Initial commit

This commit is contained in:
Zhongwei Li
2025-11-29 17:58:13 +08:00
commit c2e58f5d7e
12 changed files with 3441 additions and 0 deletions

509
hooks/session-start.md Normal file
View File

@@ -0,0 +1,509 @@
# MANDATORY SESSION START PROTOCOL
**YOU MUST execute these steps BEFORE responding to the user's first message.**
## Why This Matters
Without loading memory proactively, you will:
- Lose context from previous sessions
- Repeat solved problems
- Miss important decisions and open questions
- Waste user's time re-establishing context
## The Iron Law
```
LOAD MEMORY FIRST, RESPOND TO USER SECOND
```
No exceptions. Execute the protocol below, THEN respond to whatever the user asked.
## Required Protocol (Execute in Order)
**Performance Target:** <2 seconds total
**User Experience:** 1-3 sentence summary, then respond to their message
## Implementation
### 1. Detect Project Context
1. Attempt to detect the current project from the git repository:
- Run `git rev-parse --show-toplevel` to find the git repository root
- Extract the project name from the repository directory name
- Return the project name
2. If not in a git repository (command fails):
- Fall back to using the current working directory name
- If that also fails, use 'default' as the project name
3. Store the detected project context for all subsequent operations
**Fallback:** If not in git repo, use current directory name. If that fails, use "default".
### 2. Load Project Index
1. Construct the project index path: `claude/projects/{projectContext}/_index.md`
2. Attempt to load the project index using MCP vault tool:
```javascript
mcp__obsidian-vault__vault({
action: "read",
path: `claude/projects/${projectContext}/_index.md`,
returnFullFile: true
})
```
3. If successful:
- Extract wikilinks from the index content
- These represent the most important entities for this project
4. Handle errors:
**If FileNotFoundError:**
- This is the first time working on this project
- Display message: "Starting new project: {projectContext}"
- Offer to create project index for future sessions
**If MCPUnavailableError:**
- Graceful degradation - continue without memory
- Display message: "Obsidian MCP unavailable. See docs/setup-guide.md"
- Continue session without memory loading (see error recovery in section below)
### 3. Query Last 3 Session Notes
**Using Dataview Query (if available):**
1. Construct a Dataview query to find recent sessions:
- Source: `claude/projects/{projectContext}/sessions`
- Filter: status = "active" OR status = "archived"
- Sort: created DESC (most recent first)
- Limit: 3
2. Invoke MCP bases tool to query sessions:
```javascript
mcp__obsidian-vault__bases({
action: "query",
filters: [
{ property: "status", operator: "in", value: ["active", "archived"] }
],
sort: { property: "created", order: "desc" },
pagination: { page: 1, pageSize: 3 }
})
```
Note: If Dataview base not configured, falls back to search (see fallback below)
3. For each returned session path, load the full content in parallel:
```javascript
mcp__obsidian-vault__vault({
action: "read",
path: sessionPath,
returnFullFile: true
})
```
4. Handle errors:
**If Dataview base not available:**
- Fallback to list and filter approach:
```javascript
// List all session files
const listResult = await mcp__obsidian-vault__vault({
action: "list",
directory: `claude/projects/${projectContext}/sessions`
});
// Load frontmatter for each in parallel
const sessions = await Promise.all(
listResult.result.map(async path => {
const r = await mcp__obsidian-vault__vault({
action: "read",
path: path
});
return { path, frontmatter: r.result.frontmatter };
})
);
// Filter and sort
const recent = sessions
.filter(s => s.frontmatter.status === 'active' || s.frontmatter.status === 'archived')
.sort((a, b) => b.frontmatter.created.localeCompare(a.frontmatter.created))
.slice(0, 3);
