PR Review Summary

Status: Request Changes

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

• Decision: Request Changes
• Rationale: Request Changes because the root-cause model switch and schema-supplied REM paths are aligned with #13853, but the OpenAI-compatible fallback still preserves the LM Studio-rejected json_object mode that the close-target requires this PR to remove.

Peer-Review Opening: Grace, the strategic direction is right: switch the REM chat model to the MoE, make hidden reasoning a per-task lever, and move structured consumers onto schema-backed output. The remaining issue is narrow but blocking because it keeps the old provider failure mode alive.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #13853 issue body/ACs, current #13857 head/checks, changed-file list, ADR 0019/AiConfig authority via KB, prior benchmark memories for gemma-4-26b-a4b, and exact-head source for OpenAiCompatible, SessionService, SemanticGraphExtractor, Ollama, and KeepAlive.spec.mjs.
• Expected Solution Shape: Config should select the MoE for the OpenAI-compatible REM route, summary/extraction should pass schema + per-task reasoning_effort, and the provider boundary should emit json_schema when a schema exists while not sending LM Studio-rejected json_object on the default REM path. Existing Ollama/native parity remains sibling work (#13856/#13855), not folded into this close target.
• Patch Verdict: Partially matches. The schema-supplied path is implemented for summary and extraction, but ai/provider/OpenAiCompatible.mjs:113-117 still emits response_format: {type: 'json_object'} when JSON is requested without a schema, and the unit tests still assert that old behavior.
• Premise Coherence: coheres: verify-before-assert is reflected by measured model-choice evidence and focused tests; conflicts: the fallback branch contradicts the close-target evidence that json_object is rejected by the target endpoint, so the implementation preserves a known-bad path instead of converting the friction into a complete provider contract.

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #13853
• Related Graph Nodes: #12740, #13854, #13855, #13856, #13750, ADR 0019, OpenAI-compatible provider contract, REM summarization, tri-vector extraction

🔬 Depth Floor

Challenge OR documented search (per guide §7.1):

• Challenge: The patch correctly adds json_schema for callers with responseSchema, but it leaves the schema-less JSON fallback on json_object. Because buildChatModel forwards generation options through to the provider, any remaining responseMimeType: 'application/json' caller against LM Studio can still hit the rejection #13853 identified.

Rhetorical-Drift Audit (per guide §7.4):

Verify symmetry between stated framing and mechanical implementation:

• PR description: framing matches the model switch and schema-supplied REM wiring.
• Anchor & Echo summaries: config/caller comments mostly match the durable intent.
• [RETROSPECTIVE] tag: N/A.
• Linked anchors: #13853 establishes the model/schema contract.

Findings: Drift flagged with Required Actions: the body says the json_object rejection is fixed, but the fallback branch and tests preserve it.

🧠 Graph Ingestion Notes

• [KB_GAP]: None found; KB pointed to ADR 0019, SharedDeployment, and the relevant graph/session source surfaces.
• [TOOLING_GAP]: The formal review validator rejected my first draft until I used the exact template anchors; this is reviewer/tooling process friction, not a PR author issue.
• [RETROSPECTIVE]: Model-choice evidence is strong; provider fallback semantics need to be treated as part of the close-target contract, not as backward-compat background noise.

🎯 Close-Target Audit

For every issue named as close-target, verify it does NOT carry the epic label:

• Close-targets identified: #13853
• For each #N: confirmed not epic-labeled.

Findings: Close target valid, but not complete until the json_object fallback is removed/gated and CI is green.

📑 Contract Completeness Audit

• Originating ticket (or parent epic) contains a Contract Ledger matrix.
• Implemented PR diff matches the Contract Ledger exactly (no drift).

Findings: Contract drift flagged. #13853 says OpenAiCompatible.mjs should use json_schema with caller schema and retain a text fallback when no schema is supplied; this patch still emits json_object for schema-less JSON requests.

🪜 Evidence Audit

• PR body contains an Evidence: declaration line.
• Achieved evidence is stated as L2 + L3, with L4 live heavy-session REM digestion left as post-merge validation.
• If residuals exist: close-target issue body has the residuals annotated as [L<N>-deferred — operator handoff needed].
• Two-ceiling distinction: the body names manual local benchmark evidence and post-merge live verification separately.
• Evidence-class collapse check: review language does not promote L2/L3 to L4.

Findings: Evidence is acceptable for the model switch, but runtime gate remains red and RA1 means the structured-output AC is not complete.

🔌 Wire-Format Compatibility Audit

• PR alters native API wire format for OpenAI-compatible chat payloads.
• Schema-supplied path emits response_format: {type: 'json_schema', json_schema: ...}.
• Legacy JSON fallback no longer sends endpoint-rejected json_object on the target REM path.

Findings: Wire-format compatibility gap flagged in RA1/RA2.

N/A Audits — 📡 🔗

N/A across listed dimensions: no OpenAPI tool descriptions, skill files, startup rules, AGENTS substrate, or MCP tool surfaces changed.

🧪 Test-Execution & Location Audit

• Branch checked out locally via exact-head worktree at af9917a4b9ab38d8e386d58ba6f5b7dc88481759.
• Canonical Location: changed test remains in test/playwright/unit/ai/provider/KeepAlive.spec.mjs.
• If a test file changed: ran the specific test file.
• If code changed: verified related provider tests and current CI state.

Findings: Focused provider spec passes locally (npm run test-unit -- test/playwright/unit/ai/provider/KeepAlive.spec.mjs, 13/13). GitHub unit is currently red on run 27947584214; the failed job reports one failing test/playwright/unit/ai/graph/Database.spec.mjs test and two flaky retries elsewhere. The head is not mergeable until unit is green.

📋 Required Actions

To proceed with merging, please address the following:

• Remove or gate the surviving json_object fallback in ai/provider/OpenAiCompatible.mjs:113-117. The comment at lines 96-101 says schema-less JSON falls back to prompt-driven/no response_format, and #13853 says LM Studio rejects json_object; the implementation still sends payload.response_format = {type: 'json_object'}. The default REM/OpenAI-compatible route must not keep using the rejected mode.
• Update test/playwright/unit/ai/provider/KeepAlive.spec.mjs:374-384 and :527-542 so the tests stop locking json_object as the expected JSON fallback. The new json_schema test is useful, but the old assertions preserve the failure mode this PR is meant to retire.
• Get the GitHub unit check green on the current head before re-requesting review. If the current CI failure is unrelated baseline flake, rerun/rebase and keep that evidence in the review response.

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 82 - The lane targets the measured REM bottleneck and uses the AiConfig/provider boundary, but fallback semantics are still contract-drifting.
• [CONTENT_COMPLETENESS]: 68 - Main model/schema path is present; close-target fallback and green CI are incomplete.
• [EXECUTION_QUALITY]: 62 - Focused unit coverage exists and passes, but it also encodes the old json_object behavior as expected.
• [PRODUCTIVITY]: 78 - High-value load-reducer once the provider contract is closed.
• [IMPACT]: 86 - Addresses the REM digestion choke point behind the Golden Path freeze.
• [COMPLEXITY]: 72 - Crosses config, provider wire format, summary, extraction, and local-provider operational behavior.
• [EFFORT_PROFILE]: Heavy Lift - Root-cause model switch plus provider contract repair across runtime consumers.

Please re-request after the fallback branch and tests match the #13853 contract and unit CI is green.