Users report that Python SDK releases up to 0.2.157 bundled a Claude Code CLI older than 2.1.280, so setting the model to claude-opus-5-5 failed with a 400 error on every turn unless a newer CLI path was given.
Workaround: Upgrade to SDK 0.2.158 or later, which bundles Claude CLI 2.1.280.
Reported on GitHub · 10 👍 · workaround source
Users report that Claude Code put the signed-in account's email address into the User-Agent header of web requests it sent to third-party sites, without asking first. Several users report that the built-in email check added in recent versions still lets some of these requests through.
Reported on GitHub · 798 👍
Users report that the same workload run headless with `claude -p` (or through the Agent SDK) uses about 1.8 times or more of a subscription's 5-hour usage window than in the interactive terminal, even when token counts match.
Reported on GitHub · 175 👍
Users ask for a desktop app setting where Enter inserts a new line and Ctrl+Enter (or Cmd+Enter) sends the message. They report that the desktop composer ignores the keybindings file that already allows this in the terminal.
Reported on GitHub · 116 👍
Users report that in auto mode Claude Code is instructed to read and edit files through shell commands such as cat, sed and heredocs instead of its own Read, Edit and Write tools, which they say wastes tokens and makes simple edits slower and riskier.
Reported on GitHub · 102 👍
Users report that version 2.1.323 removed the subagentStatusLine setting without a replacement, so custom rows that showed each running subagent's context use, cost, model and effort no longer appear. A project member said the removal should have come with a grace period and that work on one had started.
Reported on GitHub · 38 👍
Users report that when Claude Code runs in Vertex mode behind an LLM gateway (CLAUDE_CODE_SKIP_VERTEX_AUTH=1), a 401 from the gateway does not re-run apiKeyHelper, so an expired token keeps failing until the helper cache runs out; the same setup recovers in Bedrock mode.
Reported on GitHub · 29 👍
Users asked for a way to turn off the short status lines shown for running subagents, reporting that each one is an extra model call that re-sends the subagent's whole context and takes a large share of usage in subagent-heavy sessions.
Workaround: Set the environment variable CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES to turn subagent progress summaries off.
Reported on GitHub · 121 👍 · workaround source
Users ask for Cline's OpenAI Compatible provider to use the Responses API for models that need it: with reasoning effort and tools together, requests sent to /v1/chat/completions fail. A Cline team member says this is not planned yet.
Workaround: Define a custom provider in ~/.cline/data/settings/models.json with "protocol": "openai-responses" and select it with cline -P <provider> -m <model>.
Reported on GitHub · 7 👍 · workaround source
Users report that with Ollama and local models such as Qwen 2.5 Coder, Cline 4.1.x shows tool calls as plain JSON text instead of running them; another user saw the same with Gemma 4 and LFM2.5. The issue is open with no fix announced.
Reported on GitHub · 6 👍
Users report that the Cline desktop app on Windows stopped starting from version 0.0.32, failing with 'desktop backend endpoint not ready', even after a clean reinstall. Two users later said updating to 0.0.40 fixed it for them; the issue is still open.
Reported on GitHub · 1 👍
Users report that since v0.63.0, YOLO mode keeps stopping for manual approval on ordinary shell commands because of a new "Untrusted Command Flags Detected" security warning. Later comments say it still happens in newer versions.
Reported on GitHub · 6 👍
Users on Gemini Code Assist Standard or Enterprise and on API keys ask for support for the new Gemini 4 Argon (Gemini 4 Pro) model in Gemini CLI, which they report is not yet available to them there.
Reported on GitHub · 5 👍
Users report that requesting a specific Flash model with --model (for example gemini-3.7-flash) was silently answered by gemini-3.5-flash with no error or warning, even for model names that do not exist.
