feat(setup): skip LLM scan on effectively-empty projects

[George-iam]

Summary

Bug: running axme-code setup in a freshly-created empty folder sits for 1-3 minutes with no progress output, then completes with empty oracle/decisions output anyway. Reported 2026-05-17.

Root cause: the four LLM scanners (oracle / decision / safety / deploy) launch in parallel regardless of project content. Each round-trips to Claude generating boilerplate text + empty section markers, even when there's literally nothing in the project tree to analyse. The user has zero progress output between "starting scanners" and "scanners complete".

Fix

A pre-scan check isEffectivelyEmpty(projectPath) that walks the project tree (capped depth 4, early exit at threshold, skips .git/node_modules/build outputs/etc) and counts non-trivial files. If fewer than 3 non-trivial files exist outside boilerplate (README*/LICENSE*/.gitignore/CLAUDE.md/.editorconfig etc), we skip the LLM scanner block entirely and fall through to the existing deterministic init paths (bundlesToDecisions presets + initOracleDeterministic), which run instantly.

User sees a clear message instead of silent waiting:

[project] Project appears empty (< 3 non-trivial files outside
.git/node_modules/etc) — skipping LLM scanners, writing presets +
deterministic oracle only. Re-run with --force after adding code to
get an LLM scan.

Trade-offs

• Heuristic, not a sandbox. False positive (real project misidentified as empty) results in presets-only setup — user can re-run with --force after adding code.
• False negative (truly empty project misidentified as non-empty) means we run the LLM scan unnecessarily — annoying but not broken.
• Threshold of 3 is conservative: a tiny real library is usually 5+ files; a fresh git init with README+.gitignore is 0-1 non-trivial files.

Result-shape stability

The deterministic-oracle fallback at src/tools/init.ts:268-273 already exists for the "no claude CLI installed" path, so the InitResult shape is unchanged when we skip:

• oracle: { files: 4, llm: false }
• scannersRun: 0, scannersFailed: 0
• Telemetry consumers see this exactly like the no-claude path.

Test plan

• Merge to main
• mkdir /tmp/empty && cd /tmp/empty && axme-code setup — completes in seconds, prints the "Project appears empty" message, .axme-code/ populated with presets
• cd path/to/axme-code && axme-code setup --force — still runs full LLM scan (existing behaviour)
• Existing core test suite (npm test) passes

🤖 Generated with Claude Code

[George-iam]

Reverting on user feedback — the oracle scanner reads more than just local project files: it pulls in ~/.claude/projects/<encoded-path>/memory/ (cross-session auto-memory) plus CLAUDE.md / AGENTS.md / .cursorrules cascade. Skipping the LLM block on empty-looking project folders would silently lose all that accumulated context. Better to keep the slow path than ship a regression. Closing without merge; branch feat/skip-llm-scan-empty-project-20260517 left in place in case we revisit with a more selective skip later.