Community · source page ↗ · last checked —

Change history

• Signal Jul 5, 2026, 12:29 PM

  roborev: free-tier gemini backup scoped to PUBLIC repos only (fail-safe opt-in)

  JohnGavin/llm — ## Context Calibrated privacy position (session discussion): free-tier gemini (aistudio) may train on submitted prompts. roborev submits code diffs. For PUBLIC repos that's moot (already world-readable); for PRIVATE repos it would leak proprietary strategy code (finance/betting/healthcare) + risk secrets in diffs. So gemini backup is acceptable on public repos, forbidden on private. Repo visibility (audited 2026-07-04) PUBLIC: llm, llmtelemetry, historical PRIVATE: BigBetData, od

• Signal Jul 4, 2026, 11:28 PM

  Hardcoded model ID 'claude-sonnet-4-20250514' will break when deprecated

  morrisstephon51/Enrollment_Funnel_Agent — ## Bug src/lib/claude-client.ts line 68 hardcodes a dated model ID: \\\ts model: 'claude-sonnet-4-20250514', \\\ This ID format (-YYYYMMDD) is a snapshot alias that Anthropic will retire. The current stable ID is claude-sonnet-4-6. When the snapshot is retired the API will return a 404/error on every report run with no obvious error message. Fix Update to the current model ID: claude-sonnet-4-6 Pull the model name from env with a sensible default so it

• Signal Jul 4, 2026, 3:54 PM

  grace_daily_limit_reached triggers inconsistently for claude --print, independent of account usage % shown in UI

  anthropics/claude-code — I'm hitting error_type: grace_daily_limit_reached when invoking claude --print non-interactively (spawned as a child process from a Node.js server, with ANTHROPIC_API_KEY deliberately deleted from the child env to force the OAuth/subscription session, per the documented pattern for using Claude Code as a local automation backend). Error: {"data":{"modal_data":{"elapsed_days":1},"type":"unlock_full_access_notice"},"error_type":"grace_daily_limit_reached"} What doesn't add

• Signal Jul 4, 2026, 2:12 PM

  Proposal: distinguish a Claude usage-limit/overload failure from a generic internal error (and tell an admin)

  swampratnz/community-agent — ## Problem Who it helps: every member whose turn fails because the bot's Claude subscription hit its usage/rate limit — plus the admins who currently have no signal that it's happening. Theme: Reliability & ops (graceful degradation — the same principle already applied to two sibling cases in router.ts/core.ts, just not this one). Evidence, read directly from the code: src/agent/core.ts's execTurn catches *any* query() exception and returns the exact same text for al

• Signal Jul 4, 2026, 1:37 AM

  fix(llm): deprecated model IDs pass validation silently, then 404 mid-generation — data to warn already exists

  gfargo/coco — From the AI-core audit. Severity: HIGH — the fix data already exists and is unused. Defect validation.ts:78-105 + modelValidity.ts:22-84 — validateModel only rejects a CROSS-provider mismatch (detectProviderMismatch, exact membership in a DIFFERENT provider's list). It never consults DEPRECATED_MODELS. getDeprecatedReplacement/DEPRECATED_MODELS are referenced ONLY by coco doctor (commands/doctor/checks.ts:201), never in getLlm/validateModel. A config carried over with model: 'gpt-4

• Signal Jul 3, 2026, 12:31 PM

  Fleet auto-fail-over on Claude/Fable usage-limit: built-in detection for real Claude signatures + build fable→opus→codex→glm→kimi chain

  BeFeast/maestro — ## Problem When the default claude backend hits an account-level usage limit, the fleet does not fail over to another model — workers respawn on the same exhausted backend, burning each issue's retry budget until it is marked retry_exhausted/blocked. On 2026-07-03 claude-fable-5 was account-exhausted and ok-player/ok-folio issues (#151, #152, folio-118..127) looped and got wrongly blocked. Two gaps (the ordered fallback_backends walk + respawn machinery are correct — the failur

• Signal Jul 3, 2026, 4:08 AM

  please give opsion to disable DISABLE CONTEXT CACHING, we can't use free tier gemini

  agent0ai/agent-zero — _when i used gemini free tier API i got this, but eferything is oke, when i used Paid Tier Gemini Please give option to disable context caching in future update_ ================================= litellm.exceptions.RateLimitError: litellm.RateLimitError: litellm.RateLimitError: GeminiException - { Traceback (most recent call last): Traceback (most recent call last): File "/opt/venv-a0/lib/python3.12/site-packages/litellm/llms/vertex_ai/context_caching/vertex_ai_context_

