Files
Comic Chang
49f40b463f
Configure per-model context windows and output limits (#279)
A provider that serves a model the catalogue does not list, or lists at the wrong size, has a window and a reply limit magpie cannot know: both fall back to whatever the catalogue says about the model's maker, which for such a provider is nothing at all. magpie model context relay-a/gpt-5.5 262144 magpie model output relay-a/gpt-5.5 131072 magpie model context 'relay-a/*' 200000 magpie model output relay-a/gpt-5.5 --reset Both are looked for in the same order as each other: this model, then `<provider id>/*`, then the provider's own list, then the catalogue. A model's own value wins over its provider's. The reply limit is new; the window is not — `Provider.Contexts` has been there, with a `"*"` of its own, settable through `magpie provider set <id> context.<model>=`, and this reuses it rather than shadowing it. That reuse is the point rather than a detail. A context is a routing input and not only a number agents are shown: at 95% of a held member's window a request moves to a member that takes more, so two sources of truth would mean the gateway routing on one number while agents are told another. `magpie model context` is a second way to set the same field, not a second field. Both reach the agents. Pi keeps a model's window and reply limit in its own models.json, OpenCode keeps its limit, and Crush, Droid, Cline, Qoder and Zcode keep their own; an agent reads those files when it starts, so a limit set here that never reaches them is a limit that does not take effect. `magpie provider set` has always told the agents, and this behaves the same way rather than introducing a second, quieter path. The test asserts each setter reaches them exactly once — the two paths differ (one saves the provider file, the other settings and then says so) and that difference is worth pinning. An end-to-end pass confirms it: with a Pi models.json in place, `magpie model context relay-a/sol 999999` writes `contextWindow: 999999` into `providers.magpie.models`. A `0` is refused rather than read as a reset, naming `--reset`; a mistyped model id is refused the way `magpie model name` refuses the same id; a reset says what is actually in force instead of claiming to be back to the vendor's list while a provider-wide value still applies; and `magpie model context <provider>/*` reports that value rather than saying the provider serves no such model. Validation happens where the entry is written, so one hand-edited bad key cannot stop a theme or a window size from saving, and #288's walk moves these on a provider rename without a line here.
agents: omp's roles pick their own models (subagents, smol, slow), keep their thinking level, and get back what they had before magpie (#325)
tests: TestDriftUnwiredEveryAgent leaves WSL agents alone — it walks every agent and picks a magpie model to see the wiring come back, and on a Windows machine with a real distro running, the WSL agents in All() are the user's own: the pick wrote fake/m1 into their real ~/.codex/config.toml, ~/.claude/settings.json and ~/.pi/agent/settings.json through \wsl.localhost. Their files live in the distro, beyond the sandbox home, so the test skips them (a.WSL set). Verified on such a machine: the full ./internal/agent suite leaves the three files' hashes untouched. (#283)
agents: omp's roles pick their own models (subagents, smol, slow), keep their thinking level, and get back what they had before magpie (#325)
agents: Claude Code on a magpie model is told the model's context window (#311: its auto-compact settings stay as they were for another model) — Claude Code takes any model it doesn't know by name (all but Claude's own) for 200K, and CLAUDE_CODE_AUTO_COMPACT_WINDOW is only ever the smaller of its value and that, so a 128K model compacted too late and a 400K one far too early; CLAUDE_AUTOCOMPACT_PCT_OVERRIDE is a share of the window and holds for any model (read in Claude Code 2.1.285). magpie now writes CLAUDE_CODE_MAX_CONTEXT_TOKENS, the window Claude Code takes those names for, from the model's known context (a group's smallest), with the model: another model's replaces it, one magpie doesn't know the window of, a switch back to a Claude model or a reset removes it; with a [1m] main model, the smallest of the tiers not marked. A value the user set, or changed magpie's to, is left as it is (magpie keeps the one it wrote). TestClaudeContextWindow fails without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass.
agents: Claude Code's effort reaches Opus 5.5 and Sonnet 5.5, follows each model's levels, and gets an ultracode switch (#351, #352, #353, #354) — Claude Code 2.1.251 on reads Opus 5.5's, Sonnet 5.5's and any newer Claude's effort only from modelSettings.<model>.effortLevel, so the top-level effortLevel magpie wrote left them at medium; it now writes modelSettings for the model (and the top key for the models that still read it), and a reset takes out only its own level. The levels offered follow what Claude Code 2.1.285 sends each model: Opus 4.5 low–high, Opus/Sonnet 4.6 without xhigh, Haiku 4.5, Sonnet 4.5 and older none; a level a model doesn't take is refused. A new ultracode square writes "ultracode": true, on models taking xhigh, and magpie claude ultra no longer writes "ultra" as the model. An open session keeps the effort it started with, so a change says to restart it while Claude Code runs. Checked against real Claude Code 2.1.285 in a sandbox HOME with a fake API recording the effort sent. TestClaudeEffortModelSettings, TestClaudeEffortPerModel, TestClaudeUltracode and TestClaudeEffortNotice fail without the change; claude-effort.test.cjs (Chromium, WebKit, en/zh) added. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass; omp-roles, panel-effort, click-scroll, old-webkit pass.
agents: Claude Code's effort reaches Opus 5.5 and Sonnet 5.5, follows each model's levels, and gets an ultracode switch (#351, #352, #353, #354) — Claude Code 2.1.251 on reads Opus 5.5's, Sonnet 5.5's and any newer Claude's effort only from modelSettings.<model>.effortLevel, so the top-level effortLevel magpie wrote left them at medium; it now writes modelSettings for the model (and the top key for the models that still read it), and a reset takes out only its own level. The levels offered follow what Claude Code 2.1.285 sends each model: Opus 4.5 low–high, Opus/Sonnet 4.6 without xhigh, Haiku 4.5, Sonnet 4.5 and older none; a level a model doesn't take is refused. A new ultracode square writes "ultracode": true, on models taking xhigh, and magpie claude ultra no longer writes "ultra" as the model. An open session keeps the effort it started with, so a change says to restart it while Claude Code runs. Checked against real Claude Code 2.1.285 in a sandbox HOME with a fake API recording the effort sent. TestClaudeEffortModelSettings, TestClaudeEffortPerModel, TestClaudeUltracode and TestClaudeEffortNotice fail without the change; claude-effort.test.cjs (Chromium, WebKit, en/zh) added. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass; omp-roles, panel-effort, click-scroll, old-webkit pass.
