Skip to content

Make Codex and Claude Code subscriptions go further: a local model router that gives routine agent work to economical models and keeps your strongest model for the hard parts.

12.6 KB
Raw Blame

Article illustrations

The article's explanatory slides have Chinese and English versions:

These are explanatory presentation graphics created with the built-in image_gen tool, not Desktop screenshots or measured output. Their text describes product roles and the routing policy discussed in the article. The body introduces the product, then presents two case studies, the local architecture and the download entry. Configuration pictures accompany the first case; competence, handoff and observation pictures accompany the second.

The original generation and correction prompts and architecture prompts record the requested text and visual constraints. Chinese and English text and arrows were visually checked. The product overview appears once in the introduction to explain the positioning and the slogan. The local architecture figure near the end explains runtime responsibilities and the computer boundary. Both are retained because they serve different purposes.

The cost chart and recorded HTTPX handoff graphic are generated separately by experiments/render_figures.py from the published evidence. Their measurements come from the experiment records, not image generation.

The earlier native Desktop screenshots (Chinese / English) were captured on 2026-10-04 from an isolated macOS app displaying an exact subset of the long-task observation metadata: one session, 40 requests, four quality segments, their receipts and safe execution facts. This is historical metadata replay, not a new model run or a screenshot captured during the original experiment. No account credentials, conversation bodies or agent connection configuration were imported.

The app was resized and its built-in zoom set to 150% for readability. PNG pixels were not retouched or generated. The native capture tool reported restricted screen access; the returned app images were visually checked against the app's accessibility snapshots. Existing partial-assessment and missing-content indicators were retained. The raw score is 0.485; the app formats it to two decimal places. These screenshots illustrate observation, not configuration or independent acceptance. They have been removed from the article body pending a corrected native capture: the model labels expose internal identifiers, and the evidence badges do not explain their coverage. The original session also has content-capture gaps; omitted replay content is not a complete explanation for its incomplete-content status. The assets remain for provenance, not as a claim that those presentation defects have been fixed.

Configuration screenshots

The article's three setup steps use six native macOS captures made on 2026-10-04:

Capture Chinese English
Model management Models Models
Smart Saving plan Routing Routing
Codex model-routing setup Client access Client access

The app source revision is 89925c86bc8b4993759a8ce137c2e905f17d5e72. The isolated Desktop and daemon used the managed Debug Pilot build at that revision. The final captures use the app's 100% text setting, as requested, and a native 1920 × 1050 window. The capture output is 1920 × 1080 including the top margin. PNG pixels were not cropped, retouched or generated. The capture tool reported restricted screen access; every returned image was visually checked against the real app and accessibility snapshot.

This is a configuration demonstration reconstructed through the normal Desktop UI, not a capture taken during the original experiment. The route Astra × Qwen · Smart Routing, alias hiroute-smart, was published and reopened at revision 2: Qwen-3.8 Flash in Economy at xhigh, GPT-6 Astra in Primary at medium. The configuration uses the built-in classifier to illustrate the model groups; it does not stand in for the experiments' recorded decision policy or request traces.

The source named Research models used a loopback, catalog-only endpoint and a non-production key. Model names and configurable capabilities were entered through the source editor; the demo uses a 131,072-token context limit and 32,768-token output limit. These settings are demonstration declarations, not measurements of upstream capabilities. The endpoint served only the two model IDs and rejected inference. No upstream model calls, subscription refreshes or account credential imports were performed. No price was assigned, so the source page correctly shows it as unpriced.

The Codex picture shows the ordinary setup dialog with the published route selected. The final client-enable action was not applied to the operator's Codex configuration. It demonstrates the setup entry, not a new end-to-end model test. The original model calls, cost figures and acceptance results remain in the frozen experiment files.

Observation evidence

The final Astra repair stage has no recorded later competence score. Keep that state; do not replace it with 0.775 from the earlier summary stage. A screenshot can display both models' stages but cannot establish a same-task score comparison. The independent 343-assertion result is separate evidence and should be presented as such.

The article uses the recorded handoff figure and the linked execution notes to explain the second experiment. It does not reuse the earlier flawed observation screenshots as current product evidence or substitute a new score into the frozen case.

