MCP preview after the server lost a session (it expired, or the MCP
process restarted):
- Every server state has an id, made when the state is created. The tab
notices a new id even when the version numbers happen to match, and
every push names the state it was based on, so one based on a lost state
is refused, also when it comes before the tab's first poll (the server
recovers the saved file first).
- The tab keeps the newest canvas XML, saved or not. When the server knows
nothing (no file) or exactly what the tab last saved, the canvas wins and
is saved, so edits made while the server was down are kept. Otherwise
the server's diagram (an AI write the tab missed, a cleared document
that was saved) is shown and the tab's copy goes to History.
- Late answers to an old state's push or poll are dropped; a failed push
says the server is unreachable; Download as .drawio saves the canvas.
Settings and server:
- Saved providers this version does not know stay in storage with their
keys, and sending no longer trips over them.
- The desktop "Ollama (Local)" preset with a key goes to local Ollama
again; a server model's Ollama URL variable is read; the admin panel
writes Ollama Cloud's URL for a key without one.
- Provider error texts show again in the desktop app and for EdgeOne.
- .env: a quoted value followed by a comment ending in a quote is read as
dotenv reads it; unquoted values are unchanged.
- Desktop app: the next launch opens the port where a chat was last
saved; a launch elsewhere that saves nothing does not move it, and a
page with no chats lets the next launch try the other port once.
- The Test button no longer stays busy after another tab changed the key.
- A completed append_diagram is no longer undone by an earlier failed
edit's preview; a file read once in vain is saved again once it is read
or gone.
From the first batch's review:
- The admin panel's Test of an entry without a URL now tests the server's
<P>_BASE_URL, where chat sends the entry's key; chat is unchanged (the
first fix rerouted working setups).
- The model list ends downloads that are too large, accepts answers
without a body, and keeps the "redirects are not allowed" explanation.
- A test covers the preview's History rendering.
- Bedrock: a request's AWS region must be a region name. It becomes part
of the endpoint's host name, so a value such as
"us-east-1.attacker.example/" sent the server's bearer token or signed
request to another host.
- MCP preview server: only the preview page itself (Origin equal to the
Host) or a non-browser client may call it; a page on another localhost
port could replace the diagram with a plain text POST. History builds
its thumbnails element by element and shows only SVG data images, so a
stored value can no longer run script in the preview.
- chat, validate-model, validate-diagram, provider-models and parse-url
take JSON bodies only, so another website cannot make the user's own
server (the desktop app, a local install) run models with their keys;
the desktop app also refuses a foreign Host (DNS rebinding).
- The model list reads at most 2 MB, also through the Gateway SDK, and
answers only with its own error texts: the URL is the caller's and may
be an internal address.
- An admin panel provider with its own key and no URL no longer inherits
the global <P>_BASE_URL, which may be a proxy for another key; OpenAI
then gets the official endpoint, as its Test. Azure keeps the server's
resource.
- Bedrock: a user's AWS keys no longer go to an endpoint the server sets
in AWS_ENDPOINT_URL_BEDROCK_RUNTIME / AWS_ENDPOINT_URL (read by the
upgraded SDK), and admin panel keys win over AWS_BEARER_TOKEN_BEDROCK,
as the Test button checks them. Checked with Bedrock.
- Ollama: a server key without a base URL (admin panel, OLLAMA_API_KEY)
goes to Ollama Cloud, as env.example says, instead of 127.0.0.1.
- Quota: EdgeOne counts whatever key header comes along, keyless Ollama at
a private address counts, and their provider texts stay in the log.
- An EdgeOne server model (admin panel, ai-models.json) works: the route
checked the raw provider header, which holds the name's slug.
- parse-url ends downloads it does not read (too large, PDF, errors).
- Desktop app: the port follows where the chats are (IndexedDB per
origin) instead of a remembered port, which could hide them for good;
a same-port restart tells the page to refetch the server models; a
failed preset switch no longer undoes a newer choice; a presets file
removed after a failed read can be saved again; .env values quoted from
start to end keep their inner quotes, as dotenv reads them.
Found by the second PR review:
- With AWS_BEARER_TOKEN_BEDROCK set on the server, a request with the
user's AWS keys ran on the server's token: the Bedrock SDK prefers it.
Checked with Bedrock: invalid user keys used to get an answer.
- An OpenAI key with the official URL filled in (the settings form does
that) went to the Responses API. Back to main's rule: a configured base
URL uses Chat Completions.
- A user's Ollama key went to the server's OLLAMA_BASE_URL, for chat and
for the model list. Like every other provider, it goes to the user's
base URL or Ollama Cloud.
- The server's keyless Ollama and EdgeOne were not counted in the quota.
- AI_MODEL models on the server's keys ran on any provider with a server
key, not only on AI_PROVIDER.
- A user's Azure key without a base URL used the server's resource name.
- The admin panel's Test button failed whenever access codes were set.
- DeepSeek's errors in the stream (plain text) were shown as they were,
without a hint and also on the server's keys. Bedrock's throttling in
the stream was not recognised as a rate limit.
- The EdgeOne function accepted text/plain; x=application/json, which
other sites can send without a CORS preflight.
- Desktop app: a launch that found the old port taken for a moment (the
previous version still quitting after an update) remembered the new
port for good. The new port is kept only when Windows reserves the old
one. A failed read of the presets file moved it aside as corrupt, and a
save could then replace the presets. Switching presets on the same port
now reloads the page. The dev launcher no longer misses a preset change
made before or during a restart.
Found by the PR review, each with a test that failed first:
- Quota: any key header skipped it, even one the provider never reads
(x-aws-access-key-id with OpenAI), so a request ran on the server's
key without being counted. The check now runs after the model is
resolved and uses usesServerCredentials. On main already.
- usesServerCredentials read the raw base URL; "/" cleans up to none, so
an Ollama request ran on the server's key past the server-model check.
- SGLang's default 127.0.0.1:8000 only fills the settings form. Chat and
the model list used it as a real address, so the server called its own
machine even with private URLs blocked. Now a base URL is required.
- With a user's OpenAI key and no base URL, the SDK read the server's
OPENAI_BASE_URL. The official endpoint is now passed. On main already.
- The Test button refused nothing on the server's keys (Ollama Cloud),
and a 15 s timeout reported "connected, no tool call".
