Ollama native Gemma 4 MoE format schema slice

Context

This leaf exists because #13854 is intentionally broader than the repo-local change already ready to ship. The parent ticket also contains #13853-dependent reasoningEffort work and a cloud benchmark, so a PR that only changes native Ollama defaults / structured-output payload shape must not use #13854 as its close target.

Release classification: post-release (repo-local provider parity slice; not release-blocking).

Live latest-open sweep: checked latest 20 open issues at 2026-06-22T10:31:31Z. Closest related tickets were #13854 (broader Ollama/cloud parity parent), #13853 (OpenAI-compatible / LM Studio reasoning-effort sibling), and #13852 (Ollama serving lifecycle). No equivalent closeable leaf for native Ollama format schema passthrough + gemma4:26b default was present.

A2A in-flight claim sweep: checked latest 30 A2A messages at 2026-06-22T10:31:31Z; only overlapping claim was @neo-gpt's own [lane-claim][#13854] Ollama MoE default + native format schema slice.

The Problem

Native Ollama dispatch still defaults NEO_OLLAMA_MODEL / aiConfig.ollama.model to the dense gemma4:31b model while Ollama publishes the 26B A4B MoE under the registry tag gemma4:26b. The provider also only maps JSON-output intent to format: "json"; callers that can provide a JSON schema have no native Ollama format: <schema> path and schema hints leak into payload.options instead of the top-level /api/chat request.

The Architectural Reality

• ai/config.template.mjs is the canonical Tier-1 config template. ADR 0019 requires preserving the leaf(default, env, type) SSOT shape instead of adding parallel env parsing.
• test/playwright/fixtures/aiConfigDefaults.mjs and test/playwright/unit/ai/config.template.spec.mjs mirror the Tier-1 defaults for deterministic assertions.
• ai/provider/Ollama.mjs owns native Ollama /api/chat payload construction. Ollama's documented API accepts format as either "json" or a JSON schema object, and accepts think as a top-level request field for supported models.
• test/playwright/unit/ai/provider/KeepAlive.spec.mjs already covers native provider payload shape without requiring a live daemon.

The Fix

1. Change the native Ollama chat-model default from gemma4:31b to gemma4:26b while keeping NEO_OLLAMA_MODEL as the override leaf.
2. Normalize provider-neutral schema hints into native Ollama top-level format:
   • responseSchema / response_schema -> schema object;
   • OpenAI-compatible response_format: {type: "json_schema"} -> nested schema object;
   • OpenAI-compatible response_format: {type: "json_object"} and JSON mime hints -> "json".
3. Promote explicit caller-provided think to the native top-level field without inventing a reasoningEffort mapping before #13853 lands.
4. Add hermetic unit tests plus an env-gated live Ollama probe that is skipped when NEO_TEST_SKIP_CI is set and only runs when NEO_RUN_LIVE_OLLAMA_TESTS=1 is provided.

Contract Ledger Matrix

Decision Record impact

aligned-with ADR 0019. The change updates the existing config leaf default and does not add a parallel provider alias, formula, or env parser.

Acceptance Criteria

• Native Ollama default model is gemma4:26b while NEO_OLLAMA_MODEL remains the override.
• Native Ollama provider emits JSON schemas via top-level format instead of leaking schema hints into payload.options.
• Existing plain JSON extraction still emits format: "json".
• Explicit think passes through top-level when provided, without mapping unresolved reasoningEffort leaves.
• Focused unit coverage verifies the config and payload contracts.
• Manual live Ollama verification is available through env vars without making CI depend on Ollama.

Out of Scope

• Cloud benchmark and production deployment measurement; remains on #13854.
• #13853 reasoningEffort config leaves and OpenAI-compatible / LM Studio path.
• Ollama serve lifecycle, preload, and context-window residency; owned by #13852.
• Pulling gemma4:26b during tests. Operators should run ollama pull gemma4:26b before the manual live probe.

Avoided Traps

• Do not close #13854 with only repo-local L1/L2 evidence; its cloud benchmark remains unproven.
• Do not invent a reasoningEffort config surface ahead of #13853.
• Do not make CI perform live local-model inference or pull large Ollama artifacts.

Related

Related: #13854 Related: #13853 Related: #13852 Related: #12740

Origin Session ID: b9a8f817-9a9e-4243-abfb-62e762a94964

Handoff Retrieval Hints: Ollama native format schema gemma4:26b NEO_TEST_SKIP_CI #13854