docmd-assistant@0.1.18

This release expands the multi-turn reasoning budget to 6 turns, introduces a real-time streaming replacement protocol (meta.replace: true) for live tokens, and adds architectural context synthesis fallbacks.

Key Features & Enhancements

Multi-Turn Reasoning Budget Expansion

• Extended Turn Ceiling: Increased maximum turns to 6, allowing tool-assisted investigations (e.g. searching structure, inspecting navigation, reading target pages) to complete without premature tool shutoff.
• Guaranteed Final Synthesis: Tools remain active across intermediate turns and are selectively disabled only on the final turn to guarantee clean synthesis.

Real-Time Streaming Replacement Protocol

• Unblocked SSE Token Streaming: Tokens stream immediately without intermediate suppression, emitting { replace: true } markers when transitioning across reasoning turns.
• In-Place Live UI Editing: Downstream chat clients can smoothly overwrite intermediate thought chunks with synthesized responses in real time.

Architectural Context Synthesis Fallback

• Multi-Tier Context Synthesizer: When tool searches yield no dedicated documentation page, the engine automatically synthesises answers by referencing overarching project architecture and layout conventions.

Full Changelog: 0.1.17...0.1.18

docmd-assistant@0.1.17

This release makes docmd-assistant truly site-agnostic and universal, adds version and subproject filtering to documentation search, introduces token-economy document capping, strengthens context window boundaries, and improves network error resilience.

Key Features & Enhancements

Universal & Site-Agnostic System Prompt

• Authentic Knowledge Discovery: Shifted from hardcoded documentation manifest assumptions to a universal, site-agnostic prompt. The assistant dynamically explores and grounds its answers using available tools (search_documentation, read_documentation_page), making it work seamlessly across any documentation site, blog, or project.
• Retained Core Identity: Maintains the clean, professional "docmd assistant" persona without biasing retrieval assumptions toward specific build systems.

Version & Project Filtered Search

• Targeted Version Filtering: Extended search_documentation parameter schema to support optional version (e.g. "0.9.0", "0.8.0", "latest") and project filters, enabling precise retrieval in multi-version or multi-project documentation setups.
• Custom Delegate Passthrough: Forwarded (query, project, version) arguments to consumer-provided onSearch implementations.

Smart Content Capping & Context Window Safeguards

• DOM Reader Token Economy: Added an 8,000-character ceiling with a clean continuation indicator (...[content capped for token economy]) in read_documentation_page, preventing oversized pages from exhausting LLM context limits while preserving essential content.
• Context Window Floor: Enforced a minimum 1,000-character floor for contextWindow options, preventing accidental micro-truncation from stripping retrieved documentation context.

Resilient Network Error Handling

• User-Friendly Error Translation: Intercepts low-level browser network failures (Failed to fetch, Load failed) in both streaming and non-streaming modes and returns friendly, actionable error messages.
• Clean Abort Handling: Gracefully handles cancelled requests and AbortError without polluting runtime logs.

Full Changelog: 0.1.16...0.1.17

docmd-assistant@0.1.16

This release significantly improves documentation context delivery for AI agents, eliminates quote artifacts in user query forwarding, and adds multi-turn tool follow-up flagging for clean analytics logging.

Key Features & Enhancements

Expanded & Configurable Context Window

• Full Context Retention: Removed artificial 40,000-character caps on documentation page readers and increased default context window to 200,000 characters, delivering comprehensive documentation context to modern large-context LLMs.
• Configurable contextWindow: Added optional contextWindow parameter to AssistantOptions allowing users to tune retrieval length for token-constrained setups.

Clean Analytics & Tool Follow-Up Handling

• Tool Follow-Up Gating: Multi-turn tool execution calls now set isToolFollowUp: true, allowing relays and analytics loggers to distinguish internal tool retrieval steps from actual visitor queries.
• Orphaned Quote Elimination: Fixed prompt formatting where user questions were wrapped with double quotes that broke downstream log sanitizers when search context was appended.
• Normalized Tool Status Indicators: Replaced raw tool identifiers with clean human-readable activity labels ("Reading page...", "Searching docs...").

Full Changelog: 0.1.15...0.1.16

docmd-assistant@0.1.15

This release hardens LLM output sanitisation, prevents accidental stripping of documentation JSON snippets and code blocks, and enforces strict tool registration gating during autonomous execution loops.

Key Features & Enhancements

Strict Tool Registration & Name Validation

• Registered Tool Gating: Tool parser (parseAssistantOutput) and execution loops (DocmdAssistantEngine) now strictly validate tool names against registered tools (knownToolNames / this.tools).
• Markdown JSON Codeblock Preservation: JSON objects containing a "name" property (such as package.json, configuration examples, or OpenAPI schemas) are preserved as documentation content rather than falsely parsed and stripped as tool calls.
• Prose Keyword Protection: Hardened bracket and tag extraction to require explicit invocation formatting, preventing natural English sentences starting with Action:, Call:, or Tool: from being stripped from the response text.

