens=0:
  - Log "tool-based agent stats" with tool_calls count
  - Log "tool-based agent completed" with tools used
- Preserve original logging for agents with real token stats

Before:
  Agent: got normal stats... prompt_tokens=0 completion_tokens=0
  Agent: applied normal stats... prompt_tokens=0 completion_tokens=0

After:
  Agent: tool-based agent stats tool_calls=1 duration_ms=5423
  Agent: tool-based agent completed tool_calls=1 tools=schedule_query

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Align comment formatting in Warmup function for consistency.

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Use ListMessages API instead of ListBlocks for pagination support
- Reduces duplicate DB queries when switching conversations
- React Query now caches response for both hooks

Before: useBlocksWithFallback → ListBlocks + syncMessages → ListMessages
After:  Both hooks use ListMessages → cached response

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Major improvements:
- Add "SOTA Agent 工程实践" section with:
  - Thinking Protocol for complex decisions
  - Tool usage strategy matrix
  - Metacognition: self-healing protocol when stuck
  - SOTA reasoning patterns (CoT, ReAct, Self-Refinement)
- Restructure for progressive disclosure
- Add freshness header (v0.93.1)
- Simplify tables for better scannability
- Emoji icons for visual hierarchy

Key additions:
- When to use Task tools vs direct operations
- Explicit thinking criteria (architecture changes, unfamiliar APIs)
- Self-healing flow: re-read → visualize → clarify → confidence → document

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

BlockManager already logs "Created block for chat" with round_number.
Removed redundant log in handler.go to avoid duplication.

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

The lockfile had duplicated mapping keys causing CI to fail.
Regenerated with pnpm install.

Refs #87

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Add duration_ms to track request processing time
- Shorten service name (remove memos.api.v1. prefix)
- Use underscore case for status (ok, client_error, server_error)
- Use empty message for cleaner structured log output

Before: INFO OK method=/memos.api.v1.AuthService/GetCurrentUser
After:  INFO type=API method=AuthService/GetCurrentUser status=ok duration_ms=12

Refs #79

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>