Poo Code (v3.58.8)

Poo-Code is a fork of Zoo-Code which is a fork of Roo-Code which is a fork of Cline.

Release v3.58.8 — Bulk Mode API Config: assign API profiles to all modes at once with optimistic UI (see Changelog below)

(The truth is I am too lazy to chunk it into smaller commits — the full pile lives in the selfimproving branch)

⚠ EXPERIMENTAL — This fork adds a full self-improving AI layer + parallel execution engine + codebase intelligence + verification system on top of Zoo-Code. All new features are gated behind experiment toggles. Enable at your own risk. Source: selfimproving branch.

---

The Problem

1. I can't sleep well because of anxiety due to the wrong decisions it made by always selecting the first choice as the answer.
2. It ruined my morning because when I woke up I found it having an unauthorized day off during a busy day (silently stuck because of an error).

The ultimate goal is to totally replace you, so you can be permanently "Ooo" (Out of Office) and jobless like I am.

What's different from Zoo-Code

This fork adds ~31,000 lines of new infrastructure across 230+ files, all behind experiment toggles. Every new feature is gated — Zoo-Code main's behaviour is preserved with everything off.

Core Features

Feature	Poo-Code (this branch)	Zoo-Code main
Self-improving loop	SelfImprovingManager — background review pass every N turns/tool calls. Learns from mistakes, curates skills, suggests optimizations.	❌ No automated self-review
Pattern analysis	PatternAnalyzer — detects recurring tool-use patterns, error signatures, and skill gaps from execution history.	❌ No pattern detection
Curator service	CuratorService — tar.gz-backed skill store (backup/rollback). Decides when to create/update/merge skills from learned patterns.	❌ Manual skill authoring only
Skill automation	ActionExecutor + ImprovementApplier — auto-creates and updates skills from reviewed patterns.	❌ No auto skill creation
Insights engine	InsightsEngine — generates project-level insights (dead code, stale configs, architecture notes).	❌ No project insights
Resilience	ResilienceService — streaming backoff, tool error healer, auto-retry with learned recovery strategies. Separate streaming failure counter prevents cross-contamination.	❌ Basic retry only
Question evaluation	QuestionEvaluatorService — evaluates user questions for clarity/completeness; auto-selects best answer when choices are offered.	❌ Always picks first choice
Trust service	TrustService — learns tool-approval patterns over time. Full-trust mode auto-approves known-safe tools.	❌ Static auto-approval rules
Review team	ReviewTeamService — multi-agent review (innovator + critic + decider) scores every learned pattern before applying it.	❌ No pre-apply validation
Agent memory	AgentMemoryAdapter + MemoryStore + MemoryBackendFactory — pluggable memory backend (SQLite default, configurable). Bounded context injection via memory.ts types.	❌ No persistent agent memory
Learning store	LearningStore — stores/retrieves learned patterns with confidence scoring. Schema-versioned for forward compat.	❌ No learning storage
Transcript recall	TranscriptRecall — retrieves past conversation context for pattern learning.	❌ No historical context
Skill usage tracking	SkillUsageStore — tracks which skills fire, success rate, frequency. Feeds curator decisions.	❌ No usage metrics
Auto-mode orchestrator	AutoModeOrchestrator — automatically switches between VS Code modes based on task type.	❌ Manual mode switching
Mode factory	ModeFactoryService — generates custom modes from learned workflows.	❌ Fixed mode set

