TRAE CLI

  • ID: trae
  • Store: ${TRAE_HOME:-~/.trae}/cli/sessions/YYYY/MM/DD/rollout-<timestamp>-<uuid>.jsonl, plus archived_sessions/ for traex archive <id> and history.jsonl beside them
  • Read override: DEJA_TRAE_ROOT replaces the cli directory; TRAECLI_HOME moves that directory and TRAE_HOME moves TRAE's whole home, and deja follows both
  • Format: JSONL — Codex rollouts: session_meta, turn_context, response_item, event_msg, plus TRAE's own history_mutation
  • Needs: nothing (zstd only for a compressed rollout)
  • Resume: traex resume <id>
  • Wiring: deja install trae adds the MCP server to ${TRAE_HOME:-~/.trae}/traecli.toml; deja install trae-auto adds Codex's hooks to ${TRAECLI_HOME:-${TRAE_HOME:-~/.trae}/cli}/hooks.json as well, which TRAE runs once they are trusted at start-up

TRAE CLI 2.0 ships as traex and reports itself as traecli 0.200.x. It is a

closed-source fork of codex-rs and writes the same rollout files Codex does,

with originator: "codex-tui", one level further down at ~/.trae/cli. deja

reads them with the Codex reader and files the sessions under trae, not

codex, because they resume with traex and three things in them are read

differently.

Who said it. TRAE writes its own runtime injections — environment context,

process-limit notices — as response_item messages under the user role, so

that record is not evidence a person typed anything. A prompt is taken only

from event_msg user_message or, from 0.201.4, event_msg item_completed

carrying a UserMessage item. The two can describe the same prompt, in either

order, and the older one sometimes has no turn_id; a pair is read once — same

text and turn id, or same text within five seconds when one side has no id — and the same

words in a later turn are a second prompt.

The answer. It is taken from the first of these that the rollout has:

assistant response_item messages, assistant messages in history_mutation,

item_completed AgentMessage items (text blocks typed Text, text or

output_text), and agent_message events.

Tool calls. They ride in history_mutation records with `operation:

"append". The shell is exec with an argv under command`

(["bash","-lc","<script>"], the script is what is kept); a result's output

is a block list whose text blocks are joined; apply_patch is the Codex patch

format. A call seen both there and as a response_item is read once, by

call_id. A replace re-sends history the file already holds and is skipped.

Last verified: 2026-10-05

Known quirks and drift

  • The shapes come from other readers, not from a running install. The

user-turn rule, the item_completed dialect and history_mutation are read

from botmux's TRAE reader (src/services/traex-transcript.ts and its tests);

the store layout and the resume command from agentsview's TraeX provider

(internal/parser/traex.go, types.go, server/resume.go). Nothing here

has been checked against a rollout TRAE wrote on this machine.

  • An append that splits a pair is read whole. When the new bytes hold the

second record of a prompt, or a repeat of a call, read before the offset,

the incremental pass reads the file from the start, so neither lands twice.

  • The 1.x store is not read. TRAE CLI 1.x (traecli, internally CoCo)

wrote events.jsonl under ~/Library/Caches/coco/sessions/ on macOS and

~/.cache/coco/sessions/ on Linux, a different format.

  • The IDE store is not read, but the IDE is wired. TRAE, TRAE CN and TRAE

SOLO keep chat in a SQLCipher-encrypted ModularData/ai-agent/database.db.

deja install trae-ide adds the server to <user data>/User/mcp.json

(~/Library/Application Support/Trae, %APPDATA%\Trae or ~/.config/Trae;

Trae CN for the CN build) with no type key, which its schema rejects, and

the skill to ~/.trae/skills (~/.trae-cn/skills); with both builds installed,

each is wired in its own files.

deja install trae-ide-auto adds Claude-shaped hooks to ~/.trae/hooks.json. The IDE ships

with hooks off and keeps that switch in its own settings store, so deja cannot

tell whether they run: turn them on in Settings > Hooks and run them locally.

  • ~/.trae/hooks.json is the IDE's, not the CLI's. TRAE CLI 1.x read it;

2.0 reads cli/hooks.json and says the old file is no longer read.

deja reads this format and thirty-seven 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.