Minimal cave-code personality + /caveman command (no tool suppression)

[romgenie]

Goal

Ship the cave-code personality preset, /caveman slash command, and bundled skill as a minimal additive PR. Treat the hard clientTools suppression path in #19 as a separate follow-up that can be added if the persona alone proves insufficient to stop tool calls on the mode-switch turn.

Scope (what's in)

• src/agent/prompts/persona_caveman.mdx
• src/skills/builtin/caveman/SKILL.md
• Register the new prompt in src/agent/promptAssets.ts, document in src/agent/prompts/README.md
• Add caveman to PersonalityId, PERSONALITY_OPTIONS, PERSONALITY_ALIASES (cave-code), and the existing content/human/block-definition branches in src/agent/personality.ts (no table-lookup refactor — minimal branches only)
• New src/cli/commands/caveman.ts with CAVEMAN_MODES, normalizeCavemanMode, buildCavemanCommandPrompt, isCavemanCommandInput. Does not export suppressPreparedClientTools.
• /caveman dispatch in src/cli/App.tsx that calls processConversation with the mode-switch prompt. No new refs, no queued-approval metadata changes, no approval-cancel reset.
• Register /caveman in src/cli/commands/registry.ts with a plain string output for non-interactive use (no CommandResult type change).
• Update --personality help in src/cli/subcommands/agents.ts to list cave-code.
• Tests: personality.test.ts additions for the new preset + a focused caveman-command.test.ts that covers mode parsing, prompt building, and SKILL.md example alignment.

Scope (what's out — defer as separate PRs if needed)

• Hard clientTools: [] suppression during the mode-switch turn (pending-approval metadata, cancel-path reset, suppressPreparedClientTools). The persona + skill + prompt already instruct the agent not to call tools; only add enforcement if we observe the model calling tools on the switch turn anyway.
• personality.ts refactor from ternary chains to PERSONA_TEMPLATE_BY_ID / HUMAN_TEMPLATE_BY_ID / PERSONA_CONTENT_OVERRIDES tables.
• registry.ts handler return-type widening to string | CommandResult.
• Unrelated cleanups bundled into Add cave-code command and skill support #19: build.js Windows chmodSync fix, reconcileExistingAgentState.ts dead areToolSetsEqual, generate-memory-viewer.ts ConversationListItem typing, runtimeDeps.test.ts dead helper.

Tradeoff

Without clientTools suppression, the agent could occasionally attempt a tool call on the /caveman turn despite the persona telling it not to. If this happens in practice, open a follow-up adding the suppression plumbing.

Validation

• bun run typecheck
• bun run lint
• bun run build
• bun test src/tests/agent/personality.test.ts src/tests/cli/caveman-command.test.ts

Related: #19 (larger variant — replaces #13–#18).