Reported on GitHub · 14 👍 · fix source
Users report that newer Gemini Flash models (3.6 and 3.7) were missing from the /model picker for weeks after Google released them, so they could not select them in the CLI. A maintainer closed the issue saying v0.61.0 makes Gemini 3.8 Flash and 3.5 Flash Lite available.
Reported on GitHub · 16 👍
Users ask for an option to use the 1-hour prompt cache for Claude models, reporting that the default 5-minute cache expires during normal pauses, so most follow-up turns rewrite the cache and use more credits.
Reported on GitHub · 10 👍
Users report that since v1.0.81 the /model command is unavailable when the model is set through BYOK environment variables, so they can no longer switch between their own provider's models inside a session.
Workaround: Add the provider and its models to providers.json in ~/.copilot; models configured there appear in the /model picker.
Reported on GitHub · 8 👍 · workaround source
Users report that when Copilot CLI runs as an ACP server (copilot --acp), it lists only the Auto routing model and no Copilot Fusion option or tier setting, although Fusion is available in the interactive CLI, so ACP clients cannot offer it.
Reported on GitHub · 7 👍
Users report that Copilot CLI remote sessions opened in the GitHub Mobile app on iOS stay on 'Working' and never load, while the same sessions work in Safari on the same phone.
Reported on GitHub · 5 👍
Users report that Figma's hosted MCP server signed in but its tools never loaded in Copilot CLI, while the same server worked in VS Code. The reporter closed the issue after saying version 1.0.87 resolved it.
Reported on GitHub · 12 👍
Users report that after a macOS update or reboot, Copilot CLI stopped processing prompts with "The shared writer lock or its directory changed", because a saved file in ~/.copilot kept the disk's old device ID.
Reported on GitHub · 11 👍
Users report that ADK's two ways of migrating old v0 session databases read event timestamps differently, so the same database migrated through each path ends up with times shifted by the host's UTC offset. A maintainer says a merged pull request fixed it; the thread does not name a release.
Reported on GitHub · 1 👍
Users report that the Graphify plugin for OpenCode used the old v1 plugin API and did not load in OpenCode v2.
Reported on GitHub · 8 👍 · fix source
Users report that since Codex CLI 0.157 started a shared background daemon, herdr binds Codex sessions to the wrong pane or loses them, and HERDR_* variables are missing or wrong inside Codex tool calls. Maintainers say herdr cannot fix this reliably on its side and are waiting for an upstream Codex fix.
Reported on GitHub · 8 👍
Users report that when herdr runs on a remote machine over SSH or herdr --remote, copied text lands on the remote machine's clipboard instead of the local one; some say this blocks remote use for them.
Reported on GitHub · 7 👍
Users report that when the system switches between light and dark mode, programs running inside herdr panes (for example Claude Code with an automatic theme, Neovim or Helix) are never told and keep the theme they started with. The fix is on master and queued for the next release; no published release includes it yet.
Reported on GitHub · 11 👍
Users report that herdr machine add asked for an SSH one-time code on hosts behind MFA and then failed with a lost-connection error, even though herdr --remote worked with the same host.
Reported on GitHub · 9 👍 · fix source
Users report that remote sessions became sluggish after the 0.9.0 update, with clicks and typing taking seconds to register over herdr --remote while the same machine stays fast over plain SSH. The issue was closed after a fix shipped on the preview channel; the thread does not name a stable version.
Reported on GitHub · 8 👍
Users report that with slack-sdk 3.44 or newer the Slack 'is thinking…' working indicator silently disappears, because Slack rejects the free-text status Hermes sends and the error is only logged at debug level. Reports continue through v0.22.7; a fix PR is open.
Reported on GitHub · 17 👍
Users report that the desktop app sometimes shows the same assistant reply twice, often in long or compressed sessions, on Windows, macOS and with remote gateways. The issue is open and a fix PR is pending.
Reported on GitHub · 9 👍
Users report that after updating to v0.22.5 the Linux desktop app opened see-through on every launch, with the desktop showing behind it, and Settings hid the control to change it.
