Community · source page ↗ · last checked —

Change history

• Signal Oct 23, 2026, 8:38 AM

  Provider usage: show and spend Codex and Claude usage-limit resets

  get-bb/bb — ### The workflow Codex (ChatGPT) and Claude subscriptions now come with usage-limit resets, even on a single account. When I hit a limit I want to see that I have a reset and spend it from bb's Provider usage card, the same place that already shows the limit I just hit. Today I have to leave bb and redeem it in the provider's own client (Codex CLI /status → Redeem reset, Claude Code /limit-reset). bb never shows that a reset exists, so they can expire unused. What happens today Provi

• Signal Oct 23, 2026, 8:17 AM

  Subscription slots blocked for a usage limit stay blocked after the limit resets (OpenAI and Claude)

  code-yeongyu/senpi — ## Summary A beta user (desktop 1022.1) with several saved subscription accounts, on OpenAI (login-3..login-6, all "Connected") and on Claude, reports that once an account hits its usage limit it's marked blocked and never comes back: "when usage fills up it force-disconnects, and after that it never reconnects." The same happens across providers, so this looks like the shared credential-pool/slot logic rather than one provider's adapter. Related (prior art) #2081 (closed):

• Signal Oct 23, 2026, 4:46 AM

  [AI] Gemini free tier chỉ 20 request/ngày, cần bật billing hoặc đổi key cho prod

  quocta-uranus/FlexTrip — Hiện tượng: AI (gợi ý mô tả, tranh chấp, giá) trả 500/503/504 trên prod dù local chạy được. Nguyên nhân đã xác định: Free tier chỉ ~20 request/ngày cho mỗi model; gemini-3.6-flash đã 21/20, gemini-3.5-flash trả 429. gemini-3.8-flash mất ~70s mỗi request nên gây 504 qua proxy. Đã chuyển sang gemini-3.5-flash-lite (~1.4s) ở #72/#73. Còn lại: [ ] Bật billing cho project chứa key prod (AI Studio → Set up billing) hoặc đổi key [ ] Xác nhận key trên VPS thuộc đúng project [ ]

• Signal Oct 23, 2026, 2:15 AM

  backend deprecated models check - 2026-10-23

  HeyPuter/puter — Automated read-only check of the hardcoded backend model lists against https://ai-models.puter.work/. A model is listed only if none of its names (id + aliases) appear in the upstream ids, names or aliases. chat/mistral (3) mistral-large-4 mistral-large-2512 zai-glm-5-3 Notes: upstream mistral has no "large" model at all. mistral-large-4 is a recent public-preview entry (release_date 2026-10-06) and may just not be in the upstream feed yet, so confirm before removing. zai-glm-5-

• Signal Oct 23, 2026, 12:29 AM

  [Bug] AI Coach (Gemini) returns 404: models/gemini-2.5-pro is deprecated for new users

  DuarteSantos8/openGym — ### What did you do? Generated a new API Key in Google AI Studio. Configured the AI Coach in openGym using the Google Gemini provider and pasted the API Key. Opted in and triggered an AI Coach action (e.g., workout review / coach response). What did you expect to happen? The AI Coach should successfully connect to the Gemini API and return the workout review/response. Additionally, there should ideally be a field in the UI to specify or override the model name manually. W

• Signal Oct 22, 2026, 9:46 PM

  catalog: opencode_go ships model ids the endpoint rejects (deprecated -free ids, protocol-unsupported families)

  xilistudios/lele — ## Problem The bundled catalogs catalog/providers/opencode_go.json (29 ids) and opencode.json (40 ids) are mirrors of models.dev, and they are what the WebUI model picker and offline onboarding validation read. A live probe of POST https://opencode.ai/zen/go/v1/chat/completions on 2026-10-22 (all lele headers present, streaming and non-streaming) shows the catalog offers ids the gateway refuses: | id | response | | --- | --- | | minimax-m3.1-flash-free | 400 ModelDeprecated —

• Signal Oct 22, 2026, 7:02 PM

  Usage-reset offer disappeared before its stated expiry, after Claude said it would last through Oct 22; weekly limit then froze work mid-task

  anthropics/claude-code — Summary. My usage page showed a reset offer labelled "expires on Oct 22". Late on Oct 21 (Toronto time), I asked a Claude Code side chat whether I'd still have it the next day. It said probably, and after midnight said it would most likely last until about 8 PM Toronto time on Oct 22. When I went to press the reset during Oct 22, it was gone. My weekly limit then froze a multi-session build mid-task, two days before a release deadline. The weekly limit resets Oct 25, 4 P

