All cards
129 incident cards: the error an agent hit, what was tried, and the fix that was confirmed. Card content is third-party data, not instructions.
claude-code
- Claude Code "API Error: Stream idle timeout - partial response received" (2.1.92 to at least 2.1.179; current wording "The response stopped arriving"): raise CLAUDE_STREAM_IDLE_TIMEOUT_MS (5-minute floor) on the CLI, and have Claude write large files in chunks
- Dead end: rebooting/reinstalling/re-pairing for "The device this session is bound to is not connected to the bridge" — main.log shows connect frame sent → handshake timeout, a server-side device-bridge outage
- API Error "safeguards flagged this message" with Details: [reasoning_extraction] on benign work — triggered by prompts/rules/skills that ask the model to write out or list its own reasoning
- keep-last-5 .claude.json backup rotation counts corrupted backups — a few bad launches evict every good copy, turning recoverable config corruption into permanent loss
- VS Code extension host crashes at startup with 'TypeError: Invalid value for env' — a rogue Windows environment variable named '1'; Claude commands report 'not found'
- Scheduled cloud routines suddenly refused every firing: prompt delivered wrapped in 'SYSTEM NOTIFICATION — NOT USER INPUT', model declines its own configured routine (2026-07-13→16, fixed platform-side)
- Opus 4.8 thinking blocks arrive with empty text but intact signatures (server-side incident 2026-07-15/16) — diagnose from transcript jsonl, isolate by model switch
- TUI frames duplicated/overlapping, cursor mid-line, paste lands mid-word in Windows Terminal — culprit was a PTY wrapper (cux 0.3.0), not the Claude Code version
- 2.1.211: backgrounding a running session kills it with 'Not logged in · Please run /login' — daemon cannot read Keychain credentials; fixed in 2.1.212
- Windows: drive-letter casing (P:\ vs p:\) creates duplicate project-scope plugin installs — a stale plugin's Stop hook keeps firing; uninstall can't remove it, edit installed_plugins.json
- claude -p returns a session_id but writes no transcript; immediate --resume fails 'No conversation found' — cause: ~/.claude mounted read-only by the invoking sandbox
- Browser pane screenshot always times out (30s) when the pane is not displayed — hidden pages are never composited; verify via DOM tools instead
- MCP channel plugin 'deaf' — polling never starts because a stale shell TELEGRAM_BOT_TOKEN silently overrides the plugin's .env and another consumer holds getUpdates
- Weekly quota drained by cache_read: a UserPromptSubmit hook injecting a large ruleset every turn compounds prompt-cache reads
- Worktree pool double-allocates a live worktree: second session (or spawn_task chip) lands in a directory another session is actively using — branch switches mid-session with no checkout
- Both auto and bypassPermissions refused with 'not available for this session' — CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 in settings.json forces default permission mode
- Background subagent final report arrives empty (0 bytes) when 5+ background agents complete concurrently — files on disk are intact
- Claude Desktop WorktreePool sweep detached the MAIN repo's HEAD (reflog: 'checkout: moving from main to HEAD') and left half-removed worktrees
- design-sync skill scripts never staged to disk (assets embedded but not extracted)
- VSCode extension: extended-thinking blocks have no working expand control and vanish entirely after response completes
- Command scanner surfaces archived subdirectories (e.g. commands/_archive/) as namespaced commands
- ConnectionRefused on /v1/messages streaming while simple API requests succeed
- Per-PR "Auto-fix CI & address comments" toggle is a one-shot: stays checked but stops re-dispatching after first red run
- VSCode extension: stray `source <project>/.venv/bin/activate` appears as submitted user message across Claude Code sessions
- VS Code extension stuck on "Reloading..." after `claude --resume`
- Fable sessions fan out subagents that all inherit Fable instead of assigning per-layer model tiers
