Config Prompts

Prompts for configuration management: TypeScript config data files, environment variables, model routing, and feature flags.

Prompts

#	Prompt	Purpose
1	Update Model Registry	Add a new AI model or change tier assignments
2	Update Taxonomy Rules	Per-topic extraction guidance, vocabulary, and voice
3	Update Challenge Voice	Voice constitution, format, style, and difficulty rules
4	Update Seed Controls	Seeding pipeline mode, volume, and quality thresholds
5	Add Banned Patterns	Add CQ violation regex patterns

How does the configuration system work?

Configuration data lives in TypeScript data files within their respective packages (e.g., packages/ai/src/config/, packages/config/src/).
Each data file exports typed constants validated with Zod schemas at import time.
To modify configuration, edit the TypeScript file directly -- no sync step is needed.
Configuration is imported by TypeScript modules at runtime with full type safety.

What are the main config data files and where do they live?

packages/ai/src/config/categories.ts -- Topic categories and seeding prompts (44 CategorySpec definitions across 33 root categories).
packages/config/src/model-registry-data.ts -- AI model registry and tier routing (30 models across 7 providers).
packages/ai/src/config/challenge-voice.ts, challenge-style-rules.ts, challenge-style-voices.ts, challenge-difficulty.ts, challenge-banned-patterns.ts -- Voice, style, difficulty, quality rules.
packages/ai/src/config/taxonomy-rules-data.ts, taxonomy-voices-data.ts -- Per-topic guidance, vocabulary, voice.

Where are environment variables defined and how do I add a new one?

Central config: packages/config/src/index.ts -- all env vars are loaded and typed here.
All env vars loaded from .env.local at the monorepo root.
Example file: .env.example (must be kept in sync with actual usage).
To add a new var: define it in packages/config/src/index.ts, add to .env.example, document in APP-CONTROL.md.

What do the env:check scripts validate?

bun run env:check-example -- Ensures .env.example has entries for all used env vars.
bun run env:check-local -- Validates .env.local against .env.example (warns on missing vars).
bun run env:check-typos -- Checks for common env file typos (e.g., SUPABASE_ULR instead of SUPABASE_URL).
env:check-example runs in CI to enforce completeness.

Which env vars control AI budget and cost limits?

How does the three-tier model routing system work?

Three tiers: default (92% of calls, cost-efficient), mid (5%, balanced), high (1%, top-tier reasoning).
Tier-to-model mapping: packages/config/src/model-registry-data.ts.
Runtime: packages/ai/src/model-router.ts reads config with a 60-second cache from the ai_model_tier_config DB table.
Each model has a ModelAdapter in packages/ai/src/models/adapters/ for per-model prompt optimizations.
Current default mapping (fallback when DB unavailable): all tiers → gemini-3-flash-preview.
Thinking budgets: models that support extended thinking (e.g., Gemini 3 Flash) have configurable thinking token limits per adapter.
Registry: packages/config/src/model-registry.ts.

How do I add a new AI model to the registry?

Add a new entry to packages/config/src/model-registry-data.ts with: model_id, provider, status, input_price, output_price, deprecation_note.
Add the model identifier to the ai_model DB enum via migration: ALTER TYPE ai_model ADD VALUE IF NOT EXISTS '<model-id>'. The enum must include the model before it can be set in ai_model_tier_config.
Optionally create a ModelAdapter in packages/ai/src/models/adapters/.
Adapters added Feb 2026: gemini-3-flash-preview, gpt-5-nano, grok-4-1-fast-non-reasoning, MiniMax-M2.5.
Test: bun scripts/seed/llm-fact-quality-testing.ts --all --models <model_id> --limit 50.

How do I change which model is used for a specific tier?

DB override (production): Update the ai_model_tier_config table directly — this is the authoritative source at runtime. Change propagates within 60 seconds (DB cache refresh), no restart needed.
Code fallback: Edit DEFAULT_TIER_CONFIG_DATA in packages/config/src/model-registry-data.ts — this is the fallback used when the DB is unavailable.
The model must exist in both MODEL_REGISTRY_DATA (code) and the ai_model DB enum before it can be assigned to a tier.

What is Opus escalation and how is it controlled?

Allows routing to Opus (highest-tier model) for top-1% complex tasks.
Controlled by: OPUS_ESCALATION_ENABLED (default false) and OPUS_MAX_DAILY_CALLS (default 20).
When enabled, the model router may select Opus for high-tier requests.
Config: packages/config/src/index.ts.

Where are feature flags stored and how do I toggle one?

Database table accessed via packages/db/src/drizzle/feature-flags.ts.
Admin UI: admin.eko.day/feature-flags -- toggle switches with immediate effect.
Runtime checks throughout the codebase via feature flag queries.

What are the key operational feature flags?

EVERGREEN_ENABLED -- Master switch for evergreen fact generation (env var, not DB flag).
Feature flags in DB control user-facing features (specific flags depend on current implementation).
Admin dashboard provides toggle UI with immediate effect.
Note: some "flags" are env vars (like EVERGREEN_ENABLED), others are DB-stored runtime flags.