Read Agent Sessions

Turn a loose session reference — an id prefix, a title someone remembers, or a fragment of a prompt — into a verified handoff another agent can act on. All four harnesses keep complete transcripts on the machine where the session ran, so the work is resolution and faithful extraction, never reconstruction from memory. The skill applies regardless of which agent runs it; loading it grants no write access to any session store.

Quick Start

1. Identify the deliverable: locate a session, extract a conversation, or produce a takeover summary. Preserve the requester's harness if named, and their intended consumer — another agent, the user, or both.
2. Identify the harness before touching storage. The identifier often decides it: a sess_ prefix means Zcode, YYYYMMDD_HHMMSS_<hex> means Hermes, and a bare UUID is Codex or Claude Code and needs the requester's intent or both stores searched.
3. Use the resolution map below to choose the lookup route. Treat a match as a hypothesis until the file or row is opened and its first records confirm the expected conversation.
4. Follow the identified harness's reference before extracting. Each store has its own schema, duplicate-content traps, and subagent linkage; the recipes there were verified against real stores on the author's macOS machine at the versions and dates each reference states, so confirm them against the store in front of you.
5. Extract the human-readable conversation: real user turns and assistant text, skipping tool plumbing, system-injected context, and reasoning blocks. Follow subagent or child sessions when the main transcript delegates work the requester cares about.
6. Verify the summary against the extracted records before reporting. Dates, paths, decisions, and quoted claims in the handoff must appear in the transcript; a plausible summary that outruns its evidence is a failure, not a shortcut.
7. Deliver the handoff with its provenance: harness, session id, project path, time range, the resolution chain (which identifier matched which file, via which index), and any gaps such as encrypted reasoning or missing titles.

Prerequisites: jq for the JSONL recipes, and the sqlite3 CLI with JSON1 (json_extract) and FTS5 (messages_fts) for the SQLite recipes. When a tool is missing, do not install it without the user's approval; read the same files with a short Python standard-library script instead (json line by line for JSONL, the sqlite3 module with a file:...?mode=ro URI for databases), and fall back from FTS5 to LIKE filters if that build lacks it.

References, one per harness: