Beta → 1.0 hardening + data-loss repair (24 hardening items + C1-C6)

[senna-lang]

Implements three openspec changes: reliability/safety/performance hardening toward the 1.0.0 release, repair of an ongoing distillation data-loss incident, and a rewrite of the loci prime injected prompt to drive agent-initiated memory usage.

release-1.0-hardening — 24/24

• B1 WAL + busy_timeout / B2 user_version-based migration mechanism / B3 transactional distillation / B4 atomic settings.json writes + .bak backup / B5 meta table (model/prompt versions) + drift warnings
• H1 distill_status (pending/skipped/distilled) / H2 flock-based distill lock / H3 server double-start prevention / H4 search connection leak fix / H5 full socket reads / H6 .jsonl cleanup on claude timeout / H7 loci hook uninstall / H8 memory.db 0o600
• P1 FK indexes / P2 sequential indexer reads / P3 tree-sitter cache
• Q1 sha256 consolidation / Q2 datetime.now(UTC) / Q3 error reporting in distill_all / Q4 assert→raise / Q5 finer-grained config exceptions / Q6 constant naming / Q7 llm.py tests / Q8 tobytes()

repair-distill-and-revive-context — 6/6

• C1 collect file paths from tool_use (exchange_files) / C2 many-to-many symbols / C3 repair migration (reset lost distillations to pending + orphan cleanup + drop bm25_text) / C4 drop INSERT OR IGNORE, verify persistence / C5 lighter loci context output + --full / C6 link only symbols appearing in conversation bodies

improve-prime-injection — P1-P3

Rewrites the prompt that loci prime injects at session start, so agents actually reach for loci search / loci context on their own (real-world logs showed near-zero spontaneous usage):

• P1 Triggers rewritten from user-question-driven ("when asked where X is") to agent-action-driven: before editing/refactoring a function, before starting a new implementation, on encountering a known error
• P2 loci context (reverse lookup) promoted to an independent section alongside loci search, with the design intent spelled out: touching a symbol = recalling memory about that symbol
• P3 Concrete, non-placeholder command examples embedded for both commands (loci search "BM25 RRF fusion ranking", loci context --symbol "SymbolResolver.extract")
• CLAUDE_MD_SECTION collapsed to a minimal redirect — PRIME_TEXT is now the single source of behavioral guidance, removing the duplicated-and-drifting wording between the two
• New tests/test_prime_cmd.py: 8 tests covering the PRIME_TEXT contract and inject_claude_md idempotency
• P4 (when-not-to-search guidance) deliberately deferred

Verification

• All CI gates green: pytest 226 passed / pyright 0 errors / ruff clean
• C3 migration applied to real DBs (logosyncs / arxiv-newspaper): lost distillations restored to pending, orphaned rooms/symbols removed, legacy bm25_text column dropped

Notes

• The 1.0.0 tag is not cut yet
• Re-distilling exchanges reset to pending (claude API cost) is left to each project's hook

🤖 Generated with Claude Code

[claude]

Code review

Two bugs found and validated.

---

Bug 1: Ply-skip boundary mismatch after malformed JSON lines

File: src/codeatrium/indexer.py lines 196-204

Codeatrium/src/codeatrium/indexer.py

Lines 195 to 205 in a4b7425

	continue
	if ply <= last_ply_end:
	raw_entries.append(None)
	ply += 1
	else:
	try:
	raw_entries.append(json.loads(line))
	ply += 1
	except json.JSONDecodeError:
	continue

The skip loop increments ply for every non-empty line unconditionally, including malformed JSON. However, last_ply_end was written by the old parser which only advanced its counter on successfully-parsed entries — malformed lines hit except json.JSONDecodeError: continue and were never appended or counted. The two values live in different coordinate systems.

Impact: For any existing DB whose already-indexed region contains malformed JSON lines, the skip boundary fires one position too early per bad line, pushing a previously-indexed entry back into the parse region. The newly-stored ply_end then diverges further on each incremental run, causing progressive boundary drift.

Suggested fix: Only advance ply in the skip region when the line parses successfully, so both regions use the same coordinate system as the stored last_ply_end:

```python
if ply <= last_ply_end:
try:
raw_entries.append(json.loads(line))
except json.JSONDecodeError:
continue
ply += 1
```

---

Bug 2: WAL sidecar files not chmod 0o600 — sensitive data potentially world-readable

File: src/codeatrium/db.py lines 197-207

Codeatrium/src/codeatrium/db.py

Lines 196 to 208 in a4b7425

	con.enable_load_extension(False)
	con.execute("PRAGMA journal_mode=WAL")
	con.execute("PRAGMA busy_timeout=10000")
	con.row_factory = sqlite3.Row
	return con
	def init_db(db_path: Path) -> None:
	"""DB を初期化してスキーマを作成する（冪等）"""
	db_path.parent.mkdir(parents=True, exist_ok=True)
	con = get_connection(db_path)
	os.chmod(db_path, 0o600)

`init_db` calls `os.chmod(db_path, 0o600)` only on the main `memory.db` file, but `get_connection` enables WAL mode (`PRAGMA journal_mode=WAL`), which causes SQLite to create `memory.db-wal` and `memory.db-shm` sidecar files containing the same verbatim conversation and code data. These sidecar files inherit the process umask (commonly `0o644`), leaving them readable by other local users on a shared system.

Suggested fix:

```python
os.chmod(db_path, 0o600)
for suffix in ("-wal", "-shm"):
sidecar = db_path.parent / (db_path.name + suffix)
if sidecar.exists():
os.chmod(sidecar, 0o600)
```