Skip to content

AI-powered virtual executive team — a single coherent executive persona backed by 8 specialist agents (FastAPI + Next.js).

Honcho config: default to Anthropic direct, fix ignored table keys

Rufus651d7510a415
diff --git a/docker/honcho/Dockerfile b/docker/honcho/Dockerfile
index b4a539bc..ecb8beec 100644
--- a/docker/honcho/Dockerfile
+++ b/docker/honcho/Dockerfile
@@ -2,7 +2,7 @@
#
# Mirrors plastic-labs/honcho's official Dockerfile (Python 3.13-slim,
# uv 0.9.24) but pins the Honcho source to a specific release tag, bakes
-# in OE's config.toml (OpenRouter routes every LLM call, dreaming on,
+# in OE's config.toml (Anthropic direct for every LLM slot, dreaming on,
# AUTH_USE_AUTH=true), and bundles an embeddings sidecar
# (`docker/honcho/embed_server.py`) that runs as the `embed` process
# serving `BAAI/bge-small-en-v1.5` locally via fastembed/ONNX so
diff --git a/docker/honcho/README.md b/docker/honcho/README.md
index 7f5e7722..7ecbfd47 100644
--- a/docker/honcho/README.md
+++ b/docker/honcho/README.md
@@ -44,7 +44,7 @@ deriver fail every embed call with a dimension mismatch.
|---|---|
| `DATABASE_URL` | Postgres connection string |
| `AUTH_JWT_SECRET` | Every issued `HONCHO_API_KEY` derives from this — store it in a password manager; rotating it invalidates all keys |
-| `LLM_OPENAI_API_KEY` | Honcho's `openai` transport accepts any OpenAI-format endpoint, so this can be an OpenRouter key with `base_url` set per slot under `[…model_config.overrides]` in `config.toml` |
+| `LLM_ANTHROPIC_API_KEY` | Every LLM slot in `config.toml` uses Honcho's `anthropic` transport. To route through OpenRouter or another OpenAI-compatible endpoint instead, set `LLM_OPENAI_API_KEY`, `LLM_OPENAI_BASE_URL` and the per-slot `<SLOT>_MODEL_CONFIG__TRANSPORT=openai` / `__MODEL=` overrides listed at the top of `config.toml` |
| `EMBEDDING_MODEL_CONFIG__OVERRIDES__BASE_URL` | Point at the `embed` process, e.g. `http://embed:8001/v1` |
| `EMBEDDING_MODEL_CONFIG__OVERRIDES__API_KEY` | The sidecar needs no auth, but the OpenAI client requires some string |
| `EMBEDDING_VECTOR_DIMENSIONS` | `384` for `bge-small-en-v1.5`. Honcho's default schema is `Vector(1536)` |
diff --git a/docker/honcho/config.toml b/docker/honcho/config.toml
index f56d432a..a59cd1a5 100644
--- a/docker/honcho/config.toml
+++ b/docker/honcho/config.toml
@@ -5,16 +5,25 @@
# come from environment variables at runtime — see README.md.
#
# Provider strategy:
-# - All LLM calls (deriver, summary, dialectic levels, dream deduction)
-# use Honcho's `openai` transport but point at OpenRouter. OpenRouter
-# fronts Anthropic / OpenAI / Voyage / etc., so a single secret
-# (`LLM_OPENAI_API_KEY` = OE's existing OpenRouter key) drives every
-# outbound LLM call. The `model` field uses OpenRouter slugs
-# (`anthropic/claude-haiku-4.5`, etc.).
+# - 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 + the OpenRouter fallback are documented in
-# README.md ("Embedding quality ladder").
+# Quality ladder is documented in README.md ("Embedding quality").
[app]
LOG_LEVEL = "INFO"
@@ -38,56 +47,39 @@ POOL_SIZE = 10
USE_AUTH = true
# -------------------------------------------------------------------- #
-# LLM routing — everything via OpenRouter using the openai transport.
-# The endpoint goes under each slot's `[…model_config.overrides]`: Honcho
-# ≥ 3.1 reads `base_url` only from there (a `base_url` directly under
-# `model_config` is silently ignored and calls fall through to
-# api.openai.com).
-# Cheap-tier work (deriver, summary, low-tier dialectic) → Haiku for
-# cost; medium+ dialectic and both dream passes → Sonnet for quality.
+# 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 = "openai"
-model = "anthropic/claude-haiku-4.5"
-[deriver.model_config.overrides]
-base_url = "https://openrouter.ai/api/v1"
+transport = "anthropic"
+model = "claude-haiku-4-5"
[summary.model_config]
-transport = "openai"
-model = "anthropic/claude-haiku-4.5"
-[summary.model_config.overrides]
-base_url = "https://openrouter.ai/api/v1"
+transport = "anthropic"
+model = "claude-haiku-4-5"
[dialectic.levels.minimal.model_config]
-transport = "openai"
-model = "anthropic/claude-haiku-4.5"
-[dialectic.levels.minimal.model_config.overrides]
-base_url = "https://openrouter.ai/api/v1"
+transport = "anthropic"
+model = "claude-haiku-4-5"
[dialectic.levels.low.model_config]
-transport = "openai"
-model = "anthropic/claude-haiku-4.5"
-[dialectic.levels.low.model_config.overrides]
-base_url = "https://openrouter.ai/api/v1"
+transport = "anthropic"
+model = "claude-haiku-4-5"
[dialectic.levels.medium.model_config]
-transport = "openai"
-model = "anthropic/claude-sonnet-4.6"
-[dialectic.levels.medium.model_config.overrides]
-base_url = "https://openrouter.ai/api/v1"
+transport = "anthropic"
+model = "claude-sonnet-5"
[dialectic.levels.high.model_config]
-transport = "openai"
-model = "anthropic/claude-sonnet-4.6"
-[dialectic.levels.high.model_config.overrides]
-base_url = "https://openrouter.ai/api/v1"
+transport = "anthropic"
+model = "claude-sonnet-5"
[dialectic.levels.max.model_config]
-transport = "openai"
-model = "anthropic/claude-sonnet-4.6"
-[dialectic.levels.max.model_config.overrides]
-base_url = "https://openrouter.ai/api/v1"
+transport = "anthropic"
+model = "claude-sonnet-5"
[dream]
# Dreaming ON. This is Honcho's offline reflection pass — it synthesizes
@@ -95,25 +87,20 @@ base_url = "https://openrouter.ai/api/v1"
# 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 on OpenRouter — see the runbook's cost-monitoring
-# section before going to prod.
+# 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 OpenAI model name.
+# passes on Honcho's default model name.
[dream.deduction_model_config]
-transport = "openai"
-model = "anthropic/claude-sonnet-4.6"
-[dream.deduction_model_config.overrides]
-base_url = "https://openrouter.ai/api/v1"
+transport = "anthropic"
+model = "claude-sonnet-5"
[dream.induction_model_config]
-transport = "openai"
-model = "anthropic/claude-sonnet-4.6"
-[dream.induction_model_config.overrides]
-base_url = "https://openrouter.ai/api/v1"
+transport = "anthropic"
+model = "claude-sonnet-5"
# -------------------------------------------------------------------- #
# Embeddings — local fastembed sidecar (embed_server.py running as the