- The model list for Ollama without a base URL came from ollama.com while
chat went to the server's Ollama.
- Bedrock's "Too many tokens, please wait" counted as context too long.
- On the server's keys the provider's error text stays in the server log;
it can name the server's AWS account, role or internal hosts.
- Desktop app: the preset keys are the user's own (NEXT_AI_DRAWIO_DESKTOP),
so Max Output Tokens can be raised and keyless models in settings work
again. A launch that found the remembered port taken no longer replaces
it, which hid the user's chats and settings for good.
- Claude Opus 5.5 sent an edit with invalid JSON, then the same edit
again. The first call's streamed preview was never undone: its input
has no operations, and the undo sat behind that check. The second edit
then started from the preview, failed on a duplicate id, and the model
had to try a third time. The undo now runs first, and an edit that
starts in the same render uses the undone diagram.
- The SDK passes an invalid tool call's error as a string, which was
wrapped as a provider error. streamErrorText keeps it as the text the
model reads.
- Bedrock's "on-demand throughput isn't supported" gets the model id hint.
- The thinking header uses the page language ("Thought for 1 second" in
English), from the dictionary entries that were already there.
- An error object sent inside the stream (OpenRouter's { code, message })
showed as "[object Object]"; its message and status code are read now.
- A problem+json "detail" is added to the message: NVIDIA only said
"Gone" for a retired model. 410 counts as model not found.
- "Cannot connect to API" from the SDK gets the connection hint.
- Text that is only whitespace (Kimi K2.6 sends a space before a tool
call) no longer shows an empty bubble.
- allowSystemInMessages stops the warning on every request. Our system
messages carry cache points; a client's own system messages are already
dropped by the empty-content filter.
The "Fetch models" button asks the provider for its models (OpenAI-style
/models, Anthropic, Google, Ollama, OpenRouter, Vercel Gateway, AIHubMix)
and shows them in a searchable picker. This replaces the route that only
worked for AIHubMix.
A snapshot of models.dev (MIT) says which models support tool calls.
Models without them get a "no tool calls" badge in the picker and a hint
in the model list, since drawing needs tool calls. Refresh the snapshot
with scripts/update-model-catalog.mjs.
- lib/llm-errors.ts sorts an error into about a dozen kinds (key
rejected, no access, unknown model, no credit, rate limited, context
too long, no image input, no tool calls, output cut off, provider down,
cannot connect, timeout): first texts that name the cause precisely,
then the HTTP status code, then general texts. It unwraps RetryError and
hides keys and Bearer tokens in the provider's message
- The chat route uses it for errors before the stream and, through
toUIMessageStreamResponse's onError, for errors in the stream. Errors
of the model's own tool call stay as they are: the same text goes back
to the model so it can fix the call
- The chat shows the hint in the user's language, then the provider's
message; a rejected key, missing access or unknown model adds an "Open
model settings" button. The Test button shows the same hints
- Fixes: our message "API key is required when using a custom base URL"
was replaced by "Authentication failed" because it contains "key"; a
provider's "Rate limit exceeded" opened this site's quota toast; an
error body like {"error": ...} was shown as raw JSON; the Test button
matched "401" in the message, where providers rarely put it
- Remove the string matching fallbacks in the chat panel
- getAIModel resolves credentials (client key, server env vars, the
existing SSRF rules) and createModel builds the model by SDK. The
provider-by-provider switch shrinks from 24 cases to the few that
differ (lib/ai-providers.ts 1531 -> 1106 lines)
- /api/validate-model calls getAIModel instead of its own 24-case switch
(503 -> 175 lines), which had drifted from the chat: it built Azure
with createOpenAI, Kimi and MiMo with createOpenAI instead of
createDeepSeek, and the official OpenAI endpoint with Chat Completions.
A passing test now means the chat works
- Plain OpenAI-compatible providers (SiliconFlow, SGLang, ModelScope,
GLM, Qwen, Qiniu, Novita, Atlas Cloud, EdgeOne, Doubao, MiniMax in
OpenAI mode, AIHubMix on a custom URL) use @ai-sdk/openai-compatible,
which reads reasoning_content, so their reasoning shows, and accepts
SGLang's stream as is (its 95-line stream rewrite is gone).
includeUsage keeps token usage for quotas. <think> tags in their text
become reasoning (extractReasoningMiddleware)
- SGLang without a base URL used OpenAI's endpoint; it now defaults to
http://127.0.0.1:8000/v1 like the Test button did
- Chat requests to a client base URL refuse redirects, as the Test
button already did (redirectGuardedFetch moves to lib/ssrf-protection)
- The Test button streams like the chat (the ModelScope special case is
gone), times out after 15 s, does not retry, asks the model to call a
ping tool and warns when it answers without one, and tests all models
at once. The time each test took shows on its check mark
- Unknown provider names are rejected with Object.hasOwn, and the error
texts list providers from PROVIDER_INFO instead of hand-kept lists
- useChat throttles streamed message updates (experimental_throttle,
150 ms), replacing the two hand-written 150 ms timers of the
display_diagram and edit_diagram previews (94 lines less). The preview
now only runs while the input streams; once it is complete the tool
handler takes over, so a queued preview can no longer redraw an edit
the handler rejected and rolled back. Measured on a streamed 60-cell
diagram: 41 redraws at least 97 ms apart, before 37 with gaps down to
48 ms
- The diagram check endpoint uses streamText with Output.object instead
of the deprecated streamObject, and returns its fixed result as a plain
text response; new route test
- Import createGateway/gateway from ai and drop the direct
@ai-sdk/gateway dependency
- The per-request message structure logs only print with
DEBUG_LLM_PAYLOAD=true
- Remove an empty onFinish callback
- Update ai to 6.0.300 and the @ai-sdk providers to their latest v6-line
versions. @ai-sdk/anthropic 3.0.47 did not know claude-opus-4-7/4-8
and capped their output at 32000 tokens; 3.0.127 allows 128000
- Drop the fine-grained-tool-streaming beta header for the Anthropic API:
the provider now streams tool input per tool (eager_input_streaming)
- Claude 4.7 and later reject a non-default temperature/top_p/top_k and
the extended thinking budget with a 400. A middleware retries once
without them, so TEMPERATURE and *_THINKING_BUDGET_TOKENS no longer
break those models
- Prompt caching also reaches Claude on the Anthropic API and OpenRouter;
before, only Bedrock got a cache marker
- GOOGLE_TOP_K and GOOGLE_TOP_P never reached Gemini: they were sent as
Google provider options, which drops them. They are call settings now.