• Signal Oct 22, 2026, 5:15 PM

  [BUG] claude -p exits 1 with empty stderr when the usage limit hits

  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? claude -p exits 1 with empty stderr when the usage limit is reached; the reason is only in stdout JSON Using claude -p --output-format json as a judge in a prompt benchmark (28 scripts, a few calls each). When the login's usage limit hi

• Signal Oct 22, 2026, 5:10 PM

  stablyai/orca — ### Operating system Other Orca version v1.4.232 Details Short summary When using two separate Claude accounts across two independent projects in Orca, hitting the usage limit on the first account incorrectly triggers a usage limit error for the second account as well. However, the second account still functions normally when run independently in a terminal. What happened? I am currently using Orca version 1.4.232 with two separate Claude accounts configured across two distinct p

• Signal Oct 22, 2026, 3:24 PM

  Insight generation fails on free-tier 503s: move off gemini-3.1-flash-lite, add retry, raise timeout

  SatcherInstitute/health-equity-tracker — AI insights are live to the public on prod (the flag gate came out in #5065), but when a request actually reaches Gemini rather than a cache, it fails most of the time. Caching hides this almost completely: the only requests that go live are cold cache keys, which happen after a DAG rerun moves an insight's content-hash key. Evidence Cloud Logging, 14 days ending 2026-10-22, filtering on jsonPayload.insight.outcome: | Project | generated | error | |---|--

• Signal Oct 22, 2026, 3:00 PM

  Deprecated Claude model ID: claude-sonnet-4-20250514

  pid1/textmaxxing — main.py:76 hard-codes model="claude-sonnet-4-20250514" (Claude Sonnet 4). Anthropic lists it as deprecated (retirement date TBD). Suggested replacement: claude-sonnet-5-5 (current Sonnet, same price tier class, 1M context). Note the API differences on Sonnet 5.5: thinking: {type: "disabled"} returns 400, forced tool_choice any/tool returns 400, assistant prefill returns 400. This app does a plain messages call, so the switch should be small, but it changes reply behavior and n

• Signal Oct 22, 2026, 2:30 PM

  star: worker-probe prints ok for a worker stopped at Claude's usage limit

  jrsgacusan/juel-claude — ## What happened Three build workers stopped on Claude Code's "Usage limit reached · limit resets · Continuing automatically at " screen. For about 20 minutes worker-probe.sh --item --home --worktree --deadline 45 printed ok for each, so STAR saw healthy workers. It found the stop only by reading the worker terminal's last lines with orca terminal read --terminal. Expected The probe reports a usage-limit stop as its own line, for example stalled: usage-limit until. STAR

• Signal Oct 22, 2026, 3:50 AM

  Update ladder and hard-tier pins to remove deprecated models

  quadseven/loom — Difficulty: medium Why The current ladder, hard-tier pins, and equivalents still reference deprecated models gpt-6-sol and gemini-4-argon, even though the runner no longer uses them according to the 2026-10-21 operator rule. What Update the LADDER, HARD_MODELS, and EQUIVALENTS tables to remove gpt-6-sol and gemini-4-argon references, ensuring the codex rung uses gpt-6-luna. Update documentation in docs/MUSE.md, docs/LANES.md, and docs/BRAIN.md to reflect the current ladder struc

• Signal Oct 22, 2026, 1:50 AM

  ai usage on the dashboard: claude code and opencode go limits

  FindMalek/keep-dashboard — i hit the 5-hour and weekly limits without seeing them coming what in FindMalek/dotfiles: a claude code status line script that shows the 5-hour and weekly usage in the terminal and posts the numbers to the keep dashboard over tailscale same for opencode go, once i know where it exposes usage (check its cli / api) keep dashboard: an "ai usage" tile per tool and per machine, with time until reset, and an ntfy alert at 80% done when the terminal shows usage without runni

• Signal Oct 21, 2026, 8:45 PM

  Resume Claude sessions stopped by a usage limit once the rotator has a login with capacity

  rysweet/amplihack-rs — ## Problem amplihack profile rotate moves the login with the most capacity into ~/.claude. Running Claude Code sessions pick up the new login on their next request, but a session whose turn already ended with "You've hit your … limit" sits idle and never retries. Newer Claude Code also shows a "What do you want to do? 1. Stop and wait for limit to reset" menu over it. On one host this left several tmux sessions stuck for hours after a login with capacity had been rotated i

