GitHub Issues
Community · source page ↗ · last checked —
Change history
- Signal Oct 9, 2026, 8:32 AM
robhunter/agentdeals — ## What is wrong OpenAI cut its API usage tiers from five to three on 2026-10-06. Two hand-typed sentences in src/serve.ts still describe the old tiers: /gemini-api-pricing-2026 (1,898 people in the 7 days to 10-09), section 4, src/serve.ts:23003: "... and OpenAI sets each organization a monthly usage limit ($100 on Tier 1)." There is no Tier 1 now, and the entry tier's limit is $500 a month. /llm-api-pricing (39 people a week), "Rate Limits Follow Usage Tiers", src/serve.
- Signal Oct 8, 2026, 11:56 PM
Omit deprecated Gemini sampling parameters for newer models
groupthinking/EventRelay — Prevent Gemini 3+ requests from sending deprecated sampling overrides (e.g., temperature) while preserving overrides for older models and other providers.
- Signal Oct 8, 2026, 11:20 PM
Anthropic LLM provider fails on current Claude models: "temperature is deprecated for this model"
OneUptime/oneuptime — ### Anthropic LLM provider fails on current Claude models: "temperature is deprecated for this model" Describe the bug The Anthropic path in LLMService.getAnthropicCompletion (packages/Common/Server/Utils/LLM/LLMService.ts) always sends temperature (request.temperature?? 0.7; incident/alert features pass 0.2, the provider test passes 0). Current Claude models reject any value other than 1 with HTTP 400: {"type":"error","error":{"type":"invalid_request_error","message":"temp
- Signal Oct 8, 2026, 7:31 PM
Switch the solver to free models (Gemini 3.8 Flash free tier + one free fallback)
AbhishekBalija/MathGPT — ## What Move the solver to free models: Primary: Gemini 3.8 Flash on the Gemini API free tier (same key, GEMINI_MATH_AI_API). Fallback: one free OpenRouter model, picked by the test below. Candidates seen on 2026-10-09: nvidia/nemotron-3-super-120b-a12b:free (structured output), google/gemma-4-31b-it:free (images). Why Both current backups (google/gemini-2.0-flash-001, tngtech/deepseek-r1t2-chimera:free) no longer exist on OpenRouter, and gemini-2.5-flash is legacy. USE_
- Signal Oct 8, 2026, 5:40 PM
fix(engine): read the reset time of a Claude Code usage limit correctly
Mobius-Toolkit/Mobius — Goal Make the pause of a Claude Code usage limit end at the correct reset time. Cause (see the result in #347) decodeObject (internal/engine/transcript.go) calls decoder.UseNumber(), so each JSON number is a json.Number. Agent.record (internal/engine/agents.go) reads _meta._claude/rateLimit.resetsAt with.(float64). This type assertion always fails, so resetHint stays zero. Then detect uses resetsAt (internal/engine/limits.go). It reads "resets 5:10pm (Europe/Warsaw)", ign
- Signal Oct 8, 2026, 5:07 PM
TabularisDB/tabularis — ### Describe the bug With the Anthropic provider, AI Assist / Generate SQL fails whenever the default model is a Claude 5.5 model (claude-sonnet-5-5, claude-haiku-5-5). The Anthropic API rejects the request: Anthropic Error: {"type":"error","error":{"type":"invalid_request_error","message":"temperature is deprecated for this model."},"request_id":"req_011Cfq7zz6YeCsc1BR8UZJpx"} The same setup works with Claude 4.7 models, so the API key and configuration are fine. The cau
- Signal Oct 8, 2026, 3:24 PM
Gemini 3: thinkingBudget: 0 sent when reasoning is disabled; deprecated, will 400 on upcoming models
NousResearch/hermes-agent — ## Summary _build_gemini_thinking_config (agent/transports/chat_completions.py) sends thinkingBudget: 0 for gemini-3* and gemini-flash-latest when reasoning is disabled. Google has announced that thinking_budget is deprecated for Gemini 3: requests that set it will return 400 INVALID_ARGUMENT on upcoming models (today it is still silently remapped to thinking_level). Google AI Studio is already emailing users ("[Action Required] Update thinking_budget and sampling par
