Release v0.11.0

What's Changed

Added

• Apple Reminders agent tools (reminders_status, reminders_lists, reminders_show, reminders_add, reminders_edit, reminders_complete, reminders_delete) in agentforge/tools/reminders_tools.py
• remindctl backend (brew install steipete/tap/remindctl) with osascript / AppleScript fallback when remindctl is absent
• Split-deploy support: reminders tools register on every worker so remote agents pass has_tool() and cross-dispatch to the Mac local worker. Execution uses EventKit on Darwin only
• Server-side due-date normalization (tomorrow, today, tomorrow 09:00) and rejection of past absolute ISO dates
• Title-to-ID resolution for reminders_delete and reminders_complete (exact title match against open reminders)
• Agent prompt rule and per-turn reminder-query suffix in _run_agent to steer models toward reminders_* tools
• tests/test_reminders_tools.py: registration, remindctl arg building, date validation, title resolution
• @notes custom agent example documents Apple Reminders alongside kb_search (markdown/custom-agents/notes.md)

Changed

• @agent and @pipeline base tool sets include all reminders_* tools (macOS execution and omitted from tool specs on non-Darwin workers)
• custom_agents.example.yaml and markdown/custom-agents/notes.md expanded with Reminders tool guidance and scheduler/monitor boundaries

Full changelog: https://github.com/bulletinmybeard/agent-forge/blob/master/CHANGELOG.md

Release v0.10.0

What's Changed

Added

• Attachment file storage: HEAD/GET/POST /knowledge/entries/{id}/file for checking, downloading, and uploading original attachment files (e.g., PDFs stored alongside their extracted text)
• KnowledgeFileService: filesystem-backed store for original attachments with configurable size limit
• Multi-collection routing via X-Knowledge-Collection header: the KB SPA uses knowledge_entries (default), AgentForge Notes uses kb_note_entries
• knowledge_registry module: collection-scoped service caching, FastAPI dependency injection, and ensure_all_collections() startup hook
• force_unique flag on CreateEntryRequest. Bypasses content-hash dedup so the same text can be stored under multiple parents
• Auto-relink: creating a duplicate entry with a parent_id reattaches the existing entry under the new parent instead of returning 409
• metadata field on UpdateEntryRequest (was already on create and now on updatable)
• Config keys: knowledge.notes_collection_name, knowledge.files_dir, knowledge.max_attachment_bytes
• @kb and @notes custom agent examples in custom_agents.example.yaml
• markdown/custom-agents/notes.md system prompt for the Notes agent
• Notes-aware WebSocket session routing: source=notes sessions auto-scope kb_search to the notes collection
• extracted_bytes in /knowledge/extract response metadata
• pdftotext page-marker formatting (--- Page N ---) and scaled timeouts for large PDFs

Changed

• Knowledge API routes use FastAPI Depends() injection instead of a module-level singleton. Each request resolves the correct collection-scoped service
• KnowledgeVectorService accepts an optional collection_name parameter (defaults to settings.knowledge.collection_name)
• PDF extraction prefers pdftotext for large files (>5 MiB) and pdfplumber for smaller ones (was pdfplumber-first with pdftotext fallback)
• delete_entry also removes stored attachment files for the entry (*optionally)

Full changelog: https://github.com/bulletinmybeard/agent-forge/blob/master/CHANGELOG.md

