Acts on a second triple-check audit.
`normalized` means one thing. The driver stamped it on both the OpenAI-shape
path and the legacy `response.normalize` path, which converts toward Anthropic
blocks — so a caller could get `normalized: true` alongside array content, and
the flag told them nothing they could branch on. The legacy branch no longer
sets it. That branch is reachable only by a direct driver call with
`response.normalize` and no `normalize`; the four wire routes pin
`normalize: false`, which skips it.
Mistral streaming, split by concern. Flattening a reasoning model's chunked
`delta.content` to a string is a correctness floor and stays ungated — the
shared handler passes the value straight to `addText`, so an array reaches the
caller as stringified objects. Splitting the thinking text out into a
`reasoning` delta is the dialect change and now sits behind the policy gate
like the non-streaming remap. On the native path the thinking text is kept
inline rather than dropped.
Mistral `finishReason` is deleted only once its value carried over. The delete
ran unconditionally, so a non-string `finishReason` with no `finish_reason`
left the choice with no finish reason at all.
Four doc claims corrected against the code paths they cover: `content` is
string-or-null on normalized responses (tool-only turns carry no text, and the
`// always a string` example comment was wrong); `reasoning_details` is not
scoped to normalized responses, since Responses models emit it either way; and
the release-date rule depends on the serving provider's own dates — OpenRouter
derives them from its live API, so models newly listed there from 2026-09-01
normalize by default.
Adds the test the copy-on-write fix was actually for: one messages array sent
through two sequential calls, asserting the caller's array is untouched and
both attempts carried the thinking signature. That is the fallback hazard; the
harness wires one provider per model, so the fallback loop itself cannot be
driven from a provider test.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Acts on a triple-check audit of this branch.
Gate the Mistral dialect remap. The camelCase→snake_case rewrite and the
chunked-content flattening were firing for every Mistral call regardless of
`normalize` or the cutoff, deleting `finishReason` and `message.toolCalls` out
from under any caller reading them. Both now sit behind the policy resolution
the driver already used, extracted as `shouldPresentAsOpenAI` so the provider
and the driver cannot drift. The streaming chunk-array split stays ungated:
handing an array to `addText` is a plain bug, and streamed chunks are
provider-uniform by design. The conformance matrix now passes `normalize: true`,
which is the contract it was always testing.
Unify the reasoning join. Three code paths produced two separators while one
doc sentence described them all: the coercer joined thinking segments with '',
the Responses handler and Mistral with '\n\n'. The coercer now matches, and
chatresponse.md's claim is true for every path it covers. Text blocks still
join with '' — Anthropic splits prose mid-sentence across them.
finish_reason is an open set. chat.md's normalize bullet and the SDK
ChatMessage typedef still declared a closed four-value set, contradicting the
documented pass-through of unmapped vendor reasons and the coercer that
implements it.
Stop mutating caller messages. Both reasoning-replay input paths deleted
output-only fields from the caller's own message objects, which the driver
reuses across fallback attempts. Both strip a copy now; tests pass a frozen
message through each.
Drop three dead things the type cleanup left: the no-op ChatProvider
checkModeration stub (no subclasses, no callers — its removal restores a
pre-existing baselined TS2420), the redundant second normalizeReasoningContent
call in BytePlus and ZAI, and the coercer's bare-string branch that no provider
reaches. A bare string now passes through by reference instead of being
coerced.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Commit 6b4f5b01f dropped 31 entries (49 error instances) from
tools/typecheck-baseline.json as a side effect of fixing the AI-code type
errors they suppressed. The file is a ledger the maintainer owns, and shrinking
it inside a feature PR ships a CI-gate change nobody asked for.