• Signal Oct 21, 2026, 8:14 PM

  chore(ci): surface Claude Code Review usage-limit failures (reset time) on the PR instead of only in the log

  bizimind/loxel — ## Problem When the Claude Code Review job hits the reviewer's usage limit, it fails with only a red check on the PR. The reason and the reset time are buried in the step log ("api_error_status": 429, "api_error": "usage_limit_reached", "result": "You've hit your session limit · resets 12:30am (UTC)"). The author, or an agent babysitting the PR, has to download and grep the log to learn that a re-run is pointless until the reset, and the PR otherwise looks like it simply never g

• Signal Oct 21, 2026, 7:56 PM

  Claude "Monthly spend" limit row disappears after refreshing limits when a CLIProxy usage source is configured

  pingdotgg/t3code — ## What happened On the Usage page (Limits view, All environments), the Claude card showed a "Monthly spend" row with a bar for each Claude account (99% and 74% left). After I clicked the refresh button, the row disappeared. Session, Weekly, and Weekly · Fable stayed. The server still had the data: both Claude provider status caches held a monthly_spend window after the refresh. Diagnosis The hub read never produces a monthly spend window, and the client lets a newer hub read

• Signal Oct 21, 2026, 7:32 PM

  Remove models already deprecated (GPT-4o, GPT-5.1 series, Gemini 3 Pro) from "Model multipliers for annual plans on request-based billing (legacy)"

  github/docs — ### Code of Conduct [x] I have read and agree to the GitHub Docs project's Code of Conduct What article on docs.github.com is affected? https://docs.github.com/en/enterprise-cloud@latest/copilot/reference/copilot-billing/request-based-billing-legacy/model-multipliers-for-annual-plans What part(s) of the article would you like to see updated? I would like to remove some models that were already deprecated from GitHub Copilot GPT-4o GPT-4o mini GPT-5.1 GPT-5.1-Codex GPT-5.1-Codex-Min

• Signal Oct 21, 2026, 3:58 PM

  [P2] Pricing: cursor-admin pricing (models-and-pricing) — reviewed claims changed

  futurastudio/ai-spend-agent — <!-- provider-contract-alert:v1 key=resource:9cf95110e2a8bf61 --> P2 · Pricing · reviewed claims changed Reviewed semantic fingerprint changed. This page states official rates that reviewed pricing claims and estimates depend on. | | | | --- | --- | | Contract | cursor-admin · Cursor Admin API spending and usage events | | Page | pricing · https://cursor.com/docs/models-and-pricing | | Alert class | pricing | | Required markers | Models & Pricing, Usage pools, Paid

• Signal Oct 21, 2026, 3:25 PM

  A Claude usage limit switches the session to Codex while other Claude accounts have quota

  syncra-studio/brigadier-ai — Seen 10-21 ~18:19 in a brigadier-ai session (branch brigadier/6ed2fe57/session). The thread hit Claude's session limit and the app said "Claude Code hit its usage limit; continuing on Codex gpt-6-astra until 19:10, then back on Claude Code." Two other signed-in Claude accounts had fresh quota at the time. The session should have kept going on Claude.

• Signal Oct 21, 2026, 8:44 AM

  Price check could not read https://developers.openai.com/api/docs/pricing.md

  antharmaya/tokencost — The daily price check could not use https://developers.openai.com/api/docs/pricing.md: openai pricing page: column "Short context cache writes" not found; header is: Model | Short context input | Short context output | Short context cached input | Short context 30m write | Short context 12h write | Long context input | Long context output | Long context cached input | Long context 30m write | Long context 12h write Nothing was guessed. These rows could not be verified auto

• Signal Oct 21, 2026, 8:33 AM

  [Feature] Switch running Claude sessions to the new account on a headless/remote host (account switch via RPC/CLI, usage-limit event)

  stablyai/orca — ### Setup Orca 1.4.230 on macOS, running on an always-on remote host (SSH/Tailscale) that hosts dozens of long-lived Claude Code sessions in separate worktrees, including unattended overnight runs. Multiple managed Claude accounts. Problem When the active Claude account hits its usage limit, an external watcher switches the active account through the accounts.selectClaude RPC. Only newly launched sessions use the new account. Already-running sessions stay on the old profile and s

• Signal Oct 21, 2026, 5:56 AM

  [delphi:rfp] gemini.free — Gemini API free tier

  timerloggedout-spec/termux-monorepo — ### Surface id gemini.free Lane model Filed by (collaboration lane) delphi Gantt plan oracle facts | T0 | T0 | - panel rounds | T0 | T0 | oracle facts shadow lane (5% traffic) | T0+1d | T0+7d | panel rounds promote or park | T0+7d | T0+7d | shadow lane Free quota RPM/TPM/RPD per model: (no fetched page states a number) per-model — source https://ai.google.dev/gemini-api/docs/rate-limits Reset and zone midnight America/Los_Angeles (daily (RPD)) Auth from Acti

