feat(router): Executive Router v1 — intelligent multi-provider model routing#662
Open
tonyskifo wants to merge 5 commits into
Open
feat(router): Executive Router v1 — intelligent multi-provider model routing#662tonyskifo wants to merge 5 commits into
tonyskifo wants to merge 5 commits into
Conversation
…ider
Add the foundational layer for Executive Router v1. No dispatch logic
is wired yet; these are the schema, config, and catalog primitives that
Steps 4-7 will build on.
src/types/router-config.ts (new)
Complete type system for the routing pipeline: TaskType, ContextLen,
Urgency, TaskClassification, RouteDecision, RouterPoolEntry,
PolicyMatch, PolicyRule, RouterEscalationConfig, RoutingConfig.
src/server/hermes-config-store.ts (additive)
- SetRoutingConfigPatch action added to HermesConfigPatch union
- DEFAULT_ROUTING_CONFIG constant (Sonnet default, Opus threshold
0.75, 5 USD/day cap, empty pool/policy)
- readRoutingConfig(paths?) safe parser for the routing: block in
~/.hermes/config.yaml; validates every field, falls back to defaults
on missing or malformed values, never returns partial data
- applySetRoutingConfig shallow-merges patch into existing routing:
block without touching other config keys
- Existing applyHermesConfigPatch switch extended; exhaustive check
still passes
src/lib/provider-catalog.ts (additive)
DeepSeek entry added between MiniMax and Ollama. Uses OpenAI-
compatible API so no new backend handler is needed; baseUrl override
in openai-compat-api.ts already supports it.
Tests: 23 tests, all passing (vitest run)
src/server/hermes-config-store.routing.test.ts: 15 tests covering
readRoutingConfig and set-routing-config patch with real temp-dir
YAML I/O
src/lib/provider-catalog.test.ts: 8 tests covering DeepSeek catalog
entry, getProviderInfo, getProviderDisplayName
No changes to gateway-capabilities.ts, chat dispatch, or PR outsourc-e#661.
Pure-function pipeline for intelligent model routing: task-classifier.ts - parseManualOverride(): use:opus/sonnet/codex/gemini/deepseek/local/fast and model:<id> (model: takes precedence over use:X) - classifyTask(): token estimation, keyword-based task_type detection, log-scaled complexity 0–1, urgency from last user message executive-router.ts - route(): 6-step decision tree — manual override → routing disabled → policy rules (first match, skip disabled providers) → complexity escalation to Opus → default; Opus budget cap enforced at every path Tests: 52 tests across task-classifier.test.ts and executive-router.test.ts No new tsc errors introduced. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Persists per-request usage to ~/.hermes/model-usage.json. - ModelUsageRecord: id, ts, provider, model, input/output tokens, cost_usd, latency_ms, success, error - estimateCost(): pricing table for 13 models across Anthropic, OpenAI, Google, DeepSeek, Ollama; prefix-match for versioned IDs; zero for local - recordUsage(): computes cost, appends to JSON array; never throws - readUsageLog(): reads full log; returns [] on missing/corrupt file - sumCostToday(): aggregates today's cost with optional provider/model filter - getOpusSpendToday(): convenience wrapper for router budget enforcement All paths injectable for testing. 29/29 tests pass. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Wires the executive router pipeline into the hot chat path. Router is opt-in (routing.enabled: false by default) and never overrides an explicit user-selected model. - Import readRoutingConfig, classifyTask, parseManualOverride, route, getOpusSpendToday, recordUsage, RouteDecision - After local-model detection: classify task, call route(), apply RouteDecision — sets chatMode/localBaseUrl for non-Anthropic providers - Thread routerDecision?.model into Responses API, openaiChat, and streamChat dispatch points - Record usage (tokens, cost, latency) on openaiChat success and failure - Router block is wrapped in try/catch — failure always falls through to existing defaults; no change in behaviour when routing disabled Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds a Smart Routing settings panel at /settings?section=routing. settings-sidebar.tsx - Add 'routing' to SettingsNavId (fixes pre-existing tsc error at line 360) - Add 'Smart Routing' nav item between Agent Behavior and Voice hermes-config-route.ts - Import readRoutingConfig; expose routingConfig field in GET response - Add set-routing-config to PatchActionSchema (zod-validated) and ACTION_MESSAGES settings/index.tsx - Add RoutingConfigSnapshot type and routingConfig? to ClaudeConfigData - Add 4 routing state vars; sync from fetchConfig via syncInputsFromData - Add saveRoutingConfig() — sends action: set-routing-config PATCH - Replace renderSmartRouting() with renderExecutiveRouting(): - Executive Router section: enabled toggle, default provider, default model - Opus Escalation section: complexity threshold, daily USD budget - Pool & Policy section: read-only entry count when config has entries src/routes/api/-routing-config.test.ts (new, 7 tests) - GET returns routingConfig with defaults - GET reflects persisted routing block - PATCH writes fresh config, merges without losing keys, saves escalation - PATCH 400 on missing routing field, 503 when capability unavailable Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds a fully opt-in Executive Router pipeline to Hermes Workspace. When
routing.enabled: truein~/.hermes/config.yaml, every chat request isclassified and dispatched to the most appropriate provider and model based
on task type, complexity, and a configurable daily Opus budget cap.
Routing is disabled by default — no behaviour changes until opted in.
What's included
RoutingConfig,PolicyRule,RouterPoolEntry), routing config schema, DeepSeek provideruse:opus,use:deepseek,model:<id>)send-stream.tswith automatic routing and usage recordingKey design properties
Test coverage
111 tests added
Test plan
use:opus,model:...) workNotes
This feature is completely opt-in and preserves existing behaviour unless routing is explicitly enabled.