prime-agent (PrimeIntellect)

Discovery

prime-agent (PrimeIntellect-ai/prime-agent) keeps one file per session

directly under ~/.prime/agent/sessions/, named by the session's uuid7. The

root is flat: unlike pi and omp there is no encoded project directory, so the

project comes from the cwd the header line carries.

prime-agent relocates the root with two variables of its own,

PRIME_AGENT_SESSION_DIR and PRIME_AGENT_CODING_AGENT_SESSION_DIR, and deja

reads both — a machine that has moved its sessions has moved them for deja too.

DEJA_PRIME_ROOT overrides all of it.

rlm.spawn writes each child session as its own transcript beside the root,

under ~/.prime/agent/session-artifacts/<parent-id>/sub-<n>/<child-id>.jsonl,

with parentSession and an rlmDepth above 0 in its header. deja reads those

too and files each as a subagent of the parent, read as a Claude Code subagent

is: the task, what it changed and its last turns. DEJA_INCLUDE_SUBAGENTS=1

takes the whole run, =0 leaves children out. The semantic-edges.jsonl next

to it is prime's event log, not a transcript.

Older installs kept sessions under ~/.pi/agent/*.jsonl and in a --cwd--

directory beneath this root. prime-agent migrates both into the flat root when

it starts, so the flat layout is what a live install has.

Shape

The first line is the session header:

{"type":"session","version":3,"id":"<uuid7>","timestamp":"...","cwd":"...","parentSession":null,"rlmDepth":0}

The lines after it are typed entries — message, model_change,

thinking_level_change, service_tier_change, compaction, branch_summary —

chained by parentId. Message entries carry content text blocks.

That is the envelope pi writes, which is why deja reads it with the same parser

pi and omp share: prime-agent descends from the same codebase

(@earendil-works/pi-coding-agent) and kept the format.

The model has one tool, ipython, so a change to a file is Python in a cell:

the bundled edit skill's await edit(path, old_str, new_str). Each change it

writes lands on the cell's toolResult as `details.diffs[{path, oldStr,

newStr, startLine}]`, and deja records the files, the replaced spans and the

written lines from there (#4526). A command run with bash() inside a cell is

only in the cell's code and is not recorded as a command.

What deja does with it

Sessions are indexed and searchable like any other harness.

• MCP: deja install prime adds a stdio entry under mcpServers in

~/.prime/agent/settings.json (under PRIME_AGENT_CODING_AGENT_DIR when

that is set).

• Skill: the shared ~/.agents/skills/deja-history/SKILL.md, which

prime-agent loads.

• Auto-recall: deja install prime-auto also writes

~/.prime/agent/extensions/deja.ts. At before_agent_start it returns the

session digest on the first turn and per-prompt recall after that;

session_start shows a footer status while the first index builds, and

session_compact runs deja hook-precompact and session_shutdown

deja hook-session-end. It registers `/deja

<query>, which runs deja search. tool_result fires for the ipython`

cell, in --print too (0.9.8): the handler adds the earlier fix when a

bash() in the cell exited non-zero, else the line deja keeps for a file the

cell's code opens or edits.

• Resume: cd <cwd> && prime-agent --resume <id>, with the cwd from the

header; prime-agent resumes a session only from the project it ran in.

• Handoff: exec.

Reported and specified from source by @iMaxTomas in #2529.

Last verified: 2026-08-30

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.