Tool calls to an OpenAPI tool server put every argument the model returned into the JSON request body, including the parameters that were already substituted into the URL. Servers that validate their input strictly (additionalProperties: false) answered 422 "unexpected property", so reads worked and every write through an endpoint with a path or query parameter failed.
The body is now built from the model's arguments minus the operation's declared parameters, keeping any name the requestBody schema declares as a property of its own, so an endpoint that wants the resource id in the body as well as in the path still gets it.
The filter only runs when the resolved body schema lists its properties. A free-form, composed or non-JSON body offers nothing to check a name against, so those requests go out exactly as before.
src/lib/apis/index.ts carries the same request builder for direct tool server connections and had the same bug, so it gets the same fix.
Fixes#29716
Some endpoints stream a tool call whose function name is JSON null instead of a string. Nothing normalized it, so the null stayed on the tool call, was written into the stored message, and was sent back to the endpoint in the assistant message on the next turn, where a null is not a valid function name.
The delta accumulator now replaces a null name with an empty string, at the same point it already normalizes the arguments field. The call still fails as an unknown tool, which is the right outcome for a call that has no name, so the result is one failed tool call instead of a follow-up request the endpoint has to reject.
This is done where the delta enters the accumulator rather than at the consumers, because the name is emitted to the client and persisted while the response is still streaming, before anything downstream could clean it up.
Checked against 1261 streaming delta sequences: behaviour is unchanged except where the delta that creates the tool call carries a null name.
Seen in #29686 with a custom sglang build.
Two OAuth failure paths interpolated the raw token object into their log message. On the callback path that object is a live credential set, so a provider returning no user data wrote an access token, and usually a refresh token, straight into the application log.
Both messages now log without the payload. The token-exchange failure keeps its error level and its client_id binding and reports the provider's error description instead of the raw response body, which by that branch's own condition never contained an access token anyway. The callback failure keeps its warning level and identifies the provider, matching the other failure logs in that handler.
Sending a message to an arena model failed with "'JSONResponse' object has no attribute 'body_iterator'" whenever the backing provider answered with an HTTP error, so the real error never reached the user. Non-streaming requests on an arena model, such as title and tag generation, broke the same way with "'JSONResponse' object is not a mapping".
The arena wrapper assumed the sub-model call always returns a stream for a streaming request and a dict for everything else, but the OpenAI-compatible router returns a plain response object as soon as the provider answers 4xx or 5xx. Both arms now hand that response straight back, which is exactly what the non-arena path already does, so the existing error handling turns it into the usual error message in the chat.
Verified against a matrix of streaming and non-streaming requests with the sub-model returning a stream, a dict, a JSONResponse and a PlainTextResponse: both crashes are gone and the two success paths are unchanged, including the selected_model_id prelude on the stream.
Fixes#29658
A web search is not something that can be cited. The search engine
returns a title, a link and a one line snippet for each hit, and the
model never opens any of those pages. Emitting them as citation sources
produced one <source> tag per result, all named search_web with empty
bodies, and the citation template then instructed the model to cite them
by id. Models either hesitated visibly or attached an id to content from
a different result, which the citations panel then resolved to a title
that looked plausible, so the misattribution read as correct.
Web search results now stay in the tool output the model reads, and stop
being offered as things to cite. Where the model needs to cite a page it
calls fetch_url, whose citation names the URL and already works.
Web search results no longer appear in the citations panel. That is the
point of the change: the panel was offering pages that nothing had read.
Scoped to the native tool-calling path. The legacy handler cites every
tool result as one opaque source and does not single out web search, so
it is left alone rather than special-cased.
A direct-connection streaming request registers a per-request socket.io handler before it asks the browser to start the completion, and that handler was only removed once the response had been fully streamed. Every other way the request could end left it behind for the life of the process: the call to the browser raising, a non-success status, an ack with no arguments or an ack that is not an object, cancellation while waiting for that ack and a response body closed or garbage-collected before it finished. A client that repeatedly hits a failing direct connection grows the server's handler table and the closures it holds without bound, and nothing ever cleans it up.
