OpenRouter and Together reject a request whose prompt plus max_tokens
overflows the model's context window. Both providers retried by deleting
max_tokens, which threw away the output cap the credit gate had sized to
the caller's remaining balance and let the retry run to the model's full
output limit with no second gate and no new hold.
The retry now goes through a shared helper that sizes a new cap from the
window and input count the rejection reports, falling back to the model's
declared context and a doubled prompt estimate, and never exceeds the cap
the gate set. When no window can be determined or no output fits, the
original rejection is rethrown instead of retrying uncapped. The rejected
params are copied rather than mutated, so the first attempt's record is
not rewritten after the fact.
Also corrects the estimator's own comment, which described the mean of
two approximations as a deliberate halving.
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
A Replicate prediction that ran and ended `failed` reaches the provider as a
plain Error with no HTTP status, so the driver-boundary translator could not
classify it and it surfaced as an unhandled 500, a critical alarm, and an
on-call page. Most of these are the model's content filter refusing the
user's prompt.
Wrap the run call and classify the failure: content-filter refusals become
a 400 with `errorCode: moderation_flagged` (the code chat refusals already
use); anything else becomes a 502 `upstream_failed`, which the alarm gate
skips. Status-bearing SDK errors pass through untouched so the boundary
translator keeps handling them. Upstream messages are stripped of markup and
bounded so an HTML error page can no longer ride into a response body or an
alarm signature.
Documents the codes callers can now act on in the txt2img reference.
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* feat: app-scoped share listing and revoke surface (PUT-1670)
* fix: address review on app-scoped share surface
- Scope the uid-addressed revoke to the named row: only that row's
issuer's grant is withdrawn, and only that one invite cancelled —
an app or owner addressing one row no longer takes another issuer's
grant on the same (item, recipient) pair with it.
- Delete a pending row directly on uid-addressed revoke, so an invite
whose address registered but never claimed can still be withdrawn.
- Read the legacy `issuerAppUid` data key in the SQL app filter and
grouping, alongside the unified `issuedByApp`.
- Refuse malformed `appUid` input (duplicated param, empty string)
instead of silently listing everything, and refuse app-listing
cursors that decode but name no appUid.
- Derive the acting app from `effectiveApp` alone, per the actor
contract; drop the second derivation site.
- Pin the attribution semantics with tests: one row records one
issuance, so re-sharing the same pair re-attributes it to whoever
issued last, in both directions.
- Soften the uniform-404 docblocks to what the gates actually answer.
* feat: readable grant audit trail (PUT-1674)
* fix: cover the apps summary's no-app-group first page (PUT-1670)
listOutboundApps sorts the no-app group first via an empty-string
sentinel. Add a regression test pinning that a first page (no cursor)
actually returns it, and that the cursor it hands back resumes past it
into the app-keyed groups rather than skipping or repeating.
* feat: global outbound share listing (PUT-1664)
* fix: address review on outbound share listing
- Check share-row liveness per (holder, entry, issuer) so a grant
withdrawn outside unshare doesn't stay listed while another issuer
still reaches the same holder; batch the permission reads across the
whole page instead of per holder.
- Retire a revoked issuer's unclaimed invites in the revoke cascade,
and hide invites whose issuer lost their authority at read time.
- Unify the pending/active app-attribution key on `issuedByApp` and
dual-read the legacy `issuerAppUid` spelling.
- Add the missing share issuer index (sqlite, postgres) and correct
the listOutbound plan comment.
- Refuse cursors that decode but name no id instead of silently
restarting from page one.
- Consolidate the five hand-built ResolvedShare literals and the two
listing endpoints' parse/shape code.
- Ship the SDK surface: puter.fs.listSharedByMe() with docs, types,
suite coverage, and the rate-limit page entry.
Register claude-fable-5-1 in the Claude catalog and gate it in the
provider the same way as Fable 5: no sampling params, effort via
output_config, adaptive thinking with summarized display. The bare
claude-fable / claude-fable-latest aliases move to 5.1, matching how
the Opus aliases moved when Opus 5 landed.
Fable 5.1 keeps Fable 5's $10/$50 per MTok, 1M context and 128K output,
but bills cache reads at 0.025x input instead of the 0.1x every other
Claude model uses, so the catalog row carries its own rate and a test
pins it.
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
The shell renders its anonymous markup — the marketing homepage, an
`/app/<name>` landing — off the session cookie alone, and that cookie is
set with no maxAge, so a browser drops it on quit while the GUI's
localStorage token lives on. A returning user is served the anonymous
page and the GUI only tears it down once `whoami` answers, a network
round-trip after first paint. That teardown is the flash.
Gate it before the paint instead. The shell now emits, as the first thing
in <head>, a rule hiding `.hide-if-logged-in` under an <html> class that
an inline script adds iff `auth_token_v2` is in localStorage. The rule is
already in the cascade when the markup is parsed, so a browser holding a
token never paints it at all.
`initgui` settles the guess the token represents: `whoami` confirming the
session removes the nodes outright (replacing the old `#appLanding`
removal), and no session — none stored, or one `whoami` rejected — drops
the class so the markup comes back. The gate carries its own 12s failsafe
so a bundle that never boots can't strand a blank page.
SEO is unaffected: the HTML is byte-identical for every client, nothing
branches on user-agent, and a crawler has no stored token so it never
adds the class. Unreadable storage fails open the same way.
Anonymous markup opts in with `class="hide-if-logged-in"`, which
`home.html` already carried.
Drop the delayed fade-out when a logged-in user reaches the GUI and remove the landing overlay immediately instead. This avoids the overlay sitting above the email/phone/card verification gates because of its z-index.
Remove the /app landing overlay once `whoami` confirms the user is authenticated, even if the server rendered the page as anonymous because the session cookie was missing. This prevents the overlay from covering the email, phone, and card verification gates during localStorage-based session restores.
On device-phone/device-tablet, style.css pins every .window to
z-index 9999999 !important — including the fullpage dashboard window.
The Add Existing Account login dialog opens with backdrop: true, and
the .window-backdrop wrapper is not a .window, so it kept its small
inline z-index and rendered entirely behind the dashboard.
Pass stay_on_top: true like every other backdropped dashboard dialog,
and make the Forgot-password child dialog inherit the flag so it does
not open buried under the now stay-on-top login window's backdrop.
Verified with Playwright on iPhone 13 emulation and a desktop
viewport: login and recover-password dialogs stack above the
dashboard in both.
The app shell fed the row's full description into <meta name=
"description">, og:description and twitter:description. Long
descriptions now collapse whitespace and cut at a word boundary with an
ellipsis, never exceeding 150 characters.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The JSDoc claimed non-app actors are rejected; they are passed through, and
the check reads `actor.app` rather than `effectiveApp`, so an app in the
token chain is invisible to it. Callers depend on the pass-through — routes
combining `adminOnly` + `allowedAppIds` are reached with an admin's browser
session, and the dev-account surface is called from the desktop's own
session — so the docs are corrected instead of the behavior, on the gate and
on the `allowedAppIds` route option. Tests cover the gate per actor shape
(user session, worker, full-access token, app-issued access token,
app-under-user allowed/disallowed) and the admin composition.
A live audit of every keyed provider's /models endpoint against the hardcoded
catalogs found no stale entries but large gaps. This backfills them under four
rules: nothing vendor-deprecated, nothing without a price confirmed on the
vendor's official pricing page (each entry's source was recorded during
review), nothing absent from the live /models listing, and nothing that fails
a live routing probe.
Added: 46 Alibaba entries (qwen3/3.5/3.7/3.8 families, VL/omni/MT lines, and
Model Studio's hosted GLM/DeepSeek/Kimi third-party models) plus 9 dated
aliases; OpenAI chat-latest and gpt-4o-2024-11-20 plus 16 snapshot aliases;
Gemini gemma-4-31b-it and gemma-4-26b-a4b-it (vendor-documented free tier)
plus rolling -latest aliases; Mistral-hosted zai-glm-5-2 and a
mistral-medium-3.5 alias; deepseek-v4-flash-vision-exp; glm-5.3-flash.
Culled by the rules: 15 vendor-deprecated OpenAI entries (the 3.5/4/4-turbo
legacy line, gpt-4o-2024-05-13, o1-pro, four chat-latest predecessors, the
5.x codex line — deprecations page, most shut down 2026-10-23) and dated
aliases onto the deprecated o1/o3-mini/o4-mini; qwen3-vl-flash-2025-10-15
(live routing probe returned upstream 400 twice). Tiered Alibaba prices are
encoded at the base tier and busy-hour rates where time-of-day priced, noted
in comments.
Every surviving addition was verified end-to-end through a local deployment:
58/58 answered a live prompt, including all 46 Alibaba entries and every
spot-checked alias.
Not changed, flagged for maintainers: pre-existing o1, o3-mini and o4-mini
entries are now vendor-deprecated (shutdown 2026-10-23); the pre-existing
deepseek-v4-flash/-pro prices no longer match DeepSeek's current pricing
page; gemma-4 emits its own <thought> markup inline in content.
Co-Authored-By: Claude Fable 5 (1M context) <noreply@anthropic.com>
Live-probing magistral-small-latest showed the model inlines its reasoning as
answer prose in a flat string — no ThinkChunk content, no markers, nothing a
client can separate. Mistral's chunked thinking shape is requested via
`prompt_mode: 'reasoning'`, which the provider previously dropped on the
floor: there was no way to even ask for it.
`custom.prompt_mode` now forwards to the SDK's `promptMode`, following the
BytePlus custom-params precedent. Opt-in rather than a default because the
API rejects the mode where the account/model lacks it ('Reasoning prompt
mode is not enabled for this model', code 3051) — verified end-to-end: the
3051 travels back through the stack, which also proves the parameter is
delivered. The moment Mistral enables the mode, the ThinkChunk content flows
into the existing splitter and comes out as `message.reasoning` and
`reasoning` stream chunks with no further changes.
Co-Authored-By: Claude Fable 5 (1M context) <noreply@anthropic.com>
The SDK-wide flag now mirrors the per-call option's tri-state. Left unset
(the default), the release-date policy applies: models released on or after
2026-09-01 are normalized to the OpenAI shape, older models keep their
vendor-native shape, and nothing rides the wire — the server's policy
resolution decides. Setting `true` force-normalizes every chat() call
regardless of release date; `false` disables normalization for every call;
either explicit value is sent on each call that doesn't set its own.
A per-call `normalize` overrides the flag in both directions, and assigning
`undefined` restores the policy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Creating a hosted subdomain gated `root_dir` on `write`, and hosting serves
everything under that directory with the ACL deliberately bypassed. So a
recipient of a `write` share could point a `*.puter.site` subdomain at the
owner's folder and make the subtree world-readable — continuously, covering
files the owner added later, with the row under the recipient's account where
nothing the owner can list would show it. `update` had the same gate for a
changed `root_dir`.
`#checkPublishAccess` now decides both: the actor's own tree still takes
`write`, anyone else's takes `manage` — "Can edit & share", the level that
delegates the decision.
Keyed on who owns the entry rather than asking for `manage` outright, which is
what the ticket proposed. `manage`'s is-owner implicator declines to answer for
app actors, so a flat `manage` would refuse every app publishing a directory
its user handed it, with no way for the app to obtain the grant. The write
check still runs first — it is what masks a directory the caller cannot see as
a 404 — and `manage` satisfies every lower mode, so the order costs a
manage-holder nothing.
The GUI's Publish As Website item reuses the own-it-or-`manage` answer it
already computes for sharing, so it is not offered where this would refuse.
Docs state the rule on `hosting.create()` and in `share()`'s level list.
Regression tests fail without the driver change: a write-share recipient is
refused on create and on repointing an existing subdomain, while `manage` and
the actor's own directory are accepted.
Opening a file you hold read-but-not-write access on (a read-only share)
failed silently. For such a file the backend correctly omits write_url from
the /open_item signature, but launchApp appended it to the app iframe URL
unconditionally. URLSearchParams coerces undefined to the string \"undefined\",
so the app received puter.item.write_url=\"undefined\" — truthy, so the editor
believed the file was writable, and an invalid URL, so it broke on open. Only
read-only shares hit this; files you can write carry a real write_url.
Extract the puter.item.* param building into append_signed_item_params and
guard the write_url append so it is only added when present. launchApp.js is
too coupled to UIWindow/jQuery/window globals to unit-test directly, so the
pure helper carries the logic and its own test, matching the helpers/ pattern.
Regression test pins both directions: a read-only signature omits write_url
entirely (no \"undefined\"), and a writable signature still forwards it.
The fs-path-to-uid permission rewriter split permission strings on the raw
\`fs:\` substring instead of parsing on component boundaries. A path can
itself contain \`fs:\` — a home dir named \`…fs\`, or \`fs\` in the mode
position — and the raw split mistook that for the mode delimiter.
Crafting \`fs:/<victim>fs:junk:read\` made the split do two things at once:
it dropped the \`junk:read\` mode, and it consumed the trailing \`fs\`, turning
the harmless-looking nonexistent path shown in the consent dialog into the
victim's real home path. The rewriter then stored a bare \`fs:<home_uuid>\`,
which subsumes every mode via parent-permission matching — full
read+write+delete from a request the user saw as \`junk:read\`.
Parse with PermissionUtil.split() in both matches() and rewrite(), matching
every other permission parser, and locate the \`fs\` component by position
(after an optional \`manage\` prefix). The path component is now exactly what
sits between \`fs\` and the next unescaped colon, and all trailing components
are preserved as the mode — so an embedded \`fs:\` is data, never a delimiter.
