feat(cli): harden and polish loci init command

[senna-lang]

Summary

This PR hardens loci init across multiple dimensions — UI, prompt robustness, resilience to failures, automatic hook installation, and better error messages for environment-level problems. Re-running loci init after a failure is now safe.

Changes

UI: richer startup banner

• Renders codeatrium in the pagga half-block figlet font with a blue vertical gradient (#7bb8ff → #1b45a8), ⦿ mascot glyph, tagline, and version — framed by a rich Panel. Fits in 80 cols without wrapping.

Interactive prompts: no more silent fallback

Previously, invalid prompt input silently defaulted — "y" for Yes was treated as No because only "2" worked. Now:

• Re-prompt on invalid input for all 4 interactive prompts (min_chars, skip-count choice, custom counts, distill priority).
• yes/no aliases: 1/2/y/n/yes/no (case/whitespace tolerant) for the "run distillation now?" prompt.
• Range validation: custom exchange count must be 1..total; custom min_chars must be >= 0.

Resilience: safe failure & interrupt

• Execution-phase try/except: if anything fails mid-init, .codeatrium/ is removed (shutil.rmtree) so re-running loci init is safe. Previously the half-created DB was picked up by the Already initialized fast-path and blocked retry.
• KeyboardInterrupt: exits cleanly with code 130 and the same cleanup.
• Per-file resilience: one bad .jsonl no longer aborts the indexing loop — it logs ⚠ skip <file>: <err> and continues.
• git_root catches FileNotFoundError: no crash when git isn't on $PATH.
• parse_exchanges returns [] for missing files instead of raising.
• find_project_root no longer traverses parents when outside a git repo (prevents picking up an unrelated .codeatrium/).

Automation: hooks auto-installed

• loci init now calls install_hooks at the end (writes to ~/.claude/settings.json). loci init + loci hook install → just loci init.
• Opt out with --no-hooks.
• Hook install failure is warn-only; DB is preserved and user can retry with loci hook install.

Embedder fail-fast with remediation hints

Previously a broken sentence_transformers import (common cause: numpy 2.x + older pyarrow) produced a full traceback followed by 242 per-row error lines. Now:

• New EmbedderSetupError distinguishes environment-level failures from per-row errors.
• _ensure_model wraps import/init failures with a remediation-aware message (pip install 'numpy<2' or pip install -U pyarrow).
• distill_all re-raises EmbedderSetupError to stop the loop on first occurrence.
• loci init's distill handler catches it specifically and prints a single clean message — no traceback, no per-row spam.

Test plan

• Banner renders with pagga characters and subtitle
• Invalid prompt input re-prompts with "Invalid choice" hint
• y / n aliases accepted; custom count range-validated
• Execution-phase failure triggers .codeatrium/ cleanup
• KeyboardInterrupt exits 130 with cleanup
• Per-file index failure doesn't abort init
• Hooks installed by default
• --no-hooks skips install
• Hook install failure warn-only, DB preserved
• EmbedderSetupError wraps import failure
• init shows friendly message on EmbedderSetupError, preserves DB
• All 142 tests pass (pytest, ruff check, pyright clean)

🤖 Generated with Claude Code

[claude]

Code Review

One issue found.

Bug: failure-cleanup can delete pre-existing .codeatrium/ contents

File: src/codeatrium/cli/__init__.py, lines 121-124

The guard at line 88 only checks db.exists() (memory.db), not whether .codeatrium/ itself pre-existed. Meanwhile, the config-write at line 127 intentionally skips overwriting a user-authored config.toml but both exception handlers at lines 194 and 199 unconditionally rmtree the entire directory, including any files init did not create.

Concrete scenario: a user pre-creates .codeatrium/config.toml with custom settings before running loci init. The db.exists() guard is bypassed (no memory.db yet), line 127 correctly skips overwriting their config but a mid-init failure (Ctrl-C or disk error) calls rmtree on the whole directory, destroying that file.

The inconsistency is clear: line 127 goes out of its way not to overwrite a pre-existing config.toml, yet the cleanup at lines 194/199 destroys it anyway.

Suggested fix: capture whether .codeatrium/ pre-existed before the try block and guard the cleanup with if not dir_preexisted: shutil.rmtree(...).

[senna-lang]

@claude
fix

[claude]

Claude finished @senna-lang's task in 1m 46s —— View job

---

Fixed: pre-existing .codeatrium/ preserved on failure cleanup

• Gather context and understand the request
• Read src/codeatrium/cli/__init__.py — found the issue at lines 122, 194, 199
• Implement fix: capture dir_preexisted before the try block, guard rmtree with it
• Add regression test test_init_preserves_preexisting_dir_on_failure
• Commit and push changes

What changed:

src/codeatrium/cli/__init__.py — one line added before the try block:

dir_preexisted = codeatrium_dir.exists()

Both shutil.rmtree calls in the except handlers now check if not dir_preexisted: before removing the directory. If .codeatrium/ existed before loci init ran, it is left untouched on failure — consistent with the existing logic that already skips overwriting a pre-existing config.toml.

tests/test_init.py — new regression test test_init_preserves_preexisting_dir_on_failure: pre-creates .codeatrium/config.toml with custom content, injects a mid-init failure, then asserts the directory and file survive intact.
| Branch: fix/init-hardening