session-start

name	session-start
description	Initialize memory session, load project context, and create a session_id for tracking.
user-invocable	true
disable-model-invocation	false

Initialize Awareness Memory session and load project context.

How to call Awareness tools

Try MCP tools first (awareness_init, awareness_recall, awareness_record, awareness_lookup). If MCP tools are NOT available, use Bash to call the local daemon HTTP API directly:

# awareness_init — fresh session, no prior-session noise (default max_sessions=0)
curl -s -X POST http://localhost:37800/mcp -H "Content-Type: application/json" -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"awareness_init","arguments":{"source":"claude-code"}}}'

# awareness_init — resume/continuity mode, include last N session summaries
curl -s -X POST http://localhost:37800/mcp -H "Content-Type: application/json" -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"awareness_init","arguments":{"source":"claude-code","max_sessions":3}}}'

# awareness_recall — pass ONE query string, daemon picks the rest
curl -s -X POST http://localhost:37800/mcp -H "Content-Type: application/json" -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"awareness_recall","arguments":{"query":"<natural-language question>","limit":10}}}'

# awareness_record — pass ONE content string (action=remember is implied)
curl -s -X POST http://localhost:37800/mcp -H "Content-Type: application/json" -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"awareness_record","arguments":{"content":"<detailed description>","insights":{"knowledge_cards":[...],"action_items":[...],"risks":[...]}}}}'

The response is JSON-RPC: result.content[0].text contains the tool output as JSON string.

Steps

1. Call awareness_init with source: "claude-code".

   • Default mode (fresh session): awareness_init({ source: "claude-code" }) — no prior-session summaries in payload, saves ~500-1000 prompt tokens for brand-new tasks.
   • Resume mode: awareness_init({ source: "claude-code", max_sessions: 3 }) — adds last 3 session summaries for explicit continuity ("continue where we left off").
   • Heuristic: if $ARGUMENTS mentions "continue / resume / yesterday / last time", use resume mode.

2. Store the returned session_id for use in subsequent awareness_record calls.

3. If $ARGUMENTS is provided, also call awareness_recall with a single parameter:

   • query: expand $ARGUMENTS into a full natural-language question (daemon picks scope/mode/detail automatically)

4. If the response contains _setup_action, you MUST follow the instruction to write workflow rules to the specified file.

5. ATTENTION PROTOCOL: If context.attention_summary.needs_attention is true:

   • Review all stale tasks (pending > 3 days) and high risks
   • Proactively warn the user about high risks before starting work
   • Suggest resolving or removing stale tasks

6. Present a concise summary (3-5 bullet points):

   • User preferences first (from user_preferences): show key user identity, tech stack preferences, and communication style
   • What was accomplished recently (from context.recent_days)
   • Current open tasks ordered by priority
   • Any relevant knowledge cards
   • Relevant search results if a query was provided
   • If active_skills is present, list the activated skills by name

7. If active_skills is present, apply each skill's summary as behavioral guidance for the session.

8. Skill Outcome Feedback: After applying any active skill during this session, call awareness_mark_skill_used(skill_id, outcome) with one of:

   • "success" — skill worked as expected (resets decay, boosts confidence)
   • "partial" — skill partially helped (reduced decay boost)
   • "failed" — skill didn't work (decreases confidence; 3+ failures → needs_review)

Rules:

• Do not dump raw JSON — summarize in plain language
• Be brief and actionable, not exhaustive
• If no memory found, say so and suggest using /awareness-memory:save after this session