Read Session

Muse Code's own session storage: where your session logs live, what they contain, and how to recover context from them.

Your Store, Not Anyone Else's

You are Muse Code (binary muse). Your sessions are stored under YOUR data directory:

${XDG_DATA_HOME:-$HOME/.local/share}/muse/sessions/YYYY/MM/DD/<session-id>/

Each session directory contains:

• session.jsonl — the main event log (one JSON record per line).
• subagent/<child-session-id>/session.jsonl — one log per delegated subagent session.
• tool-outputs/ — full tool outputs that were too large to keep inline.

Muse sessions NEVER live in another coding agent's store. Do not look for a Muse session under ~/.claude, ~/.claude/projects, ~/.local/share/claude, ~/.codex, or ~/.grok — those belong to Claude Code, Codex, and Grok. Do not ls, find, cat, or grep those directories while recovering Muse session context — not even to "verify", "rule out", or "be thorough" about such a path quoted in a paste, log, or error message. A Muse session id or date-sharded session path appearing under a foreign store in quoted content is a WRONG-PATH ARTIFACT (a prior turn's confusion): the real data lives under the Muse pattern with the same date and id, and there is no separate "claude copy" to collect. Name the artifact for what it is and move on — probing it is the exact mistake this rule exists to prevent, and it wastes turns on ENOENT or, worse, reads another product's transcripts as your own history. Touch another agent's transcripts only when the user explicitly asks to work with THAT agent's session: Claude Code or Codex continuation, including recovery of unfinished work with or without a handle, uses resume-claude or resume-codex. Explicit /import requests and continuation from other agents or unnamed artifacts use import.

Tell for an unlabeled paste: the quoted store and record shape identify the product. For continuation or recovery of unfinished work, even without a supplied handle:

• Claude Code records under ~/.claude/projects use resume-claude.
• Codex records under ~/.codex use resume-codex.
• Grok records under ~/.grok use import.

Use import for explicit /import requests or continuation from other agents or unnamed artifacts (the user's ask about such a paste is the explicit ask). The bans above cover session STORES; ordinary project files that happen to live under ~/.claude (e.g. skills you are developing) are file work, not session probing.

Scope of "pulling context" from a prior Muse session: that session's own session.jsonl and subagent/ logs. External agents or planner processes the session MENTIONS (e.g. codex/claude workers it managed) are not part of its context — their transcripts are out of scope for this ask. If one looks load-bearing, say so and let the user ask for it. Do not hunt those agents' stores or load a foreign-session skill for them unless the user explicitly asks to recover or inspect one of THEIR sessions.

Finding The Right Session

1. Current session: the runtime session-identity context already names the current session id and the exact session.jsonl path. Use it; do not guess or search for it.

2. A pasted or quoted absolute path under .../muse/sessions/...: use that path directly.

3. A known session id without a path: the store is sharded by local date, so check the likely dates first, then search only the Muse sessions root:

   MUSE_SESSIONS="${XDG_DATA_HOME:-$HOME/.local/share}/muse/sessions"
   ls -d "$MUSE_SESSIONS"/*/*/*/*/ 2>/dev/null | grep <session-id>
   

4. "Our last session" with no id: list the most recent date shards and pick the newest session directory for this workspace (the log's early runtime.session.metadata record carries workspace_root).

5. Not found under the Muse sessions root: ask the user for the id or path. Never widen the hunt to other agents' directories or home-wide scans.

Reading A Session Log

Each session.jsonl line is an event-log envelope:

{"schema_version":1,"id":"…","stream":{"kind":"session","id":"<session-id>"},
 "sequence":42,"recorded_at":1771088000123456,"record_type":"event",
 "durability":"durable","causation_id":null,"payload_type":"runtime.session",
 "payload_schema_version":1,"payload":{"kind":"run","run_id":"…",
 "event":{"kind":"assistant_message_committed","text":"…"}}}

The useful payload.event.kind values for context recovery:

• started — what the user asked: every submitted prompt is recorded here (text in payload.event.prompt). A user_prompt_display record follows only when a differing user-facing form exists (an attachment placeholder such as [Image 1], a composer form that differs from the sent text) — prefer its text when quoting the user. A prompt typed while a run was active is instead an inbox_item_queued whose source.source is "user_steer" (text in payload.event.payload.prompt, or in payload.event.body when the record carries no payload); background and scheduled runtime deliveries share that kind, so never quote those as the user.
• assistant_message_committed — what the agent concluded (decisions, summaries, handoffs usually live here, late in the log).
• assistant_tool_calls_committed / tool_result_batch_committed — what was actually done and what it returned.
• terminal — turn boundaries.

Read discipline for long logs: read the TAIL first (later records matter most), then only enough earlier evidence to understand context. Use bounded tail/grep slices; never load a whole multi-megabyte log into context. Subagent findings live in subagent/<id>/session.jsonl, not the main log. For product debugging of the current session, prefer the doctor skill's session-evidence helper.

Restoring Lost Work

When the user asks to restore or recover lost, wiped, or overwritten work and a prior session's log holds the only copy of that content (as recorded tool-call mutations), reconstruct each requested file exactly; do not guess between versions:

1. Enumerate every mutated workspace path first — write_file, edit_file, apply_patch, and shell writes — and build the complete file list before restoring anything. Restore the full scope of the ask: for a general "restore what was lost" ask that means every lost file, while an explicitly narrower ask wins as stated.
2. A file's final state is its mutation history replayed in record order (sequence/recorded_at): the LAST write_file content for the path in record order — or the newest full-file apply_patch/shell write when that came later — with every LATER edit_file delta applied in the same order (patch hunks and shell edits likewise). Skip an edit_file/apply_patch delta whose tool result reported an error; an errored shell command may still have mutated the file first, so check whether later records reflect its write. Recency comes from record order, never by content length or size: the longest version is often a superseded draft, and refactors make the final version SHORTER. Reconstructions built from write_file records alone silently drop every later edit.
3. Restore the exact recorded content, not a paraphrase. Re-typing from memory, rewording, "improving", or summarizing content that is recoverable verbatim is data loss, not a restore — extract the bytes from the record and write those. (A user asking only to summarize a prior session is not a restore; this rule governs restoring files.)
4. Report the result per file: which records it was restored from (the last full write plus the deltas applied), and what was restored and what was not, with a reason for anything skipped.
5. If two candidate final versions are genuinely ambiguous (e.g. divergent edit branches after a fork or resume), present both candidates with their record timestamps and let the user pick; never silently prefer the longer one.

First-Party Commands

Prefer these over hand-rolled parsing when they fit:

• muse resume <session-id> or muse resume --last — interactive continuation.
• muse exec --session-id <session-id> "<follow-up>" — headless continuation, only on explicit request.
• muse export --session <id-or-session.jsonl> --redacted --out <file> — a shareable, redacted export.
• muse trace inspect --session-log <session.jsonl> --render-mode compact — model-call level inspection.

Treat session logs as read-only evidence: never modify, move, or delete them.