Autonomous Loop Stability & Graceful Fallback

• Unregistered Tool Prevention: Execution loops now filter tool calls to ensure only tools actually registered on the engine are executed, preventing "Tool not registered" loops.
• Graceful Multi-Turn Termination: If the multi-turn autonomous execution loop reaches maxTurns (5), the engine safely returns the latest generated clean response text rather than falling back to an empty string.

Resolved Issues

• AI Chat Response Stripping: Resolved an issue where documentation code snippets containing "name" caused the assistant to respond with "No response generated." after search retrieval.
• Tool Parsing False Positives: Fixed accidental stripping of natural prose containing Action: or Tool: keywords.

Full Changelog: 0.1.14...0.1.15

docmd-assistant@0.1.14

docmd-assistant v0.1.14 introduces a universal, provider-agnostic tool parser that reliably intercepts and executes AI tool calls across all major LLMs and output formats.

Key Features & Enhancements

Universal Multi-Format Tool Call Parser

• Balanced JSON Scanner: Reliably extracts inline JSON tool objects regardless of key ordering ({"name": ..., "parameters": ...} or {"parameters": ..., "name": ...}), nested objects, or string-encoded argument payloads.
• Markdown Codeblock Parsing: Detects and extracts tool payloads formatted inside standard Markdown code fences (json ... , tool_call ... ).
• Bracketed Tool-Call Syntax: Added comprehensive regex extraction for bracketed and labeled tool expressions, including:
  • [Tool Call: search_documentation({"query":"..."})]
  • [Tool Call: get_site_structure()]
  • [TOOL_CALL: search_documentation(...)]
  • [Function Call: name(...)]
• Zero-Argument & Functional Syntax: Added support for zero-argument tool invocations (get_site_structure()) and functional argument lists (tool_name(query="value")).

Registered Tools Matching

• Dynamic Registry Matching: Assistant engine execution loops now pass registered tool names (search_documentation, get_site_structure, and custom tools) directly into parseAssistantOutput for deterministic recognition.

Zero Raw Tool String Leakage

• Clean User Chat Experience: Completely prevents unparsed tool-call markup from reaching human users in the chat interface; all tool invocations are executed autonomously and synthesised into clean, fact-grounded responses.

Resolved Issues

• docmd #206: AI chat returns tool-call string instead of real responses.

Full Changelog: 0.1.13...0.1.14

docmd-assistant@0.1.13

docmd-assistant v0.1.13 introduces robust payload management and clean context truncation for multi-turn RAG search synthesis.

Key Features & Enhancements

Clean Paragraph-Boundary Context Truncation

• Smart Paragraph Alignment: Context strings retrieved from documentation search results are truncated cleanly at double (\n\n) or single (\n) newline boundaries, preventing arbitrary mid-sentence cuts.
• Automatic Code Block Balancing: Detects unclosed triple-backtick fences (```) in truncated context blocks and automatically appends closing code fences to preserve Markdown syntax integrity.
• Expanded Context Capacity: Context window support expanded up to 15,000 characters per synthesis turn.

Sliding Window Conversation History

• History Management: Implemented automatic sliding window context retention for long multi-turn conversations, preserving up to the last 12 chat entries while preventing request payload bloat.
• System Prompt Safety Constraints: Updated system prompt instructions enforcing clean, structured Markdown output without raw unescaped HTML or script tags.

Cloud Relay Capacity Expansion

• 10x Relay Payload Capacity: Expanded the api.docmd.io cloud relay payload limit from 5,000 to 50,000 characters, enabling comprehensive multi-page documentation RAG retrieval without request rejections.

Full Changelog: 0.1.12...0.1.13

docmd-assistant@0.1.12

docmd-assistant v0.1.12 delivers critical fixes for multi-turn tool execution during SSE streaming, ensuring seamless tool calls without leaking raw syntax into the chat UI.

Key Features & Enhancements

SSE Streaming Tool Loop Fixes

Fixed an issue in runRelayStreamLoop where Server-Sent Events (SSE) streaming responses containing tool calls exited early without executing local tools:

• Full Tool Execution in SSE: The SSE streaming reader now parses both structured dataObj.tool_calls and text-based tool calls ([Tool Call: ...]), executes registered tools locally, and proceeds to Turn 2 synthesis.
• Turn 1 Chunk Buffering: Tokens generated during Turn 1 tool evaluation are buffered until tool parsing completes, preventing raw tool call syntax from flashing inside user chat bubbles.
• Live Turn 2 Synthesis Streaming: Real-time token streaming (callbacks.onChunk) is active during Turn 2 synthesis, delivering instant token-by-token rendering of final synthesized answers.
• Status Badge Emission: Seamlessly emits tool status updates ("Searching documentation for...") during SSE tool evaluation turns.

