- Chat (StudioConfigPane): add Provider + Model selects reusing the translator
hooks; order Endpoint -> Provider -> Model; ConfigState gains optional provider.
- Compare (CompareTab): add a user-prompt input and include it in the request
body; throttle per-column stream updates via requestAnimationFrame to stop the
UI freeze (was setColumns per chunk x N columns with an empty user message).
- Build (BuildTab + build/BuildWizard): redesign as a guided 3-step wizard
(What to test -> Configure -> Run) reusing ToolsBuilder/StructuredOutputEditor;
all run/tool-call handlers preserved.
- i18n: playground.build.* (18 keys) in en + pt-BR.
- MemoriesTab: auto-run health check on mount + poll every 30s (was manual-only).
- page: add enable/disable memory toggle (role=switch) via useMemorySettings.save({ enabled }).
- MemoryEngineStatus: show "npm install sqlite-vec" hint when vector store backend is "none".
- i18n: memory.memoryEnabled + memory.engine.vectorStoreInstallHint (en + pt-BR).
- Add compliance.eventTypes (36 labels) to en.json + pt-BR.json.
- ComplianceTab: render translated label via t.has/t fallback instead of raw entry.action.
- A2aAuditTab: render translated task state via a2aState* keys instead of raw task.state.
- Memory type/strategy dropdowns needed no i18n change — keys already exist; the
Phase 1 Select fix makes them render.
- search-tools: export modal no longer opens by default / stuck — guard on
exportOpen and drop the invalid isOpen prop the modal never read.
- logs: remove duplicate proxy/console tabs + SegmentedControl (dedicated
/dashboard/logs/proxy and /console pages already exist in the menu).
- memory: order tabs Memories -> Engine -> Playground.
- ui(Select): render children and suppress the placeholder/options branch when
children are provided — fixes the "empty" memory type/strategy dropdowns
(children were being shadowed by the component's own option list).
- test: source-level regression guards for all four.
Root cause: model string provider prefix (e.g. "xiaomi" from
"xiaomi/mimo-v2-flash") differs from the credential DB provider ID
(e.g. "opengate") when a combo target has a custom providerId. The
pre-selection and execution flows both looked up credentials using the
model-inferred provider, which didn't match any DB entry.
Fix A (chat.ts): checkModelAvailable now uses target.providerId when
available, falling back to modelInfo.provider.
Fix B (chat.ts): handleSingleModelChat now accepts runtimeOptions.providerId
and preferentially uses it for credential lookup instead of re-resolving
the provider from the model string.
Fix C (model.ts): add "xiaomi" alias for "xiaomi-mimo" so direct
(non-combo) model requests to xiaomi/mimo-* also resolve correctly.
The CPA executor now cloaks non-Claude-Code tool names (not just mcp_*), so the
prior "does not rewrite non-mcp_ tool names" assertion no longer holds:
my_tool is aliased to MyTool and restored on the response via _toolNameMap.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Addresses confirmed findings from an adversarial review of the prior commits:
- schema sanitizer: a truncation placeholder in a SCALAR annotation keyword
(description/title/pattern/format) was coerced to {}, which is itself invalid
draft-2020-12 and re-triggered the exact "input_schema is invalid" 400 the
sanitizer exists to prevent. Placeholders are now only coerced to {} in
subschema-expecting positions; scalar keywords are left untouched.
- schema sanitizer: numeric-string coercion is folded into
stripInvalidSchemaConstructs so it also covers contains / propertyNames /
additionalItems (which coerceSchemaNumericFields never visited).
- schema sanitizer: stop stripping the valid `default` keyword on the Claude
native/passthrough surface (the #1782 default-strip is a translator concern;
tool schemas here were previously forwarded verbatim). sanitizeClaudeToolSchema
is now a single stripInvalidSchemaConstructs pass.
- tool-name cloak: consult TOOL_RENAME_MAP / EXTRA_TOOL_RENAME_MAP before the
generic PascalCase fallback, so the CLIProxyAPI path uses the established
fingerprint-evasion aliases (subagents->SubDispatch, session_status->CheckStatus,
webfetch->WebFetch, ...) identically to the native path instead of weaker
first-letter casing.
- kill-switch: CLAUDE_DISABLE_TOOL_NAME_CLOAK is now honoured inside
cloakThirdPartyToolNames, so BOTH the native and CLIProxyAPI executor paths
respect it (previously only base.ts did); .env.example + ENVIRONMENT.md updated.
Regression tests added for each. Verified end-to-end through the live CPA path:
mixture_of_agents, subagents, and a tool carrying placeholder descriptions and
`default` values all return 200 with original names restored on the response.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The /dashboard/tools/agent-bridge page (Server Component) passed ALL_TARGETS
directly to AgentBridgePageClient (a Client Component). Each MitmTarget carries
a `handler: () => Promise<...>` function, which Next.js forbids across the
Server/Client boundary, raising at SSR time:
"Functions cannot be passed directly to Client Components ..."
This broke the whole page ("erro ao carregar").
Fix: introduce MitmTargetView = Omit<MitmTarget, "handler"> and pass a
sanitized array (ALL_TARGETS.map(({ handler, ...rest }) => rest)). The UI never
invokes handler, so behavior is unchanged. Adds a regression test asserting the
sanitized targets are function-free and JSON-serializable.
The native Claude OAuth guard in executors/base.ts is bypassed when
`upstream_proxy_config.mode = cliproxyapi` routes the request through the
CliproxyAPI executor — it has its own execute()/transformRequest() and never
reaches BaseExecutor.execute(), so the cloak/sanitizer never ran for that
(common) deployment. Wire the same guards into
CliproxyapiExecutor.transformRequest (Anthropic-shape branch), composing with
the existing bisected `mcp_*` reserved-namespace rewrite:
- sanitizeClaudeToolSchemas() on transformed.tools.
- cloakThirdPartyToolNames() with skip = mcp-reserved, so applyMcpToolNameRewrite
keeps authority over `mcp_*` (its bisected `Mcp_X` form) and the two reverse
maps stay disjoint / single-hop. Both merge into the non-enumerable
_toolNameMap the response stream already uses to restore the caller's names.
cloakThirdPartyToolNames is now non-mutating (clones changed entries) to respect
transformRequest's no-input-mutation contract, and takes an optional `skip`
predicate.
Verified end-to-end through the live CPA path: a real ~100-tool harness payload
that returned the "out of extra usage" placeholder now returns 200 with original
tool names restored on the response stream; `mcp_*` tools and genuine PascalCase
Claude Code tools are unaffected.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Follow-up commit on PR #2943 review:
- Preserve boolean schemas in `sanitizeClaudeToolSchemas` (Gemini Code Assist,
high severity). `additionalProperties: false` is the canonical JSON Schema
lock-down for object tools; the previous coercion silently turned it into the
permissive `{}`, which would invite models to hallucinate extra arguments
during tool calling. Same rule now applies to per-property boolean schemas
under `properties`. Placeholder strings still get the permissive `{}` slot —
booleans get preserved verbatim.
- Defensive null guards in `cloakThirdPartyToolNames` for `tools[]` and
`messages[]` entries that might be `null`/`undefined`. Prevents a runtime
`TypeError` if a malformed payload reaches the cloak.
- Document `CLAUDE_DISABLE_TOOL_NAME_CLOAK` in `.env.example` and
`docs/reference/ENVIRONMENT.md` (env/docs contract was failing in CI).
- Regression tests covering all of the above (5 boolean preservation cases,
2 null-tolerance cases).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Extract 3 high-value CPU/RAM optimizations from perf branch:
1. estimateSizeFast() — fast object-tree size estimator replacing
JSON.stringify().length in isSmallEnoughForSemanticCache(). Walks
object tree with a stack, zero string allocation, early exit at 256KB.
2. Consolidate settings reads — move getCachedSettings() to a single
early read in handleChatCore(), eliminating a redundant second read
200 lines later. Also removes the isDetailedLoggingEnabled() wrapper
call (reads settings internally) in favor of direct field check.
3. Registry Proxy→direct export — convert 8 registries from lazy
Proxy+getOrCreate pattern to simple exported const objects. Eliminates
Proxy trap overhead on every provider property access during routing.
Affected: audio, embedding, image, moderation, music, rerank, search,
video registries (-451 lines of Proxy boilerplate).
