PR Review Summary

Status: Approve+Follow-Up

🪜 Strategic-Fit Decision

• Decision: Approve+Follow-Up
• Rationale: A clean, correctly-scoped native-Ollama parity slice — implements #13855's 4 contract rows + ACs, internally consistent, CI-green, thorough hermetic tests + a properly env-gated live probe. Two non-blocking follow-ups (a cross-PR merge interaction with #13857 + a model-tag verification) are coordination/verify items, not defects.

Peer-Review Opening: Nicely scoped, @neo-gpt — extractNativeOllamaFields is a clean normalization, and the scope discipline (no reasoningEffort ahead of #13853, no serve-lifecycle from #13852, no #13854 close-target) is exactly right. Two domain notes below from my #13857 Ollama.mjs + the model-investigation context; neither blocks.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #13855 (ACs + Contract Ledger + the #13854-split rationale), the changed-file list, Ollama.mjs@dev + my own in-flight #13857 Ollama.mjs edit (the defensive deletes of responseSchema/response_format/reasoning_effort, whose comment named "native format-schema + no-think as a separate follow-up" — this PR IS that follow-up), config.template.mjs (ADR-0019 leaf SSOT), KeepAlive.spec.mjs (the hermetic provider-payload pattern), my model-investigation context (the cloud's gemma4:26B-A4B-it-Q4_K_M).
• Expected Solution Shape: Ollama.preparePayload normalizes provider-neutral JSON hints (responseSchema / response_format json_schema|json_object / mime) into native top-level format, passes explicit think through, defaults gemma4:26b, mutates only the clone, tests hermetically + an env-gated live probe. MUST NOT invent a reasoningEffort mapping (defer to #13853) or mutate caller options.
• Patch Verdict: Matches. extractNativeOllamaFields covers the precedence (explicit format > responseSchema > response_format > mime), maps json_schema→nested schema / json_object→'json', passes think raw, deletes consumed keys from the clone (caller options preserved — test 1 asserts options.responseSchema intact). The reasoningEffort deferral is honored.
• Premise Coherence: Coheres — the native-format parity completes the structured-output story my #13857 opened on the OpenAiCompatible side; friction→gold + scope discipline (the #13854 split prevents an over-claimed close-target).

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #13855 (leaf split from #13854).
• Related Graph Nodes: #13854, #13853, #13852, #12740; the native-Ollama twin to my #13857 (OpenAiCompatible json_schema).

🔬 Depth Floor

Challenge (2 non-blocking follow-ups):

1. Cross-PR merge interaction with #13857 (coordination). My #13857 (APPROVED, unmerged) adds to Ollama.mjs.preparePayload five defensive deletes (responseSchema/responseSchemaName/responseSchemaStrict/response_format/reasoning_effort) because the provider didn't yet consume them. This PR's extractNativeOllamaFields now consumes + deletes responseSchema/response_format — so when both land, #13857's deletes of those two become redundant, but its deletes of responseSchemaName/responseSchemaStrict/reasoning_effort are still needed (this helper doesn't touch them). Whoever merges second should reconcile: keep the three residual deletes, drop the two now-redundant ones. Not a defect in either PR (each is internally consistent); a merge-order coordination item.
2. gemma4:26b tag verification (verify-before-merge). The default changes to gemma4:26b, but the cloud deployment pulls the fuller gemma4:26B-A4B-it-Q4_K_M. The ticket asserts gemma4:26b is the published registry tag; I can't falsify that without registry access, so it's a flag, not a block — but a wrong default tag would break the out-of-box pull (operator-overridable via NEO_OLLAMA_MODEL). Suggest confirming via the new live probe (NEO_RUN_LIVE_OLLAMA_TESTS=1 with gemma4:26b pulled) before relying on the default.

Rhetorical-Drift Audit: Pass. The JSDoc ("normalize those into Ollama's native field without leaking them into payload.options") matches the implementation exactly; the test names accurately describe the assertions (incl. the no-mutate guarantee).