agent: Claude Desktop runs on magpie in its third-party mode (#156: Claude Desktop 不会解锁 Claude Code 面板 — CC Switch 可以，咱这个支持吗？) — signed in with a Free account, Desktop's Code tab is only "Get Claude Code"; in its 3p deployment mode it needs no Anthropic sign-in and Code and Cowork run on the gateway. A Claude Desktop agent whose provider field switches it: on writes deploymentMode "3p" into Claude/ and Claude-3p/claude_desktop_config.json, a magpie profile in Claude-3p/configLibrary/<id>.json (inferenceProvider gateway, inferenceGatewayBaseUrl the gateway's address, inferenceGatewayApiKey magpie-claude-desktop, inferenceGatewayAuthScheme bearer; disableDeploymentModeChooser and coworkEgressAllowedHosts ["*"] only where unset) and its entry in _meta.json, applied — the keys and files CC Switch writes (claude_desktop_config.rs). No inferenceModels: Desktop lists /v1/models, which the gateway already gives it by ids it keeps. Off puts each deploymentMode back as it was (gone where there was none), the profile applied before applied again, magpie's profile and entry removed, and files and folders magpie made gone if nothing else went in; other keys and profiles are never touched. Folders as CC Switch finds them: ~/Library/Application Support, %LOCALAPPDATA% (a Claude… folder there if not Claude itself), $XDG_CONFIG_HOME. Desktop reads them at start-up, which the picker and notice say. The Library keeps giving Desktop's MCP servers to %APPDATA%'s file on Windows, listed once. Tested in a sandbox home: fresh, with the user's settings and 3p profiles (off restores every byte), on twice, on → off → on, drift, odd files refused, each system's folders; and against Claude Desktop 2.9939.2 on Windows 11 (winget): on, it started in 3P mode (provider gateway, apiHost the gateway) with the Code tab and no sign-in, and asked the gateway's /v1/models; off, it went back to claude.ai
library: Claude Desktop's MCP servers go into Claude-3p/claude_desktop_config.json as well, where its 3p mode reads them (#297: Claude Desktop 使用 magpie 三方网关（3p）后，library 写入的 MCP 在 Desktop 里看不到) — in 3p mode, which magpie's own provider switch turns on, Desktop's whole userData is Claude-3p (~/Library/Application Support/Claude-3p, %LOCALAPPDATA%\Claude-3p on Windows, ~/.config/Claude-3p), its MCP servers included, so the servers the library wrote into Claude/'s file never showed in Settings → Developer. Where the Claude-3p folder is, its file is now written with Claude/'s, in one edit with each file's own keys kept, read for servers to bring in, and cleared on removal even when Claude/'s no longer has the entry — both, not whichever mode is on, so switching the provider leaves nothing stale and Desktop finds them in either mode; no Claude-3p folder is made for a Desktop never in 3p, and one only ever in 3p (no Claude folder) gets its servers there alone. A server given before reaches Claude-3p on the next sync. Tested with new desktop3p_test.go in a sandbox HOME (both files written and cleared, the user's servers and keys kept, sync of an older server, 3p only); each fails without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass. Not tried against a real Desktop in 3p.
agent: OmO (oh-my-openagent, omo-ai), a fork of Pi that keeps its own ~/.omo/agent, gets its own row: a magpie model sets settings.json's defaultProvider/defaultModel and magpie's provider in models.json there as Pi's does, $OMO_CODING_AGENT_DIR / $SENPI_CODING_AGENT_DIR are honoured, the Library writes its AGENTS.md, mcp.json and skills, and
omo update updates it (#264 rhyme227). Pi's own ~/.pi is left alone. piIn becomes piLike(at, id, name, dir) so both share one wiring. Tested with omo_test.go in agent and library (fail without the change); go vet, go build -tags nogui (darwin, GOOS=linux, GOOS=windows), go test -tags nogui ./... pass.
agents: Codex signed in with only an API key gets magpie as its provider, its picker listing magpie's models (#322, dakango: switching the Codex app to one of its own models went to OpenAI, 502) — magpie counted an API key in auth.json as a ChatGPT sign-in and only set openai_base_url, counting on Codex to ask the gateway for its models; but Codex asks for none with an API key (models-manager refresh_available_models returns early), so the picker showed Codex's built-in models and the one magpie model, and a built-in one went to OpenAI with the relay's key. Now only a ChatGPT sign-in takes that route; with an API key alone magpie is the model_provider with model_catalog_json listing every magpie model on for Codex, as when signed out, and a sync moves a setup an older magpie wrote; going back to Codex's own model restores the user's model_provider. A Codex model whose OpenAI can't be reached now fails naming the model and what to do instead of a bare 502. TestCodexAPIKeyUsesProvider, TestCodexAPIKeyBaseURLMoved and TestCodexOwnModelUnreached fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass.
agents: Pi's models go to PI_CODING_AGENT_DIR (#304), and the Codex app offers magpie's models' efforts (#310) — Pi keeps settings.json and models.json in PI_CODING_AGENT_DIR when set (~ expanded; @earendil-works/pi-coding-agent config.js), and magpie wrote ~/.pi/agent regardless; omp is read by its own rules (PI_CONFIG_DIR, OMP_PROFILE/PI_PROFILE, then PI_CODING_AGENT_DIR), and the library's Pi and omp targets follow. The Codex app's model picker offers only the efforts in [desktop] enabled-reasoning-efforts of config.toml (ChatGPT.app 26.924: default low…xhigh, ultra, persistent), so a magpie model's max or minimal was never offered; routed through magpie, the efforts its models take are added to the user's list (or the default), nothing else changed, only values the app knows (it drops the setting for any other), comments kept, nothing written when nothing is missing. TestPiAgentDir, TestOmpAgentDir and TestCodexDesktopEfforts fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass.
toml: Codex's config.toml with a multi-line array (notify = [ … ]) is edited right, and an edit that fails leaves every file as it was (#253 kerwinliu00001: "edited TOML is invalid: line 2, column 1: incomplete number", config.toml left with model_provider gone and [model_providers.magpie] added but no model) — the top-level key edits (GetTOMLTop, SetTOMLTop, DelTOMLTop) went line by line with regexes: a new key went after the last line that looked like
key = …, which for notify = [ was inside the array; key = value lines inside a multi-line string were read as keys and a line starting with [ as a table; replacing or deleting a multi-line value touched only its first line. They now locate the top-level keys with the TOML parser, as the table edits already did, each with its value's whole range however many lines it takes (parseTOMLFile, the unnamed root table beside the tables); SetTOMLTop finds all its keys in the file as read and writes once. A file the parser refuses is still read line by line, so a broken config keeps showing its model, and writes to it are refused untouched. And a pick is several writes (Codex: drop model_provider, add magpie's provider table, set openai_base_url and model), each checked, so one failing left the earlier ones on disk: edit.Atomically(fn, paths…) puts each file back to its bytes and mode, or removes one it created, when fn fails, and every field set and Sync of Codex (config.toml and magpie-models.json, so Codex in WSL too), Kimi and Grok goes through it, as do Codex's MCP server writes (putCodex, delCodex). Tested in temp dirs and a temp HOME: the reporter's file read, set and deleted directly and through Codex's model field (signed in, with model_provider and model above the array, with magpie as the provider), the array's text kept and the TOML valid, and back to a Codex model; nested arrays, trailing commas, comments in arrays, [x] and a = b in strings, an inline table in an array; """ and ''' strings holding model = … and [table] lines not read as keys, new keys after their closing quotes, one deleted whole; a forced failure of Codex's last write leaving config.toml byte-identical; TestAtomically. They fail on main (the Codex ones with the reporter's error). go vet ./internal/..., go build -tags nogui (and GOOS=linux, GOOS=windows) and go test -tags nogui ./... pass. Not tried against a real ~/.codex, in WSL or in the GUI; multi-line inline tables (TOML 1.1) are refused by the parser and so left untouched with an error.
