switchboard

changelog

What's changed.

Current switchboard-fyi@0.2.17

v0.2.17·2026-07-02·claude

Claude Sonnet 5 joins Claude Code routing.

Switchboard now routes the Sonnet tier through Claude Sonnet 5, with updated aliases, effort support, and gateway compatibility cleanup.

What's new

  • Claude Sonnet 5 is now available for Claude Code routing, including sonnet, sonnet[1m], and claude-sonnet-5[1m] aliases.
  • New Claude Code configs now default difficulty 2 and 3 to claude-sonnet-5, and the catalog exposes xhigh effort for Sonnet 5.

What's fixed

  • Sonnet 5 routes drop unsupported manual thinking budgets, sampling parameters, and trailing assistant prefill before forwarding.
  • Sonnet 5 routes preserve thinking: { type: "disabled" }, which Anthropic documents as supported for turning adaptive thinking off.
  • Hosted routing analytics now normalize the sonnet alias to claude-sonnet-5.
v0.2.16·2026-06-12·claude

Claude Fable 5 joins the model catalog.

Switchboard now knows Claude Fable 5 aliases and long-context variants, while tightening token accounting, route breaker recovery, and a few local CLI cleanup paths.

What's new

  • Claude Fable 5 is now available for Claude Code routing, including fable, fable[1m], and claude-fable-5[1m] aliases.
  • The gateway advertises Claude Fable 5 as a 1m adaptive-thinking model with supported effort levels.

What's fixed

  • OpenAI cached input tokens are counted as part of processed input, not added on top.
  • Claude routes to always-on adaptive-thinking models drop incompatible thinking fields and trailing assistant prefill.
  • Half-open route breaker trials now expire if a process disappears during the trial.
  • Routed streams abort upstream reads when the downstream client closes after headers.
  • switchboard difficulty set, switchboard login, and the local inspector handle edge cases more cleanly.

What's changed

  • The npm README now uses the updated terminal screenshot.
v0.2.15·2026-06-01·fallback

Fallback now returns to your harness default.

Switchboard now treats fallback as a return to each harness's configured best profile, not whatever routed model happened to be in the request body.

What's fixed

  • Router errors, classifier failures, guarded routes, and mutated upstream retries now use the configured best harness profile.
  • Codex fallback applies the configured reasoning effort, and Claude fallback applies the configured output effort.
  • auto, missing models, and previously routed request models no longer become the fallback target.
v0.2.14·2026-06-01·routing

Classifier outages now use the configured default.

When the hosted classifier fails or times out, Switchboard now routes through each harness's configured best profile instead of leaving a lower-tier original model in place.

What's fixed

  • Classifier failures and timeouts now send Codex and Claude Code requests to the configured best harness profile.
  • Workers AI classifier errors are handled with the same fallback policy across harnesses.
  • Non-classifier API errors still preserve the original request as router_error.
v0.2.13·2026-06-01·release

Homebrew releases now prove they shipped.

Switchboard now verifies the public Homebrew tap during release and confirms the installed binary version before an updater claims success.

What's fixed

  • switchboard update now checks that the installed command reports the target version after Homebrew or npm finishes.
  • Homebrew updates fail loudly if brew upgrade exits successfully but leaves Switchboard on the old version.

What's changed

  • npm run release:homebrew now updates, commits, pushes, and verifies the public tap formula from the npm registry tarball.
  • npm run release:postpublish verifies that npm and the remote Homebrew tap point at the same released version.
v0.2.12·2026-06-01·routing

Gemma classification stops timing out.

Switchboard now gives the hosted Gemma classifier enough time to answer and asks it for routing JSON without spending the output budget on model thinking.

What's fixed

  • The API disables Gemma thinking for classifier calls, avoiding empty routing responses from tiny output budgets.
  • Local gateways now allow at least 15 seconds for Switchboard API classification, even when an older config still says 2500ms.
  • Gemma routing is confirmed live against Cloudflare Workers AI and the deployed /v1/classify endpoint.
v0.2.11·2026-06-01·update

Updates now respect how Switchboard was installed.

Switchboard now detects Homebrew installs and points updates through Homebrew instead of npm, with a manual update check available from the home screen.

What's fixed

  • Homebrew-managed installs now use Homebrew update instructions instead of npm install -g.
  • switchboard update runs the right updater for the active install source: Homebrew for brew installs, npm for npm installs.

What's new

  • The interactive Switchboard home screen now includes a check updates action.
  • Manual update checks refresh update state and show the exact command users should run.
v0.2.10·2026-06-01·usage

Actual token usage shows up in Switchboard.

Switchboard now records actual upstream token usage when providers return it, labels estimated versus actual totals in the CLI, and keeps Codex continuation routing steadier.

