Address bot review feedback: use .finally() instead of .then()/.catch()
so limiters.delete() runs regardless of whether stop() succeeds or
throws (e.g. already stopped by concurrent 429).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Multiple concurrent requests can receive 429 simultaneously, causing
stop() to be called on an already-stopped limiter. Add .catch() to
prevent unhandled rejection.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When a provider returns 429 (rate limit exceeded), the rate limit manager
was setting reservoir=0 and waiting for reservoirRefreshInterval before
releasing queued requests. For providers with long rate limit windows
(e.g. Codex with hours-long resets), this caused all queued requests to
hang indefinitely — they never timed out or returned an error.
This prevented upstream callers (e.g. LiteLLM) from triggering fallback
to alternative providers, effectively making the entire model unavailable
until the rate limit window expired.
Fix: on 429, call limiter.stop({ dropWaitingJobs: true }) to immediately
fail all queued requests, then delete the limiter from the Map so
getLimiter() creates a fresh instance for subsequent requests.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add preserveDeveloperRole option and model compat override
- Normalize developer→system in roleNormalizer when not preserving
- Translator runs normalizeRoles for Responses API with option
- UI: ModelCompatPopover with do not preserve developer toggle
- Add ZWS_README_V2 documenting cause and fix
Made-with: Cursor
The Bailian Coding Plan provider page may render a dialog on load
that blocks pointer events on the Add API Key button. Add pre-dialog
dismissal (Escape key) before attempting to click.
Also triages #485 (Claude Code tool calls — needs-info).
- #462: Mark gemini-cli provider as deprecated in providers.ts
Add deprecated, deprecationReason, hasFree, freeNote, authHint, apiHint
to Zod provider schema
- #471: Add VM_DEPLOYMENT_GUIDE.md to DOC_SOURCE_FILES in generate-multilang.mjs
Delete 29 stale PT-language copies and regenerate from EN source
for all 30 locales (29 auto-translated + 1 Czech from PR #482)
formatSSE() in streamHelpers.ts explicitly returned 'data: null' for
null/undefined data. This violates SSE protocol and causes
AI_TypeValidationError in strict clients (Zod-based AI SDKs).
Now returns empty string, silently skipping null chunks.
- New API: /api/logs/export?hours=24&type=call-logs
- UI: Export button with dropdown on /dashboard/logs page
- Supports export of request-logs, proxy-logs, and call-logs
- Downloads as JSON file with Content-Disposition header
Previously resolveModelAlias() output was used only for getModelTargetFormat()
but the original model was sent in translatedBody.model and to the executor.
Now effectiveModel is propagated to all downstream operations.
saveCallLog only read prompt_tokens/completion_tokens (OpenAI format).
When sourceFormat=claude, the openai-to-claude translator writes
input_tokens/output_tokens instead, causing all cross-format requests
(Codex-via-Claude, Kiro-via-Claude, etc.) to show 0|0 tokens in
call_logs.
Also includes cache_read and cache_creation tokens in tokens_in total
so heavily-cached requests don't show misleadingly low input counts.
Changes:
- Read prompt_tokens || input_tokens (supports both formats)
- Read completion_tokens || output_tokens (supports both formats)
- Sum cache_read_input_tokens + cache_creation_input_tokens into total
logUsage stored only non-cached input tokens in usage_history.tokens_input.
For heavily-cached Claude requests (common with Claude Code), this shows
near-zero input when the real total is 150K+, causing the analytics
dashboard to severely underreport input token usage.
Now sums: input = prompt_tokens + cache_read + cache_creation
chatCore.ts injects translatedBody.model for all providers after
translation. Kiro API (AWS CodeWhisperer) has strict schema validation
and rejects unknown top-level fields — only conversationState, profileArn,
and inferenceConfig are valid. This causes 100% of Kiro requests to fail
with "Improperly formed request".
Strip the injected model field in KiroExecutor.transformRequest().
* feat: add api-key Kimi Coding provider support
* fix(kimi-coding): honor apikey auth header in executor
Ensure DefaultExecutor sends x-api-key for kimi-coding-apikey at runtime
and deduplicate shared kimi coding config blocks in registry and models
config to reduce drift between oauth and apikey variants.
---------
Co-authored-by: OmniRoute Agent <agent@omniroute.local>
- fix(budget): BudgetTab sent integer percentage (80) but schema validated
fraction (0-1). Now divides by 100 on POST and multiplies by 100 on GET (#451)
- fix(combos): expose Agent Features UI in combo create/edit modal — fields for
system_message override, tool_filter_regex, and context_cache_protection were
implemented server-side (#399/#401) but missing from the dashboard UI (#454)
- fix(combos): strip <omniModel> tags from messages before forwarding to provider.
The internal cache-pinning tag was being sent to the provider, causing cache
misses as providers treated each tagged request as a new session (#454)
- fix(docker): copy pino-abstract-transport + pino-pretty in standalone (#449)
- fix(responses): remove initTranslators() from /v1/responses route (#450)
- chore(deps): commit package-lock.json with each version bump
- fix(docker): copy pino-abstract-transport and pino-pretty explicitly in
runner-base stage — Next.js standalone trace omits them, causing
'Cannot find module pino-abstract-transport' crash on startup (#449)
- fix(responses): remove initTranslators() call from /v1/responses route —
bootstrapping translator registry from a Next.js Route Handler worker
caused 'the worker has exited' uncaughtException on Codex CLI requests.
Translators are already bootstrapped server-side via open-sse (#450)
- chore: include package-lock.json in commit (was being left behind on
version bumps, causing npm ci to install inconsistent deps in Docker)