codex: a subscription switched off narrows nothing, its models kept whole (switching the codex provider off left the one picked ChatGPT model as the only one Codex listed) — the codex provider's picks narrow the ChatGPT account's own model list as well as magpie's added models (CodexNativePicked,
aed12ac), and ownCodex narrows the picker's OpenAI half the same way. Neither looked at Off, so a provider switched off — which serves no agent anything, by Provider.Off's own promise — still narrowed the account's list to the picks it kept, and a single pick left Codex listing that one model alone. Both now treat a switched-off provider as picking nothing: the account's list is left whole, and the picks stay in providers.json, narrowing again as soon as it is switched on. Tested with a new TestCodexModelListNotNarrowedWhileOff in internal/gateway (a fake backend listing two models; switched off with picks=[gpt-6-sol] both come through, switched on just the pick; fails on main with [gpt-6-sol]) and a new TestCodexOwnModelsNotNarrowedWhileOff in internal/agent (Codex's own picker: both models while off, just the pick while on; fails on main with [gpt-b]). go build -tags nogui ./..., go vet ./internal/agent ./internal/provider ./internal/gateway and go test -tags nogui -count=1 ./... pass (#274)
agents: OpenChamber gets its own row, its model and small model set to magpie's (#321) — OpenChamber keeps its own default model (preferences.json, a copy in settings.json, under $OPENCHAMBER_DATA_DIR or ~/.config/openchamber) and sends it with every session over OpenCode's, so setting OpenCode's model changed nothing it used. Its row writes defaultModel (a fresh updatedAt, defaultVariant cleared as its own settings do) and the small-model override, keeping every other key and refusing a preferences.json it can't read; magpie's provider goes into the OpenCode config it runs on, and stays there while either uses it. OpenCode's config now follows $OPENCODE_CONFIG_DIR: the OpenCode 2 OpenChamber bundles reads only that folder, so what magpie wrote to ~/.config/opencode was never read; OpenCode 1 reads it last, so it wins there too. A project with its own model in OpenChamber is named in the row. Checked against @openchamber/web 2.0.4 on OpenCode 2.0.18 in a sandbox HOME: its settings API returned magpie's model and OpenCode listed magpie's provider. TestOpenCodeConfigDir, TestOpenChamberModel and TestOpenChamberSettingsOnly fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass; click-scroll, old-webkit pass.
codex: with magpie as Codex's provider, openai_base_url points at magpie as well, so a thread started on Codex's built-in provider and reopened sends a magpie model to magpie, not to ChatGPT's backend (#259 isymbo: ChatGPT Desktop, "The 'xiaomi/mimo-v2.6-pro' model is not supported when using Codex with a ChatGPT account")
agents: Codex signed in with only an API key gets magpie as its provider, its picker listing magpie's models (#322, dakango: switching the Codex app to one of its own models went to OpenAI, 502) — magpie counted an API key in auth.json as a ChatGPT sign-in and only set openai_base_url, counting on Codex to ask the gateway for its models; but Codex asks for none with an API key (models-manager refresh_available_models returns early), so the picker showed Codex's built-in models and the one magpie model, and a built-in one went to OpenAI with the relay's key. Now only a ChatGPT sign-in takes that route; with an API key alone magpie is the model_provider with model_catalog_json listing every magpie model on for Codex, as when signed out, and a sync moves a setup an older magpie wrote; going back to Codex's own model restores the user's model_provider. A Codex model whose OpenAI can't be reached now fails naming the model and what to do instead of a bare 502. TestCodexAPIKeyUsesProvider, TestCodexAPIKeyBaseURLMoved and TestCodexOwnModelUnreached fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass.
agent: Command Code — its model picked in magpie like any agent's; a model through magpie adds magpie as a keyless provider in ~/.commandcode/providers.json with the catalog as its models, and sets settings.json's model to magpie/<provider>/<model>; its own models stay as they are, and default takes magpie's entry out; Command Code still wants its own sign-in (#32)
agents: Factory's Droid CLI is an agent magpie sets up (#242 tasselx) — droid keeps its settings in ~/.factory/settings.json ($FACTORY_HOME_OVERRIDE/.factory when set), its BYOK models the customModels array ({model, displayName, baseUrl, apiKey, provider "anthropic" | "openai" (Responses) | "generic-chat-completion-api", maxOutputTokens, noImageSupport …}), picked as "custom:<id>". Picking a catalog model through magpie appends one entry per catalog model after the user's own (so their ids never move), each with an explicit id "custom:magpie/<provider>/<model>", the catalog's label as displayName, the gateway and token
magpie, its context (maxContextLimit) and output when known, noImageSupport always written (droid takes a Chat Completions model as blind otherwise), and the API its provider speaks natively so the gateway relays it as it is: Responses → "openai" at /v1, Messages → "anthropic" at the gateway root (Anthropic's SDK adds /v1), anything else "generic-chat-completion-api" at /v1; and sets sessionDefaultSettings.model. The model shows in magpie as magpie/<ref>; the user's sessionDefaultSettings.model is stashed and put back on reset (a top-level "model", the docs' spelling, is left alone and is what droid falls back to), magpie's entries go and an emptied customModels or sessionDefaultSettings with them; everything else in the file, the user's entries included, is kept byte for byte. The picker offers the user's own custom models by the id droid gives them (its "id", else the displayName with whitespace as dashes and a count of the same name before it), including the legacy ~/.factory/config.json custom_models, which magpie never writes. Sync rewrites magpie's entries as providers change; Check reports a missing entry or a changed baseUrl/apiKey; its requests are Droid's in usage (User-Agent factory-cli/<version>). Icon is Factory's own favicon glyph (factory.ai/favicon.svg); lobehub has none. From the docs: docs.factory.ai/cli/byok/overview (customModels fields, the three providers, config.json legacy snake_case merged under settings.json), /cli/configuration/settings (model, sessionDefaultSettings.*), /droid-exec/overview (custom:<Display-Name>-<index>). Verified by reading the shipped droid 0.229.0 bundle (@factory/cli-linux-x64, unpacked, never installed or run): the zod schema of an entry (also id, index, maxContextLimit, enableThinking, reasoningEffort, bedrock-converse), ids as custom:${displayName.trim().replace(/\s+/g,"-")}-${n} with n counting equal names, an entry's own id winning, the lookup by id, top-level model/reasoningEffort/specModeModel moved into sessionDefaultSettings only when unset there, getModel reading sessionDefaultSettings.model, noImageSupport defaulting by provider, FACTORY_HOME_OVERRIDE, and BYOK requests carrying factory-cli/<version>. Tested in a sandbox HOME (internal/agent/droid_test.go): set/sync/reset on a settings.json as droid leaves it with a Groq BYOK model (its entry byte-identical, the file restored exactly), each protocol's provider and baseUrl, a moved baseUrl caught by Check and healed by Sync, picking the user's own model while on magpie's, no file at all, a top-level model, FACTORY_HOME_OVERRIDE and legacy config.json ids (custom:Local-Qwen-0/-1) offered and never written. go build -tags nogui (and GOOS=linux, GOOS=windows) and go test -tags nogui ./... pass (TestClaudeSignedOut and TestClaudeAccountNamedAsItsLogin flaked under load and pass alone). Not tried with a real droid: whether droid's /model lists the entries, that it reaches the gateway on each protocol, and how it sends thinking for a custom model (magpie sets no reasoningEffort/enableThinking) are unverified.