Restoring is free: tools/typecheck.mjs fails only on *regressions*
(count > baseline), so stale entries are non-fatal — the gate prints "49
baselined error(s) fixed. Run npm run typecheck:update to lock that in." and
exits 0. Nothing else in the repo reads the file, vitest included, and
tsconfig.build.json's noCheck:true means emit is unaffected. The errors stay
fixed in the source either way; only the ledger's own bookkeeping is deferred
to whenever the maintainer chooses to regenerate it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`magistral-*` returns `message.content` as a ContentChunk[] rather than a
string, with the thinking text nested one level deeper inside `thinking`
chunks. The camelCase remap did not touch it, so a non-streamed magistral
response reached the caller as an array with no `reasoning` — the one case
left where a provider did not produce the equalized shape this branch
promises. Streaming had the matching bug: the chunk array was handed to
addText, which would have stringified it into the text stream.
Both paths now split chunked content into a string `content` plus a
`reasoning` string, joining multiple thinking chunks with a blank line as the
Responses handler and the Anthropic coercer do. The streaming fix rides the
existing Mistral-only `chunk_but_like_actually` hook, so no new deviation is
introduced.
The conformance matrix had no Mistral reasoning fixture, which is why it
missed this; it now has one carrying chunked content, verified to fail
without the flattening.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Rules from the post-normalization review: don't silently change what existing
models return, verify doc claims against every code path they cover, record
self-disclosed defects in the PR draft, finish with a fresh build + suites, and
check git stash before concluding an edit was lost. Plus the standing
typecheck-baseline no-go.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Closes the reasoning gaps left open by the normalization work.
Reasoning replay. The coercer dropped Anthropic thinking-block signatures and
the Responses handler dropped reasoning item ids/encrypted_content, so a
normalized reasoning turn could not be replayed — Anthropic rejects an
extended-thinking tool-use continuation whose thinking blocks lost their
signature. Both now ride `message.reasoning_details` verbatim, and both input
paths accept them back: ClaudeProvider splices the blocks ahead of the content
(Anthropic requires them to lead), and the Responses input processor expands
them into standalone `reasoning` items. Output-only fields a replayed message
carries (`reasoning`, `refusal`, `normalized`) are stripped on both paths,
since neither upstream accepts them. The docs caveat recommending
`normalize: false` for agentic Claude loops is gone; it is no longer true.
Unmapped stop reasons. chatresponse.md promised a vendor `finish_reason` with
no OpenAI analog "passes through unchanged" — true for the Mistral remap, false
for the Anthropic coercer, which discarded it. Anthropic's `pause_turn` means
"continue this turn", so flattening it to `stop` destroyed the signal. The
coercer now passes unmapped values through verbatim, matching both the doc and
the Mistral path, and the docs gain the full Anthropic stop-reason table.
Reasoning summaries. Multi-part summaries joined with '' instead of a blank
line, and the streaming Responses path emitted no reasoning at all;
`response.reasoning_summary_text.delta` now feeds the same `reasoning` stream
channel the chat-completions handler uses.
Types. `text?: string & { verbosity?: ... }` was an uninhabitable intersection
(providers read `text?.verbosity` as an object), and the verbosity enum was
`'concise' | 'detailed'` where OpenAI accepts `'low' | 'medium' | 'high'`.
Adds `reasoning`, `reasoning_details`, and `refusal` to the SDK ChatMessage
typedef.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Adds the upstream Hoonify provider (feat #3499, merged after this
branch was cut) to the cross-provider output conformance suite. It
speaks the OpenAI chat-completions dialect and conforms out of the box.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Introduces a new `normalize` option for chat completions, plus release-date based default normalization (post-2026-09-01) to coerce provider-native outputs into a consistent OpenAI-style shape. Adds shared normalization utilities, extensive driver/provider consistency tests, and controller safeguards that pin provider-native output where route-specific translators are used. Also wires the option through puter.js (`chat` options and `puter.ai.normalize` default), updates AI/chat response docs and examples, and resolves related TypeScript typing issues reflected in the typecheck baseline.
The realtime fan-out resolved its audience straight from the `share`
index, which has no live-grant check. `/auth/revoke-user-user` deletes
the permission and leaves the index row, so a revoked recipient's socket
kept receiving name, size, masked path and mtime for every write and
move under the folder, with no expiry.