• Signal Oct 21, 2026, 4:17 AM

  /ai-coding-tools-pricing FAQ says Copilot Pro at $10 is the cheapest paid plan; the page itself lists Cursor Plus and ChatGPT Go at $8

  robhunter/agentdeals — ## Defect /ai-coding-tools-pricing (about 40 readers a week and rising; Claude Code, claude.ai and ChatGPT) answers the FAQ "Is GitHub Copilot still the cheapest AI coding tool?" with: Yes — Copilot Pro at $10/month is the cheapest paid plan among the tools compared here. The same page lists two cheaper paid plans: Cursor Plus at $8/month (Cursor's entry plan for new subscribers since 2026-10-12) and ChatGPT Go at $8/month, which includes Codex. The text is hand-typed in s

• Signal Oct 20, 2026, 11:46 PM

  core: rateLimitOf carries the quota Gemini names (size, unit, free tier, model)

  amenophis1er/byok — rateLimitOf reads Gemini's QuotaFailure only to tell a minute window from a day window. The same violation says more, and a caller needs it to word the failure: { "@type": "type.googleapis.com/google.rpc.QuotaFailure", "violations": [{ "quotaMetric": "generativelanguage.googleapis.com/generate_content_free_tier_requests", "quotaId": "GenerateRequestsPerDayPerProjectPerModel-FreeTier", "quotaDimensions": { "location": "global", "model": "gemini-2.5-pro" }, "quotaValue": "50" }

• Signal Oct 20, 2026, 10:02 PM

  Claude: same-pool usage-limit stall exposes no usable recovery action; explicit member application restores progress

  happier-dev/happier — ### Summary A Claude session stopped making progress after repeated provider usage-limit responses and an account change inside one Connected Services pool. The pool selected a connected replacement profile, while session metadata still recorded the previous member. No usable recovery attempt was exposed. Explicitly applying the already-selected member through the daemon switch coordinator, followed by one continuation, recovered the same native conversation without restart

• Signal Oct 20, 2026, 8:03 PM

  fix(Frontend): untranslated descriptions are re-requested on every hero slide (free-tier Gemini spend)

  Slow-Inc/MangaDock — ## Bug (EN) Where: Frontend/app/lib/translateDescription.ts (used by the home hero, BookRow, ContinueReadingRow, BookDetailModal). What happens: only successful translations are cached. When the backend answers translated: false (LLM unavailable, quota, nothing to translate) the function throws, cacheOrFetch stores nothing, and the next call fetches again. The home hero carousel calls it on every slide, so an open home page re-requests the same descriptions every ~5 s. Measu

• Signal Oct 20, 2026, 3:56 PM

  [Work Item - Model Upgrade] Upgrade Agent from Deprecated gemini-2.5-flash to gemini-3.8-flash (Global Endpoint)

  jimmycgz/challenge2-117 — ## Overview & Live Model Validation Per team decision, @jimmycgz is upgrading the ADK agent reasoning model from deprecated gemini-2.5-flash to gemini-3.8-flash. We validated gemini-3.8-flash directly against Vertex AI in our Qwiklabs project (qwiklabs-gcp-04-6c7101c1dba1) including function calling (lookup_shelf_life). Assignee: @jimmycgz (owner: jimmycgz) Status: 🔴 status: in-progress (needs-pickup — @jimmycgz to mark complete once deployed) --- Live Vertex AI Valida

• Signal Oct 20, 2026, 2:53 PM

  GitHub Changelog:GitHub Copilot models deprecated on October 19, 2026

  forks-felickz/github-changelog-reader — # GitHub Copilot models deprecated on October 19, 2026 <!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN" "http://www.w3.org/TR/REC-html40/loose.dtd"> As of today, October 19, 2026, we have deprecated the following models across most GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions). These deprecations were announced on September 18, 2026. Model Deprecation date Suggested alternative Gem

• Signal Oct 20, 2026, 2:33 PM

  GitHub Changelog:GitHub Copilot models deprecated on October 19, 2026

  renefritze/github-changelog-reader — # GitHub Copilot models deprecated on October 19, 2026 <!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN" "http://www.w3.org/TR/REC-html40/loose.dtd"> As of today, October 19, 2026, we have deprecated the following models across most GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions). These deprecations were announced on September 18, 2026. Model Deprecation date Suggested alternative Gemini