agents: Droid, Qoder, OpenHanako, WorkBuddy, OpenCode and ZCode no longer get an output limit above the model's context window (#347)
dsh: DeepSeek Harness's thinking takes low as well — off, low, high, max, as llm-deepseek's reasoningEffort has since 0.1.1 (#269) — magpie offered off/high/max from an older dsh. Checked in dsh 0.1.1 to 0.2.0-rc.2 and master, nothing between; in a sandbox HOME dsh 0.2.0-rc.2 with reasoningEffort low sent output_config.effort low to a fake server. TestDshEffort takes low, still refuses medium; fails on the old list.
agents: "Hide all" in an agent's model list keeps its line, and an uninstalled Gemini CLI goes from the Agents page (#356, #230) — the line under an agent ("Showing 0 / N models") came only when a picker still listed one of magpie's models, so hiding them all took away the one way back; it now stays while any is hidden. Gemini CLI was taken as present from what it leaves in ~/.gemini (settings.json, .env, installation_id…), which stays after it is uninstalled, so its row stayed, offering to set up a CLI that can't run; it is now its gemini binary. TestAgentModelLineAllHidden and TestGeminiNotAntigravity fail without the change; agent-models-all-hidden.test.cjs (Chromium, WebKit, en/zh) added. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./internal/agent ./internal/gui ./internal/library ./internal/provider pass; agent-models, click-scroll, old-webkit pass.
agents: "Hide all" in an agent's model list keeps its line, and an uninstalled Gemini CLI goes from the Agents page (#356, #230) — the line under an agent ("Showing 0 / N models") came only when a picker still listed one of magpie's models, so hiding them all took away the one way back; it now stays while any is hidden. Gemini CLI was taken as present from what it leaves in ~/.gemini (settings.json, .env, installation_id…), which stays after it is uninstalled, so its row stayed, offering to set up a CLI that can't run; it is now its gemini binary. TestAgentModelLineAllHidden and TestGeminiNotAntigravity fail without the change; agent-models-all-hidden.test.cjs (Chromium, WebKit, en/zh) added. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./internal/agent ./internal/gui ./internal/library ./internal/provider pass; agent-models, click-scroll, old-webkit pass.
toml: Codex's config.toml with a multi-line array (notify = [ … ]) is edited right, and an edit that fails leaves every file as it was (#253 kerwinliu00001: "edited TOML is invalid: line 2, column 1: incomplete number", config.toml left with model_provider gone and [model_providers.magpie] added but no model) — the top-level key edits (GetTOMLTop, SetTOMLTop, DelTOMLTop) went line by line with regexes: a new key went after the last line that looked like
key = …, which for notify = [ was inside the array; key = value lines inside a multi-line string were read as keys and a line starting with [ as a table; replacing or deleting a multi-line value touched only its first line. They now locate the top-level keys with the TOML parser, as the table edits already did, each with its value's whole range however many lines it takes (parseTOMLFile, the unnamed root table beside the tables); SetTOMLTop finds all its keys in the file as read and writes once. A file the parser refuses is still read line by line, so a broken config keeps showing its model, and writes to it are refused untouched. And a pick is several writes (Codex: drop model_provider, add magpie's provider table, set openai_base_url and model), each checked, so one failing left the earlier ones on disk: edit.Atomically(fn, paths…) puts each file back to its bytes and mode, or removes one it created, when fn fails, and every field set and Sync of Codex (config.toml and magpie-models.json, so Codex in WSL too), Kimi and Grok goes through it, as do Codex's MCP server writes (putCodex, delCodex). Tested in temp dirs and a temp HOME: the reporter's file read, set and deleted directly and through Codex's model field (signed in, with model_provider and model above the array, with magpie as the provider), the array's text kept and the TOML valid, and back to a Codex model; nested arrays, trailing commas, comments in arrays, [x] and a = b in strings, an inline table in an array; """ and ''' strings holding model = … and [table] lines not read as keys, new keys after their closing quotes, one deleted whole; a forced failure of Codex's last write leaving config.toml byte-identical; TestAtomically. They fail on main (the Codex ones with the reporter's error). go vet ./internal/..., go build -tags nogui (and GOOS=linux, GOOS=windows) and go test -tags nogui ./... pass. Not tried against a real ~/.codex, in WSL or in the GUI; multi-line inline tables (TOML 1.1) are refused by the parser and so left untouched with an error.
routing: a group can name the reasoning levels agents are offered, so one member with few doesn't take the rest from the others (#295) — a group offered only the levels every member has, so deepseek-flash (low/high/max) beside models with medium and xhigh left the group without them. Group.Levels (CLI
magpie group set <id> levels=low,medium,high,xhigh, levels= clears; TUI l; GUI group editor's Levels: its models' shared / named) is offered instead in /v1/models, Codex's catalog, Pi and the rest; only levels magpie knows are taken, stored lowest first; a group in a group that names its own offers them for its models; unset is as before. A member without the level asked is sent its nearest, as it already was. The GUI's group save also keeps a family set from the CLI. TestGroupLevelsOverride, TestGroupLevelsReachAgents, TestGroupLevelsNamed (passthrough and translated: medium→high, xhigh→max on the low/high/max member, as asked on the one with all), TestGroupSetLevels and routing-levels.test.cjs (Chromium, WebKit, en/zh) fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass (TestUserPath, a known flake under load, passes alone); click-scroll, old-webkit, routing-effort(-row), routing-levels, settings-groups GUI tests pass.