The removal now runs on every exit from the request, through one named helper that pops with a default so it is safe to run twice on the paths where both the generator's finally and the response's background task fire. The guard covers the exchange up to the status read and catches BaseException, because cancellation is not an Exception, and re-raises it unchanged.
Measured across thirteen exit paths: twelve leak a handler on dev and none of those leak here. The thirteenth, a response body that is never iterated at all, behaves the same on both. 200 failing requests leave 200 handlers on dev and none on this branch. Streamed bytes on the success path, exception types and cancellation behaviour are unchanged.
* perf: stop re-parsing the whole tool-argument buffer on every streamed chunk
Converting an OpenAI stream to Anthropic events buffers each tool call's arguments and, to find out when the JSON is complete, parsed the entire buffer again on every chunk. A tool call with large arguments pays that parse thousands of times, and the cost grows with the square of the argument size.
The parse now runs only when the buffer could actually be complete. A JSON object can only close on its final brace, so a chunk that does not end there cannot complete it. Arguments that are not an object, or that start with whitespace, keep parsing on every chunk exactly as before.
Measured on CPython 3.12 with 130 KB of tool arguments over 7648 chunks:
| | before | after |
|---|---|---|
| parses | 7648 | 1 |
| time | 382 ms | 0.82 ms |
The block closes on exactly the same chunk as before, verified by replaying randomized fragmentations of objects with braces inside strings, escaped characters, unicode escapes, arrays, bare scalars, leading and trailing whitespace and a buffer that never completes, against both JSON backends.
* refactor: read tool['arguments'] directly in the JSON completion guard
Restores the pre-existing comment above the guard to its original wording and drops the `buffered` local, so the guard and the parse call both read `tool['arguments']`, the name the rest of the file already uses for that buffer. Behaviour is unchanged: same three conditions in the same order, same short-circuit result.
* perf: strip whitespace in the tool-argument completion guard
The character guard only looked at the first and last byte of the buffer, so a
chunk that ended in a space still triggered a full parse and a tool argument
with leading whitespace fell back to parsing on every chunk. Stripping first
collapses both cases to a single parse at the end of the stream.
Measured on a streamed tool call, parses and wall time for the whole stream,
orjson on the left of the slash and stdlib json on the right:
| argument shape | before | after |
|---|---|---|
| 20 KB string, char-by-char deltas | 3678 parses, 56 / 28 ms | 1 parse, 3.1 / 3.1 ms |
| 200 KB, 20-char deltas | 1473 parses, 176 / 63 ms | 1 parse, 3.1 / 2.8 ms |
| 8 KB prose, leading whitespace | 715 parses, 5.5 / 2.5 ms | 1 parse, 0.18 ms |
| 8 KB of spaces inside a value | 713 parses, 6.0 / 2.9 ms | 1 parse, 0.83 / 0.72 ms |
The strip costs about 20 ns per delta on arguments that have no whitespace at
either end, which is where the old form was already optimal: a 20 KB compact
argument goes from 191 to 216 us over 1786 deltas. Soundness is unchanged, the
guard can still only skip a parse that would have failed: 2660892 buffers
(exhaustive to length 6 over a JSON-lexical alphabet, every prefix of 26 named
cases with a trailing byte appended, and every codepoint below U+3000 after a
complete document) with zero cases where a parse would have succeeded.
Every chat completion request re-loads the target conversation's entire
message history from the database inside drain_approved_tool_calls() before
discovering there is nothing to drain: a fresh message always points at a
newly minted assistant message with no stored output, so the full-history
read (one SELECT of every chat_message row plus building the message map,
uncached, on top of the identical read process_chat_payload already did) is
pure overhead on every message.
Queued tool approvals can only ever be acted on by a resume or continue
request, and exactly those requests carry assistant_message_id in their
payload. The drain now returns early when the field is absent, removing one
O(conversation length) query per chat message while resume, continue, reject
and pause flows behave exactly as before, independent of the approval mode.
