Adds opt-in automatic approval for projects you trust, and refines the workspace.
- A failed Codex turn says why, for example that the configured model is not available for the account or that a usage limit was reached, instead of stopping with only "Failed". Long messages, and any that look like they carry a credential, keep the general advice to check the CLI.
- Opt-in automatic approval. Approval settings (select the approval note
beside the model picker) lets a trusted project apply file edits and/or run
shell commands without a prompt; a session can override its project. It is
off by default, and while it is on the composer note turns amber or red and
the header reads "Auto-approve on". Every operation is still recorded and
marked Auto-approved; protected paths stay blocked, and edits outside the
project, to Switchya's own state, gateway or agent account settings, to
files other programs load by themselves (shell profiles, Git, editor and
agent settings), or anywhere in a project that contains your home folder
still ask. The same project setting covers complete Codex and Claude Code
edit and command requests in Existing agents, read when each request
arrives, so turning it off reaches runs already in progress; deny-only
requests always wait for you. From the terminal, use
switchya trust --edits on --commands onor--auto-approveonchatandrun. - Approval settings opened from Existing agents shows only the project setting, which those runs follow. It no longer offers the override of a built-in session that is selected but out of view. The approval note now reads "Edits and commands need approval", the same words as its other states and the dialog; screen readers hear each checkbox by its title with the explanation as its description, and the project and session sets are named groups.
- Existing agents, the first-run view, says "No agent CLI found" when no Codex, Claude Code or Antigravity CLI is installed, and opens the list of what it checked, instead of showing only a disabled agent picker.
- Keyboard focus no longer drops to the page after Allow once, Deny or Close details: it returns to the review heading. On wide screens, opening a transcript row moves focus to its details, and Review agent request focuses the review pane on screen instead of opening the narrow-window drawer over it.
- The provider form says why a configuration name is rejected (lowercase letters, digits, hyphens or underscores, starting with a letter) instead of only disabling Save. Cancelled and timed-out commands read "Cancelled" and "Timed out" in transcript rows, as in their results, and a tab opened without its launch token says where the launch link comes from.
- Transcript rows, notices and the composer share one reading column. Existing-agent conversations show a repeated agent notice once, with one thin rule per finished turn, and tool rows use readable names such as "Read file" while Details keeps the exact tool name.
- Pending approvals stand out: hold amber appears only on the transcript row and approval card that need a decision, and the review header says "Approval needed" or "Nothing to approve". The first-run screen offers Connect a provider, Use an existing agent and Refresh models.
- Phone, keyboard and contrast fixes: setup buttons no longer break words, touch targets are 44px, the Built-in agent / Existing agents switch announces its state, small text is at least 12px, dark-theme panes are easier to tell apart, and copied file previews no longer contain hidden zero-width characters.
- On macOS, a command that finished normally is no longer reported as failed with "Operation not permitted (os error 1)" when its process group had already exited by the time Switchya cleaned it up.
- Sharing switchya.com shows a large preview card instead of a bare text link, and the site has an Apple touch icon.
- The download section lists the Windows requirements, the ZIP size and its SHA-256 hash next to the button, followed by the known limits.
- The product illustration uses the app's real wording and keeps the review pane at every width; tabs show a clear selected state, and small text, focus rings and the 404 page are fixed.
- A "From download to first task" section walks through the four steps: extract the ZIP and its runtimes, open a project, connect a model, then send a task and review it.
- A Questions section answers what Switchya costs, what data leaves your
computer, which models work, whether files change without asking, where
data is stored, why Windows warns about the unsigned ZIP (with a
Get-FileHashcheck), Mac and Linux availability, and how the app relates to Switchyard Gateway. - The page names local Ollama models next to OpenAI, Anthropic and Gemini,
follows the visitor's light or dark system theme, and publishes
SoftwareApplication and FAQ structured data for search engines. The footer
links the user guide, changelog, security model and MIT License, and
responses add a
Cross-Origin-Opener-Policyheader.
A gateway release with a clearer dashboard. Configuration files from 0.1.1 load unchanged.
- The Overview headline says "Degraded" when a model is cooling down on every credential or 5% or more of the last minute's requests failed, and names the model and when it comes back.
- A Routes panel on the Overview shows where each model's requests go now; rails turn amber while a route cools down.
- Request lists on Requests and the Overview hold still while the pointer or keyboard focus is on them; new requests wait behind "N new requests".
rustlsis updated to 0.23.46.
A bug-fix preview from a full audit of the app, CLI, agent tools and adapters. It includes the gateway fixes listed under 0.1.1.