GOOGLE_CANDIDATE_COUNT and GOOGLE_REASONING_EFFORT, which the provider
does not support, are removed
- Add @ai-sdk/openai-compatible as a direct dependency
- The three XML examples (swimlanes, two edges, waypoints) that the web
system prompt, the display_diagram description and the MCP drawing
guide each had a copy of now live in packages/mcp-server/src/xml-examples.ts
- The shape library group list and reader come from the MCP
shape-library.ts; getShapeLibrary takes the folder, and the web route
passes docs/shape-libraries under the working directory as before.
Only the 30 known library names are read. Error texts differ slightly
- next.config.ts traces docs/shape-libraries/*.md for /api/chat, since
the read now happens in another module
- Every prompt is byte-for-byte unchanged: the web system prompt for five
model ids with and without minimal style, and the MCP drawing guide,
compared before and after; the library list and swimlane example match
the old tool description text exactly
- edit_diagram runs the MCP server's editDiagram: every new_xml is checked
first, one cell per operation, and after the edit only the target page
is checked, rejecting only errors this edit introduced. An unrelated
problem elsewhere in the document no longer blocks every edit. The
error lists each failed operation
- The streaming edit preview uses the MCP applyDiagramOperations
- Delete applyDiagramOperations (292 lines) and wrapWithMxFile from
lib/utils.ts, and the unused hand-copied scripts/test-diagram-operations.mjs
- One blank document (BLANK_MXFILE) for the web app and the MCP preview,
replacing four copies
- Saving a .drawio wraps a bare model with normalizeToMxfile
- The empty-diagram check uses hasCells, which also counts cells wrapped
in a UserObject/object
- DiagramOperation is the MCP type
- The wrapped-cell and empty-diagram tests now run against the MCP code
- New e2e test: edit_diagram changes the canvas, and a failing edit
leaves it as it was
- /api/log-save accepted only drawio, png and svg, so saving as Editable
SVG (xmlsvg) got a 400 and was never recorded in Langfuse
- The default download name and the template export name used the UTC
date, which is the previous day on mornings in East Asia
- Remove the Electron openFile/saveFile IPC handlers, their preload
bindings and types; nothing in the app calls them
- Shared checkAccessCode for validate-diagram, validate-model, parse-url, verify-access-code
- parse-url: 5 MB streamed body limit; validate-diagram: 5 MB image limit
- validate-model refuses redirects when private URLs are blocked
- Admin settings state shared across module instances via globalThis
- Server model ids: unique slugs (non-ASCII names encoded), duplicates rejected
- Panel Bedrock credentials stored as ADMIN_AWS_* so the DynamoDB client keeps its own
- Locale redirect keeps basePath and query; EdgeOne function drops open CORS and checks the access code
- Providers payload reports whether .env sets a default model
- Vertex: a client-supplied base URL only works with the client's own Vertex key
- Accept only data: URLs for file parts in every message, so the server never downloads them
- Output budget retry accounts for the thinking budget Bedrock/Anthropic add, and reads
Volcengine, DashScope, SGLang and vLLM rejections; falls back to 16000 once
- x-max-output-tokens can only lower the budget on server credentials
- On server credentials only server models or AI_MODEL entries can be used
- Drop tool results together with the invalid tool calls they belong to
- Count quota tokens as input + output (cached tokens were counted twice)
- Private-URL check for custom base URLs, end Langfuse traces on error/abort/early return
- Fix repairToolCall ordering and placeholder, align edit_diagram prompt with operations
- Panel Bedrock keys are read from ADMIN_AWS_*; forward the access code to EdgeOne
- isMinimalDiagram only treats root cells as an empty canvas
* fix: raise the output budget so reasoning models reach the tool call
A reasoning model spends the output budget in order: thinking first, then prose,
then the tool call. With 16000 the thinking alone can consume all of it, so the
turn ends with finishReason "length" before display_diagram is ever called. The
canvas stays empty and nothing surfaces in the UI, because no tool call means no
tool error, and the client never reads finishReason.
Measured on openrouter deepseek/deepseek-v4-flash, the model from the report:
- max_tokens=800 with reasoning on returns reasoning_tokens=800, empty content,
finish_reason length. So reasoning is billed against this budget, not exempt.
- refining an existing diagram (19k chars of XML in the input) produced 49142
chars of reasoning, zero tool calls, finishReason "length" at 16000
- the same request at 40000 finished and called edit_diagram with 12 operations
64000 cannot just be sent to every model: bedrock claude-3-haiku caps at 4096,
nova-lite at 10000, and the openrouter deepseek-r1 endpoint counts input and
output against one 64000 ceiling. All three name the real limit in the 400, so
parse it and retry once. Verified: nova-lite logs "64000 rejected, retrying with
10000" and then completes its tool call.
Also expose the budget in Settings. It is sent as a header rather than read from
env only, so desktop users can raise it themselves without an env file.
vercel.json goes back to the 300s it had before #238 traded it for $2-4/month.
That is now Vercel's own default, and billing pauses while the function waits on
the model, so the saving that motivated 120s no longer applies. edgeone.json is
left alone: its 120 may be that platform's actual ceiling.
* fix: only reinterpret an error as a budget rejection when it says so
Review of the first commit found the retry could fire on errors that have
nothing to do with the budget, which would replace a readable provider error
with a truncated response: exactly the symptom this PR exists to remove.
- Drop the generic "lower than N" pattern. For the Bedrock message it was dead
code, since "model limit of N" matches first with the same number. Left live,
it would read a number out of any message shaped like "must be lower than 2".
- Skip errors whose status is not 400 or 422, so auth and rate-limit failures
are never reinterpreted.
- Require the parsed ceiling to be at least 1024. Below that a diagram cannot
come out whole, so retrying would hide the error behind broken XML.
- Validate MAX_OUTPUT_TOKENS from env the same way as the header, so a stray
"-1" falls back instead of reaching the provider.