Full Changelog: 0.1.11...0.1.12

docmd-assistant@0.1.11

docmd-assistant v0.1.11 brings critical stability and reliability enhancements to the multi-turn autonomous tool execution loop, context-grounded response synthesis, non-streaming JSON fallback parsing, and keyword-first search decomposing.

Key Features & Enhancements

Multi-Turn Autonomous Tool Loop Stabilization

The relay streaming and standard HTTP execution loops (runRelayStreamLoop and runRelayLoop) have been re-architected for deterministic multi-turn tool execution:

• Original Query Preservation: The user's original query is explicitly tracked and maintained in the conversation history across all intermediate tool executions, ensuring the synthesis turn grounds its answer on the true user intent.
• Context-Grounded Synthesis Turn: When local tools (e.g. search_documentation) execute, the retrieved documentation chunks are cleanly formatted and injected into a dedicated synthesis turn that instructs the LLM to provide a direct, comprehensive Markdown answer with clickable links.
• Tool-Calling Loop Termination: Tool definitions are explicitly deactivated during the final synthesis turn to eliminate infinite tool loops and force immediate text output.

HTTP Stream & JSON Parsing Resiliency

Eliminated intermittent "No response returned" errors when interacting with serverless or proxy relay backends:

• Protocol Fallback: When a relay endpoint returns a standard JSON payload rather than a Server-Sent Events (text/event-stream) stream, the reader automatically falls back to full-body parsing.
• Buffer Recovery: The stream parser includes secondary buffer flushing to ensure complete message assembly even if the network connection terminates abruptly.

Keyword-First Search Query Decomposition

• Refined system prompt guidance and tool calling instructions to decompose natural language queries into targeted, high-precision search keywords (e.g., "install", "config", "cli") to maximize match accuracy on static site documentation indexes.

Enhanced Output Sanitizer Recovery

• Hardened cleanAssistantReply and parseAssistantOutput to gracefully recover text if the model generates reasoning blocks (<think>) without separate markdown content, preventing empty bubbles from reaching the chat UI.

Full Changelog: 0.1.10...0.1.11

docmd-assistant@0.1.10

This release introduces an autonomous multi-turn ReAct execution loop and universal model sanitization to ensure reliable, synthesized AI responses across all LLM providers.

Key Highlights

• Autonomous Multi-Turn Agent Loop:
  • Automatically intercepts structured and text-based tool calls (search_documentation, get_site_structure, etc.).
  • Executes tool handlers locally against documentation and search indexes.
  • Feeds tool execution results back into the conversation context to produce fully synthesized, human-readable answers in a single user turn.
• Provider-Agnostic Output Sanitizer:
  • Replaced hardcoded provider filters with a universal tag cleaner that automatically strips internal reasoning/thinking blocks (<think>, <thought>, <reasoning>, etc.) without printing internal monologues to the user.
  • Universal JSON tool-call fallback parsing for models emitting raw text blocks.
• Live Streaming API Foundation:
  • Added sendMessageStream(query, { onStatus, onChunk, onToolCall, onToolResult, onFinish }) for real-time token streaming and status updates.

Full Changelog: 0.1.9...0.1.10

docmd-assistant@0.1.9

docmd-assistant v0.1.9 introduces comprehensive LLM output sanitisation, autonomous and proactive search execution, smart version/locale relevance boosting, and enhanced chat flow stability across local and relay runtimes.

Key Features & Enhancements

LLM Output Sanitisation & Artifact Stripping

cleanAssistantReply() now aggressively strips reasoning/thinking tokens (<think>, <mm:think>), raw tool XML tags (<tool_call>, <request>), and internal provider metadata markers (e.g. minimax wrappers) before saving messages or rendering them to the user:

import { cleanAssistantReply } from 'docmd-assistant';

const clean = cleanAssistantReply(rawModelOutput);

Proactive Tool Execution & Autonomous Search

The system prompt and execution pipeline now mandate autonomous tool usage for technical, version, and release-note queries:

• No Permission Loops: The assistant directly executes search_documentation or get_site_structure to retrieve accurate facts rather than asking the user for confirmation ("Would you like me to search?").
• Release Notes & Patch Intelligence: Automatically performs targeted searches on queries regarding the latest release or patch notes (e.g. v0.9.1) without relying solely on static branch identifiers.

Version-Scoped Ranking & Filter Optimisation

• Active Version Boosting: Enhanced search scoring logic prioritising documents matching the active documentation branch and active locale.
• Legacy Exclusion: Automatically excludes or deprioritises older archived documentation branches unless the user explicitly requests historical information.

Enhanced Chat State Stability

• Hardened internal turn tracking in runRelayTurn and runAiplugTurn.
• Guaranteed message synchronisation and error boundary protection preventing broken chat UI states during relay interruptions.

Full Changelog: 0.1.8...0.1.9