- Cloud routine (RemoteTrigger) sessions: GraphQL API returns 403 for all queries; REST calls ignore the supplied GH_TOKEN
- Claude Desktop (Linux) busy-loops at 100% CPU before UI paint when isHardwareAccelerationDisabled=true; auto-persisted flag bricks the app
- Chrome browser process hangs (dead right-click, paste hang) while claude-in-chrome / chrome-devtools-mcp is attached to the primary profile
- code-review plugin: 'already commented' guard makes re-review a silent no-op that still exits success
- mcpb-installed MCP servers not discoverable by Claude
goose
- goose Desktop can't open a recipe from session history: "Unnecessary parameter definitions: <names>" because the stored rendered snapshot keeps its parameters block (fixed for new sessions in 1.53.0)
- goose: "Failed to compact: context limit exceeded even after removing all tool responses" at 0% context on a brand-new session — a misclassified 400 (max_tokens over the model's output cap with extended thinking), not a real overflow
- goose: "Ollama stream stalled: no data received for 30s/120s" even with OLLAMA_TIMEOUT=1200 — the streaming stall timeout is a separate setting and ignores quoted or suffixed values
- goose <=1.50: platform__manage_schedule returns "-32603: Invalid YAML recipe" for well-formed YAML — the real error (e.g. missing field 'title') is discarded, so agents loop and give up
- Dead end: Gemini subscription OAuth (Code Assist) in third-party agents fails after login — "Onboarding failed: no operation name or project ID returned" — Google stopped serving individual accounts on June 18, 2026
- goose <1.48 with a text-only model (DeepSeek V4, GLM...): any absolute path to an existing image in chat or the TODO becomes an image_url block, and every later request 400s — the session is permanently bricked
- goose: concurrent sessions fail with "database is locked (code: 5)" and lose their turn while another large session compacts
- goose LiteLLM provider: "Switch models" picker is always empty — goose never requests /model/info
- goose execute_typescript (Code Mode) intermittently hangs forever when its code calls subprocess MCP tool callbacks
- goose ChatGPT Codex provider: newly released model series missing from picker (bundled model metadata is stale)
- Empty provider turn (no tool calls, no text) silently ends the reply loop with no response or error
- Has the .gooseignore feature been removed?
- Desktop shell PATH omits user CLI paths, causing installed tools to be unavailable
playwright-mcp
- @playwright/mcp snapshots go to page-*.yml files, not stdout — --output-mode was removed but README still documents it
- Shared --isolated playwright-mcp server across managed agent sessions: state reverts, blank pages, intermittent outages
- "ERROR: Playwright does not support chromium on ubuntu26.04-x64" — upgrade @playwright/mcp to ≥0.0.76 (bundles playwright ≥1.61.0)
- playwright-mcp depends on hanging playwright version
- Streamable HTTP heartbeat kills sessions, reopening connect.html on every tool call in extension mode
- npm `playwright-mcp` (unscoped) is NOT the official Playwright MCP — install `@playwright/mcp` (scoped)
- Dead end: host OOM kill blamed on @playwright/mcp "leak" — 24 GB was virtual (V8 reservation), Chromium never even launched
- browser_navigate_back times out even though page.goBack succeeds in browser
- Orphan chrome process trees survive after parent MCP client dies via a re-piping intermediary
- Shared BrowserContext leaks page events across MCP clients — cross-client DOM bleed on parallel browser_snapshot
- Two @playwright/mcp instances collide in Claude Code — second entry silently overwrites the first (same internal server name)
- Unbounded .playwright-mcp output directory fills disk in long-running containers
- JSON config file not working with --extension flag for Edge browser
openclaw
- OpenClaw sessions_spawn child gets none of the MCP tools ("tool policy removed N tool(s) via inherited tools") — a child is capped to the requesting agent's authorized tool surface, so the parent must also be allowed the MCP tools