Adds tests for the retry wrapper itself, which had none: it retries once with
the named ceiling, leaves a 401 alone, does not retry when the ceiling is not
smaller, propagates a second rejection, and preserves the other call options.
Re-verified against the live APIs: bedrock nova-lite still logs "64000 rejected,
retrying with 10000" and completes its tool call, and deepseek-v4-flash still
finishes normally at 64000.
Unset does not mean the model's maximum — the provider fills in its own, and
Bedrock's is 4096 (measured: converse with no inferenceConfig on
us.anthropic.claude-opus-5 returns stopReason=max_tokens at exactly 4096).
A 30-cell diagram is ~3000 tokens of XML, so anything larger arrived as truncated
JSON and nothing reached the canvas. Small diagrams fitted, which made it look
intermittent.
* feat: add MiMo (Xiaomi) as AI provider
* fix: correct MiMo default base URL, suggested models, and reasoning support
- Default base URL was the China Token Plan endpoint (tp- keys only);
switch to https://api.xiaomimimo.com/v1 which works with standard
pay-as-you-go sk- keys. Token Plan users can override in settings.
- Replace deprecated mimo-v2-flash suggestion with mimo-v2.5
(v2 series was deprecated on 2026-06-30).
- Use createDeepSeek instead of createOpenAI so reasoning_content is
passed back during multi-turn tool calls (MiMo returns 400 without
it), matching the existing Kimi implementation.
- Add mimo to SINGLE_SYSTEM_PROVIDERS so system messages are merged.
- Fold validate-model case into the shared OpenAI-compatible group.
- Drop the Bot icon special case; models.dev serves a real xiaomi logo
via PROVIDER_LOGO_MAP.
- Document MIMO_API_KEY/MIMO_BASE_URL in env.example and
docs/{en,cn,ja}/ai-providers.md.
* feat: show base URL hint for MiMo in provider settings
MiMo has two endpoints tied to key type: pay-as-you-go keys (sk-...)
use the default api.xiaomimimo.com/v1, while Token Plan keys (tp-...)
require token-plan-cn.xiaomimimo.com/v1. Surface this under the Base
URL field like the existing MiniMax hint, in all four locales.
---------
Co-authored-by: mapengfei <[email protected]>
Co-authored-by: dayuan.jiang <[email protected]>
* fix: resolve DNS before SSRF check and block redirects in parse-url
isPrivateUrl() did string-only hostname matching and never resolved DNS,
so a public-looking name that maps to an internal IP (e.g.
127-0-0-1.sslip.io -> 127.0.0.1) passed the check while fetch/extract
later resolved it and reached internal services (GHSA-wqcv-5qvx-vx75).
- isPrivateUrl is now async: it keeps the fast string/literal-IP path,
then resolves the hostname via DNS and rejects if any address is private.
- parse-url now fetches the page itself with redirect: "error" and parses
via extractFromHtml(), since article-extractor follows redirects
internally and drops a redirect option, which allowed a public URL to
302 to an internal host.
- Update validate-model call site to await; add regression tests.
* fix: preserve charset detection and block CGNAT range in parse-url SSRF fix
Follow-up to the multi-reviewer review of the SSRF fix:
- Restore charset handling lost when switching from extract() to
response.text(): non-UTF-8 pages (Shift_JIS/GBK/EUC/Big5, common on CJK
sites this project targets) decoded as mojibake. Now read the body as
bytes, detect charset from Content-Type / <meta charset>, and decode
with TextDecoder before extractFromHtml.
- Wrap extractFromHtml in try/catch: it throws (not returns null) on
empty/non-HTML bodies, which previously surfaced as a 500 instead of the
intended 400.
- Add 100.64.0.0/10 (RFC 6598 CGNAT) to isPrivateIp; it is routable inside
some cloud internal networks and was a residual SSRF target.
- Add tests for CGNAT, its boundaries, 0.0.0.0, and DNS-resolved IPv6.
supportsImageInput() guessed multimodal capability from the model id
string. The heuristic misfired on newer models (e.g. kimi-k3.6, qwen36),
either wrongly rejecting images for capable models or letting them through.
The AI SDK does not emit a warning when an OpenAI-compatible endpoint
silently drops an image, so the guess was the only signal — but an
unreliable one. Drop the detection entirely and let the real provider
error surface instead (already translated to a friendly message in
chat-panel.tsx). Validation falls back to "valid" on any model error.
- Remove supportsImageInput() and its pre-send check in chat route
- Drop the vision-capability throw in getValidationModel()
- Remove the corresponding unit tests
* feat: add file-based admin settings panel at /admin
Settings saved in the panel are written to data/settings.json and
overlaid onto process.env, taking precedence over environment
variables and applying immediately without restart. Enable by setting
ADMIN_PASSWORD; on serverless platforms without persistent disk the
panel degrades to read-only.
* polish: admin panel UI improvements
- Provider logos in credential rows (shared ProviderLogo component,
extracted from model-config-dialog)
- Scroll-spy active state in the sidebar nav
- Green success state in the save bar that clears after a few seconds
- Wider content column (max-w-6xl) for less wasted space on desktop
* polish: admin panel section toggles and reorder
- Move Quota & Rate Limits to the end of the settings page
- Add enable switches to Observability and Quota sections; default off
with fields grayed out, auto-on when any field is already configured
* polish: make section enable switch more visible
Wrap the switch in a labeled pill ('Enabled'/'Disabled') with border
and background so the off state is clearly visible.
* refactor: derive admin registry from PROVIDER_INFO, simplify page state
- Provider options, labels, and base-URL placeholders now come from
PROVIDER_INFO instead of hand-copied lists (fixes SiliconFlow .com/.cn
placeholder drift; panel names now match the model-config dialog)
- Replace free-text subgroup strings + SUBGROUP_PROVIDERS reverse map
with a typed provider field on SettingDef
- Precompute SETTINGS_BY_GROUP and PROVIDER_SUBGROUPS at module level
- Merge justSaved into saveMessage, drop unused mainRef, hoist
fetchSettings out of the component, dedupe savedText logic
- Serialize from SETTINGS_REGISTRY directly; json validators in a map
instead of a hardcoded key check
- Make allowPrivateUrls a function so ALLOW_PRIVATE_URLS edits in the
admin panel apply without restart
* feat: graphical model management in admin panel
Replace the provider credential fields and raw AI_MODELS_CONFIG JSON
textarea with a Models section mirroring the in-app model settings UI:
provider instance list with logos, credential fields per provider type,
model add/remove with suggestions, per-model connectivity test, and a
default-provider star.
