Provider-agnostic context-limits via localModels.{chat,embedding} role-keyed split + retire SessionService hidden default

Premise

6 consumer sites hardcode aiConfig.openAiCompatible.contextLimitTokens / safeProcessingLimitTokens. The actual model-context-window axis is model-role (chat vs embedding), not provider-namespace. For deployments using ollama as graphProvider, friction telemetry reflects openAiCompatible's 262144 regardless of the actual Ollama context window → wrong substrate, wrong calibration.

Operator direction (synthesis): remote providers (gemini) are API-bound — no operator control over context cap, so no template keys needed. Local-model context caps belong on the model-role axis, shared across local providers — one config for chat, one for embedding. Provider-namespace duplication is over-the-top.

Substrate shape (cross-family V-B-A converged with @neo-gpt)

• Remote providers (gemini, future API-only): no contextLimitTokens in template
• Local-model-role-keyed: localModels.chat.{context,safe}LimitTokens + localModels.embedding.{context,safe}LimitTokens
• Namespace boundary (localModels.*) states the actual constraint: locally controlled context caps; top-level chat/embedding would be too close to provider-selector primitive

Surface enumeration (V-B-A 2026-05-27)

Template current state: aiConfig.openAiCompatible.contextLimitTokens=262144 + safeProcessingLimitTokens=200000. aiConfig.ollama namespace has NO context-limit keys.

Contract Ledger

Acceptance Criteria

• AC1: localModels.chat.contextLimitTokens=262144 + safeProcessingLimitTokens=200000 added to ai/config.template.mjs (preserves current shipped defaults per @neo-gpt V-B-A)
• AC2: localModels.embedding.{context,safe}LimitTokens added to ai/config.template.mjs with concrete defaults — value requires V-B-A per AC6
• AC3: 4 consumer sites migrated from aiConfig.openAiCompatible.{context,safe}* to aiConfig.localModels.{chat,embedding}.* based on consumer role
• AC4: SessionService.mjs:612 hidden default fallback (|| 32768) deleted; loud-fail on missing config (matches feedback_no_hidden_default_fallbacks operator contract)
• AC5: SemanticGraphExtractor.mjs:121 JSDoc rationale (openAiCompatible-as-conservative-baseline-for-both-families) rewritten to reflect role-keyed substrate; pre-existing ticket-anchor archaeology (#11965 AC5 / #12059) stripped per AGENTS_ATLAS.md §15.6
• AC6: Embedding context-cap concrete default value verified against the configured embedding model (LM Studio qwen3-8b-embedding OR Ollama equivalent) BEFORE implementing PR opens. Placeholder 8192/6144 acceptable in template if V-B-A defers to follow-up.
• AC7: Open question — aiConfig.openAiCompatible.contextLimitTokens legacy retention strategy (keep as deprecated passthrough vs delete in same PR). Operator/Stage-1 sequencing call.

Timing + scope

Recommendation (lean B-with-escape-hatch per @neo-gpt): fold into Epic #12101 Stage 1+ rewrite to avoid double migration. Stage 1 (#12103) establishes meta-leaf primitives; Stage 2 (#12104) rewrites template; this ticket's substrate semantic naturally rides on Stage 2's template rewrite. If operator requires the SessionService hidden-default-fallback retirement (AC4) before Stage 1/2 lands, carve a tiny interim PR with explicit retirement path documented.

Empirical anchor

PR #12113 (just-shipped) preserved the existing aiConfig.openAiCompatible.contextLimitTokens reference for interim consistency with pre-existing pattern. Operator surfaced the substrate-shape issue during cycle-1 V-B-A: "client wants to use ollama". This ticket retires those references as the next substrate-evolution step.

Related

• Parent: Epic #12101 (greenfield aiConfig substrate redesign — graduated from Discussion #12100)
• Sibling: #12104 (Stage 2 template rewrite — this ticket's substrate semantic rides on Stage 2)
• Sibling: #12090 (provider multi-model coexistence; orthogonal but adjacent)
• Sibling: #11976 (graphProvider routing — the consumer-side discriminator)
• PR #12113 (interim consistency baseline; this ticket retires the deferred reference)

Discipline anchor

@neo-gpt's V-B-A surfaced: chat default = preserve current 262144/200000 (already live after #12064). Embedding default = NOT memory-asserted — requires V-B-A against actual model before AC6 closure. SessionService.mjs:612 hidden default fallback retirement (AC4) = highest-value invariant per feedback_no_hidden_default_fallbacks operator contract (2026-05-27).