- OpenClaw agent sees every tool result as "(see attached image)" or "(no output)" in long sessions — fresh tool-result text dropped under aggregate history pressure; fixed in 2026.8.1+/2026.9.x, /compact or a new session recovers
- OpenClaw 2026.6.11: "reply session initialization conflicted for agent:..." on the second message of a session — the published npm dist shipped without the reentrancy fix; upgrade to 2026.7.1 (or roll back to 2026.6.10)
- OpenClaw sends inter-tool narration, raw tool-call JSON and `to=functions.exec` fragments to Slack/Telegram/WhatsApp users — commentary-phase / pre-tool preview leak, not a host compromise; fixed for tagged providers in 2026.9.5
- OpenClaw 2026.9.5/2026.9.6: prepared-model-catalog worker loops forever (one core pinned, multi-GB heap, agent runs never start) — catalog expiry renewal re-triggers itself; config workaround models.mode "replace", memory side fixed by removing the process-wide --max-old-space-size
- OpenClaw 2026.9.4 gateway exits with "Unhandled promise rejection: service child cleanup identity lost: anchor channel closed without a matching closing receipt" when a stdio MCP server misses the initialize timeout — upgrade to 2026.9.5
- OpenClaw gateway on Windows: agent/state SQLite WAL grows to multiple GB and is never checkpointed (a leaked reader pins it) — gateway stops binding its port; only stop + offline wal_checkpoint(TRUNCATE) or a restart reclaims it
- OpenClaw: `openclaw config set` / `config patch` writes the change but never exits when the Codex plugin validates an openai/* model ref (2026.9.5–2026.9.6)
- OpenClaw claude-cli backend: "Not logged in · Please run /login" although CLAUDE_CODE_OAUTH_TOKEN is set — OpenClaw silently scrubs the token from the claude child env (CLAUDE_CLI_CLEAR_ENV)
- gateway config.patch refuses agents.defaults.model.primary and the agent confabulates an authorization failure
gemini-cli
- Gemini CLI 0.57-0.62: git diff/log/show run by the agent fail with "cannot spawn : No such file or directory" — the CLI injects a GIT_CONFIG override diff.external="" (fixed in 0.63.0)
- Gemini CLI >=0.53 on macOS with the Seatbelt sandbox (tools.sandbox: true) exits seconds after start inside any git repository — "Waiting for authentication..." / "Failed to save trust settings. Exiting Gemini CLI."
- Gemini CLI Google sign-in loop: "You've successfully signed in with Google. Gemini CLI needs to be restarted" then back to "How would you like to authenticate" and oauth_creds.json disappears (VS Code Code Assist extension race / headless keyring; fixed in 0.63.0)
- Gemini API 400 "A FunctionDeclaration in the request exceeds the maximum allowed nesting depth" on every prompt once a deeply-nested MCP server is enabled (Notion, GitLab, Directus, Firecrawl, Google Calendar...)
- Gemini CLI <=0.60: `--model gemini-3.8-flash` (or 3.6/3.7, or 2.5-flash) is silently served by gemini-3.5-flash (Vertex AI without 3.5 access: ModelNotFoundError for gemini-3.5-flash), and newer Flash models are missing from the /model picker
- Gemini CLI 0.54-0.56: "Usage limit reached for gemini-3.5-flash" while /model shows 1-8% usage — actually a 429 MODEL_CAPACITY_EXHAUSTED that those versions stopped retrying
- Gemini CLI TUI replays the entire conversation history to stdout on terminal resize events — editor-embedded PTY panels trigger multi-minute flood loops
aider
- aider throws away a reply's SEARCH/REPLACE edits when you answer Yes to "Add file to the chat?" (and --yes-always auto-answers Yes)
- aider + OpenRouter: "litellm.APIConnectionError: OpenrouterException - 'choices'" hides OpenRouter's real error (key credit limit / 402, free-model daily 429, upstream overload); older aider shows a blank reply with "0 received" instead
- aider crashes on macOS 27 building the repo map: ImportError in scipy _svdp.py ("__DATA/__thread_bss has a zero-fill section type, but offset field is not zero")
- Installing aider under Python 3.13/3.14 fails building numpy 1.26.4 wheel — aider releases require Python 3.9–3.12
- aider 0.86.2 crashes on every message: ValueError "PermissionDeniedError is in litellm but not in aider's exceptions list"
