Claude Code users report read:edit regressions, OpenClaw disconnects, and settings.json fixes
HN follow-up added a 6,800-session regression analysis, reproduced OpenClaw-triggered disconnects, and surfaced settings.json mitigations like high effort or disabled adaptive thinking. Track the same path for repo reads, refusals, and subscription-versus-API differences.
TL;DR
- the GitHub regression report claims Claude Code's workflow shifted from 6.6 reads per edit before February 12 to 2.0 reads per edit by March 8 to 23, based on more than 6,800 saved sessions.
- In Anthropic's April 23 postmortem, the company said the API was unaffected and traced product quality issues to three separate changes: a March 4 default-effort downgrade, a March 26 stale-session thinking bug, and an April 16 prompt change that reduced verbosity.
- the OpenClaw incident page and the HN discussion summary both describe repos where the string "OpenClaw" in commit history or files could trigger refusals, abrupt disconnects, or quota draining into separate usage.
- the subscription-policy thread says Anthropic stopped covering third-party harnesses such as OpenClaw under Claude subscriptions on April 4, while the follow-up discussion framed local MCP setups as outside that restriction.
- the main HN thread surfaced two operator workarounds before Anthropic's postmortem landed: preserve session logs for regression analysis, and override defaults with higher effort or disabled adaptive thinking in
settings.json.
You can read the full regression issue, Anthropic's own postmortem, and the live settings reference. The oddest thread is that three different failure modes stacked on top of each other: one changed default effort, one quietly dropped prior reasoning after idle sessions, and one made any repo mentioning OpenClaw feel like a tripwire.
Read:edit ratio
Reported Regression in Claude Code Performance Following Thinking Content Redaction
A GitHub issue reported a performance regression in Claude Code starting in February 2026, where users found the model became unreliable for complex engineering tasks. Quantitative analysis of over 6,800 session files suggested that the February 12, 2026, implementation of thinking content redaction caused a shift in model behavior from research-first to edit-first, characterized by a 70% reduction in research actions (reads) per code edit. The issue has been closed and locked by the maintainers.
Discussion around Issue: Claude Code is unusable for complex engineering tasks with Feb updates
Thread discussion highlights: - noxa on author analysis and log-based regression claim: The author says they can spot shallow thinking via read:edit ratio shifts and character changes in logs, and recommends backing up session logs because defaults only keep 20 days. - bcherny on Anthropic response on redaction: Anthropic staff says `redact-thinking-2026-02-12` is UI-only, hides thinking summaries to reduce latency, and does not change thinking budgets or extended reasoning under the hood. - thrtythreeforty on user-reported Opus 4.6 degradation: A commenter says switching to Opus 4.6 quickly felt like it was post-trained to hack something together, with repeated ‘simplest fix’ behavior and less research.
The regression report split saved sessions into three windows: Jan. 30 to Feb. 12, Feb. 13 to Mar. 7, and Mar. 8 to Mar. 23. In the issue writeup, the headline number is a fall from 6.6 reads per edit to 2.0, alongside a drop from 46.5 percent reads to 31.0 percent.
According to the HN discussion summary, users were not just counting tool calls. They were comparing read:edit ratios, character deltas, and session traces to argue that Claude Code had shifted from research-first edits toward immediate local changes.
Default effort, stale thinking, verbosity
Issue: Claude Code is unusable for complex engineering tasks with Feb updates
The useful engineering signal is about how Claude Code behavior may have changed in practice: the report claims fewer repository reads, more shallow edits, and the need to preserve session logs to audit regressions. The thread also surfaces concrete operational knobs users are trying, like `showThinkingSummaries`, effort settings, and disabling adaptive thinking.
Anthropic's postmortem turns what looked like one broad regression into three dated product changes.
- March 4: Claude Code's default reasoning effort moved from
hightomedium, then Anthropic reverted that on April 7 after users said they preferred higher intelligence over lower latency. - March 26: a stale-session optimization started clearing old thinking on every subsequent turn instead of once, which Anthropic says made Claude seem forgetful and repetitive; the fix shipped April 10 in v2.1.101.
- April 16: a system-prompt instruction to keep tool-call text under 25 words and final responses under 100 words hurt coding quality; Anthropic reverted it on April 20.
That also resolves one contested claim from the HN thread. Boris Cherny, Claude Code lead at Anthropic, said the redact-thinking-2026-02-12 header was UI-only, and the postmortem points instead to effort defaults, stale-session context loss, and prompt tuning.
