session-start

Two-phase boot: show what exists (Phase 1), then load deep context for the chosen focus (Phase 2).

Step 1: Read Cache

Read ~/.claude/cache/session-context.json (written by session-context-preloader hook).

Extract:

• detectedProject -- project matched from cwd (may be null)
• vaultProjectPath -- vault path for detected project
• recentSessions -- previews of last 3 session notes
• projectIndex -- all scanned vault directories with name, vault, status, codebase, summary, lastSession

If cache is missing, unreadable, or lacks projectIndex, fall back:

• Glob <vaultPath>/Projects/*/README.md (vault path from agent-os.config.json)
• Present folder names as a plain list (no metadata)

Step 2: Present (Phase 1)

If a project was detected from cwd:

Skip straight to Phase 2 (Step 3) for the detected project. After showing the deep briefing, append a condensed "Other Projects" list:

• Active projects as a simple bulleted list (name + last session date), not a full table
• Omit dormant items

If no project detected (home dir or unmatched):

Display the full index. Sort all entries by lastSession date descending (most recently worked on first). Items with no session date go to the bottom.

Format:

For each key in projectIndex (e.g., "projects", "ventures"), show a table:

Session Context Loaded

## Projects
| Project | Last Worked | Summary |
|---------|------------|---------|
| {name} | {M/D or --} | {summary} |
...

## Ventures (if present in index)
| Venture | Last Worked | Summary |
|---------|------------|---------|
| {name} | {M/D or --} | {summary} |
...

## Dormant
{comma-separated list of items with status complete/deprecated/planned}

Recent Sessions:
- {M/D} -- {title extracted from session filename} ({project tag})
- ...

What are we working on?

Rules for the index:

• Convert lastSession dates to short format: "3/13" not "2026-03-13"
• Summaries should be under 10 words (truncate if needed from the cache data)
• Active = status is "active" or "unknown" with a recent session
• Dormant = status is "complete", "deprecated", "planned", or "unknown" with no session in last 60 days
• Put dormant items on a single comma-separated line with their status in parens, not a table row
• Target: 25-40 lines total for the index output

Step 3: Deep Context Load (Phase 2)

When the user names a project to work on (or when auto-detected from cwd):

1. Find the vault path from the matching index entry's vault field
2. Read the README: <vaultPath>/{vault}/README.md
   • Extract: current status, what's built, blockers, next steps
   • Summarize in 4-6 bullets max -- do NOT dump the raw README
3. Find related sessions: Glob <vaultPath>/Sessions/* for files containing the project name. Also check the lastSessionFile from the index. Read first 20 lines of the 2 most recent matches.
4. Check for extras in the vault folder:
   • CODEBASE.md -- mention "CODEBASE.md available" if it exists (don't read unless asked)
   • STATUS.md -- read if exists, include key points
5. Check error-patterns: Read <memoryDir>/error-patterns.md and surface any entries that mention the project name or its key technologies
6. Present deep briefing:

Deep Context: {Project Name}

Status: {status}
Codebase: {codebase path}
Last Session: {date} -- {topic from session filename}

Current State:
- {bullet from README}
- {bullet from README}
- {bullet from README}

{If blockers exist:}
Blockers: {list}

{If CODEBASE.md exists:}
CODEBASE.md available for architecture reference.

{If relevant error patterns:}
Watch out for: {brief warning}

What's the focus?

Step 4: Cross-reference Check

After loading deep context, quickly compare:

• Does MEMORY.md (already in system prompt) have info about this project that contradicts the README?
• If drift detected, flag it: "Note: MEMORY.md says X but README says Y -- which is current?"

Rules

• Keep all output concise -- this is a briefing, not a report
• Don't read files that don't exist -- check gracefully with Glob first
• Phase 1 (index) = exactly 1 file read (the cache JSON)
• Phase 2 (deep load) = 2-4 file reads max (README + sessions + error-patterns)
• Never dump raw file contents -- always summarize
• If user asks to work on something not in the index, search the vault with Grep before saying it doesn't exist
• All paths come from agent-os.config.json -- never hardcode user-specific paths