Skip to content

Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.

fix(agent-runtimes): offer Jarvis tools to OpenClaw directly

rubenluetke10-beep008558fef053
diff --git a/docs/agent-runtimes.md b/docs/agent-runtimes.md
index 58cacb8940..2c1a6d8972 100644
--- a/docs/agent-runtimes.md
+++ b/docs/agent-runtimes.md
@@ -54,8 +54,13 @@ replayed during `session/load` is swallowed because the chat already shows it.
`<VENDOR>_API_KEY` derived from the host), in that one process' environment
only. A keyless local server gets Hermes' `no-key-required` placeholder.
- **Tool search:** Hermes defers MCP schemas behind its own `tool_search` by
- default. Jarvis keeps that native behaviour; tests switch it off
- (`tools.tool_search: false`) so a scripted model sees the tools.
+ default; OpenClaw does the same for local models. Both are switched off for
+ agent profiles (`tools.tool_search: false`, `tools.toolSearch: false`): in a
+ live turn a 9B local model never found the deferred Jarvis tools, and with
+ them offered directly both runtimes called `society_wiki_note` correctly.
+ The price is prompt size — about 69k input tokens per model call instead of
+ about 27k on a fresh Hermes session (prompt caching covers roughly half of
+ it from the second call on).
## OpenClaw
@@ -100,6 +105,12 @@ Hermes 0.20.6 and OpenClaw 2026.9.8 on Windows: new session, an MCP tool call
with the right session header, resume in a new process, unknown-session
handling.
+Live, 2026-10-06, on a headless dev instance with a local Ollama model
+(`qwen3.5:9b`, no cost): a Hermes and an OpenClaw agent answered with the
+agent's Jarvis identity, read a fact from its Jarvis memory after switching
+runtimes, and wrote a new memory entry through `society_wiki_note`; killing
+the app reaped the OpenClaw Gateway.
+
`scripts/spikes/agent_runtimes_e2e.py hermes|openclaw` runs the same fake
model through Jarvis' own drivers (profile and config writing, process
environment, the OpenClaw Gateway supervisor): two turns in one conversation,
diff --git a/jarvis/agent_runtimes/openclaw.py b/jarvis/agent_runtimes/openclaw.py
index 84d6b6a77c..83af8b0417 100644
--- a/jarvis/agent_runtimes/openclaw.py
+++ b/jarvis/agent_runtimes/openclaw.py
@@ -302,7 +302,9 @@ class OpenClawRuntime:
}
},
"session": {"reset": {"mode": "none"}},
- "tools": {"deny": denied, "exec": {"mode": exec_mode}},
+ # Jarvis' tools offered directly, never behind OpenClaw's tool search
+ # (on by default for local models; smaller models miss deferred tools).
+ "tools": {"deny": denied, "exec": {"mode": exec_mode}, "toolSearch": False},
}
if turn.mcp_url and turn.control_key:
from jarvis.agent_chat.jarvis_harness import HEADER_NAME
diff --git a/tests/unit/agent_runtimes/test_runtime_configs.py b/tests/unit/agent_runtimes/test_runtime_configs.py
index 28cb76eb6b..49df9a787d 100644
--- a/tests/unit/agent_runtimes/test_runtime_configs.py
+++ b/tests/unit/agent_runtimes/test_runtime_configs.py
@@ -137,6 +137,7 @@ def test_openclaw_never_runs_background_turns_or_persona_files(tmp_path):
assert defaults["model"]["primary"] == "jarvis/gpt-5.2"
assert config["session"]["reset"]["mode"] == "none"
assert "automations" in config["tools"]["deny"]
+ assert config["tools"]["toolSearch"] is False
assert config["gateway"]["bind"] == "loopback"