ReadonlyappendExecution chunk size, independent of the number of tokens injected.
ReadonlydecodeOptional token→text decoder for the proposal trace (set by the serve layer, which owns the tokenizer).
ReadonlyplanReadonlystatsAre strict rows still armed? (Cleared by noteParseFailure.)
Record what the engine actually emitted from proposal (all of it for an
assert; the accepted prefix for a verify; 0 when the engine declined).
MUST be called exactly once per proposal returned by push().
The tool-call parser rejected this request's markup. Strict rows are disarmed for the rest of the request — the template's own rendering disagreed with what the model actually emits, so every further row is suspect. (Echo rows are unaffected: they never claimed to be schema.)
REALITY NOTE: the served parse (ToolAwareStream.toolCalls → parseGeneratedToolCalls) runs at sink FLUSH, i.e. after the generation has ended, so within one request this is telemetry and the disarm has no remaining tokens to act on. The seam is here so a future incremental parse (or a plan cached across requests) disarms for real.
A verify proposal reached the forward (the checkpoint was taken).
A verify proposal was dropped: this model's caches cannot rewind.
The engine dropped an in-flight sampled token to take an ASSERT fill.
Record sampled output independently of when the method requests proposals.
Select a continuation after a committed output boundary. Methods may request only verified proposals while retaining ordinary sampled policy.
Optionalpolicy: ProposalPolicyFeed one SAMPLED token, then ask the sources what follows. budget is
the caller's remaining allowance (max_tokens − generated).
One request's proposal sources plus the running match state.
CONTRACT. push() returns a proposal; the caller MUST then call commit() exactly once with the number of ids it actually emitted (all of them for an assert, the accepted prefix for a verify, 0 when it declined). Nothing enters the session's history until commit — so a partial accept needs no rollback here.