Repository navigation
GoneBy 1.1.0 — the ctx-audit release
An external comparison (gottz.de/ctx-vs-goneby.html, by the ctx project) catalogued nine ways this project's documentation outpaced its code. Every claim was independently verified against a4e326e, found true, and is closed here with a named regression test. Full claim→fix→test ledger: docs/AUDIT.md. Credit where due — it was a good audit.
Fixed
- Project-scoped memory IDs — identical text in two projects no longer overwrites each other's entries.
- Judge fails open for real — short or empty verdict lists KEEP the unjudged entries, warn, and surface as
judge_fallback_batches(plugin and standalone judge both had the silent-drop bug). - Master capture switch now gates
wall_append(it computed the verdict and discarded it). - One shared WAL default (
~/.goneby) for CLI, MCP server, and Hermes plugin — the split-brain paths are gone;scripts/migrate_wal.sh(dry-run default, sha256-verified, non-destructive) moves existing stores. - Restart-safe recording — message identities persist in WAL metadata; dedup state re-seeds per WAL (builtin
hash()was salted per process and could never survive). - Deterministic SHA-256 dedup + trailing-punctuation merging (Latin & CJK); docs state exactly what merges.
- Qdrant works with its pinned client —
search()→query_points(); live-server testing then exposed a second defect: non-UUID string point IDs are rejected server-side, so IDs now map deterministically (UUIDv5) with the original recoverable from payload. - No provider hardcoded anywhere — the MCP judge is configured via
GONEBY_LLM_*(OpenAI-compatible default via the already-declared httpx; Anthropic as declared opt-in extragoneby[anthropic]). Unconfigured walks now return an actionablestage=judgeerror.
Added
- Position-bound WAL (v2 framing):
v2:<seq>:<base64>with file name + hashed directory token + sequence authenticated into the AEAD tag. Deleted lines report[GAP], reordered lines are excluded from judging, spliced-in lines from other logs fail their tags. Legacy lines stay readable, flaggedintegrity: unbound. Wrong-key reads say "wrong key" via the new.keyidsidecars. Honest limits: a trailing deletion is not detectable without a head-hash chain (documented). - Live integration tests against real Qdrant and Valkey containers (
@pytest.mark.integration, auto-skip when unreachable). - Backward-walk selection-quality harness (
scripts/eval_walk.py+ labeledtests/fixtures/walk_golden.json+docs/BASELINE.md) — the audit's strongest methodological hit was that selection had never been measured; now it can be, reproducibly. - Test suite: 172 → 205 passed, zero warnings, green under both
pytestandpython -m pytest, ruff clean.
Docs
The security/threat-model sections now say what the code actually does: the transcript log is encrypted and position-bound, and the searchable vector store holds plaintext by design (semantic search needs it — protect the store with store-side controls). "Keeps ~10-20%" and scale figures are labeled design targets until docs/BASELINE.md has measured rows.
Migration note: existing Hermes installs keep entries under <hermes-home>/plugin-data/goneby/ until you run scripts/migrate_wal.sh (review dry-run, stop the gateway, --apply). Until migrated, new writes go to the shared ~/.goneby default; GONEBY_WAL_PATH/GONEBY_DATA_DIR override.