Reasoning models at high effort could spend the entire max_tokens budget
on thinking and return zero visible content with finish_reason=length;
the empty article flowed writer -> editor (reject) -> empty draft post,
with the cost booked as success (2026-06-11 incident: prod run acf24029,
draft post 921; diagnosis: convo pipeline-v2-phase1 seq 12).
- Empty-completion guard at the provider seam (new
OpenRouter_Completion_Guard): empty/whitespace content is a failure;
one retry with reasoning disabled (per-call override) when
finish_reason=length or reasoning was active, then the stage fails.
- finish_reason=length always warning-logged (model, stage, tokens).
- Failed/retried attempts booked to generation_log with
response_status=error + real token burn; run_decisions verdict rows
(retry_reasoning_disabled / failed_empty_completion) for v0.18.0 runs.
- Writer-path guard in Content_Generator::execute_stage and
belt-and-braces empty check in Publisher::create_post.
- Reasoning budget sanity: reasoning.max_tokens cap + completion
headroom via new prautoblogger_reasoning_max_tokens setting (default
constant 2048, 0 = legacy uncapped effort mode); cap replaces effort
when active; governor reservations use the effective ceiling.
- Body assembly moved to Request_Builder::build_body (300-line cap).
- Tests: guard verdict matrix, builder budget shaping, writer/publisher
guards, full provider retry path.
Agent-Session: prab-empty-draft-guard-20260611
2bdc763dc0