AI-powered virtual executive team — a single coherent executive persona backed by 8 specialist agents (FastAPI + Next.js).
# Honcho server config baked into the Open Executive Docker image.
#
# Env vars override every key in this file (e.g. `DB_CONNECTION_URI=...`
# wins over `[db].CONNECTION_URI`), so secrets and per-environment values
# come from environment variables at runtime — see README.md.
#
# Provider strategy:
# - Every LLM slot (deriver, summary, the dialectic levels, both dream
# passes) uses Honcho's `anthropic` transport with the same Claude
# models Open Executive itself defaults to, so `LLM_ANTHROPIC_API_KEY`
# is the only LLM secret the sidecar needs.
# - To route through OpenRouter (or any OpenAI-compatible endpoint)
# instead, override per slot from the environment; Honcho reads
# `<SLOT>_MODEL_CONFIG__TRANSPORT`, `<SLOT>_MODEL_CONFIG__MODEL` and
# `<SLOT>_MODEL_CONFIG__OVERRIDES__BASE_URL`, e.g.
# DERIVER_MODEL_CONFIG__TRANSPORT=openai
# DERIVER_MODEL_CONFIG__MODEL=anthropic/claude-haiku-4.5
# DIALECTIC_LEVELS__LOW__MODEL_CONFIG__TRANSPORT=openai
# DREAM_DEDUCTION_MODEL_CONFIG__TRANSPORT=openai
# LLM_OPENAI_API_KEY=<key> LLM_OPENAI_BASE_URL=https://openrouter.ai/api/v1
# A `base_url` written directly under `[…model_config]` is silently
# ignored (Honcho ≥ 3.1): in a file it belongs under
# `[…model_config.overrides]`.
# - Embeddings come from a local fastembed sidecar (see
# docker/honcho/embed_server.py) — no external embedding vendor.
# Quality ladder is documented in README.md ("Embedding quality").
[app]
LOG_LEVEL = "INFO"
# Embeddings ON. Honcho uses them for the dialectic chat's internal
# retrieval and the deriver's similarity-based fact dedup. Without them
# the knowledge-graph synthesis quality degrades meaningfully — links
# between related facts across sessions/peers stop being made.
EMBED_MESSAGES = true
[db]
# CONNECTION_URI is set per-environment via the DB_CONNECTION_URI
# variable (a managed Postgres usually injects DATABASE_URL; the README covers
# the postgres:// → postgresql+psycopg:// scheme rewrite). Pool sizing
# tuned for the 1-vCPU shared dev VM.
POOL_SIZE = 10
[auth]
# Auth-in-depth: even on a private network the API requires a
# valid JWT. JWT_SECRET comes from the environment; OE's HONCHO_API_KEY is
# a token derived from it via src.security.create_admin_jwt().
USE_AUTH = true
# -------------------------------------------------------------------- #
# LLM routing — Anthropic direct. Cheap-tier work (deriver, summary,
# low-tier dialectic) → Haiku for cost; medium+ dialectic and both dream
# passes → Sonnet for quality. Same model ids as Open Executive's own
# defaults (packages/core/openexecutive/config.py).
# -------------------------------------------------------------------- #
[deriver.model_config]
transport = "anthropic"
model = "claude-haiku-4-5"
[summary.model_config]
transport = "anthropic"
model = "claude-haiku-4-5"
[dialectic.levels.minimal.model_config]
transport = "anthropic"
model = "claude-haiku-4-5"
[dialectic.levels.low.model_config]
transport = "anthropic"
model = "claude-haiku-4-5"
[dialectic.levels.medium.model_config]
transport = "anthropic"
model = "claude-sonnet-5"
[dialectic.levels.high.model_config]
transport = "anthropic"
model = "claude-sonnet-5"
[dialectic.levels.max.model_config]
transport = "anthropic"
model = "claude-sonnet-5"
[dream]
# Dreaming ON. This is Honcho's offline reflection pass — it synthesizes
# higher-order patterns from accumulated peer data (e.g. "Alice is
# consistently anxious about cash flow", "the team disagrees on hiring
# velocity"). Without dreaming, peer cards are reactive snapshots; the
# knowledge-graph value largely lives in this layer. Cost: continuous
# background LLM spend — watch it before going to prod.
ENABLED = true
# Both dream passes. The table names are `deduction_model_config` and
# `induction_model_config` (Honcho ≥ 3.1); the older
# `[dream.deduction.model_config]` spelling is ignored, which left both
# passes on Honcho's default model name.
[dream.deduction_model_config]
transport = "anthropic"
model = "claude-sonnet-5"
[dream.induction_model_config]
transport = "anthropic"
model = "claude-sonnet-5"
# -------------------------------------------------------------------- #
# Embeddings — local fastembed sidecar (embed_server.py running as the
# `embed` process). The deriver hits it over the internal
# 6PN at `embed.process.<app>.internal:8001/v1`. The base_url is set
# via the EMBEDDING_MODEL_CONFIG__OVERRIDES__BASE_URL secret (Honcho's
# nested-pydantic-settings convention — env_prefix=EMBEDDING_,
# env_nested_delimiter=__), so the deploy can swap to OpenRouter
# without rebuilding the image. See README.md.
#
# IMPORTANT: bge-small-en-v1.5 emits 384-dim vectors but Honcho's
# schema defaults to Vector(1536). The runbook step "Configure vector
# dimensions" runs `scripts/configure_embeddings.py` to ALTER the
# column once per deploy that changes the embedding dim. Run it as a
# release step on every deploy
# (idempotent — no-ops when current dim matches).
# -------------------------------------------------------------------- #
[embedding.model_config]
transport = "openai"
model = "BAAI/bge-small-en-v1.5"
# overrides.base_url + overrides.api_key intentionally omitted —
# set via EMBEDDING_MODEL_CONFIG__OVERRIDES__BASE_URL and
# EMBEDDING_MODEL_CONFIG__OVERRIDES__API_KEY variables.