Anthropic accepts at most 4 blocks carrying cache_control per request. When the
client had already spent that budget, the re-anchor added a 5th marker and the
request was rejected with a non-retryable 400 that the failure path treated as
an account problem, retrying the same malformed body across the whole pool until
every account locked.
anchorClaudeCache now normalizes bare-object content, strips the invalid
cache_control carried by defer_loading tools, pins the 1h head anchors on the
last system block and last cacheable tool, then trims an over-budget body to 4
markers. The trim holds those head anchors and fills the remaining slots with the
tail-most message markers: a plain "keep the last four in document order" rule
drops the anchors first even though they lead document order, and skipping the
re-anchor at a spent budget left system/tools on the 5m default instead of 1h.
Some clients send content as a single block object rather than a one-element
array. Such a turn was dropped or zeroed on every leg that reads messages,
silently losing conversation history. normalizeMessageContent wraps it as a
one-block array on all four paths, and hasValidContent keeps it.
Cline (api.cline.bot) wraps non-stream chat completions in
{"success":true,"data":{...choices...}}, which both the dashboard model-test
ping and the proxy non-stream path read at top level, producing "Provider
returned no completion choices for this model" (#3644). Unwrap the envelope
before usage extraction and response translation; the error envelope
({"success":false,...}) never matches and passes through untouched.
Scoped through `transport.quirks.clineEnvelope` so only cline/clinepass opt
in — no other provider's response body is ever rewritten.
Also adds a live Cline catalog: `fetchClineRawModels()` is shared between
`resolveClineModels()` (full catalog, including free-tier ids such as
z-ai/glm-5.3-flash) and `resolveClinepassModels()` (cline-pass/* only), wired
into /v1/models, the per-provider models route, and the combo selector's
model picker with the static catalog kept as fallback.
Refreshes the dead api-airforce free models (anthropic/claude-3.7-sonnet,
moonshot/kimi-k2.6, google/gemini-2.5-flash) with the live gpt-oss-120b,
gpt-oss-20b and kimi-k2.7-code, plus passthroughModels, forceStream and a
suggested-models filter.
Cline/ClinePass requests failed with HTTP 401 ("Please make sure you are using
the latest version of Cline and re-authenticate your Cline account", #3230 /
#2333 / #3644). `getClineAccessToken()` unconditionally prefixed every token
with `workos:`, which is correct for Cline OAuth access tokens (WorkOS JWTs)
but wrong for ClinePass API keys — those are opaque strings (e.g. `clp_…`)
that the API accepts only verbatim, so the `workos:`-prefixed value was
rejected.
Only prefix tokens that look like a WorkOS JWT (`eyJ…`); API keys and other
opaque tokens pass through untouched, and an existing `workos:` prefix is
never doubled.
Also register `clinepass` in the token-refresh handlers. ClinePass shares
Cline's WorkOS auth endpoints, but without the entry expired ClinePass OAuth
tokens were never rotated, so every request kept 401ing. Finally, list
`apikey` first in the ClinePass `authModes` (ClinePass is meant to be used
with an API key from app.cline.bot/settings/api-keys), and add an "Import
from /models" button that pulls the live Cline catalog into custom models.
DeepSeek's Anthropic-compatible endpoint accepts only the built-in
web_search_20250305 / web_search_20260209 tools and rejects client-defined
`custom` tools (MCP / Read / Bash) with HTTP 400 "unknown variant `custom`".
The generic non-Claude filter in prepareClaudeRequest dropped the offending
tools but also dropped the web_search_* ones DeepSeek does accept.
- Add an opt-in per-provider transport quirk `claudeSupportedToolTypes`; when
declared it becomes a strict allow-list for Anthropic tool `type` values
- Stop stripping the `type` discriminator from surviving tools under that
quirk, since DeepSeek needs it to route built-ins
- Declare the quirk on the deepseek transport with the two web_search_* types
- Providers without the quirk keep the previous filter and normalisation
behaviour byte-for-byte; openai-format targets never reach this path
Codex's /responses validator has no Unicode property escapes, so a tool
`pattern` containing `\p{...}` 400s the whole request with `Invalid schema
for function ... is not a 'regex'` — identically on every account, costing a
full combo failover per turn (#3922).
- Add open-sse/utils/codexToolSchema.js: copy-on-write walk that drops only
`pattern` values carrying a property escape, returning the original
reference when nothing changed so the caller's schema stays intact for a
retry against another provider
- Treat `properties` keys as property names, so a field literally called
`pattern` is never read as the schema keyword; skip escaped literals via
backslash-parity counting
- Apply it in normalizeCodexTools for both function and namespace sub-tool
parameters, and log the strip count via dbg
- Add three cases to tests/unit/codex-tool-normalization.test.js
kiro.dev rejects any body carrying a top-level systemPrompt with
400 REQUEST_BODY_INVALID. The translators stopped emitting the field in
v0.5.59 (the prompt travels in the first user turn via contentPrefix),
but two paths kept writing it back downstream of the translator:
- rtk/systemInject.js::injectKiroSystem() appended the RTK prompt to
body.systemPrompt, so every kr/ model failed whenever an RTK injector
(caveman, ponytail) was active. It now appends to the first history
user turn's content (else currentMessage), reusing
dedupStringAppend/hasPrompt so retries stay idempotent.
- executors/kiro.js::appendRepairInstruction() wrote the tool-call repair
instruction to systemPrompt on the retry, turning every repair into a
hard failure. It now appends to currentMessage.userInputMessage.content.
isKiroBody() no longer requires a string body.systemPrompt — that marker
is gone from the wire shape — and sniffs the conversation turn shape
instead, keeping the stray-conversationState guard intact. Stale comments
in both kiro translators corrected: the systemPrompt local is only a
session-replay cache key, not a wire field.
Also drops the mirror/rollback repair heuristic the injector no longer
needs: net -52 lines.
Fixes#3641, #3845, #2890, #2901, #2939, #3109, #3459, #3749
- Coalesce Qoder's empty finish-in-delta frame with the later choices:[] usage
frame so OpenAI and Claude clients receive prompt_tokens, completion_tokens
and cache-hit tokens (the dashboard already saw them)
- Upload inlined images through /api/v2/image/upload like qodercli, and stub
oversized non-image files instead of stuffing 30MB+ data URIs into
agent_chat_generation
- Emit response.completed -> response.usage for chat-native upstreams so
/v1/responses clients (Codex CLI, sub2api) no longer log 0/0/0
- Keep Claude message_delta.usage working when usage arrives without choices[0]
- Escalate to the smallest advertised Qoder context tier (200K/400K/1M) when
the estimated prompt no longer fits max_input_tokens
- Pass apiKey for PAT connections and list hidden enable:false catalog keys
from /v1/models
- Add gpt-image-1.5, gpt-image-2, gpt-image-2.5, gpt-image-2.5-flare and
gpt-image-2.5-sunburst as Codex image models with multi-image support
- Add gpt-image-2.5, gpt-image-2.5-flare and gpt-image-2.5-sunburst to the
OpenAI provider catalog
- Route tool-backed image models through the Codex responses model while
passing the selected model to the image_generation tool, pinning
tool_choice and deriving generate/edit from the presence of references
- Cover the Codex gpt-image-2.5 request shape with a unit test
Video generation was xAI-only. Adds an adapter layer under
open-sse/handlers/videoProviders/ so /v1/videos/* can target OpenRouter or
Google Cloud credentials. A provider with no adapter keeps the exact previous
behaviour (raw body to {baseUrl}/{action}, poll {baseUrl}/{id}, verbatim
passthrough), so the xAI path is unchanged.
- openrouter: async job shape identical to xAI; creation POSTs to the /videos
collection root (no /generations suffix) and the registry HTTP-Referer /
X-Title headers are applied. Bodies pass through verbatim.
- vertex: two-way translation, since Veo does not speak the OpenAI-ish videos
shape. create -> :predictLongRunning { instances[], parameters{} }, poll ->
:fetchPredictOperation (Veo has no REST GET poll). The operation resource
name is base64url-encoded into the job id so GET /v1/videos/{id} stays a
flat path. Access tokens are minted from Service Account JSON via the
existing refreshVertexToken; raw API keys are rejected up front. The
operation response maps back onto the { id, status, video, videos } shape
clients already poll.
- videoCore: the request plan is rebuilt per attempt, so the 401 -> refresh
once -> retry once path picks up the refreshed token. Adapter validation
errors return 400 before any upstream call, so a malformed request can never
create a billable job.
- videoGeneration: GET /v1/videos/{id} resolves the provider from the pinned
x-connection-id connection, then ?provider=, then falls back to the xAI
default.
- registry: openrouter and vertex gain videoConfig, the video serviceKind and
video-kind models (Veo 3.1 / 3 / 2, Sora 2 Pro, Seedance 2.0).
The image handler's `version` header was commented out, so Codex image
requests reached chatgpt.com without the Version identity the backend
expects. Restore it and route every Codex identity header through one
constant.
The CLI version now lives on registry codex.transport as `cliVersion`
(the same pattern gemini-cli uses) and is re-exported as CODEX_CLI_VERSION,
so the registry User-Agent, the image handler and the connection test can
no longer drift apart. Bumped 0.136.0 -> 0.154.0 (current stable).
Co-Authored-By: Claude Code <noreply@anthropic.com>
Add the models the provider docs now list but the registry lacked:
chat/completions glm-5.3, kimi-k3, deepseek-flash, longcat-2.0,
hy4-preview, hy3
+ /messages qwen3.8-max, qwen3.8-flash
responses only grok-4.6, gpt-5.6-luna
Endpoints follow the table at https://opencode.ai/docs/go/. chat-only
models stay on the sourceFormat-matched transport guard so a Claude
client is never routed to /messages for a model that lacks it.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Clear stale connection health state (modelLock_*, backoffLevel,
rateLimitedUntil, errorCode) whenever a connection is explicitly
marked active after successful validation or OAuth re-login.
Closes#3810
- Add muse-spark-1.2-contributor as responses-only model on OpenCode Go
- Normalize object tool schemas without properties in OpenCode Go executor
- Make fallback Responses call_ids unique across same-millisecond calls
- Make Responses output coercion fail-soft for circular and non-stringifiable values
- Group Antigravity Gemini text models into single 'Gemini (Flash / Pro)' quota
- Group Claude models into single 'Claude (Sonnet / Opus)' quota
- Prune stale or legacy model keys from hidden quota visibility list
Co-Authored-By: Claude Code <noreply@anthropic.com>
- Add muse-spark-1.3-contributor as responses-only model on OpenCode Go with dedicated executor
- Key Responses→chat streaming tool calls by item_id to prevent parallel tool calls merging into index 0
- Standardize tool coercions and call_id clamping in Responses API translation
- add a dedicated OpenCode Go executor that always sends x-opencode-session
- preserve a valid caller-provided native OpenCode session header
- translate downstream Agent session IDs into opaque, stable, Agent-scoped IDs
- forward the original provider session seed and client tool on both initial and credential-refresh requests
- Registry/constants: drop qmodel_preview/gm51model, add lite,
qmodel_38max (Qwen3.8-Max), qfmodel (Qwen3.8-Flash), gmodel (GLM-5.3),
gfmodel (GLM-5.3-Flash)
- capabilities: add PROVIDER_CAPABILITIES['qoder'] so opaque internal
ids resolve to their real models' context windows and limits
- executor: preserve image blocks instead of flattening away, convert
Claude-style image blocks, and hash images into chat_record_id
- tests: cover image preservation, data-URI and Claude-block conversion
- build(docker): use CN mirrors for apk and npm
Non-streaming codex traffic recorded cached_tokens: 0 even when upstream
prompt caching worked. The Claude-format branch (which OpenAI Responses
usage also matches) never read input_tokens_details, and the OpenAI
branch ignored a top-level flat cached_tokens. Read both in both
branches; Responses prompts are cache-inclusive so canonicalizeUsage
passes the value through without folding. 5 new regression tests.
Support single model lookup by replacing the one-segment models route
with a catch-all route that preserves capability kind paths while
allowing provider-prefixed IDs like cc/claude-sonnet-5.
Adds a client-side status filter (All / Active / Inactive / No
connection) to the Providers page, applied over the already-fetched
provider + connection list. Status derives from getProviderStats
(total, allDisabled); noAuth providers count as Active. Filter composes
with the existing search across all provider sections. Part of #3699.
Google's quota API can report remaining quota while generation endpoints
keep returning 429 (sprint/weekly dual-pool mismatch). handleAntigravityQuotaError
trusted remainingPercentage > 0 as healthy and returned null, causing 429 retry
loops across multi-account pools.
Add a strike-based circuit breaker to the optimistic and unavailable quota paths:
- After 3 strikes (429/409) within 60s for the same connection+model, cache-block
that pair for 15 minutes by synthesizing an entry in the shared RAM quota cache.
- Re-assert active strike blocks across refreshes so optimistic readings cannot
resurrect a broken pair prematurely.
- Reset strikes and clear synthesized cache entry upon successful request.
- Keep exact-resetAt handling for genuine 0% exhausted readings.
Closes#3681
Route all Muse Spark models (not just 1.2) on OpenCode Free to
/zen/v1/responses via isMuseSparkModel(), fixing HTTP 500 on
muse-spark-1.3-contributor-free. Declare vision:true on Muse Spark
models so image input is no longer stripped; register 1.3 in the
registry and capabilities. Scoped to opencode only — other providers
keep Chat Completions routing.
Closes four SSRF guard bypasses reported in #3714:
- Block alternate IPv6 encodings (hex format, NAT64, IPv4-compatible, IPv4-mapped) by parsing to 16-bit groups
- Normalize trailing dots on hostnames to prevent FQDN bypasses
- Add assertPublicUrlResolved() with DNS resolution to block wildcard DNS domains resolving to private/metadata IPs
- Add fetchPublic() to safely handle and validate HTTP redirects
- add claude-fable-5-1 to the Claude Code model catalog (1M context,
permanent adaptive thinking)
- centralize the spoofed Claude Code version and update both request
and billing identities to 2.1.257 (Fable 5.1 rejects < 2.1.251)
- send output_config.effort without the redundant thinking switch for
permanently adaptive models
- add regression coverage for capabilities, headers, billing identity
and adaptive-effort payload
# Conflicts:
# open-sse/providers/registry/claude.js
# open-sse/providers/shared.js
# open-sse/utils/claudeCloaking.js
# tests/__baseline__/providers-baseline.json
With the 1M-context beta enabled, Claude Code sends model: "claude-opus-5[1m]".
