v5.3.1 - Security update of the bundled database image

Security update. No code changes, no database migrations and no configuration
changes. Updating is recommended for all installations.

Security

• Updated the bundled PostgreSQL/pgvector image in docker-compose.yml to a
  current release with the latest security fixes.

Upgrade notes

• Pull the new image and recreate the database container:
  docker compose pull && docker compose up -d. Your data stays in the
  mounted data directory (./data/postgres by default).

v5.3.0 - Surgical note editing + mcp SDK v2

Added

• Surgical note editing. New note_edit_string MCP tool replaces an exact
  substring inside a note instead of rewriting the whole content. This
  directly counters the append-culture that grows notes unbounded: a stale line
  can be replaced in place rather than by appending a correction block. Matching
  is literal (not regex) and mirrors an exact-match code editor - it refuses
  when old_string is not found, or matches more than once while replace_all
  is False, forcing the caller to add surrounding context or opt in explicitly.
  Results exceeding the 500 KB content limit are refused, never silently
  truncated. The replacement is computed in Python and written through the
  normal parametrized UPDATE, so there is no SQL-injection surface.

• Search within a note. New note_grep MCP tool returns only the matching
  lines plus surrounding context (with 1-based line numbers, overlapping windows
  merged), letting a caller locate the exact text to feed into note_edit_string
  without pulling a large note into context. The line-number prefix is display
  only and must not be copied into old_string.

• Partial note reads. note_show gained optional start_line / end_line
  parameters to return only a line range of a large note, using the same line
  numbers as note_grep. Without them the full note is returned as before.

  Together these turn a big note from an all-or-nothing load into something a
  session can inspect and maintain a slice at a time - fewer tokens on read, and
  a smaller blast radius than a whole-content rewrite on edit.

Changed

• Upgraded to the mcp Python SDK v2 (~=1.0 -> ~=2.2). In SDK v2 the
  high-level server class FastMCP was renamed to MCPServer and moved to
  mcp.server.mcpserver; the server now imports and instantiates MCPServer.
  The tool and resource surface is unchanged - all tools keep their names,
  parameters and generated schemas, and the stdio transport is unaffected - so
  there is no change for MCP clients. Installing this release pulls SDK v2 and
  its dependencies (notably httpx2, opentelemetry-api and pydantic>=2.12);
  the minimum Python version is unchanged (>=3.10).

v5.2.0 - context_delete + Now item rename

Added

• Delete context entries via MCP. New context_delete MCP tool removes a
  single context entry (action plus its content, metadata and relations) by ID.
  The delete capability already existed in the backend (delete_entry) and CLI
  (cm delete); only the MCP surface was missing, leaving context entries the
  odd entity out next to notes, now-items and infra, which all expose delete.
  The tool is deliberately ID-only - there is no delete-by-query - to prevent a
  single call from wiping many entries. Context entries remain non-editable in
  place by design: to correct a wrong entry, delete it and context_save a
  corrected one, which keeps the entry log a record of what was actually saved
  rather than a silently-rewritten surface. (No context_edit is provided.)
• Rename Now items in place. New cm now edit <id> <title> subcommand and
  now_edit MCP tool. Previously a title could only be changed by deleting the
  item and re-adding it, which silently reset its position, creation time and
  linked-entity reference. Editing is title-only by design (bucket changes go
  through move, status through done/remove) and reuses the same title
  validation as add, so the single-line 200-character limit is enforced the
  same way on both paths.

v5.1.2 - Now titles no longer silently truncated

Fixed

• Now item titles are no longer silently truncated. A title longer than
  the 200-character limit was quietly cut to fit, discarding everything past
  the cutoff. Since Now items have no body field, that text was lost with no
  warning. Over-long titles are now rejected with a clear error that reports
  the actual length and suggests linking a note for long detail. The
  200-character single-line limit itself is unchanged.

Internal

• Removed the dead setup_schema() / setup_fts() methods from the database
  backend (PostgreSQL implementation + abstract declarations). They were never
  called; schema creation runs exclusively through the migrator
  (run_migrations). No behaviour change.

v5.1.1 - Complete schema on fresh installs (plus v5.1.0 Now backlog & infra CRUD)

[5.1.1] - 2026-06-07

Fixed

• Complete schema on fresh installs. On a clean database several objects
  were missing, so todos, MD indexing and semantic search were unavailable out
  of the box, and some epoch-timestamp columns used INTEGER (2038 overflow).
  Databases upgraded from earlier versions already had them and were
  unaffected. Migration 008 reconciles the schema (missing tables, views and
  columns) and widens epoch-timestamp columns to BIGINT. The migration is
  idempotent and safe to run on an existing database.

