Command Code

  • ID: commandcode
  • Store: ~/.commandcode/projects/<encoded-cwd>/<session>.jsonl
  • Read override: DEJA_COMMANDCODE_ROOT replaces the project root
  • Format: JSONL — a session header, then one message envelope per line
  • Needs: nothing
  • Resume: cmd --resume <id> (cmdc on Windows), run in the header's cwd, since it finds a session only under the current directory's project (#4372). With that directory gone, or none recorded, cmd --session <id>, which searches every project and continues the same transcript where you run it (checked against 1.73.4, #4460)

Command Code writes a Claude-Code-style project directory with one transcript

per session. Since 1.73 (format version 3) the first line is a header,

{"type":"session","version":3,"id","timestamp","cwd"}, and every line after

it is {"type":"message","timestamp",…,"message":{"role","content"}}, the

content in Claude's blocks (text, tool_use, tool_result). The project is

the header's cwd; the directory name is a lossy slug of it.

Tool calls use Command Code's own names with Claude's input keys:

shell_command (command), read_file, edit_file (file_path,

old_string, new_string), write_file (file_path, content) and

read_multiple_files (paths). All of them are indexed as commands, files,

replaced spans and written lines (#4370). A failed command's tool_result

opens with "Exit code: N" (or "Exit code: N (No matches found)" and the like)

and has no is_error; that code goes on the command the result answers as

→ exit N. A clean run writes no such line and gets no code. A result that

lands in a later index pass than its call is read with the call (#4539).

1.74 adds powershell (command)

and monitor_command, and shell_command and monitor_command take an args

list beside command; the command is recorded with its arguments as the client

shows it, go test ./... for {command: "go", args: ["test", "./..."]}.

read_file takes a paths list too, where an entry with *, ?, [ or {

is a glob the client expands, not a file, and is left out (#4540).

The older shape, one flat role/content/timestamp/sessionId line per

message, is still read. The client migrates such a file to v3 in place the

next time it opens the session and keeps the old copy as <session>.v2.bak.

Last verified: 2026-10-07 (command-code 1.77.0)

Known quirks and drift

  • A sibling with the same extension. <session>.checkpoints.jsonl sits

beside the transcript and is a snapshot stream, not a conversation — read as

one it adds a session with no words in it that then competes for a recall

slot. It is skipped by name, and a test pins that. So are

<session>.prompts.jsonl and <session>.v2.bak, which is the client's own

transcript filter, and doctor counts them and the <session>.meta.json

sidecar as files beside the transcripts rather than as unread.

  • A line whose role is neither user nor assistant is tool output, which is

where a command's error text lives. It is indexed as tool output rather than

dropped, so a user can search for an error they have already hit.

  • Wiring: deja install commandcode writes all four of its user-scoped

surfaces, and commandcode-auto adds the hooks —

~/.commandcode/mcp.json for the server, skills/deja-history/SKILL.md,

commands/deja.md for /deja, and a hooks key in settings.json in

Claude's shape.

  • Two things there fail quietly if copied from another harness. The

timeout unit is seconds, not milliseconds — the numbers deja passes to

other settings.json targets would be ten minutes each, past the documented

600-second maximum. And the tool matcher is a regex over Command Code's own

display names — SHELL, READ, EDIT, WRITE, SEARCH, GLOB, LIST

— so a Claude-shaped Bash matcher never fires at all. Both are pinned by

tests. The matcher is case-insensitive, so SHELL also fires on

powershell; the hook payload carries the internal tool name and

tool_input.args after command, and both are read (#4540).

  • There is no per-prompt, compaction or session-end event among the shell

hooks: the four documented are SessionStart, PreToolUse, PostToolUse

and Stop, so the digest rides SessionStart. The rest goes through a mod,

~/.commandcode/mods/deja.ts, which commandcode-auto writes and Command

Code loads at the start of every session. Its transformContext adds the

question's recall to the question on each model call of the turn (it never

writes back, so the transcript keeps the person's words), onSessionEnd

ends the live stamp, and cmd.ui.setStatus puts deja statusline in the

TUI footer. A PreToolUse cannot answer the question: a turn reaches the

transcript only when it commits, after its tool calls (1.77.0 stand).

  • Compaction appends {"type":"compaction","summary","firstKeptEntryId"} and

keeps every turn above it. The next matched PreToolUse reads the session

as it stood before the newest one and answers with the packet, once. The

summary is indexed as a summary.

  • The transcript format above was read off files command-code 1.73.4 wrote

and its bundle (toStoredEntry, parseV3Lines, migrateToV3). The wiring

paths come from rulesync's commandcode-paths.ts and types/hooks.ts, and

akitaonrails/ai-memory's support matrix.

deja reads this format and 42 others, and turns what it finds into memory your agents can search. See the harness matrix for what is wired where, or install it and search your own history.

Found this useful? Star deja-vu on GitHub.