Competence product views

The four images are native captures of HiRoute Desktop on macOS:

  • Session model performance (Chinese / English): actual execution models, reasoning settings, stage scores and evidence actions.
  • Routing-plan model performance (Chinese / English): model configurations grouped by execution branch, scoped averages, and scored/unrated stage counts.

Desktop reads an isolated copy of the unchanged observation records retained from this HTTPX case. The images come from the production UI and local daemon, without mock IPC, generated interface text or image edits. Both languages were selected through the app's settings, with 100% text size in all four captures. The Chinese routing list translates plan names and purposes in the isolated replay catalog; the English list retains the source names. Plan IDs, revisions, model candidates, execution identities and scores are unchanged. The main competence views display gpt-6-astra; the full CPA transport ID remains in expandable stage details. Reasoning settings use the original tier names (medium and xhigh) in both languages. Native validation covered stage and assessment-trigger navigation, exact model drill-down, unrated filtering, scope changes, refresh on re-entry, and preserving draft settings without changing published-plan statistics. Capture hashes, binary hashes and the source evidence archive are recorded in the capture provenance.

Qwen has one scored stage at 0.485; Astra has one at 0.775 and one unrated stage. The existing two-decimal display produces 0.48 and 0.78. The earlier 0.410 assessment lacks an attributed visible execution stage and is excluded from this product-view average. A prior 0.615 assessment of the later Qwen stage is not another independent stage. No score has been added to the final Astra stage.

Stage scores describe different stretches of work, not a controlled model ranking or whole-delivery acceptance. Retained diagnostic states and raw identities are available in expanded stage details. The captures validate the product's observation interface; they do not constitute a new model-quality experiment. Each article uses matching localized images and the same original numerical values and stage counts.

Local architecture

The architecture figures show one concrete call path: an existing agent calls the local HiRoute Gateway, which connects to cloud or locally hosted model services. Desktop/CLI configures the gateway, and execution facts feed local observation. The computer boundary includes the agent, management, gateway, records and local models; cloud services remain outside it. Connected model services perform inference.

The Rust core, shared Desktop/CLI service and local record storage are described by the public source references in the architecture prompt record. The figures explain those responsibilities rather than prescribing internal process or crate boundaries. The article describes the Higress engineering lineage separately from HiRoute's Rust implementation. The existing product overview remains in the introduction; the historical prompt file also retains an earlier, unexecuted redraw brief.

Open-source introduction

The Chinese and English introduction preserve the reviewed launch article's six chapters. It is an undated backfill, rather than a new launch announcement replacing the homepage case study.

  • The new Chinese routing overview and local architecture are explanatory image-generation illustrations from the reviewed launch draft. The matching English overview and architecture are editable SVGs with the same independent routing layers, local boundary and call directions. Decision services are external; CPA connects subscriptions. Neither diagram depicts an embedded Higress/Envoy dependency.
  • The ecosystem figure comes from the user-supplied AGNTCon + MCPCon China keynote. Its original labels are English, so both article versions use the same image. The ecosystem is background context, not a claim that HiRoute integrates every depicted project.
  • Context continuity has Chinese and English vector versions. A rebuilt complete prefix differs even if the same model continues; an unchanged common prefix may remain reusable. New user requests are reassessed, and different models do not share KV Cache. These are mechanism illustrations.
  • The new native stage-performance captures (Chinese, English) reuse the current homepage's exact localized captures. They show constructed documentation data for Qwen Flash and Qwen Max, not new benchmark measurements; see the original capture provenance.
  • Research cost, unattended handoffs and historical HTTPX observation reuse the existing case study's localized assets and original evidence. Their numerical values, stage counts and unrated final stage remain unchanged.

The three setup steps use actual native macOS product captures, with matching language for each article:

Step Chinese English
Add models Capture Capture
Smart saving Capture Capture
Agent connection Capture Capture

The Chinese smart-saving capture was made on 2026-10-08; the other setup captures were made on 2026-10-09. Both smart-saving views show an unsaved Qwen 3.8 Flash economy group and Qwen 3.8 Max primary group. No plan was saved, enabled or published for the captures. Agent views show the connection controls, without changing agent configuration. The English interface was selected through the product's language setting, then restored to its original setting. Native JPEG captures were encoded as PNG with the decoded pixels and dimensions verified unchanged. Product text was not translated by retouching screenshots. These setup pictures illustrate the real configuration flow; they are not inference or acceptance evidence.