- Signal Oct 8, 2026, 12:50 PM
MervinPraison/PraisonAIDocs — ## Context Upstream PR MervinPraison/PraisonAI#5344 (merged on 2026-10-08) changes the default model for the Python realtime capability and its CLI from the OpenAI-deprecated gpt-4o-realtime-preview to the GA (non-deprecated) gpt-realtime. The SDK is now the source of truth for gpt-realtime; existing docs still show the deprecated id and must be brought back in sync. Upstream PR: https://github.com/MervinPraison/PraisonAI/pull/5344 Upstream issue: https://github.com
- Signal Oct 8, 2026, 9:11 AM
OpenAI provider sends deprecated max_tokens, which newer models reject
nucket/NekoAI — ## Summary OpenAIProvider sends max_tokens. Newer OpenAI models (o-series, gpt-5 family) reject this parameter with HTTP 400 and expect max_completion_tokens. A user who types one of those models into Settings gets "Sorry, something went wrong". Where src/ai/providers/openai.ts (line 23) Proposed fix Send max_completion_tokens (accepted by current chat models). Consider updating the default model (gpt-4o-mini) as well. Acceptance criteria [ ] Chat works with both gpt-4o-mini and
- Signal Oct 8, 2026, 9:05 AM
repo-ci-baseline: Claude Review on Opus can strand a PR when the subscription's usage limit is hit
mpaulosky/dotfiles — ## Problem Statement Claude Review authenticates with CLAUDE_CODE_OAUTH_TOKEN, a subscription token from claude setup-token, not an API key. Since #120 it runs Opus 5.5 at high effort, which draws on the plan's Opus usage limits. ADR 0004 priced it as about $1.20 a review, which is API-style billing. In a re-Apply round (six repos, a review on every new head while review:claude is on, several minutes of Opus each), the account can hit its usage limit. A refused run fails the
- Signal Oct 8, 2026, 7:39 AM
beeltec/aiproxy — ## Problem aiproxy does not expose the native OpenAI Decisions API. Model-only pricing cannot distinguish Decisions billing from normal Luna requests. Scope Add native POST /v1/decisions for OpenAI API-key connections and gpt-6-luna. Preserve native payloads and answers, model routing, permissions, limits, cancellation, and usage accounting. Add endpoint-aware price selection, versions, overrides, long-context rates, and regional billing facts. Update endpoint and pricing docum
- Signal Oct 8, 2026, 4:49 AM
KempnerInstitute/hpc-agentic-recipes — ### Describe the Feature Claude Code now marks ANTHROPIC_SMALL_FAST_MODEL as deprecated and uses ANTHROPIC_DEFAULT_HAIKU_MODEL instead (see the environment variables reference ). The README, docs/quickstart.md, the recipe README template, and 27 recipe README and client.env files still set the old name. Switch them to ANTHROPIC_DEFAULT_HAIKU_MODEL. Use Case Users who copy a recipe's client settings get the supported variable, and the computing handbook can
- Signal Oct 8, 2026, 3:00 AM
aiur-team/aiur — ## Follow-up to #3029 / #3038 When an account hits its usage limit, the session handoff (#3038) moves only claude-repl sessions to another account. Headless claude workers, which run through the aiur-claude app-server, stay on wait-for-reset. The app-server keeps its thread map in memory, and thread/start cannot seed a moved session transcript (see handoff_candidate? in src/lib/aiur/orchestrator/rate_limit_fallback.ex). Done when Resume support. aiur-claude can resume a session
- Signal Oct 8, 2026, 2:02 AM
Add retry with exponential backoff for Gemini API free tier
andreshg112/etoro-trading — Currently, calls to the Gemini API don't retry on 429/503 errors. Adding an exponential backoff wrapper would make the bot much more reliable when running on Google AI Studio's free tier.
