pub(crate) const SEARCH_HISTORY_TOOL: &str = r#"{
"name": "SearchHistory",
"description": "Search YOUR OWN past conversations — including sessions that have ended or been compacted away from your current context. Use it before asserting anything about what you previously did, said, or sent: your context window is a lossy view of your own history, so after a reset or compaction you can confidently misreport your own actions. Typical questions it answers: 'did I already message X', 'have I made this change before', 'what did I conclude about Y last week'. Matches message text, and — via the `tool` filter — actual tool calls, so 'did I call SendMessage to Agent4' is answered structurally rather than by hoping the prose mentions it. Scans your most recently modified sessions first; `since`/`until` narrow the window cheaply before any file is read. Searches every session in AgentMux's own record of you (all accounts, channels and versions), including sessions whose provider transcript is gone. Returns {hits: [{session_id, timestamp, role, snippet, tool_name, source, attribution}], complete, incomplete_reasons, total_sessions, sessions_scanned, sessions_skipped, sessions_unreadable, sessions_partly_read, discovery_errors, ledger_sessions, record_only_sessions, inferred_sessions_excluded, truncated}. IMPORTANT: only `complete: true` with no hits means it did not happen (within your window). `complete: false` means unknown, never 'it didn't happen' — `incomplete_reasons` says why (hit_limit, max_sessions, unreadable_sessions, unreadable_records, discovery_errors). Read-only. Searches only your own history; another agent's conversation is governed by the transcript_request visibility protocol instead.",
"inputSchema": {
"type": "object",
"properties": {
"query": { "type": "string", "description": "Case-insensitive substring to find in message text and tool arguments. May be empty ONLY when `tool` is set (meaning: every call to that tool)." },
"tool": { "type": "string", "description": "Only match calls to this tool, e.g. \"SendMessage\" — matches the built-in tool and MCP tools recorded with a server prefix (mcp__agentmux__SendMessage) alike. Structural — matches the recorded tool name, not prose mentioning it." },
"role": { "type": "string", "enum": ["user", "assistant"], "description": "Restrict to your own messages (assistant) or what you were told (user)." },
"since": { "type": "integer", "description": "Unix time in seconds (milliseconds, e.g. a hit's `timestamp`, also accepted). Only messages at or after it; sessions last written earlier are skipped WITHOUT opening them." },
"until": { "type": "integer", "description": "Unix time upper bound, seconds or milliseconds. Only messages at or before it; sessions started later are skipped without opening them." },
"max_sessions": { "type": "integer", "description": "Max sessions to open, newest first (default 20, cap 100). Sessions beyond it are counted in sessions_skipped and make the answer incomplete." },
"limit": { "type": "integer", "description": "Max hits before reporting truncated (default 50, cap 200)." },
"include_inferred": { "type": "boolean", "description": "Also search sessions found only by your working directory or account that your own record doesn't name (inferred_sessions_excluded counts them). Default false: they may be another agent's." }
},
"required": ["query"]
}
}"#;