Skip to content

Changelog

All notable changes to the LEAI project are documented here.

[0.3.15] — 2026

📋 Audit Manifest Logs in TUI /update & Counter Fixes

  • Automatic Audit Log Generation in TUI (_run_update):
  • The interactive /update command in the TUI now automatically builds update manifests (latest.json, latest.md, and update_YYYYMMDD_HHMMSS.json).
  • Automatic upload of audit manifests to the SeaweedFS S3 bucket (logs/updates/latest.json and latest.md) as well as local disk ./logs/updates/ when not in remote-only mode.
  • Modified Objects Summary Counter Fix:
  • Fixed accumulator count for Modified Objects in the final TUI update panel, ensuring exact reporting of modified database entities.

[0.3.14] — 2026

🌐 Full Internationalization (DocEditor & Doctor)

  • Complete Localization of TUI Documentation Studio (doc_editor.py):
  • Localized all object catalogs, pagination, live search filters, doc completeness meters, action menus (1-7 and 0), table headers, and interactive prompts for editing columns, routines, business rules, tags, and technical warnings.
  • Complete Localization of Environment Diagnostics (doctor.py):
  • Fully translated subsystem diagnostic checks (Oracle, S3 SeaweedFS, Git, AI Providers), filesystem integrity summaries, and troubleshooting tips across Portuguese (pt-BR) and English (en-US).
  • 100% Translation Key Parity:
  • Strict 1:1 key parity verified between pt-BR and en-US locale dictionaries.

[0.3.12] — 2026

🛡️ Full SSE Streaming Resilience & Incremental Multi-Byte UTF-8 Decoding

  • Multi-Byte Incremental UTF-8 Decoder (iter_sse_lines):
  • Fixed a critical issue where multi-byte UTF-8 accented characters (ã, õ, ç, é, 🎯, •) bisected across TCP/HTTP SSE packets caused silent UnicodeDecodeError exceptions and dropped token chunks.
  • Eliminated broken words and mid-word splits (e.g. Reg\n\nras, HAD_SQL_PADRA O, substit uição, du plas).
  • Implemented iter_sse_lines with codecs.getincrementaldecoder("utf-8") across all LLM clients (OpenAI, Gemini, Anthropic, Ollama, DeepSeek, Grok).
  • Strict Output Integrity & HTML Tag Escaping:
  • Added mandatory system prompt guidelines to wrap HTML/XML tags in backticks (e.g. ``, ``, `
    `), preventing Markdown renderers from misinterpreting unclosed tags as raw HTML block containers.

    [0.3.11] — 2026

    ⚡ Inline Directives Architecture & In-Chat Lineage/Tuning Analysis

    • Support for Inline Directives (/trace, /tune, /validate, /rule, /plsql, /doc, etc.) Anywhere in Prompt:
    • Implemented an inline prompt tokenizer and directive preprocessor inspired by Antigravity CLI / 2.0.
    • Users can freely combine database object mentions (@TABLE, @PACKAGE), business rules (#RULE), and slash directives (/trace, /tune, etc.) in natural phrasing (e.g. explain @TGOVPE_EPB__VANTAGENS /trace).
    • Deterministic Lineage Precomputation & Change Risk Level:
    • Directives /trace and /lineage pre-calculate dependency graphs and evaluate Change Risk Levels (CRITICAL, HIGH, MEDIUM, LOW).
    • Instead of solely saving external files to disk, the agent delivers the full functional explanation, blast radius, and Mermaid diagrams (flowchart TD) directly within the conversation response.
    • Specialist Persona Overlays & SQL Diagnostics:
    • Automatic contextual injection of sargability diagnostics with /tune, Oracle SQL validation with /validate, canonical business filters with /rule, and specialist personas (/plsql, /doc, /patch, /impact, /catalog).
    • Real-time visual indicator badges in the interactive TUI (⚡ Directivas em linha identificadas: /trace).