- Signal Oct 7, 2026, 10:45 PM
Zenvi-pro/zenvi-core —
- Signal Oct 7, 2026, 5:35 PM
Let users redeem Claude usage limit resets (experimental)
rock3r/headroom — Headroom shows Claude's usage limit resets, but cannot use them yet. Codex and Grok resets can already be used from the reset sheet. What Claude's API needs (from research, not tried live) Redeem with POST https://api.anthropic.com/api/organizations/{org}/reset_rate_limits, with the same OAuth token and anthropic-beta: oauth-2025-04-20. Body: {"program":"cedar_ember","grant_id":" ","request_id":" "}. Only the grant named next_grant_id can be used. The org id is organization.uui
- Signal Oct 7, 2026, 3:32 PM
Remove deprecated Gemini generation parameters before the next model change
cinatra-ai/gemini-connector — ## Context Google announced that the Gemini API will stop accepting some generation parameters on its upcoming models: thinking_budget (SDK: thinkingConfig.thinkingBudget) will return 400 INVALID_ARGUMENT instead of being remapped. Use a thinking level (SDK: thinkingConfig.thinkingLevel, values minimal, low, medium, high) or omit it. Supported levels differ per model. Google already documents intermittent 400 errors for the deprecated budget on Gemini 3 series model
- Signal Oct 7, 2026, 11:55 AM
fix(registry): audit and models-generate ignore models.dev status: deprecated
laurigates/pal-mcp-server — ## What scripts/audit_model_registry.py and scripts/generate_model_entries.py treat every model in a models.dev provider slice as live. Neither reads the per-model status field, which models.dev sets to "deprecated" for withdrawn models. Why it matters After the OpenCode Go sync (the PR that adds space-bunny), just models-audit-one opencode_go_models.json lists space-bunny-free and grok-4.5 as CANDIDATE ADDITIONS. Both carry "status": "deprecated" in models.dev, and b
- Signal Oct 7, 2026, 8:53 AM
LLM: add a free-tier Gemini key to the provider rotation
VishnujanNarayanan/Job_Application_Bot — ## Goal Add a second, free-tier Gemini key to the LLM rotation, so more parses run at $0 and fewer depend on OpenRouter. Details The key comes from a Google AI Studio project with no billing attached, so it can only ever use the free tier. It needs its own provider name, gemini-free. llm.rotation matches on provider names, so a second entry called gemini would pull the paid key into the rotation. It uses gemini-3.5-flash-lite, because gemini-2.5-flash-lit
- Signal Oct 7, 2026, 5:12 AM
Fix Production Planner deprecated Workers AI model before Research AI
toki1031/creator-tools — PR #419 introduced @cf/meta/llama-3.1-8b-instruct. Cloudflare's May 2026 deprecation notice says this model was deprecated May 30, 2026. Before adding Research AI: replace Planner model with a current Workers Free-compatible multilingual text model; preserve JSON ProductionBrief behavior; add a safe diagnostic/contract check where practical; run regression tests, Quality Gates and Cloudflare Preview; do not merge without explicit confirmation. Current Cloudflare docs lis
- Signal Oct 6, 2026, 11:50 PM
[BUG] claude-cli のセッション上限到達（429 usage_limit_reached）検知と復帰時刻に応じた待機・通知
Saltmu/orchestune — ## 不具合の概要 をターゲットとしてタスクを実行中、Anthropic API のセッション上限（You've hit your session limit · resets 1pm / 429 usage_limit_reached）に達してプロセスが終了した場合、Orchestune ディスパッチャーはこれを汎用の LOCAL_PROCESS_DEAD（プロセス消失）としてのみ扱い、エラー原因がセッション制限であることやリセット時刻を認識できません。 そのため、次のサイクルで即座に再起動を試みて再失敗するか、max_task_reclaims を不必要に消費してエスカレーションされてしまいます。 再現手順 dispatch-target = 'claude-cli' でタスクをディスパッチする。 セッション上限に達すると claude コマンドが You've hit your session limit · resets を出力して終了コード 1 で終了する。 ディスパッチャーが LOCAL_PROCESS_DEAD として検知する。 期
- Signal Oct 6, 2026, 11:31 PM
