cc-safe-setup

The #1 Max complaint cluster (#16157, 720+ reactions). 20-second check: how much of your fast quota burn is self-caused (shared quota, Opus, big context) versus a possible anomaly worth reporting. Shareable card, browser-only.

The full explainer behind the burn check: why you can't see your real quota, the version boundaries where consumption changed, the false "Usage credits required for 1M context" gate and how to clear it, what you can measure free, and when it's a platform report worth filing.

Field notes from an 800-hour run: the five small habits that reliably brought the bill down and kept sessions calm — fresh session per task, Sonnet for routine work, compacting on purpose, watching the numbers, and a lean context. Each is a win you can copy today, free.

20-second check: does the June 15 programmatic-billing split hit you? See your plan's new credit cap, a live countdown, and a shareable exposure card. Browser-only.

The precise version: drop your ~/.claude/projects logs and it splits your usage into interactive (Pool 1) vs programmatic (Pool 2) by the entrypoint your sessions actually recorded — not a guess. Browser-only, nothing uploaded.

5-minute self-assessment: paste your past-30-day usage, get estimated monthly cost after June 15 with Pool 1 / Pool 2 split. Browser-only, no signup, no telemetry.

10-question audit returns a readiness score (0-100) with each gap mapped to a specific Migration Playbook chapter. Browser-only, no signup.

Structural analysis of $47K/3-day subagent runaway, 887K-tokens/min parallel-49 burn, and $6K-overnight cache-TTL surprise. Three operator-side preventions mapped to PR #298, #286, #283.

7-step preparation checklist, hook patterns for closure-claim arrest and budget gates, parallel-agent cost reality, and the four official-doc blind spots. CC BY 4.0, ~1,800 words.

A permission rule that silently doesn't match is worse than no rule. Windows path normalization, settings precedence, the UI mode overriding defaultMode — and how to see your real rules with /permissions.

The 529 overload vs. a real usage limit vs. 429 acceleration limits — how to tell them apart, confirm with /usage, and what you actually can (and can't) do.

When Opus 4.8 reports output for a tool that never ran — the parallel-cancellation trigger, how to spot a fabricated "done", and the operator-side fixes (#63538, #63884).

When Claude Code says "done" or "tests passed" — would your setup catch it if that weren't true? 8 questions → a 0-100 score for your defenses, with the hook pattern that closes each gap (#33781, #44955, #27430).

Legitimate firmware/sysadmin/security work false-positived, then the whole session dies and bills you for the restart. Why it happens, how to keep working, and the refund framing (#64405, #61185).

Auto-approve safe git, project reads, and test runs. ~60% fewer permission prompts.

rm -rf blocker, force-push guard, secret leak prevention, compaction breaker, token saver. Copy-paste ready.

50 GB destroyed, SSH keys wiped, production databases killed. One hook prevents all of it.

15+ compactions overnight consumed entire budget. Circuit breaker hook stops it in 30 seconds.

Thinking phase consumed entire quota with no output. Stall detector hook catches it early.

/cost の出力と settings.json と CLAUDE.md を貼ると、費用の内訳・キャッシュの効き方・CLAUDE.md が毎回載せている常駐コストが出ます。日本語の文字数で数えるので、実際より軽く出ません。

SKILL.md は、そのスキルが発火しなくても毎セッション文脈に載ります。1つ貼ると、毎回かかる常駐トークンと、外へ出せる重い塊が分かります。

拾ってきたリポを「信頼しますか」で はい と押す前に。.claude/settings.json と .mcp.json を貼ると、hook の自動実行・鍵の漏洩・MCP の自動起動を赤黄緑で判定します。

permissions.deny に .env や鍵を書いても、cat や python 経由では抜けられます。自分の deny ルールを貼ると、守れているものと素通りするものが並びます。

claude daemon status の出力を貼ると、止めたはずの背景の作業が生き残っていないか、自動更新で孤児になっていないかを判定します。

自分のログ（.jsonl）を読み込ませると、tool call could not be parsed が CLI の版ごとに何回起きたかを集計します。最悪の版は環境ごとに違うので、自分の実績で選べます。

allow/deny を書いたのに無視される。設定したつもりで守られていないのは、ルールが無いより危険。Windows のパスの正規化・優先順位・モードの上書きの原因と、/permissions での確認の仕方。

原因の調査だけ頼んだのに、確認なしで docker restart や ALTER SYSTEM、マイグレーションを実行してしまう。「原因を見つけた=直してよい」で読み書きの境界が崩れる原因と、PreToolUse の hook で読み取り専用を強制する防ぎ方。

$200 を払って作業の途中で締め出される前に。何が枠を食っているか調べて長持ちさせる、公式に基づく検証済みの方法。

利用枠を数分で使い切る前に、トークンと Actions の暴走を止める。

Claude Code の rm -rf 事故を、実行の前に確認で止める。

DROP DATABASE / DELETE の暴走を、実行の前に止める。

未コミットの変更が消える事故を防ぐ。

鍵・トークンの読み取りと流出を止める。

一時ファイルとログの蓄積が原因のときの、原因と直し方。

設定したのにツールが出ない。最大の原因は承認待ち。/mcpで切り分ける手順。

停止・凍結の原因と対処。公式の出典つき。

回復できるか、報告するしかないかを切り分ける。

固定したはずのモデルが揺れるとき、ログから検証する。

事故を防ぐ設定の全体像（日本語）。

Pressed Esc Esc and the default "Restore code and conversation" wiped your edits (#64615)? Exactly what /rewind can and can't touch, how to recover from git now, and the free hook that makes any rewind recoverable.