    [0.3.10] — 2026

    🌐 Complete TUI Internationalization & SeaweedFS Remote-Only (no_cache) Mode

    • Comprehensive Interactive TUI Localization:
    • Localized all 30+ slash command descriptions (/doc, /rule, /tune, /validate, /thoughts, /extract, /update, /compile, /annotate, /enrich, /chat, /serve, /trace, /tables, /schema, /changes, /model, /provider, /agent, /workflow, /copy, /save, /audit, /tools, /git, /seaweed, /doctor, /init, /clear, /help, /exit) and their autocomplete hints.
    • Localized all subarguments, table headers, audit trace panels, DDL change inspectors, and interactive model switchers.
    • Strict 1:1 key parity enforced between en-US and pt-BR translation catalogs.
    • Full Support for local_cache=False / no_cache in SeaweedFS Storage:
    • Fixed unwanted creation of the ./annotations folder and local YAML files when operating in remote-only mode with SeaweedFS (no_cache: true or --no-cache).
    • Functions save_annotation, ensure_annotation_stub, sync_schema_annotations, and save_glossary now respect local_cache=False, operating 100% in-memory and via the SeaweedFS S3 API.

    [0.3.9] — 2026

    🖥️ Terminal Word-Wrapping Fix & Clean Markdown Rendering

    • Native Markdown Rendering for Workflows & Subagents:
    • Fixed mid-word line-wrapping and broken word splits (e.g. Reg\n\nras, rot ina, bloque io, P_ROW\n _NEW, V\n ANTAGENS) in both interactive TUI and CLI outputs.
    • Autonomous workflow reports (/workflow run, leai workflow run) and specialized subagent replies (/agent, leai agent run) now render through Rich's native Markdown engine with Monokai syntax highlighting bordered by subtle horizontal Rule dividers, removing boxed Panel containers that caused premature word-splitting and wrapping glitches across Windows terminals.
    • Clean mouse selection and copy-pasting without vertical box border interference.

    [0.3.8] — 2026

    🛡️ Read-Only / System Directory Resilience & Multi-Tier Fallback

    • Multi-Tier Directory Fallback for SessionAuditLogger & TUI History:
    • Fixed PermissionError: [WinError 5] Access is denied when launching leai from protected or system directories (such as C:\WINDOWS\system32).
    • The session audit persistence engine (SessionAuditLogger) and terminal history manager (InteractiveTUISession) now employ a resilient multi-tier directory cascade: ./.leai (local project directory) ➔ ~/.leai (user home directory) ➔ %TEMP%/leai (system temporary directory) ➔ InMemoryHistory.
    • Ensures that LEAI starts seamlessly from any working directory or terminal prompt without permission crashes.

    [0.3.7] — 2026

    🎯 100% Deterministic @OBJECT RAG Architecture & Dynamic Tool-Calling

    • Elimination of Plain-Text Regex Heuristic Guessing: Removed autonomous entity guessing from free natural language (extract_entities_from_question), preventing common vocabulary words (table, view, procedure, column, date, user, etc.) from colliding with homonymous internal routines in giant corporate packages.
    • Strict Single/Focal Object Scoping on Explicit @ Mentions:
    • When the user mentions @OBJECT (e.g. In table @FUNCIONARIOS...), the RAG engine scopes resolution 100% and exclusively to the mentioned object, avoiding secondary entity pollution.
    • Preserved full detailed dossier depth of up to 8,000 characters per focal entity (columns, official DBA_COL_COMMENTS, primary/foreign keys, and immediate dependencies).
    • Top-Level Object Precedence & Clean Disambiguation: Tables and Views take absolute priority over internal package routines with matching names.
    • Token Optimization & Latency Drop:
    • Natural language questions without @ generate lean initial prompts (~400 tokens), preventing prompt bloat (reducing initial prompt size from 72k tokens down to < 1k tokens) and dramatically speeding up response time.
    • The Agent utilizes native tools (get_table_schema, search_column_comments, search_database_objects, trace_object_dependencies, explain_plsql_object) dynamically and surgically on-demand.
    • Package Skeleton Compression: extract_package_skeleton synthesizes packages with over 15 routines using compact summaries to avoid token explosion.