receptron/mulmocast-cli — Follow-up to the verification on #1605 (comment ). #1605 itself covers the OpenAI image default; this issue covers the rest. Rule A default model that is shut down, or has an announced shutdown date, is replaced by the successor its provider names. A model that is already shut down is removed from the models lists. A model with a shutdown date still ahead stays listed until that date. Checked against the providers' model-list APIs (2026-10-07), not only the docs | Provi
- Signal Oct 6, 2026, 9:27 PM
[M3] Wait and resume Claude implementation after usage limits
rcpassos/mergeyard — ## Parent 60: M3 hardening specification What to build Make a Claude implementation survive a temporary usage limit end to end: visible durable waiting, account-wide dispatch gating, bounded waits, explicit Retry, and same-phase resumption across restart. Acceptance criteria [ ] Classify unsuccessful native execution using the verified Claude contract; unknown failures remain ordinary failures and successful output or tool text mentioning a limit cannot trigger a wait. [ ] P
- Signal Oct 6, 2026, 9:27 PM
[M3] Verify Claude headless usage-limit evidence
rcpassos/mergeyard — ## Parent 60: M3 hardening specification What to build Establish an observed Claude contract for recognizing temporary usage limits before shipping its classifier. Acceptance criteria [ ] Capture a real headless usage-limit failure with binary version, invocation, relevant native records, exit/completion behavior, session identity availability, and any reset-time source and units. [ ] Distinguish observed facts from documentation-derived expectations and unobserved variants;
- Signal Oct 6, 2026, 9:00 PM
serious-alchemy/arbiter — Seen on grok run 4a76953c (bd-dv0nhn, 2026-10-06 20:52Z). Investigated read-only; evidence on main (= v0.2.21 + #450). 1. Provider label hardcoded. apps/arbiter/lib/arbiter/worker/claude_session.ex:1893-1901: init_summary/1 and result_summary/1 emit "⚙ claude session started/…" for every provider that runs through ClaudeSession. Grok normalizes into Claude shape (Agents.Grok.Stream.normalize_event/1) and the session knows provider: "grok" (:177,:534), but the summaries
- Signal Oct 6, 2026, 10:31 AM
memU memorize fails on Bedrock Claude 5 models: 'temperature is deprecated for this model'
ClickHouse/nerve — ## Summary With provider.type: bedrock and a Claude 5 model as memory.memorize_model, every memorize call fails: [ERROR] nerve.memory.memu_bridge: memU LLM error [memorize/extract_items]: Error code: 400 - {'message': 'temperature is deprecated for this model.'} after 177ms (prompt=5283 chars) [ERROR] nerve.memory.memu_bridge: memU memorize_file failed for /root/.nerve/memu-conversations/session-cron:memory-maintenance:20261005-090000-1791190805.json: Error code: 400 - {'messa
- Signal Oct 6, 2026, 9:26 AM
Cockpit for Claude Code: usage limits, context and chain progress in the terminal
specnaut/specnaut-cli — ## Why Since v5.0.0 the /specnaut chain runs on autopilot after the plan: implement, review, merge and push without stopping. A developer working only in Claude Code has no in-terminal view of how close they are to the 5-hour and weekly usage limits, and a limit hit mid-chain can cut the run between merge and push. Claude Code (v2.1.287+) ships mods: plugins whose hooks module runs inside Claude Code and can draw a band above the prompt, a pane, a status-line entry and to
- Signal Oct 6, 2026, 6:13 AM
failure_class: Claude Code "weekly limit" message is classified indeterminate, not usage_limit
rysweet/amplihack-rs — ## Problem classify_failure_text (crates/amplihack-cli/src/commands/recipe/run/failure_class.rs) classifies a Claude Code weekly-limit message as indeterminate instead of usage_limit. USAGE_LIMIT_MARKERS holds "session limit", "usage limit", "usage_limit_reached", "quota exceeded", "monthly limit" and "credit balance is too low". The message Claude Code prints when the weekly allowance runs out matches none of them: You've hit your weekly limit · resets Oct 8, 12am (Americ
- Signal Oct 6, 2026, 12:44 AM
What does a running task in the Claude app do when a usage limit is reached?