The service already solves this elsewhere — `#reachingHolders` returns
exactly the holder/entry pairs whose grant still stands, and
`listSharedWithMe` was moved onto it for the same reason. The realtime
path never got the same treatment; it does now.
Free on the unshared path: with no share rows reaching the entry there
are no holders to check, so the write path every user takes is unchanged.
Pinned by a test that counts permission reads.
A grant on the entry itself is keyed on uuid, so it follows the entry
into the owner's Trash. Both ends of the move then resolved, and the
recipient was told the shared item had moved — to the GUID name Trash
gave it. Their own copy got renamed to a GUID and stayed on screen.
`shared-with-me` has always omitted trashed entries, so the listing and
the event disagreed; only the event was wrong. Trashing now reports
item.removed at the path the recipient knew, which is also what the
desktop's data-path selector needs to find the row. Restoring out of
Trash reports item.added.
A move that leaves the recipient's masked address unchanged now stays
quiet — a share masks its own root, so the owner shuffling it around
their tree is invisible to the recipient and the event carried nothing.
Both ways a file can vanish were silent for anyone holding the folder
above it, so a third party's window kept showing a file that was gone
and 404'd on click.
A delete built its audience from the permission rows it removed, and a
file inside a shared folder has no grant of its own — only the folder
does — so the audience was empty. A move resolved its audience from the
entry's new path, and the GUI's Delete is a move to the owner's Trash,
where no recipient has a share.
Resolve the audience from where the entry was rather than from the
grants that went with it:
- Deletes also fan out to holders reaching the entry through an
ancestor, coalesced by parent so a subtree stays a couple of queries.
A holder covered by both passes is told once.
- Moves resolve both ends. Reaching both is item.moved, only the
destination item.added, only the origin item.removed.
Recipients are named by the path they knew, masked through their own
share rather than the owner's tree.
Browsers only allow documentPictureInPicture.requestWindow() from a
top-level document, and an app lives in an iframe, so an app calling it
gets NotAllowedError ("only allowed from a top-level browsing context").
The `document-picture-in-picture` token in the iframe's `allow` list does
nothing — it is not a policy feature the browser knows. Video PiP
(video.requestPictureInPicture) already works inside apps.
The GUI is the top-level document, so a new PictureInPictureService opens
the window for the app and fills it with an iframe of a page the app names,
which must come from the app's own origin (checked against the message's
origin, now carried on the IPC caller context). One window per app
instance; it closes with the app's window, and the app hears about a close
it didn't ask for. The window's opener is the GUI, so the page inside it
can reach its app's frame through parent.opener.frames and share objects
directly — a MediaStream included, which postMessage cannot carry (tracks
are not transferable between windows in Chromium).
puter.js gains puter.ui.requestPictureInPicture({ url, width, height,
onClose }) and puter.ui.exitPictureInPicture(), with docs.
Below 480px the Files tab hides its directories sidebar, which was the only
entry point to Shared (a query, not a directory) and Trash (filtered out of
the Home listing). Home now carries a row for each; CSS shows them only at
the widths where the sidebar is hidden, using the complementary breakpoint so
the two never both show or both disappear.
The rows borrow the item markup for layout but aren't `.item`, so sorted
insert, selection restore, share-link selection and socket updates ignore
them. The footer count and keyboard select-all exclude them explicitly, and
the placeholder-removal sweeps leave them in place. They navigate on tap and
offer the same menu as their sidebar entry via the ⋯ button or long-press;
the Trash icon tracks empty/full alongside the sidebar's.
Also guard the `window.user.directories` lookup in renderDirectory: it is
undefined for some sessions, and since `puter://shared` doesn't look like a
path it always reached that branch, threw outside the try, and left the
spinner up with navigation stuck — from the desktop sidebar as well.
Replace the Files tab's viewport-height sizing with flex-based layout sizing through the dashboard content chain. This keeps the internal file list within the visible content area on mobile browsers, preventing the bottom rows from being clipped when browser toolbars are shown.
Raised in review: could an event from another node re-trigger the
fan-out? Not today — broadcast carries outer.* and pubsub.* only, so
fs.* never crosses a node boundary, and the emitted outer.gui.* is
consumed on the peer by SocketService while ShareService listens to
fs.* alone, so nothing re-enters.
