Synara

• IMPROVEPermission prompt redesigned — newline + cyan > prompt, Kiro-style UX
• IMPROVEDeny with feedback — type anything other than y/t to deny and send feedback to LLM
• IMPROVEMulti-line paste support — debounce collects all pasted lines into single message
• IMPROVEAuto-compact now shows full summary and preserved message count
• IMPROVE/load now triggers LLM summary of restored context
• IMPROVEShared container instance "synara" — instant deploys, no version bump needed
• FIXThinking spinner now appears immediately after content ends (before tool args stream)
• FIXEmpty Enter no longer auto-denies permission — loops until valid input
• FIXStdin buffer drained before permission prompt — no more phantom denies
• FIXRelaxed bash block list — project-level rm -rf allowed after user confirms

• NEWLayered Harness architecture — 4-layer prompt system (Identity, Context, Skills Index, Active Skills)
• NEWSkills system — on-demand skill loading via activate_skill tool from R2
• NEW3 built-in skills: ui-ux-pro-max, cloudflare-deploy, superpowers (TDD methodology)
• NEWShared container instance — CLI sessions use fixed "synara" container, instant deploys
• IMPROVEGuardrails hardcoded into Layer 0 — no R2 dependency, zero-latency, tamper-proof
• IMPROVEError recovery rules in system prompt — stops blind retries, forces root cause analysis
• IMPROVERelaxed bash block list — only system-level destruction blocked, project rm -rf allowed after confirm
• FIXcontent_end now always sent — thinking spinner resumes correctly between content and tool calls
• FIXactivate_skill no longer sent as tool_request to CLI — handled internally by Harness

• NEWMCP server instructions injection — servers provide their own usage guidance to the LLM
• NEWMCP registry expanded: Context7 (live library docs) and sequential-thinking (structured reasoning)
• NEW! prefix for local shell commands (!ls, !git status) without going through LLM
• NEW/load tab completion — shows saved session names
• NEWDouble Ctrl+C to exit — first press shows hint, prevents accidental exit
• IMPROVEMCP parallel loading — all servers connect simultaneously, ~50% faster startup
• IMPROVEMCP tool display: dynamic verb (search/fetch/query) with blue content
• IMPROVETrusted write_file/edit_file now shows diff even when auto-approved
• IMPROVE/tools table with headers, grouped by server, aligned columns
• IMPROVEWelcome screen: SYNARA logo fixed, Cloudflare in orange, Built by Bowen Liu
• IMPROVEThinking spinner resumes after content_end — no gap before tool calls
• IMPROVEHarness forces LLM to summarize if it returns empty after tool operations
• IMPROVESystem prompt: LLM knows it can fetch URLs, must explain before tool calls
• FIXMCP server crash no longer crashes CLI — graceful disconnect and cleanup
• FIXCancel after cloud tool no longer triggers phantom thinking spinner
• FIXhandleMessage scope crash (ReferenceError) after esbuild minification
• FIXContent fragments leaking between tool calls (LLM streaming artifact)
• FIX/load now works — cloud REPL handles session_data message from CLI
• FIXsynapse-cli-worker CONTAINER_VERSION synced with harness-agent

• NEWMCP (Model Context Protocol) support — connect external tool servers for documentation, APIs, and more
• NEW/mcp command — view, enable, and disable MCP servers at runtime
• NEWBuilt-in MCP registry: aws-docs, cloudflare-docs pre-configured and ready to enable
• NEWMCP server state persisted — enabled servers auto-reconnect on next startup
• NEWTab completion for /mcp subcommands (/mcp aws-docs enable)
• NEWCommand dependency check — missing tools (uvx, npx) show install instructions
• IMPROVE/help redesigned — vertical layout with detailed descriptions
• IMPROVE/tools now shows MCP tools with server name and permission status

