gajae-code

  • ID: gjc
  • Store: ${GJC_CODING_AGENT_DIR:-~/.gjc/agent}/sessions/<project-slug>/<session>.jsonl
  • Sub-agent passes: …/<project-slug>/<session>/N-*.jsonl
  • Read override: DEJA_GJC_ROOT replaces the session root
  • Format: pi's session JSONL
  • Needs: nothing

gajae-code (gjc) is a pi descendant: a session header carrying the id and

the cwd, then one message line per turn, so the parsing is pi's. The encoded

project directory names the project, and the header's cwd wins when it is there.

Last verified: 2026-09-21

Known quirks and drift

  • Resume: gjc --resume <id>. From its session-operations document:

--resume <id|path> at startup opens an existing session, and a session

belonging to another project forks into the current one — so deja prints

the command without a working directory rather than guessing at one.

  • Sub-agent passes sit one directory deeper, under a directory named for the

session they belong to, one file per pass. They are skipped: a sub-agent's

transcript repeats the parent's work in its own words, and indexed as a

session of its own it competes with the parent for the same recall slot.

DEJA_INCLUDE_SUBAGENTS=1 takes them, the switch Claude Code's and Cursor's

sub-agents already use.

  • service_tier_change lines are not turns and are dropped rather than read as

empty messages.

  • Wiring: deja install gjc writes the server into ~/.gjc/agent/mcp.json

and the skill into ~/.gjc/agent/skills/deja-history/SKILL.md. Both paths

are from gjc's own surface table (docs/customization.md), and the skill

location matters: gjc loads its native skills directory, while Claude's and

Codex's are import candidates it does not read, so a skill written there

would be a file no session ever sees.

  • Auto-recall is deja install gjc-auto, and it is pi's extension unchanged:

loadExtensionModules in src/discovery/builtin.ts discovers modules in

<agent dir>/extensions (native .gjc/.pi only, a *.ts file or a

directory with an index), and the events in

src/extensibility/extensions/types.ts are pi's — session_start,

before_agent_start, context, tool_result, session_compact — with the

same result shapes.

  • The directory hooks stay unwired, and not for want of a path. gjc's two

documents disagree on it (~/.gjc/hooks/{pre,post} against

~/.gjc/agent/hooks/{pre,post}), and the loader settles it:

resolveScopePaths puts the user-scope hooks at <agent dir>/hooks/<pre|post>,

so docs/hooks.md's ~/.gjc/hooks is stale. What stops a hook being written

there is the shape: a directory hook is a module exporting

default (api) => api.on("tool_call", …) whose only documented return is

{block, reason} — allow or refuse, with no channel for adding context. The

extension has that channel, so that is where recall goes.

Measured on a live install

gajae-code 0.17.1 (which wraps @gajae-code/coding-agent), in a hermetic

HOME:

  • deja install gjc writes the three paths this entry claims:

~/.gjc/agent/mcp.json, ~/.gjc/agent/skills/deja-history/SKILL.md and

~/.gjc/agent/commands/deja.md.

  • gjc's own --help confirms the MCP path in its own words — --no-mcp

disables "conventional MCP autoload (native user ~/.gjc/agent/mcp.json and

project .gjc/mcp.json registrations)" — and -r, --resume[=<value>] takes an

ID prefix, a path, or opens a picker.

  • On 0.17.1 its screens could not be read: every entry point, gjc mcp list

and gjc -p included, exited with `Cannot find module

'../../../../node_modules/mupdf/dist/mupdf-wasm.wasm'`.

0.17.2 runs (it wants Bun 1.4 and says so), and the rows above are its own

screens rather than its documentation:

  • gjc customize doctor reports ~/.gjc/agent/extensions/deja.ts as "a

trusted filesystem extension module discovered for session-start loading",

and ~/.gjc/agent/commands/deja.md as a loaded slash command.

  • The same screen is where the skill location proves itself: a copy under

~/.agents/skills is listed source-ignored, "GJC loads skills only from

.gjc (native) locations", which is why deja install gjc writes gjc's own

directory.

  • gjc -p against a mock provider carried deja's <deja-recall> block into

the request — the past session, the question it answered and the conclusion —

on both the OpenAI (/v1/responses) and Anthropic (/v1/messages) paths,

and the server in mcp.json was connected in the same run: the tool

inventory in the system prompt lists deja/deja: mcp__deja_deja.

deja reads this format and thirty-three 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.