    [0.3.6] — 2026

    🛠️ Non-Destructive leai.yml Layout Update & Smart Config Migration

    • Smart Installer Config Update (install.ps1): Configuration step now checks for existing leai.yml files and safely merges missing sections and schema layout fields (such as updates_log_path, generate_update_log, new AI providers, git, and storage.seaweedfs), preserving 100% of existing user-configured values (dsn, schemas, API keys, models) with automatic .bak backup creation.
    • Deep Merge Engine & Canonical Config Renderer (leai.template): Implemented deep_merge_missing_fields, render_canonical_config, and update_existing_config to regenerate configuration files with clean numbered sections and detailed comments without dropping custom user settings.
    • CLI leai init --update (-u): New CLI option to non-destructively refresh existing configuration layouts from the terminal.
    • TUI /init Command: Slash command /init inside interactive chat sessions now safely updates existing layouts instead of prompting for destructive overwrites.

    [0.3.5] — 2026

    🚀 SQL Performance Tuning, Strict Oracle Dialect Validation, Reverse-Engineering Workflow, Thought Streaming & Modernized TUI

    • Modernized Interactive Terminal TUI (Gemini CLI / Claude Code Experience):
    • Smart Prefix Autocompletion: Autocomplete database objects with type icons (@ with 📋 Table, 📦 Package, ⚙️ Procedure, 👁️ View, etc.), business glossary rules (# for glossary.yml), and terminal slash commands (/).
    • Dynamic Bottom Status Toolbar: Real-time bottom bar displaying active provider, model, connected schema, thought streaming toggle status (thoughts:on/off), and help shortcut.
    • Live Reasoning Streaming Panel (🧠 thoughts): Real-time chain-of-thought display from reasoning models prior to final synthesis, with /thoughts [on|off] toggle.
    • Visual Tool Execution Cards: Animated spinners and real-time execution duration timers during tool calling (⚡ search_database_objects (0.34s)).
    • Multiline Keybindings: Use Alt+Enter or Ctrl+J to insert newlines for long queries and Enter to submit.
    • New In-Session Slash Commands: /tune [sql], /validate [sql], /thoughts [on|off], /provider [name], and /workflow enhancements.
    • Query Tuning & Explanation Tool (explain_and_tune_sql): Static heuristic analysis for SQL queries identifying sargability pitfalls (TRUNC, TO_CHAR, UPPER, NVL), Full Table Scan (FTS) risks, NOT IN subquery NULL traps, correlated scalar subqueries (N+1 effect), compound index column ordering (equality filters placed before range filters), and AI-driven query rewrite recommendations.
    • Strict Oracle Dialect Validator (validate_oracle_sql): Dedicated tool to enforce Oracle Database SQL compliance by catching and correcting non-Oracle constructs (LIMIT/OFFSET -> ROWNUM/FETCH FIRST, BOOLEAN -> CHAR(1), ILIKE -> UPPER/REGEXP_LIKE, IFNULL -> NVL, DATEADD -> date arithmetic, GETDATE -> SYSDATE, AUTO_INCREMENT -> IDENTITY/Sequence, + string concatenation -> ||) and cross-validating tables/columns against loaded schemas.
    • Autonomous Reverse-Engineering Workflow (reverse-procedure): New 5-stage automated pipeline (leai workflow run reverse-procedure <TARGET>, aliases: reverse, decomp) to decompile stored procedures, functions, and packages: extracts source code, synthesizes a CRUD matrix (SELECT, INSERT, UPDATE, DELETE, MERGE), traces system packages (DBMS_OUTPUT, UTL_FILE) and external routine calls, decomposes business logic into discrete rules, and outputs Mermaid flowcharts (flowchart TD) with comprehensive functional specifications in Markdown.
    • Real-Time Streaming with Tool Calling & Thought/Reasoning Tokens (on_thought): Native support for token-by-token streaming during tool calling iterations (stream_chat_with_tools), streaming internal reasoning blocks (<think> / reasoning_content in DeepSeek-R1/Qwen 2.5/Ollama, thought: true in Gemini, and thinking_delta in Claude) to the terminal and Web Studio via SSE (event: thought).
    • Local Model Inference Optimization: Support for num_ctx, keep_alive, max_tokens, top_p, and options in leai.yml for Ollama and LM Studio, preventing silent prompt truncation on extensive schemas.
    • Subagent Specialist Upgrades: plsql_analyst and patch_generator updated with access to explain_and_tune_sql and validate_oracle_sql.