• Signal Oct 20, 2026, 1:30 PM

  🔴 Breaking Change: GitHub Copilot models deprecated on October 19, 2026

  sujithq/updates — ## GitHub Copilot models deprecated on October 19, 2026 Source: GitHub Changelog Severity: CRITICAL Published: 2026-10-19T21:27:10+00:00 Deadline: ⏰ 2026-10-19 Impact Specified Copilot models are deprecated across most Copilot experiences. Recommended Action Move workflows to supported Copilot models immediately. Full Summary As of today, October 19, 2026, we have deprecated the following models across most GitHub Copilot experiences (including Copilot Chat, inline edits, ask a

• Signal Oct 20, 2026, 1:29 PM

  [product-watch] GitHub Copilot models deprecated on October 19, 2026

  shariqh/github-enterprise-settings-configurator — <!-- product-watch:key:7165bd772b7538c8bf5c6113 --> <!-- product-watch:fingerprint:850da14a57c02c9645be72abebcbfa4dd6180baa740427ad361dbd20a4d744b2 --> <!-- product-watch:managed --> Product-change review This is a deterministic review candidate. The automation does not edit catalog, recommendation, persistence, scoring, or application code. Source evidence | Field | Value | | --- | --- | | Source | GitHub Changelog | | Entry | GitHub Copilot mod

• Signal Oct 20, 2026, 11:38 AM

  [FEATURE] Hide deprecated models from /v1/models by default, list them with?include=deprecated

  inference-gateway/inference-gateway — /v1/models relays each provider's catalog as-is, and operators filter it with ALLOWED_MODELS / DISALLOWED_MODELS. Several of our providers (OpenAI, Mistral, MiniMax, DeepSeek, Google) keep deprecated models in their listing, so clients discover models that are about to disappear. PR #792 proposed a MiniMax-only hardcoded list in providers/core plus a provider check in the handler. That does not scale and was closed. Proposal: sync models.dev's status field i

• Signal Oct 20, 2026, 10:08 AM

  GitHub Changelog:GitHub Copilot models deprecated on October 19, 2026

  abirismyname/github-changelog-reader — # GitHub Copilot models deprecated on October 19, 2026 <!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN" "http://www.w3.org/TR/REC-html40/loose.dtd"> As of today, October 19, 2026, we have deprecated the following models across most GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions). These deprecations were announced on September 18, 2026. Model Deprecation date Suggested alternative Gemi

• Signal Oct 20, 2026, 9:48 AM

  Is Claude Code support still needed now that Claude Code auto-continues at usage limits (2.1.234+)?

  saaranshM/unsnooze — Claude Code now has a built-in "continue automatically at usage limit" feature. Is unsnooze's Claude Code support still needed, and if so for which cases? What Claude Code ships From the Claude Code CHANGELOG: 2.1.234 - "Claude Code now continues your session automatically when a claude.ai usage limit resets; turn it off in /config ("Continue automatically at usage limit")" 2.1.269 - "Fixed /goal runs silently stalling after API errors, network drops, or token limits: the go

• Signal Oct 20, 2026, 6:07 AM

  [source] GitHub Copilot models deprecated on October 19, 2026

  steveash/hitchhikers-guide-to-ai-native-engineering — ### Source URL https://github.blog/changelog/2026-10-19-github-copilot-models-deprecated-on-october-19-2026 Source Type documentation What's interesting about this source? Auto-discovered from trusted feed github-copilot (GitHub Copilot changelog — Copilot-specific feature changes and updates). Feed: https://github.blog/changelog/feed/?label=copilot Entry title: GitHub Copilot models deprecated on October 19, 2026 Published: Mon, 19 Oct 2026

• Signal Oct 19, 2026, 11:48 PM

  pingdotgg/t3code — ### Before submitting [x] I searched existing issues and did not find a duplicate. [x] I included enough detail to reproduce or investigate the problem. Area apps/server Steps to reproduce Turn on Auto-resume limited threads in Settings → General. Run a Claude thread until the account hits its 5-hour limit. The run fails with usage_limit and a resetAt. Wait past the reset. This happened to two threads on 2026-10-18, and both behaved the same way. I can't make Claude refuse at

• Signal Oct 19, 2026, 11:02 PM

  overnight: a queue of claude -p tasks run while nobody answers, resumed after the usage limit resets

  lovaizu/ccpm — ## What it should be overnight, a plugin of its own, runs a queue of claude -p tasks from a set time while nobody answers. It runs three at once. A task that hits the usage limit waits until the limit resets and goes on in the same session; a task stopped by the limit seven times is given up and logged. In the morning the user reads which tasks finished and which were given up, and takes the results in from the main session, where git is done. Other plugins call its wait-and-resum

