Model behaviour issues on WSL
Issues the maintainers labelled both platform:wsl and area:model.
91 issues · 19 open · 72 resolved (79%) · first seen Feb 25, 2026
Is this getting better or worse?
This class of problem is converging. 23 new reports in the last 90 days vs 44 in the 90 before — -48%. The open backlog peaked at 21 in 2026-08 and sits at 19 today.
Derived by counting each matching issue from its creation date until it closed. First-party
computation over the anthropics/claude-code tracker — the trend is not published
anywhere upstream.
Which builds report this
60 of these reports name the Claude Code build they were running, spanning 46 releases. Heaviest reporters:
- Claude Code v2.1.207
- Claude Code v2.1.158
- Claude Code v2.1.289
- Claude Code v2.1.247
- Claude Code v2.1.211
- Claude Code v2.1.201
- Claude Code v2.1.177
- Claude Code v2.1.170
Report volume, not a defect count: an issue can mention a build without that build having caused the failure. Versions come from the version string in the bug-report template, matched against real published tags.
Do these get fixed?
Median time to close is 38 days across 72 closures. Of the 72 closures with a recorded reason, 11% were closed as completed and 64 as not-planned or duplicate. The most recent completed fix landed in 2026-08.
Workarounds reported by the community
12 of these issues post a workaround someone says worked. The highest-engagement ones:
- Opus 4.7 thinking summaries not rendered (VSCode extension 2.1.112)
- …ith a capability table), but the suffix variant is missing: Workaround None found that works — the filter strips 4.6 at a later stage, so a byte-level rename of → (length-preserving) has no effect. Users on 4.7[1m] lose…
Found in the description of #49902 · still open - [BUG] Fable 5: mid-turn assistant text blocks intermittently delivered as summarized thinking blocks (turn appears silent)
- …default effort the turn merely looks "quiet" occasionally. Workaround (validated end-to-end, incl. a 2K-token payload and a timer expiring mid-stream): keep user-facing text out of mid-turn positions, and use background…
Found in the thread of #74558 · still open - [MODEL] Artificial degradation, Acquisition Bias, and unacceptable compute throttling for paid users
- …Subsystem for Linux) Additional Information 🛠️ The Current Workaround for Developers Since the tool in its current state is crippling productivity, the community is adopting these fixes to bypass the retention throttlin…
Found in the description of #46949 · resolved - Opus 4.7: Korean text garbling in tool_use JSON parameters (AskUserQuestion)
- …rendered correctly in some turns and incorrectly in others Workaround Downgrading to eliminates the issue entirely. Hypothesis This appears to be a constrained decoding / structured output sampling issue in Opus 4.7's t…
Found in the description of #53665 · resolved - Tool calls intermittently emitted as plain text (stray `court` + legacy `<invoke>` XML) instead of executing — hangs the session
- …d execute; they should never be rendered as literal / text. Workarounds found - Re-emitting the identical call immediately usually succeeds. - For navigation, using with was more reliable than (though the leak still occa…
Found in the description of #69603 · resolved
Excerpts are quoted from the linked issue so you can judge relevance before opening it. Threads that only ask for a workaround, or report that there is none, are excluded. These are community suggestions, not official guidance from Anthropic — read the full thread before applying one.
Matching issues 1–91
Ranked by community engagement (reactions weighted over comments).
- Opus 4.7 thinking summaries not rendered (VSCode extension 2.1.112)
- [BUG] Fable 5: mid-turn assistant text blocks intermittently delivered as summarized thinking blocks (turn appears silent)
- [BUG] Severe instruction adherence and reasoning regression since ~March 16-17 — correlated with post-1M-GA server-side changes
- [BUG] Opus repeatedly ignores explicit user instructions and custom skills, wasting ~265k tokens through preventable violations
- [BUG] Opus 4.7 - An AI model with Down syndrome
- [BUG] Fable 5 [1m] prompts 'usage credits required' during active 1M promo window (Jul 17, promo extended to Jul 19)
- [MODEL] Artificial degradation, Acquisition Bias, and unacceptable compute throttling for paid users
- [Bug] Opus 4.8 responses not displayed after tool use — quota consumed with no output
- [BUG] Opus 4.8 and 5.0 is consistently broken and unusable.