• Signal Jul 3, 2026, 3:06 AM

  Gemini free-tier OAuth deprecated upstream by Google (IneligibleTierError); steer to API key / Vertex

  rappdw/sandy — ## What Google has retired the free-tier OAuth login for gemini-cli ("Gemini Code Assist for individuals"), redirecting users to the Antigravity product suite. A sandy session with SANDY_GEMINI_AUTH=oauth (or auto-probe falling to OAuth) now fails at Google's tier check: IneligibleTierError: This client is no longer supported for Gemini Code Assist for individuals. To continue using Gemini, please migrate to the Antigravity suite of products: https://antigravity.google ineligibleT

• Signal Jul 2, 2026, 4:13 PM

  Add adaptive polling and persisted rate-limit backoff for Claude usage

  tjones-gss/ai-usage-overlays — ## Summary Replace the fixed Claude polling cadence with adaptive polling and explicit 429 handling inspired by claude-meter. Why The overlay currently refreshes on a fixed timer and treats 429s as stale data. A smarter cadence would stay responsive during active/high usage while reducing unnecessary calls and avoiding repeated rate-limit hits. Suggested approach Poll faster only when values change or 5-hour utilization is high. Slow down when snapshots are unchang

• Signal Jul 2, 2026, 11:04 AM

  [provider_openai] Sends deprecated max_tokens — 400 on reasoning models when maxOutputTokens is set

  ananmouaz/flutter_ai — Severity: critical openai_provider.dart:86-87 sends 'max_tokens': options!.maxOutputTokens. Chat Completions deprecated max_tokens; reasoning models (o-series, gpt-5 family) hard-reject it with a 400. Since the provider explicitly supports reasoning_effort (:82-84), reasoning models are clearly in scope — so reasoningEffort + maxOutputTokens together always 400. Fix: send max_completion_tokens (accepted by all current models), or gate by model family. While in there, consi

• Signal Jul 2, 2026, 9:49 AM

  Show Claude Fable scoped quota from usage limits

  majiayu000/quotabar — ## Problem Claude Code's current OAuth usage response exposes Fable weekly usage inside limits[], not as a top-level seven_day_fable5 quota window. Observed shape: group: weekly kind: weekly_scoped scope.model.display_name: Fable percent: Fable usage percentage resets_at: weekly reset time QuotaBar only parsed top-level Fable aliases, so the app logged seven_day_fable5=None and did not render the Fable 5 (7-Day) card even when the API had Fable usage. Done When QuotaBar par

• Signal Jul 2, 2026, 2:53 AM

  cascade: Gemini score rung is dead (free-tier limit 0 on gemini-2.5-pro) — retarget or demote

  Senkichi/job-cannon — ## Symptom (observed live 2026-07-01 during the model sweep) When an Ollama score call fails, the cascade falls to Gemini, which now returns 429 RESOURCE_EXHAUSTED with limit: 0 on every free-tier quota metric for gemini-2.5-pro — Google has zeroed free-tier quota for the pro model. The rung is dead weight: each fall-through burns a 15s retry + a second 429 (~30-60s of doomed latency) before reaching claude_code_cli, which then serves the call fine ($0). This is PREMORTEM d

• Signal Jul 1, 2026, 6:26 PM

  GPT-4.1 is being deprecated soon – Request to upgrade model version to GPT-5.4

  ruoccofabrizio/azure-open-ai-embeddings-qna — Hi Team, The application is currently configured to use GPT-4.1, which is scheduled for deprecation. To ensure continued functionality and avoid potential service disruptions, please review the repository and update the model configuration from GPT-4.1 to GPT-5.4. Repository: https://github.com/ruoccofabrizio/azure-open-ai-embeddings-qna Requested Actions: Identify all references to GPT-4.1 in the codebase and deployment configurations. Upgrade the m

• Signal Jul 1, 2026, 4:11 PM

  Claude code error that caused a hit limit and 211% of the usage credits

  anthropics/claude-code — Bug Description Claude code error that caused a hit limit and 211% of the usage credits Error kept showing each time claude did anything until finishing the 5h limit and also 211% of my usage credits!! API Error: an image in the conversation could not be processed and was removed. Double press esc to edit your message, or re-read the file if you still need it. The error messages tells that the image was removed, but apparently it didn't Environment Info Platform: darwin

