Claude Code on Windows desktop: env flags, autoCompactWindow, output-cap hook and Bash backslash quirk

Claude Code on Windows desktop: token and config facts

Verified September 2026 on Claude Code CLI 2.1.260 and the Claude desktop app 1.46388.4 (Windows 11).

Env flags the desktop app injects

• DISABLE_MICROCOMPACT=1 is hardcoded into every desktop-spawned session, but CLI 2.1.258 and later contain no reference to it. It is a dead flag and there is nothing to override. Microcompact in these builds is a server-hint-driven "keep-recent" tool-result clear with no user switch.
• ANTHROPIC_BASE_URL=https://api.anthropic.com is set by the desktop app. It is not a proxy, and tool deferral stays active.
• CLAUDE_CODE_PROMPT_CACHE_TTL=1h is set for desktop sessions (subagents get 5m).

The "Model auto-fallback" toggle

• The desktop preference is mirrored from the claude.ai account setting enabled_model_auto_fallback. Local file edits are reverted.
• When off, the desktop only sets CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1. It governs refusal fallback (a model refuses, the request is retried on another model). It does not govern usage-limit switching. Limit hits show synthetic "Switch to another model" messages and the switch is manual.
• Local equivalent: "CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK": "1" in the settings.json env block (settings env replaces inherited shell values).

Context size is the cost driver

• With 1M-context models the auto-compact window moves to about 1M tokens. One measured session ran at 842k tokens of context per turn.
• Fix: set "autoCompactWindow": 265000 (numeric tokens, accepted range 100000 to 1000000) in ~/.claude/settings.json. The env var CLAUDE_CODE_AUTO_COMPACT_WINDOW and the --autocompact flag take precedence over the key.
• Compaction runs inside a live session and re-reads the whole context, so "scheduled compaction of idle chats" spends tokens rather than saving them.

PreToolUse hook that caps command output

A hook with matcher Bash|PowerShell can rewrite long-output commands (test, build, install, git log) to keep only the tail:

• Bash form: set -o pipefail; { cmd } 2>&1 | tail -n 150
• PowerShell form: & { cmd } 2>&1 | Select-Object -Last 150
• The hook skips commands that are already piped or redirected, heredocs, backgrounded commands, --json, and commands ending in a # full marker.
• Configure it in exec form with forward-slash Windows paths (C:/path/to/python.exe plus args).
• updatedInput replaces the entire tool input object, so unchanged fields must be echoed back.
• A running session picked the hook up without a restart. Overhead was about 26 ms per call.

Bash tool quirk in Claude Desktop on Windows

The Bash tool collapses \\ to \ inside the command text, even in single quotes and quoted heredocs. JSON, regexes and Windows paths with escaped backslashes get corrupted.

• Build backslashes with chr(92) in Python, or use forward slashes in Windows paths.
• Python on Windows does not understand Git Bash /c/... paths. Pass C:/....
• Observed September 2026: a Python heredoc editing a path containing \\t wrote a literal TAB (the \\t had collapsed to \t), and \\b became a backspace. The damage was silent and selective, because invalid escapes such as \C survive. A workable rule is to put any inline script that contains a Windows path or a regex into a .py file written with the file-write tool, then run python <file>.

Desktop versus terminal settings precedence

• The desktop app launches claude.exe with --effort and --model taken from its own preferences, so the settings.json model and effortLevel keys only affect terminal-launched sessions. autoCompactWindow, env, hooks and plugins apply to both surfaces.
• Keeping settings.json model and effortLevel equal to the desktop preferences makes both surfaces behave the same.
• In auto permission mode the classifier can refuse edits to ~/.claude/settings.json through both the Edit and Bash routes. A desktop restart cleared the refusal; retrying variants did not help.

Auditing token use

• Session logs are JSONL files under ~/.claude/projects/<project>/<session>.jsonl. Deduplicate assistant lines by message.id before summing usage.cache_read_input_tokens and the other usage fields.
• Entries whose model is <synthetic> are rate-limit or error notices, not API calls, and are not model switches.
• Verify env-var and setting names against the CLI binary (grep -a on claude.exe) and the docs markdown rather than recalling them.