feat(config): add strict_tools to send tool schemas in strict mode

[mudler-agent]

What changed

• strict_tools (top-level, default false): turns on cogito's EnableStrictToolSchemas (feat(tools): add EnableStrictToolSchemas and keep MCP additionalProperties cogito#81). Tools are sent with "strict": true and a schema that follows strict mode's rules: additionalProperties: false everywhere, every property required, formerly optional ones nullable. cogito drops the resulting nulls before it calls the tool.
• Named endpoints inherit it, like reasoning_effort, and can set strict_tools: false to opt out. /settings lists it.
• cogito bump to the feat(chat): add ephemeral todo_write tool for task tracking #81 commit. With or without strict_tools, MCP tools now carry the additionalProperties their server declares, so bash tells the model that unknown arguments are rejected.
• README config reference and the named-endpoint inheritance paragraph, synced to selfdoc/.

Why

Models called bash with command instead of script, and the MCP server rejected every call (unexpected additional properties ["command"]). On LocalAI with a model that uses the backend's chat template, llama.cpp's autoparser builds the tool-call grammar, and under that grammar argument names are literals from the schema, so command cannot be generated. The grammar is lazy under tool_choice: auto, though: it starts only when the model writes the format's trigger marker. A call written without the marker comes back from the autoparser as plain text, and LocalAI's Go-side text-parsing fallback (core/http/endpoints/openai/chat.go, "Content-based tool call fallback") turns it into a tool call without checking it against the schema. That is how {"command": ...} got through.

Correction: an earlier version of this description said strict mode is the fix for LocalAI. It is not the right lever there: a strict tool makes LocalAI send its own grammar, which replaces the autoparser's pipeline and can break tool calls on template-based models (LocalAI issue #10052). The LocalAI-side fix is to validate fallback-parsed tool calls against the schema (PR in progress on mudler/LocalAI). strict_tools stays useful for OpenAI and for servers without a native tool parser, which is why it is off by default. #129 (accepting command) covers the gap meanwhile.

Before merging

• go.mod pins cogito at bca971d from the feat(chat): add ephemeral todo_write tool for task tracking #81 branch. feat(chat): add ephemeral todo_write tool for task tracking #81 has merged as d35c7af; chore(deps): pin cogito to the merged strict-schemas commit on main #133 re-pins to it.
• Off by default on purpose. Some OpenAI-compatible servers reject strict. On LocalAI, a model using the backend's chat template may conflict with LocalAI's grammar (see the comment about issue #10052 in LocalAI's model_config.go). This needs a real run against a LocalAI model before recommending it.

How to verify

• go test ./... with an empty HOME. New tests:
  • TestStrictToolsSendsBashAsAStrictTool: the real bash server behind a fake OpenAI endpoint; the first request sends bash with strict: true, additionalProperties: false and timeout typed [integer, null].
  • TestStrictToolsOffByDefault: nothing is strict by default, and bash still carries additionalProperties: false.
  • TestConfigNamedEndpointInheritsStrictTools: the default endpoint, an inheriting endpoint and an opted-out one.
• By hand: set strict_tools: true on a LocalAI endpoint and check that tool calls still work, and that LocalAI's debug log shows grammar in use.

🤖 Generated with Claude Code