• Signal Jul 1, 2026, 3:25 PM

  Research: proactively detect deprecated/retired models and surface results with a "deprecated" annotation instead of failing

  weval-org/app — ## Background In #24, a sandbox run failed because the hardcoded default model anthropic:claude-3-haiku-20240307 has been retired by Anthropic. The direct Anthropic API now returns: 404 Not Found - {"type":"error","error":{"type":"not_found_error","message":"model: claude-3-haiku-20240307"}} This caused generation to fail and the Macro Coverage Overview to show 0% (no gradable output). The immediate fix (#24 branch claude/github-issue-24-cnk7a6) just swaps the hardcoded default t

• Signal Jul 1, 2026, 3:30 AM

  feat: patrol GLM model updates — auto-detect when Coding Plan docs list new or deprecated models

  telleroutlook/claude-bot — ## Background GLM Coding Plan models change over time. Currently GLM-5.2, GLM-5-Turbo, and GLM-4.7 are available, with older models (GLM-5.1/GLM-5) auto-redirecting to GLM-5.2. When new models are released, the bot's model aliases need to be updated manually. Reference: https://docs.bigmodel.cn/cn/coding-plan/overview Current model tier mapping (as of 2026-07-01) | Tier | Current model | Role | |---|---|---| | hard | GLM-5.2 | Equiv. Claude Opus — complex implementatio

• Signal Jul 1, 2026, 1:26 AM

  immersive-translate/immersive-translate — ### 插件版本号 | Version 1.30.3 翻译引擎 | Translation Engine Claude Sonnet 5（内置 AI 大模型服务） 平台 | Platform macOS 浏览器 | Browsers Chrome 插件类型 | Extension Type 浏览器插件 | Browser Extension 请描述这个Bug | Describe the bug 在插件的 AI 大模型服务列表中，点击「Claude Sonnet 5」卡片的"点此测试服务"，测试失败，报错： 服务返回错误， 请求参数错误，请检查相应配置。 400: temperature is deprecated for this model. 查阅 Anthropic 官方文档确认：从 Claude Opus 4.7 / 4.8、Claude Sonnet 5 起，temperature、top_p、top_k 采样参数已被移除/受限，非默认值会直接返回 400（Fable 5 / Opus 4.7

• Signal Jun 30, 2026, 4:44 PM

  Deprecated: no longer needed on Opus 4.8+ (silent-stop fixed model-side)

  evnchn-agentic/claude-code-stop-gate — *Posted by Claude Code on evnchn's behalf.* TL;DR: Deprecated — this hook is no longer needed on Claude Opus 4.8+. The silent-stop failure this gate guarded against — the model ending a turn text-only when the task still needed a tool call — was fixed model-side in Opus 4.8, not in the Claude Code harness. The guard is now redundant on 4.8+. Keep it only if you run Opus 4.7 or older, where the failure mode returns and you'd be unguarded. Archived as a depre

• Signal Jun 30, 2026, 12:39 PM

  Replace deprecated Groq model (llama-3.3-70b-versatile)

  codedaily04/commit-craft — ## Description The project currently uses the llama-3.3-70b-versatile Groq model, which has been deprecated by Groq. According to Groq's announcement, the model will be decommissioned on August 16, 2026. After that date, API requests using this model will no longer be served. Suggested Fix Replace llama-3.3-70b-versatile with one of Groq's recommended replacement models, such as: gpt-oss-120b qwen3.6-27b It would also be beneficial to make the model name configurable t

• Signal Jun 29, 2026, 6:30 PM

  Bug: Metabase model created as standard question due to deprecated dataset flag

  vojaka/scheduling-agent-backend — ## Bug Description The "Cleaned Shifts Dataset" was created as a standard Question instead of a Model because the dataset boolean flag has been deprecated in Metabase v0.49+ in favor of the type: "model" field. Fix Applied Sent a PUT /api/card/46 request to set type: "model" directly. Metabase has successfully updated the card type to model. Status Closed.

• Signal Jun 29, 2026, 6:30 PM

  Bug: Metabase model created as standard question due to deprecated dataset flag

  vojaka/comforthub-backoffice-backend — ## Bug Description The "Cleaned Shifts Dataset" was created as a standard Question instead of a Model because the dataset boolean flag has been deprecated in Metabase v0.49+ in favor of the type: "model" field. Fix Applied Sent a PUT /api/card/46 request to set type: "model" directly. Metabase has successfully updated the card type to model. Status Closed.