The marker is a client-side annotation — it matches no combo name, no alias and
no provider/model pair — so the request dies at model resolution with
"Invalid model format" and the client reports "There's an issue with the
selected model". Every request from that session fails until the beta is
switched off.
New open-sse/utils/modelMarkers.js exporting stripModelContextMarker(modelStr)
-> { model, contextMarker }. handleChat strips the marker before resolution and
normalizes body.model so downstream logging and translation see the real name.
Only a trailing marker is stripped, so a model whose name genuinely contains
brackets is left alone.
The capability itself travels in anthropic-beta: context-1m-2025-08-07, which
the default executor already forwards untouched — only the routing key needed
cleaning.
Fixes#3690.
Tests: tests/unit/model-context-marker.test.js (6 cases).
Restrict the legacy IDE-version override to generation endpoints so
catalog and other passthrough requests keep their original User-Agent
and metadata.ideVersion, letting newer Antigravity releases see current
models like Gemini 3.7 Flash in the MITM selector.
Fixes#3414
Anthropic validates server_tool_use.id against ^srvtoolu_[a-zA-Z0-9_]+$
and 400s the whole request when one does not match. A combo that falls
back to a provider with its own built-in tools (z.ai/glm emits
OpenAI-style call_ ids for analyze_image) leaves such blocks in the
history, so every later Claude turn fails.
Extend normalizeClaudePassthrough to drop those blocks (reusing the
existing loop), drop the paired tool_result / web_search_tool_result
referencing a dropped id, and drop empty text blocks plus messages left
with no content. Well-formed srvtoolu_ blocks and regular tool_use ids
are untouched.
buildUsage() only emits cache reads under prompt_tokens_details, so the
top-level-only read dropped the count for every Responses-format provider
(codex, grok-cli, ...), persisting cached_tokens: 0 and billing cache hits
at the full input rate. Mirror the cache_creation fallback already used
just above.
Registry/pricing/CLI catalog already had gemini-3.7-flash-{high,medium,low}
but MITM_TOOLS.antigravity.defaultModels was missing them, so the tiers
never showed up in the dashboard model-mapping table.
The 3000 ms timeout on /v1/compress was fixed, so busy or slow machines
timed out often and sent the LLM an inconsistently compressed body,
hurting prompt caching. Add a headroomTimeoutMs setting, thread it from
the chat handler down to compressWithHeadroom, expose it in the Token
Saver dashboard, and normalize invalid values back to the 3000 ms default.
Prefer the sourceFormat-matched runtime transport over a model's
declared targetFormat when both apply. MiniMax-M3 previously resolved
to a Claude-shaped body while being posted to the already-selected
OpenAI endpoint, silently dropping image_url blocks from OpenAI
clients. Fixes#3418.
Add a Zed usage handler so connected Zed accounts appear on
/dashboard/quota. Reads GET /client/users/me for plan, edit
predictions, optional hosted model requests and billing-cycle reset.
Render unlimited rows as "N used · Unlimited" instead of 0 / ∞, and
surface overdue-invoice / token-billing messages.
CommandCode returns errors as a type:"error" event inside an HTTP 200
NDJSON stream instead of a non-200 status, so the existing combo/account
fallback logic (keyed off response.status) never triggered and the error
text was streamed to the client as if it were content.
Peek the first NDJSON events before committing to a stream; on a
type:"error" event, abort and return a proper 4xx/5xx Response instead.
Normal streams are replayed losslessly (buffered prefix + rest of the
stream) through the existing translator, so the happy path is unchanged.
Add CommandCodeExecutor.parseError() so parseUpstreamError() can extract
a clean message/status from the synthesized error body.
translateResponse() short-circuited untouched on claude->claude streaming,
so OAuth-cloaked tool names (CLAUDE_TOOL_SUFFIX) leaked to the client and
every tool call was rejected as unknown. Add decloakStreamChunk(), the
streaming counterpart of decloakToolNames(), and call it on the same-format
path using the already-plumbed state.toolNameMap.
GLM quota parsing only accepted TOKENS_LIMIT and wrote every limit to a
single "session" key, so credit-based plans showed nothing and later
intervals overwrote earlier ones. Accept CREDIT_LIMIT too and derive the
quota key from the limit unit (5h session, 7d weekly, tokens, custom).
Moves the parser into its own usage/glm.js, re-exported from misc.js.