    [0.3.4] — 2026

    🛡️ Anti-Hallucination Guardrails, Resilient Tool Call Parser & English System Prompts

    • Resilient Embedded Tool Call Parser (extract_embedded_tool_calls): Added robust extraction of tool call JSON blocks ({"name": "...", "arguments": {...}}) even when embedded inside free-form conversational text (common quirk in local models such as Qwen in Ollama), <tool_call> tags, and markdown code blocks.
    • Auto-Correction Critique Reprompting Loop: When the model outputs conversational preambles promising to search ("I will check the table...") on iteration 1 without executing tools via the API, LEAI interceptively reprompts the model internally to force immediate execution, preventing empty conversational responses to the user.
    • Zero-Preamble Rule & Few-Shot Demonstrations: Enforced strict zero-preamble rules during tool investigation along with in-context learning demonstrations showing direct tool calls versus prohibited preambles.
    • Global English System Prompts with Localized Output Directive: Unified all agent, subagent, workflow, and enrichment system prompts strictly in English to maximize LLM reasoning fidelity, coupled with a mandatory closing directive ensuring the final synthesized response is delivered in the project's configured language (language: "pt-BR" or "en-US").

    [0.3.3] — 2026

    ⏱️ Real-Time Elapsed Timer Updates During Metadata Extraction (TUI & CLI)

    • Elimination of Frozen Status Timer (LiveStatusTicker): Introduced the LiveStatusTicker component with a periodic daemon thread in leai/status.py. The elapsed duration counter [MM:SS] in Rich's status message now ticks continuously every second in real time during long Oracle database queries (Views (querying...), Tables, etc.), eliminating the false perception of an application freeze.
    • Decoupled Architecture & Thread Safety: The status ticker runs asynchronously without interfering with database network I/O (which releases Python's GIL in the oracledb driver), and features full Context Manager support for deterministic and safe shutdown upon schema completion.
    • Comprehensive Test Coverage: Added dedicated unit tests in tests/test_status.py covering duration formatting, background periodic ticks, and exception resilience.

    [0.3.2] — 2026

    🎨 UI Polish, Prompt Localization, and Updater Resilience

    • Update Prompt Localization: The interactive update prompt now displays [S/n] when running in Portuguese (pt-BR) and [Y/n] in English (en-US), accepting both s/sim and y/yes.
    • Updater Timeout Extension & Resilience: The PyPI version lookup timeout was increased from 2.0s to 5.0s, and GitHub changelog highlights lookup from 1.5s to 4.0s, preventing false negative timeouts on slow or high-latency networks.
    • Animated SeaweedFS Loading Spinner: Connecting to SeaweedFS S3 and catalog synchronization now features an animated status spinner (console.status), brief success confirmation (✓ Schema catalog synchronized with SeaweedFS S3 successfully!), and automatic status cleanup, ensuring a seamless visual transition to the TUI and Web Studio.