• Signal Jun 28, 2026, 11:50 PM

  feat(ai-dictionary): add Gemini free-tier fallback model chain

  u88803494/my-website — ## Background AI Dictionary currently uses gemini-2.5-flash-lite as a single model. Recent Free Tier usage can hit quota / rate limit / high-demand errors quickly, which reduces reliability. Goal: keep the solution free-tier friendly while adding a Gemini model fallback chain so the API can try another free Gemini model when the primary model fails. Proposed approach Primary: gemini-3.1-flash-lite Fallback 1: gemini-2.5-flash-lite Fallback 2: gemini-2.5-flash Avoid using l

• Signal Jun 28, 2026, 12:11 PM

  Pricing-page change detected: https://openai.com/api/pricing/

  modelmeters/modelmeter — Source URL: https://openai.com/api/pricing/ Reason: HTTP 403 fetching https://openai.com/api/pricing/ Affected entries: openai/gpt-5-4 openai/gpt-5-4-mini openai/gpt-5-5 Please open the URL, verify current pricing, update pricing/current.json if needed, and close this issue.

• Signal Jun 28, 2026, 9:06 AM

  Tối Ưu Hóa Định Tuyến Hybrid & Tự Kiểm Định Model (Gemini Free-Tier Optimization)

  hdangprod/prj226 — # Tối Ưu Hóa Định Tuyến Hybrid & Tự Kiểm Định Model (Gemini Free-Tier Optimization) 🚨 Bối cảnh & Vấn đề (The Context & "Why now?") Hệ thống hiện tại (V2.0) đang phụ thuộc hoàn toàn vào Gemini 3.5 Flash (PRO) cho các tác vụ quan trọng như lập kế hoạch tuần (/weekly_planning) và báo cáo tuần (/weekly_report). Vấn đề: Giới hạn API của Free-tier cho model PRO cực kỳ khắt khe: 20 Requests Per Day (RPD). Việc sử dụng PRO cho mọi luồng lập lịch sẽ dẫn đến tình trạng cạn kiệt tài ngu

• Signal Jun 28, 2026, 6:09 AM

  [backend] [HIGH] codex harness ships unusable model ids + deprecated --full-auto

  fxmartin/claude-code-config — ## Description The codex harness shipped in Epic-20 Feature 20.7 (Stories 20.7-003/004) is not runnable out of the box: Fictional model ids. controller/config/harnesses.yaml hardcodes a per-stage models: map of gpt-5.4-codex / gpt-5.4-codex-mini / gpt-5.4-codex-high. These are rejected on a ChatGPT-account Codex: ERROR: {"type":"error","status":400,"error":{"type":"invalid_request_error", "message":"The 'gpt-5.4-codex' model is not supported when using Codex with a

• Signal Jun 28, 2026, 2:37 AM

  Together.ai models to be deprecated July 10

  earendil-works/pi — ### What do you want to change? Together.ai is deprecating two models on July 10, 2026: zai-org/GLM-5.1 Qwen/Qwen3-235B-A22B-Instruct-2507-tput The recommended model alternatives by notice emails: zai-org/GLM-5.2 (not supported) MiniMaxAI/MiniMax-M3 (supported) Why? Notices via email: Important Notice_ Deprecation of zai-org_GLM-5.1.eml Important Notice_ Deprecation of Qwen_Qwen3-235B-A22B-Instruct-2507-tput.eml Other references: zai-org/GLM-5.1 will have no upgrade path by J

• Signal Jun 27, 2026, 8:36 PM

  Pricing-page change detected: https://openai.com/business/pricing/#api

  modelmeters/modelmeter — Source URL: https://openai.com/business/pricing/#api Reason: HTTP 403 fetching https://openai.com/business/pricing/#api Affected entries: openai/gpt-5-4 openai/gpt-5-4-mini openai/gpt-5-5 Please open the URL, verify current pricing, update pricing/current.json if needed, and close this issue.

