Settings reference

Production default (2026+): query router + route & symbol registries + knowledge lookup + EOP on; optional graph executor and intent classifier for staging. LLM inference for evidence selection is off by default (last resort only). api_surface is a legacy rollback bridge — not the primary path. After enabling EOP, re-index and run B2 cold (symbol definition, no open file) before sign-off.

Rollout scenarios

Dependencies

• Structural questions use index-time knowledge objects first (route registry), then search + expand (windows, grep, graph neighbours) — not repo-specific path injection.
• plan_selector_enabled is the query router master switch; keep route_registry_enabled on for RouteObject lookup.
• knowledge_object_lookup_enabled requires plan_selector_enabled and registry hits.
• evidence_obligation_platform_enabled requires symbol_registry_enabled for DEFINITION cold answers; re-index after enabling.
• evidence_receipt_eop_fields_enabled is useful only when EOP is on — exports obligation audit fields on Context Receipt.
• obligation_slot_ratios apply inside resolve_rag_budget when EOP is enabled — tune together with max_context_tokens.
• retrieval.discovery_policy is read per primary obligation type (overview, definition, registration, semantic).
• plan_executor_enabled is for optional graph reconstruct when lookup misses — off by default in production.
• retrieval.inference_enabled (server defaults) gates LLM evidence selection — last resort only, off by default.
• retrieval.trace_enabled collects decision trace per repo-lane turn; trace_export_enabled writes JSON to retrieval.trace_export_dir (operator-tunable, not hardcoded).
• Set retrieval.trace_include_query false before sharing trace exports; cap fields with retrieval.trace_max_field_chars.
• retrieval.live_progress_enabled emits pipeline_progress SSE for the VS Code live timeline — off until dogfood; no separate extension toggle required.
• inline.pause_mode is the pause SSOT (off | shared_runtime | on_embed_busy); inline.pause_while_indexing is a deprecated bool mirror.
• Extension contextmint.inline.pauseWhileIndexing can only tighten to on_embed_busy — it cannot weaken a stricter server pause_mode.
• api_surface_enabled is a legacy rollback bridge — not the primary production path.
• runtime.ollama_ensure_model_loaded preloads cold models before chat; runtime.ollama_keep_model_loaded prevents eviction on sovereign servers.
• server.ollama_warmup_read_timeout_sec caps preload wait — separate from server.ollama_chat_read_timeout_sec stream read.
• Server-only keys are not in VS Code Settings — use Engine Ollama panel, Server config tab, YAML overlay, or env vars, then restart the API.

How to check

• Ask how an API route is wired — the answer should trace handler → router → application entry from registry facts or expanded excerpts.
• B2 cold: ask where a PascalCase symbol is defined with no file open — answer should cite the definition site; receipt should show selection_path R1 and source=registry.
• B1 overview: receipt should show pack_seed coverage and obligations_requested including overview when EOP is on.
• With retrieval.live_progress_enabled=true, a slow retrieval turn should show pipeline phases in the VS Code timeline before the first answer token.
• On a shared Ollama host, set inline.pause_mode=shared_runtime and confirm ghost-text pauses only while embed/reindex is busy — not on every idle status poll.
• Retry the same question after re-indexing if route or symbol facts look stale.
• If quality drops, disable the newest flag you turned on (or roll back to EOP off / search-only) and re-index the workspace.
• Confirm effective flags after restart — a stale ~/.contextmint/server.defaults.yaml overlay can override repo defaults.
• Use Engine → observability or your metrics endpoint to watch structural query volume after changes.

Controls citation-bound LLM narration on sanitized excerpts, clarification recovery when evidence is insufficient, and the final honest stop after max clarification turns. See master plan §6.1 in documentation/CONTEXTMINT-EVIDENCE-NARRATION-MASTER-PLAN.md.

Edit in Engine → Server config under Evidence narration & clarification.

Decision audit for repo-lane structural turns — ordered steps with inputs, outputs, and reasons. Export writes retrieval-trace-*.json under the operator-configured folder (retrieval.trace_export_dir, default documentation/evidence). Path is never hardcoded in product code. Pair with context-receipt-*.json from the same folder. See master plan §6.1b (Retrieval & evidence decision trace) in documentation/CONTEXTMINT-EVIDENCE-NARRATION-MASTER-PLAN.md.

Edit in Engine → Server config under Retrieval, expand & infer, or in ~/.contextmint/server.defaults.yaml.

Optional real-time pipeline_progress SSE events during chat retrieval so the VS Code extension live timeline can show what the server is doing while the user waits. Off by default until RT-VIEW dogfood. Enable retrieval.live_progress_enabled on the API host — the extension renders events when the server emits them (no separate extension toggle). See documentation/REALTIME-SERVER-VIEW-EXTENSION-PLAN.md and documentation/IC-PAUSE-UX-RT-VIEW-ARCHITECTURE-PLAN.md.

Edit in Engine → Server config under Retrieval, expand & infer.