tiavelum/claude-mechanics — Question: When the account reaches its five-hour or weekly usage limit while a task in the Claude app is running, does the task stop, wait and continue after the reset, or continue on usage credits when they are on? The documentation describes this only for threads of redesigned projects, which wait and continue after the reset (PRJ-089), and for Claude Code. The statement set fable-5-1-2026-10-05 observed, in one app session, subagents stopping with a session-limit m
- Signal Oct 5, 2026, 7:46 PM
audit_content_freshness: deprecated-model table stops at Claude 3.x; Haiku pricing note is stale
brandon-behring/guides-tooling — tooling/audits/guide/audit_content_freshness.py holds a table of deprecated model names.:111-126 flags Claude 3 and 3.5 IDs as "replaced by Claude 4.x" and nothing newer. Models retired since then pass unflagged, for example claude-3-7-sonnet-20250219 (retired 2026-02-19) and claude-sonnet-4-20250514 (retired 2026-06-15).:168-171 says "Haiku 4 is $0.80/$4". Anthropic's models page lists Haiku 4.5 at $1 / $5 per MTok. Options: extend the table by hand at each rele
- Signal Oct 5, 2026, 5:19 PM
Migrate deprecated Groq model to GPT-OSS 120B
Mystify7777/Halotask-Pro — ## Context HaloTask Pro currently hard-codes Groq model llama-3.3-70b-versatile in halotasks-server/src/utils/groqClient.ts. That model is no longer a valid free/developer-tier production choice. Groq's current deprecation documentation says llama-3.3-70b-versatile was deprecated with a shutdown date of 2026-08-16, and the deprecation applies to free and developer-tier usage. Requests using it should therefore be treated as unsupported for this project unless the deplo
- Signal Oct 5, 2026, 4:41 PM
Bino5150/lumina — ### What happens If ANTHROPIC_DEFAULT_MODEL (or the saved default_model for the Anthropic cloud credentials) is set to one of the current Claude models, such as claude-sonnet-5, claude-opus-5 or claude-fable-5-1, every request the Anthropic backend makes comes back as: 400 invalid_request_error: temperature is deprecated for this model That covers normal chat turns, streaming, and the utility calls (titles, dreaming notes, continuity compiler), because all of them put a tempera
- Signal Oct 5, 2026, 7:23 AM
Claude Usage → Limits shows only two of three enabled accounts; third reports Could not read limits
pingdotgg/t3code — ### What happened Three Claude accounts are configured and enabled, but Usage → Limits consistently shows bars for only two. The omitted account has a warning saying “Could not read limits.” This affects the Session, Weekly, and model-specific Weekly sections. It is the usage display that is incomplete; the third provider has not disappeared from configuration. Verified observations All three accounts are enabled in local settings. All three appear in T3's live provider/model
- Signal Oct 5, 2026, 6:37 AM
Claude-Mods: Kontext-% und Plan-Limits aus $.session.usage() statt /usage-Probe
erwins-enkel/shepherd — Ziel: Kontextfüllung in Prozent und die Plan-Limits (5h/Woche) kommen direkt aus der laufenden Claude-Sitzung statt aus einem Wegwerf-claude, in das /usage getippt wird. Heute: Plan-Limits: src/usage-probe.ts:121-198 startet alle 5 Minuten eine interaktive Claude-Instanz, tippt /usage, wartet 900 ms und schickt \r; braucht vorab gesetztes Ordner-Vertrauen (#1075). Tokens pro Sitzung werden nachträglich aus der Transkript-JSONL gelesen (src/usage.ts). Ein Kontext-Prozentwe
- Signal Oct 4, 2026, 2:35 PM
Claude usage-limit and other CLI API errors appear as the bot's reply
Huc06/kind-meitner — On production, Spend Scout 'said' "You've hit your session limit · resets 10am (UTC)". The Claude CLI reports its own failures as assistant frames flagged is_api_error_message/error; only the no-login case was treated as an error.