• Signal Jun 27, 2026, 7:05 PM

  Gemini free tier daily limit

  SanDiego2049/ingredient-iq — ## Description AI Studio dashboard confirms the actual current limit for this project is 20 requests/day (RPD 22/20 observed 27 June 2026). Google's own rate limits documentation no longer publishes fixed per-model numbers and explicitly states limits aren't guaranteed, account-specific figures must be checked via the AI Studio dashboard directly. Acceptance criteria [x] Development and testing workflow adjusted to treat live Gemini calls as a limited resource, favou

• Signal Jun 27, 2026, 1:51 PM

  [Phase 3] Model-tier routing (free Gemini → paid frontier)

  punyamsingh/ARGUS — Phase: 3 (margins & scale) Run most briefs on the free/cheap tier; reserve a paid frontier model for deep or high-value synthesis — via the provider abstraction, so it's config, not rewrites. Tasks [ ] Heuristic/classifier to route by brief complexity (evidence volume, ambiguity, deal value) [ ] Default: Gemini Flash (free); escalate to a stronger model (e.g. Gemini Pro, or Claude Opus once there's budget/credits) for deep/ambiguous briefs [ ] Per-route token + cost logging t

• Signal Jun 26, 2026, 9:53 PM

  Fireworks AI models are deprecated/unavailable

  decolua/9router — ## Problem The Fireworks AI provider (open-sse/providers/registry/fireworks.js) ships 3 LLM models that are no longer available on Fireworks: | Model | Status | |-------|--------| | accounts/fireworks/models/deepseek-v3p1 | Deprecated — superseded by deepseek-v4-pro/flash | | accounts/fireworks/models/llama-v3p3-70b-instruct | No longer served | | accounts/fireworks/models/qwen3-235b-a22b | No longer served | Users selecting these models get upstream errors. The provider also l

• Signal Jun 26, 2026, 11:17 AM

  [FEATURE] Claude Code mobile/web: show model, context (K), effort level & usage limit

  anthropics/claude-code — ### Preflight Checklist [x] I have searched existing requests and this feature hasn't been requested yet [x] This is a single feature request (not multiple features) Problem Statement On the mobile / web Claude Code surface there's no persistent indicator of session state. You can't tell at a glance which model is active, how full the context is, what effort level you're on, or how close you are to a plan/rate limit. This is especially acute when driving a remote session

• Signal Jun 26, 2026, 8:59 AM

  Release notes fail to generate because the AI model is hardcoded and deprecated

  awinogradov/code-assistants — ## Context The weekly release action invokes the Anthropic API to turn the changelog into human-readable release notes. The model id is hardcoded in releaseNotesPrompt.ts as claude-opus-4-20250514, which Anthropic has now deprecated (end-of-life June 15 2026). The latest release run logs The model 'claude-opus-4-20250514' is deprecated followed by 404 not_found_error and AI release notes not generated, using full changelog. Because the Generate Release Notes step is

• Signal Jun 26, 2026, 4:08 AM

  Wire Vertex AI so Gemini spends the $1,300 Google Cloud credits (AI Studio keys are free-tier only)

  alex-place/lantern-os — ## Problem The $1,300 Google Cloud credits (billing account on project My First Project / 843848914143) are NOT being spent by Gemini. Every AI Studio API key tested — both legacy AIza… and the new AQ. express keys — is metered on the free tier: quotaMetric: generativelanguage.googleapis.com/generate_content_free_tier_requests quotaId: GenerateRequestsPerDayPerProjectPerModel-FreeTier quotaValue: 20 (per day, per model) The AI Studio Generative Language API (generativelan

• Signal Jun 25, 2026, 2:23 PM

  Bedrock: Update default streaming chat model from deprecated anthropic.claude-v2

  quarkiverse/quarkus-langchain4j — ## Problem When using Bedrock streaming chat without explicitly configuring a model ID, the application fails with a 404 error because the default model anthropic.claude-v2 has reached end of life:

• Signal Jun 25, 2026, 1:24 PM

  roborev silently failing: gemini free-tier dead + health checks read wrong metric

  JohnGavin/llm — ## Summary roborev reported verdicts.failed: 0 / roborev:ok while 28/28 reviews on 2026-06-25 crashed. The "clean" report is a false negative — reviews crash *before producing a verdict*, so verdicts.failed stays 0. Surfaced when /bye reported "roborev is clean (0 failed)". ## Root cause (evidence) summary --json has two distinct sections: - overview: job outcomes → failed: 16 (and 28/28 today) - verdicts: only reviews that produced a pass/fail → failed: 0,

