Claude Code users report a 70% read:edit drop, OpenClaw disconnects, and tighter limits
Claude Code discussions centered on a 70% read-to-edit drop, OpenClaw-linked disconnects or 100% session usage, and tighter subscription boundaries. Users should audit logs, check commit-message triggers, and compare billing or settings workarounds.
TL;DR
- Across issue 42796 and the main HN thread, users centered the Claude Code quality drop on a measured shift from research-first to edit-first behavior, including a reported 70% collapse in the read:edit ratio across 6,852 session logs.
- According to the HN discussion summary, Anthropic said the
redact-thinking-2026-02-12header was a UI-only change, while An April 23 postmortem later tied the broader regression to three separate Claude Code changes that were resolved in v2.1.116. - the OpenClaw report and the follow-up HN discussion described a separate failure mode: repositories with commit text mentioning "OpenClaw" could trigger an immediate disconnect and consume an entire session's usage bar.
- In the OpenClaw subscription thread, the bigger argument was not only about one string trigger, but about Anthropic drawing a harder line between first-party Claude subscriptions and third-party harnesses, with commenters comparing that shift to straight API billing and rival capped plans.
You can read issue 42796 on GitHub, skim Anthropic's postmortem, and check the current Claude pricing page. The official Claude Code cost docs now make one useful distinction explicit: the /usage dollar estimate is for API users, while Pro and Max subscribers are meant to read plan usage bars instead.
The read:edit collapse
Performance Regression in Claude Code (GitHub Issue 42796)
GitHub issue 42796 documents a significant performance regression in the Claude Code tool observed by users starting in February 2026. Telemetry analysis of over 6,800 sessions revealed a correlation between this quality decline and the introduction of thinking content redaction. Specifically, the model shifted from a "research-first" to an "edit-first" behavior, with the ratio of file reads to edits dropping by 70%. Users reported that this reduction in reasoning depth impaired the tool's ability to execute complex engineering tasks, leading to issues such as instruction ignoring and suboptimal code changes.
The core claim came from a GitHub bug report that analyzed 17,871 thinking blocks and 234,760 tool calls across 6,852 Claude Code session files. That report said the ratio of file reads to edits fell from 6.6x to 2.0x after the February redaction rollout, a roughly 70% drop, and framed the result as a shift from research-first work toward blind edits.
Issue: Claude Code is unusable for complex engineering tasks with Feb updates
This is relevant as a case study in agentic coding-tool reliability: users are debating whether a model update, UI redaction change, or workflow mismatch is causing worse code changes, and they’re sharing concrete mitigations like higher-effort settings, disabled adaptive thinking, and log-based audits.
The HN thread turned that telemetry into a workflow argument. One linked comment from noxa said old Jan and Feb logs were enough to compare read:edit ratios and thinking-character counts, while another user said phrases like "simplest fix" had become a reliable tell for shallower code changes the discussion summary.
Anthropic's postmortem
Discussion around Issue: Claude Code is unusable for complex engineering tasks with Feb updates
Thread discussion highlights: - noxa on report author explains the analysis: Use logs from Jan/Feb to compare read:edit ratios and thinking character shifts; add `cleanupPeriodDays: 365` if you want to keep enough session history to audit it. - bcherny on Anthropic response on the redaction header: `redact-thinking-2026-02-12` is described as a UI-only change that hides thinking summaries and reduces latency, not a change to reasoning budgets or the underlying model. - robeym on workaround settings for higher effort: A commenter shares `CLAUDE_CODE_EFFORT_LEVEL=max` plus disabling background tasks and adaptive thinking in `settings.json` as a way to keep Claude from shortchanging itself.
Boris Cherny, Claude Code lead at Anthropic, told HN that redact-thinking-2026-02-12 was a UI change meant to hide thinking summaries and cut latency, not a direct reduction in reasoning budget. That mattered because the original bug report had treated the redaction marker as the pivot point.
Anthropic's postmortem later described a messier root cause. The company said three separate Claude Code, Claude Agent SDK, and Claude Cowork changes were involved, and that the API was not affected.