    [0.3.1] — 2026

    📋 Audit Log System & Object Manifest for leai update

    • Modified Objects Audit Trail: The leai update command now formally tracks and records every Oracle database object modified within the specified time window (tables, views, mviews, packages, procedures, triggers, sequences, indexes, synonyms) along with timestamps (last_ddl_time), modifier (last_modified_by), and comments.
    • Dual Manifest Format (JSON + Markdown): Automatic generation of structured JSON manifests (update_YYYYMMDD_HHMMSS.json and latest.json) for agent and pipeline consumption, alongside human-readable Markdown tables (update_YYYYMMDD_HHMMSS.md and latest.md).
    • Remote SeaweedFS S3 Persistence: When remote storage is enabled, update reports and manifests are automatically replicated to the S3 bucket under logs/updates/.
    • Configuration & Command-Line Flags: Supported --log / --no-log and --log-dir flags in leai update, plus updates_log_path and generate_update_log configuration options in leai.yml.

    [0.3.0] — 2026

    ⚡ Metadata Search Optimization & Network I/O Elimination via annotations_index.json

    • Elimination of N+1 Network Bottleneck in SeaweedFS / S3: Assistant search tools (search_database_objects, search_column_comments, get_table_schema, and search_business_documentation) now leverage a centralized in-memory index (annotations_index.json). This eliminates hundreds of sequential HTTP network requests in remote modes (--no-cache), reducing search latency from ~150s to under 0.05s.
    • Preservation of Empty Stubs: All .yml annotation files (including empty or default stubs) continue to exist as-is on disk and in S3. The annotations_index.json includes only enriched objects (with custom comments, business rules, or tags), accurately classified by is_annotation_enriched.
    • Automatic & Incremental Index Sync: The leai annotate command automatically compiles the index, and dedicated methods in SeaweedFSStorage maintain the index updated on individual annotation writes and deletions.
    • Smart Table Direct Match in search_column_comments: When querying directly for a known table name (e.g. WEB_DADOS_FUNC), the tool immediately isolates that entity and returns all its documented columns.
    • Code Quality & Ruff Linting: Canonical formatting and linting fixes applied across the codebase according to project guidelines.

    [0.2.30] — 2026

    🛡️ Windows Entrypoint File Lock Handling & CI Cross-Platform Compatibility

    • Graceful Windows Active Process Lock Handling (os error 32): During interactive auto-updates via uv tool or pip on Windows, the active launcher leai.exe cannot be overwritten while executing. leai/updater.py now detects this file lock condition and verifies whether the Python package was successfully installed in the virtual environment (sys.executable or uv tool list), allowing the update to complete smoothly and restart without spurious errors.
    • CI Test Runner Isolation: Mocked shutil.which in Windows entrypoint lock unit tests, preventing Linux/Ubuntu runners on GitHub Actions from erroneously attempting to invoke the Windows-only _winapi C module.

    [0.2.29] — 2026

    🌐 Complete TUI Interface & Interactive Banner Localization

    • 100% Dynamic TUI & Banner: Welcome banner, database & AI status cards, quick action shortcuts, starter query suggestions, and bottom toolbar now instantly adapt to the active language (pt-BR or en-US).
    • Bilingual /help Guide & Inspection: The 28-command interactive reference guide (/help) and catalog inspection tables (/tables and /schema) now feature localized column titles, categories, and descriptions.
    • Consistent Locale Propagation: Locale initialization integrated directly into InteractiveTUISession and CLI root callback, guaranteeing proper inheritance in subcommands such as leai chat with --lang / -L flag support.
    • Parity Assurance & Testing: New automated test test_tui_localization_and_session verifying that language settings in leai.yml or environment variables correctly propagate to the TUI.

    [0.2.28] — 2026

    🛡️ Test Runner Execution & Cross-Platform Restart Fix

    • Restart Argument Sanitization: Filter test runner keywords and flags (discover, run, pytest, unittest, .py files, -m, --cov) when rebuilding the restart command in leai/updater.py, preventing spurious arguments from being passed to the LEAI CLI.
    • os.execv Isolation in Unit Tests: Mocked os.execv in tests/test_updater.py, eliminating accidental process replacement in Linux/CI (ubuntu-latest) environments and resolving No such command 'discover'.

