okooooooo/magpie
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
2,515 Commits 1 Branch 0 Tags
main
Commit Graph
6 Commits
Author SHA1 Message Date
yetone 40c5b92a93 gateway: structured output crosses every translation: a Responses text.format or a Chat response_format reaches a Chat, Responses, Messages or Gemini upstream in its own words (frostming on Discord: responses.parse through magpie to OpenRouter came back as plain text) 
The IR held a schema only from Anthropic's output_config.format, and only
Messages (and Claude Code's --json-schema) were asked for one. A Responses
request to a provider magpie speaks Chat or Messages to (OpenRouter has no
Responses URL) lost text.format; a Chat request translated to Responses or
Messages lost response_format; an Anthropic schema translated to Chat or
Responses was lost too.

Format (format.go) now carries json_schema (name, description, schema,
strict) or json_object from Responses, Chat, Anthropic and Gemini, and each
builder asks for it in its upstream's field. Where a target can't say it
(json_object on Messages or Claude Code; the vendor bridges Cursor, Devin,
Kiro, Qoder, Command Code, Zed) it is told in the system prompt. An upstream
answering 400/422 naming the field (DeepSeek's "This response_format type is
unavailable now") is asked again with it told in words, and so from then on.
Relayed requests are unchanged.
 2026-10-08 10:37:02 +08:00
yetone b56bfb5a94 gateway: an upstream that answers 200 with a web page, with nothing at all, or with JSON that isn't, has failed as a 502 does, so a group asks its next member (#1012, cth123456: Codex on group/medium, agentrouter/deepseek-v4-flash now and then answered 200 text/html; the log said status=200 done=true in=0 out=0 served=null, codex/gpt-6-luna was never asked, and Codex failed with "stream disconnected before completion: stream closed before response.completed") — the candidate loop judged a try by its status alone, and holdWriter passes a 200 that is neither text/event-stream nor application/json straight to the agent, so the page went to Codex as the reply. notAnAPIReply now reads every 2xx (but 204) an upstream gives, in forwardOnce (every relayed and translated request, every API, plugins included) and in the built-in clients' own requests (the ChatGPT backend's /responses, Kiro, Zed, Qoder, Devin, Cursor, Command Code): a text/html reply, a body that begins as a web page whatever its Content-Type, an empty body, or one sent as JSON that doesn't begin with { or [ is turned into the 502 it stands for, with a JSON error saying what came back ("answered 200 OK with a web page (text/html), not an API reply: “Just a moment...”", the firewall hint for a block page). It then fails over, rests a minute and is logged as any 502 before the reply begins does; a provider with nobody after it hands the agent that 502 rather than the page. Only the first bytes are peeked, which a stream waits for anyway; a read that fails is left to the reader, a compressed body and anything else as it came. TestNotAnAPIReplyFailsOver (Chat, Responses and Messages, streamed and not, a member answering a Cloudflare page as text/html, as JSON and as SSE, an empty body as JSON and as SSE, and plain text as JSON; the next member must answer and the log name why) fails without the change in all 36 cases (the page relayed to the agent, or b never asked, or no reason logged) and passes; TestNotAnAPIReplyAlone (a 502 with the reason, logged 502) and TestCodexBackendWebPage (the ChatGPT backend's page reached Codex as a 200) fail without it and pass; TestAPIReplyUntouched keeps JSON after whitespace, SSE with and without a Content-Type, gzip, XML, a 204 and a 404 as they came. -race -count=10 on them; go test -tags nogui ./internal/gateway ./internal/provider, go vet (darwin, windows), gofmt -l, linux and windows (nogui) builds pass. Not tried with the reporter's relay (agentrouter) itself; the page is a Cloudflare challenge's, as the issue has only its Content-Type. docs/subsystems/gateway-routing.md says it. 2026-10-06 18:26:31 +08:00
yetone fa9daa6945 docs: every built-in subscription a plugin serves once moved says so in its code, and AGENTS.md says where a fix for it goes (user: 对内置的订阅 provider 在代码层面和文档层面加上注释，防止又有别的 agent 在改内置订阅 provider 的代码而不知道生效的其实是对应的插件代码) — a moved Cursor, Devin, Grok, Command Code plan, Kiro, Zed, Qoder, Factory, Xiaomi MiMo, ZCode or WorkBuddy (provider.Moved; since v0.1.642 also the default for a new sign-in, provider.Adopt) is signed in, listed, requested and metered by its @magpie-community plugin, and none of the built-in's code in internal/provider/<name>*.go, internal/gateway/<name>.go, internal/zed or internal/qoder serves it; a fix made there alone never reaches those users. Each of those files now opens with a PLUGIN-SERVED notice naming the subscription's id, its plugin package, where the plugin lives (magpie-community/plugins, packages/<name>) and the mover whose min to raise so moved users get the fixed plugin (keepMovedCurrent); Qoder's adds that Qoder CN isn't moved and still runs there. AGENTS.md (CLAUDE.md imports it) lists the moved built-ins with their packages and movers, how to tell whether a user's is moved (move.state "plugin", onPlugins), the order a fix goes in (the plugin, publish, raise min, the built-in only for those not moved), the magpie code that does run for a moved one (internal/plugin, plugins.go, plugin_usage.go, pluginsignin.go, migrate*.go, app.js's subOf/pluginSubs/startPluginSignIn), and the parity tests. movers in migrate.go and SUBS in app.js point there too. TestMovedBuiltinsSayTheirPlugin fails for a mover whose code carries no PLUGIN-SERVED notice naming its id and package (it failed while a wrapped package name was split). Comments and docs only, no behaviour changed. go vet ./internal/..., go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui -count=1 ./... under a temp HOME pass (TestProvidersAnswerWhileListsComeIn and TestClaudeSignedOut failed once on their timing under load, pass alone; untouched). 2026-10-02 19:37:02 +08:00
