Activates a persona. Activating several personas in one conversation to make them answer each other is a debate: use start_debate.
`persona_slug`: the persona you speak as, `soov` by default. A former name
(`rodin`, `occam`, `leonard`) still works; the answer then gives the
current slug and an `alias_note`.
`purpose` (a string): `session` (default) when the user opens a
conversation with this persona; `debate` or `consult` in a subagent opened
for a debate or a consultation. Only `session` reads and returns
`previous_context`: apply it, then call `clear_previous_context`.
`topic` (debate or consult only): the question; the memory pack then
serves Core and the most relevant entries on it.
Before `archetype_md`, the answer gives `rules_md` (the immutable debate
rule) and `team_md` (the team block, then the team sheet of this
account), then the memory pack. The answer stays under a budget (about
22,000 characters, 49,500 with the Large setting of the Mémoire page,
12,000 for `debate` and `consult`): the rule, the team and the archetype
are served in full, the memory pack fills the rest. `mode`: `lite`
(default: the `Core` entry, the most used procedural rules, an index of
the rest), `medium` (semantic and procedural entries) or `full` (plus the
last `recent_events_n` events). `Core` is the semantic entry titled
exactly `Core` in the global memory, served first in every mode;
maintain it with `update_semantic("Core", ..., project_scope="global")`
(on plans without general memory, the server keeps it in the project).
The general memory is read only on the Pro plan, and outside a project it
is the only memory read.
`memory_note` counts what is not served: `search_memory` fetches it on
demand. `pack_chars` counts the whole answer.