- Signal Oct 4, 2026, 8:28 AM
bug: failed to import sessions with deprecated models
getpaseo/paseo — ### What's broken I tried to import old sessions using omp and pi, but it failed. Because old sessions use old models like 'xx/deepseek-v3.2-flash', and my current providers do not provide the model anymore. It failed to find the exact model id when importing. But in omp/pi cli, it can import the session with another default model. Steps to reproduce create a new session with an existing model using any harness, like omp/pi. close the session in your paseo app make the model inv
- Signal Oct 4, 2026, 3:09 AM
One meeting exhausts the Gemini free tier: add an OpenRouter fallback
Daniel-c137/stormhacks-2026-MoE — Found in the local end-to-end run of the brain on the Gemini free tier (3 Oct 2026). What happens: one meeting's worth of brain calls exhausts the free tier's per-minute quota. The write-up after the meeting then fails, and the host has to retry. That meeting covers the agenda rewrite, timekeeping, fact-check, three questions, the write-up and the Home follow-ups: about 15–20 model calls in a minute, plus embeddings. Evidence: brain log of the failing run. The m
- Signal Oct 4, 2026, 12:13 AM
[provider-review] gemini: gemini-3-flash-preview now reported free-tier eligible
digithings-ai/digithings — <!-- provider-review --> <!-- dedup-key: gemini:better_free --> Provider change detected: gemini — gemini-3-flash-preview now reported free-tier eligible Trigger: better_free Affected config: N/A (# llm-decision: gemini better_free quota_unconfirmed) Current model: gemini-2.5-flash Finding: gemini-2.5-flash verified ok (1080ms, down from 8449ms). Multiple trackers now list gemini-3-flash-preview as free (1500 RPD) alongside 2.5-flash. Trackers also report 10 RPM / 250K
- Signal Oct 3, 2026, 2:39 PM
GitHub Changelog:Selected models in GitHub Copilot deprecated
forks-felickz/github-changelog-reader — # Selected models in GitHub Copilot deprecated <!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN" "http://www.w3.org/TR/REC-html40/loose.dtd"> As of today, October 2, 2026, we have deprecated the following models across all GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions). Model Deprecation date Suggested alternative Gemini 3.5 Flash 2026-10-02 Gemini 3.8 Flash Gemini 3.6 Flash 2026-10-
- Signal Oct 3, 2026, 2:37 PM
[product-watch] Selected models in GitHub Copilot deprecated
shariqh/github-enterprise-settings-configurator — <!-- product-watch:key:e71ae90584bb71b0d6ae6d40 --> <!-- product-watch:fingerprint:e0890bf5fd5dedf2e3390790867e47d735ea754040d3e0a28a67fbc7fbba916b --> <!-- 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 | Selected models in
- Signal Oct 3, 2026, 2:27 PM
GitHub Changelog:Selected models in GitHub Copilot deprecated
renefritze/github-changelog-reader — # Selected models in GitHub Copilot deprecated <!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN" "http://www.w3.org/TR/REC-html40/loose.dtd"> As of today, October 2, 2026, we have deprecated the following models across all GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions). Model Deprecation date Suggested alternative Gemini 3.5 Flash 2026-10-02 Gemini 3.8 Flash Gemini 3.6 Flash 2026-10-02
- Signal Oct 3, 2026, 12:14 PM
fix(ai-service): Deprecated Gemini and Groq model IDs cause API 404/500 failures
sanjayjakhar/FreshersCompass — ### Summary The AI service configured model names that are now deprecated or unavailable: gemini-1.5-flash (404: model not found) gemini-2.0-flash (404: model deprecated) gemini-flash-latest (500/504: deadline exceeded / internal error) Groq llama-3.3-70b-versatile & llama-3.1-8b-instant (404: model not found) This affected all core AI service modules: Resume parsing (gemini.py) Codebase RAG and Q&A (rag_service.py) LinkedIn profile optimization & post generation (
- Signal Oct 3, 2026, 10:08 AM
GitHub Changelog:Selected models in GitHub Copilot deprecated
abirismyname/github-changelog-reader — # Selected models in GitHub Copilot deprecated <!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN" "http://www.w3.org/TR/REC-html40/loose.dtd"> As of today, October 2, 2026, we have deprecated the following models across all GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions). Model Deprecation date Suggested alternative Gemini 3.5 Flash 2026-10-02 Gemini 3.8 Flash Gemini 3.6 Flash 2026-10-0
- Signal Oct 3, 2026, 6:25 AM
Usage-limit stops resume at the provider's reset time (Claude Code, Codex)
