fix: empty-content guard + finish_reason=length handling + reasoning budget (v0.18.1)

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
This commit is contained in:
peptiderepo
2026-06-11 13:15:58 +00:00
parent f5658f1c97
commit 2bdc763dc0
19 changed files with 1196 additions and 34 deletions

View File

@@ -64,6 +64,7 @@ class PRAutoBlogger_Chief_Editor {
'temperature' => 0.3,
'max_tokens' => 5000,
'response_format' => array( 'type' => 'json_object' ),
'stage' => 'review',
)
);

View File

@@ -78,6 +78,7 @@ class PRAutoBlogger_Content_Analyzer {
'temperature' => 0.5,
'max_tokens' => 8000,
'response_format' => array( 'type' => 'json_object' ),
'stage' => 'analysis',
)
);

View File

@@ -197,6 +197,12 @@ class PRAutoBlogger_Content_Generator {
PRAutoBlogger_Run_Stage_State::start( $this->run_id, $stage, '', $this->item_key );
}
// v0.18.1: stage metadata for the provider's empty-completion
// guard — warning logs and failed-attempt audit rows carry the
// real stage + registry key (never sent upstream).
$options['stage'] = $stage;
$options['prompt_key'] = $prompt_key;
$ctx = PRAutoBlogger_Opik_Trace_Context::current();
$span_id = $ctx->start_span(
@@ -223,6 +229,21 @@ class PRAutoBlogger_Content_Generator {
$options
);
// v0.18.1 writer-path guard (belt over the provider seam, which
// already throws for OpenRouter): no empty article may ever leave
// a writing stage — it would flow to the editor, be rejected as
// "draft text is missing", and land as an empty draft post.
if ( '' === trim( (string) $response['content'] ) ) {
throw new \RuntimeException(
sprintf(
/* translators: 1: writing stage name, 2: model identifier. */
__( 'Writing stage "%1$s" produced empty content (model %2$s) — failing the stage instead of passing an empty article downstream.', 'prautoblogger' ),
$stage,
(string) $response['model']
)
);
}
$ctx->end_span(
$span_id,
array(

View File

@@ -200,6 +200,7 @@ PROMPT;
array(
'temperature' => 0.7,
'max_tokens' => self::REWRITER_MAX_TOKENS,
'stage' => 'image_prompt_rewrite',
)
);

View File

@@ -82,6 +82,23 @@ class PRAutoBlogger_Publisher {
?string $run_id = null,
?PRAutoBlogger_Cost_Tracker $cost_tracker = null
): int {
// v0.18.1 belt-and-braces: never create a post whose content is
// empty once tags and whitespace are stripped — an empty "draft
// for human review" (2026-06-11 incident, post 921) reviews
// nothing and hides the underlying generation failure. The
// provider/writer guards make this unreachable on OpenRouter
// paths; this protects every other path into post creation.
if ( '' === trim( wp_strip_all_tags( $content ) ) ) {
throw new \RuntimeException(
sprintf(
/* translators: 1: intended post status, 2: article title. */
__( 'Refusing to create %1$s post "%2$s": generated content is empty.', 'prautoblogger' ),
$post_status,
$idea->get_suggested_title()
)
);
}
// v0.18.0 idempotency: post creation is keyed by run + idea
// (check-before-insert) so a retried/resumed run cannot create a
// duplicate post. One run_id spans all N articles of a batch, so