lucia-chat
Skill6 starsChanged 3 months agoMITView on GitHub
Set two agent sessions talking through the Luciazero Agent Bus and watch it live in a terminal: pick the pair, open the windows, read the transcript. Use for \"ให้ codex กับ claude คุยกัน\" or \"watch the bus\".
In this repository12 other files it gives its agents.
- bisectSkill
- debugSkill
- discipline-reportSkill
- doneSkill
- experimentSkill
- imouto-modeSkill
- lucia-busSkill
- lucia-relaySkill
- planSkill
- readySkill
- retroSkill
- showSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 0 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 12
After ./install.sh (on Windows, node bin\luciazero.js from the checkout), lucia is that command from anywhere and luciazero-agentd is the …
- Very specific: line 38
Read the answer back to the user and offer to run the command it names.
- Very specific: line 136
Ask the bus instead — lucia sessions for every live binding, lucia terminal list for what each provider session is bound to, lucia whoami …
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Lucia Chat
/lucia-bus is how you take part in the bus. This is how the user sets up a
conversation between two other sessions: one window per agent, and an optional
read-only pane showing every message as it lands.
After ./install.sh (on Windows, node bin\luciazero.js from the checkout),
lucia is that command from anywhere and luciazero-agentd is the same
program under its long name. Without the launcher, each command below is the
same one run as python3 -m luciazero_agentd (on Windows, python -m
luciazero_agentd) from the repository's agentd/ directory, which is the form
the bus itself prints; where Windows has only the py launcher on PATH, type
py -3 -m luciazero_agentd instead. Run them in the user's terminal, or hand
them over to paste. Never start a provider session on their behalf without being asked to.
1. Start here, always
lucia nextIt reads the bus and answers the only question the user actually has — what is waiting on whom — as the command that unblocks it, most blocking first: a delivery nobody could deliver or a task that ran out of budget (both need a person to decide something), then each agent with unread messages and the command that opens its session in its own worktree. If the daemon is down it says so and nothing else, because nothing else can happen first:
lucia serveRead the answer back to the user and offer to run the command it names; do not
paraphrase it into different commands. status is still there for the full
picture, and next never writes anything.
2. Open one window per agent
lucia claude
lucia codexThe provider is the verb: each starts that provider with the binding already in
place, as the agent whose id is the provider's own name. A second window for the
same provider needs a name of its own or the bus refuses the duplicate —
lucia claude --as reviewer is the agent claude-reviewer. An id that is not a
provider's name takes the long form, which is what chat prints:
lucia run --agent codex-architect -- codexEach agent's git checkout must be its own: start the second session from a
separate worktree, or worktree_bind refuses it. A daemon is started if this
state directory has none; --no-autostart refuses instead of starting one.
3. What starts the other session's turn
A session runs code during a turn and not one moment otherwise, so a delivery
landing while it sits at its prompt is read by nothing. run, and the short
forms above, hold the provider's terminal, so the bus can type one line into it:
the literal check your bus inbox, and in brackets only what it counts itself —
check your bus inbox (2 new tasks from codex). No word of a payload is ever
typed; payloads arrive through message_inbox, where /lucia-bus treats them
as untrusted input.
The knock is narrow on purpose: nothing is typed until the agent has used the
bus in this session, only a delivery arriving after that session started counts,
each knock waits out a 20-second cooldown, and knocking stops after 8 in a row
with nobody at the keyboard — any keystroke starts that count over, and
--max-nudges changes the cap. A delivery the cap holds back is not lost; it
knocks as soon as a person is back.
--no-nudge is the pull-only flow: that session is never typed into and reads
its inbox when its own turn next starts — which somebody has to start. It is
also what happens wherever there is no terminal to type into, such as a piped
run or a dispatched turn.
4. When the pair is not obvious
lucia chat
lucia chat --between codex-architect claude-implementer
lucia roster add claude-implementer claude implementerchat lists the roster with the terminal each agent currently holds, asks which
two, and prints the exact command for each window; --between skips the
questions. It reads the database read-only and writes nothing, so it is safe to
run mid-conversation. roster add names an agent that has never been seen.
5. Watch it happen (optional)
lucia watch --between codex-architect claude-implementerNeither session needs this. Open it first and the conversation is visible from
its first message. --payload full shows the whole body, --payload none only
who spoke to whom, and --agent X (repeatable) widens the filter beyond one
pair. It shows traffic and never touches it: it acknowledges nothing, because
acknowledged_at has to keep meaning that an agent opened the message itself.
6. Reading the pane
17:42:22 codex-architect -> claude-implementer [task] M7a: read-only inbox watcher
17:53:22 claude-implementer opened it after 11mThree different numbers live in that gap and they are not interchangeable:
- delivery latency — send to the peer's acknowledgement, the second line
above. Under a knock it is the knock plus what that session takes to start a
turn; under
--no-nudge, however long until somebody gives it one. - completion latency — send to the answer coming back: the reply, or
task_completewith its artifacts. - user-attributed blocking cost — how long the user's own work stood still waiting. An 11m gap is a person waiting, a session already mid-turn, or a window nobody sat at, and nothing recorded here tells those apart.
Report the first two from the timestamps, ask the user for the third, and say which one any number is. A retro counts toward the beta gate only when a human names the blocking cost.
7. Who is who
Starting a session prints no bus banner inside the provider: run names the
binding in the terminal that started it, before the provider takes the screen.
Ask the bus instead — lucia sessions for every live binding, lucia terminal
list for what each provider session is bound to, lucia whoami for the
terminal it runs in. Inside a session, /lucia-bus asks with agent_whoami.
8. Letting them answer each other
Turns started by the dispatcher instead of by a person are managed dispatch (M6). Each turn starts a real provider process and spends real quota, so it is never set up without the user asking for it in that many words:
lucia chat --between codex-architect claude-implementer --autoThat prints the commands and runs nothing. Enrol each side as a worker in its
own worktree, keep --approve workspace so a turn can work without being able
to accept whatever it is asked, and cap the run. A human approval nonce is still
unskippable for sensitive operations. An agent cannot be dispatched and hold a
human terminal at the same time: the managed turn opens its own session.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

