Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
/**
* Helper agents on a model the person names: the Harness `subagent` tool with a forgiving model choice.
*
* The pinned Harness's `subagent` tool takes an exact `provider` route key and model id, and its discovery tool lists them as
* `api-1b2c3d4e-5f/deepseek/deepseek-v3.2`. Nobody (and no small local model) can guess that from "ask DeepSeek". This plugin keeps the
* Harness's tool and its checks and changes three things at the seam where the tool is called:
*
* - the choice: the model id, `provider/id`, the name in the model picker, or "the local model" (any letters) is turned into the route
* the Harness needs; a close guess is not started but answered with "Did you mean ...?" for the person to confirm; nothing that fits
* is answered with the plain list of models that can be used (server/agent/model-routes.mjs);
* - the words the assistant reads: the tool says when to use it ("ask a subagent", "ask model X") and that the model must never be looked for
* with shell commands or in the environment;
* - the discovery tool lists the models by the names the person knows.
*
* It rewrites no Harness file: the two tool definitions the Harness registers for each chat are wrapped in place (their `execute` and their
* text), once each. A Harness update that renames them leaves the tools as they were and the test that runs the real Harness fails.
*
* Plain ESM, loaded by absolute path like strata-gate.mjs. `config.routes` is the table the profile wrote (model-routes.mjs documents a route).
*/
import { modelProblem, resolveModel, routeList } from '../model-routes.mjs';
export const name = 'strata-models';
export const inject = ['tools'];
/** What the assistant is told in the tool's own text, so every model (a small local one and a hosted one alike) knows when and how. */
export const SUBAGENT_GUIDANCE = ' Use this tool whenever the person asks you to ask a subagent, a helper or another model ("ask a subagent", "ask DeepSeek", "use the local model", "ask model X"): start it with this tool and do not look for the model yourself with shell commands, in the environment or on this computer. To use a particular model, give `model` as its name or id (for example "DeepSeek V3.2" or "deepseek/deepseek-v3.2", or "local" for the model running on this computer); `provider` is optional. `list_subagent_models` shows what can be used. After starting a helper, do not wait with sleep or commands: end your turn and its answer arrives as a new message.';
const MODEL_WORDS = 'The model to use for the helper: its name as the person knows it ("DeepSeek V3.2"), its id ("deepseek/deepseek-v3.2") or "local" for the model running on this computer. Omit it to use this chat\'s model.';
const PROVIDER_WORDS = 'Optional: the connection the model belongs to ("OpenRouter"). Usually the model alone is enough.';
const wrapped = new WeakSet();
/** The two wrappers' behaviour, apart from the Harness objects (the tests call these directly). */
export function subagentArgs(routes, args) {
const result = resolveModel(routes, args ?? {});
if (result.status !== 'exact') return { ok: result.status === 'inherit', args };
const { provider, model } = result.route;
return { ok: true, args: { ...args, provider, model } };
}
export function apply(ctx, config = {}) {
const routes = Array.isArray(config.routes) ? config.routes : [];
const wrapSubagent = def => {
if (wrapped.has(def)) return;
wrapped.add(def);
const original = def.execute;
def.description = `${def.description}${SUBAGENT_GUIDANCE}`;
const props = def.parameters?.properties;
if (props?.model) props.model = { ...props.model, description: MODEL_WORDS };
if (props?.provider) props.provider = { ...props.provider, description: PROVIDER_WORDS };
def.execute = function execute(args, exec) {
const request = args && typeof args === 'object' ? args : {};
const problem = modelProblem(routes, request);
if (problem) throw new Error(problem);
const settled = subagentArgs(routes, request);
return original.call(this, settled.args, exec);
};
};
const wrapList = def => {
if (wrapped.has(def)) return;
wrapped.add(def);
def.description = 'List the models a helper agent can be started on, by the names the person knows. Call it with no arguments, then give `model` (a name or id from the list, or "local") to the subagent tool.';
def.execute = function execute() {
return routes.length
? `Models a helper can use:\n${routeList(routes)}\nStart one with the subagent tool: model = its name or id.`
: 'No other model is set up. A helper agent uses this chat\'s own model.';
};
};
const wrapAll = () => {
let agents;
try { agents = ctx.get('agents')?.list?.() ?? []; } catch { return; }
for (const agent of agents) {
try {
const sub = ctx.tools.get('subagent', agent);
if (sub) wrapSubagent(sub);
const list = ctx.tools.get('list_subagent_models', agent);
if (list) wrapList(list);
} catch { /* a tool that is not there yet is wrapped at the next change */ }
}
};
ctx.on('tools/change', wrapAll);
wrapAll();
}