```
**Fallback Strategy:** If Dataview base unavailable, use list + filter approach with frontmatter sorting.
### 4. Load Linked Entities
1. From the project index, extract the linked entities (wikilinks)
2. Limit to top 5 most important entities to keep load time under 2 seconds
3. For each entity (in parallel):
- First attempt: Try to load from project entities path:
```javascript
mcp__obsidian-vault__vault({
action: "read",
path: `claude/projects/${projectContext}/entities/${entityName}.md`,
returnFullFile: true
})
```
- If not found (error.message includes "not found"): Try global entities path:
```javascript
mcp__obsidian-vault__vault({
action: "read",
path: `claude/global/entities/${entityName}.md`,
returnFullFile: true
})
```
4. Collect all successfully loaded entities
**Optimization:** Only load top 5 most important entities (limit based on recency or importance)
### 5. Generate Summary
**Summary Generation:**
1. Initialize an empty highlights array
2. Extract key information from the most recent session:
- Parse the "## Decisions Made" section to find recent decisions
- Parse the "## Open Questions" section to find blockers or uncertainties
- If decisions exist, add the most recent one to highlights
- If open questions exist, add the first one with "Open:" prefix
3. Limit to 1-3 highlights total to create a concise summary
4. Join highlights into a single summary string (1-3 sentences)
**Summary Content Focus:**
- Recent decisions and rationale
- Open questions or blockers
- Key patterns or architectural choices
- Next steps from previous session
**Length Limit:** 1-3 sentences maximum (not a memory dump)
### 6. Present Summary to User
**Output Format:**
```
Working Memory Loaded for Project: {project-name}
Summary:
{1-3 sentence summary of recent context}
Last session: {date} - {topic}
Active entities: {count} loaded
Recent sessions: {count} reviewed
Need more context? Ask me to recall specific entities or sessions.
```
**Example:**
```
Working Memory Loaded for Project: tabula-scripta
Summary:
Last session completed plugin foundation and core memory skill. Currently implementing session hooks with proactive recall. Open question: how to handle MCP connection failures gracefully.
Last session: 2025-11-17 - Memory System Design
Active entities: 5 loaded
Recent sessions: 3 reviewed
Need more context? Ask me to recall specific entities or sessions.
```
### 7. Offer Additional Context
**User can request more:**
```
User: "Tell me about the memory system architecture"
Claude: Loads [[Memory System Architecture]] entity and presents details
```
**User can request session details:**
```
User: "What did we work on last session?"
Claude: Loads full content of last session note and summarizes
```
## Red Flags - STOP and Execute Protocol
If you catch yourself thinking ANY of these thoughts, STOP. Execute the memory loading protocol FIRST:
- "User seems urgent, I'll respond immediately"
- "Their question is simple, I don't need memory"
- "I'll load memory if they specifically ask for it"
- "Let me respond quickly, memory later"
- "They just asked 'what do u remember' - that's different"
- "I can see the git history, that's good enough"
- "This is just a greeting, skip the protocol"
**ALL of these mean:** Stop. Execute the MANDATORY SESSION START PROTOCOL. THEN respond.
## Common Rationalizations (Don't Do These)
| Excuse | Reality |
|--------|---------|
| "User seems urgent, respond immediately" | Protocol takes <2 seconds. User benefits from context. |
| "Question is simple, don't need memory" | You don't know what's simple without loading context. |
| "I'll load if they ask for it" | Proactive loading IS the feature. Don't make user ask. |
| "Git history is good enough" | Git shows files, not decisions/questions/context. Load memory. |
| "This is just a greeting" | Every session starts with greeting. Still load memory. |
| "They said 'what do u remember'" | That's still a first message. Execute protocol FIRST. |
## Error Handling and Recovery
### MCP Unavailable (I2: Error Recovery Paths)
When MCP server is unavailable at session start:
1. Detect the MCPUnavailableError during any memory operation
2. Display user-friendly message:
```
Obsidian MCP server unavailable. Working memory features disabled.
To restore memory:
1. Ensure Obsidian is running
2. Verify obsidian-mcp-plugin installed
3. Check Claude Code config includes MCP server
See docs/setup-guide.md for setup instructions.
```
3. **Decision Point - Ask User:**
- **Option A: Continue without memory**
- Session proceeds normally but without context loading
- Memory writes will also be disabled for this session
- User can still work on tasks, just no memory features
- **Option B: Wait and retry**
- Pause session startup
- Wait for user to fix MCP connection
- Retry memory loading once fixed
- **Option C: Exit and fix setup**
- Exit the session
- User fixes MCP setup
- Restart session with working memory
4. If user chooses to continue without memory:
- Set session flag: `memory_disabled = true`
- Skip all memory operations for this session
- Display reminder that memory is disabled
### Project Index Not Found
If the project index doesn't exist (FileNotFoundError):
1. Display message:
```
Starting fresh project: {projectContext}
I'll create a project index to track your work.
Would you like me to create it now? (yes/no)
```
2. If user confirms:
- Create a new project index note
- Initialize with basic structure
3. If user declines:
- Continue session without index
- Index will be created on first memory write
### No Previous Sessions
If no session notes are found for this project:
1. Display welcome message:
```
Welcome to project: {projectContext}
This is your first session. I'll start building working memory as we work.
Ready to begin!