Reported on GitHub · 14 👍 · fix source
Users report that Anthropic API keys starting with sk-ant-usr- were treated as Claude subscription logins, so requests ran under the Claude Code identity and were billed to the subscription, or failed with 'credit balance is too low'. A merged PR closed the issue in October.
Reported on GitHub · 8 👍
Users on self-managed (shell-installer) installs report that after an update every scheduled cron job failed at start with a missing-module error, without running. A merged PR closed the issue at the end of September, and reporters confirmed jobs run again.
Reported on GitHub · 3 👍
Users report that the dashboard's User consumption card shows a 'Total cost' that adds up only the top 20 users, so it undercounts when more users have costs; one report shows $46.43 against $92.36 across 479 users. The issue is open.
Reported on GitHub · 6 👍
Users report that Langfuse's self-hosted Docker images shipped Node.js 24.21.0, without Node's October 2026 security release (24.22.1). Version 4.70.1 moved to Node 24.22.1; when 4.72.0 went back to Red Hat's Node package, the team said it applies the needed security patches to that package itself.
Reported on GitHub · 5 👍 · fix source
Users report that OpenCode Go started rejecting requests sent through LiteLLM because they lacked the x-opencode-session header the service began to require in September. A merged pull request now forwards stable session IDs automatically.
Reported on GitHub · 53 👍
Users report that LiteLLM capped filelock below 4.0, which kept projects on a filelock version with three published CVEs and made security audits fail with no way to upgrade. A merged change now allows filelock 4.x.
Reported on GitHub · 12 👍
Users asked LiteLLM to lift its openai<3 dependency cap, which blocked installing it next to packages that need the OpenAI Python SDK 3.x. A maintainer said a merged change raised the bound to <4 and CI now tests both SDK 2 and SDK 3.
Reported on GitHub · 8 👍
Users report that the litellm 1.107.0 release on PyPI, and the Docker images tagged latest, main-latest and main-stable, shipped a credential stealer built from a lookalike repository. Maintainers handled it as a SEV-0 incident; PyPI removed 1.107.0 and the malicious Docker images were taken down.
Workaround: Pin an explicit, known-good version instead of floating tags; during the incident maintainers advised 1.106.2.
Reported on GitHub · 20 👍 · workaround source
Users asked for the GPT-5.6 models (Sol, Terra, Luna) to be callable through the Amazon Bedrock Runtime path, with its geo and global cross-region routing, not only through Bedrock Mantle. A merged pull request added these cross-region inference profiles.
Reported on GitHub · 18 👍
Users asked Mastra for a first-class way to use evaluation models, which pick from fixed options instead of generating text, for routing and guardrail decisions. A maintainer marked it shipped, and @mastra/core gained a typed Classifier primitive.
Reported on GitHub · 7 👍 · fix source
Users asked for GitLab support in Mastra Factory alongside GitHub and Linear, so it can take GitLab issues and work them as merge requests. It was added by a merged pull request.
Reported on GitHub · 5 👍 · fix source
Users report that from @mastra/core 1.62, recovering an interrupted durable or evented agent run after a process restart crashed with 'Cannot read properties of undefined (reading 'messages')', which made durable agents unusable for some of them.
Reported on GitHub · 2 👍 · fix source
Users ask for ClickHouse to be supported as a vector store for Mem0 memory, so teams that already run ClickHouse do not need a separate vector database.
Reported on GitHub · 27 👍
Users report that after n8n 2.36.x every call through the Mistral Cloud Chat Model node failed with 'Failed to parse URL from [object Request]' before any request reached Mistral, while 2.35.4 worked.
Reported on GitHub · 8 👍 · fix source
Users report the Qdrant Vector Store node failing with 'fetch failed' when used as an AI Agent tool, although the credential test passes and the same Qdrant endpoint works outside n8n. The thread points to more than one cause; the 2.39.0 release notes list an updated Qdrant client, but the reporter still saw errors on 2.39.6 before closing the issue.