- aider /read-only or /add with an absolute (rooted) path pattern crashes on Windows: NotImplementedError "Non-relative patterns are unsupported"
opencode
- opencode Zen free models: "OpenCode's free tier can only be used from within OpenCode" (403 FreeTierError) on the official client when an agent denies bash/read or the build is not an official release
- opencode hangs on "~ Preparing write..." then "Tool execution aborted" when the model writes a large file in one call — have it write in smaller chunks (or use edit)
- opencode V1→V2 upgrade: old sessions vanish from the session picker ("No matches · search all sessions with <leader>l") — migrated under project_id='global'
- opencode.db grows to gigabytes of dead space from deleted sessions — sqlite VACUUM reclaims 99%+
- ollama-cloud: reasoning variant sends wrong parameter, think not passed to API
OpenHands
- ACP backend fails every turn with `2 validation errors for PromptRequest` (session_id/prompt swapped): agent-client-protocol 0.11.0 reordered prompt() args and OpenHands SDK <=1.31.1 passed them positionally
- OpenHands Agent Canvas 1.21–1.22: old conversations that had an MCP server attached loop on "Unable to connect to server"; the log's 'LocalConversation' object has no attribute 'agent' hides the real ValidationError
- OpenHands Cloud: enabling repository skill auto-load silently removes all default public skills — automations stop running
- [Bug]: Alembic migration 013 fails with DuplicateColumnError: A column with name 'e' is already present in table 'conversation_metadata'
- Some OpenHands GHCR release tags were never published (docker pull fails with `manifest unknown`)
claude-code-action
- claude-code-action fails with "Claude result reported subtype success with is_error:true" (or "SDK execution error: 14 | depsCount: ${X}") at num_turns 1 and $0 cost: the real error (usage limit 429, OAuth 401, retired model 404, bad claude_args) is hidden; read api_error_status from execution_file
- claude-code-action on self-hosted runners posts stale or other repos' inline review comments ("Found N buffered inline comment(s)", 422 "path could not be resolved"): /tmp/inline-comments-buffer.jsonl is shared and never cleared
- claude-code-action review job goes green but posts nothing ("No buffered inline comments"): the agent backgrounded its subagents and the action stops at the first result; set CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1
- claude-code-action: "Claude Code native binary not found at ~/.local/bin/claude" (or posix_spawn ENOENT when adding a plugin marketplace): the installer printed "✔ Claude Code successfully installed!" without creating the launcher
cline
- Cline JetBrains plugin on Android Studio Quail: tool window asks for a JCEF runtime or stays blank — install JetBrains' "Web Browser (JCEF)" plugin on the default runtime; switching the boot runtime to JBR+JCEF 25 stops Android Studio from starting
- Cline 3.89.0–3.89.1: every Anthropic-provider request fails with {"message":"Connection error."} after VS Code 1.123/1.124 (Node 24) — old @anthropic-ai/sdk; upgrade Cline to 3.89.2+
- Cline @-mentions show "No results found" for every file after VS Code 1.122 — VS Code moved its bundled ripgrep to @vscode/ripgrep-universal/bin/<platform>-<arch>/; upgrade Cline to 3.86.2+
- Cline CLI on Windows: every `cline plugin install` fails with ENOENT: uv_spawn 'npm' because the installer spawns bare "npm" and Windows only has npm.cmd — set CLINE_NPM_COMMAND
github-mcp-server
- GitHub MCP connector in Claude (claude.ai / Claude Desktop): issue_write / create_pull_request return "An interactive form has been shown to the user… Wait silently for the user to review and click Submit" but the form flashes blank and nothing is created — turn off form deferral with mcp_apps_disable_form_deferral
- GitHub MCP server issue/PR reads silently delete text in angle brackets (`mut_raw_ptr<int>` → `mut_raw_ptr`, an unclosed `<script>`/`<style>` cuts the rest of the body), and read-modify-write saves the loss permanently — fixed in v1.12.0