* fix: a rejected ask_user call ending the turn with no reply
The documented behaviour of the built-in ask_user tool is that a call breaking its rules comes back to the model as an error. Instead the reply stopped there: the error was recorded as the tool result, the model was never asked again, and the user was left with a dead chat and no answer.
The rejection is now handed back like any other failed tool result, so the model sees it and can correct itself within the normal tool-call iteration limit. Any ordinary tool the model emitted in the same turn still runs.
A call rejected for arriving alongside other ask_user calls also left those siblings without a result, which the UI shows as a tool call stuck on "Executing..." forever. Every invalid call now gets its own result. Two ask_user calls on their own also reported the wrong reason, saying the call must be made by itself rather than that only one is allowed per turn.
Fixes#29077
* Keep the original ask_user validation order
Restores the pre-existing check order and the unchanged output id fallback, so this change only alters the return shape needed for staging, and trims a comment that narrated the lines below it.
* Correct the ask_user sibling-call error message
* Shorten the ask_user sibling-call error message
* Drop the untrue sibling-call claim from the ask_user error
The ask_user error text told the user and the model "The others ran.", but that sentence is written into the turn output before any sibling tool call has executed, so it can be plainly false. Under a saved chat with tool approval set to ask, the turn pauses right afterwards and the siblings sit at pending/queued, so the user reads "The others ran" directly above the approval prompt for tools that have not run, and reads it again beside the rejection result if they decline. When the model sends two ask_user calls and nothing else, nothing runs at all and the sentence is emitted twice.
The staging helper cannot see what happens to the sibling calls, so it no longer narrates it. The remaining two sentences hold in every flow: ask_user really is dropped from the executed calls whenever this error is set, and calling it on its own is always the right retry.
The Redis-backed model registry skips its write when the content signature matches what is already stored. That skip has never worked across processes. Two of the values it hashes come out of Python sets, and set iteration order varies with each process's hash seed, so every worker computed a different signature for identical content and every worker rewrote the whole registry on every refresh.
Sorting both makes the signature depend on content alone. Measured on a 120 model registry, 522 KiB serialized: a refresh whose content already matches drops from GET, HKEYS, HSET and SET at 5.1 ms to a single GET at 2.2 ms per worker, and the 522 KiB write leaves the wire entirely.
Verified across 12 child processes with 12 distinct hash seeds: 12 different signatures before, 1 after. Filter execution order is unaffected, because the filter pipeline re-sorts by priority and id before running.
* fix: stop streaming responses breaking on a duplicate output key
With reasoning-capable models the chat froze mid-stream: the first chunk of the answer appeared, nothing followed, and the whole message only showed up once generation finished. The browser console showed a Svelte each_key_duplicate error.
When a stream event addresses an output slot past the end of the array, the missing slots were filled with the event's own item, id included, so a gap of two left two entries claiming the same id. The next chunk for that item was matched by id, landed in the first of the two, and the rendered list ended up with two items sharing a key, which Svelte refuses to update.
Only the addressed slot now takes the event's item, and the slots before it are anonymous placeholders. Replayed the reported event sequence against the real code: keys are unique again and the chunks stay in order instead of being split across the copies.
* fix: stream reasoning deltas when the provider also sends reasoning_details
Providers such as OpenRouter emit reasoning_details alongside the reasoning
text on the same delta. Merging those details cleared the pending event
unconditionally, discarding the response.reasoning_text.delta that had just
been built, so the client received no reasoning until the response completed
and the thinking block only appeared after generation finished.
The event is now only dropped when the details were all there was to report.
Details persistence is unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014uuEg4AXPs9zE3vVUfN1Fj
---------
Co-authored-by: Claude <noreply@anthropic.com>
After a tool call, the model's thinking was streamed into the chat as if it were the main response, and only jumped into the collapsed Thoughts section once the turn finished. Every further tool call repeated it.
