AI: surface an empty final answer as an error instead of a silent success

[juanlentino]

Scoped to the adapter change green-lit in #517 (plus the belt-and-braces runner guard).

What

A final turn with no function calls whose text cannot be extracted — toText() throws, or the response carries no text part (observed live: a hard agent task spends the entire max_tokens budget inside a thinking block, stop_reason: "max_tokens") — was swallowed into a "successful" generation with text: null. The agents runner then reported success with an empty answer, bypassing both the transient retry and every error surface.

openstation_ai_client_generate() now returns WP_Error( 'openstation_ai_empty_answer' ) for that case, with the original extraction detail preserved in the error data.

Consumers, walked

• Search agentic loop: the duplicate local ! is_string( $text ) check (the old openstation_ai_empty branch) is removed — the centralized error reaches the loop's existing is_wp_error handling at both generate sites. Net behavior change: an empty-string final turn now reports openstation_ai_empty_answer instead of falling through to openstation_ai_result_parse.
• Command follow-up: the deliberate degrade is preserved exactly as requested — this path matches the new code and still falls back to "Done." with the fallback flag, so a command that DID run never surfaces as a failure, and no openstation_ai_search_error fires for it (same as today).
• Agents runner: belt-and-braces guard in the toolless-turn branch (trim-empty text ⇒ error, never a success with ''), plus a humanized mapping (openstation_agent_empty_answer) telling the user the model likely spent its output budget reasoning. The new code is deliberately NOT in the transient-retry signature list — the failure is task-shaped, not a flap; happy to add it if you'd rather retry once.
• stop_reason isn't reachable from the SDK result surface this adapter touches, so per your note I didn't block on it — the provider's exception message lands in detail.

Tests

Three additions to Tests_OpenStation_AgentsRunner via the openstation_agent_runner_generate pre-filter seam: a text: null final turn is an error (not an empty success), whitespace-only text is the same failure, and the humanize mapping preserves the extraction detail. php -l clean on all four files.

Honest caveat: I couldn't run npm run test:php locally (no Docker on this machine) — the suite additions mirror the file's existing patterns and CI runs the same scripts.

Evidence for the mechanism (raw provider captures): #517

Follow-up issues for the other three threads from the maintainer reply (provider config filter, async runner shape, replayed placeholder) are being filed separately as requested.

🤖 Generated with Claude Code

[AllTerrainDeveloper]

There is a NIT that can be dismissed if there is no way of checking the error number :)

[AllTerrainDeveloper]

Hey! I'm wondering if this can be replaced with an error number check