Check the protections you have, get an honest X/8 coverage card you can screenshot and share. Runs in your browser.

Does every AI tool actually read your one AGENTS.md? Get an honest X/6 interop score you can screenshot and share. For the #6235 pain.

Paste your settings.json. Watch 10 real attack scenarios play out. See what your hooks catch, and what slips through.

5 questions, 30 seconds. How exposed is your Claude Code setup? Based on 90 real incidents.

90 real safety incidents from GitHub Issues. Filter by category, severity, and date.

Safe in 5 minutes. Step-by-step for beginners. No coding required.

Max Plan running out too fast? 5 questions to diagnose why. Free.

Paste /cost output + settings.json for deep analysis. Cache efficiency, safety score, hook recommendations.

Paste settings.json → instant check for 7 known Opus 4.7 vulnerabilities.

4x token burn, 50GB data loss, classifier bugs. Hooks and fixes for every known issue.

65+ bugs tracked. 50GB+ lost. Classifier failure, cache destruction, 4x token burn. Full analysis.

Pin Claude Code to Opus 4.6. 4 methods: CLI, settings.json, project config, alert hook.

How much is the 4.7 upgrade costing you? Input your usage → estimated cost increase.

Cut your token bill in half. 800 hours of data. CLAUDE.md patterns, hooks, model selection. Opus 4.7 updated.

800時間の運用データに基づくトークン節約ガイド。Opus 4.7対応。

3 rules to cut token costs by 40%. Before/after examples and templates.

Paste /cost output → instant cache diagnosis. Detects #46917 inflation bug.

Is your Claude Code version affected by cache inflation (#46917) or TTL regression (#46829)?

6 questions to find vulnerabilities. Based on $1,800+ in real losses (April 2026).

Score your Claude Code setup out of 100. Find gaps and get fix suggestions.

14-point checklist before autonomous operation. Interactive, printable, shareable.

Real problems → copy-paste solutions. Permission fatigue, silent edits, syntax errors, each with a working hook.

Paste your settings.json → get instant score, error detection, and improvement suggestions.

Paste settings.json → find broken paths, missing scripts, Windows backslash issues. Fix "Permission denied".

Write hooks, test with sample inputs, see results instantly. Browser-based hook IDE.

30+ copy-paste hook patterns. Block, approve, monitor, with one-click copy.

Answer 5 questions. Get your personalized safety hook setup in 30 seconds.

Type the error message or version you are hitting. Searches 109 incident write-ups, fix recipes, and prevention hooks.

Paste your settings.json. Find which safety hooks you're missing. Get install commands.

Detect common hook configuration mistakes: wrong exit codes, missing jq, invalid event types.

Cut token consumption in half, save $20-80/month on Max plan. 10 chapters, 800+ hours of data, copy-paste templates. Ch.1 free.

Incident 1 of 10: The 1h to 5m Cache TTL Silent Downgrade. Full book live on Gumroad since 2026-05-05 (see product page for current price), 10 production Claude Code failures reverse-engineered from Issues #46829/#52648/etc with detection hooks.

Monthly Claude Code incident digest. 4-8 new incidents per issue, fix-it hooks, safety checklist, product updates. ¥500/month. Delivered around the 15th of each month; each issue has a free preview chapter.

Free 6,000-character HTML preview of the May 2026 Safety Lab Part 1: nine incident articulations (cache_control, usage display, Cowork, Worktree, Skill, UI, client switching, child worker identity leak, 2.1.126 silent regression). HTML companion to the Gist preview; better suited for non-technical readers comparing the depth of monthly content.

Free 6,500-character HTML preview of the June 2026 Safety Lab Part 1: the June 15 billing cliff preparation guide with 7 ordered steps (receipt-notice readiness, automation inventory, current-usage baseline, pre-flight check protocol, ceiling-stop instrumentation, official-key switch decision, before/after diff capture) and 3 warning omens (OAuth-key misuse, asking the model about billing, missing the receipt email). Pairs with the May 2026 preview for back-to-back depth comparison.

Free 6,500-character HTML preview of the July 2026 Safety Lab Part 1: the overengineering complexity trap, articulated from thecatfix's 16-month 62-issue corpus (2,109 plan documents, 19% implementation rate, 968M tokens over 4 months, 60 agents spawned in 5 days then deleted). Three self-checks (plan-to-implementation ratio, agent-to-output ratio, design-doc-to-runtime ratio) and three warning omens (future-extension vocabulary frequency, daily agent-spawn spike, articulation-verb accumulation). Detection requires cumulative metrics — single-incident inspection cannot surface this failure mode.