    [0.2.27] — 2026

    🌐 Full Localization Completion & Canonical Standardization

    • 100% Canonical English Terminal by Default: Comprehensive codebase scan and migration of all residual Portuguese terminal messages in CLI and TUI to the leai/i18n catalog, ensuring clean English output by default and seamless Portuguese when language: "pt-BR" is active.
    • Git Status Tables & TUI Indicators: The /git status table (properties, columns, synchronization states, ahead/behind commit indicators), smart clipboard copy messages, and footer latency hints are 100% localized.
    • Bilingual Markdown Documentation Headers: Documentation generators (leai/docs.py) now dynamically render sections and Markdown headings (## Overview, ## Columns, ## Primary Key, ## Foreign Keys, ## Business Rules) according to active locale (## Visão Geral, ## Colunas, etc. in pt-BR).
    • Web Studio Fallbacks & Loading Screens: Catalog synchronization loaders, SeaweedFS error messages, and Mermaid lineage rendering fallback controls are now fully integrated into the Web UI i18n dictionary.
    • Import Hygiene & Scope Collision Fixes: Cleaned up redundant local t imports that could trigger UnboundLocalError and renamed iteration variables shadowing the t translator function.

    [0.2.26] — 2026

    🌐 Complete Internationalization (i18n) Architecture (en-US / pt-BR)

    • Native i18n Subsystem (leai/i18n): Lightweight, typed internationalization engine with zero external dependencies, featuring complete message catalogs for English (en-US, canonical default) and Brazilian Portuguese (pt-BR), with support for Rich markup styling and hierarchical fallback.
    • Flexible Resolution Precedence: Active locale resolution follows: CLI flag (--lang / -L) > Environment variable (LEAI_LANG / LEAI_LANGUAGE) > Config file (language: in leai.yml) > Canonical default (en-US).
    • Standardized Terminal, TUI & Updater: 100% migration of interactive updater messages, SeaweedFS/Git synchronization alerts, and TUI slash commands (/copy, /rule, /git, /init) to the i18n catalog, displaying English by default or Portuguese when configured.
    • Bilingual Configuration Templates: leai init --lang pt-BR generates starter leai.yml with Brazilian Portuguese comments and examples, while leai init generates the default English version.
    • Studio Web Integration: /api/config endpoints (GET & POST) synchronize the language setting in real time, and the Settings modal features a visual language selector with live UI re-rendering.
    • Bilingual AI System Prompts: Table enrichment, code object enrichment (procedures/packages/triggers), and Copilot RAG/chat prompts dynamically adapt based on project language, generating business descriptions and rules in Portuguese when language: "pt-BR" is active while strictly preserving technical Oracle SQL identifiers.

    [0.2.25] — 2026

    🛡️ Auto-Update Restart Fixes

    • Robust Module-Based Process Restart (python -m leai): Replaced direct sys.argv script invocation with [sys.executable, "-m", "leai"] + sys.argv[1:]. This completely resolves Windows [Errno 2] No such file or directory where the binary executable launcher (leai.exe / ~/.local/bin/leai) was erroneously passed to Python as a script path.
    • Synchronous Foreground Execution on Windows: Used subprocess.call instead of os.execv on Windows so the terminal console session remains in the foreground without prematurely handing control back to PowerShell.

    [0.2.24] — 2026

    🛡️ Fixes & Stability

    • Oracle Cursor Leak Prevention: Added explicit finally: cursor.close() blocks in fetch_schema_metadata, fetch_focal_trace, and fetch_available_schemas within leai/oracle.py, ensuring deterministic cursor cleanup even on failures.
    • Clipboard Security & Reliability: Removed shell=True on clip.exe calls and replaced PowerShell command string interpolation with direct stdin piping ($input | Set-Clipboard) to prevent quoting and injection issues.
    • Loop Variable Closure Fixes: Resolved late-binding loop variable issues in async/UI progress callbacks across leai/web/server.py, leai/cli.py, and leai/tui/session.py.
    • Code Quality & Cleanups: Fixed iteration variable reuse in leai/ai/tools.py and simplified dictionary iteration in leai/workflows/__init__.py.