- Security: on Windows, the agent's file tools refuse 8.3 short names such as
GIT~1orENV~1. These aliases could read, search or patch repository metadata and credential files that the tools block by name. - Antigravity conversations can be continued. A follow-up message resumes the
saved
agyconversation; ifagyreports a different conversation, the run stops for review instead of silently starting a new thread. - Claude Code replies that follow a thinking block no longer appear twice in
the workspace or in
switchya agentoutput, and their streamed fragments are replaced by the completed reply in saved history. - On macOS and Linux, Open Antigravity launches the CLI for managed Antigravity profiles instead of failing with "Account profile does not match the selected agent."
- Agent discovery detects an installed Oh My Pi again (in
~/.bun/binor onPATH) instead of always reporting it as not installed. switchya agent runandswitchya agent continuework with managed Antigravity profiles. Like the workspace, the CLI no longer refuses them for lacking a confirmed sign-in, which Antigravity cannot report; the Antigravity CLI reports sign-in failures itself.switchya agent loginchecks a profile whose status is not yet known before it starts a sign-in, so signing in from the terminal works without the app open, and it says clearly when a profile is already signed in. It exits with an error when the sign-in fails or is cancelled. Without--launch-info, Ctrl-C says that the sign-in stops with the command.- Removing a profile whose account metadata was changed outside the app no longer deletes its sign-in: the removal is refused and the profile keeps working. A removal that cannot be saved no longer stops the app from starting.
switchya mcp --projectresolves a relative path such as.against its own working directory, also when it joins the open app's host.- Project search (the built-in agent's
search_files, the local API'sGET /api/projects/:id/searchand MCPswitchya_search) no longer fails when one subfolder is unreadable or has a reserved name such asaux: that entry is skipped and counted in a newomittedfield. Search also finds matches in large UTF-8 files whose 64 KiB read limit falls inside a multi-byte character. - On Linux and macOS, files created by the agent's
apply_patchtool follow your umask (usually 0644) instead of being readable only by your user. - Ctrl/Cmd+Enter in an existing-agent conversation that is still running, awaiting approval or recovery review, or unsaved no longer starts a separate run. The shortcut follows the same rules as the disabled Send button.
- Claude Code Edit approvals that replace every occurrence of the text
(
replace_all) say so above the before/after text. - A task sent from New session no longer reappears in the composer the next time New session is opened. Changes shows the current diff or file preview once a later refresh succeeds, instead of keeping an earlier transient error such as "Project inspection is busy".
switchya-desktop.exeshows invalid command-line options in an error dialog instead of exiting silently, and shows--helpand--versiontext in a dialog when started without a console.- The Windows portable preview README says Codex, Claude Code and Antigravity
conversations run inside the workspace and can be shown, continued and
recovered with
switchya agent; opening Claude Code in its own terminal is optional. The archive includes the MCP guide that the user guide, local API guide and desktop README link to. - Corrected the documentation for preview.3. The adapter contract describes
the Codex device-code and Claude Console/SSO sign-in methods, the built-in
system-antigravityprofile and managed profile removal, and the local API and account guides acceptantigravityas an agent and terminal target. The guide's verification-status links work again, and the workbench, MCP and Claude workspace notes no longer tell preview.3 users to build from source.
A gateway bug-fix release from the same audit. Configuration files from 0.1.0 load unchanged.
- Security: the admin API refuses (403) a direct loopback request (one not
relayed by a reverse proxy on this machine) whose
Host, HTTP/2 authority orOriginnames a host other thanlocalhost, a loopback address,server.hostor the listener's own address. A web page using DNS rebinding can no longer make the operator's browser send wrong secrets and lock the operator out of the dashboard. If you reach a loopback gateway through a hosts-file alias, setserver.hostto that name. - Security: a client that sends the first bytes of the HTTP/2 preface and then stops is disconnected after the two-minute opening timeout. Before, such a connection stayed open indefinitely.
- Structured output requests from OpenAI, Responses and Gemini clients served
by Claude no longer return an empty
{}for map fields (for example pydanticDict[str, int]orpatternProperties) or free-form objects. These schemas are described to the model in a system instruction instead of being forced into a schema that only allows an empty object. /v1/completionsreturns log probabilities in the legacy completions shape (tokens,token_logprobs,top_logprobs,text_offset) instead of the Chat shape, in full and streamed answers. Offsets count from the start of the generated text and continue across stream chunks./v1/completionsrefuses token-id prompts, and any other prompt that is not text, with a 400invalid_request_erroronprompt. Before, the prompt was silently replaced by a placeholder and answered.- Gemini requests that list very many
allowedFunctionNamesor declare very many functions are no longer slow to process, so one large request cannot stall the gateway. - Tool parameters whose union choices are objects (discriminated unions such
as pydantic
Unionmodels) keep every choice when sent to a Gemini upstream. Before, only the first choice survived. videoMetadata(clip offsets, frame rate) andmediaResolutionon a Gemini client's media part are kept when the request is re-encoded for a Gemini upstream.- A
POST /v1/responsesrequest withprevious_response_idorconversationfor a model served by a Chat Completions, Anthropic or Gemini provider is refused with a 400 (previous_response_not_foundforprevious_response_id); send the full conversation ininputinstead. Before, it was answered without the earlier conversation while the reply claimed to continue it. - Responses WebSocket follow-up turns that drop a tool call the client never answered also drop the reasoning in front of that call, which OpenAI otherwise rejected with a 400.