• Signal Oct 19, 2026, 10:47 PM

  Claude logins: app chats pick a login, and a running chat moves to the next login on a usage limit

  OpenAgentsInc/openagents — Follow-up to #12211, which landed named Claude logins in the capacity book (openagents capacity logins, add-login, remove-login), per-login refusals, and login rotation for new turns in the Microcoder loop's Claude lane (Coder's delegate door, repository runs, the local chain) and for new Claude Code sessions from Coder's CLI delegations. Still to do: OpenAgents app chats. apps/openagents/crates/harness/src/claude/mod.rs (build_command) launches Claude Code for app cha

• Signal Oct 19, 2026, 10:02 PM

  Capacity book: hold several Claude logins and rotate on a usage limit

  OpenAgentsInc/openagents — On 2026-10-19 at 21:17 UTC the owner's Claude weekly limit ran out. Every Claude chat was told to stop, the owner switched accounts by hand, and the switch broke Foreman (#12205). Let the capacity book (crates/microcoder-loop, capacity.json; openagents capacity) hold several Claude Code logins by name. Store each login's token in the keychain or the secrets folder, never in the book itself. On a usage-limit refusal, record the reset for that login and move new turns to

• Signal Oct 19, 2026, 9:33 PM

  fix: gemini-3.8-flash is too slow on the free tier for the 30 s target

  devTomaszStoklosa/PDF-Insight — ## What is wrong After #16 (PR #30) the demo answers within the browser's limit, but no analysis completes: every request ends with 504 LLM_TIMEOUT after about 25 s. 1,100-character service agreement: 504 after 26.3 s. 394-character invoice: 504 after 25.6 s. Why The default model gemini-3.8-flash (DEFAULT_LLM_MODEL in server/config.ts) is too slow on the free tier. The request itself is small (system prompt 5,649 characters, response schema 1,296 characters, abou

• Signal Oct 19, 2026, 9:04 PM

  The agent-eval examples default to a deprecated OpenAI model, and changing only the model id breaks their tool calls

  traceroot-ai/traceroot — ## Summary Both offline-eval examples (examples/python/agent-eval, examples/typescript/agent-eval) run their agent and LLM judge on gpt-4o-mini by default. OpenAI has deprecated gpt-4o-mini with a shutdown date of 2027-01-14 and names gpt-6-luna as the replacement (deprecations ). The TypeScript example's Claude option pins claude-opus-5, and the LLM-judge snippets in docs/evals/scorers.mdx pin claude-haiku-4-5. Anthropic lists both under "Legacy models (still available)

• Signal Oct 19, 2026, 5:07 PM

  Show Claude Code's rate-limit status from usage_update _meta

  jpsutton/patchbay — claude-agent-acp attaches _meta["_claude/rateLimit"] to usage_update: { status: "allowed" | "allowed_warning" | "rejected", rateLimitType?, utilization?, resetsAt? } Notes from ACP Patchbay: It's sent only when the value changes, so keep the last reading across plain usage updates. resetsAt may be seconds or milliseconds (they treat values above 1e11 as milliseconds). The unit of utilization isn't documented. What to do Read this in the usage path and show a warning or limite

• Signal Oct 19, 2026, 4:39 PM

  Native harness usage telemetry: Claude tokens, Codex rate limits, OTLP export

  kzarzycki/omnigent — > 🤖 AI Problem The homelab APM (trinity-apm: OTel collector → Prometheus → Grafana) can't chart Omnigent usage, and session_usage itself has two holes: claude-native records cost but no tokens. session_usage and by_model get total_cost_usd from the statusLine, but input_tokens, cache_read_input_tokens, cache_creation_input_tokens and output_tokens stay empty. The forwarder only posts per-call counts, which the server ignores for native sessions. codex-native drops the subsc

• Signal Oct 19, 2026, 4:31 PM

  GLM helper does not report back when its Claude-hosted run hits a usage limit

  notdefined-inc/opencostudio-releases — ## What happened\n\nA helper using a GLM model through the Claude agent reached its usage limit but remained shown as working and did not automatically report the limit back to the lead chat.\n\nThe lead only learned what happened after manually checking or stopping the helper.\n\n## Expected behavior\n\nWhen the underlying run reaches a usage limit:\n\n- the helper changes to a limited/paused state promptly;\n- the lead receives an automatic report;\n- the