yetone cb761abdd1 commandcode: a reply is asked for within Command Code's 200000 tokens, and the agents are told no more (Discord: every Command Code model, free or paid, failed with "Too big: expected number to be <=200000 at params.max_tokens") — magpie passed the client's max_tokens to /alpha/generate as it came, and told the agents each Command Code model's reply limit as models.dev gives it: DeepSeek V4's 384000, MiniMax M3's 512000, Space Bunny Alpha's 524288, Kimi K2.x's 262144, so an agent asking for what it was told had the whole request refused. params.max_tokens is now the asked number within the model's own limit as models.dev gives it and within 200000; less is sent as it was, and none asked is still the CLI's 64000. The Command Code subscription's models are listed to the agents (their configs, /v1/models) with a reply limit of at most 200000, a user's own limit included. Prompts and headers are unchanged. TestCommandCodeGoMaxTokens and TestCommandCodeOutputCapped fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass. 2026-10-02 06:58:18 +08:00
yetone f654bec468 gateway: a model's web search goes to a search API of the user's own when no provider can search (#419 by HwFee: let me set up Tavily, Brave or SearXNG with my key rather than lose the search tool) — with no provider that searches by itself and no provider magpie could search with, the gateway took the client's web search tool away. Tavily, Brave Search, Exa, Firecrawl and SearXNG (an address of one's own, a key only if it asks for one) can now be set up, in an order: the injected web_search tool is then offered, and a search goes to the searcher's model as before, or, with none or when it fails, to each search API in turn, called directly with no model, its results handed back as the same hits and text the searcher path gives, so clients see the same web_search_tool_result blocks. The order is provider searches itself, searcher provider, search APIs, strip. They are kept in providers.json (0600) like provider keys; a backup carries them, their keys only when it carries keys, and one restored without keys keeps the key here; sync does the same, and a backup or sync from a magpie before them leaves the ones here alone. magpie search lists them, magpie search add <api> [key] [url=…] adds or updates one, magpie search rm <api> removes one; Settings has a Web search section (联网搜索) to add and remove them. Prompts, identity lines and headers to model providers are unchanged. TestWebSearchBySearchAPI, TestNoSearchAPINoTool, TestBackupSearchAPIs, TestSyncSearchAPIs and search-api.test.cjs fail without the change; TestSearchAPIsEachAsked checks each vendor against a fake, TestSearchCLI the command. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass; search-api, provider-search, click-scroll, old-webkit pass. 2026-10-01 19:44:44 +08:00
yetone a08b061b2f commandcode: a Command Code Go-plan account is served through the gateway in the CLI's own format, at /alpha/generate (Discord #general, Mikan: 有计划支持command code的go plan吗，走他们的私有格式) — Go has no Provider API access, so magpie's Command Code Plan subscription, which asks api.commandcode.ai/provider, could not serve a Go account. Read from command-code 1.72.2's bundle: the CLI asks POST /alpha/generate for the models Command Code hosts, with {config, memory:null, taste:null, skills:null, permissionMode, params:{model, messages, tools, system, max_tokens, stream, temperature?, reasoning_effort?}} (createModelClient, toWireMessages: text / tool-call / reasoning parts, tool results as role "tool" tool-result parts, images as data URLs), the headers of buildCommandAuthHeaders (User-Agent cli, x-command-code-version, x-cli-environment production, x-project-slug, x-taste-learning, x-session-id, Bearer key), and reads newline-delimited JSON back (text-delta, reasoning-delta, tool-call, finish with totalUsage, error, abort). Its plan table lets individual-go / individual-go-v1 use the "opensource" models less a blocked list. Now: an account's plan is read from /alpha/billing/subscriptions and kept 10 minutes per key (the Usage page's read refreshes it; a failed read falls back to the sign-in's plan and is retried after a minute; GOAT is not taken for Go). A Go account is asked at /alpha/generate, translated to and from Anthropic Messages, OpenAI chat and Responses, streamed or not, tool calls included, and lists Go's models from the CLI's table (GPT-6 Luna, GPT-5.6 Luna, DeepSeek V4 Pro/Flash, Kimi K3, GLM-5.3, MiniMax M3, Qwen 3.8 Max/Flash, MiMo V2.6 Pro); MODEL_NOT_IN_PLAN is a 403 and exhausted credits a 402, so routing moves on; token counts are estimated locally. Pro, Max, Ultra, GOAT and Teams accounts go on the Provider API as before. Tested with internal/gateway/commandcode_test.go and internal/provider/commandcode_go_test.go against a fake Command Code in a temp HOME: a Go account routed to /alpha/generate and a Pro one to /provider/v1/messages, the request body and headers, streamed Anthropic, chat and Responses (streamed and not) with text, reasoning, a tool call and usage, MODEL_NOT_IN_PLAN over HTTP and in the stream, credits run out, a stream error, a reply cut short, the plan cached, re-read and falling back; also under -race. They fail with the routing off. go vet, go build -tags nogui (and GOOS=linux, GOOS=windows) and go test -tags nogui ./... pass. Not tried with a real Go account: which models Go allows today, whether the server checks config, project slug or session id, the usage arithmetic with cache tokens, real errors' codes, and pause_turn (ended rather than resent). 2026-09-30 08:37:57 +08:00