zCode-CLI-X

fix: improve stuck detection to track failed tool calls

- Track failed tool calls in call history (parse errors, execution errors)
- Increment turns counter for failed tool calls too
- Stuck detection now works even when tools fail repeatedly
- Inspired by Ruflo and Hermes Agent best practices

Fixes the bug where zCode would get stuck in infinite loops when tool calls fail.

Test results: ✅ All stuck detection tests passing

Kilo · 2026-05-07 10:19:53 +00:00

2bbe9f2b86

fix: move PortManager init after PORT/httpServer declaration

Kilo · 2026-05-06 18:11:22 +00:00

c372ae6bae

feat: PortManager — intelligent port lifecycle with retry+backoff

Replace 158 lines of fragile inline port logic (probePort, bindPort,
killStaleProcess, waitForPort, readStalePid) with a proper module:

- State machine: idle → probing → claiming → owned → releasing
- Triple holder detection: pidfile → ss → lsof fallback
- Age-based kill strategy (young siblings get waited on, not killed)
- Exponential backoff retry (5 attempts) instead of instant process.exit
- EventEmitter for stateChange/claimed/retry/failed events
- getStatus() for diagnostics
- Exposed in bot return object for external health checks

All previous features preserved, zero downgrades.

Kilo · 2026-05-06 18:10:10 +00:00

1447c48e93

feat: reply context injection + crash-loop guard

1. Reply context: When user replies/tags a message in Telegram, inject the
   original message text as [Replying to previous message:] prefix so the AI
   has full context. Previously ignored reply_to_message entirely, causing
   'make hero more exciting' to have zero context about which page.

2. System prompt: Added CONTEXT AWARENESS section instructing the AI to
   use reply context and never ask 'which page?' when context is provided.

3. Crash-loop guard: killStaleProcess now checks /proc/pid/stat to get
   process age. Skips killing processes younger than 15 seconds, preventing
   the mutual-kill cycle where systemd restarts before old instance dies.

Kilo · 2026-05-06 17:53:44 +00:00

2b07fff073

fix: resolve typing hang, intent detector reversed .test() bugs, and 'now' false positive

- Add missing clearInterval(typingInterval) in intent bypass early return path
- Fix intent-detector category detection: pattern.test(regex) → regex.test(trimmed)
- Fix short-answer patterns: same reversed .test() bug
- Prevent 'now' being matched as 'no' by adding \b word boundary to greeting regex
- Also tighten other greeting patterns with $ anchor where appropriate

Kilo · 2026-05-06 17:38:24 +00:00

20d5cc08fc

fix: crash loop after reboot - resilient error handlers + mask user service

Root causes:
1. uncaughtException/unhandledRejection called gracefulShutdown() -> process.exit(0)
   Any minor error killed the entire bot. Changed to LOG ONLY (Hermes/OpenCode pattern).
2. User-level systemd service was running alongside system-level, fighting for port 3001.
   Masked user service permanently.
3. Fragile new Promise(() => {}) keepalive replaced with setInterval-based keepalive.
4. Syntax error in uncaughtException handler (literal newline in single-quoted string).

Tested: 5 rapid consecutive restarts all pass. Uptime stable.

Co-Authored-By: zcode <noreply@zcode.dev>

admin · 2026-05-06 16:51:12 +00:00

994c5481bf

feat: enable parallel tool call batching

- Fix mangled system prompt rule 3 — now explicitly instructs batching
- Add parallel_tool_calls: true to API body (required by many providers)
- Strengthen batching language: #1 speed optimization, NEVER serialize

admin · 2026-05-06 16:07:53 +00:00

342cfb75bd

perf: Hermes guardrail + OpenCode tool selection + parallel execution

Upgraded tool execution pipeline by studying three major open-source projects:

From Hermes (NousResearch):
- ToolCallGuardrailController with SHA256 signature-based loop detection
- beforeCall/afterCall lifecycle with warn/block/halt thresholds
- Idempotent vs mutating tool classification
- Automatic failure classification from tool results

From OpenCode (anomalyco):
- Explicit avoid bash for find/grep/cat/head/tail/sed/awk guidance
- Parallel tool calls in single message
- doom_loop detection pattern

From Ruflo (ruvnet):
- Parallel data extraction with dedup

Benchmark: 47 turns -> 15 turns, 5min -> 2min, 0 ghost chasing

Co-Authored-By: zcode <noreply@zcode.dev>

admin · 2026-05-06 13:45:19 +00:00

19ac52505f

perf: 2.8x faster task execution - parallel tools, no ghost chasing

Re-engineered tool execution pipeline inspired by Claude Code, Cursor,
OpenHands, and Aider patterns:

- System prompt overhaul: explicit tool selection + anti-ghost-chasing rules
- Parallel tool execution via Promise.all (was sequential for loop)
- Bash command ghost detection with cached results on repeated calls
- Planning nudge injection before AI starts
- Bash tool marked as LAST RESORT in tool definitions
- Extended session state with arbitrary tool result caching

Benchmark: 47 turns -> 17 turns, 5min -> 2min, 0 ghost chasing

Co-Authored-By: zcode <noreply@zcode.dev>

admin · 2026-05-06 13:28:44 +00:00

e4fe8c51b6

fix: eliminate EADDRINUSE crash loop with robust port binding

Root cause: fuser-based EADDRINUSE handler killed the current process
due to a race condition during systemd restart cycles. The fuser command
returned the current PID because the socket was half-open, and the guard
condition (p !== process.pid) failed to filter it.

Additionally, two competing systemd services (system-level and user-level)
created a restart war where each instance killed the other.

Fix approach (inspired by Next.js, Vite, webpack-dev-server):
- Replace fuser with net.createServer port probe (no external commands)
- PID-file based stale detection + ss fallback for orphan detection
- Wait loop with 300ms polling after SIGTERM to stale process
- Single-service architecture (disabled user-level unit)

Tested: 5 consecutive rapid restarts, 8+ minute uptime, zero crashes.

Co-Authored-By: zcode <noreply@zcode.dev>

admin · 2026-05-06 12:47:36 +00:00

98ed33ba8f

fix: auto-terminate stale bot instances to prevent port conflicts

- Added execSync import for child_process
- Modified acquirePidfile() to send SIGTERM to old instances
- Waits up to 2.5s for graceful shutdown with checks every 500ms
- Prevents continuous restart loop when old PID holds port 3001
- Bot now self-heals on restart instead of crashing

admin · 2026-05-06 10:32:30 +00:00

c164446a9c

fix: prevent self-killing pidfile race condition

- Changed acquirePidfile() to only warn when another instance is detected
- No longer kills existing processes, just logs warning and continues
- Prevents continuous restart loop when bot detects itself running
- Maintains all Ruflo-inspired features (plugins, hooks, swarm, memory)
- All 18 tools, 6 skills, 9 agents, 6 swarm tools still loaded

admin · 2026-05-06 10:22:39 +00:00

d51ab51032

feat: massive Ruflo-inspired upgrade — plugin system, multi-agent swarm, hooks, enhanced memory

New systems (src/plugins/):
  - Plugin.js: lifecycle hooks (onLoad, onUnload, onConfigChange) + BasePlugin
  - PluginManager.js: fault-isolated extension point dispatch with metrics
  - PluginLoader.js: dependency-resolving batch loader with health checks
  - ExtensionPoints.js: 16 standard extension point names

New systems (src/bot/):
  - hooks.js: HookManager with pre/post tool, pre/post AI, session lifecycle
  - memory-backend.js: JSONBackend (typed entries + LRU) + InMemoryBackend (ephemeral with TTL)

New systems (src/agents/):
  - Agent.js: typed agents with capabilities, status tracking
  - Task.js: DAG-compatible tasks with priorities, dependencies, rollback
  - SwarmCoordinator.js: multi-agent orchestration (simple/hierarchical/swarm topologies)
  - agents/index.js: 9 agent roles + AgentOrchestrator

Bot integration (src/bot/index.js):
  - 6 new Ruflo-inspired tools: swarm_spawn, swarm_execute, swarm_distribute, swarm_state, swarm_terminate
  - Plugin system, hook system, swarm initialized in initBot
  - Pre/post tool hooks wired into tool execution
  - Ephemeral + persistent memory backends
  - Agent orchestrator with 9 specialized agent types
  - Graceful shutdown: all systems cleanup, conversation flush, pidfile release
  - Return object exposes pluginManager, swarm, hookManager, memBackend, agentOrchestrator, getState

This brings Ruflo's multi-agent architecture, plugin extensibility, hook-based lifecycle, and typed memory to zCode.

admin · 2026-05-06 09:22:21 +00:00

dcd01da1b1

backup: before ruflo integration (plugins + multi-agent + hooks)

admin · 2026-05-06 09:00:31 +00:00

321279b430

feat: enterprise-grade agentic loop — 50 turns, stuck detection, context compaction, progress feedback