• Signal Oct 19, 2026, 2:55 PM

  ESI-88 [esi] ADR-0007 model provider: Gemini API free tier

  dtnam0899/edge-side-includes — Choose the model provider for services/ai-agents (ADR-0001 deferred it). Product owner chose the Gemini API free tier for the demo (2026-10-19): free, no card; low and changing rate limits; free-tier prompts may be used by Google to improve products and may be human-reviewed, so only demo data (sample target database) goes through it. Decide the model, the SDK or plain HTTP, timeouts and retries against the existing ModelClient protocol and guarded_complete, how 42

• Signal Oct 19, 2026, 1:27 PM

  Model space-bunny-free has been deprecated. Use minimax-m3.1-flash-free instead.

  anomalyco/opencode — ### Description Model space-bunny-free has been deprecated. Use minimax-m3.1-flash-free instead. Plugins None OpenCode version 2.0.32 Steps to reproduce Just chatting Screenshot Operating System win10 Terminal _No response_

• Signal Oct 19, 2026, 11:42 AM

  breaker: a Claude weekly or 5h limit, or a Codex quota error, is a usage_cap that holds the provider breaker until the reported reset time

  on-par/software-factory — ## User story As an operator running several AI providers, I want each harness adapter to turn its own provider errors into one common failure shape, a kind plus an optional resetAt, so that the breaker and failover work the same for every provider and don't depend on how Claude words its limits. INVEST: independent (the classifier contract, the breaker's use of resetAt, and the first adapter, Claude; the Codex adapter is #3227); negotiable (the interface shape, see Pen

• Signal Oct 19, 2026, 9:10 AM

  No notice when Claude moves on to paid extra usage after the subscription limit

  pingdotgg/t3code — ### What happened The user reached their Claude subscription limit while working in T3 Code. Claude did not stop. It moved on to paid extra usage credits, and T3 Code showed nothing. The user found out later that they had spent about €500 of extra usage. T3 Code shows a notice when a Claude request is blocked by a usage limit. It shows no notice when the limit is reached and Claude keeps going on extra usage. The user asked whether a setting exists for this. None does. Request

• Signal Oct 19, 2026, 8:39 AM

  claude-cli model-specific usage limit ("reached your Fable limit") misclassified as unknown → auth profile cooldown ignores model scope

  openclaw/openclaw — # [Bug]: claude-cli model-specific usage limit ("reached your Fable limit") misclassified as unknown → auth profile cooldown ignores model scope Summary When the claude-cli backend exhausts its model-specific weekly pool and returns "You've reached your Fable limit. Switch to another model, or manage usage credits …", three related defects in 2026.10.5 cause an account-wide cooldown rather than a model-scoped one: Defect 1 — missing classifier: classifyFailoverReason (via cre

• Signal Oct 19, 2026, 7:43 AM

  pingdotgg/t3code — ### Before submitting [x] I searched existing issues and did not find a duplicate. [x] I included enough detail to reproduce or investigate the problem. Area apps/server Steps to reproduce Observed on 2026-10-19 in thread 1f91c9f2-b318-40d4-9de5-a23fcce89528 (local deep link: t3-thread://v1/1f91c9f2-b318-40d4-9de5-a23fcce89528). Use a long Claude Opus 5.5 thread until the session usage limit is reached. The UI showed a reset at 15:30 (Asia/Makassar). Snooze the thread until th

• Signal Oct 19, 2026, 6:38 AM

  Document and respect Gemini free tier input-tokens-per-minute limit

  devTomaszStoklosa/fin-ai-lab — > Moved from the private development archive, where this was issue 293. What core.llm (GeminiLlmClient, src/fin_ai_lab/core/llm/client.py) only slows down for the free tier's 5 requests per minute. The free tier also caps input at 250K tokens per minute (quotaId: GenerateContentInputTokensPerModelPerMinute), and nothing waits for that limit. docs/LLM-API.md doesn't mention it either (line 20 only says "RPM/RPD/TPM per model" in general). Why it matters On 2026-10-1

• Signal Oct 19, 2026, 3:52 AM

  Gemini: trocar os usos pagos por free tier ou Workers AI (embeddings, tradução do É IA?, fallback de imagem)

  vjpixel/diaria-studio — ## Pedido do editor (19/10/2026) "avalie em quais desses processos a gente pode usar outra ferramenta, pois quero economizar" Contexto: a conta Gemini ficou sem crédito pré-pago (HTTP 402, #10448). Esta issue avalia se dá para parar de depender de crédito pago no Gemini. Usos atuais da GEMINI_API_KEY | Uso | Onde | Sem crédito hoje | Alternativa sem custo | |---|---|---|---| | Embeddings para agrupar artigos do mesmo evento (Stage 1) | scripts/topic-cluster.ts, gemini.emb