kontourai/station — ## Outcome When Claude Code or Codex stops on a usage limit and the provider reports when the limit resets, Station shows the reset time in the conversation and, if the user opted in, resumes the same Session after the reset. Unknown reset times stay manual. Why The scheduling half already exists (archive#1236): packages/contracts/src/connection-recovery.ts has a wait-until-reset decision, recovery-ledger.ts persists intents, and session-recovery-coordinator.ts dispatches the
- Signal Oct 3, 2026, 6:10 AM
[source] Selected models in GitHub Copilot deprecated
steveash/hitchhikers-guide-to-ai-native-engineering — ### Source URL https://github.blog/changelog/2026-10-02-selected-models-in-github-copilot-deprecated 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: Selected models in GitHub Copilot deprecated Published: Fri, 02 Oct 2026 18:24:19 +0000
- Signal Oct 3, 2026, 3:36 AM
Move question text-to-speech off the deprecated OpenAI speech models before 2027-01-06
yutaasakura96/suburi — ## Why Found while building #45 (the spoken question). The worker checked OpenAI's documentation on 2026-10-03: the speech-endpoint text-to-speech models, including the one suburi pins, were deprecated on 2026-10-01 and are removed on 2027-01-06. The named replacement is only available through the Realtime API. #45 builds on the current speech endpoint behind a one-implementation port, so the voice is swappable, and records the deprecation in docs/06. What it needs Move th
- Signal Oct 2, 2026, 10:56 PM
Bug: Gemini API 429 Rate Limit - Quota exceeded on free-tier preview models
Harshjsh02/Sabha- — ## Description Location: pp/api/meeting/summarize-and-email/route.ts, pp/api/translate/route.ts`n gemini-3-* preview models enforce a strict 20 requests/day quota on the free tier. Live translation calls were hitting Gemini repeatedly during calls, causing HTTP 429 errors and failed translations. Resolution Resolved.\n- Completely removed Gemini from /api/translate in favor of a \ free public translation gateway with memory caching (0 API calls during live calls).\n- Upgrad
- Signal Oct 2, 2026, 6:53 PM
hardbeat920/monocode — Problem Claude Max plans have a separate weekly limit for Fable, on top of the shared 5-hour and weekly limits. MonoCode only shows the 5-hour and weekly bars: in the footer chip, its popover, and the Settings → Accounts row. Fable usage doesn't appear anywhere. That leaves me with two bad options: switch to the CLI and run /usage, or wait until I hit "Fable limit reached" and Claude Code starts drawing from usage credits. On 6bd432c, the Claude usage path only reads two w
- Signal Oct 2, 2026, 3:47 PM
joshualeestone/kosmos — Found in review round 5 of #5029, by reading Claude Code 2.1.287's own strings. Two limit messages the vendor binary contains are not matched by RATE_LIMIT_MARKERS (engine/status.js), so a pane showing either would read idle and the Guide's hosted fallback (#3660) would not switch on: You've hit your team's shared budget. /model to switch models. You're out of usage credits. /model to switch models. Neither has been seen on a live pane. The standing rule in status.js is t
- Signal Oct 2, 2026, 11:31 AM
Hikari9/auto-office — ## Observed Run 25906a65, dispatches D0e891eeb (T7) and D3252759d (T8), both Claude executors launched in herdr at 09:58Z. Both panes stopped mid-task on Claude's session limit and sat idle showing: ⚠ Usage limit reached · limit resets 9:30pm You've hit your session limit · resets 9:30pm (Asia/Manila) Nothing in Office identified the cause. The operator noticed by reading the panes and sent office prompt -- continue by hand after the reset (events prompt seq 3898/3899, 10:2
- Signal Oct 2, 2026, 3:44 AM
eh-homelab/ScadBuddy — Since at least 03:09Z on 2026-10-02, the claude-review Merge-gate classification step fails on every PR review run (8 of the last 9 runs, across feat/951-…, docs/durable-printing-agents-flows, feat/940-…, docs/tracing-spec, feat/844-…, feat/931-…). That fails the claude-review required check regardless of the review's content. From run 36961370920: Action failed with error: --json-schema was provided but Claude did not return structured_output. Result subtype: success The
- Signal Oct 2, 2026, 2:37 AM
Warn when a request targets a model its provider has deprecated or retired
pydantic/pydantic-ai — > This issue was posted by Claude Code using claude-opus-5-5 on behalf of David. Pydantic AI should warn when a request targets a model its provider has deprecated or retired. Why Users find out a model is going away when requests start failing. Only the Anthropic SDK warns today, and only for its own DEPRECATED_MODELS table. Our test suite runs with filterwarnings = error. A warning from Pydantic AI would fail every recording made against a deprecated model as soon as the