- Requests translated for an OpenAI Responses upstream (for example from Chat Completions clients) no longer fail with a duplicate reasoning item when the client splits one response's parallel tool calls over several assistant turns: the shared reasoning is restored once per request.
- When a vendor answers a pooled credential with a 400 because it disabled or
restricted the organisation behind it (Anthropic's "This organization has
been disabled.", Groq's
organization_restricted), that credential rests and the request is retried on another one. Before, the client got the 400 and the dead credential stayed in rotation. - Rate limits that arrive while a model already rests on a credential after
a 5xx, a transport error or a 404 no longer escalate its backoff, so a mixed
burst of 5xx and 429 replies no longer rests that model on the credential
for much longer than
rate_limit_base_secs, doubled once per rate-limit window up torate_limit_max_secs, would. - An alias that is ignored because none of its targets is routable no longer hides the real model of the same name from other aliases that name it.
- Round-robin keeps separate cursors for models whose names differ only in case, and for requests only some providers can serve (raw side endpoints, the Realtime relay), so every credential gets its share.
- A request refused while a model is cooling down names the failure that started the rest on its record, not a later failure that did not lengthen it.
- Embeddings, image, speech and moderation requests whose JSON body is
labelled as something else (for example curl's default form type) reach
the upstream as
application/json, with the upstream's own model id. - A model listing that ends while a provider's endpoint is being changed no longer leaves the old endpoint's models routed to the new one.
- After a key's
rate_limit_rpmis lowered, a refused request'sRetry-Aftergives the real wait instead of about a second at a time. - Realtime relay: with
auth.required = false, a key offered in theopenai-insecure-api-key.<key>WebSocket subprotocol authenticates even when the browser also sends anAuthorizationheader that holds no key, such as a reverse proxy's Basic credentials. Before, such clients were treated as anonymous: refused withorigin_not_allowedon a public host name, or recorded without their key. - Provider
headersvalues written asenv:NAMEor${NAME}are read from the environment, as API keys are. Before, they were sent upstream as the literal text. If the variable is unset, the provider's credentials are shown as unusable with the reason and no request is sent; a reference that names no variable is refused when the configuration is checked. - Saving a provider through the admin API while moving one credential into
api_keysno longer hands it the key of a labelled credential with the same mask. Before, that silently moved the labelled credential's disabled flag, weight, priority and proxy to the other key. switchyard import-cliproxyno longer turns CLIProxyAPI'srequest-retryinto one attempt per retry round.routing.max_attemptsnow covers every imported credential (ormax-retry-credentialsof them) in each round, and never less than the default, so failover between keys keeps working after import, and the report explains the value.switchyard import-cliproxyimports a model pool of a prefixed OpenAI-compatible provider asprefix/pool, the name CLIProxyAPI served it under, plus the bare name unlessforce-model-prefixis on. A pool member whose name an earlier entry already gives to a different upstream model, or that another pool of the same provider is named after (earlier or later in the list), is left out and reported; before, the pool silently sent requests to that model or pool.switchyard import-cliproxyleaves out provider headers written asenv:NAMEand lists them in its report, since Switchyard now reads such values from its own environment.- Unquoted passwords and other secrets that contain brackets (for example
DB_PASSWORD=x7{Kq9!mZ#2) are redacted in logs and captured bodies instead of being mistaken for source code. - Captured request and response bodies over the size cap no longer keep the
start of a secret that was cut off at the cap, including when
logging.request_log_max_body_kbis 8 MiB or more. - A usage record written after a write that failed part-way (for example on a full disk) is no longer lost when usage history is reloaded.
- Dashboard: the Overview setup steps keep a reverse proxy's path prefix in
the example curl request, so it reaches the gateway when the dashboard is
opened at an address like
https://host/gw/admin/. - Dashboard: payload rules show a read error instead of loading a copy whose
large integers and decimals such as
1.0had been silently rewritten, which the next save would have written back to the configuration. - Dashboard: on the Requests page, a status filter the gateway refuses (for
example
6xx) no longer lets live requests fill the list as if no filter were set. - Dashboard: "Paths to remove" in a payload filter rule accepts paths that differ only in letter case, as the gateway treats them as different fields.
