v0.2.0

Four parameters — context_window, effort, temperature and top_p — were read once, when a session started, and never looked at again. They now resolve from layers, at every moment that can change the answer. The rule is one sentence: when the config speaks the config decides; when the config is silent your hand-set value stands.

What changes

• /model re-evaluates the four parameters. Switching model used to move the provider's model id and the id in the session meta, and nothing else: a chat that went from a 400k model to a 128k one kept counting against 400k, and an effort inherited from the entry it left followed it to a model that had never asked for one. The switch now asks the model it arrives at what it declares, applies whatever moved to the provider and the token budget, and says so one notice at a time. Choosing the FIRST model does the same — a run that started on a provider:* wildcard had no entry to read them from until then.
• A value the chat inherited is dropped on a switch; one you set by hand is kept. Which is the same rule from the session's side. A window that came from a declaration gives way to the built-in default on a model that declares none, rather than silently following you to a model half its size; a window you typed in /model survives. Adjusting a knob there is the answer for what no declaration covers, not a permanent override of one — to change a value for good, write it into the config.
• defer_mode is not one of the four. The deferring wrapper is built once, around the MCP dispatcher, and there is no seam to rebuild it mid-chat, so switching to a model that asks for another mode prints a note that this session keeps the one it started under. A context_window: on the model being switched to that does not parse is a warning for the same reason: the chat is already running, and the honest answer to a bad value is to leave that key silent and say so.

What is new

• agents.<name>.context_window. The same key as the model's, parsed the same way, and the top layer: an agent that knows how long its conversations run says so once instead of forking a models: entry per usage.
• Three tiers, two evaluating moments. agents: first, then what the model the chat is RUNNING declares, then — lowest — the value the session is already running under. A new session and a model switch are the only two moments that evaluate them.
• A resume is not one of them. iota resume restores the bundle's own values and the origin of each, so an old session continues exactly as it was and a config edited in between reaches it the first time you switch models inside it. Only a parameter the bundle never recorded is evaluated.

On disk

• meta.json gains top_p, which now replays beside effort and temperature as it could not while nothing wrote it down, and a record of where each of the four values came from — a config declaration, your own hand, or the built-in default. Both keys are omitted when they have nothing to say, so a bundle written by 0.1.0 keeps its byte shape and still resumes; its silence is read as every value here is the user's own, the reading that cannot lose something a person chose.

Upgrading

Nothing to edit: a 0.1.0 config loads unchanged and a 0.1.0 session resumes. The one habit worth revisiting is a knob set in /model and relied on to outlast a model switch — it now yields to a model or an agent that declares that parameter. If it should always hold, it belongs in models: or agents:, and the config file has the table.