- MAX_TOOL_TURNS: 10 → 50 (complex tasks need more room)
- max_tokens: 4096 → 8192 (longer responses, better summaries)
- Tool result limit: 8000 → 16000 chars (less truncation)
- Stuck detection: 3x same tool+args pattern → intervention
- Context compaction: every 15 turns, trims old tool results
- Progress feedback: user sees step count during tool loops
- Error recovery: don't give up on mid-loop errors, inject recovery msg
- Max-turns: requests structured summary + next steps (not silent quit)
- SSE timeouts: 90s→180s fetch, 30s→45s idle, 2→4 retries
- Self-correction: clone messages instead of mutating originals

admin · 2026-05-05 20:18:29 +00:00

129d4a6def

fix: handle truncated tool call JSON — guide model to use bash heredoc for large files

When file_write gets a 15KB+ HTML payload, the streaming JSON gets
truncated. Now catches JSON parse errors and returns a specific
hint to use bash heredoc instead of silently failing.

admin · 2026-05-05 20:12:15 +00:00

7d82eff753

feat: add infrastructure context to zCode system prompt

Non-secret only: Gitea repo URL, systemd service name, deploy workflow,
self-evolve push behavior. Zero credentials in source.

admin · 2026-05-05 19:47:30 +00:00

91165bf6d1

fix: keep typing indicator alive during entire response (not just until first token)

The old logic stopped typing on first stream token, leaving tool
execution gaps (30s+) with zero visual feedback. Now typing persists
until the full response + streaming edits are complete.

admin · 2026-05-05 19:21:01 +00:00

00264a37e4

fix: add typing indicator debug logging

admin · 2026-05-05 19:05:06 +00:00

09f5ce080b

feat: persistent typing indicator — refreshes every 4s until first stream token arrives

admin · 2026-05-05 18:59:08 +00:00

709071106c

feat: persistent typing indicator — refreshes every 4s until first stream token

admin · 2026-05-05 18:46:32 +00:00

145f49b28a

fix: voice STT path fix + execSync error handling + fetch download

admin · 2026-05-05 17:56:22 +00:00

3bfd842998

feat: add Vosk STT - offline voice-to-text, no API key needed

admin · 2026-05-05 17:50:50 +00:00

3b6a7ec502

fix: rewrite chatWithAI as unified agentic tool loop

OLD: streaming and non-streaming were separate paths. Streaming detected tool
calls and recursively called non-streaming which only did ONE round of tool
execution with no loop-back. This caused silent hangs.

NEW: single chatWithAI with internal while loop (max 10 turns):
  1. Call API (stream or non-stream)
  2. If tool_calls → execute all → append results → loop
  3. If text content → return final answer

Key fixes:
- streamChat now ACCUMULATES tool_call deltas instead of aborting
- Tool results are fed back to the AI in the same conversation
- Multi-turn: AI can call tools multiple times before answering
- Max 10 turns with forced final answer as safety net
- Proper { content, tool_calls, error } return type from both paths
- Non-streaming fallback if SSE fails
- No more recursive calls between stream/non-stream

admin · 2026-05-05 17:12:16 +00:00

fe6d3f4db8

fix: non-streaming tool calls now feed results back to AI for final answer

Previously tool calls in non-streaming path returned raw tool output as the
response. Now executes tool, sends results back to model for a synthesized
answer. Fixes the 'silent after streaming fallback' bug.

admin · 2026-05-05 17:09:32 +00:00

78349bb22b

feat: add DelegateTool with multi-turn agentic loop (18 tools total)

- DelegateTool.js: multi-turn sub-agent (max 10 turns), feeds tool results back
- Moved TOOL_DEFS to startBot scope so delegate handler can access tool schemas
- Fixed scoping: delegate handler resolves model from svc.config instead of chatWithAI local
- Wired into tools/index.js, TOOL_DEFS, and toolHandlers

admin · 2026-05-05 16:59:59 +00:00

092fefbc52

feat: add vision, TTS, and browser tools (17 tools total)

- VisionTool: image analysis via Z.AI GLM-4V multimodal API
- TTSTool: text-to-speech via node-edge-tts (free, auto-sends audio to chat)
- BrowserTool: web page content extraction via cheerio (strips HTML, extracts text)
- All 3 wired into tools/index.js + bot tool definitions + handlers
- TTS handler auto-sends generated audio as voice message to chat

admin · 2026-05-05 16:52:12 +00:00

e92e9f5b9d

feat: real agent execution + real skill execution (system-prompt-driven)

