| name | start |
| description | Lightweight session reorientation — queries session-rag + mem0 for recent session context, checks git state, and surfaces 5 most recent open issues. Fast alternative to /prime for continuing same-day work or resuming after a short break. No indexing, no code search, no agent dispatch. Triggers: "start", "quick start", "resume", "orient me", "where were we".
|
Start
Lightweight session reorientation. Use this when you're continuing recent work and just
need to know where you left off — not when you need a full environment boot.
Under 15 seconds. No code indexing, no agent dispatch. Two fast MCP calls
(session-rag + mem0) replace the old digest file read.
When to use /start vs /prime
| Situation | Use |
|---|
| Continuing same-day work in a new terminal | /start |
| Coming back after a short break (< 1 day) | /start |
| Starting fresh after days away | /prime |
| First session on a new project | /prime |
| Need full code indexing (Milvus/CGC) | /prime |
| Need comprehensive mem0 search | /prime |
| Just need context from last session | /start |
The Flow
Step 0 — Verify project identity (MUST run first, blocking)
cat .claude/project.json 2>/dev/null
pwd
If .claude/project.json is missing OR cwd differs from the expected project root (e.g.,
launched from a subdir whose parent is a different git repo — common when a project like
Devops/ sits inside an outer workspace repo), STOP the flow and warn prominently:
⚠ project.json missing or cwd mismatch
cwd: <pwd output>
project.json: <found path or "none">
Likely cause: .claude/ is gitignored; file dropped during a merge/reset/worktree switch.
Fix: create .claude/project.json with name/tag/issuePrefix, or cd to correct project root.
Offer to create .claude/project.json from a nearby sibling's template (read
../<sibling>/.claude/project.json, adjust name/tag/issuePrefix, write). Do not
continue to Steps 1-3 until resolved — otherwise digest paths and project scoping break
silently (the digest lands under Root/ instead of the real project folder).
Step 1 — Retrieve recent session context (session-rag + mem0)
Query session-rag and mem0 in parallel for recent session context. These are the only
MCP calls /start makes — both are fast (<2s each).
1a — session-rag (recent turns, project-scoped):
IMPORTANT: Do NOT pass filesystem paths as project_root — silently returns empty.
Omit project_root (defaults to current project via HTTP header) or use "*" for
cross-project search. Include the project name in the query for relevance.
mcp__session-rag__search_all_sessions(
query="last session summary recent work decisions next steps <project-name>",
n=5
)
Extract from results: what was worked on, decisions made, next steps noted, handoff items.
1b — mem0 (curated episodic memory):
mcp__mem0__search_memories(
query="recent work decisions handoffs",
limit=10
)
Post-filter results by metadata.project matching the project tag (case-insensitive).
Merge with session-rag findings — mem0 adds retro learnings and cross-session decisions
that verbatim turns don't capture.
Degradation: If session-rag is unreachable (port 7102 down), use mem0 results only
and append a warning: ⚠ session-rag unavailable — context from mem0 only.
If both are unavailable, note "No recent session context available" and continue.
Step 2 — Git state snapshot
git status --short
git branch --show-current
git log --oneline -10
git worktree list
Surface:
- Current branch
- Last 10 commits (quick orientation on recent work)
- Uncommitted changes (if any)
- Active worktrees (if any beyond main)
Step 3 — Recent open issues
Find the 5 most recently updated open issues from the vault:
ISSUES_DIR="${DOCVAULT_PATH}/Issues"
grep -rl 'status:.*\(open\|backlog\|todo\|in-progress\|in-review\)' "$ISSUES_DIR" 2>/dev/null \
| grep -v '_Index' | xargs ls -t 2>/dev/null | head -5
For each file, read the frontmatter to extract: id, title, status, updated.
Display as a compact list.
If there is no vault Issues directory or no open issues, note it briefly and move on.
Also check the project-specific Issues path if the global one is empty:
${DOCVAULT_PATH}/Projects/{ProjectName}/Issues/
Output Format
Print a compact orientation report:
## Session Start — <ProjectName> (<date>)
### Last Session
<3-5 sentence recap synthesized from session-rag turns + mem0 memories — what was worked on, decisions made, next steps>
**Handoff:** <Any handoff/next-session items surfaced from session-rag or mem0 — omit if none>
---
### Git State
- **Branch:** <current branch>
- **Recent commits:** <last 3-5, one-liners>
- **Uncommitted:** <count or "clean">
- **Worktrees:** <list or "none">
---
### Open Issues (5 most recent)
- [<ID>] <title> — <status>
- [<ID>] <title> — <status>
- ...
---
_Use /prime for full environment boot with code indexing and comprehensive mem0 search._
Keep the whole output under 50 lines. If session context or issues are missing, compress
those sections to a single "none found" line rather than expanding on the absence.
REQUIRED: The final line of every /start report MUST be exactly:
_Use /prime for full environment boot with code indexing and comprehensive mem0 search._
Do not replace it with a custom summary or action item. It must always appear.
Rules
- session-rag + mem0 MCP calls only. No indexing MCPs (Milvus, CGC, Codacy). No agent dispatch.
- No code indexing. Do not trigger claude-context, CGC, or any indexing operation.
- Read-only. Do not write, commit, or modify anything.
- If /start would take > 20 seconds, something is wrong. The two MCP calls + git
commands + issue scan should complete in under 10 seconds total.
What this skill does NOT do
- Full environment indexing (use /prime)
- Code health analysis or security scans (use /prime)
- DocVault page reading (use /prime)
- Agent dispatch of any kind (use /prime)
- Retro, wrap, or any write operations