Parallel Execution Engine

Feature	Poo-Code	Zoo-Code
Parallel execution	ExecuteParallelSubtaskTool + ExecuteParallelChildTaskTool — DAG-based parallel task execution with dependency resolution, lock-aware scheduling, and blackboard communication between subtasks. Real child agent spawning via SubtaskExecutor callback.	❌ No parallel execution
Parallel orchestrator	ParallelSubtaskOrchestrator — topological DAG wave scheduling, cycle detection, subtask heartbeat monitoring, automatic retry/failure handling, resume/skip/cancel controls.	❌ No orchestrator
Parallel dashboard UI	Slide-in overlay panel with tabbed views (Subtasks, Child Tasks, DAG, Logs). Live DAG visualization, subtask cards with status/duration/deps, intervention controls (pause/resume/cancel/retry/skip), agent thought stream, system log stream. Auto-opens on parallel task start; closeable via Escape/backdrop/X.	❌ No UI
Log forwarding	Live log streaming from orchestrator to webview via onLog callback. Every subtask lifecycle event (start, complete, fail) pushed to dashboard in real-time.	❌ No equivalent

Verification System

Feature	Poo-Code	Zoo-Code
Requirements extraction	RequirementsVerifier — parses user prompts for bullet points, numbered lists, and narrative action verbs. Splits "Create a login page with auth and session management" into 3+ granular requirements.	❌ No requirement tracking
Auto-verification	Cross-references requirements against actual tool_use blocks in API conversation history. Marks requirements as verified when matching file paths found, failed when no corresponding file changes.	❌ No equivalent
File-changes gate	VerificationEngine runs git diff --diff-filter=ACMRTD at completion time to verify files were actually modified (not just claimed).	❌ No file-change validation
Content volume check	Parses git diff --stat to count lines inserted+deleted. Warns when <5 lines changed across all files (suggests stubs/placeholders).	❌ No equivalent
Stub detection	Scans modified files for TODO, FIXME, HACK, throw new Error('not implemented'), @ts-ignore, @ts-nocheck. Flags files where >5% of lines match.	❌ No stub scanning
Build config integrity	SHA256-hashes 15 build config files (package.json, Cargo.toml, pyproject.toml, etc.) at task start. Re-checks at completion — fails if agent tampered with build scripts.	❌ No integrity check
Claim cross-reference	Extracts file names from the agent's attempt_completion result text. Blocks completion when >50% of claimed files were not actually modified by tool calls.	❌ No claim validation
Result substance check	Rejects empty or evasion-language result text. Minimum 20 meaningful characters, blocks "nothing/failed/unable to" patterns under 80 chars.	❌ No validation
Test coverage gate	Supports coverageCommand + minCoverage config. Parses XX% from coverage tool output (matches last occurrence for total). Auto-detects coverage command per language.	❌ No coverage check
Escalation	After 5 consecutive verification failures across any gate, prompts user to bypass, retry, or cancel. Cross-call tracking persists across attempt_completion retries.	❌ No escalation
Bypass mode	Single verificationLevel: "bypass" skips ALL 5 gate sections (requirements, auto-verify, claim check, substance check, code quality). Opt-out from verification.	❌ No equivalent
Child task scoping	Aggregates tool call files from delegated children into parent's verification. Children store files under direct parent taskId; no sibling pollution. Cleaned up on delegation completion/denied/error.	❌ No equivalent
Gate consistency	vLevel resolved once, enforced uniformly across all 5 checks. No gate runs when bypassed.	❌ No equivalent

Codebase Intelligence & Developer Experience

Feature	Poo-Code	Zoo-Code
Codebase mapping	@zoo-code/codebase-mapping package — AST parsing via web-tree-sitter, dependency graph with circular detection, symbol extraction, token compression for LLM context, multi-format serialization (JSON/Mermaid/Markdown/DOT), security layer (PII/secret scanning), AST cache with TTL. 3 VS Code commands: refresh/show/export map.	❌ No equivalent
Prompt compression	compressPrompt() — lossless compression for child agent prompts (verbose phrase shortening, whitespace optimization, markdown stripping, JSON compaction, file path shortening). Prevents child agent context overload.	❌ No compression
i18n translation	All 18 locales fully translated for parallel UI (dashboard, subtask cards, intervention controls, resume panel, detail panel, DAG labels). Languages: ca, de, en, es, fr, hi, id, it, ja, ko, nl, pl, pt-BR, ru, tr, vi, zh-CN, zh-TW. Fixed defaultNS: "common" so dashboard renders labels instead of raw keys.	❌ Partial EN-only

Orchestrator Modes

Feature	Poo-Code	Zoo-Code
ONE-SHOT Orchestrator	Autonomous 8-phase sequential build agent — handles entire projects from requirements to verification in a single pass	❌ No equivalent
KAIZEN Orchestrator	Continuous improvement agent with 7-step iteration loop (Analyze → Identify → Fix → Verify → Enhance → Git Push → Re-evaluate) and self-evolving mini-goals	❌ No equivalent
Proactive Error Prevention	Pre-execution tool call validation, structured error classification (12 categories), cascading failure detection, and prevention hint injection — catches errors BEFORE they happen	❌ No equivalent
Git Auto-Push	KAIZEN mode auto-commits and pushes every cycle, enabling CI/CD pipelines to apply changes to staging/production automatically	❌ No equivalent
Self-Evolving Mini-Goals	Mini-goals automatically evolve upward as each is achieved, with healing that reverts to fixing regressions first	❌ No equivalent

Infrastructure

Feature	Poo-Code	Zoo-Code
Skill name validation	Validates skill name format (1-64 chars, lowercase alphanumeric/hyphens) in SkillManageTool (create/update/delete/merge) and ActionExecutor. Hash-truncated names pass validation with safe fallback.	❌ No validation
MemoryManager init	Calls initialize() before consolidate() in F3 cycle — prevents silent init failures on cold start.	❌ No equivalent
ConfidenceScorer wiring	Fixed: only scores when patterns exist; passes proper typed args to calculateScore().	❌ No equivalent
Always-available tools	list_files + read_file promoted to always-available tools (available in any mode).	❌ Gated per mode
Streaming failure resilience	Separate streamingFailureCount counter so consecutive tool mistakes don't exhaust streaming retry budget.	❌ Single counter
Verification ON by default	Verification engines created on extension activation. Opt-out with experiments.disableVerification: true. Previously required opt-in.	❌ No default

Experiment gate reference

Toggle	Enables
selfImproving	Master switch — enables the entire learning loop
selfImprovingAutoSkills	Auto-create/update/merge skills from learned patterns
selfImprovingAutoMode	Auto-switch VS Code modes based on task
selfImprovingReviewTeam	Multi-agent review before applying learned patterns
selfImprovingFullTrust	Auto-approve tools that TrustService considers safe
selfImprovingQuestionEvaluation	Evaluate user questions for clarity; auto-select best answer
oneShotOrchestrator	Enable ONE-SHOT Orchestrator mode for autonomous project builds
kaizenOrchestrator	Enable KAIZEN Orchestrator mode for continuous improvement
proactiveErrorPrevention	Enable pre-execution tool call validation and cascade detection
gitAutoPush	Enable auto-commit and push in KAIZEN mode
selfEvolvingMiniGoals	Enable self-evolving mini-goals with regression healing
fullUICoverage	Enable full UI coverage for all self-improving services
parallelExecution	Enable DAG-based parallel task execution
structuredOutputParsing	Enable structured output parsing for model responses
dependencyGraph	Enable dependency graph analysis for task ordering
multiRootWorkspace	Enable multi-root workspace support
parallelSubtask	Enable parallel subtask execution with blackboard communication
disableVerification	Disable ALL verification gates (verification ON by default)

Use Case Examples

Example 1: One-Shot Project Build

Scenario: You need to build a complete REST API server from scratch.

1. Switch to ONE-SHOT Orchestrator mode
2. Describe your requirements: "Build a FastAPI REST API with PostgreSQL backend, JWT auth, and CRUD endpoints for users and products"
3. The agent autonomously executes 8 phases:
   • Requirements analysis → Architecture design → Project scaffolding → Core implementation → Integration → Testing → Bug fixing → Verification
4. Result: A fully tested, production-ready API server with zero manual intervention

Example 2: Parallel Task Execution

Scenario: You need to update multiple independent files simultaneously.