The postmortem's concrete timeline:
- March 4: Claude Code's default effort moved from
hightomediumto reduce latency. - March 26: a prompt-caching change cleared Claude's earlier reasoning from conversation history after inactivity.
- April 7: the default-effort change was reverted.
- April 20: Anthropic said all three issues were fixed in v2.1.116.
That explanation does not fully erase the earlier dispute, but it does resolve one important question. Anthropic's public account blamed operational changes around effort defaults and memory of prior reasoning, not a hidden model downgrade.
OpenClaw triggers and disconnects
Claude Code Reportedly Flags Git Commits Mentioning "OpenClaw"
Developer Theo Browne reported that Anthropic's Claude Code tool may refuse requests or charge users extra if their git history contains a commit mentioning "OpenClaw" in a JSON blob. Users and observers speculate this is caused by an automated filter or prompt-level instruction within the tool that flags the term, potentially affecting workflows for developers working with related code or documentation.
A second cluster of complaints had a different shape. Theo Browne reported that Claude Code could refuse requests or burn through session usage when git history contained the string "OpenClaw," even inside a JSON blob.
Discussion around Claude Code refuses requests or charges extra if your commits mention "OpenClaw"
Thread discussion highlights: - abdullin on reproduction: I reproduced this on my account... git commit ... with OpenClaw in the message ... Immediate disconnect and session usage went to 100% - stingraycharles on implementation bug: it looks like a simple regex... Could have been as easy as ... automatically whitelisting it. - scottbez1 on subscription economics: Subscription models only work when marginal costs are low... minority of customers keeps paying for the all you can eat buffet and staying for hours...
HN commenters said they could reproduce it. One cited test described adding "OpenClaw" to a commit message and getting an immediate disconnect with session usage jumping to 100%, while another guessed the filter might be a simple regex rather than a deeper repo classifier the follow-up HN discussion.
The thread also carried its own epistemic brake. A separate commenter pushed back on the speed of the pile-on and asked who had verified the claim before assuming intent the HN overview.
Subscription boundaries
Tell HN: Anthropic no longer allowing Claude Code subscriptions to use OpenClaw
Claude-based agent and coding tool builders should read this as a billing and platform-boundary change: Anthropic is separating first-party Claude products from third-party harnesses, which can change cost assumptions, reliability, and how you architect around hosted subscriptions versus direct API usage. The thread also surfaces migration ideas like alternative models, local-first harnesses, and competing usage-based plans.
The larger policy fight was about where Anthropic wants flat-rate Claude plans to stop. In that HN thread, users treated the OpenClaw restriction as a boundary between first-party interactive use and third-party harnesses that can turn a subscription into effectively continuous agentic compute.
Discussion around Tell HN: Anthropic no longer allowing Claude Code subscriptions to use OpenClaw
Thread discussion highlights: - _pdp_ on alternative model/tooling: “The solution as usual is open source... moved a very expensive sonnet 4.6 agent to step-3.5-flash... we are most likely going to see a lot of micro coding agent architectures everywhere.” - password4321 on competitive billing alternatives: “GitHub Copilot supports Anthropic models... they have a monthly usage cap after which it is pay-per-prompt... $40/month for 1500 requests.” - davesque on support for Anthropic restricting harness abuse: “They should not be expected to shoulder the burden of community tools like OpenClaw... I’m happy to see they're doing something about this.”
That thread surfaced three concrete reactions:
- Some users said open-source harnesses and smaller model mixes would get more attractive if Claude subscriptions no longer subsidized heavy autonomous runs.
- Others compared Anthropic's shift with competitors that pair monthly caps with pay-per-prompt overflow, including a commenter citing GitHub Copilot's capped Anthropic access.
- Supporters of the restriction argued that subscription economics break when a minority of users run token-hungry agents through unofficial wrappers.
The official product pages now reflect that split more clearly than the April arguments did. The current pricing page says Pro includes Claude Code and Max offers 5x or 20x more usage than Pro, while the Claude Code costs docs say /usage cost estimates are for API users and are not the relevant billing number for Pro and Max subscribers.