Skip to content

CHANGELOG.md

Latest commit

 

History

History
72 lines (60 loc) · 5.74 KB

CHANGELOG.md

File metadata and controls

72 lines (60 loc) · 5.74 KB
Raw

Change Log

All notable changes to the NeuroCode VS Code extension are documented in this file.

Format based on Keep a Changelog.

[Unreleased]

Added

  • Agent session state — AgentToolLoop rebuilds a bounded prompt each step (task + session log + cached file previews); input tokens stay ~2–4k per step instead of growing with history
  • search_replace tool — line-level edits with low output token cost; preferred over write_file for import fixes and small changes
  • LLM-first Auto routing — IntentRouter with seed_paths from stack traces and error text; Auto mode uses agent tool loop (not read-only investigate)
  • Corruption detection — agent detects files containing tool-call JSON and prompts restore before fix
  • Write safety guards — isAgentToolArtifact blocks tool JSON in createOrApplyFile, applyPendingWrites, and AgentTools.write_file; agent loop skips markdown auto-apply
  • neurocode.llm.maxOutputTokens — default raised to 2048; getAgentOutputTokens() for agent steps (up to 8000 cap)
  • OpenAI-compatible LLM gateway — single OpenAICompatibleAdapter for any /v1/chat/completions endpoint (LiteLLM, vLLM, RunPod proxy, OpenAI, custom gateway)
  • LLM config refactor — neurocode.llm.mode, apiBaseUrl, apiKey, model, gatewayLabel; legacy vllmUrl / openaiUrl / provider auto-migrate via resolveLlmConfig()
  • Cursor-style model picker — Auto (optimal model per task) or Manual from GET /v1/models; ModelSelector.js, GET /llm/models, POST /llm/resolve
  • Chat attachments — attach current file, editor selection, or browse files; injected as priority-0 shards (neurocode.chat.maxAttachments)
  • Analytics panel — token usage, latency, thumbs up/down feedback per response
  • Change review — Cursor-style Accept / Reject per file, diff editor, Accept All / Reject All
  • Drift panel — semantic drift alerts UI wired to GET /drift/status and acknowledge
  • Genome panel — edit genome stats, consent, and JSONL export in sidebar
  • FileQueue — serialized file reads for indexer and shard assembly (I/O storm prevention)
  • Task queue DAG UI — improved TaskNode visualization in Task Queue panel
  • LlmConnectionBadge — replaces RunPod-specific status badge; shows gateway vs Ollama connection
  • Cursor-like chat UI on the right secondary sidebar (configurable via neurocode.ui.chatLocation)
  • Chat mode toolbar: Auto, Ask, Plan, Edit, Agent
  • IntentResolver — history-aware natural language routing (explain / plan / implement)
  • Cursor-style intent routing — IntentRouter with permissions (readOnly, allowWrites, investigate); debug/config questions use read-only Investigate loop (read_file, search_code, reply) instead of rewriting files; investigate UI shows progress instead of raw tool JSON
  • Intent router settings — neurocode.chat.intentRouter (heuristic | hybrid | llm), neurocode.chat.investigateMaxSteps
  • Agent tool loop (POST /agent/loop/stream) with tools: read_file, search_code, search_replace, write_file, reply
  • Auto-apply — Implement mode writes files when batch completes (neurocode.chat.autoApply)
  • Auto-continue — Cursor-style continuation for truncated outputs (neurocode.chat.autoContinue)
  • Fix on check — Incomplete file + review/check auto-routes to implement (neurocode.chat.fixOnCheck)
  • Collapsible code cards in chat — large blocks collapsed by default, expandable (Cursor-style)
  • Auto-index on open with retries and per-project health (neurocode.indexing.autoIndex)
  • Settings: neurocode.chat.mode, neurocode.chat.agentMaxSteps, neurocode.chat.agentToolMaxSteps

Changed

  • Auto mode execution — routes to AgentToolLoop with LLM intent resolution; Ask pill uses read-only investigate loop only
  • Agent prompt strategy — session cache holds full file bodies; LLM sees previews only (max 3 files × ~1.2k chars)
  • Seed paths — stack trace and ./path.tsx extraction merged with LLM seed_paths in ChatOrchestrator
  • RunPod is optional — GPU pod lifecycle (neurocode.runpod.*) is independent of LLM routing; gateway URL is primary
  • Chat orchestration via ChatOrchestrator.js with SSE streaming (POST /agent/chat/stream); intent event includes resolved model
  • LLMRouter.js modes: gateway | ollama; provider id: gateway | ollama
  • Plan step execution uses implement prompts (buildMessagesForIntent('edit'))
  • Improved diff parsing for unfenced // path/to/file.ts headers
  • Gateway stream errors surfaced with actionable messages
  • Right sidebar tabs: Overview | Chat | Analytics | Tasks | Shards | Review | Memory | Drift | Genome | Debug

Fixed

  • Agent streamed tool JSON auto-applied to source files via maybeAutoApplyEdits / parseCodeBlocks
  • Truncated write_file at output token limit staging invalid or tool-shaped content
  • Agent loop echoing internal tool labels as final reply ([Called read_file:…])
  • Agent token bloat from append-only message history re-sending full file contents each step
  • SSE client swallowed sidecar error events → showed misleading "stream ended without a response"
  • SSE trailing buffer not flushed → missed done events
  • thanks / social acknowledgments incorrectly triggered implement + file writes
  • Intent classified before context assembled (incomplete-file check skipped)
  • Duplicate plan-agent loop when Agent mode uses tool loop

[0.1.0] — 2026-03-01

Added

  • Initial MVP: sidecar, indexer, shards, RunPod vLLM + Ollama routing
  • Ask Agent, Task Queue, Shard Visualizer, Code Review, Memory, Debug panels
  • Attention heatmap, RunPod lifecycle, air-gap mode, edit genome (opt-in)