That safety is a property of what broadcast happens to replicate, which
is not this service's to rely on. The handlers now skip anything tagged
from_outside: the node that did the write has already told the audience,
and a second fan-out would only duplicate it.
A recipient's client keeps its cache fresh from fs events pushed over
their socket, and ShareService fans those out to holders — but only for
write, move and delete. A new entry emits fs.create.<flavor>, not
fs.write.file, and an in-place rename emits fs.rename; neither had a
listener, so a recipient watching a shared folder never learned that a
file appeared in it or was renamed. Part of why: those keys and
outer.gui.item.renamed were missing from the typed event map, so a
listener for them did not compile.
Delivering the event is only half of it. Paths were masked against the
entry itself, so item.added named a parent no cached listing was keyed
on, and the payload carried no dirpath, which is how the desktop finds
the container to render into — the event would have arrived and changed
nothing. Paths are now masked at the share the holder reached the entry
through, which is the address their own reads returned, and from_path on
a move and old_path on a rename travel the same way (dropped when the
move started outside the share, self-masked when the share is on the
entry itself, where the root already carries the new path).
Creates fire per entry, so an upload would have cost one share lookup
per file; they are coalesced by parent folder the way subtree deletes
already are. Measured on a 25-file burst into one folder: 25 lookups
before, 1 after. A holder with a share on both a folder and something
inside it is told once, by the nearer of the two.
A peer server could only be reached by the invite code the signaller
minted for it, which is unknowable ahead of time and dies with the
server's socket. Any app that wanted a link to share in advance or reuse
had to run a registry of its own that maps a stable code to whatever
invite is live — and keep it live, since the signaller socket can drop
without the library saying a word (`onclose` was an empty function),
leaving a host serving an invite nobody can dial.
`serve({ name })` serves under a room name of the caller's choosing;
`connect(name)` dials it. A name is held by whoever serves it and free
once they stop; serving a held name from another identity rejects with
`name_in_use`, from the same identity it takes over. A connection to a
room nobody serves fails with `no_host`, so a lobby has a definite answer
to poll on instead of guessing between "not yet" and "gone".
`serve({ guestGrant })` and `server.setGuestGrant()` leave a guest grant
with the signaller, which hands it to every anonymous guest in the connect
reply; the connection redeems it before making its offer, so a guest with
no session gets relays without the app publishing the grant anywhere.
A server whose signaller socket drops now re-registers on its own with
backoff — under the same name, or a fresh code announced by the new
`reconnect` event — and pings the socket to keep idle proxies from cutting
it. It fires `close` (`replaced`, `name_in_use`) when it has to stand
down for good. Existing connections are peer-to-peer and never affected.
Signaller refusals reach the connection's `error` event as an Error with
a `code` next to the message. Keepalive replies and stray frames no longer
throw in the message handlers.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A share email lands on `/?shared=…`, and its recipient is signed out or
in a temporary session more often than not. Password login reloads the
same URL, so the link came back with it — but an OIDC login leaves the
origin and returns to a URL the backend builds from a whitelist of bare
paths. `return_to` was a pathname only, and `/` wasn't on the list, so
the recipient came back signed in on the Dashboard's Apps tab with
nothing to say what had just been shared with them.
`shared` now travels with the path. The GUI sends it along with the
return path; the backend accepts it as the one parameter a `return_to`
may carry, and only values shaped like the masked share path the mail was
built from — everything else rejects the whole value, the path is still
matched as a raw string so a protocol-relative one can't smuggle an
origin through, and the query is rebuilt rather than reflected. The items
ride onto the error page too, so a failed attempt retries on the link
instead of losing it.
The root joins the whitelist for this, but only when it names something:
on its own it is where the flow already lands.
src/gui/src/helpers was mostly snake_case, disagreeing with itself and
with the camelCase file convention in AGENTS.md. Rename the 45 offenders
with git mv and repoint every import specifier at the new paths.
