Goal

Recover useful past context instead of answering from guesswork. Use this skill when the user refers to earlier work, another chat, another context, a subagent run, or a remembered fact that is not visible in the current chat.

Use when

• The user says things like "we discussed this before", "what was the last status", "did we already decide this", or "in the other project...".
• You need message content from another chat, from your own current chat, or from any of the user's chats.
• You need to see which contexts and chats exist, or what a specific context contains.
• You need the outcome or trace of an earlier subagent run.
• You need to confirm whether a stable fact, preference, or decision is already stored in long-term memory.

Tool Map

• search_message_chains: message-level text search and chain reading. The workhorse for "what was said where".
• get_contexts_overview: browse all contexts (names + keys). Entry point when you do not know the context.
• get_context_details: one context in depth - metadata, linked chats (names + keys), linked skills, tools, files, and LTM spaces.
• get_subagent_execution_trace_by_initiating_message: full trace of a subagent run from the tool-call message key that started it.
• read_agent_audit_log: action/audit history with filters (action, tool, status, since/until, chat, context) - "what was done where and when".
• search_background_processes: locate background processes and runs by name, type, or status.
• search_longterm_memory: durable facts, preferences, and decisions. When the explore-longterm-memory skill is loaded, follow its strategy for detailed LTM work.

Workflow

1. Check the current chat context first; answer from it if it already contains the answer.
2. Known chat or context key: run search_message_chains scoped with chat_key or context_key.
3. Unknown context: get_contexts_overview to list contexts, then get_context_details for the candidate, then scope search_message_chains with one of its chat keys.
4. Subagent runs: find the tool-call message that started the run via search_message_chains, then call get_subagent_execution_trace_by_initiating_message with that message key.
5. Durable facts, preferences, and settled decisions: search_longterm_memory.
6. Action-level recall (which tools were run, where, when): read_agent_audit_log. Live or finished background runs: search_background_processes.
7. If sources disagree, say so and surface the conflict instead of silently choosing one.

Search Strategy

• search_message_chains is an AND search: every term must occur in the same message. Start narrow with names, project terms, decision words, or exact phrases; broaden only when a search returns nothing.
• Scope with chat_key or context_key as soon as you know them to cut noise.
• Continue paginated results with the returned cursor; keep limit between 10 and 20.
• Narrow read_agent_audit_log results with its filters (action, tool, status, since/until) instead of scanning raw entries.
• For long-term memory, use focused natural-language queries; add reference_time_iso only when the user anchors the request to a specific past time.

Output Rules

• Report where the answer came from: chat (name/key), context, subagent trace, audit log, or long-term memory.
• Quote or summarize only the evidence needed to answer the user.
• If nothing reliable is found, say that clearly and offer the closest partial context.