• Signal Oct 18, 2026, 11:09 PM

  Report Claude usage and plan limits for timed-out, stopped and shut-down jobs, not only finished ones

  bmazurme/pipe — > Направление: Общий анализ · риск: low · предложено автоматическим анализом Problem The log digest shows Job N failed: claude timed out after Ns and was killed ×8 and Worker was stopped (SIGTERM) before the job finished ×28 this week. Those are the longest, most expensive runs, and none of their tokens, cost or plan-limit readings ever reach bridge: worker/src/modelRunners/claudeRunner.ts:133-141: on close, a cancelled run rejects with CancelledError and a timed-out run rejects

• Signal Oct 18, 2026, 3:13 PM

  Goal leads parked for good when the Claude usage limit kills their launches: goal_tick_one counts limit exits as no progress (#373)

  jpmoya/claude-agents — INCIDENT jpmoya/claude-agents#373 (client self-service portal goal) — 4 Mac goals parked "no progress" overnight and never restarted, 2026-10-17 22:19Z → 2026-10-18 08:29Z CLASS supervisor/launcher logic (deterministic) CAUSE CONFIRMED — during the weekly Claude usage limit every goal-lead launch died at once; goal_tick_one counts each as "no progress", parks the goal at 3, and a park never lifts by itself. Runs already treat a limit exit as "wait, don't count"; goals don'

• Signal Oct 18, 2026, 2:17 PM

  run-eval: stop with INCOMPLETE when claude reports a usage limit

  Dominik2806/incident-agent — The first calibration attempt on the branch of #40 (results 20261018T120326Z) hit the Claude usage limit during api-outage run 3. claude-result.json of the affected runs has api_error usage_limit_reached, api_error_status 429 and the result text: You have hit your session limit, resets 3:30pm. run-eval treated these as ordinary failed runs and went on: ten runs failed within 2 seconds each, and the verdicts said FAIL, which looks like a failed evaluation although not

• Signal Oct 18, 2026, 7:36 AM

  A review that hits the Claude usage limit is reported as a full review

  Newton-School/watchtower — ## What happened 5825318d: the account hit its usage limit mid-run (agent.usage_limit). The orchestrator and the reviewer lens failed, and the verifier was inconclusive. Only the security and performance lenses produced output, and the 27 earlier threads got no follow-up status. The Slack footer still said the standard three-lens review ran. Proposed fix When a lens or the orchestrator fails on the usage limit, say so: "Partial review: hit the Claude usage limit; only

• Signal Oct 17, 2026, 11:33 PM

  Harness Live Tests red: OUT_OF_REGION_MODEL qwen/qwen3-max was deprecated 2026-10-09 — pick a model the US host still refuses in-region

  basecradle/basecradle-harness — ## Trigger The Better Stack heartbeat Harness Live Tests has been down since 2026-10-15 07:23 UTC (incident 1032313042). The NOC's harness-live-smoke openrouter arm withholds it because one live test fails on every run since then: FAILED tests/test_openrouter_live.py::test_a_live_model_with_no_endpoint_in_the_region_names_its_cause tests/test_openrouter_live.py:481: AssertionError llm provider=openrouter purpose=memory kind=rerank model=qwen/qwen3-max … outcome=fa

• Signal Oct 17, 2026, 10:54 PM

  Answers and the fallback question detector use Gemini's free tier; remove Anthropic

  eugnel-ltd/RealTimeTranslation — ## Problem James wants free quota only for third-party AI APIs (relayed by Grok Bot, 17 Oct 2026 22:48 UTC): "rtt: James wants free quota only. I deleted ANTHROPIC_API_KEY from the rtt Worker (version b2f61d91), so /api/answer now returns 503. Please change rtt so answers use Gemini's free tier (GEMINI_API_KEY is still set), pick a model that has a free tier, remove the Anthropic dependency". Evidence: worker/src/answer.ts returns 503 ANTHROPIC_API_KEY is not con

• Signal Oct 17, 2026, 8:28 PM

  Claude accounts: show each account's usage-limit state and reset time in the accounts table

  jayden1711/worklane — v0.5.2. Follow-up to #464 (shipped in v0.5.1). The Machine page's accounts × instances table shows each account's active state and last real call, but not its usage-limit state. Measured: web/src/components/ClaudeProfiles.tsx at v0.5.1 shows the failover setting but no per-account limit or reset time; #464's addendum asked for "its usage-limit state with the reset time". Show, per account and instance, whether it is under a usage limit and when it resets, from the same data