feat(routing): honor models.for via metadata modelFor
This commit is contained in:
@@ -3660,6 +3660,18 @@
|
|||||||
"docs/plans/state.json"
|
"docs/plans/state.json"
|
||||||
],
|
],
|
||||||
"test_status": "pnpm test:run src/channels/line/adapter.test.ts src/channels/zalo/adapter.test.ts + pnpm typecheck passing"
|
"test_status": "pnpm test:run src/channels/line/adapter.test.ts src/channels/zalo/adapter.test.ts + pnpm typecheck passing"
|
||||||
|
},
|
||||||
|
"model-for-tier-routing-consumption": {
|
||||||
|
"status": "completed",
|
||||||
|
"date": "2026-02-17",
|
||||||
|
"updated": "2026-02-17",
|
||||||
|
"summary": "Implemented runtime consumption of `models.*.for` via `metadata.modelFor` tier selection in channel routing (including audio capability resolution path), and added routing precedence/behavior regression coverage.",
|
||||||
|
"files_modified": [
|
||||||
|
"src/daemon/routing.ts",
|
||||||
|
"src/daemon/routing.test.ts",
|
||||||
|
"docs/plans/state.json"
|
||||||
|
],
|
||||||
|
"test_status": "pnpm test:run src/daemon/routing.test.ts + pnpm typecheck passing"
|
||||||
}
|
}
|
||||||
},
|
},
|
||||||
"overall_progress": {
|
"overall_progress": {
|
||||||
|
|||||||
@@ -51,28 +51,96 @@ describe('daemon agent routing integration', () => {
|
|||||||
expect(router.resolve('telegram', '123')).toBeUndefined();
|
expect(router.resolve('telegram', '123')).toBeUndefined();
|
||||||
});
|
});
|
||||||
|
|
||||||
it('model tier precedence: metadata > agent config > global default', () => {
|
it('model tier precedence: metadata > metadata modelFor > agent config > global default', () => {
|
||||||
// This test documents the tier resolution precedence used by createMessageRouter.
|
// This test documents the tier resolution precedence used by createMessageRouter.
|
||||||
// The actual resolution logic: tierFromMetadata ?? agentConfig?.modelTier ?? primary_tier ?? 'default'
|
// The actual resolution logic:
|
||||||
|
// tierFromMetadata ?? tierFromMetadataModelFor ?? agentConfig?.modelTier ?? primary_tier ?? 'default'
|
||||||
function resolveTier(
|
function resolveTier(
|
||||||
metadataTier: ModelTier | undefined,
|
metadataTier: ModelTier | undefined,
|
||||||
|
metadataForTier: ModelTier | undefined,
|
||||||
agentTier: ModelTier | undefined,
|
agentTier: ModelTier | undefined,
|
||||||
globalTier: ModelTier | undefined,
|
globalTier: ModelTier | undefined,
|
||||||
): ModelTier {
|
): ModelTier {
|
||||||
return metadataTier ?? agentTier ?? globalTier ?? 'default';
|
return metadataTier ?? metadataForTier ?? agentTier ?? globalTier ?? 'default';
|
||||||
}
|
}
|
||||||
|
|
||||||
// With all three set, metadata wins
|
// With all three set, metadata wins
|
||||||
expect(resolveTier('fast', 'complex', 'default')).toBe('fast');
|
expect(resolveTier('fast', 'default', 'complex', 'default')).toBe('fast');
|
||||||
|
|
||||||
|
// Without explicit metadata tier, modelFor-resolved tier wins
|
||||||
|
expect(resolveTier(undefined, 'complex', 'default', 'fast')).toBe('complex');
|
||||||
|
|
||||||
// Without metadata, agent config wins
|
// Without metadata, agent config wins
|
||||||
expect(resolveTier(undefined, 'complex', 'default')).toBe('complex');
|
expect(resolveTier(undefined, undefined, 'complex', 'default')).toBe('complex');
|
||||||
|
|
||||||
// Without metadata or agent config, global wins
|
// Without metadata or agent config, global wins
|
||||||
expect(resolveTier(undefined, undefined, 'default')).toBe('default');
|
expect(resolveTier(undefined, undefined, undefined, 'default')).toBe('default');
|
||||||
|
|
||||||
// Without anything, falls back to 'default'
|
// Without anything, falls back to 'default'
|
||||||
expect(resolveTier(undefined, undefined, undefined)).toBe('default');
|
expect(resolveTier(undefined, undefined, undefined, undefined)).toBe('default');
|
||||||
|
});
|
||||||
