caveman update, 2026-10-24

Upstream: JuliusBrussee/caveman, pinned at tag v4.0.1 (d262ab3882952a59312ff3537cd86d894523c5ab, 2026-10-13, latest release tag; main was 1f951fb, a day old, not used). Previous base: no commit was recorded. Our copy matched upstream around 2026-04-09 (a9e62b48, "Add mode-tracker hook"), then diverged by local rewrite.

A monorepo (Go engine, proxy, browse, CLI, 30+ agent installer). The Claude hooks are now six files, about 150 KB (src/hooks/), with per-session mode state, repo/user config files, a SubagentStart hook, a SessionEnd stats recorder (spawns child processes) and a statusline. Levels were replaced by three skills: caveman (lite/full), ultracave (ultra) and megacave (wenyan). The caveman skill was rewritten (nine rules, answer first, bounded tool status, no arrows, no invented abbreviations).

Full adoption would mean re-grafting ultra-plus and the English lock onto a 150 KB state machine we do not otherwise need. We kept our small hooks and ported the upstream behaviour that matters:

• SessionStart no longer resets the mode on compact/resume/fork (upstream #691): a mid-session ultra-plus or an explicit "stop caveman" survives compaction. startup and clear still auto-activate at full.
• Mode is per session (~/.claude/.caveman-sessions/<session_id>.mode), .caveman-active stays as a last-write-wins mirror that never holds "off". session_id is whitelisted before it touches a path.
• Symlink-safe atomic writes (O_EXCL temp + rename, refuses symlinked file or parent), as upstream safeWriteFlag.
• Stdin read returns on the first complete JSON object with a 2s watchdog (upstream #729/#833, Windows pipe lag).
• One-shot skills (commit/review/compress) restore the displaced prose mode on the next prompt.
• /caveman status (read-only), /caveman off, <command-name> envelope handling, natural-language on/off phrases (short prompts only, so a long prompt that merely mentions "stop caveman" is not a command).
• Skill text: caveman SKILL.md gains the v4 rules (never drop negations, bounded tool-run status, never perform caveman, persisted-outside-chat and question exemptions, pre-send check). caveman-commit gets the user-attribution-trailer clause. caveman-compress gets the explicit-filepath rule, secret-file refusal and wider type lists.

• ultra-plus level and its hook text; English only, no wenyan/megacave; levels lite/full/ultra/ultra-plus kept (upstream removed lite as a separate level).
• Auto-activate every session. No statusline nudge.
• caveman-compress stays in-session (no scripts/): upstream's script sends the file to a model provider.
• Not taken: caveman-config/parse/stats/statusline/host-session-start, SubagentStart and SessionEnd hooks, cavecrew, and the engine/proxy/browse/mem skills.

Read: README and CLAUDE.md of upstream, src/hooks/caveman-activate.js (stdin/state/emit path), grep of all four hook files for require/network/exec. Taken code is our own (caveman-core.js, activate, tracker): only fs, path, os; no network, no child_process, no telemetry; writes only under $CLAUDE_CONFIG_DIR (default ~/.claude). Upstream's tracker requires child_process (stats command) and its compress scripts call model providers; neither was taken. The remaining upstream hook files (config, parse, stats, 150 KB) were not read line by line because they are not installed.