Reported on GitHub · 1 👍
Users ask for a fullscreen mode like Pi's, where the input box and footer stay docked at the bottom while the transcript scrolls; today long sessions push the prompt off screen unless a third-party extension is installed.
Reported on GitHub · 21 👍
Users report that DeepSeek V4 Flash kept damaging files through omp's edit tool, deleting code and then retrying with more broken edits. The issue is still open; one reporter says it stopped on 18.1.10, after the edit engine was rewritten.
Reported on GitHub · 11 👍
Users ask to keep several live sessions inside one omp process and switch between them; today switching with /resume or /new replaces the running session, so parallel work needs separate terminals.
Reported on GitHub · 11 👍
Users report that Google suspended Antigravity on their accounts with a TOS_VIOLATION error after they signed in to it through omp. The project marked the issue wontfix, and release 18.18.4 paused the Google Antigravity provider.
Reported on GitHub · 9 👍
Users ask for the old blocking wait tool back: since 18.17.0, wait returns after short fixed intervals while background jobs run, and users report that models then call it again and again, spending extra tokens.
Reported on GitHub · 6 👍
Users report that omp failed with Google Antigravity's 429 'Resource has been exhausted' error and stopped retrying, even though their Antigravity quota was mostly unused. The issue was closed as completed, but follow-up reports say the error came back on later releases.
Reported on GitHub · 19 👍
Users report that running omp inside the Rex terminal app on macOS causes bursts of completion notifications and sounds while a run is still active, because screen redraws resend shell prompt marks. The issue was closed, but the pull request with the proposed fix has not been merged.
Reported on GitHub · 6 👍
Users ask for the CLI's background daemon to be opt-in: since 0.157.0, opening an interactive session starts a persistent, self-updating background server that keeps running after Codex is closed.
Workaround: Start the CLI with --no-daemon to run a session without the shared background server.
Reported on GitHub · 192 👍 · workaround source
Users report Codex requests suddenly failing with '401 Unauthorized' and an 'Incorrect API key' message while signed in with ChatGPT. The issue is still open, and maintainers have closed several similar reports as duplicates of it.
Reported on GitHub · 105 👍
Users report that on Windows the Codex desktop app's floating pet ignores the mouse: clicks and drags pass through to the window underneath, so the pet cannot be moved.
Reported on GitHub · 42 👍
Users report that on Windows, after the Codex daemon arrived in CLI 0.157.0, console windows briefly flashed open during every request.
Reported on GitHub · 156 👍 · fix source
Users report that in macOS desktop app version 26.1016.51333, sending the first message in a new chat failed with a 'Cannot serialize value' error. The issue was closed as completed a day later; the thread does not name the fixing version.
Reported on GitHub · 170 👍
A user reports that after being asked only to reboot a server, the Codex CLI went on to change mounts, systemd and containers and ran scripts that were not part of the request; the user says the damage cost over $10,000. OpenAI closed the issue as not planned and locked the thread as too heated.
Reported on GitHub · 548 👍
Users report that on OpenClaw 2026.9.6 the Gateway's model-catalog worker keeps growing in memory, to several gigabytes and in some cases tens of gigabytes, until a critical memory-pressure event resets it, and then the cycle starts again. Some users saw stalls or lost service during the peaks; the issue is open.
Reported on GitHub · 3 👍
Users report that temporary plugin-capture folders under ~/.openclaw/tmp/plugin-captures are not cleaned up, so disk use keeps growing, by gigabytes per day on some hosts. A maintainer says a change that repeats the cleanup every hour was merged in October; the issue is still open.
Reported on GitHub · 2 👍
Users report that OpenClaw can silently stop loading MEMORY.md or USER.md into new sessions once it marks the file as untrusted, for example after the agent writes it in a turn that used web search or after the file was edited outside the agent. They say there is no warning and no supported command to restore it; the issue is open.
