Brings safari-mcp (https://github.com/achiya-automation/safari-mcp)
to CLI-Anything as a schema-driven harness. Every one of the 84 MCP
tools is auto-generated as a Click command from the bundled tool
schema (resources/tools.json) — guaranteed 1:1 parity with the
upstream server.

Architecture highlights:
- Schema-driven CLI generation via scripts/extract_tools.py (offline
  parser of safari-mcp's Zod schemas with depth-aware modifier
  detection for nested z.array(z.object(...)).describe() patterns)
- Dynamic Click command registration at import time from tools.json
- TestCLISubprocess with _resolve_cli() per HARNESS.md Phase 5
- URL validation with scheme blocking and optional private-network
  lockdown (utils/security.py)
- 95 tests: 16 unit + 36 security + 24 parity + 19 E2E (all pass
  against real Safari; SAFARI_E2E=1 to enable live tests)

Honest positioning:
The harness is STRICTLY slower than using safari-mcp directly for
MCP-compatible agents (Claude Code, Cursor, Cline). Per-call
latency: 119ms (persistent MCP session) vs 3,023ms (CLI subprocess
spawn). See HARNESS.md "Performance Tradeoffs" for the full
benchmark — MCP wins by ~25× on latency. The CLI is optimized for
a different audience:
- Agents that do not speak MCP (Codex CLI, GitHub Copilot CLI)
- Bash scripting and jq pipelines
- CI/CD and cron jobs
- Long-running Opus sessions where tool-definition token overhead
  (~8K tokens per API call with MCP vs ~100 with CLI) dominates cost

Both paths return byte-identical output (verified live with Unicode
titles via list_tabs).

Files:
- safari/agent-harness/HARNESS.md         architecture deep-dive
- safari/agent-harness/SAFARI.md          software-specific SOP
- safari/agent-harness/setup.py           PyPI package metadata
- safari/agent-harness/scripts/
    extract_tools.py                      offline Zod parser → tools.json
- safari/agent-harness/cli_anything/safari/
    safari_cli.py                         Click CLI with dynamic tool group
    core/session.py                       in-memory REPL state
    utils/safari_backend.py               sync MCP stdio client wrapper
    utils/security.py                     URL validation + DOM sanitization
    utils/tool_registry.py                loads tools.json, kebab↔camel
    utils/repl_skin.py                    copied verbatim from plugin
    resources/tools.json                  84-tool schema bundle (safari-mcp 2.7.8)
    skills/SKILL.md                       agent discovery manifest
    README.md                             user-facing docs
    tests/                                11/36/24/19 test files + TEST.md

…anguage

- Remove unused `sanitize_dom_text` and `_PROMPT_INJECTION_PATTERNS`
  from security.py (and 11 corresponding tests). The function was
  never called in the CLI pipeline.
- Add explicit `return` after `_handle_error` in `_validate_url_or_exit`
  so control flow is unambiguous if `sys.exit` is mocked.
- Replace manual `__name__`/`__doc__` copy with `functools.wraps` in
  `handle_error` decorator.
- Document `asyncio.run()` limitation in `call()` docstring.
- Add `safari` entry to `_ACCENT_COLORS` in repl_skin.py (was falling
  back to default).
- Add conflict-of-interest disclosure to HARNESS.md and SAFARI.md.
- Strip promotional language, cost comparisons, and benchmark tables
  from HARNESS.md, SAFARI.md, SKILL.md, and README.md. Keep factual
  per-call overhead note; remove dollar-amount positioning.
- Remove DOMShell comparison table from HARNESS.md.

All 49 non-E2E tests pass.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

# Conflicts:
#	.gitignore