v0.6.0 — pin gbrain v0.54.1.1

v0.6.0 — pin gbrain v0.54.1.1

Engine pin lineage (continuing from v0.5.0)

• v0.4.0: mnemosyne-memory 3.15.1 → 4.0.0b3 — recall provenance logging (upstream PR #957, shipped in Mnemosyne v0.7.3)
• v0.5.0: release hardening — health-gated start, liveness vs auth doctor split
• v0.6.0 (this release): gbrain pin v0.50.0.0 → v0.54.1.1

What v0.54.1.1 carries over v0.50.0.0

• 0.50.5 security hardening wave
• 0.51.0 durable concurrent writes
• 0.51.7 search fixes
• 0.52–0.54 repair and preview UX

Note: upstream gbrain has since published newer releases; this pin tracks the version validated in a 24h dogfood of the live stack. A fast-follow bump is expected.

Also shipped from dev

• Prompt-injection CI gate: scripts/check_prompt_injection.py (stdlib, in-repo, auditable — hidden Unicode/BIDI, homoglyphs, LLM delimiters, injection/exfil patterns, base64 blobs in markdown; --self-test fixtures). Required check on every PR and pushes to main; docs-only PRs skip the heavy code gates via a changes-detection job.
• Fix: hooks recall uses top_k (was limit)
• Docs: documented deliberate divergence from upstream capture policy

Full Changelog: v0.5.0...v0.6.0

MnemoBrain v0.5.0

MnemoBrain v0.5.0

Reliability release, driven by real-world breakage reported in issue #6 (Windows/WSL: gbrain serve can look dead to Cursor MCP).

Fixed

• Health-gated start — mnemobrain start gbrain no longer trusts a live pid over a dead listener. Before claiming "already running" it GETs /health; a wedged process is stopped and restarted; if the service is still unhealthy 60s after spawn, start exits non-zero and points at the log file instead of printing a green lie.
• Liveness vs auth split in doctor — new mcp-auth check: a minimal MCP initialize POST with the configured Bearer (env GBRAIN_ADMIN_BOOTSTRAP_TOKEN or <home>/gbrain-admin.token). Verdicts: PASS (200) / FAIL (401 invalid_token → "process alive but token rejected") / SKIP (no token configured, or listener down — liveness is the real problem then). Decision tree credit: @effemmeffe (issue #6).
• Windows/WSL troubleshooting doc — symptom → actual state table + decision flow in docs/OPERATIONS.md, including the known gap that gbrain.log lines carry no timestamps.
• doctor probe reports its real version — MCP clientInfo now comes from the package __version__ instead of a hardcoded string that had already drifted.

Compatibility

No pin changes (mnemosyne-memory 4.0.0b3, GBrain v0.50.0.0), no env-contract changes. Existing installs: pull and re-run mnemobrain install only if you vendor the CLI; nothing else to do.

v0.4.0 — mnemosyne-memory 4.0.0b3 (recall provenance logging)

MnemoBrain v0.4.0

Upstream engine upgrade: mnemosyne-memory 3.15.1 → 4.0.0b3.

Why 4.0.0b3

Upstream Mnemosyne shipped v0.7.3 (2026-09-23), the first release containing recall provenance logging (mnemosyne-oss/mnemosyne#957, merged 2026-09-21). The matching core artifact on PyPI is mnemosyne-memory 4.0.0b3 — no stable 4.0.0 exists yet, so we pin the beta rather than wait.

What the provenance feature gives MnemoBrain users

• Every recall injection is now logged: one JSONL line per turn at <db>.recall_provenance.jsonl with query, injected results, scores, and top_k.
• Enables mnemosyne_provenance recall of what the agent actually saw when it answered.
• Install note: install verifies bun >= 1.3.11; pip-installs mnemosyne-memory==4.0.0b3 into the venv; bun add github:garrytan/gbrain#v0.50.0.0 --ignore-scripts.

Changes

• Bump default MNEMOBRAIN_MNEMOSYNE_VERSION pin 3.15.1 → 4.0.0b3 (src/mnemobrain/config.py, docs/CONFIG.md, SKILL.md)
• Bump MnemoBrain version 0.3.0 → 0.4.0 (pyproject.toml, src/mnemobrain/__init__.py)
• Test suite updated for the new pin

Upgrade

git pull
./.venv/bin/python -m pip install -e . && ./.venv/bin/mnemobrain install
./.venv/bin/mnemobrain doctor   # should PASS on mnemosyne line

Full Changelog: v0.3.0...v0.4.0