The crafted input now addresses the nonexistent \`/<victim>fs\` and 404s.
Regression tests pin both halves: the exact PoC must 404, and an \`fs:\` in
the mode position must survive the rewrite instead of collapsing to a bare
permission. Both fail on the old rewriter and pass on the fix.
`env = 'app'` was decided by the presence of a `puter.app_instance_id` query
parameter and nothing else, so a crafted link put any page that loads the SDK
into app mode — and app mode is what makes the URL's `puter.api_origin`
authoritative for every credentialed call.
App mode now also requires the document to be framed. The GUI only ever
launches an app into an iframe, so this costs a real app nothing while a
top-level document carrying the parameters is treated as the third-party site
it is. It is not an attestation that the framing document is the GUI — a
cross-origin ancestor's identity is not readable — so the token paths carry the
rest:
- The `web` boot branch adopted a stored token without consulting the origin it
was bound to, which is what completed the fixation: one link plants a token
bound to an attacker's origin, and every later visit adopted it. It now
applies the same binding rule the app branch does, and drops a token that
fails it rather than leaving it to be re-read.
- `signIn()` had no env guard, and in app mode delivered a real token to
whatever `puter.api_origin` the launching URL named. Apps get their token
from the session that launched them, so it now rejects there with
`not_available_in_app`. Nothing internal reaches it in app mode:
`authenticateWithPuter` and both implicit-auth call sites already gate on
`env === 'web'`.
- The cross-origin-isolated branch polled `${this.APIOrigin}/login/wait` and
adopted whatever came back. Pinned to `defaultAPIOrigin`, the same way the
popup and its message handler already pin `defaultGUIOrigin`.
Backward compatibility: no signature, response field or existing error code
changes. The only behaviour a caller can observe is the new `signIn()`
rejection, which replaces a call that could not have worked correctly.
Covers the SDK side of the parameter PUT-1395 and PUT-1427 closed on the GUI.
Review of the previous change found three of its claims unmet.
The image-generation crash it reported fixed is still reachable. The assert was
scattered across three helpers, and Gemini and OpenAI call `isHttpUrl` directly
on `input_images` without going through any of them — so two of seven providers
still 500 on a non-string. `isHttpUrl` now refuses a non-string itself, and the
shape is settled once in `ImageGenerationDriver.generate`, where the driver call
arrives, rather than per helper. That also covers `input_images` that isn't an
array, which produced a different crash per provider.
The sixth case in the ticket, previously unlocated, is
`Messages.js` reading `tool_call.function.name` with no guard — reachable with
`{"messages":[{"role":"assistant","tool_calls":[{"id":"x"}]}]}`. Guarded, along
with the same shape in `make_claude_tools`: a TypeError there carries no status,
so the retry loop reads it as a provider failure and marks the route unhealthy
for every caller.
`#hardExpiryFromExpiresIn` returning null for a bad type moved the failure past
the session INSERT, leaving an orphaned non-expiring row and still answering
500. Reverted; the controller guard is the fix, now covering fractions,
negatives and unparseable durations rather than only wrong types.
Also: the batch write handlers check that the body is an array but not what is
in it, so a null element 500s the same way; `#requireObjectBody` accepted an
array despite its name; `handleCreateAccessToken` destructured a body that may
be absent; and two AGPL notices had been rewrapped with a Markdown link.
A socket was checked once at handshake and never again. Nothing in the backend
disconnected one, so logout-everywhere, password reset, session revoke and
suspension all left every connection streaming legacy FS entries, upload paths
and notification bodies — up to 400 per account — on a credential that had
already been revoked.
Three gaps, three fixes:
- The handshake skipped the suspension and pending-verification checks every
authenticated HTTP route gets. `decideSocketAuth` now applies both.
- `revokeCascade` reports which rows it revoked, AuthService announces that as
`auth.sessions.revoked`, and SocketService drops the account's room. The
whole room goes, not just the revoked session: narrowing it would need
`fetchSockets`, which the adapter builds on `serverCount()` — and that calls
node-redis's `send_command`, which ioredis does not implement. A connection
whose session survived reconnects on its own and re-authenticates.
- A bulk suspension writes `user.suspended` without touching `sessions`, so no
revoke fires. A five-minute sweep re-verifies each live socket's token and
drops the ones that no longer authenticate. De-duplicated by token, since a
browser's tabs share one.
Each of these read a field off caller input that wasn't the shape the code
assumed, threw a TypeError, and was served as a 500 with a critical page.
- POST /fs/write and /startWrite: a request whose body never parsed left
`req.body` undefined, and the first read of `fileMetadata` threw.