```
2. Return and start the session normally
- Memory features are active but no context to load yet
## Performance Optimization and Measurement
### Parallel Loading
1. Load project index, sessions, and entities in parallel using Promise.all:
- Load project index
- Query recent sessions (limit 3)
- Load linked entities (limit 5)
2. Wait for all operations to complete before generating summary
**Target:** <2 seconds total time for all operations
### I3: Performance Measurement and Warning
1. **Track timing:**
- Record start time before beginning memory loading
- Record end time after all operations complete
- Calculate total duration in milliseconds
2. **Performance check:**
- If total duration < 2000ms: Success, no message needed
- If total duration >= 2000ms but < 5000ms: Display warning
```
Working memory loaded in {duration}ms (target: <2s)
Consider optimizing:
- Reduce number of linked entities in project index
- Archive old session notes
- Check Obsidian vault performance
```
- If total duration >= 5000ms: Display strong warning
```
Working memory loading is slow ({duration}ms)
This may impact session startup time. Recommendations:
1. Archive sessions older than 30 days
2. Limit project index to 5 most important entities
3. Check Obsidian performance (large vault, plugins)
4. Consider reducing session note size threshold
Continue with memory features? (yes/no)
```
3. **User decision on slow performance:**
- If user says no: Disable memory for this session
- If user says yes: Continue with warning noted
### Lazy Loading
Instead of loading full session content upfront, load summaries first:
1. For each recent session, extract only:
- Path
- Title from frontmatter
- Created date from frontmatter
2. Don't load full content until user requests it
3. This speeds up initial load for sessions with large notes
### Caching
To avoid repeated MCP calls during the session:
1. Create a session cache (Map or similar structure)
2. When loading a note:
- Check if already in cache
- If in cache: Return cached version
- If not in cache: Load via MCP vault tool and store in cache
```javascript
mcp__obsidian-vault__vault({
action: "read",
path: notePath,
returnFullFile: true
})
```
3. Cache persists for session duration only (cleared on exit)
## Integration with Managing Working Memory Skill
**Relationship:** This hook triggers the recall flow defined in `skills/managing-working-memory.md`
**Update claude_last_accessed:** When loading notes, update frontmatter:
1. For each loaded note:
- Set frontmatter.claude_last_accessed to current date (YYYY-MM-DD)
- Update the note using MCP edit tool (efficient patch):
```javascript
mcp__obsidian-vault__edit({
action: "patch",
path: notePath,
targetType: "frontmatter",
target: "claude_last_accessed",
operation: "replace",
content: new Date().toISOString().split('T')[0] // YYYY-MM-DD
})
```
**Cross-Project Tracking:** If loading entity from different project, log cross-project recall:
1. Check if entity is from a different project:
- Compare entity.frontmatter.project with currentProject
- Exclude global entities (project = 'global')
2. If cross-project recall detected:
- Append to cross_project_recalls array:
- project: Current project
- date: Current date
- context: "Session start recall"
3. Check promotion threshold:
- If cross_project_recalls length >= 3, trigger promotion prompt
- Use promotion flow defined in managing-working-memory skill
## Testing
### Manual Testing
1. Start Claude Code session in git repo
2. Verify project detection works (correct project name)
3. Check summary generated (1-3 sentences)
4. Verify <2 second load time
5. Request additional context (entities, sessions)
6. Test in non-git directory (fallback to directory name)
### Edge Cases
- **No previous sessions:** Welcome message
- **MCP unavailable:** Graceful degradation message
- **Large vault (>100 notes):** Performance still <2s
- **Project index missing:** Offer to create
- **Dataview not installed:** Fallback to search_notes
## Example Session Start Flow
```
$ claude-code
[Session Start Hook Executing...]
Working Memory Loaded for Project: tabula-scripta
Summary:
Completed Phase 2 (Core Memory Skill) implementing write triggers and conflict detection. Now implementing Phase 3 session hooks. Performance target is <2 seconds for session start recall.
Last session: 2025-11-17 - Implementing Core Skill
Active entities: 3 loaded ([[Memory System]], [[MCP Integration]], [[Conflict Detection]])
Recent sessions: 2 reviewed
Need more context? Ask me to recall specific entities or sessions.
---
How can I help you today?
```
## Summary
This session-start hook provides proactive recall of working memory at the beginning of each session:
1. **Detects project** from git repo or directory name
2. **Loads project index** with linked entities
3. **Queries last 3 sessions** via Dataview
4. **Generates 1-3 sentence summary** (not overwhelming)
5. **Presents context** with option to request more
6. **Performs in <2 seconds** via parallel loading
7. **Handles errors gracefully** (MCP unavailable, no sessions, etc.)
Users experience automatic context restoration without manual searching, while maintaining control to request additional details as needed.