    [0.2.23] — 2026

    🌟 Added & Improved

    • Granular AI Configuration in leai.yml: Support for temperature and timeout configured either globally or overridden per-provider under ai.providers.<name>.
    • Agent Operational Limits: Added max_history_turns (chat memory window), max_agent_iterations (maximum reasoning tool iterations per turn), and max_subagent_iterations (maximum steps for specialized subagents).
    • Local Model Presets: Native support for local (LM Studio, vLLM at http://localhost:1234/v1 with qwen2.5) and custom (http://localhost:8000/v1), plus updated default model for ollama to qwen2.5-coder:latest.
    • Expanded Documentation: Full updates to leai.example.yml, README.md, and MkDocs documentation covering local runtimes and AI configuration limits.

    [0.2.22] — 2026

    🌟 Added

    • Interactive Auto-Update via PyPI: The root leai command now automatically checks PyPI (https://pypi.org/pypi/leai/json) for newer releases using a fail-safe, 2-second timeout check that never blocks execution when offline.
    • Smart Installation Manager Detection: Automatically distinguishes between uv tool (uv tool upgrade leai) and standard pip (pip install --upgrade leai), with protection for local git clones (editable).
    • Friendly Interactive Terminal Prompt: Formatted Rich panel showing current vs latest version, release highlights, changelog link, an interactive [Y/n] prompt, and seamless automatic restart via os.execv.
    • Flexible Bypass Options: --no-update-check CLI option, LEAI_NO_UPDATE_CHECK=1 environment variable, and update_check: false configuration key in leai.yml for automated scripts and CI/CD pipelines.
    • Unified Windows Installer Script (install.ps1): Complete one-liner installation script covering Astral uv, Ollama, qwen2.5-coder model, LEAI CLI, and automated user workspace configuration.

    [0.2.21] — 2026

    🌟 Added

    • 100% Remote AI & Subagent Operations via SeaweedFS S3: All database inspection tools (DATABASE_TOOLS_DEFINITIONS), subagents, and autonomous workflows now operate directly against SeaweedFS S3 storage without requiring local snapshots or YAML files on disk (raw/, annotations/, docs/).
    • Remote Glossary Fallback in lookup_business_term: Instant resolution of business rules, domain definitions, and canonical SQL filters directly from the remote S3 bucket (annotations/glossary.yml).
    • Remote Annotation Enrichment for Schemas and Columns: get_table_schema, search_column_comments, and search_database_objects fetch business descriptions, rules, tags, and human column comments persisted in SeaweedFS for Tables, Views, and Materialized Views.
    • Remote Semantic Documentation Search: search_business_documentation scans remote S3 annotation manifests via list_annotated_objects() when local files are absent.
    • Remote Dossiers and RAG: build_rag_context accesses remote SeaweedFS annotations to assemble focal entity context dossiers.
    • Storage-Aware CLI Commands: leai agent run and leai workflow run now support --seaweed (-W) and --no-cache, pulling technical snapshots seamlessly from SeaweedFS.

    [0.2.20] — 2026

    🌟 Added

    • Direct Universal Rules Sync with SeaweedFS in Web Studio: The Web Studio interface (leai serve / /serve) now directly queries, creates, and removes business rules and glossary terms against the remote SeaweedFS S3 bucket (annotations/glossary.yml).
    • Visual Sync Feedback in Web Studio: Refresh button with dynamic rotation animation (spin) and instant toast notifications reporting synchronized rule counts.
    • Real-time Progress & Execution Timing in leai update and /update: Step-by-step progress tracking for tables, views, packages, and S3 uploads with [current/total] counters and elapsed duration metrics per schema and in aggregate.
    • Native S3 Bucket Versioning on SeaweedFS: Automated support and verification for S3 bucket versioning (Status: Enabled) ensuring revision history and rollback capabilities for annotations and metadata.

