Patch Ollama.stream() to inject top-level keep_alive (mirror Ollama.generate() line 108)

Parent context

Follow-up surfaced by @neo-gpt PR #12076 cycle-2 review (Epic #12065 Sub 8 / #12074). The keep-alive-probe.mjs benchmark cannot characterize native Ollama keep_alive control because of an SDK-level payload-shape gap; this ticket owns the SDK fix.

V-B-A finding (banked from PR #12076 cycle-2)

ai/provider/Ollama.mjs:

Native Ollama /api/chat reads keep_alive at TOP LEVEL of the request body. The nested payload.options.keep_alive is silently ignored by Ollama's REST API. This makes caller-supplied keep_alive via provider.stream() unobservable on the native Ollama path.

Prescription

Mirror Ollama.generate() line 108's pattern in Ollama.stream():

async *stream(input, options = {}) {
    const payload = this.preparePayload(input, options, true);
    // Promote top-level keep_alive (mirror of generate() line 108) so streaming
    // callers can control KV-cache retention. Without this, keep_alive nested
    // under payload.options is silently ignored by Ollama's /api/chat.
    if (options.keep_alive !== undefined) {
        payload.keep_alive = options.keep_alive;
        if (payload.options) delete payload.options.keep_alive;
    }
    // ... existing fetch logic
}

(Final shape may also extract keep_alive from payload.options to remove the nested duplicate.)

Acceptance Criteria

• AC1: Ollama.stream() promotes caller-supplied keep_alive from nested payload.options to top-level payload.keep_alive before serialization
• AC2: Unit test covering top-level keep_alive injection (mock fetch verifies payload.keep_alive is set when caller passes options.keep_alive)
• AC3: ai/scripts/benchmark/keep-alive-probe.mjs Ollama early-exit (added in PR #12076 cycle-2) is REVERTED — probe can now characterize native Ollama keep_alive control empirically
• AC4: learn/agentos/gemma4-rem-benchmark.md "Substrate V-B-A findings" table row for Ollama.mjs:91-92 updated to reflect the fix
• AC5: Verify OpenAiCompatible.stream() does NOT have the same nesting bug — current cycle-2 V-B-A shows OpenAiCompatible.preparePayload() line 105 already uses Object.assign(payload, clonedOptions) so top-level IS the natural shape. AC5 may close as "no-op confirmed" once verified

Avoided Traps

• Don't expand scope to a generic provider-options-promotion pattern — keep_alive is the documented Ollama-specific request-body field; other payload.options.* fields are correctly Ollama's options model-parameters object (temperature, top_p, etc.)
• Don't change Ollama.generate() line 108 — it already does the right thing; this ticket only adds parity to the stream() path
• Don't bundle with broader provider refactor — Epic #12065 has Sub 3 (executeRemCycle) which may touch provider lifecycle; this ticket is a focused SDK fix

Related

• Epic #12065 — Orchestrator-as-SSOT REM pipeline
• #12074 / PR #12076 — gemma4 benchmark Sub 8 (this ticket unblocks AC3 native-Ollama keep_alive characterization)
• @neo-gpt cycle-2 review on PR #12076 (2026-05-27) — original V-B-A surface
• learn/agentos/gemma4-rem-benchmark.md — substrate V-B-A findings table