[5.1.0] - 2026-05-23

"Now" sprint backlog: curated cross-project shortlist with WIP limits;
plus the four missing infrastructure CRUD MCP tools.

Added

• New "Now" sprint backlog: a curated cross-project shortlist of what's
  actively in flight. Three active buckets (today / week / later)
  plus a 24h done holding bucket, each with a configurable WIP limit
  (defaults 7 / 20 / 50). Items can either stand alone or reference an
  existing CM entity (todo, action, note, snippet, ai_instruction, host,
  service); the listing JOINs on the referenced table so callers see
  live status (e.g. a referenced TODO that has since been closed).
  Migration 007 adds the now_items and now_settings tables. CLI
  subcommands: cm now {add,list,show,move,done,remove,reorder,settings}.
  MCP tools: now_add, now_list, now_show, now_move, now_done,
  now_remove, now_reorder, now_settings_get, now_settings_set.
• Four new MCP tools complete the infrastructure CRUD surface:
  infra_edit_host, infra_delete_host, infra_edit_service,
  infra_delete_service. The underlying backend methods already
  existed; only the MCP wrappers were missing, forcing manual SQL or
  CLI fallbacks when entries had to be renamed or removed. Edits
  update only the fields you pass; infra_delete_host requires
  force=True to cascade-delete services on a host that still has
  any.

Fixed

• delete_host raised KeyError: 0 because the result row from the
  RealDictCursor was indexed positionally instead of by column
  name. The bug had been latent since the method was first written
  (no callers existed). The CLI command cm infra delete-host and
  the new infra_delete_host MCP tool both depend on this path.
• setup.sh --install-only could silently leave the cm tool on
  the previous version. Two latent bugs that together hid this since
  the V5 rebrand:
  1. uv tool install --force only replaces the installation from
     uv's wheel cache. With a Hatch dynamic version (read from
     src/version.py at build time) the cached wheel can carry the
     previous version, so --force reinstalls that. Switched to
     --reinstall, which forces a fresh build.
  2. The post-install verification ran the editable .venv/bin/cm,
     which always reports the current source version - so it printed
     success even when the global install was still on the old
     version. The check now targets the global cm on PATH.
     Users who bumped from any V5.0.x release should now see the new
     version after git pull && bash setup.sh --install-only. If you
     were stuck on a prior bump, this is likely why.

v5.0.4 - Embedding worker health monitoring

Embedding worker health monitoring - failures become visible without manual log-tailing.

Added

• Migration 006: new embedding_worker_runs table. Every worker invocation
  records success/failure, duration, and processed count. Retention is
  30 days, trimmed by the worker itself after each successful batch.
• cron_embed.sh now fires a desktop notification on failure (macOS
  osascript, Linux notify-send), deduplicated to once per calendar
  day so the user notices the problem without spam.
• cm save and the context_save MCP tool emit a one-time-per-day
  warning when the worker's latest successful run is older than 48h.
  Claude Code surfaces the MCP variant to the user automatically.
• cm doctor reports embedding coverage percentage, last successful
  run age, and failure count from the last 24h. Stale workers show a
  clear fix command (uv sync --extra embeddings).

Fixed

• The embedding worker previously failed silently when numpy /
  onnxruntime were missing (e.g., after uv sync without --extra embeddings). Failures now surface via notifications, DB records,
  and user-facing warnings on the next save.

Internal

• src/core/embedding_health.py: shared stale-detection helper used
  by both the CLI and the MCP server. Tolerant of missing tables on
  older installs (migration 006 not applied yet).

v5.0.3 - Pinned items + schema hardening

Pinned items (GUI feature integration) and schema hardening.

Added

• New CLI command cm pinned - read-only listing of items curated in
  the optional GUI (context-wolf-ui). Output as grouped Markdown, with
  --project filter and --json output. Returns empty silently when
  GUI tables are missing.
• New optional MCP tool pinned_list - conditionally registered at
  server startup. Only appears in the tool schema when the pinned_items
  table exists (i.e., when the GUI is installed). Keeps the MCP schema
  lean for CLI-only installations.
• Migration 003: pinned_item_projects table for many-to-many project
  scopes on pinned items (an item with no scope rows is global).
• Migration 005: CHECK constraints on ai_instructions.scope and
  priority columns to enforce the canonical vocabulary at the DB level.