- Release tags with a prerelease suffix (for example
v0.2.0-rc.1) are published as GitHub prereleases and no longer become the Latest release or move thelatestcontainer image tag.
switchyard.example.tomlshows every setting, including a client key'senabledflag, an alias'shide_targets, the Vertexprojectandlocationoptions and the thinkingdynamic_allowedfield.- The design notes no longer claim the container image enables remote admin
access; pass
-e SWITCHYARD_ADMIN_ALLOW_REMOTE=trueat run time. - The README, design notes, security model, adapter contract, admin API reference, gateway API notes, agent-tools README, example configuration and the Switchya, account, MCP and local API guides describe the changes in this release and in Switchya 0.1.0-preview.4.
- Managed Codex and Claude Code profiles can be removed from Accounts. Removal deletes the profile's private CLI home (its stored sign-in), returns a removed default to the Existing CLI entry, and is refused while the profile has an active run, refresh, sign-in or terminal launch. Conversations that used a removed profile can no longer continue.
- The built-in agent's "Connect a coding model" notice also offers the existing Codex or Claude Code setup, opening Accounts in Existing agents mode.
- Sign-in methods: Codex profiles can sign in with a browser or a device
code; Claude profiles can use a Claude subscription, Anthropic Console or
SSO. The CLI gained
switchya agent login --profile ID [--method …]andswitchya agent use --profile IDfor terminal-first account switching. - Antigravity (
agy) is a third existing agent: named profiles, terminal sign-in, and read-only plan-mode conversations in the workspace and CLI (switchya agent run --agent antigravity). Its CLI has no account-status command, so status stays unknown and runs surface the CLI's own sign-in failures. Gemini CLI is not supported: Google has ended Gemini Code Assist for individuals and directs users to Antigravity. - Website: corrected the model-access section — Codex and Claude Code conversations run inside the workspace; opening Claude in its official terminal is optional for managed profiles.
First release.
- Client APIs: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages
and Gemini
generateContent, streaming and non-streaming, with token counting and model listings in each vendor's shape. Legacy/v1/completionsis served through chat. - Any client protocol can be served by a provider that speaks another one. Requests in the provider's own protocol preserve provider fields, with JSON normalization, model routing and configured reasoning or payload changes.
- Providers: OpenAI, Anthropic, Gemini, Vertex AI (service account or API key), any OpenAI-compatible server, and a built-in mock provider for trying things out.
- Several credentials per provider with round-robin, fill-first, weighted or least-latency selection, priorities, session affinity, cooldowns with backoff and automatic failover. Streaming requests are retried on another credential until the first event has been sent.
- Model aliases with fallback chains, provider prefixes, exclusions, model discovery and a built-in catalog of model capabilities.
- Reasoning suffixes —
model(high),model(8192),model(none),model(auto)— converted to each provider's own settings and fitted to what the model accepts. - Payload rules that set default values, override values or remove fields in upstream request bodies.
- Client API keys with model allow-lists and per-minute rate limits.
- Outbound proxies (HTTP, HTTPS, SOCKS5) globally, per provider or per credential. Optional TLS for the listener.
- The OpenAI Responses API over WebSocket (
GET /v1/responses) for any model, with per-connection conversation state so follow-up turns can send only their new input. - A relay for the OpenAI Realtime API (
GET /v1/realtime). - A live event feed for the dashboard.
- Single-use browser connection tickets, with current key permissions checked during long-lived client sessions.
- Embedded in the binary at
/admin/: live overview, request log with attempt timelines and captured bodies, provider and credential management, model routes and aliases, client keys, usage and cost charts, a playground for all four protocols and the WebSocket endpoint, live logs, settings forms and a raw configuration editor. Light and dark themes; works on phones.
- One TOML configuration file, reloaded when it changes. Edits made from the dashboard keep the file's comments and formatting.
- Usage history kept on disk, with estimated cost from a configurable price table. Optional capture of request and response bodies.
switchyard init,checkandimport-cliproxy(converts the API-key sections of a CLIProxyAPI configuration).
- Enforced current key policy and resolved-model permissions; anonymous browser requests check their Origin and Host independently of CORS configuration.
- Bounded schema conversion, tool bookkeeping, discovery responses and JSON display work; expanded secret redaction and private Unix file permissions.
- Completed dashboard forms, focus handling, exact usage/request drill-down, WebSocket tickets, configuration error mapping and credential preservation.
- Scoped release job permissions and resolved release tags to an immutable commit. Documented shared-account, trusted-browser and Windows ACL limits in the security model.
- Windows x64 archive, built and verified locally. Linux/macOS archives and container publication remain pending while GitHub Actions is unavailable.
- Signing in to consumer subscriptions or imitating vendor command-line clients. Providers are reached with API keys or service accounts.