Files only — exported symbols are unchanged. applyItemAddedToContainers.js
still exports apply_item_added_to_containers and no call site is edited
beyond its import line; those identifiers cross module boundaries and some
hang off window, so renaming them is a separate, riskier change.
* Land share-email links in the Dashboard's Shared view
A share email's links opened the item on the desktop, and "Open Puter"
went to the bare origin — the recipient arrived on Home with nothing to
say what had just been shared. Every link now lands in the Dashboard's
Files tab, on Shared, with the shared items selected.
Each named file still links to itself. "Open Puter" carries every item
in the mail (?shared=a&shared=b), so the whole batch lands highlighted;
the digest records now keep each item's masked path for that. The
button's href is rendered raw like the item links, so `=` and `&` read
the same in the html and text parts, and the unsubscribe link is built
from a separate origin value.
On the GUI side `/?shared=` boots the dashboard rather than the desktop;
`/desktop?shared=` keeps opening the item in place. The Files tab
selects the rows by uid, which survives a rename. A share only reaches a
real account, so a share link never mints a temp user and an existing
temp session is asked to sign in first, as the desktop already did. The
address-bar cleanup both shells need moves into a helper.
* Keep the share email's button link within what mail clients tolerate
The "Open Puter" link names every item in the mail, capped at twenty. But
a single item's parameter is ~150 characters once the owner, uuid and an
encoded name are in it, so twenty of them run to several kilobytes — past
the ~2000 characters where older mail clients cut a URL off or stop making
it clickable — and the count alone couldn't do what its comment promised
for the mail's primary button.
Add a length budget beside the count: items go in, in digest order, only
while the whole link stays under it, so the button always works and the
first items are the ones highlighted. A single item always fits.
* Keep a share link's own item however long its name runs
`shareDeepLink` now builds through `sharedViewLink`, whose length budget
applied to the first item too: a parameter that alone overran it was
dropped, and the link came out as a bare `?shared=`. A name of a few
hundred characters — the GUI allows five hundred, and encoding triples
every non-ASCII one — was enough, so that item's own link in the mail
landed on Shared with nothing picked out, where it used to open the item.
The first item now always goes in; the budget only decides how many more
join it. One long link is still the item the mail is about, and it is no
worse than a bare origin for the clients that truncate it.
Sharing was one item at a time: the only way to give someone four files
was to open the dialog four times. A multi-selection now offers Share in
its context menu and in the mobile selection bar, and UIShareModal takes
a list of items.
With several items the access list folds into one row per person rather
than repeating per file, so a grant, a mode change or a revoke is one
decision about a person. A row says what it can't otherwise show: how
much of the selection the person reaches ("On 2 of 4 items", with an
"Add to all" that extends them), a mode select that rests on a
placeholder when their grants disagree rather than presenting one item's
mode as the batch's, and inherited grants left uncontrollable where they
belong. Changing a mode touches only the items the person already holds
— nothing here widens access without saying so.
Requests are chunked to the documented 50-items-per-request cap, and a
selection past it skips the per-item listing instead of firing one
request per file on every refresh.
The eligibility rule the Share entry already used moves into a shared
can_share() helper, since the multi-select menu needs the same answer
for every row.
Merge the .md extension case into the plain-text branch so it suggests ['editor', 'code'] instead of ['markus', 'editor', 'code']. Update the related test to use 'viewer'/'png' as the built-in guard example since 'markus'/'md' no longer applies.
- Add modelLookupNames() and route every provider's list() through it so
repeated or self-referential aliases can't double-list a model.
- Stop the chat/image/video drivers' #buildModelMap from writing through
to the module-level catalog constants (ids/aliases accumulated across
instantiations); bucket copies preserve the models() wire shape.
- Remove self-aliases and exact duplicates from the claude, deepseek,
groq, mistral, and openai catalogs; fold the duplicated
gemini-3.7-flash entry — this corrects the live cached-token price
from 8 to 7.5 and the knowledge date to 2026-03 (the previously
effective first entry carried stale values).
