v2.3.18

Fixed — bound MCP responses so heavy load can't crash the stdio pipe

Under heavy recall usage the MCP could disconnect mid-session (the whole Claude
Code session lost the tool). Root cause: search_messages returned each message's
full text verbatim via _row_to_hit; a single ~172k-char tool-result blob (times
limit rows) produced a JSON-RPC response large enough to break the stdio pipe.

• _row_to_hit caps bulky free-text fields (text/content/value/preview/
  body/snippet) at 1500 chars with <key>_truncated + <key>_full_chars markers
  — one chokepoint bounds every recall/message tool.
• search_messages clamps limit to ≤50 and applies a 256 KB total-response
  backstop that trims rows and appends a non-silent _meta.
• Internal Python callers bypass _row_to_hit, so they keep full text — no logic change.
• New tests/test_mcp_payload_bounds.py pins the bounds.

v2.3.0

Fixed

• LLM refinement crash: '<' not supported between instances of 'NoneType' and 'int' — sort keys in cmd_rebuild.py now use or 0 to handle SQLite NULL frequency/message_count values.

Changed

• fastembed>=0.4 and sqlite-vec>=0.1.6 moved from optional [vec] extra to core dependencies — vec/hybrid search works out of the box with no extra install step.
• [vec] and [llm] extras emptied (kept for backwards compatibility); ollama is an external daemon, not a Python package dep.
• Dev dep floors raised: pytest>=9, pytest-cov>=7, ruff>=0.15, mypy>=2, build>=1.5, twine>=6.
• Removed Operating System :: Microsoft :: Windows classifier (hooks are bash-only; no Windows CI).

v2.2.0

Added

• Goose CLI source adapter (lib/sources/goose.py): reads sessions from SQLite DB at ~/.local/share/goose/sessions/sessions.db. Supports 10 CLI clients total.
• Grok CLI source adapter (lib/sources/grok.py): reads JSONL chat history from ~/.grok/sessions/. URL-decodes CWD from directory names.
• docs/install/goose.md, docs/install/grok.md: install guides for new adapters.
• Support matrix updated: 10 clients (was 8).

Changed

• hooks/lib/common.sh: GPU-aware ollama resolver. Bundled binary takes priority over system PATH. RECALL_OLLAMA env override bypasses GPU check entirely.
• MCP: added title= to all 26 tools; annotated all with readOnlyHint=True via ToolAnnotations; bumped mcp>=1.28.0 SDK floor.
• CI: bumped checkout@v7.0.0, setup-python@v6.2.0, Python 3.13 across all workflows; added workflow_dispatch trigger.

v2.1.0

v2.1.0 — gte embeddings, worktree pooling (schema v5), compaction con…

v2.0.1

v2.0.1 — dependabot CI action bumps

v2.0.0

fix(v2.0.0): complete the vec backfill wiring + finish version bumps

The v2.0.0 commit (e7578b1) shipped INCOMPLETE: the _backfill_vectors edit
failed on a stale anchor so the function was never added (rebuild couldn't
import it), and only __init__.py got bumped — plugin.json / marketplace /
pyproject / uv.lock stayed at 1.6.0, failing the version-consistency gate.

This finishes the job:
- total_recall/cmd_rebuild.py: add _backfill_vectors() and call it after the
  LLM-refinement block in rebuild_cmd (cold path; TOTAL_RECALL_VEC gate;
  optional + non-fatal).
- bump plugin.json / marketplace-entry.json / pyproject.toml / uv.lock to 2.0.0
  so all 5 version sources agree.

Gate (with this applied):
- unit: 1320 passed, 15 skipped (exit 0) — version-consistency tests now green
- vec eval: 2 passed (rebuild->embed e2e + FTS5 0.133 -> hybrid 0.933, +0.800)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

v1.6.0

v1.6.0 — vec hybrid eval harness (+0.800 P@5 measured)

v1.5.0

v1.5.0 — near-duplicate dedup + score floor

v1.4.1

v1.4.1 — actually wire composite re-rank + tests

v1.4.0

v1.4.0 — composite re-rank (recency + kind priority)