• NEWKiro-style tool descriptions: "I will read the file: counter.py (using tool: read_file)"
• NEWPurpose field — LLM explains why it's using each tool
• NEWSmart diff — only changed lines shown in red/green, unchanged lines as gray context
• NEWTab completion for all / commands
• NEW/tools shows live trust status, /tools reset to restore defaults
• NEW/help now shows compact two-column layout, use /command help for details
• NEWCloud guardrails loaded from R2 with 1-hour TTL cache (python3, pip3, etc.)
• IMPROVEPermission prompt y/n/t highlighted in cyan for visibility
• IMPROVETool results show full path and "✓ Successfully" in bright green
• IMPROVECompleted time on separate line for cleaner output
• FIXPermission prompt no longer auto-approves from stdin buffer — messages queued during wait
• FIXCtrl+C during permission prompt now correctly denies the action
• FIXCancel (Ctrl+C) now truly stops — ignores stale messages from cancelled turn
• FIXThinking spinner uses ANSI erase-line for clean removal

• NEW/undo — undo last file write or edit, restores previous content or deletes newly created files
• NEWMulti-line input — end a line with \ and press Enter to continue on the next line
• NEWCtrl+C to cancel — interrupts current LLM generation or tool execution, returns to prompt
• NEWMarkdown rendering for /tools, /context, /compact and other non-streamed responses
• IMPROVEgrep now supports context_lines parameter — shows N lines before and after each match
• IMPROVEread_file now has 100KB size protection — large files show truncation notice with line count
• IMPROVEWelcome screen version number now auto-syncs with package.json
• IMPROVERenamed package to synara — install with npm install -g synara, run with synara

Internal releases — npm package rename (synapse-cli → synara), esbuild bundling migration, bin path fixes.

• NEW/save [name] [--force] — save session locally to .synapse/sessions/
• NEW/load <name> — restore a previously saved session
• NEW/sessions — list all saved sessions with message count and date
• NEWASCII art welcome screen with cyan branding and info box
• NEWTool results show file size and elapsed time: "✓ Successfully read 7.0 KB - Completed in 0.1s"
• NEW/compact now displays the LLM-generated summary of compressed messages
• NEWContainer versioning with rollout_active_grace_period (10min) for zero-downtime deploys
• IMPROVESystem prompt rewrite — LLM asks clarifying questions before complex tasks, stops on user deny
• IMPROVEPermission prompt shows full action: "Read ../file.html (using tool: read_file)"
• FIXProject-outside read/list operations now correctly require permission
• FIXThinking spinner shows consistently between all tool calls

• NEWSystem prompt overhaul — LLM now asks clarifying questions before complex tasks instead of guessing
• NEWPermission prompt shows full action description: "Read ../file.html (using tool: read_file)"
• NEWContainer versioning — new deploys route to fresh instances, old sessions get 10min grace period
• FIXUser deny now stops ALL remaining tool calls in the batch, not just the denied one
• FIXPermission input requires Enter to confirm — no more accidental single-keypress execution
• FIXThinking spinner now shows consistently between all tool calls, not just the first one
• FIXPlaywright browser path fixed for Container (PLAYWRIGHT_BROWSERS_PATH + libXfixes)
• IMPROVECloud tool results (web_search/web_fetch) show brief status only, full content goes to LLM

• NEWCore architecture — cloud-powered LLM with local tool execution via WebSocket
• NEW7 built-in tools: read_file, write_file, edit_file, grep, glob, list_dir, bash
• NEWPermission model — auto-approve reads in project, bash never trustable, destructive commands blocked
• NEWLLM streaming — real-time token output via WebSocket
• NEWContext compression — LLM-powered summarization at 80% capacity, reactive compact on overflow
• NEWSession commands: /compact, /context, /tools, /usage, /quit
• NEWCloud tools: web_search (Serper) and web_fetch (Scrapling + Playwright)
• FIXArrow keys in input — switched permission prompt from raw mode to readline
• FIXStreaming tool_calls arguments truncation — re-serialize JSON after parse
• FIX400 error from extra fields in assistant messages — clean before re-send to API
• IMPROVECloud tool results show brief status in terminal, full content only sent to LLM