On save the server derives everything the runtime needs into
settings.json: credential env vars (with _2 suffixes for multiple
instances of one provider), AI_MODELS_CONFIG, and AI_PROVIDER/AI_MODEL
for the default. Secrets round-trip as masked markers and are never
sent back to the browser. The general settings registry now only
covers non-provider settings (generation, access, features,
observability, quota).
* fix: allow testing unsaved providers in admin panel
The test button previously looked up credentials by providerId in the
saved settings, so testing a newly added (unsaved) provider failed with
'Unknown provider or model'. The test endpoint now accepts the client's
current provider state; newly typed secrets are used as-is and masked
markers are resolved against the stored values, so testing works both
before and after saving.
* fix: merge env AI_MODELS_CONFIG with admin panel providers
Previously, saving in the admin panel wrote a complete AI_MODELS_CONFIG
into settings.json, which (by overlay precedence) replaced any config
from .env or ai-models.json — admins lost their env-configured models.
The panel no longer writes AI_MODELS_CONFIG. Instead its providers are
merged with the env baseline at read time in loadRawServerModelsConfig,
and panel credentials go to ADMIN_-prefixed env vars wired up via
apiKeyEnv/baseUrlEnv so they never shadow standard vars. Env-based
providers now appear read-only in the panel, name clashes are rejected,
and a panel default overrides the env default. data/ is now gitignored.
* fix: block global-credential providers already managed via env
Bedrock, Vertex AI, and Ollama credentials live in fixed env vars with
no apiKeyEnv redirection, so a panel instance of one of these would
silently override the credentials that env-configured models rely on.
The API now rejects saving such a provider when the env config already
uses that type, and the Add Provider dropdown disables it with a
'managed via env' note.
* fix: address admin panel review findings
- Security: test-model no longer resolves a stored secret when the
request's baseUrl/provider differs from the stored entry, closing a
path where a tampered baseUrl could exfiltrate a saved key
- Save failures are now visible: the save bar shows the error in red
(was masked by the persistent 'Unsaved changes' text), and per-field
validation errors from the settings API are surfaced under each field
- The Observability/Quota enable switch is now real: toggling off stages
deletion of the group's saved values, and the toggle no longer snaps
back to Enabled after saving
- Env provider's default star is hidden when a panel provider is the
active default (no more double star)
- Clearing a credential field reverts to the stored value instead of
silently deleting it; an explicit X button removes a stored secret
- Form inputs are disabled during an in-flight save
* refactor(admin): split 1549-line admin page into focused modules
Extract admin-shared.ts (types + fetch helper), setting-field.tsx
(registry-driven fields), and models-section.tsx (provider/model
manager) from page.tsx. Pure mechanical move, no behavior change.
* feat(admin): share credential fields with user dialog and localize panel
Extract ProviderCredentialsFields (display name + per-provider
credential inputs) used by both the user ModelConfigDialog and the
admin Models panel; secret input passed via renderSecret (plaintext
vs masked), test button via footer slot. Add full i18n for the admin
panel across en/zh/ja/zh-Hant, reusing modelConfig.* for shared parts.
* fix(admin): address Copilot review findings
- Reflect built-in defaults for boolean settings (ALLOW_PRIVATE_URLS
defaults on) and allow clearing a saved boolean back to default,
so the SSRF toggle matches actual runtime behavior.
- Harden JSON loading: filter settings values to strings only, and
schema-validate stored ADMIN_PROVIDERS entries, dropping malformed
ones instead of letting them reach runtime code.
- Set beforeunload returnValue so the unsaved-changes prompt shows in
all browsers; reject non-finite numbers in settings validation.
- Fix README/CN/JA docs that claimed the panel auto-generates
AI_MODELS_CONFIG (providers are merged at read time, not written).
- Add unit tests for corrupted-file value filtering and provider
schema validation.
* docs: move admin panel details to dedicated docs/{en,cn,ja}/admin-panel.md
The READMEs now carry a short blurb + link, matching the existing
per-topic docs (docker.md, ai-providers.md, ...). Removes the ~22-line
inline section and the duplicated data/settings.json mentions.
* fix(admin): address follow-up Copilot findings on the prior fixes
- loadAdminProviders now validates against a stored-shape schema where
secrets are plain strings, so a hand-edited ADMIN_PROVIDERS holding an
{isSet} marker is dropped instead of later crashing maskSecret().
- loadSettings guards against array values (typeof [] === 'object'),
which would otherwise overlay numeric keys onto process.env.
- Admin SecretInput uses the bare id so the shared component's
<Label htmlFor> stays associated (only one ProviderDetail mounts).
- Add tests: marker-secret rejection, array-values guard, bedrock
multi-secret round-trip.
/api/parse-url accepted any URL the user submitted, fetched it via
@extractus/article-extractor, and returned the body as Markdown. With
ALLOW_PRIVATE_URLS unset (the default after #600) the SSRF guard
short-circuited entirely, so an unauthenticated POST could probe
container ports, read AWS IMDS / GCP metadata, and reach same-VPC
internal services.
- parse-url now always rejects private URLs regardless of
ALLOW_PRIVATE_URLS. The flag's only legitimate use case is local
LLM provider baseUrl overrides (validate-model, chat); article
extraction has no business fetching internal hosts. Local LLM
setups (Ollama, LM Studio, etc.) are unaffected.
- Strip a trailing dot from the hostname before equality checks so
the FQDN form "localhost." (which still resolves to 127.0.0.1) is
caught by the existing string match.
Known follow-ups (not addressed here):
- DNS rebinding: hostnames are matched as strings; a public domain
resolving to 127.0.0.1 (e.g. localtest.me) is not caught.
- HTTP redirects: @extractus/article-extractor uses cross-fetch with
default redirect: "follow" and exposes no hook, so a public URL
302-ing to an internal host still leaks.