• Signal Jun 25, 2026, 3:15 AM

  [BUG] Claude Code-shared usage limit blocks macOS Desktop Chat while WebChat works

  anthropics/claude-code — ### Preflight Checklist - [x] I have searched existing issues and this hasn't been reported yet - [x] This is a single bug report (please file separate reports for different bugs) - [x] I am using the latest version of Claude Code / Claude desktop app available to me ### What's Wrong? Clarification on repo scope: the visible broken surface is the Claude macOS desktop app's norma

• Signal Jun 24, 2026, 11:53 PM

  feat(app,cli): surface Claude plan usage — 5h & weekly rate limits + reset time

  slopus/happy — ## Problem There's no way to see how much of the Claude subscription quota is left — the 5-hour rolling window, the weekly cap, or when either resets. On mobile this means working "blindfolded": you can't size the next task against remaining quota, which is what actually throttles work (well put in #64). #1126 asks for the same: plan session usage + reset time in the status bar. ## Status today: Happy doesn't surface it, but the SDK already provides it Happy captures Claude ra

• Signal Jun 24, 2026, 8:36 PM

  Expand the deprecated-model-identifier list for staleness detection

  ssamba1/skill-doctor — sdlib.STALE_MODEL_PATTERNS flags deprecated model IDs in skill bodies. Add more from Anthropic's deprecation list (e.g. claude-3-haiku-20240307, claude-3-sonnet-20240229, claude-3-opus-20240229, older Gemini/GPT IDs) and add a test case to tests/test_sdlib.py::test_stale_findings. Small, self-contained, no new deps.

• Signal Jun 24, 2026, 4:40 PM

  Twin: expand Gemini free-tier cascade with gemini-3-flash-preview + a Gemma rung

  Jin-HoMLee/jin-ho-lee-cv — ## Goal Before reaching for a cross-vendor fallback (#97), tap the remaining Gemini-side free-tier quota the cascade isn't using. Verified live against the API (the bucket is keyed by the *resolved* model, so aliases add nothing; only genuinely distinct model ids add headroom). ## What to add (verified live) - gemini-3-flash-preview — a distinct model id with its own free-tier daily bucket (not an alias of any current rung). Top-tier Flash quality; takes `thi

• Signal Jun 24, 2026, 4:32 PM

  AI requests fail with cryptic HTTP 404 when a project's model is deprecated on OpenRouter

  StoryThread-Dean/StorythreadStudio — ## Summary A project pinned to a model that OpenRouter later deprecates fails every AI request with a cryptic OpenRouter returned an unexpected error: HTTP 404., and the per-project model picker made it nearly impossible to self-diagnose or fix. Reported by a user whose project was pinned to x-ai/grok-3-mini (deprecated by xAI). The OpenRouter chat site worked fine for them, and switching the global model in Settings had no effect, so it looked like a

• Signal Jun 24, 2026, 12:20 AM

  [E22.X] validate model ids in workflows.yaml against the live /v1/models snapshot — accept newly-released models automatically, surface deprecated/typo'd values early, fail-open when unverifiable

  kuhlman-labs/fishhawk — ## Summary Model ids in.fishhawk/workflows.yaml (executor.model, reviewers.agents[].model) are unvalidated free strings today — workflow-v0.schema.json declares them {"type":"string"} with no enum/pattern. So a typo, a nonexistent id, or a deprecated/sunset model passes schema validation AND the (opt-in, fail-open) modelpolicy.go allow-list, then fails LATE at agent spawn with a raw vendor error — or silently records $0 cost when unpriced. Goal: newly-relea

• Signal Jun 23, 2026, 11:42 PM

  [BUG] 5 hour usage limit reached - but I have not used Claude in over a day

  anthropics/claude-code — ### Preflight Checklist - [x] I have searched existing issues and this hasn't been reported yet - [x] This is a single bug report (please file separate reports for different bugs) - [x] I am using the latest version of Claude Code ### What's Wrong? I opened claude, sent one message and get several Api errors and then it quickly says I have maxed out usage. I have the Max plan, i

• Signal Jun 23, 2026, 7:45 PM

  Model.get_or_none() and Model.delete() still use deprecated operator predicates (v0.12.0)

  syn54x/ferro-orm — ## Summary v0.12.0 deprecates operator-style where() (Model.field OP value) in favor of lambda predicates, but first-class Model helpers still construct operator predicates internally. Consumers who migrated call sites to lambdas still get DeprecationWarning from framework code they did not write. Affected APIs observed: Model.get_or_none(pk) — warns on fetch Model.delete() — builds self.class.where(getattr(self.class, pk_field) == pk_val) (same operator path) This blocks pyte

