Defaults and limits¶

The values the module uses when you configure nothing, and the caps that are fixed in the code and cannot be raised through configuration.

Default models¶

Config.Model left empty selects these:

All three are chosen by one rule: the most capable model the provider makes generally available that supports structured output, tool use and multimodal input. That favours capability over cost, and it moves when vendors release, pin Config.Model if you would rather it did not. The reasoning is in Providers.

Two consequences are worth carrying in your head. Two of the three defaults refuse Config.Temperature (claude-opus-5 and gpt-5.6-sol) while gemini-3.7-flash still accepts it, up to a maximum of 2. And gemini-3.7-flash is a genuine step below the other two, because Google ships no generally-available Pro-tier model.

Those two facts pull in opposite directions, which is the honest position: the one default that still takes sampling controls is also the least capable.

Default limits and timeouts¶

DefaultCompactionTurns is far larger than a truncation budget would be, and deliberately so: dropping turns is free, summarising them costs a round-trip, so a conversation should be allowed to get genuinely long before it starts paying for one. Compaction's other budget is a share of the model's own reported input limit rather than a constant, because a number right for a 200k window is wrong for an 800k one.

The request timeout is generous on purpose (a large single-shot generation on a flagship model runs well past a typical 30-second HTTP default) but it is bounded on purpose too, so a model stuck in a loop fails instead of hanging forever.

Caps you cannot configure¶

These are fixed in the code. There is no field, option or environment variable that raises them; changing one means changing the module.

Rate limits, per-minute token budgets and monthly quotas are not in this list because the module has none. Those belong to the provider, and the module surfaces the refusal as chat.ErrThrottled rather than pacing calls itself.

The effort ladder¶

Config.Effort accepts exactly five values, in ascending order:

Anything else is dropped at construction with the valid list in the hint, and the provider's own default effort is used. Effort.Valid() reports whether a value is on the ladder, and Effort.Rank() gives its position so an adapter can clamp onto a shorter ladder by index.

Media types the module will send¶

An attachment's type is sniffed from its bytes with net/http.DetectContentType and must appear in the allowlist below; a declared Media.MIMEType is only ever a cross-check, and the sniffed type is what is sent.

claude-local accepts no media at all: passing any attachment returns chat.ErrMediaUnsupported. Formats the standard library cannot positively identify (mov, flv, wmv, 3gpp, flac, m4a) are rejected as unidentifiable even where the vendor would accept them, because the sniffer, not the vendor, is the constraint.

Support values¶

chat.Support is three-valued, and SupportUnknown is its zero value, so an unpopulated capability map reads as "nobody could say" rather than "no".

Related¶

• Configuration fields: which field each default belongs to.
• Errors: the sentinel each rejection carries.
• Check what a model supports: asking before you set a value.