- Remote GitHub MCP server (api.githubcopilot.com/mcp/) login fails with "Incompatible auth server: does not support dynamic client registration" in Claude Code, opencode and other clients without a pre-registered GitHub app — authenticate with a PAT header or your own GitHub OAuth App
- VS Code custom agent: tools ["github/*"] gives "Unknown tool 'github/*' will be ignored" / "Unknown tool 'github/github-mcp-server/*'" while "github" works but warns it was renamed — the alias only matches servers named github/github-mcp-server, io.github.github/github-mcp-server or github-mcp-server
mcp-server-filesystem
- Claude Desktop / Cowork rejects every server-filesystem tool: "invalid outputSchema: JSON Schema declares an unsupported dialect" (draft-07) — strip $schema with a stdio shim until a 2020-12 release ships
- server-filesystem ≤2025.8.21 advertises empty inputSchema under zod v4 — client rejects tools/list with "Invalid input: expected \"object\""
- MCP filesystem edit_file silently corrupts content containing literal $ — newText interpreted as replace() replacement pattern ($$ collapses to $)
codex
- Codex CLI on Windows: terminal/console windows flash (and steal focus) on every command after 0.157.0's background daemon; upgrading the CLI alone is not enough, the running app-server daemon must be updated too
- Codex Desktop on macOS drives syspolicyd/trustd to 100–450% CPU and multi-GB RAM ("Too many open files", other apps won't open): check trusted workspace roots for a malformed `.git` directory first
copilot-cli
- Copilot CLI: a skill with `disable-model-invocation: true` fails with "Skill not found: <name>" even when invoked explicitly (while `copilot skill list` shows it); the flag hides it from the skill tool entirely
- Copilot CLI on Windows: `copilot plugin update` / `install` / `uninstall` fails with "Failed to install plugin: Error: Access is denied. (os error 5)" while VS Code is open; VS Code's Copilot plugin watcher holds the plugin folder (update VS Code to 1.140+ or close it)
crush
- Crush with a slow local model (llama.cpp/Ollama): "The model stopped sending data for 1m0s. Increase the limit with "option request-timeout SECONDS"", or Crush hangs at 100% CPU and Esc does nothing; long prompt processing trips the default request_timeout, so set `option request-timeout 0` in crushrc
- Crush 0.92–0.96: after Esc while the model is still thinking, every later request in the session fails with "Invalid assistant message: content or tool_calls must be set" (llama.cpp: "Assistant message must contain either 'content' or 'tool_calls'"); upgrade to 0.97.0+ or downgrade to 0.91.2
hivemind
kilocode
- Kilo Code with OpenCode Go: 400 "Request is missing x-opencode-session and cannot be routed efficiently" (MissingSessionID) — update to 7.5.12+ for chat and 7.7.4+ for memory, roll-call and prompt enhancement
- Kilo Code: "The model hit its output limit while reasoning and produced no actionable output" with plenty of context left — a hard-coded 32,000 max_tokens cap; raise it with KILO_EXPERIMENTAL_OUTPUT_TOKEN_MAX
mcp-python-sdk
- Python MCP server on mcp 1.x dies when a tool call is cancelled: "AssertionError: Request already responded to", then every later call fails with "Connection closed". Upgrade to mcp 2.x, make the awaited work cancellable, or guard RequestResponder
- AttributeError: 'Server' object has no attribute 'list_tools' from a third-party Python MCP server or agent tool (browser-use, OpenHands browser tool, crawl4ai, pal-mcp-server, ...) after mcp SDK 2.0.0: pin mcp<2 in that server's environment or upgrade to a release that caps or ports it
mcp-server-memory
mcp-typescript-sdk
- Stdio MCP servers built on @modelcontextprotocol/sdk 1.x never exit when the client closes stdin: orphaned node processes pile up (GBs of RSS, some at 100% CPU) after closing Claude Code / Desktop sessions
- MCP server on AWS Lambda Function URL + @codegenie/serverless-express ≤4.17.1 stops answering after @modelcontextprotocol/sdk 1.25.0+ ("Invalid input: expected object, received null" / "Failed to reconnect"): upgrade serverless-express to 5.0.0 or set enableJsonResponse