    ⚡ Fixes

    • Modified Objects Count Fix: Resolved issue in count_schema_objects query that defaulted empty schemas to reporting 1 modified object (max(1, total) removed).
    • Glossary Test Rules Cleanup: Eliminated hardcoded fallback rules ("Estágio Probatório") on server startup, preserving a clean workspace when no terms exist.

    [0.2.19] — 2026

    🌟 Added

    • leai update CLI Command & /update in TUI: Surgical incremental extraction of database objects modified within hours (--hours) or days (--days), merging technical schemas with consolidated snapshot and uploading deltas to SeaweedFS.
    • SeaweedFS Annotation Preservation & Merging: Non-destructive merge logic preserving existing human descriptions, tags, business rules, and column comments in SeaweedFS without overwriting them with blank stubs.
    • Continuous GLOSSARY.yml Sync with SeaweedFS: Automated synchronization and cloud persistence of corporate domain rules at annotations/glossary.yml.
    • leai rule del (CLI) and /rule del (TUI): Removal of business terms with immediate synchronization against SeaweedFS bucket.
    • Non-destructive Rule Merging (merge_glossaries): Prioritization of centralized SeaweedFS definitions during conflicts and union of domain tags and SQL filters.
    • TUI Autocompletion: Enhanced autocompletion for /update and /rule [list|add|del|find].

    [0.2.18] — 2026

    🌟 Added

    • leai doctor CLI Command: New command and official alias for check, providing end-to-end pre-flight health checks across Oracle Database (v$version), catalog permissions, pipeline directories, S3 bucket (SeaweedFS), AI model connectivity, and GitOps status.
    • /doctor and /check in TUI: In-terminal pre-flight diagnostics runnable directly inside the interactive session (leai chat) with structured Rich table feedback.
    • Updated Documentation: Expanded CLI and TUI slash command reference tables with /doctor, /seaweed, /git, /rule, /agent, and /workflow.

    [0.2.17] — 2026

    🌟 Added

    • SeaweedFS S3 Synchronization in Web Studio (/serve): Annotation edits made in the browser (POST /api/annotations) are synchronized directly to the S3 bucket in real time.
    • Remote Annotation Fallback in Web Studio: The GET /api/object endpoint automatically fetches annotations from SeaweedFS if the local file does not exist, populating local cache transparently.
    • Visual S3 Feedback in Web UI: Header status badge (☁️ S3: <bucket>) and sync confirmation message in the save toast notification.
    • /seaweed sync Subcommand in TUI: Smart bi-directional synchronization (push + pull with SHA-256 deduplication) directly runnable from the interactive terminal.
    • Local Disk Isolation for /doc: The in-terminal documentation editor saves exclusively to local disk, preventing unintended remote uploads.

    ⚡ Improvements

    • Support and documentation for S3 Lifecycle Rules (NoncurrentVersionExpiration) targeting annotations/ prefix to purge non-current version history.
    • Terminal autocompletion updated with /seaweed sync and --seaweed, -W, and --no-cache flags for /annotate.

    [0.2.15] — 2026

    🌟 Added

    • Official GitHub Pages Documentation: Full bilingual documentation suite (English and Portuguese) using Material for MkDocs.
    • In-Memory Autonomous Agent: Enhanced ReAct reasoning loop and PL/SQL subprogram compression inside leai chat.
    • Multi-Provider AI Architecture: Native lightweight REST clients for OpenAI, Gemini, Claude, DeepSeek, Qwen, and local Ollama.

    ⚡ Improvements

    • Multi-level dependency lineage tracing (trace) with automated risk calculations.
    • Recursive dereferencing of public and private synonyms (PUBLIC SYNONYM) and database links (@dblink).
    • Surgical skeletonization of PL/SQL packages delivering up to 95% token reduction.

    [0.2.0] — Initial Releases

    • Technical data dictionary extraction to JSON.
    • Non-destructive, human-editable YAML business annotations layer.
    • Compilation to Markdown documents with YAML Frontmatter and Mermaid diagrams.