• Signal Jun 23, 2026, 7:38 PM

  [Bug] Claude Code usage limits metered separately per surface on same account

  anthropics/claude-code — Bug Description Claude Code's 5-hour usage limit appears to be metered separately per surface on the SAME account. Desktop app Claude Code (Cowork/VM, v2.1.181) reports "hit your 5-hour limit," while the terminal CLI (v2.1.177) keeps working normally at the same time. Both are signed into the same Team org (org ID 28fdbd1e-8bfc-4e67-847e-56c4d3c341a6, tucker@archiveresale.co, team_standard seat, prepaid, no API key in use). I confirmed the org UUID is identical in bo

• Signal Jun 23, 2026, 4:43 PM

  Twin: Worker hangs (~44s) instead of failing fast when the Gemini free-tier cascade is fully exhausted

  Jin-HoMLee/jin-ho-lee-cv — ## Problem When the whole Gemini free-tier model cascade hits its daily request caps (429 across gemini-3.5-flash → gemini-2.5-flash → gemini-2.5-flash-lite), the Worker does not degrade gracefully — it hangs until the Cloudflare Workers runtime kills the request as "code had hung and would never generate a response." Observed locally (just worker-dev) while testing the twin:

• Signal Jun 23, 2026, 2:26 PM

  Migrate Gemini model lineup from deprecated 2.5 family to Gemini 3

  parleq/parleq-speech — ## Context Google has deprecated the entire Gemini 2.5 family, which is Parleq's current Gemini/Vertex lineup: - gemini-2.5-flash-lite — deprecated, shutdown Oct 16, 2026; Google's named replacement is gemini-3.1-flash-lite. - gemini-2.5-flash and gemini-2.5-pro — same 2.5 generation, same retirement track (Gemini 2.5 retiring on Vertex before Oct 2026). - The 2.5 Flash-Lite preview variants are already retired (Mar/Jul 2026). Refs: [Gemini API depr

• Signal Jun 23, 2026, 2:06 AM

  P1: real LLM-API pricing registry for the Fleet spend panel (OpenAI/Gemini/local)

  kevinthelago/base-studio-code — Part of #1078 (P1 — provider+model pricing registry). ## Problem tokens.rs model_pricing only knew the Claude families — every non-Claude model fell back to Sonnet pricing, so the Fleet spend panel reported wrong $ for OpenAI/Gemini/local sessions. The Opus entry was also stale ($15/$75; Opus 4.8 is $5/$25). ## Fix Rewrote the price table to span all four providers (list prices as of 2026-06), keyed by lowercased-model substring, most-specific family first

• Signal Jun 22, 2026, 5:13 PM

  Resolve Gemini API free-tier quota (limit: 0) blocking the Vision AI PoC

  CSE-Sudo-26/sudo-capstone-project — ## Background The Gemini Vision diet-analysis PoC (backend/services/gemini_service.py) cannot make live API calls. After hardening the script (commit 90efad8 — migrated to google-genai, GA model gemini-2.0-flash,.env loading with override=True, surfaced full errors), live calls fail with:

• Signal Jun 22, 2026, 4:12 PM

  Gemini free-tier 429s immediately under load

  Troppy2/Nosey — Severity: Medium ### Summary Gemini returns 429 across all three _with_retry attempts (1s, 2s backoff) before falling through. Combined with Groq 403 (#3), the auto chain effectively depends entirely on Claude. ### Evidence (prod logs 2026-06-22)

• Signal Jun 20, 2026, 7:56 PM

  Task marked as Done (success) despite Claude API usage limit error

  jonwiggins/optio — A Claude Code review task hit the API usage limit (Usage credits required for 1M context error from Claude API), but the task was still marked as Done (green/success) in the UI instead of showing the error. The task should have been marked as failed or needs_attention with the error message visible, but it appeared as successfully completed despite the API limit error. ## Steps to Reproduce 1. Run a Claude Code review task 2. Let it hit the Claude API usage credits l

• Signal Jun 19, 2026, 5:40 PM

  [DOCS] Model configuration docs omit deprecated/auto-updated model warning behavior for -p and subagent model frontmatter

  anthropics/claude-code — ### Documentation Type Incorrect/outdated documentation ### Documentation Location https://code.claude.com/docs/en/model-config ### Section/Topic Model-selection warning behavior when a requested model is deprecated or automatically updated to a newer model, including subagent model frontmatter and print-mode (-p) output handling ### Current Documentation The model configuration page currently says: > Switching to a blocked model with /model is rejected with an err