agents: Droid, Qoder, OpenHanako, WorkBuddy, OpenCode and ZCode no longer get an output limit above the model's context window (#347)
agent: add Hermes Agent (#49) — magpie adds itself to ~/.hermes/config.yaml ($HERMES_HOME) as the provider "magpie" under providers (the gateway as chat completions, the catalog as its models) and points model.provider at it, with the model in model.default; the provider and model the user had are stashed and put back when they pick one of Hermes's own or go back to the default. Hermes sends its own User-Agent only from a recent release on, so the entry sets it for the gateway's usage view
kimi: the new Kimi Code (2.x) is wired in ~/.kimi-code, where it reads its settings — a model picked in magpie never reached it, nor did magpie's models reach its /model (#290, kyzhouxu: Gemini 3.8 flash picked, Kimi Code still starts on the old one). magpie wrote only ~/.kimi/config.toml, the old kimi-cli's; the new CLI copies that over once when it first starts and never reads it again. KimiDir picks $KIMI_CODE_HOME, ~/.kimi-code if there, $KIMI_SHARE_DIR, ~/.kimi if there, else ~/.kimi-code. The new CLI's model tables also say tool_use (kimi-cli refuses the word, so it isn't written there), and its skills go to ~/.agents/skills, which it reads (not ~/.config/agents/skills). Tested with TestKimiCode (fails on main: path .kimi/config.toml, want .kimi-code/config.toml) and TestKimi; in a sandbox HOME the new CLI's
kimi provider list shows magpie and the new default model, kimi doctor finds the config valid, and kimi -p went through a sandbox gateway to a fake upstream asked for gemini-3.8-flash with its 26 tools. go vet, go build -tags nogui (darwin, GOOS=linux, GOOS=windows) and go test -tags nogui ./... pass (flaky tests under load pass alone).
agents: Kimi Code offers a model's thinking levels, and the level picked reaches the model (#333) — the new Kimi Code (2.x) builds its thinking picker from a model's support_efforts and default_effort, which magpie never wrote, so it offered thinking only on or off. magpie now writes the model's levels (none left out: Kimi Code's Off is that) and default_effort high when the model has it, to ~/.kimi-code only; the old kimi-cli's schema has neither. Kimi Code sends the level inside thinking ({"type":"enabled","effort":…}), not as reasoning_effort, so the gateway moves it there on a Chat request that doesn't say reasoning_effort itself, and fits it like any other agent's. Checked with the real kimi CLI (2.1.1) in a sandbox HOME against a fake upstream: kimi doctor passes, /model shows Off Low Medium High Max, a pick reaches upstream as reasoning_effort; unknown values don't make it reject the file. TestKimiEfforts and TestKimiCodeEffortReachesChatVendor fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass.
agents: Kimi Code offers a model's thinking levels, and the level picked reaches the model (#333) — the new Kimi Code (2.x) builds its thinking picker from a model's support_efforts and default_effort, which magpie never wrote, so it offered thinking only on or off. magpie now writes the model's levels (none left out: Kimi Code's Off is that) and default_effort high when the model has it, to ~/.kimi-code only; the old kimi-cli's schema has neither. Kimi Code sends the level inside thinking ({"type":"enabled","effort":…}), not as reasoning_effort, so the gateway moves it there on a Chat request that doesn't say reasoning_effort itself, and fits it like any other agent's. Checked with the real kimi CLI (2.1.1) in a sandbox HOME against a fake upstream: kimi doctor passes, /model shows Off Low Medium High Max, a pick reaches upstream as reasoning_effort; unknown values don't make it reject the file. TestKimiEfforts and TestKimiCodeEffortReachesChatVendor fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass.
provider: Antigravity's models at several levels are one model with those levels (#150: 统一 Antigravity 模型变体与客户端推理档位的映射) — Antigravity lists gemini-3.7-flash-low, -medium and -high as three models, "Gemini 3.7 Flash (Low)" …, so a client picked a level twice: once in the model, once as its reasoning effort, with no telling which won. Now ids that are one id with an effort word after it (-low, -medium, -high, -extra-low, -minimal, -xhigh) are one model, gemini-3.7-flash "Gemini 3.7 Flash", declaring only the levels there are (low,medium,high; Gemini 3.1 Pro low,high), so Codex, Pi and the rest get reasoning levels for it. The level a variant is at is what Antigravity's name says where each name says a different one (gemini-3.5-flash-extra-low is "(Low)", gemini-3.5-flash-low "(Medium)"), else its id's word; a model only named like another (gemini-3-flash-agent, "Gemini 3.5 Flash (High)"; gemini-pro-agent) or with no sibling (gpt-oss-120b-medium) stays as listed. A request asks for the variant the effort picks, an effort it lacks at the nearest level (xhigh is high; medium on low/high is high); with none asked, the id without a level if Antigravity has one, else high — what Gemini 3 does unasked — else the highest. A variant's request carries no thinkingLevel, so the id and the level can't disagree. An old variant id still works: as named with no effort, and moved to the effort asked for. Saved picks and agents set to a variant (antigravity/gemini-3.7-flash-high) are moved to the family, at that level where the agent's own is unset, as Cursor's are. For Gemini 3 models without variants, medium now goes as medium to Flash (which takes minimal/low/medium/high) and still as high to Pro (low/high). Tested against the list an account's fetchAvailableModels gave (sandbox cache)
models: a model's own name given by the user keeps its provider after it too (#203 uclort: 如果在供应商模型页自定义模型名称，在 Codex 中就不会显示供应商了) — Entry.Label() showed "name · provider" only for a model's own (vendor's or models.dev's) name; a name the user gave it (provider's model page → Names & levels) dropped the provider entirely, so a renamed model in Codex's picker (and Pi's, ZCode's, OMP's, magpie's own gateway list — everything magpieModels/Label feeds) could no longer be told apart by vendor at a glance. Label() now appends " · provider" (or "routing group") to a custom name too, the same as it does for the vendor's own, and only leaves it off when the name already contains the provider's (case-insensitively), so a name like "Sol on A" isn't doubled to "Sol on A · A". No settings toggle added: the fix makes the provider suffix unconditional (what the issue actually asked for), so a preference to hide/show it would only reintroduce the gap; there's also no existing "model list" preferences surface to hang it on cheaply — Names & levels only holds the name and reasoning-level pickers. Tested with TestModelName (internal/provider/modelprefs_test.go, sandbox HOME): a custom name now reads "My Sol · A" in Label() and in CodexListed(), and "Sol on A" (already containing the provider) is left alone; and TestModelNameAndLevelsReachAgentFiles (internal/agent/modelprefs_test.go, sandbox HOME) checks the generated Codex model_catalog_json and Pi models.json entries carry "GLM · Relay" for a custom name and don't double it to "GLM via Relay · Relay" when the name already says Relay. go build -tags nogui (and GOOS=linux, GOOS=windows), go test -tags nogui -count=1 ./... pass except TestUserPath and TestClaudeAccountNamedAsItsLogin, which are load-flaky (pass alone, unrelated to this change) same as the pre-existing rtk.go gofmt / TestConfiguredModelsAdvertiseImageInput flakiness
agent: OmO (oh-my-openagent, omo-ai), a fork of Pi that keeps its own ~/.omo/agent, gets its own row: a magpie model sets settings.json's defaultProvider/defaultModel and magpie's provider in models.json there as Pi's does, $OMO_CODING_AGENT_DIR / $SENPI_CODING_AGENT_DIR are honoured, the Library writes its AGENTS.md, mcp.json and skills, and
omo update updates it (#264 rhyme227). Pi's own ~/.pi is left alone. piIn becomes piLike(at, id, name, dir) so both share one wiring. Tested with omo_test.go in agent and library (fail without the change); go vet, go build -tags nogui (darwin, GOOS=linux, GOOS=windows), go test -tags nogui ./... pass.
