Context
#13854 stayed open after two repo-local slices merged: #13853 introduced the shared per-task summaryReasoningEffort / graphReasoningEffort leaves, and #13856 added native Ollama model/schema/explicit-think payload support. A current-source check on origin/dev found the remaining repo-local gap: SessionService and SemanticGraphExtractor pass provider-neutral reasoning_effort, but ai/provider/Ollama.mjs deletes that field instead of translating the default no-think value for native Ollama.
Release classification: boardless — small provider parity slice under #13854; cloud benchmark evidence remains a separate runtime gate.
The Problem
The openAiCompatible path can consume the shared reasoning_effort: 'none' leaf added by #13853, but native Ollama does not. Without this mapping, the native Ollama path relies on callers manually passing think:false; the shared config leaves do not actually disable thinking for Ollama-backed session summary / tri-vector extraction.
Official Ollama API docs list think as the thinking-model switch and format as the native structured-output surface: https://github.com/ollama/ollama/blob/main/docs/api.md
The Architectural Reality
ai/services/memory-core/SessionService.mjs passes reasoning_effort: summaryReasoningEffort || undefined into provider generation.ai/services/graph/SemanticGraphExtractor.mjs passes reasoning_effort: graphReasoningEffort || undefined into provider generation.ai/provider/Ollama.mjs already promotes native think to a top-level /api/chat payload field and strips provider-neutral schema hints.- Current gap:
ai/provider/Ollama.mjs strips reasoning_effort without mapping none to native think:false.
The Fix
Map provider-neutral reasoning_effort: 'none' to native top-level think:false inside the Ollama payload normalizer, while preserving explicit caller think as the stronger override. Keep non-none reasoning_effort values out of Ollama's options bag until there is a verified native mapping for them.
Contract Ledger Matrix
| Target Surface |
Source of Authority |
Proposed Behavior |
Fallback |
Docs |
Evidence |
OllamaProvider.preparePayload(options.reasoning_effort) |
#13853 shared reasoning leaves + Ollama API think field |
reasoning_effort: 'none' -> top-level think:false; explicit think wins |
unset / non-none values are stripped, not leaked into options |
provider JSDoc |
focused unit spec |
Native Ollama /api/chat payload |
official Ollama API docs |
emit top-level think only when requested by caller or mapped no-think leaf |
no think field |
this ticket + provider JSDoc |
payload-shape unit spec |
Decision Record impact
aligned-with ADR 0019: callers read config leaves at use site; provider translates provider-neutral options at the provider boundary.
Acceptance Criteria
Out of Scope
- Cloud benchmark evidence for #13854.
- Native mappings for non-
none reasoning-effort values; those need separate Ollama behavior evidence.
- Serving lifecycle / context preload; covered by #13852, #13865, and #13867.
Avoided Traps
- Do not pass
reasoning_effort through to Ollama options; it is provider-neutral OpenAI-style vocabulary, not a documented Ollama option.
- Do not claim this closes #13854; the cloud benchmark remains open.
Related
Parent: #13854. Siblings: #13853, #13855, #13856, #13865, #13867.
Live latest-open sweep: checked latest 20 open issues at 2026-06-22T13:37Z; no equivalent leaf found. Closest active item is parent #13854. A2A sweep: checked latest 30 all-status messages at 2026-06-22T13:37Z; only matching active claim was this @neo-gpt lane claim.
Origin Session ID: b9a8f817-9a9e-4243-abfb-62e762a94964
Handoff Retrieval Hints: Ollama reasoning_effort none think false provider; #13854 #13853 #13856 native Ollama no-think.
Context
#13854 stayed open after two repo-local slices merged: #13853 introduced the shared per-task
summaryReasoningEffort/graphReasoningEffortleaves, and #13856 added native Ollama model/schema/explicit-thinkpayload support. A current-source check onorigin/devfound the remaining repo-local gap:SessionServiceandSemanticGraphExtractorpass provider-neutralreasoning_effort, butai/provider/Ollama.mjsdeletes that field instead of translating the default no-think value for native Ollama.Release classification: boardless — small provider parity slice under #13854; cloud benchmark evidence remains a separate runtime gate.
The Problem
The
openAiCompatiblepath can consume the sharedreasoning_effort: 'none'leaf added by #13853, but native Ollama does not. Without this mapping, the native Ollama path relies on callers manually passingthink:false; the shared config leaves do not actually disable thinking for Ollama-backed session summary / tri-vector extraction.Official Ollama API docs list
thinkas the thinking-model switch andformatas the native structured-output surface: https://github.com/ollama/ollama/blob/main/docs/api.mdThe Architectural Reality
ai/services/memory-core/SessionService.mjspassesreasoning_effort: summaryReasoningEffort || undefinedinto provider generation.ai/services/graph/SemanticGraphExtractor.mjspassesreasoning_effort: graphReasoningEffort || undefinedinto provider generation.ai/provider/Ollama.mjsalready promotes nativethinkto a top-level/api/chatpayload field and strips provider-neutral schema hints.ai/provider/Ollama.mjsstripsreasoning_effortwithout mappingnoneto nativethink:false.The Fix
Map provider-neutral
reasoning_effort: 'none'to native top-levelthink:falseinside the Ollama payload normalizer, while preserving explicit callerthinkas the stronger override. Keep non-nonereasoning_effortvalues out of Ollama'soptionsbag until there is a verified native mapping for them.Contract Ledger Matrix
OllamaProvider.preparePayload(options.reasoning_effort)thinkfieldreasoning_effort: 'none'-> top-levelthink:false; explicitthinkwinsnonevalues are stripped, not leaked intooptions/api/chatpayloadthinkonly when requested by caller or mapped no-think leafthinkfieldDecision Record impact
aligned-with ADR 0019: callers read config leaves at use site; provider translates provider-neutral options at the provider boundary.
Acceptance Criteria
OllamaProvider.preparePayload()mapsreasoning_effort: 'none'to top-levelthink:false.thinkis preserved over the mappedreasoning_effortfallback.reasoning_effortnever leaks into the native Ollamaoptionsbag.Out of Scope
nonereasoning-effort values; those need separate Ollama behavior evidence.Avoided Traps
reasoning_effortthrough to Ollamaoptions; it is provider-neutral OpenAI-style vocabulary, not a documented Ollama option.Related
Parent: #13854. Siblings: #13853, #13855, #13856, #13865, #13867.
Live latest-open sweep: checked latest 20 open issues at 2026-06-22T13:37Z; no equivalent leaf found. Closest active item is parent #13854. A2A sweep: checked latest 30 all-status messages at 2026-06-22T13:37Z; only matching active claim was this @neo-gpt lane claim.
Origin Session ID: b9a8f817-9a9e-4243-abfb-62e762a94964
Handoff Retrieval Hints:
Ollama reasoning_effort none think false provider;#13854 #13853 #13856 native Ollama no-think.