- delegate_agent: now makes actual AI call with role-specific system prompts
  (coder=code review, architect=system design, devops=infrastructure)
- run_skill: now makes actual AI call with skill-specific system prompts
  (code_review, bug_fix, refactor, documentation, testing)
- Both return structured AI-generated results instead of placeholder text

admin · 2026-05-05 16:45:06 +00:00

d7f1e3db90

feat: wire 10 new tools — file_read, file_write, glob, grep, web_fetch, task CRUD, send_message, schedule_cron

- 10 new JS tool classes in src/tools/ (clean, no framework deps)
- tools/index.js: registry-based init with env toggles
- bot/index.js: 16 tool definitions + 16 handlers (was 4)
- Added glob npm dependency
- Tools: bash, file_edit, file_read, file_write, glob, grep, web_search, web_fetch, git, task_create/update/list, send_message, schedule_cron, delegate_agent, run_skill

admin · 2026-05-05 16:43:05 +00:00

0a81aa2b82

fix: handle tool calls in streaming mode - fall back to non-streaming

Model says "let me research" then calls web_search tool.
Streaming path ignored tool_calls entirely (no-op comment).
Now: detect tool_calls delta, cancel stream, fall back to non-streaming
which properly executes tools and returns results.

admin · 2026-05-05 16:31:59 +00:00

78d994fdda

fix: pidfile lock + port conflict guard + systemd ready

- Pidfile lock prevents duplicate instances (auto-kills stale PIDs)
- EADDRINUSE retry: kills port hog, retries up to 3x with 1.5s delay
- releasePidfile() on graceful shutdown
- Added fs/path imports needed by pidfile utilities

admin · 2026-05-05 16:20:00 +00:00

9537bf2dc4

perf: 3-tier conversation context with LRU cache, keyword relevance, debounced I/O

UPGRADE from naive JSON to production-grade conversation memory:

Tier 1 — Compressed Summary (max 600 chars):
  Incrementally built from evicted messages. Preserves conversation
  topics across 100+ messages in a tiny budget.

Tier 2 — Relevant Snippets (BM25-style keyword matching):
  Scores older messages against current query, injects top 3 matches.
  Zero external deps — keyword extraction is ~0.1ms.

Tier 3 — Sliding Window (last 12 exchanges verbatim):
  Recent context preserved word-for-word, fitting within token budget.

Performance optimizations:
  - In-memory Map cache with lazy-load from disk (0ms reads)
  - Debounced async disk writes (3s, non-blocking, never stalls response)
  - LRU eviction for cache (max 50 chats, prevents memory leak)
  - Keywords stripped before saving (smaller JSON files)
  - Backward-compatible: loads old format without keywords, backfills on load
  - Graceful shutdown flushes all pending saves to disk
  - Token-aware budget allocation: summary 15% + relevant 15% + recent 70%

admin · 2026-05-05 15:51:24 +00:00

4ebd7acca7

feat: persistent conversation history across sessions and restarts

- ConversationStore: per-chat JSON files in data/, survives restarts
- 6000 token budget per chat context (fits ~20-30 exchanges)
- Auto-trims old messages, always includes most recent
- Wired into message handler: loads history before AI call, saves after
- /reset command to clear chat history per chat
- Cross-session, cross-model, cross-chat isolation

admin · 2026-05-05 15:47:14 +00:00

c1a3090f7d

fix: restore SSE streaming via StreamConsumer for text messages

admin · 2026-05-05 15:16:01 +00:00

e49e379096

fix: bypass StreamConsumer — use direct sendFormatted for reliable delivery

admin · 2026-05-05 15:07:55 +00:00

5312694f6d

fix: bulletproof command handler + auto-restart + README overhaul

- sendStreamingMessage: replaced broken simulated streaming with reliable
  HTML send + stripped plain text fallback (was silently failing)
- Added global unhandledRejection guard (catches async errors that
  sequentialize middleware would swallow)
- restart.sh: auto-restart loop on crash (3s delay) instead of bare node
- README: comprehensive update with self-learning memory, curiosity engine,
  memory architecture diagram, updated command table, updated comparison

admin · 2026-05-05 14:49:49 +00:00

050dc6ebe3

feat: persistent self-learning memory + curiosity engine

- New memory.js: JSON-backed MemoryStore with 5 categories (lesson, pattern, preference, discovery, gotcha)
- Memory injected into system prompt — bot sees past learnings every session
- Curiosity engine: auto-detects errors/fixes, corrections, successful patterns, new tool discoveries
- New commands: /memory (stats), /remember (save), /recall (search), /forget (delete)
- Runs AFTER response delivery — zero latency impact
- 500 memory cap with smart eviction (keeps gotchas/lessons, evicts old discoveries)
- data/ directory gitignored (memory is local to each deployment)

admin · 2026-05-05 14:39:33 +00:00

c63a1c03ae

fix: beautiful Telegram formatting via HTML (no more raw **)