OpenClaw tripwire
Users Report Claude Code Anomalies When Detecting "OpenClaw" in Commit History
Users have reported that Anthropic's Claude Code agent exhibits unexpected behavior when it encounters the term "OpenClaw" in local git commit history. According to developer accounts, the presence of this string can cause the agent to either refuse tasks, abruptly terminate sessions, or trigger increased usage charges by shifting processing from quota to API billing. The behavior appears to be triggered by the mere presence of the term in commit messages or repository files, raising questions regarding how Claude Code processes repository context.
Discussion around Claude Code refuses requests or charges extra if your commits mention "OpenClaw"
Thread discussion highlights: - abdullin on Reproduction of the bug: "I reproduced this on my account" ... `claude -p "hi"` ... "Immediate disconnect and session usage went to 100%" - stingraycharles on Likely crude implementation: "it looks like a simple regex" ... "Could have been as easy as ... automatically whitelisting it" - xpe on Need for verification: "Where are the people saying 'Maybe this is true, but please tell me you considered other explanations first!'"
The second HN flare-up was stranger. The OpenClaw incident page says some Claude Code sessions refused requests, disconnected immediately, or pushed usage to 100 percent when local repo history mentioned "OpenClaw."
According to the companion discussion summary, at least one commenter reproduced the failure with a trivial claude -p "hi" test, while another guessed the filter looked like a crude regex over repository context. A third reply in that discussion pushed back on jumping to one explanation too quickly, which is about the right amount of skepticism for any repo-context bug that also touches billing.
Subscription boundary
Tell HN: Anthropic no longer allowing Claude Code subscriptions to use OpenClaw
Relevant as a developer-tooling/platform-policy change: Anthropic is separating subscription entitlements from third-party Claude harnesses, which affects how builders structure CLI agents, MCP/local integrations, and billing assumptions. The discussion also surfaces practical migration paths like local-first tools and alternative models or API-based pricing.
Discussion around Tell HN: Anthropic no longer allowing Claude Code subscriptions to use OpenClaw
Thread discussion highlights: - toniantunovi on local-first / MCP alternatives: Tools that route your API traffic through a third-party harness are dependent on Anthropic's policy toward that harness; tools that run locally and integrate via MCP are outside this restriction entirely. - _pdp_ on open-source alternatives: The solution as usual is open source... move a very expensive Sonnet 4.6 agent to step-3.5-flash and it works surprisingly well. - mariopt on capacity and token usage: OpenClaw burned 2.46 trillion tokens in the last 30 days; the commenter argues coding plans are being ruined by extreme token use and should move to API billing.
The policy change underneath the bug reports was simpler than the bug itself. The April 4 thread quotes Anthropic's notice saying Claude subscription limits would no longer cover third-party harnesses including OpenClaw, and that those flows would move to separate pay-as-you-go usage.
The follow-up discussion turned that into an architecture distinction:
- Third-party harnesses that proxy Claude traffic inherit Anthropic's policy toward that harness.
- Local-first tools that integrate via MCP sit outside that specific restriction.
- Some users responded by shifting heavy agentic runs to API-priced or open-source alternatives.
That split matters because the OpenClaw disconnect reports mixed three layers that engineers usually treat separately: repo scanning, product policy, and billing path.
settings.json knobs
Issue: Claude Code is unusable for complex engineering tasks with Feb updates
The useful engineering signal is about how Claude Code behavior may have changed in practice: the report claims fewer repository reads, more shallow edits, and the need to preserve session logs to audit regressions. The thread also surfaces concrete operational knobs users are trying, like `showThinkingSummaries`, effort settings, and disabling adaptive thinking.
The fastest community workaround was not a new model, it was configuration. The main thread surfaces people forcing higher effort, turning on thinking summaries, and disabling adaptive thinking to stop Claude Code from under-allocating reasoning on harder tasks.
Anthropic's settings docs confirm that effortLevel persists across sessions and accepts low, medium, high, or xhigh, while alwaysThinkingEnabled forces extended thinking by default. The same docs place those settings in both ~/.claude/settings.json and project-level .claude/settings.json, so the workaround path was not just folklore, it was part of the product surface.
One operational detail from the HN discussion is worth keeping with the config story: session logs are short-lived by default, so the people making the strongest regression claims were preserving them before the 20-day retention window erased the evidence.