Release v0.9.0

What's Changed

Changed

• RAG @ aliases centralized in agentforge/mode_prefixes.py (shared by mode_routing and intent_classifier).
• Botty engine reads analysis_interval, max_frequency_seconds, and dismissal_cooldown_seconds from config.
• canvas.enabled and botty.enabled in config.yaml now gate Canvas init and the /ws/botty route (defaults remain true).
• prompt_lab.enabled gates Prompt Lab DB init and /api/prompt-lab/* endpoints; canvas.enabled also gates the /api/canvas/* router (not just init).
• Chat sessions are namespaced by chat_sessions.source (web, kb, …): clients pass overrides.source and/or ?source= on /ws/chat; worker auto-create reads the active job's overrides. Agent Chat lists source=web only.
• web/server/api.py imports hoisted to module top; sql_schema_tool stays lazy (private/gitignored module).
• In-code RAG comments in ws_endpoint.py now treat @qdrant as canonical (@docs / @find as aliases).
• OllamaSettings profile resolution always delegates to agentforge.config.ConfigManager (removed duplicate _merge_profile_chain fallback).
• Config loading consolidated: app/config.py and agentforge.config.ConfigManager both use load_merged_yaml() (framework-config + config.yaml + split profiles). ConfigManager.raw exposes the merged dict. When gitignored config files are absent, load_merged_yaml() falls back to the committed *.example.yaml templates (CI / fresh clone).
• Legacy per-product Google connector plugins (gmail, google_drive, bigquery, youtube) removed; unified google connector only. Unmigrated SQLite rows are skipped at startup (see scripts/list-legacy-connections.py).
• Knowledge Database content types are now note, reference, documentation, document, cheatsheet, and snippet (replacing the earlier code / command / url / config / error_solution / api_example set). Update clients and any indexed entries accordingly.
• custom_agents.yaml is gitignored (per-deployment, like config.yaml); shipped template is custom_agents.example.yaml with example fallback in the loader.
• RAG search mode: @qdrant is canonical; @docs and @find are documented aliases (all three route to the same mode and can appear anywhere in a prompt).
• Mode prefix detection extracted from ws_endpoint.py to web/server/mode_routing.py.
• Put.io / Premiumize (cloud_tools) moved from register_core_tools to register_optional_tools (still registered by register_all_tools when credentials are set).

Added

• GET /knowledge/list: slim metadata listing for the browse view (no content body; optional limit, default 2000).
• custom_agents.example.yaml template for custom agents (copy to gitignored custom_agents.yaml).
• scripts/list-legacy-connections.py: read-only audit of legacy per-product Google connector rows.
• tests/test_mode_routing.py: prefix stripping for @qdrant / @docs / @find.
• tests/test_config_loader.py: merged YAML parity between app.config and agentforge.config.
• tests/test_feature_flags.py: default canvas.enabled / botty.enabled settings.
• tests/test_mode_prefixes.py, tests/test_botty_engine.py: shared RAG aliases and Botty rate limits.
• web/server/session_source.py, tests/test_session_source.py: shared session source resolution for WS and worker paths.

Removed

• get_connector_config() (unused after unified Google OAuth cleanup).
• strip_agent_prefix() (unused after mode-routing extraction).
• translate_legacy_locality() (empty map; inlined at call sites).
• tools.shell.sudo_password startup warning in the CLI (the key was already ignored; interactive sudo is the only path).
• AGENTFORGE_WORKER_LOCALITY env fallback in worker role resolution (use AGENTFORGE_WORKER_ROLE).
• Legacy connectors.google.gmail.credentials_dir config path for OAuth client secrets (use connectors.credentials_dir or GMAIL_CREDENTIALS_DIR).

Breaking

• Knowledge Database content type rename (see Changed above).
• Legacy per-product Google connector plugins removed (see Changed above).
• Chat session listing is scoped by source; clients must pass overrides.source / ?source= where appropriate.

Full changelog: https://github.com/bulletinmybeard/agent-forge/blob/master/CHANGELOG.md

Release v0.8.0

What's Changed

Added

• Knowledge Database (/knowledge/* on agentforge-api): a personal store for user-created entries (code, command, url, config, error_solution, note, api_example) in its own Qdrant collection (knowledge_entries), separate from the RAG index. CRUD, bulk create/delete, and semantic search with tag/type/project filters, tag faceting, and stats
• Smart re-indexing on update: re-embeds only when title, content, or notes change; metadata-only edits skip the embedding call
• Parent/child attachments via parent_id, with per-page chunking so a parent entry and its attached documents are searchable as passages
• POST /knowledge/filter: list entries by metadata filters (incl. parent_id) without a vector search
• POST /knowledge/entries/{id}/context: most relevant passages from one entry for a query, plus adjacent pages for context
• POST /knowledge/entries/{id}/rechunk: rebuild page chunks for entries indexed before chunking existed
• POST /knowledge/extract: server-side text extraction from uploaded files (PDF via pdfplumber, pdftotext fallback; text/code/config decoded as UTF-8), reusing AgentForge's extraction path instead of frontend JS
• metadata free-form field on all knowledge points (request + response)
• knowledge config block: collection_name, dedup_threshold, composite_template (env prefix KNOWLEDGE_)
• SAQ batch job for bulk knowledge ingestion

