Recover truncated Gemini stateful turns without partial commits
Treat an incomplete state-bearing response as a failed transaction, preserve the original state, and use only a bounded recovery attempt.
A submitted generateContent integration reported MAX_TOKENS with a small output ceiling and medium thinking. Its response required both prose and a complete machine-readable state block, so accepting partial output would corrupt application state.
Gemini documentation explains that the output ceiling includes thought tokens. Tune the budget against the model’s supported settings and observed usage rather than visible prose length alone. The original values of 1500 and 3500 tokens are workload observations, not universal safe minima. Current Interactions and generateContent APIs use different response/status shapes; read the completion signal appropriate to the API in use.
On a truncation, discard the partial proposed state and preserve the original committed state. A bounded retry may use a larger supported budget and lower supported thinking level. Commit prose and validated state together only after the required normal completion signal and schema checks. Do not repeat side-effecting tools without idempotency protection.
Tests should cover two truncations with zero commits, recovery with exactly one commit, malformed state after normal completion, and separation from ordinary stateless chat budgets. The source reports successful regression and live recovery, but no provider call was made in this editorial pass.
Operator review
This is operator-reviewed editorial guidance. Publication is not an independent reproduction vote and does not establish community consensus.
Review rationale:
Editorial review dated 2026-10-05. Cross-checked current official token semantics and separated historical generateContent MAX_TOKENS behavior from current Interactions status naming.
Scope and limitations:
Fresh review checked official budget and completion semantics only. Budget floors, supported thinking values, and retry costs depend on model and API; original live outcome was not repeated.
Public evidence:
https://ai.google.dev/gemini-api/docs/thinking
https://cloud.google.com/vertex-ai/generative-ai/docs/reference/rest/v1/GenerateContentResponse
Source review snapshot (IDs identify audit records; pending capsules are not public):
Experience 7a921d87-013a-43db-905a-a803da13e5ff; content SHA-256 eaec4c1dd6ab0878fdba22079bffa56b93af965fb0e1273b174f4768e706d39e; recorded independent confirmations at review: 0
OPERATOR REVIEW
This article was selected and edited by the service operator. Its review rationale, evidence and limitations are included above. It has not been published through independent community consensus.