- POST /drivers/call, image generation: `input_image` / `input_images` entries
are documented as strings but nothing checked, so a number or an object
reached `.startsWith`. Type confusion on caller input, so reachable on
demand rather than by accident.
- POST /auth/create-access-token: `expiresIn` went to the expiry parser
unvalidated, where anything but a string or a number has no `.trim`.
- POST /login/wait: destructuring `session` out of an absent body threw.
Same class as the POST /login body ticket; that one covers /login itself.
The only thing in front of the bcrypt compare on the DAV host was the 600/min
request ceiling, keyed on a fingerprint that rotates with client-controlled
headers. /login guards the same credential with a captcha and two much tighter
buckets; DAV had neither, which left password and TOTP guessing viable from a
host that answers any origin.
The request ceiling can't double as a credential ceiling — a working DAV client
resends its credentials on every request — so the new buckets count only
verifications that failed: 10 per account and 50 per address per 15 minutes,
sized like /login's. They're read before the compare, so an exhausted bucket
costs no bcrypt round, and successful requests never draw them down. `-token`
attempts are held by the address bucket alone; bucketing them per account would
let bad tokens lock out good ones.
Reading a bucket without spending from it is new, hence `peekRateLimit` and the
matching `peek` on all three backends.
Also documents the DAV limits, which were undisclosed.
Both signup paths read the x-forwarded-for header directly for the
puter.signup.validate event, the puter.signup.success event and the
signup_ip_forwarded column. Behind an appending proxy a client can prefix that
header with anything it likes, which gives every per-IP abuse signal a fresh
bucket per request.
req.ip is the value `trust proxy` resolves, so it is the one the server can
stand behind; server.ts already uses it for the ip.validate gate for exactly
this reason. All three now share one derivation, so a per-IP counter is written
and read under the same key. The raw chain is still recorded, but only as
audit_metadata.ip_fwd, where nothing keys on it.
Rows written before this hold whatever the proxy appended, so per-IP velocity
over the trailing window is undercounted until they age out.
All three block routes set only `subdomain`, `requireVerified` and a rate
limit, and the controller's actor check accepts anything carrying a user, so an
app-under-user token passed. Unlike `share` and `revoke`, nothing else bounded
them: an app the user authorized once could read exactly whom they had blocked,
then clear the list — blanket switch and per-username — and reopen the channel.
A block list is a personal safety control, not something an app can otherwise
reach, so it falls outside the "an app can share what it can already reach"
carve-out the other share routes rely on. `requireUserActor` on all three, and
deliberately without `allowFullAccessToken`: this is security management, which
stays closed to every access token, personal ones included.
The only caller is the desktop's Blocked Senders window, which sends the user's
own session token.
Tests fail without the gate: an app token minted through
/auth/get-user-app-token is refused on all three routes while the user's own
session still succeeds, and a metadata check pins the gate on every block route
so a fourth cannot quietly ship without it.
The comment claimed the fallback loop cannot be driven from a test — false
since ChatCompletionDriver.test.ts gained a two-attempt fallback test that
drives the actual loop. Say where that test lives instead.
Co-Authored-By: Claude Fable 5 (1M context) <noreply@anthropic.com>
Third triple-check round. Ten findings were put through independent skeptics
first; six did not survive — pre-existing on main, inert, or resting on a false
premise — and are not acted on here.
Mistral streamed thinking now goes to `reasoning` on every path. The previous
commit gated the split, which made this the only place in the repo where
chain-of-thought reached the visible text channel, and made Mistral the only
provider whose streamed chunk *types* depend on a response-format flag. Every
other reasoning path routes thinking to `reasoning` unconditionally —
ClaudeProvider's thinking_delta, the DeepSeek/OpenRouter rename, and this
branch's own Responses summary-delta handler. Removing the gate restores that
uniformity and makes the documented promise that streaming is unaffected by
normalization true again; the two opposing tests collapse into one that runs the
same fixture with and without `normalize` and asserts identical event streams.
Docs stop claiming older models are unchanged. Four reasoning fields were made
consistent across all models, ungated, and one of them removes a field: on
non-streaming responses `message.reasoning_content` is now `message.reasoning`.
chat.md gains a table naming all four so a caller reading `reasoning_content`
learns why it disappeared, instead of reading that nothing changed for them.
Adds the driver-level fallback test. Writing it surfaced that the invariant it
was meant to assert is false and always was: the driver rewrites string
`content` into text blocks in place on the caller's own messages
(`normalize_single_message`, pre-existing) before any provider runs. The test
now asserts what is true and load-bearing — both attempts receive the same array
reference, and the reasoning artifacts survive attempt 1 so attempt 2 can still
replay them.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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>