Full changelog: https://github.com/bulletinmybeard/agent-forge/blob/master/CHANGELOG.md

Release v0.7.0

What's Changed

Added

• Diff-preview confirmation gate for write_file and append_file tools -- shows a unified diff and requires user approval before writing, matching the existing code_edit flow
• Agent-level destructive shell/ssh guard: sed -i, perl -pi, and other in-place edits are intercepted via run_confirm() before execution
• sed -i / perl -pi patterns added to CommandGuard's deterministic _DESTRUCTIVE_PATTERNS regex
• sed -i / perl -pi added to the LLM classification prompt (command_guard.md) as explicit destructive examples
• skip_confirm parameter on _dispatch_tool() for callers that already ran confirmation at the agent level

Fixed

• Cross-role dispatch now respects skip_confirm separately from internal, preventing double-confirmation on pre-gated tools

Full changelog: https://github.com/bulletinmybeard/agent-forge/blob/master/CHANGELOG.md

Release v0.6.0

What's Changed

Added

• GitHub connector (@github / @gh): token (PAT) auth, github.com-only. Reuses the gh CLI by injecting each connection's PAT as GH_TOKEN rather than reimplementing the REST API; read-only connections apply a best-effort allowlist of read gh subcommands
• Per-connection read/write toggle for token connectors (GitLab, GitHub) via PATCH /api/connectors/{id}, re-registering the agent so its tools and prompt reflect the mode
• Answer bookmarks: command_notes extended with kind + content, so agent answers can be saved, not just tool calls
• source_branch / target_branch filters on gitlab_merge_requests
• needs_url / default_url on connector types, so SaaS-only connectors (GitHub) omit the URL field in the connect form

Changed

• Token connections are verified with a live API call before they are saved so invalid tokens or hosts now fail immediately instead of creating a potential dead connection
• /api/agents returns a source field so clients can distinguish connector connections from built-in and custom modes
• Connection listings expose read_write
• Expanded user_context.example.md with richer explanations

Fixed

• The name-greeting parser tolerates bullet variations in user_context.md

Removed

• The browser-location feature: location_service.py, the /api/location endpoint, and the location prompt-injection

Full changelog: https://github.com/bulletinmybeard/agent-forge/blob/master/CHANGELOG.md

Release v0.5.0

What's Changed

First public-ready release: cleanup and hardening, a unified Google connector, and the GitLab toolset reintroduced.

Added

• GET /api/tools endpoint listing every registered runtime tool (name, description, category)
• YouTube connector tools: youtube_search, youtube_video_details, youtube_channel_details, youtube_playlist_items, youtube_my_subscriptions
• Full GitLab connector toolset: 28 gitlab_* tools covering projects, branches, merge requests, pipelines, jobs, runners, and users, with a per-connection read/write toggle and confirmation gates on destructive actions
• Configurable per-request history character limit, so clients that send large payloads (e.g., the AskPage DOM) aren't truncated
• Botty semantic search over chat session titles and message content, wired into run calls

Changed

• Unified the Gmail, Google Drive, and BigQuery connectors into a single Google connector with per-connection product selection and a shared OAuth client
• Derive better connector labels automatically (e.g., the local part of an account email)
• Corrected search-provider settings precedence between environment variables and YAML config
• TMDB: accept a v4 API Read Access Token (sent as Authorization: Bearer) alongside the v3 API key
• Refreshed the hardcoded modes info returned by the /agent endpoint
• Improved sticky-mode handling for the dry-run feature
• Made the GET uploads endpoint a catch-all (dropped session_id from the path)

Removed

• Deprecated and outdated configs, files, and logic

Breaking

• Google connector config moved from the per-product connectors.gmail, connectors.google_drive, and connectors.big_query blocks to a single connectors.google block. Update config.yaml accordingly (see config.example.yaml).

Full changelog: https://github.com/bulletinmybeard/agent-forge/blob/master/CHANGELOG.md