• Signal Jun 19, 2026, 9:31 AM

  [Anthropic] Native structured-output model uses the deprecated structure

  zendev-sh/goai — Provider: anthropic Go version: go version go1.26.3 darwin/arm64 GoAI version: v0.8.4 (also reproduces on main HEAD) Description Native structured-output mode (structuredOutputMode: "outputFormat", or "auto" with a supported model) sends the JSON schema as the top-level output_format request field. The Anthropic Messages API has deprecated that field in favor of output_config.format, so every native-output-format request now fails with a 400. This

• Signal Jun 18, 2026, 9:19 PM

  Twin availability: gemini-3.5-flash free tier is only 20 req/day → cascade to lower models

  Jin-HoMLee/jin-ho-lee-cv — ## Problem The digital-twin chat Worker runs a single Gemini model, \gemini-3.5-flash\. Its free-tier daily cap is only 20 requests/day (confirmed against the live API — \quotaId: GenerateRequestsPerDayPerProjectPerModel-FreeTier, quotaValue: 20\). The daily digest cron (1 req) plus one curious visitor with a ~15-turn conversation exhausts the day's allowance. After that, every visitor gets a 502 → the widget's "twin's resting" fallback, until the quota resets

• Signal Jun 18, 2026, 6:24 PM

  Perplexity API key validation uses deprecated /models endpoint

  decolua/9router — ## Summary In v0.5.4, valid Perplexity API keys can be marked as invalid because provider validation and model listing still call:

• Signal Jun 18, 2026, 10:40 AM

  claude-opus-4-8: "temperature is deprecated for this model."

  The-PR-Agent/pr-agent — > Same issue for claude-opus-4-8 (and claude-fable-5) _Originally posted by @check24-axel-nitzsche in #2400 _

• Signal Jun 16, 2026, 8:14 AM

  [BUG] Claude Max usage limit continues increasing after clean Claude Code test and local disconnect

  shreyashjagtap157/claude-code — ### Preflight Checklist - [x] I have searched existing issues and this hasn't been reported yet - [x] This is a single bug report (please file separate reports for different bugs) - [x] I am using the latest version of Claude Code ### What's Wrong? ### Summary I observed abnormal Claude Max usage-limit consumption after a single controlled Claude Code request. Usage conti

• Signal Jun 16, 2026, 5:53 AM

  You've hit your monthly spend limit ┬À raise it at claude.ai/settings/usage

  shreyashjagtap157/claude-code — Bug Description Claude stormed out 101 agents at one time and burned all of my session usage and a good portion of my weekly budget. Environment Info - Platform: darwin - Terminal: ghostty - Version: 2.1.177 - Feedback ID: 4bd87f91-04bb-4afd-be79-c015dfb849e6 Errors

• Signal Jun 16, 2026, 5:00 AM

  Usage limit discrepancy: Claude Code blocked at "100% used" while claude.ai Settings shows "44% used" on same Pro account

  shreyashjagtap157/claude-code — ### Preflight Checklist - [x] I have searched existing issues and this hasn't been reported yet - [x] This is a single bug report (please file separate reports for different bugs) - [x] I am using the latest version of Claude Code ### What's Wrong? Bug Report Summary ÔÇö Claude Code / claude.ai Usage Discrepancy Title: Usage limit discrepancy: Claude Code blocked at "100%

• Signal Jun 16, 2026, 4:51 AM

  [BUG] Claude Max usage limit continues increasing after clean Claude Code test and local disconnect

  shreyashjagtap157/claude-code — ### Preflight Checklist - [x] I have searched existing issues and this hasn't been reported yet - [x] This is a single bug report (please file separate reports for different bugs) - [x] I am using the latest version of Claude Code ### What's Wrong? ### Summary I observed abnormal Claude Max usage-limit consumption after a single controlled Claude Code request. Usage conti

• Signal Jun 15, 2026, 2:25 PM

  You've hit your monthly spend limit ┬À raise it at claude.ai/settings/usage

  shreyashjagtap157/claude-code — Bug Description Claude stormed out 101 agents at one time and burned all of my session usage and a good portion of my weekly budget. Environment Info - Platform: darwin - Terminal: ghostty - Version: 2.1.177 - Feedback ID: 4bd87f91-04bb-4afd-be79-c015dfb849e6 Errors