* fix: merge system messages for custom OpenAI endpoints (fixes#734)
When using the OpenAI provider with a custom base URL (e.g., vLLM, LMStudio),
the app sends two system messages to the API. Open-source model chat templates
(Qwen, Llama, etc.) enforce that system messages must appear at the beginning
and reject multiple system message blocks, causing the error:
'System message must be at the beginning.'
Treat custom OpenAI endpoints (client-provided base URL or OPENAI_BASE_URL env
var) the same as other known single-system providers by merging both system
messages into one before sending.
* fix: also detect custom OpenAI endpoint from serverModelConfig.baseUrlEnv
---------
Co-authored-by: dayuan.jiang <[email protected]>
Adds the missing 'novita' case to the OpenAI-compatible provider
block in the validate-model API route, fixing 400 errors when
users test their Novita API key in the UI.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* feat: add custom system message setting for AI personalization
Allow users to enter custom instructions via a textarea in Settings
that get appended to the AI's system prompt. Includes server-side
validation (type check + 5000 char limit), localStorage persistence,
and i18n support for all 4 locales.
* fix: add accessibility htmlFor/id pairing on custom system message textarea
* feat: Add support for Chinese AI providers (GLM, Qwen, Kimi, MiniMax, Qiniu)
- Add minimax, glm, qwen, qiniu, kimi to ProviderName type
- Add provider configurations to PROVIDER_INFO with default base URLs
- Add suggested models for MiniMax in SUGGESTED_MODELS
- Add minimax/glm/qwen/qiniu/kimi cases to getAIModel using OpenAI-compatible SDK
- Update ALLOWED_CLIENT_PROVIDERS and error messages
- Add environment variable examples to env.example
Fixes: MiniMax API compatibility issue (invalid chat setting 2013)
* fix: Add missing providers to PROVIDER_ENV_VARS type
* fix: Handle null case in PROVIDER_ENV_VARS for new providers
* fix: Add minimax/glm/qwen/kimi/qiniu support to validate-model API
- Add getDefaultBaseUrl helper function
- Add validation cases for new providers in validate-model route
* fix: Add new providers to buildProviderOptions switch case
* fix: Merge multiple system messages into one for minimax/glm/qwen/kimi/qiniu
MiniMax API doesn't support multiple system messages.
This fix combines them into a single message for Chinese providers.
* fix: Handle null provider in system message check
* debug: Add logging for allMessages count
* fix: Use effective provider (including env var fallback) for isSingleSystemProvider check
* fix: apply biome formatting (line-wrapping)
* docs: add Chinese AI providers documentation (MiniMax, GLM, Qwen, Kimi, Qiniu)
- Add i18n translations for new providers in all language dictionaries
- Add provider configuration documentation in en/cn/ja docs
* fix: 改进 PR #722 的代码审查反馈
1. 删除重复的 getDefaultBaseUrl 函数，改用 model-config.ts 的 PROVIDER_INFO
2. validate-model 路由改用 AI SDK 的 createOpenAI + generateText
3. 修复 resolveBaseURL 回退逻辑，传入 PROVIDER_INFO 的 defaultBaseUrl
4. 删除无用的 .bak 备份文件
Co-authored-by: Shinyi <[email protected]>
* fix: 修正中国 AI provider 端点配置
- qiniu: api.qiniucdn.com → api.qnaigc.com
- qwen: dashscope.aliyun.com → dashscope.aliyuncs.com
- 更新 env.example 文档链接
Co-authored-by: Shinyi <[email protected]>
* feat: MiniMax 使用 Anthropic 兼容 API
- MiniMax 改用 createAnthropic (而非 createOpenAI)
- 支持 api.minimax.io/anthropic 和 api.minimaxi.com/anthropic
- 合并多个 system 消息为单个 (MiniMax/GLM/Qwen/Kimi/Qiniu)
- 更新默认模型为 MiniMax-M2.5 系列
- 支持 MINIMAX_BASE_URL 环境变量配置
Co-authored-by: Shinyi <[email protected]>
* docs: 更新 MiniMax 文档
- 添加 Anthropic 兼容 API 说明
- 更新默认模型为 MiniMax-M2.5
- 添加国际版/中国大陆版配置示例
- 更新 env.example 注释
Co-authored-by: Shinyi <[email protected]>
* fix: 完善 MiniMax 双端点支持及问题修复
- 支持 MiniMax Anthropic 兼容端点和 OpenAI 兼容端点自动切换
- 修正默认端点为 api.minimaxi.com (中国大陆可用)
- 修复端点路径缺少 /v1 的问题
- 添加前端 MiniMax logo 映射
- 移除调试日志
- 修正 env.example 默认配置
* chore: clean backup artifacts and align biome formatting
* fix: resolve effectiveProvider bug, deduplicate MiniMax URL logic, fix docs
- Fix critical bug: effectiveProvider was empty during auto-detection,
causing multi-system-message to be sent to MiniMax (which rejects it).
Now uses resolved provider from getAIModel instead of re-deriving it.
- Extract normalizeMiniMaxBaseURL() shared helper to eliminate duplication
between ai-providers.ts and validate-model/route.ts
- Add guard for undefined MiniMax baseURL to prevent hitting api.anthropic.com
- Fix docs: mark China mainland URL as default (matches code behavior)
- Restructure minimax/glm/qwen/kimi/qiniu validation to use shared pattern
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* feat: document MiniMax dual API formats in docs and UI
- Add hint below Base URL input when MiniMax is selected, explaining
Anthropic-compatible (/anthropic) vs OpenAI-compatible (/v1) endpoints
- Update all 3 ai-providers docs (en/cn/ja) to list all 4 endpoint options
(China/International × Anthropic/OpenAI)
- Add i18n translations for the hint in all 4 locales
Co-Authored-By: Claude Opus 4.6 <[email protected]>
* refactor: deduplicate PROVIDER_LOGO_MAP, remove unnecessary optional chaining
- Extract PROVIDER_LOGO_MAP to lib/types/model-config.ts (was duplicated
in model-config-dialog.tsx and model-selector.tsx)
- Remove unnecessary ?. on PROVIDER_INFO.minimax (it's a full Record)
---------
Co-authored-by: msga-oc <[email protected]>
Co-authored-by: Shinyi <[email protected]>
Co-authored-by: dayuan.jiang <[email protected]>
Co-authored-by: Claude Opus 4.6 <[email protected]>
* Add Ollama Cloud support with Base URL and API Key configuration
* implemented feedback
* fix: use OLLAMA_BASE_URL env fallback in validate-model endpoint
* Remove dedicated Ollama configuration block
* security(ollama): prevent API key leak to client-controlled URLs
* added test
* fix: security hardening and Ollama Cloud default URL
- Add server OLLAMA_API_KEY fallback to validate-model endpoint with
SSRF guard mirroring ai-providers.ts
- Tighten top-level SSRF exemption: only exempt Ollama when no server
OLLAMA_API_KEY is configured
- Update Electron config to support OLLAMA_API_KEY env var
- Change default Ollama URL from localhost:11434 to ollama.com/api
(Ollama Cloud) for web UI users
- Add tests for server env combo, API-key-only, and SSRF guard scenarios
---------
Co-authored-by: dayuan.jiang <[email protected]>
* feat: add Material Design Icons shape library (#685)
Add Google Material Design Icons as a new shape library using Google's
CDN. Includes top 300 most popular icons by usage, and updates system
prompts to guide the AI to call get_shape_library before using any icon
library.