agent: OmO (oh-my-openagent, omo-ai), a fork of Pi that keeps its own ~/.omo/agent, gets its own row: a magpie model sets settings.json's defaultProvider/defaultModel and magpie's provider in models.json there as Pi's does, $OMO_CODING_AGENT_DIR / $SENPI_CODING_AGENT_DIR are honoured, the Library writes its AGENTS.md, mcp.json and skills, and
omo update updates it (#264 rhyme227). Pi's own ~/.pi is left alone. piIn becomes piLike(at, id, name, dir) so both share one wiring. Tested with omo_test.go in agent and library (fail without the change); go vet, go build -tags nogui (darwin, GOOS=linux, GOOS=windows), go test -tags nogui ./... pass.
omp: each model is asked on its provider's own API (#329) — magpie wrote every model into omp's models.yml as openai-completions, so a Claude's thinking signatures had no place between tool turns. As for Pi, a Responses model now gets api: openai-responses and an Anthropic one api: anthropic-messages at the gateway's /v1/messages, thinking adaptively on Claude 4.6 and later, else on a budget (budget-effort would send output_config.effort, which Sonnet/Haiku 4.5 refuse); groups and the rest stay on Chat. A model offering max made omp 16.3.5 reject the whole file, so every magpie model vanished; max is now offered as xhigh, which the gateway fits to max. Checked with the real omp 16.3.5 in a sandbox HOME against a fake upstream: Claude → POST /v1/messages, deepseek → /v1/chat/completions. TestOmpModelsAskedOnTheirNativeAPI fails without the change; TestOmpMessagesRelayedAsSent guards the relay. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass.
agents: omp's roles pick their own models (subagents, smol, slow), keep their thinking level, and get back what they had before magpie (#325)
agents: omp's roles pick their own models (subagents, smol, slow), keep their thinking level, and get back what they had before magpie (#325)
agents: OpenChamber gets its own row, its model and small model set to magpie's (#321) — OpenChamber keeps its own default model (preferences.json, a copy in settings.json, under $OPENCHAMBER_DATA_DIR or ~/.config/openchamber) and sends it with every session over OpenCode's, so setting OpenCode's model changed nothing it used. Its row writes defaultModel (a fresh updatedAt, defaultVariant cleared as its own settings do) and the small-model override, keeping every other key and refusing a preferences.json it can't read; magpie's provider goes into the OpenCode config it runs on, and stays there while either uses it. OpenCode's config now follows $OPENCODE_CONFIG_DIR: the OpenCode 2 OpenChamber bundles reads only that folder, so what magpie wrote to ~/.config/opencode was never read; OpenCode 1 reads it last, so it wins there too. A project with its own model in OpenChamber is named in the row. Checked against @openchamber/web 2.0.4 on OpenCode 2.0.18 in a sandbox HOME: its settings API returned magpie's model and OpenCode listed magpie's provider. TestOpenCodeConfigDir, TestOpenChamberModel and TestOpenChamberSettingsOnly fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass; click-scroll, old-webkit pass.
agents: OpenChamber gets its own row, its model and small model set to magpie's (#321) — OpenChamber keeps its own default model (preferences.json, a copy in settings.json, under $OPENCHAMBER_DATA_DIR or ~/.config/openchamber) and sends it with every session over OpenCode's, so setting OpenCode's model changed nothing it used. Its row writes defaultModel (a fresh updatedAt, defaultVariant cleared as its own settings do) and the small-model override, keeping every other key and refusing a preferences.json it can't read; magpie's provider goes into the OpenCode config it runs on, and stays there while either uses it. OpenCode's config now follows $OPENCODE_CONFIG_DIR: the OpenCode 2 OpenChamber bundles reads only that folder, so what magpie wrote to ~/.config/opencode was never read; OpenCode 1 reads it last, so it wins there too. A project with its own model in OpenChamber is named in the row. Checked against @openchamber/web 2.0.4 on OpenCode 2.0.18 in a sandbox HOME: its settings API returned magpie's model and OpenCode listed magpie's provider. TestOpenCodeConfigDir, TestOpenChamberModel and TestOpenChamberSettingsOnly fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass; click-scroll, old-webkit pass.
pi: a model's thinkingLevelMap in models.json names every one of Pi's levels, the model's own mapped to its effort and the rest null, so Pi's /thinking offers only what the vendor takes (#243) — magpie wrote only xhigh and max (glm-5.3-flash at low/high/max got {"max":"max"}), and Pi offers every level a map leaves out but xhigh and max, so minimal and medium were offered and sent efforts the vendor turned away. Now glm-5.3-flash gets {"off":null,"minimal":null,"low":"low","medium":null,"high":"high","xhigh":null,"max":"max"}. Pi's off (read from pi-ai 0.99's getSupportedThinkingLevels and its openai-completions, openai-responses and anthropic-messages streams) maps to none for a model that takes none, so off really turns reasoning off; a model without none has off null, since Pi's off asks a Responses model for effort none and leaves a Chat model on the vendor's default; on Anthropic's Messages API off is left out, as before, since there Pi's off sends thinking disabled, which Claude takes, and Claude keeps its xhigh and max. A model whose levels magpie doesn't know still gets no map and reasoning false. Tested with TestPiThinkingLevels (full maps with and without none, Anthropic keeping off, none for unknown) and TestOpenCodeNoResellerLevelsForSwitchModel, which now checks the full map written for a relay's low/high/max model and no map for an unknown one (fails on main with map[max:max]). go vet ./internal/agent, go build -tags nogui (and GOOS=linux), GOOS=windows go build and go test -tags nogui ./... pass (TestUserPath, TestClaudeSignedOut, TestClaudeAccountNamedAsItsLogin, TestClaudeIdentityServedWhileAsked, TestCodexVersion and TestAddSecondOfPreset failed in the full run under load and pass alone). Not tried in a real Pi session.
agent: Pi starts on the model magpie picks, and asks a Messages-only provider's models on Messages (#160 Aimer779; Cleary on Discord)
agents: Pi's models go to PI_CODING_AGENT_DIR (#304), and the Codex app offers magpie's models' efforts (#310) — Pi keeps settings.json and models.json in PI_CODING_AGENT_DIR when set (~ expanded; @earendil-works/pi-coding-agent config.js), and magpie wrote ~/.pi/agent regardless; omp is read by its own rules (PI_CONFIG_DIR, OMP_PROFILE/PI_PROFILE, then PI_CODING_AGENT_DIR), and the library's Pi and omp targets follow. The Codex app's model picker offers only the efforts in [desktop] enabled-reasoning-efforts of config.toml (ChatGPT.app 26.924: default low…xhigh, ultra, persistent), so a magpie model's max or minimal was never offered; routed through magpie, the efforts its models take are added to the user's list (or the default), nothing else changed, only values the app knows (it drops the setting for any other), comments kept, nothing written when nothing is missing. TestPiAgentDir, TestOmpAgentDir and TestCodexDesktopEfforts fail without the change. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass.