Each tool round appended an empty placeholder message item to the output and sent it to the browser, then dropped it again from the copy used to offset the next round's item indices. The browser therefore held one item more than the backend counted, so the first thinking chunk of the next round was written into that leftover message item and rendered as normal text until the finished output replaced it.
The placeholder is removed. It was never needed: a message item is already created when actual content arrives, and dropping it also stops an empty assistant message being sent back to the model on the follow-up request.
Co-authored-by: Claude <noreply@anthropic.com>
Any `<$...>` run in a chat message was treated as an inline skill mention and removed before the request reached the model, so text like `<$(=MonthStart($(vMaxMonthEndINC)))"}, [Registration day] >` silently vanished mid-message and the model only saw the part before it.
The mention regexes accepted any character except `|` and `>` as the skill id, so they matched far more than real mentions. Skill ids are already validated as `[a-z0-9_-]+` when a skill is created, so both regexes now require that charset. Ordinary text passes through untouched while `<$id>`, `<$id|Label>` and `</id|Label>` still resolve and strip as before.
Verified against the reported message (now preserved verbatim) and the three mention forms.
Co-authored-by: Claude <noreply@anthropic.com>
Every chat request hands each builtin tool a fresh copy of its cached spec, because callers mutate what they get. That copy was a full deepcopy of a nested dict, repeated per tool per message.
The builder now caches the spec already serialized, so a request only parses it back. Parsing is what produces the independent tree callers mutate, and the cached value becomes an immutable string, so a request can no longer reach the cached object at all.
Measured on CPython 3.12 with a 1.1 KB spec and 20 builtin tools per request:
| | before | after |
|---|---|---|
| stdlib json, the default | 276.2 us | 66.4 us |
| orjson | 279.5 us | 37.5 us |
Builtin specs are plain JSON by construction: pydantic normalizes every default before it reaches the schema, so a tuple, set, enum or datetime cannot appear in one, and an unserializable default is dropped rather than embedded.
Every streamed delta saves a snapshot of the in-progress response so a reconnecting client can resume it, and each save rebuilt the assistant text from scratch. On the Chat Completions path that re-joined every accumulated chunk, including on saves carrying no new text, so a long answer followed by a large tool call re-joined the whole answer once per argument chunk. The Responses API path never collects those chunks and reads the text back out of the output items instead, where the blank check copied it in full every time.
The joined string is now kept and reused until another chunk arrives, since content_parts is only ever appended to; the nonlocal declaration that suggested otherwise was already dead and is dropped, and inlining the single-use helper removes an unreachable branch with it. The blank check in get_output_text now tests the text rather than allocating a stripped copy of it, which is equivalent for all twelve of its callers. Text streaming on the Chat Completions path is unchanged, since a text delta always appends before it saves.
| stream | before | after |
| --- | --- | --- |
| 20k-char answer, 2000 tool-argument chunks | 21.4 ms | 0.06 ms |
| Responses API, 40k deltas, 200k chars | 80.7 ms | 50.5 ms |
Without Redis nothing extra is retained, since the snapshot store already held that string; with Redis one copy of the response text stays alive while the stream runs.
Tag detection for the code interpreter ran regardless of the tool-calling mode, so a model in Native (Agentic) Mode that emitted <code_interpreter> blocks in ordinary reply text had that code sent to the executor. Native mode never teaches the tag format and exposes execute_code as a builtin tool, so the parser had nothing legitimate to pick up there.
Gates detection on the legacy mode, matching the condition that already decides whether the tag prompt is injected at all. The five authorization checks are unchanged, and native mode keeps executing through the tool.
Deployments on native mode whose models emit the tags unprompted will now see them rendered as text.
Some providers send one very large piece of a streamed answer in a single go: a long reasoning trace, a code execution result, a turn with many tool calls, or a response echo carrying a big tool list. Anything past 128 KB in one line killed the chat mid-answer with a misleading `400, message: Got more than 131072 bytes when reading`. Nothing was rejected upstream, that is our own reader giving up on an oversized line.
