benchmark-models: GeminiAdapter.parseStreamJson reads obj.text but current gemini CLI emits obj.content — empty runs row as $0 success #2159

Symptom

gstack-model-benchmark --models gemini returns output: "", tokens 0/0, costUsd 0 with NO error field — the row looks like a successful $0 run, silently mis-ranking gemini in any comparison table.

Root cause

test/helpers/providers/gemini.ts parseStreamJson reads assistant text via:

if (obj.type === 'message' && typeof obj.text === 'string') { output += obj.text; }

but current gemini CLI --output-format stream-json emits content (with role), not text:

{"type":"message","role":"assistant","content":"PONG","delta":true}
{"type":"result","status":"success","stats":{"input_tokens":19920,"output_tokens":32,...}}

Tokens are likewise under result.stats (input_tokens/output_tokens), not result.usage.input_token_count; the model name arrives on the init event.

Note test/helpers/gemini-session-runner.ts was already updated to read event.content (with the role === 'assistant' guard) — the benchmark provider adapter was missed.

Repro

gstack-model-benchmark --prompt 'Reply with exactly the word: PONG' --models gemini --output json on gemini CLI ≥ 0.x current (emits content-shape events). Verified on v1.58.5.0.

Suggested fix (verified locally)

Additive fallbacks in parseStreamJson: accept obj.content when obj.role === 'assistant' (the role guard matters — the CLI echoes the user prompt as a role:'user' message event, which would otherwise concat into output); obj.usage ?? obj.stats with input_tokens/output_tokens fallbacks; read modelUsed from init. With that patch the same probe returns output: 'PONG', tokens 24946/32.

Hardening ask

Independent of the field fix: an adapter run that produces EMPTY output with exit 0 should surface as an error row (or a loud flag), never a $0-success row — empty-success is indistinguishable from a healthy cheap run in the table.