agent: Pi starts on the model magpie picks, and asks a Messages-only provider's models on Messages (#160 Aimer779; Cleary on Discord)
pi: a model's thinkingLevelMap in models.json names every one of Pi's levels, the model's own mapped to its effort and the rest null, so Pi's /thinking offers only what the vendor takes (#243) — magpie wrote only xhigh and max (glm-5.3-flash at low/high/max got {"max":"max"}), and Pi offers every level a map leaves out but xhigh and max, so minimal and medium were offered and sent efforts the vendor turned away. Now glm-5.3-flash gets {"off":null,"minimal":null,"low":"low","medium":null,"high":"high","xhigh":null,"max":"max"}. Pi's off (read from pi-ai 0.99's getSupportedThinkingLevels and its openai-completions, openai-responses and anthropic-messages streams) maps to none for a model that takes none, so off really turns reasoning off; a model without none has off null, since Pi's off asks a Responses model for effort none and leaves a Chat model on the vendor's default; on Anthropic's Messages API off is left out, as before, since there Pi's off sends thinking disabled, which Claude takes, and Claude keeps its xhigh and max. A model whose levels magpie doesn't know still gets no map and reasoning false. Tested with TestPiThinkingLevels (full maps with and without none, Anthropic keeping off, none for unknown) and TestOpenCodeNoResellerLevelsForSwitchModel, which now checks the full map written for a relay's low/high/max model and no map for an unknown one (fails on main with map[max:max]). go vet ./internal/agent, go build -tags nogui (and GOOS=linux), GOOS=windows go build and go test -tags nogui ./... pass (TestUserPath, TestClaudeSignedOut, TestClaudeAccountNamedAsItsLogin, TestClaudeIdentityServedWhileAsked, TestCodexVersion and TestAddSecondOfPreset failed in the full run under load and pass alone). Not tried in a real Pi session.
pi: changing the thinking level writes only that level, and only when it changed — with the model on magpie it rebuilt providers.magpie, so a contextWindow (or any other field) edited by hand was replaced. Default still clears defaultThinkingLevel and leaves the model and the provider where they are. Picking a model, and the catalog sync, still refresh the provider. OmO shares this path. (#285)
agent: Pi starts on the model magpie picks, and asks a Messages-only provider's models on Messages (#160 Aimer779; Cleary on Discord)
routing, models: a rerouted request's tooltip names who answered it (#337), and model names can go without their provider (#335) — the Routing page's and the tray panel's rows named the member that answered a rerouted request, but their tooltip still said the path first resolved; it now says the provider · account · model that answered, and on a second line what it was rerouted from. A setting, Settings → Preferences "Provider in model names" or magpie model suffix off (on by default, as before), writes the models listed to Codex and the other agents (OpenCode, Pi, omp…) by their name alone, except two in an agent's list that would read the same, which keep theirs. routing-reroute-title.test.cjs and plain-names.test.cjs (Chromium, WebKit, en/zh), TestPlainNames, TestPlainNamesReachAgents and TestModelSuffixCmd fail without the change; click-scroll, old-webkit, currency, text-size, panel-routing, routing-served, routing-steady pass. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass.
agents: Droid, Qoder, OpenHanako, WorkBuddy, OpenCode and ZCode no longer get an output limit above the model's context window (#347)
agents: omp's roles pick their own models (subagents, smol, slow), keep their thinking level, and get back what they had before magpie (#325)
agents: omp's roles pick their own models (subagents, smol, slow), keep their thinking level, and get back what they had before magpie (#325)
routing, models: a rerouted request's tooltip names who answered it (#337), and model names can go without their provider (#335) — the Routing page's and the tray panel's rows named the member that answered a rerouted request, but their tooltip still said the path first resolved; it now says the provider · account · model that answered, and on a second line what it was rerouted from. A setting, Settings → Preferences "Provider in model names" or magpie model suffix off (on by default, as before), writes the models listed to Codex and the other agents (OpenCode, Pi, omp…) by their name alone, except two in an agent's list that would read the same, which keep theirs. routing-reroute-title.test.cjs and plain-names.test.cjs (Chromium, WebKit, en/zh), TestPlainNames, TestPlainNamesReachAgents and TestModelSuffixCmd fail without the change; click-scroll, old-webkit, currency, text-size, panel-routing, routing-served, routing-steady pass. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass.
agents: Droid, Qoder, OpenHanako, WorkBuddy, OpenCode and ZCode no longer get an output limit above the model's context window (#347)
agents: each agent's CLI shows its version and updates in one click (#202 JoeyMa-zh: Agents 页面增加"更新 CLI"功能 — 检查各 agent CLI 的版本并一键更新) — the Agents page now puts each CLI's installed version after its name (read with
<cli> --version, 10s timeout, no stdin) and, when a newer one is out and magpie can tell how it was installed, a small "Update to x.y.z" pill. How it was installed comes from the resolved binary: the vendor's own installer first (Claude Code's ~/.local/share/claude/versions → claude update, Codex standalone → codex update, ~/.opencode/bin → opencode upgrade), then a global node_modules (bun, pnpm, else npm, run as npm install -g --prefix <prefix> pkg@latest with that prefix's npm, and only where the prefix is writable), then a Homebrew formula or cask allowlisted per agent (brew upgrade [--cask]); anything else shows the version alone, with no button. The latest comes from the npm registry or brew info --json=v2 (no auto-update), cached 6h (15min after a failure); installed versions are cached by the binary's path, size and mtime. GET /api/agents/cli waits at most 3s and says pending, and the page asks again; POST /api/agents/cli/{id} runs the update (one at a time per agent, 10min cap), reads the version again and says so if it did not change, or gives the command's last lines when it fails. The pill goes busy in place and the page never scrolls; not in the tray panel. zh strings added. Tested with a new internal/agent/cliupdate_test.go (version order and parsing; npm/bun/pnpm/brew formula and cask/native installs, a read-only prefix, a local node_modules and unknown binaries in a fake PATH; an update against a fake registry and updater) and a new internal/gui/tests/cli-update.test.cjs in Chromium and WebKit (versions, pills only where magpie can update, row heights, an update with the view scrolled, a failing one, zh). No real update was run. go build (nogui, GOOS=windows, GOOS=linux), go test -tags nogui ./... and the GUI tests pass
agents: each agent's CLI shows its version and updates in one click (#202 JoeyMa-zh: Agents 页面增加"更新 CLI"功能 — 检查各 agent CLI 的版本并一键更新) — the Agents page now puts each CLI's installed version after its name (read with