- Signal Oct 2, 2026, 1:19 AM
zarda/home-account — ### Summary The model catalog is hand-curated and nothing watches the vendors for the changes it has to follow. The header of ai-models.ts says the check "is not optional maintenance", docs/ai-models.md says "Nothing in the repository can check them", and the date in that header covers the Gemini pages only. A retired id fails at the moment a user reads a receipt, not at build time. The vendors do announce retirements, but Anthropic notifies "customers with active deployment
- Signal Oct 1, 2026, 10:31 PM
cursor USAGE_PRICING_REQUIRED (prepaid balance used up, 429) never triggers credential rotation
can1357/oh-my-pi — ## Symptom Multi-account cursor provider, one account's Other-models pool exhausted. Every request on that session fails with: Cursor USAGE_PRICING_REQUIRED: Your prepaid balance is used up: Add funds or enable auto top-up in your billing settings to keep going. errorStatus is 429. OMP keeps retrying/failing on the same dead credential instead of rotating to sibling cursor accounts that still have balance. cursor has no ranking strategy (hash-pin + rotate-on-wall only), so a m
- Signal Oct 1, 2026, 10:26 AM
Fix deprecated Mistral OCR model usage
spring-projects/spring-ai — Mistral AI is retiring mistral-ocr-4-0 on September 30, 2026. Mistral OCR 4.1 (mistral-ocr-4-1, mistral-ocr-4-latest and mistral-ocr-4) is recommended.
- Signal Oct 1, 2026, 5:58 AM
k-k1/agent-fleet — ## Problem Claude Code shipped two usage-limit behaviours that AF's own limit handling (docs/log/47-turn-abort-auto-resume.md §4-4 to §4-12, workspace/agent/internal/sessionx/rate_limit_resume.go) does not know about. The pinned claude is already past both versions. Built-in auto-continue (v2.1.234+, on by default for interactive claude.ai-subscription sessions; docs ). When a limit stops a turn, claude waits in-process (Usage limit reached · continuing automatically at · esc
- Signal Sep 30, 2026, 11:40 PM
Free-tier media pipeline is unbounded: no upload quotas, every upload runs Demucs + Whisper + Gemini
Rocktown-Labs/mysoundkit — ## Summary Any artist account (free included) can upload unlimited 2 GiB files with no per-user, per-day, or per-file-count quota, and every finalized master triggers an expensive processing pipeline: Demucs stem separation (Cloudflare Container, standard-3), Whisper transcription (@cf/openai/whisper-large-v3-turbo), and Gemini audio embeddings. Evidence apps/server/src/routes/uploads.ts:175-177: maxFileSize: 1024 * 1024 * 1024 * 2, multipart: true, multipleFiles: true
- Signal Sep 30, 2026, 7:40 PM
Add newer LLM models and replace deprecated options
surp-hovhannes/bahk — ## Goal Update Bahk's OpenAI and Anthropic model inventory, add newer models that fit our workloads, and replace deprecated or retired choices without breaking saved prompts, moderation, or generated content. This issue tracks the work. Production usage has not been audited for this issue, and creating it does not approve production changes. Current inventory and deadlines Verified on September 30, 2026 against main 8e70184166959f256c202c2338d09ae1a6f55317. LLMPrompt.MODEL_
- Signal Sep 30, 2026, 4:33 PM
Claude Code returns HTTP 400 for API usage limit and repeatedly reconnects
somani-a/ACC — Description Claude Code cannot start a new session because the first prompt immediately fails with an API usage-limit error. The session then closes the connection and attempts to reconnect even though the error states that access will not return until the next quota reset. Session details Task: Build a Python CLI dev environment orchestrator with DAG-based dependency resolution and live dashboard Task ID: cmo7fmxut000no8j4df6csp8w Trajectory: 2 Model: Model B Model config ID: cmu
- Signal Sep 30, 2026, 1:05 PM
OpenAI integration fails with newer models due to deprecated max_tokens parameter
freescout-help-desk/AiIntegration — Hi, I found an issue with the official AI Integration module when using newer OpenAI models. The OpenAI provider is configured correctly and the API connection works, but when using: gpt-5.6-terra the module returns the following error: HTTP Error: 400 { "error": { "message": "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.", "type": "invalid_request_error", "param": "max_tokens", "code": "unsupported_