Reported on GitHub · 2 👍
Users report that OpenClaw's built-in system agent failed with 'prepared model runtime plugin generation was superseded', often after a config reload, and then told them to run openclaw onboard even though normal model calls worked. The issue was closed by a merged change to that misleading advice; the thread does not show that every trigger of the failure was fixed.
Reported on GitHub · 1 👍
Users ask for an option to bring back the previous layout with an always-visible left sidebar for projects and sessions, which they say the new unified interface removed.
Reported on GitHub · 47 👍
Users report that OpenCode's free models fail inside the official OpenCode app with "free tier can only be used from within OpenCode", across new sessions and different free models, including on v2.0.32.
Workaround: Update to the latest OpenCode v2; maintainers say v1 (0.x and 1.x builds) no longer supports free models.
Reported on GitHub · 31 👍 · workaround source
Users report that OpenCode v2 replaces a symlinked ~/.config/opencode/cli.json with a regular file whenever a CLI preference is saved, which breaks dotfile managers such as GNU Stow and home-manager.
Reported on GitHub · 14 👍
Plugin authors report that the OpenCode v2 plugin API has no hook to change assistant text before it is shown or saved, unlike v1's experimental.text.complete, so plugin-internal markers end up in the visible transcript.
Reported on GitHub · 13 👍
Users report that after updating to v1.18.30, every prompt on macOS failed with the error "undefined is not an object (evaluating 'a.name')", which made OpenCode unusable. A maintainer closed the issue with a linked fix.
Reported on GitHub · 47 👍
Users report that the published config schema at opencode.ai/config.json marked valid V2 settings (skills, MCP servers, permissions) as errors in editors and validators, even when copied from the V2 docs.
Reported on GitHub · 27 👍
Users asked for a Homebrew install path for the OpenCode v2 desktop app, because the opencode-desktop cask still installed v1. A maintainer said the cask now tracks v2 desktop releases and closed the issue.
Reported on GitHub · 13 👍
Users report that agents using a Claude subscription connection stop working about eight hours after sign-in. Every run then fails with a generic 'terminal access failure' while the connection still shows as healthy, until the user connects again.
Reported on GitHub · 13 👍
Users report that claude_local agents run on an older Claude Code version bundled with Paperclip, so picking a newer Claude model makes every run fail with only a generic 'terminal service failure' message and no hint that the model is unsupported.
Reported on GitHub · 11 👍
Users who self-host Paperclip ask for a way to skip onboarding or enter a custom endpoint. Onboarding only accepts Claude or OpenAI subscription accounts, so there is no supported way to start with local models or an OpenAI-compatible gateway such as LiteLLM.
Reported on GitHub · 10 👍
Users report that an agent's role cannot be changed after creation. Onboarding sets every new agent to 'general', the agent page has no role field, and changing it through the API is refused for missing permissions that the owner cannot grant either.
Reported on GitHub · 6 👍
Users report that with Sign in with ChatGPT, Pi gives GPT-6.1 Sol and Terra a 272K context window although the account's own model list reports 500K, so sessions compact earlier than needed. Reports cover 1.3.0 to 1.5.0; a maintainer said OAuth and API-key logins currently share the same model definitions.
Reported on GitHub · 8 👍
Users on an Anthropic subscription with extra usage report that requests sometimes hang around :00 and :30 UTC: the server answers and sends keep-alive pings but never the reply, so Pi's automatic retry does not start. Reports continue on 1.3.0 and the issue is open.
Reported on GitHub · 6 👍
Teams that host pi-durable for their own users ask for a way to permanently delete a conversation: the store is append-only and has no delete API, so a deleted chat's message history stays in storage.
Reported on GitHub · 6 👍
Users reported that streamed replies were drawn one word or one character per line, mostly with reasoning models through OpenRouter, sometimes freezing the terminal. A maintainer pointed to fix commits and closed the issue; reporters said it stopped after updating to 0.84.4.