<cli> --version, 10s timeout, no stdin) and, when a newer one is out and magpie can tell how it was installed, a small "Update to x.y.z" pill. How it was installed comes from the resolved binary: the vendor's own installer first (Claude Code's ~/.local/share/claude/versions → claude update, Codex standalone → codex update, ~/.opencode/bin → opencode upgrade), then a global node_modules (bun, pnpm, else npm, run as npm install -g --prefix <prefix> pkg@latest with that prefix's npm, and only where the prefix is writable), then a Homebrew formula or cask allowlisted per agent (brew upgrade [--cask]); anything else shows the version alone, with no button. The latest comes from the npm registry or brew info --json=v2 (no auto-update), cached 6h (15min after a failure); installed versions are cached by the binary's path, size and mtime. GET /api/agents/cli waits at most 3s and says pending, and the page asks again; POST /api/agents/cli/{id} runs the update (one at a time per agent, 10min cap), reads the version again and says so if it did not change, or gives the command's last lines when it fails. The pill goes busy in place and the page never scrolls; not in the tray panel. zh strings added. Tested with a new internal/agent/cliupdate_test.go (version order and parsing; npm/bun/pnpm/brew formula and cask/native installs, a read-only prefix, a local node_modules and unknown binaries in a fake PATH; an update against a fake registry and updater) and a new internal/gui/tests/cli-update.test.cjs in Chromium and WebKit (versions, pills only where magpie can update, row heights, an update with the view scrolled, a failing one, zh). No real update was run. go build (nogui, GOOS=windows, GOOS=linux), go test -tags nogui ./... and the GUI tests pass
agents: Claude Code's effort reaches Opus 5.5 and Sonnet 5.5, follows each model's levels, and gets an ultracode switch (#351, #352, #353, #354) — Claude Code 2.1.251 on reads Opus 5.5's, Sonnet 5.5's and any newer Claude's effort only from modelSettings.<model>.effortLevel, so the top-level effortLevel magpie wrote left them at medium; it now writes modelSettings for the model (and the top key for the models that still read it), and a reset takes out only its own level. The levels offered follow what Claude Code 2.1.285 sends each model: Opus 4.5 low–high, Opus/Sonnet 4.6 without xhigh, Haiku 4.5, Sonnet 4.5 and older none; a level a model doesn't take is refused. A new ultracode square writes "ultracode": true, on models taking xhigh, and magpie claude ultra no longer writes "ultra" as the model. An open session keeps the effort it started with, so a change says to restart it while Claude Code runs. Checked against real Claude Code 2.1.285 in a sandbox HOME with a fake API recording the effort sent. TestClaudeEffortModelSettings, TestClaudeEffortPerModel, TestClaudeUltracode and TestClaudeEffortNotice fail without the change; claude-effort.test.cjs (Chromium, WebKit, en/zh) added. go vet, go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass; omp-roles, panel-effort, click-scroll, old-webkit pass.
library: agents in a running WSL distro take skills, MCP servers and instructions (#331) — the Library knew only this machine's agent ids, so codex@wsl:Ubuntu and the like were never offered, and the id's colon was refused as an agent's name and can't be in a Windows file name. Claude Code, Codex, Pi and OmO in a distro now get them at their default places under the distro's $HOME through \\wsl.localhost, only while it runs (opening a stopped one's files starts it; its state is kept and asking for it says to start it). Skills go in as copies, made beside and renamed into place and made again when the library's changes, since a Linux agent can't follow a link to a Windows folder (project skills share the same copy code); an MCP server that runs a Windows program (C:\…, *.exe/.cmd/.bat/.ps1) is refused for them with a message; Pi there gets no MCP servers, where its version decides the file. The id's colon is a dot in file and folder names, and setup export/import turns it back. TestWSLTargets, TestWSLSkillsCopied, TestWSLServers and TestWSLInstructions (a fake distro) fail without the change. go vet (and GOOS=windows for agent, library), go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass. Not run against a real distro.
wsl: Pi installed in a WSL distro is found and pointed at magpie there, as Codex in WSL is (#251 X4fun) — magpie looked in each running distro only for Codex (~/.codex,
command -v codex), so Pi's ~/.pi/agent there was never listed. The agents magpie looks for in a distro are now a small registry (wslKinds: id, name, the folder under $HOME, the program, the agent at a place, the restart advice, a stopped distro's options), from which the probe script is built and the distro's agents listed; Codex and Pi are in it. Pi · WSL <distro> (pi@wsl:<distro>) reads and writes ~/.pi/agent/settings.json and models.json through \\wsl.localhost\<distro>, with the same models.json block as Pi on this machine (thinkingLevelMap, reasoning, compat, contextWindow, maxTokens) but its baseUrl, and the Anthropic models' /messages baseUrl, at the gateway as the distro reaches it: 127.0.0.1 when mirrored, Windows' address under NAT, with the same NAT notice as Codex; its config check and refresh keep that address. A stopped distro shows what magpie last saw (wsl.json keeps Pi's fields as pi.model and pi.effort, Codex's bare keys unchanged so existing files still read), reads nothing, and a pick starts it first. Pi on this machine is unchanged (pi(home) is piIn(here(home))). Tested with a new internal/agent/wsl_pi_test.go against fake distros in temp dirs: the probe script and its dir:.pi / bin:pi lines; a distro with Pi only by folder, only by program, with both agents and with neither (ids, order, name, icon, current model and thinking level, the two agents' values kept apart); a pick under NAT and mirrored writing settings.json and models.json equal to Pi's native block but for the address, the check passing and a refresh keeping the distro's address, the real ~/.pi never made; a stopped distro listed from wsl.json with nothing opened, a pick starting then writing. All four fail on main; the Codex WSL tests pass unchanged. go vet ./internal/... (and GOOS=windows for internal/agent), go build -tags nogui (and GOOS=linux, GOOS=windows) and go test -tags nogui ./... pass (TestUserPath and three Claude status tests timed out under load and pass alone). Not tried on real Windows/WSL nor with a real Pi in a distro.
library: agents in a running WSL distro take skills, MCP servers and instructions (#331) — the Library knew only this machine's agent ids, so codex@wsl:Ubuntu and the like were never offered, and the id's colon was refused as an agent's name and can't be in a Windows file name. Claude Code, Codex, Pi and OmO in a distro now get them at their default places under the distro's $HOME through \\wsl.localhost, only while it runs (opening a stopped one's files starts it; its state is kept and asking for it says to start it). Skills go in as copies, made beside and renamed into place and made again when the library's changes, since a Linux agent can't follow a link to a Windows folder (project skills share the same copy code); an MCP server that runs a Windows program (C:\…, *.exe/.cmd/.bat/.ps1) is refused for them with a message; Pi there gets no MCP servers, where its version decides the file. The id's colon is a dot in file and folder names, and setup export/import turns it back. TestWSLTargets, TestWSLSkillsCopied, TestWSLServers and TestWSLInstructions (a fake distro) fail without the change. go vet (and GOOS=windows for agent, library), go build -tags nogui (darwin, GOOS=linux), GOOS=windows go build, go test -tags nogui ./... pass. Not run against a real distro.