* fix: align get_shape_library guidance for non-cloud icon libraries
Support multiple API keys per provider with random selection for load
balancing. When AI_MODELS_CONFIG has multiple apiKeyEnv values for
a provider, requests will randomly select one available key.
- Update schema to accept apiKeyEnv as string or string array
- Add random key selection in resolveApiKey()
- Update validation to check at least one key exists
- Add tests for array format support
* [Feature] Add VLM-based diagram validation
Add automatic VLM (Vision Language Model) validation after display_diagram
tool execution. The system captures a screenshot of the rendered diagram,
sends it to a VLM for visual analysis, and uses feedback to improve
diagram quality through the existing retry mechanism.
Changes:
- Add /api/validate-diagram endpoint for VLM validation
- Add diagram-validator.ts for client-side validation orchestration
- Add validation-prompts.ts for VLM system prompts
- Add ValidationCard component to display validation status in chat
- Add PNG capture functionality to diagram context
- Integrate validation into tool handlers with retry support (max 3)
- Add "Improve with Suggestions" button for manual regeneration
- Add settings toggle to enable/disable VLM validation
- Add getValidationModel() helper in ai-providers.ts
* refactor(validation): use AI SDK structured outputs and address review feedback
- Replace generateText + manual JSON parsing with generateObject and Zod schema
for type-safe structured validation output
- Use AbortSignal.timeout() instead of Promise.race for cleaner timeout handling
- Add timeout validation with minimum 1000ms to handle malformed env values
- Remove unused xml parameter from validateRenderedDiagram API
- Remove parseValidationResponse function (now handled by schema)
- Clear validationStates on session switch and new chat to prevent memory leak
- Update 100ms render delay comment to clarify best-effort heuristic
- Remove unused useEffect import from ValidationCard
- Fix optional chaining lint warning in ValidationCard
- Add unit tests for formatValidationFeedback function
* refactor(validation): use AI SDK experimental_useObject hook instead of raw fetch
- Change API endpoint from generateObject to streamObject for useObject compatibility
- Create useValidateDiagram hook using AI SDK's experimental_useObject for reactive validation
- Update useDiagramToolHandlers to accept validation function as parameter
- Update chat-panel to use new useValidateDiagram hook
- Remove validateRenderedDiagram function from lib/diagram-validator.ts (now in hook)
- Export ValidationResultSchema from API route for client-side use
* fix(validation): extract schema to shared file for client/server compatibility
Move ValidationResultSchema to lib/validation-schema.ts to avoid importing
server-side modules (ai-providers) into client-side code. This fixes the
Turbopack build error caused by the hook importing from the API route.
* fix(validation): use 'Valid' instead of 'Complete' for validation success
Change ValidationCard success label from 'Complete' to 'Valid' to avoid
conflicting with ToolCallCard's 'Complete' badge in E2E tests. This fixes
the diagram-generation E2E test that expects a specific count of 'Complete'
badges.
* fix(validation): add aria-hidden to icons to prevent duplicate ID warning
* fix: improve VLM validation with bug fixes and i18n
- Fix race condition in pendingValidationRef (reject previous pending validation)
- Fix response format consistency (use streaming for all responses)
- Remove dead code (unused lastRequestRef and ValidationRequest interface)
- Consolidate duplicate types (re-export from validation-schema.ts)
- Add 'success_with_warnings' status for valid diagrams with warnings
- Fix tool card auto-collapse (only collapse once, respect user toggle)
- Set VLM validation default to disabled
- Add i18n support for diagram validation settings (en/zh/ja)
- Mark feature as experimental in settings UI
* fix: resolve TypeScript errors in electron-standalone
- Add forwardRef support to ChatInput component with ChatInputRef type
- Copy electron.d.ts to electron-standalone/electron folder
- Exclude electron-standalone from root tsconfig type checking
* fix: return empty string for valid result with no issues in formatValidationFeedback
* feat(i18n): add validation strings for ValidationCard component
- Add validation section to en.json, zh.json, ja.json dictionaries
- Update ValidationCard to use useDictionary hook
- Replace all hardcoded English strings with i18n keys
---------
Co-authored-by: dayuan.jiang <[email protected]>
* fix(quota): bypass quota for users with Bedrock credentials
The hasOwnApiKey check only looked for x-ai-api-key header, but Bedrock
users provide AWS credentials via x-aws-access-key-id instead. This
caused Bedrock users with their own credentials to still be subject to
quota limits.
* fix(quota): also bypass quota for Vertex AI users
* style: auto-format with Biome
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
* fix: allow private URLs by default for reverse proxy setups
Fixes#588 - Users with reverse proxy setups (e.g., Antigravity tools)
were getting "Invalid base URL" errors due to SSRF protection blocking
private/internal URLs.
Changes:
- Add ALLOW_PRIVATE_URLS env var (defaults to true)
- Set to "false" to enable strict SSRF protection if needed
* refactor: extract isPrivateUrl to shared utility
When the LLM generates edit_diagram tool calls, it sometimes produces
inconsistent quote escaping in XML attributes within JSON strings.
For example: y="-20\" instead of y=\"-20\"
This causes JSON parsing to fail, and jsonrepair cannot fix this pattern.