These changes are independent of the CPU leak fix (limiter eviction)
and complement it by reducing per-request CPU overhead.
Native Claude OAuth (claude->claude passthrough) forwards client tool
definitions verbatim. Anthropic's first-party Messages API then rejects:
- invalid tool input_schemas (deep-truncation placeholders such as
`enum: "[MaxDepth]"`, or index-keyed objects where arrays are required), and
- tool names it fingerprints as a third-party agent harness (specific
blacklisted names like `mixture_of_agents`, or a large enough set of
recognizable snake_case agent tool names),
both surfaced as a misleading `400 You're out of extra usage` placeholder
(the SSE stream is refused — not a real billing event). The same request
succeeds on translator-backed providers (OpenAI/Codex), which already sanitize
and re-shape tool payloads — so the gap is specific to the native passthrough.
Adds the missing guards on the native Claude OAuth path (executors/base.ts):
- sanitizeClaudeToolSchemas(): coerce/drop invalid draft-2020-12 constructs
(non-array enum/required/anyOf/..., placeholder schema slots -> {}).
- cloakThirdPartyToolNames(): deterministically alias non-Claude-Code tool
names (Claude Code canonical mapping where one exists, else PascalCase),
tracked in the existing per-request _toolNameMap so remapToolNamesInResponse
restores the caller's original names. Opt out via
CLAUDE_DISABLE_TOOL_NAME_CLOAK=true.
Genuine Claude Code tool names (PascalCase) and already-valid schemas are
left untouched, so existing first-party traffic is unaffected.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause: Bottleneck rate limiter instances in rateLimitManager accumulate
without cleanup. Each instance runs an internal heartbeat setInterval every
250ms. Under heavy load with many provider:connection:model combinations,
hundreds of limiters accumulate causing CPU to grow ~0.1%/min until server
collapse (~2% after 5 minutes of intensive use).
Changes:
- rateLimitManager: Add idle limiter eviction in watchdogTick() using the
previously defined but unused INACTIVE_LIMITER_MS threshold. Populate
limiterLastUsed on every getLimiter() call. Clean up all 3 Maps
(limiters, lastDispatchAt, limiterLastUsed) consistently.
- combo.ts: Add size-based FIFO eviction to rrCounters, resetAwareConnectionCache,
and resetAwareQuotaCache Maps. Convert per-target log.info calls in combo
execution loops to log.debug?. to reduce serialization overhead.
- chatCore.ts: Fix double-serialization in estimateTokens(JSON.stringify(x))
calls (estimateTokens already handles objects). Make trace() conditional
on OMNIRROUTE_TRACE/DEBUG env vars. Make per-request usage logging conditional.
- apiKeyRotator.ts: Add eviction guards to _keyHealth and _connectionExtraKeys
Maps (MAX 500 entries each). Ensure removeConnectionIndex cleans all 3 Maps.
- codexQuotaFetcher.ts: Add eviction guard to connectionRegistry and quotaCache
Maps (MAX 200 entries each).
Remove the AuditLogTab from the dashboard logs page now that audit logs
live under the dedicated /dashboard/audit route. Update integration wiring
expectations and add metadata frontmatter to studio framework docs.
Keep existing object-argument cleanup behavior, but avoid parsing and stripping arbitrary JSON-string arguments for unrelated tools where empty strings or arrays may be valid payloads. Add regression coverage for non-Read and non-object Read arguments.
Only translate Claude Code web_search_YYYYMMDD server tools to native Responses web_search when the final target is OpenAI Responses. Keep the Chat Completions target on function-tool shape and cover the full translateRequest path.
Translate Claude Code web_search_YYYYMMDD server tools to the native OpenAI Responses web_search tool and preserve filters/location. Convert forced Claude tool_choice for web_search to the native Responses tool choice while leaving ordinary custom functions unchanged.
Closes#2936
Buffer Claude Code Read tool calls through the existing shim layer so empty pages placeholders are removed before streaming input_json_delta to the client. Also clean JSON-string Responses tool arguments, not only object arguments.
Closes#2935
Addresses #2889
Plan 14 (#2839) test listed only cli-code/cli-agents/acp-agents/cloud-agents; #2858 added agent-bridge/traffic-inspector to TOOLS_GROUP. Align test to real code (both feature sets).