|
|
||||||
|
it('uses metadata.modelFor tags to select tier', async () => {
|
||||||
|
const processSpy = vi.spyOn(AgentOrchestrator.prototype, 'process').mockResolvedValue('ok');
|
||||||
|
const session = {
|
||||||
|
id: 'telegram:model-for',
|
||||||
|
addMessage: vi.fn(),
|
||||||
|
getHistory: vi.fn(() => []),
|
||||||
|
clear: vi.fn(),
|
||||||
|
replaceHistory: vi.fn(),
|
||||||
|
getConfig: vi.fn(() => undefined),
|
||||||
|
setConfig: vi.fn(),
|
||||||
|
deleteConfig: vi.fn(),
|
||||||
|
};
|
||||||
|
|
||||||
|
const router = createMessageRouter({
|
||||||
|
sessionManager: {
|
||||||
|
getSession: vi.fn(() => session),
|
||||||
|
} as unknown as MessageRouterDeps['sessionManager'],
|
||||||
|
modelRouter: {
|
||||||
|
getAvailableTiers: () => ['fast', 'default', 'complex', 'local'],
|
||||||
|
getAllLabels: () => ({ fast: 'fast', default: 'default', complex: 'complex', local: 'local' }),
|
||||||
|
getLabel: (tier: string) => tier,
|
||||||
|
} as unknown as MessageRouterDeps['modelRouter'],
|
||||||
|
systemPrompt: 'test prompt',
|
||||||
|
toolRegistry: {
|
||||||
|
clone() { return this; },
|
||||||
|
register: vi.fn(),
|
||||||
|
} as unknown as MessageRouterDeps['toolRegistry'],
|
||||||
|
toolExecutor: {} as unknown as MessageRouterDeps['toolExecutor'],
|
||||||
|
config: {
|
||||||
|
agents: {
|
||||||
|
primary_tier: 'default',
|
||||||
|
delegation: {
|
||||||
|
compaction: 'fast',
|
||||||
|
memory_extraction: 'fast',
|
||||||
|
classification: 'fast',
|
||||||
|
tool_summarisation: 'fast',
|
||||||
|
complex_reasoning: 'complex',
|
||||||
|
},
|
||||||
|
max_delegation_depth: 3,
|
||||||
|
max_iterations: 10,
|
||||||
|
},
|
||||||
|
compaction: { enabled: false },
|
||||||
|
models: {
|
||||||
|
default: { provider: 'anthropic', model: 'claude', for: ['chat'] },
|
||||||
|
fast: { provider: 'anthropic', model: 'haiku', for: ['search'] },
|
||||||
|
},
|
||||||
|
} as unknown as MessageRouterDeps['config'],
|
||||||
|
});
|
||||||
|
|
||||||
|
await router.handler({
|
||||||
|
id: 'm-model-for',
|
||||||
|
channel: 'telegram',
|
||||||
|
senderId: 'model-for',
|
||||||
|
text: 'find this quickly',
|
||||||
|
timestamp: Date.now(),
|
||||||
|
metadata: { modelFor: 'search' },
|
||||||
|
} as MessageRouterInput, vi.fn(async () => {}));
|
||||||
|
|
||||||
|
const keys = Array.from(router.agents.keys());
|
||||||
|
expect(keys.some((key) => key.endsWith(':fast'))).toBe(true);
|
||||||
|
expect(processSpy).toHaveBeenCalled();
|
||||||
});
|
});
|
||||||
});
|
});
|
||||||
|
|
||||||
|
|||||||
+34
-1
@@ -44,6 +44,34 @@ function buildProviderConfigMap(config: Config): Partial<Record<ModelProvider, M
|
|||||||
return providerConfigs;
|
return providerConfigs;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
function tierFromUseCase(config: Config, useCaseRaw: unknown): ModelTier | undefined {
|
||||||
|
if (typeof useCaseRaw !== 'string') {
|
||||||
|
return undefined;
|
||||||
|
}
|
||||||
|
const normalized = useCaseRaw.trim().toLowerCase();
|
||||||
|
if (!normalized) {
|
||||||
|
return undefined;
|
||||||
|
}
|
||||||
|
|
||||||
|
const mappings: Array<{ tier: ModelTier; tags: string[] | undefined }> = [
|
||||||
|
{ tier: 'fast', tags: config.models.fast?.for },
|
||||||
|
{ tier: 'default', tags: config.models.default.for },
|
||||||
|
{ tier: 'complex', tags: config.models.complex?.for },
|
||||||
|
{ tier: 'local', tags: config.models.local?.for },
|
||||||
|
];
|
||||||
|
|
||||||
|
for (const { tier, tags } of mappings) {
|
||||||
|
if (!tags || tags.length === 0) {
|
||||||
|
continue;
|
||||||
|
}
|
||||||
|
if (tags.some((tag) => tag.trim().toLowerCase() === normalized)) {
|
||||||
|
return tier;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
return undefined;
|
||||||
|
}
|
||||||
|
|
||||||
/**
|
/**
|
||||||
* Create the unified message handler for the channel registry.