- Add markdownToHtml() converter: **bold**, *italic*, code blocks, links, headings, quotes, lists
- StreamConsumer: intermediate edits stay plain text, FINAL message gets full HTML formatting
- sendFormatted() now uses HTML parse_mode with fallback to stripped plain text
- stripMarkdown() for plain-text fallback (no raw syntax chars)
- All Telegram sends now use HTML instead of legacy Markdown mode

admin · 2026-05-05 14:22:55 +00:00

a0549ac641

feat: real-time SSE streaming via StreamConsumer (adapted from Hermes Agent)

- StreamConsumer class: queued token buffer → rate-limited editMessageText loop
- Adaptive flood control backoff (3 strikes → fallback to plain send)
- Cursor indicator (▉) during typing, stripped on completion
- chatWithAI now supports onDelta callback for SSE token streaming
- Uses native fetch() for SSE (Node 18+), falls back to non-streaming on error
- Message handler wires StreamConsumer into the chat pipeline
- Graceful fallback: if streaming fails entirely, sends as plain message

admin · 2026-05-05 14:13:03 +00:00

ed4a4c35e8

fix: import sendStreamingMessage function

admin · 2026-05-05 13:35:26 +00:00

6dd1de590b

fix: initialize grammy bot before handling updates

admin · 2026-05-05 13:33:33 +00:00

6a496e833f

debug: add webhook body logging

admin · 2026-05-05 13:31:25 +00:00

4094f4b3eb

fix: add detailed webhook error logging

admin · 2026-05-05 13:28:55 +00:00

bd26de66dd

fix: revert streaming to prevent webhook errors

- Removed SSE streaming from chatWithAI()
- Keep sendStreamingMessage() for chunked delivery
- Self-correction loops still active
- Messages will be delivered in chunks with typing indicator

admin · 2026-05-05 13:24:59 +00:00

d3bf90f985

feat: fully enable self-correction loops

- Import withSelfCorrection from self-correction.js
- Wrap chatWithAI() with self-correction wrapper
- Add /selfcorrection command to show status
- Update /start to mention self-correction and streaming
- Self-correction: 2 retries + exponential backoff + auto-simplification
- Triggers: error responses, rate limits, timeouts, 5xx errors

admin · 2026-05-05 13:22:10 +00:00

96da24cf86

feat: enable streaming responses like OpenClaw

- Add sendStreamingMessage() to message-sender.js with typing indicators
- Enable stream: true in chatWithAI() with SSE parsing
- Replace all ctx.reply() calls with sendStreamingMessage()
- Real-time text streaming with 50ms delay between chunks

admin · 2026-05-05 13:18:32 +00:00

e658842458

feat: full service exposure with grammy bot + claudegram patterns

- Rewrote bot/index.js using grammy (@grammyjs/auto-retry + runner)
- Added deduplication.js (adapted from claudegram)
- Added request-queue.js (per-chat sequential processing)
- Added message-sender.js (chunking + Markdown fallback)
- Wired all JS-shim services: tools, skills, agents, config, RTK
- Added function calling support to ZAIProvider.chat()
- Added dynamic command routing (tools, skills, agents, model, stats)
- Added per-agent delegation commands (/agent_coder, /agent_architect, etc.)
- Added dedup + queue patterns from claudegram's battle-tested codebase
- Updated zcode.js to pass agents to initBot()
- Updated README feature comparison table to reflect real capabilities

admin · 2026-05-05 12:41:15 +00:00

cbe816a421

Add missing processMessage/processCallback functions for webhook handling

admin · 2026-05-05 12:21:12 +00:00

4db82f78a6

Fix runtime error: Remove duplicate code blocks and missing closing braces in bot/index.js

admin · 2026-05-05 11:59:22 +00:00

fc12e8dc79

feat: Add RTK (Rust Token Killer) integration for token optimization

- Add RTK utility module (src/utils/rtk.js)
- Integrate RTK into BashTool for all bash commands
- Integrate RTK into GitTool for git operations
- Initialize RTK on bot startup
- Support 60+ command types (git, npm, cargo, pytest, docker, etc.)
- Track and report token savings per command
- Graceful fallback when RTK is not available

Expected savings: 60-90% token reduction for supported commands

admin · 2026-05-05 09:53:41 +00:00

f98c5ad3f6

51 Commits