Changelog¶

v0.3.21 — Reasoning Loop (2026-04-06)¶

Reasoning loop engine — autonomous plan-execute-evaluate cycle
LLM decomposes goals into multi-step plans, executes whitelisted skills, evaluates results, and iterates
Configurable LLM tier (default: local) — route to stronger models via reasoning.llm_tier
Deterministic decision pre-filter skips LLM for obvious cases (0ms decisions)
Deterministic plan validation rejects malformed steps before execution
Working memory compression every 3 iterations
Evaluation score trend (improving/stable/degrading) in decision prompt
Proportional time budget (~300s per step, scales with plan size)
User intervention: pause, resume, steer, abort, step override (real-time via PubSub)
Skill outputs embedded to pgvector for future session context
Orphaned session cleanup on process death, app restart, and page load
Full audit trail: every prompt, response, skill call, rubric score, and working memory snapshot persisted
Chat page — dual-mode interface (Chat / Reasoning toggle) with plan view, step timeline, and intervention controls
Skill descriptions — clarified web_search (snippets only), web_fetch (full page content), web_search_fetch (search + full content)

RAG pipeline overhaul — embedding metadata, relevance grading, query rewriting, semantic chunking, fallback routing
Research skill — cross-store RAG with rewriting and relevance grading
GitHub Security Review — refactored as pure diff fetcher; 5 modes
Workflow step editor fixes — save no longer closes editor; nil llm_tier fix
LLM Transform — 10 prompt presets (Security Review, Code Review, Changelog, etc.)

Services page — new /services admin page showing external service status with real connectivity checks
Database — verifies PostgreSQL connectivity via SELECT 1
Google API — checks OAuth2 token status (connected/expired/not configured)
Telegram Bot — sends a real test message to the configured chat
Discord Bot — sends a real test message to the configured channel
2FA (TOTP) — triggers a challenge via Telegram, auto-updates via PubSub on code verification
Ollama — queries /api/tags, reports loaded models
LM Studio — queries /v1/models, reports loaded models
GitHub API — authenticates with stored PAT, reports username
Web Automator — checks /status on the browser sidecar
Config seeder fix — env-backed settings no longer overwrite DB values on boot; Config page is now the sole source of truth after first seed
Dashboard cleanup — removed Google status card from dashboard (moved to Services), node name moved to dashboard header next to version
Nav bar — added Services menu item, reduced spacing between AlexClaw title and menu links

Forge page (pre-alpha) — interactive skill generation with two-column UI (chat + code output), auto-iterate with configurable retries, real-time status, structural validation for external skills
Chat simplified — stripped RAG/knowledge search, now a clean conversational chat with model selection
CodeGenerator — shared skill generation module extracted from Coder skill, reusable by both Forge UI and Coder workflow skill
Executor timeout from config — timeout_ms in step config JSON overrides the 30s default SafeExecutor timeout
Scraper improvements — all 5 knowledge scrapers now support timeout_ms, delay between items, deadline-based execution, and detailed reporting (stored/skipped/failed/timeout per item)
HexDocs guides scraper — new skill scraping guide/extra pages (README, getting started, deployment docs) — 649 guide chunks indexed
Skill UI feedback — Reload/Unload/Upload buttons show "Waiting 2FA..." with pulse animation during 2FA challenge
Workflow runs counter — now refreshes automatically when a run completes
Browser User-Agent — all SkillAPI HTTP calls now include a default browser User-Agent header to prevent site blocking
Convention fixes — 164 violations reduced to 19 (all intentional process_dictionary usage)
Skill template — updated with external/0, step_fields/0, config_hint/0, config_scaffold/0 documentation

Workflow export — self-contained JSON files with definition, steps, and full resource data
Workflow import — file upload in Admin UI, resources matched by name+URL or created automatically, disabled by default with (imported N) suffix
Workflow name filter — search/filter the workflow list by name
Action buttons — workflow row actions restyled as colored pill buttons
Bug fix — duplicate_workflow now copies input_from and routes fields
Docker naming — services renamed to alexclaw-prod, db-prod, db-test for clarity
Makefile — quiet test builds, auto-teardown after tests, test-down target
Dynamic skill metadata — skills declare their own UI fields via 7 new optional callbacks (step_fields, config_hint, config_scaffold, config_presets, prompt_presets, config_help, prompt_help). Step editor renders dynamically — zero hardcoded skill knowledge in the LiveView
Docs — README, INSTALLATION, architecture, writing-skills, and readthedocs pages updated

New: Model Context Protocol integration

MCP server exposing all skills and workflows as tools via Streamable HTTP transport
6 resource URI templates for browsing knowledge, memory, workflows, runs, config, and resources
Bearer token authentication with constant-time comparison
mcp_restriction policy rule type for fine-grained tool blocking
PolicyEngine extended with :mcp caller type
AuthContext extended with tool_name field and build_mcp/2
/health and /metrics endpoints report MCP status
Architecture, security, and README documentation updated
Full test coverage for MCP modules (855 tests, 0 failures)

See git history for the complete changelog.