Releases: BlockRunAI/runcode
v2.2.1
- Fix stale closure: handleSubmit dependency array now includes lastPrompt, inputHistory, onAbort
- Simplify UI: all slash commands pass through to loop's command registry (removed 30+ case statements)
- Did-you-mean: typos like /commti suggest /commit
- InputBox: width calculation safe on tiny terminals, shorter placeholder
- Tool timer: hides "0s" for sub-second tools
v2.2.0 — Command Registry + 7 Bug Fixes
Architecture: Command Registry
60+ slash commands extracted from loop.ts (938→564 lines) into a dedicated commands.ts module with:
- Direct commands: Execute immediately (/status, /diff, /log, /stash, /doctor, etc.)
- Rewrite commands: Transform to agent prompts (/commit, /review, /test, /fix, etc.)
- Argument commands: Parse args and rewrite (/explain, /search, /find, /refactor, etc.)
Bug Fixes
- Partial response saved on Esc abort (was losing streamed content)
- Tool result aggregate cap breaks immediately (was iterating)
- AskUser returns error on EOF (was misleading string)
- SSE buffer overflow logged in debug mode
- Glob depth increased to 50 (monorepo support)
- Read offset=0 treated as 1 (1-based)
v2.1.0 — Built-in MCP, Security Hardening, 12 Bug Fixes
Built-in @blockrun/mcp
RunCode now ships with BlockRun MCP server pre-configured — zero setup needed:
blockrun_search— web searchblockrun_exa— neural researchblockrun_dex— DEX pricesblockrun_markets— prediction marketsblockrun_chat— second opinion from other models
Just run runcode start — tools are automatically available.
Security Hardening
- Project MCP trust:
.mcp.jsonfiles in project directories now require explicit trust before loading. Prevents malicious repos from executing arbitrary commands. - Write symlink check: Target file symlinks are now resolved before writing. Prevents writing to
/etc/passwdthrough a symlink.
12 Bug Fixes
- MCP: 5s connect timeout, 30s tool call timeout, transport leak cleanup
- Session: JSONL corruption recovery, prune protects active session
- Tokens: anchor sanity check on history mutation
- LLM: malformed tool JSON logged instead of silent {}
- Tools: truncation at line boundaries, imagegen download 30s timeout
- Compact: threshold capped at 8-20 messages (was unbounded %)
v2.0.0 — MCP Support + Token Management
MCP Support (Model Context Protocol)
RunCode now connects to MCP servers, giving the agent access to external tools:
// ~/.blockrun/mcp.json
{
"mcpServers": {
"blockrun": {
"transport": "stdio",
"command": "npx",
"args": ["@blockrun/mcp"]
}
}
}The agent automatically discovers and uses MCP tools:
blockrun_search— web searchblockrun_exa— neural researchblockrun_markets— prediction marketsblockrun_dex— DEX prices- Any third-party MCP server
Use /mcp to see connected servers and tools.
Token Management
- API-anchored tracking (actual counts from API, estimates for new messages)
/contextnow shows:Tokens: ~45,200 / 200k (22.6%) ✓- Anchor resets on compaction for accuracy
v1.8.0 — 40+ Commands, Sessions, Plan Mode
RunCode v1.8.0
Session Persistence
Conversations auto-save. /sessions lists them, /resume <id> continues where you left off.
Plan Mode
/plan restricts to read-only tools for safe exploration. /execute restores full access.
AskUser Tool
Agent can ask you clarifying questions with suggested answers — no more guessing.
40+ Slash Commands
Coding: /commit /review /test /fix /debug /explain /search /find /refactor /scaffold /init /todo /deps
Git: /push /pr /undo /status /diff /log /branch /stash /unstash
Analysis: /security /lint /optimize /clean /migrate /doc
Session: /plan /execute /sessions /resume /compact /context /retry /tasks /doctor
Tool Upgrades
- Read: directories return listing, binary files detected
- Grep: multiline + before/after context
- WebFetch: JSON formatting, cleaner HTML
- Task: delete action + summary counts
v1.7.2
v1.7.1
v1.7.0 — Git Workflow + Scaffolding
/push — push to remote. /pr — create pull request via gh CLI. /undo — soft reset last commit. /scaffold — generate boilerplate. /optimize — performance analysis.