- ClaudeProvider: derive default max_tokens from the resolved entry
instead of a second name-based lookup that capped dated ids at 4096.
- Add cross-provider catalog invariants (modelCatalogs.test.ts) with a
disk-scan registration guard, plus regression tests for catalog
non-mutation, alias-dedup routing, wire shape, and max_tokens ceilings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs: remove Discord and Reddit links
Drops both links from the docs README (header row and Support list) and
from the footer generated on every built docs page.
* docs: remove Mastodon link
A share email named the file but had nowhere to go: the only link was "Open
Puter", and finding what someone shared meant hunting for it under Shared. Each
named file in the digest now links to itself, and a notification covering a
single item points at that item.
The link carries one parameter, the masked path a recipient is already given:
https://puter.com/?shared=%2Falice%2F<uuid>%2Freport.txt
Its second segment is the uuid, so there is no second copy of it to disagree
with the path, and the GUI can still find the entry when a rename has left the
name segment stale - it stats the path, then falls back to the uuid.
Built from the owner, uuid and name rather than from `ResolvedShare.path`.
That path is masked for whoever made the request, and the issuer owns the
entry, so it comes back as the owner's *real* path - mailing it would tell the
recipient which folders the owner keeps things in, which is the one thing
masking exists to prevent. A test asserts the real path never reaches the mail.
`digestLines` now returns `lead`/`items`/`trail` beside `what`, so the template
can put an anchor around each name while Handlebars keeps escaping the names
themselves; the URL is machine-built from the configured origin and one encoded
path, so it stays literal. Concatenating the parts reproduces `what` exactly,
which a test pins - the linked and sentence forms must not describe different
shares.
Notifications carry the masked path rather than a URL: the recipient is already
in the GUI, which opens the item in place instead of reloading. Only a
single-item notification gets a target; folding into a group drops it rather
than picking one of five.
In the GUI, `?shared=` joins `?download=` and `?app=` as a param that keeps the
desktop booting at `/`, and the handler reuses the `/@user` public-folder flow -
extracted to `open_path_target`, which carried a TODO asking for exactly this -
so a file opens in its associated app and a folder in an explorer window. The
param is stripped from the address bar first, so a reload lands on the desktop
rather than opening the item twice.
Invites are deliberately not linked: there is no account to route to yet, and
the invite's own call to action is to create one.
Rolling-deploy safe: a digest entry queued before this has `names` and no
`items`, and still flushes - without links. New entries write both, so a node
on the previous build can flush them too.
Both were left as follow-ups on the previous commit.
Thirteen `puter.perms` method pages went when the surface collapsed onto
`request(resource, details)`, and `requestAppData` moved to `/Perms/appData`.
All fourteen had shipped, so an external link or a bookmark landed on nothing.
The build already generates meta-refresh redirects from a map; they are entries
in it now. Pages that only ever existed on the branch that removed them are
deliberately absent — nobody can hold a link to a URL that was never published.
Six documented APIs annotated a code block with a playground id but had no
example file behind it, so the build warned on every one and the Try-it link
resolved to nothing: `fs.share`, `fs.unshare`, `fs.listShared`, `fs.getShares`,
`ai.txt2speech` with Speechify, and `ui.showFeedbackDialog`. The files are the
documented blocks themselves, extracted rather than rewritten, so the example
and the page it appears on cannot drift. Each is indexed in `examples.js` beside
its siblings.
The build now runs clean: 12 warnings to none.
`app-root-dir:` was listed among the raw permission strings, but it reads as
nothing during a permission check - the rewriter yields the sentinel outside a
grant, so `check()` reports it ungranted even once it is and `request()` prompts
every call. Both contradict what the same page promises. The `'appRootDir'`
resource is the supported path and asks the server instead; the list now says so
rather than implying the string behaves like its neighbours.
The playground listed no perms examples at all: six files existed but nothing
referenced them in examples.js, so a freshly-reshaped API shipped with nothing
runnable. Adds a Perms group after UI, matching the sidebar's order, and two
examples for the features that had none - one prompt covering several resources,
and check() deciding whether to prompt.