Added pre-processing regex to detect and fix cases where the opening
quote is unescaped but the closing quote is escaped in attribute values.
* [Feature] Server side multi-pvorider/model support
* copilot suggesition implemented
* feat: improve model selector UI and auto-select default server model
- Replace emoji headers with Lucide icons (Monitor, User)
- Fix transition-all to explicit properties per web guidelines
- Use CSS padding instead of hardcoded space indentation
- Add ModelSelectorSectionHeader component for section headers
- Replace Star icon with "default" text label
- Style Configure button with muted text color
- Auto-select default server model when page loads
- Support AI_MODELS_CONFIG env var for cloud deployments
- Support custom apiKeyEnv/baseUrlEnv per provider config
* docs: update server-side multi-model configuration documentation
- Add AI_MODELS_CONFIG env var option for cloud deployments
- Document apiKeyEnv and baseUrlEnv fields for custom env var names
- Document default field for auto-selecting default model
- Remove deprecated version field from examples
- Add field reference table for clarity
---------
Co-authored-by: dayuan.jiang <[email protected]>
SiliconFlow is transitioning from .com to .cn domain. The .cn endpoint
uses Global Traffic Manager (GTM) for better global access, while .com
is being phased out.
* feat: add URL content extraction for AI diagram generation
* Changes made as recommended by Claude:
1. Added a request timeout to prevent server resources from being tied up (route.ts)
2. Implemented runtime validation for the API response shape (url-utils.ts)
3. Removed hardcoded English error messages and replaced them with localized strings (url-input-dialog.tsx)
4. Fixed the incorrect i18n namespace (changed from pdf.* to url.*) (url-input-dialog.tsx and en/ja/zh.json)
* chore: restore package.json and package-lock.json
* fix: use i18n strings for URL dialog error messages
---------
Co-authored-by: dayuan.jiang <[email protected]>
DeepSeek provider was not properly formatting image content for Doubao's
API. Now uses OpenAI provider for Doubao models (multimodal support),
while keeping DeepSeek provider for DeepSeek/Kimi models on the platform.
* test: add Vitest and Playwright testing infrastructure
- Add Vitest for unit tests (39 tests)
- cached-responses.test.ts
- ai-providers.test.ts
- chat-helpers.test.ts
- utils.test.ts
- Add Playwright for E2E tests (3 smoke tests)
- Homepage load
- Japanese locale
- Settings dialog
- Add CI workflow (.github/workflows/test.yml)
- Add vitest.config.mts and playwright.config.ts
- Update .gitignore for test artifacts
* test: add more E2E tests for UI components
- Chat panel tests (interactive elements, iframe)
- Settings tests (dark mode, language, draw.io theme)
- Save dialog tests (buttons exist)
- History dialog tests
- Model config tests
- Keyboard interaction tests
- Upload area tests
Total: 15 E2E tests, all passing
* test: fix E2E test issues from review
Fixes based on Gemini and Codex review:
- Remove brittle nth(1) selector in keyboard tests
- Remove waitForTimeout(500) race condition
- Remove if(isVisible) silent skip patterns
- Add proper assertions instead of no-op checks
- Remove expect(count >= 0) that always passes
- Remove unused hasProviderUI variable
All 14 E2E tests and 39 unit tests pass.
* style: auto-format with Biome
* fix: resolve lint errors for CI
* test(e2e): add diagram generation tests with mocked AI responses
- Add tests for generate, edit, and append diagram operations
- Use SSE mocked responses matching AI SDK UI message stream format
- Generate mxCell XML directly in tests for deterministic assertions
- Tests verify tool card rendering and 'Complete' badge state
* test: add comprehensive E2E tests for all major features
- Error handling tests (API errors, rate limits, network timeout, truncated XML)
- Multi-turn conversation tests (sequential requests, history preservation)
- File upload tests (upload button, file preview, sending with message)
- Theme switching tests (dark mode toggle, persistence, system preference)
- Language switching tests (EN/JA/ZH, persistence, locale URLs)
- Iframe interaction tests (draw.io loading, toolbar, diagram rendering)
- Copy/paste tests (chat input, XML input, special characters)
- History restore tests (new chat, persistence, browser navigation)
* refactor: extract shared test helpers and improve error assertions
- Create tests/e2e/lib/helpers.ts with shared SSE mock functions
- Add proper error UI assertions to error-handling.spec.ts
- Remove waitForTimeout calls in favor of real assertions
- Update 6 test files to use shared helpers
* docs: add testing section to CONTRIBUTING.md
* fix: improve test infrastructure based on PR review
- Fix double build in CI: remove redundant build from playwright webServer
- Export chat helpers from shared module for proper unit testing
- Replace waitForTimeout with explicit waits in E2E tests
- Add data-testid attributes to settings and new chat buttons
- Add list reporter for CI to show failures in logs
- Add Playwright browser caching to speed up CI
- Add vitest coverage configuration
- Fix conditional test assertions to use test.skip() instead of silent pass
- Remove unused variables flagged by linter
* fix: improve E2E test assertions and remove silent skips
- Replace silent test.skip() with explicit conditional skips
- Add actual persistence assertion after page reload
- Use data-testid selector for new chat button test
* refactor: add shared fixtures and test.step() patterns
- Add tests/e2e/lib/fixtures.ts with shared test helpers
- Add tests/e2e/fixtures/diagrams.ts with XML test data
- Add expectBeforeAndAfterReload() helper for persistence tests
- Add test.step() for better test reporting in complex tests
- Consolidate mock helpers into fixtures module
- Reduce code duplication across 17 test files
* fix: make persistence tests more reliable
- Remove expectBeforeAndAfterReload from mocked API tests
- Add explicit test.step() for before/after reload checks
- Add retry config for flaky clipboard tests
- Add sleep after reload for language persistence test
* test: remove flaky XML paste test
* docs: run both unit and e2e tests before PR
* chore: add type check and unit test git hooks
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Some models (Kimi K2, DeepSeek, Qwen text models) don't support image/vision
input. The AI SDK silently drops unsupported image parts, causing confusing
responses where the model acts as if no image was uploaded.
Added supportsImageInput() function to detect unsupported models by name,
and return a 400 error with clear guidance when users try to upload images
to these models.
Closes#469
dayuan.jiang