|
* Create the unified message handler for the channel registry.
|
||||||
* Each channel+sender pair gets its own AgentOrchestrator backed by a persistent session.
|
* Each channel+sender pair gets its own AgentOrchestrator backed by a persistent session.
|
||||||
@@ -83,6 +111,7 @@ export function createMessageRouter(deps: {
|
|||||||
|
|
||||||
// Cron job tier wins over agent config tier
|
// Cron job tier wins over agent config tier
|
||||||
const tierFromMetadata = metadata?.modelTier as ModelTier | undefined;
|
const tierFromMetadata = metadata?.modelTier as ModelTier | undefined;
|
||||||
|
const tierFromUseCaseMetadata = tierFromUseCase(deps.config, metadata?.modelFor);
|
||||||
|
|
||||||
// Include agent config name in cache key so different agents aren't shared
|
// Include agent config name in cache key so different agents aren't shared
|
||||||
let skillOverride = metadata?.skillOverride as string | undefined;
|
let skillOverride = metadata?.skillOverride as string | undefined;
|
||||||
@@ -100,8 +129,9 @@ export function createMessageRouter(deps: {
|
|||||||
// Read per-session model tier override (persisted in SQLite)
|
// Read per-session model tier override (persisted in SQLite)
|
||||||
const sessionTierOverride = session.getConfig('modelTier') as ModelTier | undefined;
|
const sessionTierOverride = session.getConfig('modelTier') as ModelTier | undefined;
|
||||||
|
|
||||||
// Resolution chain: metadata (cron) → session override → agent config → global default
|
// Resolution chain: metadata (explicit tier) → metadata modelFor -> session override -> agent config -> global default
|
||||||
const effectiveTier = tierFromMetadata
|
const effectiveTier = tierFromMetadata
|
||||||
|
?? tierFromUseCaseMetadata
|
||||||
?? sessionTierOverride
|
?? sessionTierOverride
|
||||||
?? agentConfig?.modelTier
|
?? agentConfig?.modelTier
|
||||||
?? deps.config.agents.primary_tier
|
?? deps.config.agents.primary_tier
|
||||||
@@ -766,9 +796,12 @@ export function createMessageRouter(deps: {
|
|||||||
let effectiveTier: string = deps.config.agents.primary_tier ?? 'default';
|
let effectiveTier: string = deps.config.agents.primary_tier ?? 'default';
|
||||||
const session = deps.sessionManager.getSession(msg.channel, msg.senderId);
|
const session = deps.sessionManager.getSession(msg.channel, msg.senderId);
|
||||||
const sessionTierOverride = session.getConfig('modelTier');
|
const sessionTierOverride = session.getConfig('modelTier');
|
||||||
|
const tierFromUseCaseMetadata = tierFromUseCase(deps.config, msg.metadata?.modelFor);
|
||||||
|
|
||||||
if (msg.metadata?.modelTier) {
|
if (msg.metadata?.modelTier) {
|
||||||
effectiveTier = msg.metadata.modelTier as string;
|
effectiveTier = msg.metadata.modelTier as string;
|
||||||
|
} else if (tierFromUseCaseMetadata) {
|
||||||
|
effectiveTier = tierFromUseCaseMetadata;
|
||||||
} else if (sessionTierOverride) {
|
} else if (sessionTierOverride) {
|
||||||
effectiveTier = sessionTierOverride;
|
effectiveTier = sessionTierOverride;
|
||||||
} else if (deps.agentRouter && deps.agentConfigRegistry) {
|
} else if (deps.agentRouter && deps.agentConfigRegistry) {
|
||||||
|
|||||||
Reference in New Issue
Block a user