What's new

  • Forwarded provider usage is recorded against Switchboard sessions, including Claude cache reads/cache creation and OpenAI cached/reasoning token details.
  • CLI receipts and dashboards now label token totals as actual, estimated, or mixed.
  • Codex continuation calls can reuse the latest routing decision for the same account/thread when there is no fresh user request.

What's fixed

  • Compatibility route breakers are cleared once after a package update so routing can recover after CLI fixes.
  • Routing, classifier, fallback, and Switchboard mentions no longer count as high-risk evidence by themselves.
v0.2.9·2026-05-28·claude

Claude Code fallback rows keep their difficulty.

Switchboard now classifies Claude Code requests before applying cooldown guards and normalizes Claude Code system messages when routing to lower Claude models.

What's fixed

  • Cooldown fallbacks now retain the classifier difficulty in the live feed.
  • Claude Code role="system" message entries are moved to Anthropic's top-level system field on routed Claude model changes.
  • Lower-tier Claude compatibility failures no longer hide the classifier result on later requests.
v0.2.8·2026-05-28·claude

Claude Code difficulty display is stricter.

Switchboard now keeps Claude Code tool output out of the active user request and only shows real classifier difficulty values in the live feed.

What's fixed

  • Claude Code tool results are treated as supporting context instead of the latest user request.
  • The live feed no longer turns best-tier or Opus routes into an implied difficulty 5 when the classifier value is missing.
  • The classifier now recognizes normal Claude Code and Opus 4.8 plumbing, including dynamic workflows, subagents, adaptive thinking, and large request envelopes.
v0.2.7·2026-05-28·claude

Claude Opus 4.8 is now the top Claude Code tier.

Switchboard now routes Claude Code's hardest work to Claude Opus 4.8 and upgrades old Opus 4.7 config values automatically.

What's changed

  • Claude Code's strongest tier now uses claude-opus-4-8.
  • Opus 4.8 is reflected in the model catalog, long-context routing, and gateway self-tests.
  • Existing Opus 4.7 config values are normalized to Opus 4.8 at runtime.
v0.2.6·2026-05-27·routing

Codex difficulty scoring is less jumpy.

Codex requests now give the classifier cleaner task signals, so normal tool use and source-code keywords are less likely to look like hard work by themselves.

What's fixed

  • Encrypted Codex reasoning is treated as normal plumbing, not automatic hard-task evidence.
  • Tool and source-code text no longer turns words like Error, router, or auth into broad risk signals by itself.
  • Real failed commands and test failures still stay visible to the classifier.
v0.2.5·2026-05-26·docs

The npm screenshot loads correctly.

The README image now uses a public npm CDN URL instead of a private GitHub raw URL.

What's fixed

  • The Switchboard dashboard screenshot on npm now loads from the public package CDN.
v0.2.4·2026-05-26·update

Update prompts find new releases faster.

Switchboard now refreshes release information after launch, so a cached check from earlier in the day no longer hides a newer version.

What's changed

  • The startup updater can still use cached release data immediately, but it now revalidates npm in the background on eligible launches.
  • The npm README now shows the routing dashboard in action.
v0.2.3·2026-05-26·fix

Multiple terminals work again.

Routed Codex and Claude Code terminals no longer block each other. You can run more than one wrapped harness window at a time again.

What's fixed

  • You can run multiple routed Codex or Claude Code terminals at the same time again.
  • The live dashboard still keeps one active owner, but individual harness windows do not.
v0.2.2·2026-05-26·fix

Updates won't break your tools.

Before, updating Switchboard could break the `claude` and `codex` commands. Now they keep working after every update.

What's fixed

  • Updating Switchboard no longer breaks claude or codex.
  • If a recent update broke things for you, this release fixes it automatically.
v0.2.1·2026-05-26·login

Sign in to use Switchboard.

You now have to sign in before using Switchboard. Most commands stop and ask you to log in first.

What's new

  • Most commands now ask you to run switchboard login first.
  • The home screen tells you to sign in instead of showing stats.
v0.2.0·2026-05-26·docs

Faster to get started.

The setup guide is much shorter now. Just install, sign in, go.

What's new

  • The npm setup guide is now just the basics: install, sign in, run.
  • Longer guides still live on switchboard.fyi.
v0.1.0·2026-05-21·launch

Hello, world.

The first Switchboard. It picks the right AI model for each task in Claude Code and Codex, shows you what's happening, and you only pay for what you use.

Smart model picking

  • Every task gets a difficulty score from 1 (easy) to 5 (hard). Easy tasks go to cheaper models, hard ones go to smarter models.
  • Just keep using claude and codex like normal.

Live dashboard

  • Run switchboard watch to see what's happening in real time.
  • Press d to see the mix of easy vs. hard tasks, which models are running, and any errors.

Pay as you go

  • Run switchboard login to sign in through your browser.
  • $1 buys 1,000 task scorings. Add more credits anytime from the command line.