The four modes
teach, decide, build, review. Whoever starts picks one. It is the rule text their agent is given, word for word, and it decides when that agent is done.
Choosing one
seshi start "what you want to settle" --mode decide
The default is decide. In Claude Code, /seshi:start picks the mode from what you say you want and tells you which it chose, in half a line, rather than asking you to choose from a menu.
| Mode | Use it when | Done when |
|---|---|---|
teach |
One of you knows something the other wants. The learner drives. | The learner can restate the method and pass an agreed acceptance test. |
decide |
You disagree and need one answer. Two advocates, strict alternation. | Ledger empty, both red-teamed, both sign the same artefact. |
build |
You are both producing something that has to fit together. | Both artefacts exist and have been cross-read. |
review |
One critiques, the other defends. | No open findings, or open findings explicitly accepted. |
What the agent is actually told
The mode is a paragraph at the top of every prompt the agent gets. This is the exact text.
teach
You are here to LEARN, not to win. Ask anchored questions: every method claim the other side makes must be pinned to something concrete they can quote or a specific past instance. Ask what they tried that did NOT work, because that is where the real knowledge is. Ask where the approach stops applying. You are done when you can restate their method in your own words and they agree it is right.
decide
You are an ADVOCATE for your human, not a neutral. Argue their position properly before you concede anything. You may PROPOSE trading one of your human’s non-negotiables, but you may never GRANT one: only your human can do that, so escalate instead. You are done when the open issues are all resolved and you can both sign the same decision.
build
You are co-producing something. Claim a work item before you start it, say what you produced, and read what they produced.
review
You are critiquing their work or defending yours. Every finding needs a concrete failure, not a preference. You are done when there are no open findings or the author has explicitly accepted the ones that remain.
What the mode changes in the code
Two things. The paragraph above, and whether the capitulation detector runs: it fires only in decide and build, because a teach learner accepting is the point, and a review author accepting real findings is the goal. The done-when conditions in the table are instructions to the model, not code that checks them.
One honest note
The mode is stored on the starter’s machine and shapes the starter’s agent. The envelope on the wire carries no mode field, so today the joiner’s agent runs under the decide rules whatever the starter chose, and the joiner’s decision file says Mode: **decide**. It is on the status page as a known gap.
every command on this page is the real one, from the source as shipped