Open WebUI already had code that assembles lines itself with no such limit, but it only ran when CHAT_STREAM_RESPONSE_CHUNK_MAX_BUFFER_SIZE was set. Unset is the default, and in that case the raw capped reader was used instead, so a default install always broke. That path now always assembles lines, and the setting goes back to being what its name says: an optional cap, off by default. It applies to the Ollama stream as well, since both now share the same reader.
The assembly loop only splits once a line actually completes, because the old one re-concatenated and re-split the whole buffer on every network chunk. Without that, allowing long lines would have traded an error for multi-second event loop stalls.
| | 20 MB in one line | 200k small lines |
| --- | --- | --- |
| before | 4249 ms | 27.3 ms |
| after | 37 ms | 25.2 ms |
* fix: apply the SSRF checks to redirect targets on every web fetch path
Two guards protect server-side fetches: a private-IP check and the operator's `WEB_FETCH_FILTER_LIST`. Neither reached a redirect hop on the aiohttp paths, and the filter list never reached one on the requests paths either.
aiohttp answers IP-literal hosts itself without consulting a resolver, so `_SSRFSafeResolver` was never invoked for a hop such as `http://169.254.169.254/` and the private-IP check simply did not run. With redirect following enabled, a submitted public URL that redirects to an IP literal reached loopback, RFC1918 and cloud-metadata addresses, and the response body was returned to the caller. The filter list was consulted only in `validate_url`, on the originally submitted URL, so a redirect to a filter-listed host was fetched without it ever being applied.
`_SSRFSafeResolver` is replaced by `_SSRFSafeConnector`, which hooks `_resolve_host` so the IP check also covers the IP-literal shortcut and both DNS cache paths. The filter list moves to a per-request hook on each transport, `connect()` for aiohttp and `send()` for the requests adapter, because those see the request destination: at the connection layer a proxied request presents the proxy's host, and a pooled connection skips resolution entirely. This covers every hop, including redirects, on all five aiohttp call sites and both requests sessions. The Playwright loader already validated each hop and is unchanged.
Both gaps required `AIOHTTP_CLIENT_ALLOW_REDIRECTS=true`, which is not the default.
Two behaviour changes for operators. The filter list now applies to redirect targets rather than only to submitted URLs. Under a forward proxy it is evaluated against the request destination instead of the proxy, which also fixes allowlist entries rejecting every fetch in proxied deployments.
* refac: match the web fetch filter list against resolved addresses
The filter list is now evaluated against the hostname together with the addresses it resolves to, at URL validation and on each connection, on both transports. An IPv6 address is also matched by the IPv4 address it carries.
* refac: screen outbound fetch addresses against reserved ranges ipaddress misses
`ipaddress.is_global` was the only test behind the web-fetch address check, and it answers a narrower question than "may we fetch this". Several special-purpose ranges are globally routable by registry while nothing on them is a legitimate destination, so they passed. Classification now screens those ranges on top of `is_global`, and applies the same screen to the IPv4 address embedded in an IPv6 transition encoding rather than only to the literal. All three checkpoints share the predicate, so they all inherit it.
The range list is the exact complement of what CPython's `ipaddress` already models, checked entry by entry against both IANA special-purpose registries. Prefixes IANA marks globally reachable are deliberately left out, so no real destination changes behaviour. Verified against 31 addresses covering every entry, their transition-encoded forms, and public controls in both families: 31/31 expected after, 18/31 before.
* refac: match web fetch filter entries that name an address or a range
A filter entry that parses as an address or a CIDR range is matched by containment rather than by DNS label suffix, so a range covers the addresses inside it and an address matches however it is spelled. A range entry previously matched nothing at all, silently.
The built-in list gains the special-purpose networks that ipaddress.is_global reports as reachable while nothing on them is a legitimate destination, so taking an address out of reach is a WEB_FETCH_FILTER_LIST change rather than a release. Those entries hold whether or not local web fetch is enabled; the private-address rule still follows the toggle.