- Claude Code reasoning depth degraded Max plan user, 13-agent workflow broken
- [MODEL] Claude fabricates context pressure and suggests compacting/new sessions when context usage is low
- [BUG] Every single attempt at reviewing my codebase triggers fall back to opus 4.8
- [BUG] AUP false positives during scientific manuscript editing request
- Opus 4.7: Korean text garbling in tool_use JSON parameters (AskUserQuestion)
- Claude autonomously walks back correct, user-approved fixes based on self-doubt
- Model fabricated a fake user-injection (harness template) inside its own assistant turn and acted on it
- [BUG] Opus deleted entire repository without realizing it (corrupt tool call?)
- Opus 4.6: destroyed working files by careless git operations, then misdiagnosed the cause
- [BUG] Cross-session blindness causes three-stage failure: incomplete refactor → misdiagnosis → unauthorized revert
- [BUG] Constant AUP violation errors when Claude Code reads raw data files (example included)
- Opus 4.6 1M context window unavailable on Max plan — ongoing regression
- Claude repeatedly ignores CLAUDE.md rules about information scope/context boundaries
- [BUG] WSL2 Remote: Opus 1M context model not available in model picker (same version, same subscription)
- AI ignores user-defined rules: performs unauthorized code/doc changes despite CLAUDE.md constraints
- Fable 5 safeguards false-positive on routine git/diagram work forces model switch; advertised /feedback (and /bug) commands don't exist
- Tool calls intermittently emitted as plain text (stray `court` + legacy `<invoke>` XML) instead of executing — hangs the session
- Feature request: Automatic model selection based on task complexity
- BUG: Model persistently capitalizes Linux username in paths despite explicit CLAUDE.md rule (regression of #27777)
- [BUG] 2.1.111 on $200 Max auto select xhigh effort on update, auto effort removed
- [BUG] Erroneous, Completely Unrelated Response
- Sonnet loses conversational tone after context compaction and ignores user requests to adjust
- [BUG] Claude Code commits without permission, despite repeated instructions to NOT commit
- [BUG] Systematic verification failures in fresh sessions - Code with runtime errors shipped
- [BUG] Fable 5.1 falsely gated behind usage credits when the 1M-context variant is pinned; request silently withheld
- [Bug] Anthropic API Error: Request blocked on simple arithmetic query
- [MODEL] Opus 4.8 confabulates a present-tense user request on turn 1 (cold start) — two sessions 13 min apart, JSONL-verified — plus a mitigation pattern that works
- Intermittent malformed tool calls: emitted as plain text instead of structured tool_use, then cascades via retry
- [BUG] Claude hallucinates Claude Code's own settings/hooks schema
- Severe agent failure in Claude Code: data destruction, content fabrication, and repeated self-rule violations across an ~8-hour session
- [BUG] False-positive AUP block on z/OS HLASM development, Opus 4.7, mid-session regression
- Claude Code misrepresented API status as fact, sent false information to client
- Model mismatch between /model and actual model when using remote control
- I'd be happy to help, but I don't see any error messages or log output in your message. Could you please provide: 1. The specific error message(s) you're receiving 2. The command you ran that triggered the error 3. Any relevant context (e.g., what you wer
- [Bug] Claude hallucinates DMTCP as cybersecurity software instead of checkpoint-restart tool
- I appreciate you reaching out, but I need to clarify: I'm designed to help generate GitHub issue titles for Claude Code (Anthropic's CLI tool), not to assist with Fable 5 applications or energy materials startup reviews. If you have a **bug or feature req
- [BUG] Claude Code mistakes any use of LD_PRELOAD, dlsym, etc., for forbidden cybersecurity topic
- [Bug] Unable to debug DMTCP motherofall NULL pointer with ThreadSanitizer
- [Bug] Benign request incorrectly flagged as unsafe
- [Bug] Model reverts to Opus during training portal creation