Changed

• Renamed feature "Promptstart" to "Pinned" across schema, GUI, CLI,
  MCP tool, and docs. The old name collided semantically with cm ai-prompt
  (different feature: auto session-starter vs. curated items). Fresh
  installs use the new naming; existing installations can upgrade via
  ALTER TABLE (manual, no downtime, no data loss).
• Migration 002 renamed: promptstart_items -> pinned_items with
  corresponding index renames.
• All cm init flows updated to reflect the new naming.

Fixed

• Migration 004: normalized ai_instructions.priority = 'may' rows to
  'nice' (CLI vocabulary), cleaning up drift from an earlier GUI
  implementation that used the RFC 2119 term.

v5.0.2 - MCP dependency fix

Dependency fix to prevent broken cm-mcp installs.

Fixed

• mcp package moved from [project.optional-dependencies.mcp] to core
  dependencies. The cm-mcp entry point is always registered, so a
  plain uv sync (without --extra mcp) would otherwise uninstall
  mcp/pydantic and break the MCP server with
  ModuleNotFoundError: No module named 'pydantic'.

Changed

• Install instructions in README and USER_GUIDE simplified: uv sync
  now installs everything needed for both CLI and MCP server. Only
  --extra embeddings / --extra all remain as opt-in extras.

v5.0.1 - English Translation & Documentation

[5.0.1] - 2026-04-18

English translation, improved Docker setup, documentation, and stability fixes.

Added

• Internationalization: all CLI and MCP output translated to English
• Animated demo GIF in README (save, duplicate detection, search, stats)
• CONTRIBUTING.md for community contributions
• CHANGELOG.md for version history
• README badges (Python, License, MCP, pgvector)
• README section: "Why not Markdown files?"
• README section: privacy and token economy details
• README section: remote database setup
• README section: architecture (two-config layout)
• .env.example template for Docker credentials

Changed

• PostgreSQL upgraded to PG18 with pgvector 0.8.2
• Docker network renamed to contextwolf-net (reusable for optional tools)
• Docker credentials now managed via .env (not hardcoded in compose)
• cm init (Docker mode) reads .env for password and port, auto-creates
  .env if missing; keeps CLI config.yaml and Docker .env in sync
• Setup script waits for PostgreSQL healthcheck before continuing
• Setup script exits cleanly on errors instead of silent failures
• Semantic search setup integrated into main setup.sh flow
• Model size references unified across documentation (~90 MB)
• MCP server migrated to FastMCP for better tool handling
• Various documentation rewrites for clarity

Fixed

• cm doctor: correct Python version check (3.12+, was 3.10+)
• cm infra: documented valid --location values (local, extern)
• Bare except: clauses replaced with specific exception handling
• Removed broken V2-V4 legacy tests and references
• Removed em-dashes from documentation (AI boilerplate cleanup)

Internal

• Code cleanup: removed unused token_tracking feature
• Improved FILE_TYPE_MAP for code snippet detection
• Dependency updates

[5.0.0] - 2026-04-03

Initial public release. Rebranded from "Context Manager" to "ContextWolf".

Added

• Python package structure with pyproject.toml and uv support
• Interactive setup script (setup.sh) with database wizard and MCP configuration
• cm doctor diagnostics command
• Demo GIF in README
• CONTRIBUTING.md
• MIT License

Changed

• All CLI and MCP output strings translated to English
• Clean Architecture: 4-layer structure (CLI, Features, Domain, Core)
• MCP server migrated from low-level API to FastMCP
• Duplicate detection with similarity warnings at 85%+

Removed

• SQLite backend (PostgreSQL only)
• Legacy V2/V4 code and tests

Technical

• 31 MCP tools with deferred loading (~356 tokens overhead)
• pgvector hybrid search (full-text + vector similarity)
• Local ONNX embeddings (all-MiniLM-L6-v2, 86MB)
• Docker-hardened PostgreSQL with pgvector

[4.3.0] - 2026-03

Added

• MCP tool usage tracking and analytics (cm tool-stats)

[4.2.0] - 2026-02

Added

• Universal search across all content types (cm smart-search)
• Advanced AI instruction search with filters

[4.1.0] - 2026-01

Added

• Session-aware AI prompt generation (smart mode)

[4.0.0] - 2025-12

Changed

• Migrated from SQLite to PostgreSQL backend
• Added pgvector for semantic search