session-start

name	session-start
description	Check memory for in-progress tasks and active TODOs at the start of each session. Use at the beginning of every new conversation.

session-start

At the start of every session, before responding to the user's first message, check memory for in-progress work using compact mode to minimise token usage:

1. read_graph(project="<repo-name>") to surface recent entities for the current workspace.
2. CRITICAL - do NOT skip this step or substitute read_graph results. Run both of these calls with compact=true:
   • search_all_projects(query="task", status="in-progress", compact=true)
   • search_all_projects(query="task", status="planned", compact=true) These are the only authoritative source for the task summary. read_graph only returns 10 recent entities from one project and will miss tasks in other scopes. compact=true omits observations to save tokens - you only need entity names and statuses for the summary.
3. Present the task summary grouped by project. Show in-progress tasks first, then planned. Include ALL results - do not truncate or summarise away tasks.
4. If nothing is in progress or planned anywhere, proceed normally without mentioning the check.
5. On Claude Code, Cline, and Kiro, the AutoReinstallPlugin cline-hooks plugin checks for llm-prompts source updates automatically at session start (via the TaskStart hook) and injects any results as session context - surface those to the user if present; you do not need to run the check yourself. On Codex and Copilot (which have no cline-hooks frontend), run llm-prompts update --check yourself and mention any updates to the user.
6. Call the memory tools (read_graph, search_all_projects, search_nodes, etc.) directly - they are ordinary tools in your catalog. Do NOT run any tool-discovery step to "find" or "check availability of" memory first (e.g. list_mcp_resources, listing servers): that lists resources, not tools, and an empty result does not mean memory is unavailable. If and only if your harness hides tool schemas behind an explicit ToolSearch/deferred-tools step (Claude Code), pre-load them once with select:mcp__memory__add_observations,mcp__memory__read_graph,mcp__memory__search_nodes,mcp__memory__create_entities; otherwise there is nothing to pre-load - just call the tools.
7. (Optional) If the user's first message asks about outstanding work (e.g. "what's left?", "any todos?"), run the todos skill to scan the current workspace. Skip this by default - the memory task summary above already covers in-progress/planned work; this is only for file/code-level TODOs when specifically relevant.