Free 6,500-character HTML preview of the August 2026 Safety Lab Part 1: the Pro / Max 5x / Max 20x quota anomaly cluster (10 issues totaling ~2,200 reactions). Three independent root-cause paths (server-side cache_creation inflation per Issue #46917 — 49,726 to 69,922 tokens between v2.1.98 and v2.1.100; parallel sub-agent fan-out; session-start context cascade). Three self-diagnostic steps (ccusage rollups, jsonl jq analysis, per-minute token rate calculation) and three defensive hooks (cache-creation-drift-detector, quota-anomaly-detector, session-rate-monitor) covering the three paths independently.

Day-by-day preparation guide for the 2026-06-15 Claude Code Pool 1 / Pool 2 billing split. Four phases (baseline / decision framework / preparation / final readiness), one 5-10 minute task per day from May 30 through June 14, plus a Day 0 action plan. Industry signals: 12x cost multiplier for light loads, 150x+ for heavy Sonnet automation. Pairs with Migration Playbook for the decision framework.

Six-issue corroborated cluster: unauthorized Gift Max charges, silent Pro/Max downgrades, double-charges, promo misapplication. Three pattern axes (A: unauthorized Gift Max, B: silent downgrade + double-charge, C: promo misapplication). Five user-side early-detection steps from monthly check through chargeback escalation. Issues #51168/#62461/#41581/#56895/#61339/#54133 cited; corroborated against Capterra/The Register/Trustpilot reviews.

Curation page for Migration Playbook (decide), Token Book (cost control), and Incident Postmortems (forensics). Reading paths for the burn debugger, switch decider, and incident survivor. Each book is sold separately on Gumroad. Live since 2026-05-05.

Five-question interactive quiz. Two minutes. No email. Tells you whether to start with Migration Playbook, Token Book, or Incident Postmortems based on your situation right now.

Updated 5-question selector that covers all four current paid books (Migration Playbook, Claim-Verify Handbook, Sub-Agent Observability Handbook, Incident Postmortems) and surfaces the CC Safety Lab monthly subscription as an add-on when applicable. Browser-only, no signup.

Twelve specific operational pains observed in May 2026 (5-hour quota burn, sub-agent identity leak, DROP DATABASE incident, hybrid delegation, and more) with each one mapped to a free article and the book that solves it.

Cursor, Cline, Codex CLI, Aider, DeepSeek wrapper compared on release cadence, pricing, ideal user. Plus the five questions that decide stay vs hybrid vs switch.

Calculate how much Opus 4.7 bugs are costing you. Shows hidden token overhead from 5 known issues.

How much are you overpaying? Opus 4.7 defaults to xhigh, switch to save 30-50%.

3 hooks that saved work during April 6-8 outages. Auto-checkpoint, session backup, periodic commit.

Opus 4.6 vs 4.7 cost comparison (+35%). See where tokens go. 800h data.

Paste your CLAUDE.md. See token cost per turn, find waste patterns, get optimization tips.

Detect lazy rewrites, skipped reads, and untested commits. Based on 6,852-session analysis.

Describe what you want in English. Get a working hook + settings.json entry.

10 real GitHub Issues. See the before (disaster) and after (hook saves it).

All hooks in a sortable, filterable table. Find exactly what you need.

Max Plan running out too fast? Diagnose and fix hidden token sinks with two hooks.

15 questions answered. How hooks work, performance, debugging, and more.

Skip permission prompts for safe commands. Read-only, build tools, git, with PermissionRequest hooks.

Three-layer defense: block reading .netrc/.npmrc, stop credential hunting, catch hardcoded secrets.

Every settings.json field explained. Hooks, permissions, env, models.

PreToolUse, PostToolUse, Stop, UserPromptSubmit, when each fires and what it can do.

Exit 0 = allow. Exit 2 = block. Plus the bash syntax error trap that can lock your session.

10 common problems with step-by-step fixes. Hook not firing? Start here.

Answer 2-3 questions → get the exact hook you need. No knowledge required.

Block, Approve, Warn, Log, Validate, Protect, Limit, Checkpoint, Enforce, Recover.

7 mistakes that cause real incidents. Syntax error lockout, broad matchers, missing jq.

Manual setup vs cc-safe-setup vs do nothing. 15 criteria compared.

Moving from Cursor, Windsurf, Aider, or Copilot? Map your safety setup.

Type any command, see which hooks fire. Test before you install.

Auto-fix the invalid --no-stat flag bug. 44 reactions on GitHub.

Stop rules from being ignored after compaction. Re-inject CLAUDE.md automatically.

Block or conditionally control auto-compaction. 42 reactions.

Fix broken domain:* wildcard matching in sandbox mode. 46 reactions.

Detect .bashrc patterns that freeze agent-spawned shells (Angular, nvm, conda).

Block curl/wget to unauthorized domains. Defense-in-depth for sandbox gaps.

Source code, issues, releases. Star if useful.

MIT licensed. No npm dependencies (needs jq).

70 hooks from 7 projects. Search, browse, install.