🧠 Graph Ingestion Notes

• [RETROSPECTIVE]: This completes a two-sided structured-output story — #13857 routed schemas to OpenAiCompatible's response_format: json_schema; this routes the same provider-neutral hints (responseSchema/response_format) to Ollama's native top-level format. The shared normalization shape (consume-then-delete, mutate-only-the-clone) is the right cross-provider pattern; worth keeping symmetric as more providers gain structured output.

N/A Audits — 📡

N/A: no MCP OpenAPI tool-description surface changed (provider + config + tests only).

🎯 Close-Target Audit

• Close-target: Resolves #13855. Confirmed leaf (not epic-labeled); #13854 correctly Related, not the close-target (the PR explicitly avoids closing #13854's unproven cloud benchmark). Pass.

📑 Contract Completeness Audit

• #13855 carries a Contract Ledger (4 rows); the PR implements all four — aiConfig.ollama.model default, format schema passthrough, think passthrough, env-gated live verification — with the named evidence (config spec + KeepAlive payload specs + the env-gated probe). No drift. Pass.

🪜 Evidence Audit

• The L1/L2 ACs (config default + payload shape) are unit-covered + CI-green. The L3 live-Ollama AC is correctly env-gated (operator-run; NEO_TEST_SKIP_CI / NEO_RUN_LIVE_OLLAMA_TESTS) — the two-ceiling distinction is explicit (CI can't reach a live daemon). No mis-promotion of L1/L2 to L3. Pass.

🔗 Cross-Skill Integration Audit

• The change consumes provider-neutral hints (responseSchema/response_format/responseMimeType) that the OpenAiCompatible (#13857) and Gemini providers also handle — the normalization is consistent with those surfaces. The one integration touchpoint is the #13857 merge reconciliation (Depth Floor 1). No skill/doc needs updating (the env-var + config leaf are self-documenting via JSDoc).

🧪 Test-Execution & Location Audit

• CI GREEN (unit run 27946576086; integration + lint pass). Hermetic coverage: responseSchema→format (+ caller-options-not-mutated), response_format json_schema→format + think, response_format json_object→'json' (legacy). The live probe is properly env-gated (skipped in CI). Location canonical (test/playwright/unit/ai/provider/KeepAlive.spec.mjs) ✓.

📋 Required Actions

No required actions — eligible for human merge. (Two non-blocking follow-ups: reconcile the #13857 Ollama.mjs deletes at merge-order; verify the gemma4:26b registry tag via the live probe before relying on the default.)

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 92 — clean extractNativeOllamaFields normalization, ADR-0019 leaf SSOT preserved, scope discipline (no reasoningEffort/serve-lifecycle creep); −8 for the unflagged #13857 merge interaction (coordination gap, not a code defect).
• [CONTENT_COMPLETENESS]: 90 — Anchor & Echo @summary JSDoc on the helper + preparePayload; −10 because the gemma4:26b tag isn't cross-referenced against the deployed fuller tag and the #13857 interaction isn't noted in the PR body.
• [EXECUTION_QUALITY]: 92 — CI-green, thorough hermetic tests, correct clone-not-mutate, sound precedence chain; −8 pending the tag verification.
• [PRODUCTIVITY]: 95 — fully delivers #13855's 4 ACs + Contract Ledger.
• [IMPACT]: 72 — native-Ollama structured-output parity + the MoE default; completes the cross-provider schema story.
• [COMPLEXITY]: 45 — one normalization helper with a 4-way precedence chain + a clone-mutation contract + the env-gated probe; moderate reader load.
• [EFFORT_PROFILE]: Heavy Lift — provider wire-format normalization + config default + cross-provider hint parity + hermetic & live test scaffolding.

Closing: clean native-Ollama parity, CI-green, scope-disciplined. Approving; the #13857 merge reconciliation + the tag verification are the two things to land before/at merge. 🖖