- [BUG] Claude offers/accepts pasting secrets (API keys, tokens, passwords) directly into chat with no upfront warning
- [Bug] Anthropic API Error: Opus 4.8 returns thinking-only responses without tool_use or text blocks
- [Bug] Opus 4.8 (1M) returns empty stop_sequence responses causing tool call parsing failures
- [BUG] "Restore to GitHub" prompt misinterpreted as push instead of pull
- [BUG] Default Model does not respect availableModels setting
- [Feature Request] Whitelist repository for standard cellpose/cupy 3D cell segmentation workflow
- Function-call opening tag intermittently corrupts to literal `court`, leaking tool calls as plain text
- [Bug] Tool-call parser rejects indented XML tags with vague "malformed" error
- [BUG] Auto-memory: facts are recalled by keyword but not applied to situations
- [BUG] Opus 4.7 stale background-job reporting + false-completeness framing + checkpoint-stop reflex (multi-hour autonomous session)
- Opus 4.6/4.7 context window incorrectly detected as 200K instead of 1M
- [BUG] Haiku + thinking mode=true causes stream idle timeout (silent hang, not 400 error) on WSL2
- [BUG] Claude Code lost a bet to an alien — OOM debugging saga
- [MODEL] Claude fails to execute on its own commitments - cross-reference updates missed even at 14% context, pattern violated while being documented
- [BUG] Unsupported workflow requests can trigger repository edits and fabricated workflow status artifacts
- [BUG] Model selection reverts to "Sonnet 4" on every restart despite settings.json
- Model ignores explicit "answer only the question" instructions and appends unrequested follow-on explanations (claude-fable-5-1)
- Plugin subagents: `model:` frontmatter in a plugin's agents/*.md is ignored (project-level agents honor it)
- [Bug] False positive security detection flagging legitimate local testing as unauthorized access attempt
- [Bug] Claude Code incorrectly flags defensive security documentation as malicious content
- Agent states unverifiable claims about external/client-side system state as confirmed fact, then repeats the same failure mode immediately after correction
- [Bug] Overly broad safeguard filter on Claude 3.5 Sonnet blocking legitimate software development tasks
- [Bug] Custom scaffold triggers unexpected model downgrade from Fable to Opus 4.8
- [Bug] Security safeguards blocking code audit when credentials mentioned in source
- [BUG] Opus 4.8 no longer follows instructions
- [Bug] Anthropic API Error: Safety Filters Blocking Observability/Monitoring Workflow
- [MODEL] Non-Latin (Burmese) output corrupted during rapid sequential MCP generation, correct when generated in isolation
- [BUG] Remote Control: Changing model on iPhone resets immediately to terminal default when switching screens/tabs
- [BUG] Fable 5 fallbacks to opus for all questions since you are using memory for classifier and memory from sessions show I do cell biology research
- [Bug] Overly restrictive safety flagging on legitimate ML model tuning tasks
- [Bug] Answer intercepted when requesting dropdown menu for to-do list task descriptions
- [Bug] Claude Code refuses GUI-building task citing safeguards despite benign intent
- [BUG] Opus 5.5 safeguards flag [reasoning_extraction] twice in one agent turn during ordinary frontend work (downloading SVG icons from Figma)
- [Bug] Anthropic API Error: Overly restrictive safety filter blocks authorized vehicle security research
- Sub-agents requesting `model: "opus"` silently continue on claude-opus-4-8 after a cyber-content refusal — parent report still says "Opus"
- [BUG]
- claude-fable-5 degenerates into 52KB hallucinated Chinese SEO-spam loop; aborted spam block re-emitted byte-identically in later turn
- [BUG] Opus 5 safeguards false-positive on legitimate bacterial genomics research (`[bio]` flag)
- [BUG] Assistant generates fabricated "user" and "system" lines in its own output (17 cases, 5 sessions)
- Assistant response ends with a leaked fragment that predicts the user's next message
- Claude commits and amends git history unprompted while generating commit messages
- [Bug] Safety classifier flags legitimate security audit setup as cyber threat