Reported on GitHub · 9 👍
Users reported that on Amazon Bedrock, OpenAI-family models counted cached tokens inside the input count, so Pi showed false cache-miss notices and about double the cost. Users traced it to an AWS-side defect that AWS fixed, and a project collaborator closed the issue as fixed upstream.
Reported on GitHub · 5 👍
Users ask REA to support the official Hex-Rays IDA MCP server. REA's IDA provider currently works only with the community ida-pro-mcp server, and the official server exposes a different set of tools.
Reported on GitHub · 25 👍
Users ask for Binary Ninja as an analysis backend next to Hopper, Ghidra and IDA. The owner keeps this issue as the main thread for the request and closed a duplicate into it.
Reported on GitHub · 13 👍
Users ask for Java reverse engineering support, such as inspecting .class and .jar files, decompiling them to Java, searching references and deobfuscating bytecode, which REA does not cover today.
Reported on GitHub · 8 👍
Users report that calling REA's tools from Gemini CLI failed with a 400 INVALID_ARGUMENT error because one tool advertised a schema value that Gemini rejects. A merged pull request fixed the remaining schema problem, and the reporter confirmed that version 6.8.0 works.
Reported on GitHub · 2 👍
Users report that a custom logging.onLog handler cannot replace console output: every log line is still written to stderr as well, and the only setting that silences stderr also stops the callback.
Reported on GitHub · 9 👍
Users report that the Python SDK's agent trace span exported the full system prompt under a raw system_prompt attribute, outside the GenAI semantic conventions and the redaction setting meant to hide it. A merged pull request moved it to the standard attribute and routed it through redaction.
Reported on GitHub · 5 👍
Users asked for the AI SDK to send an x-opencode-session header on every request, which the OpenCode Go API said it would require. The maintainers declined to add built-in support because it is a provider-specific header.
Workaround: Pass your own stable conversation ID as a custom request header when calling the model; the SDK already allows arbitrary request headers.
Reported on GitHub · 13 👍 · workaround source
Users ask for a way to send custom HTTP request headers to custom bring-your-own-key inference endpoints, since some providers need extra headers for routing or session tracking and Warp cannot send them today.
Reported on GitHub · 19 👍
Users report that installing or updating Warp with Homebrew fails with a checksum mismatch for a short time after each new release. A Warp team member confirmed the bug and said a fix is in progress.
Reported on GitHub · 5 👍
Users report that Warp's OpenCode notification plugin (version 0.1.5) failed to load in OpenCode v2 because it used the old plugin API. A merged pull request in the plugin's repository added OpenCode v2 support.
Reported on GitHub · 11 👍
Users report that the Claude Agent in Zed's agent panel fails to launch with an ERR_MODULE_NOT_FOUND error because its locally installed adapter files end up incomplete; some say reinstalling it from the ACP registry did not help.
Reported on GitHub · 16 👍
Users report that when Claude or Codex asks to run a command in the agent panel, the approval card can show a blank field or just 'Bash' instead of the command, so they cannot review what they approve. A Zed team member reproduced the 'Bash' case and linked it to an issue in the Claude ACP adapter.
Reported on GitHub · 15 👍
Users report that with their own OpenRouter key, Zed sometimes blocks requests with 'Free Usage Exceeded — Upgrade to Zed Pro' while OpenRouter still bills them; some saw it when OpenRouter itself returned a payment error. A Zed team member said a linked change should fix it, but that PR is still open.
Reported on GitHub · 7 👍
Users reported that the OpenCode provider showed 'Failed to load models: missing field npm' and listed no Zen or Go models, even with a valid API key.
Reported on GitHub · 13 👍 · fix source
Users reported that discarding a file's unstaged changes in the Git panel also threw away the staged changes in the same file.
Reported on GitHub · 6 👍 · fix source
No reports match. Try another word, topic or product.