1. Enable Parallel Execution in Experimental Settings
2. The agent splits your request into DAG-ordered subtasks:
   • Subtask A: Update API schema (no deps) ✓
   • Subtask B: Update backend logic (depends on A) ⟳
   • Subtask C: Update frontend types (depends on A) ⟳
   • Subtask D: Update tests (depends on B, C) ○
3. The Parallel Dashboard shows live progress, DAG visualization, logs, and agent thoughts
4. Each subtask spawns a real child agent via waitForCompletion() — parent awaits real completion
5. Result: Structurally parallel execution with dependency-safe ordering

Example 3: Continuous Codebase Improvement

Scenario: You have an existing codebase with technical debt and want continuous improvement.

1. Switch to KAIZEN Orchestrator mode
2. Set your initial mini-goal: "Fix all TypeScript strict mode errors"
3. The agent enters the Kaizen loop:
   • Cycle 1: Analyzes errors → Fixes 3 type errors → Runs tests → Git push → Re-evaluates
   • Cycle 2: Fixes 5 more errors → Runs tests → Git push → Evolves mini-goal
   • Cycle N: Continues until mini-goal achieved, then evolves upward
4. Each cycle is one atomic change, verified, and pushed to CI/CD
5. Result: Continuous, safe improvement without regressions

Example 4: Proactive Error Prevention

Scenario: You're working on a large codebase and the model keeps hitting tool errors.

1. Enable Prevention Engine and Cascade Tracker in Experimental Settings
2. Before each tool call, the system validates parameters:
   • read_file with directory path → warns to use list_files instead
   • list_files recursive=true without ripgrep → suggests find/ls fallback
   • Long execute_command → warns about shell limits
3. After errors, the system classifies them and tracks cascading failures
4. If 2+ errors occur within 30 seconds, a cascade warning is injected suggesting an approach change
5. Result: Fewer wasted tool calls, faster task completion, lower API costs

Example 5: Self-Healing Production Deployment

Scenario: A production deployment has regressions and needs immediate attention.

1. Switch to KAIZEN Orchestrator mode
2. The agent analyzes logs and test results
3. Detects regressions → mini-goal auto-reverts to fixing those first
4. Each fix is verified, committed with kaizen: fix regression in X, and pushed
5. CI/CD pipeline auto-deploys each fix to staging/production
6. Once regressions are resolved, mini-goal evolves upward to the next improvement
7. Result: Self-healing deployment with zero manual intervention

Example 6: Codebase Understanding

Scenario: You inherit a large unfamiliar codebase and need to understand its structure.

1. Enable Codebase Mapping (via zoo-code.refreshCodebaseMap command)
2. The system scans all files, parses ASTs, builds a dependency graph
3. Use zoo-code.showCodebaseMap to see file count, edges, dead symbols, cache hit rate
4. Use zoo-code.exportCodebaseMap to export Mermaid diagram for documentation
5. Result: Instant codebase understanding without manual tracing

Example 7: Verification-Gated Completion

Scenario: You want to ensure the agent actually implements what it claims before marking work as done.

1. Verification runs automatically at attempt_completion (no manual toggle needed)
2. The system checks:
   • File-changes: Were files actually modified? (git diff --diff-filter=ACMRTD)
   • Build config: Was package.json tampered with? (SHA256 snapshot comparison)
   • Claims: Do the files mentioned in the result text match actual tool calls?
   • Stubs: Are there TODO/FIXME patterns in the modified files?
   • Content volume: Were at least ~5 lines changed? (anti-stub)
   • Build/Lint/Types: Do auto-detected project commands pass?
   • Requirements: Was each extracted requirement addressed by a file change?
   • Substance: Is the completion result non-empty and non-evasive?
3. Failures block attempt_completion with detailed error messages
4. After 5 consecutive failures, user is prompted to bypass or retry
5. Result: Honest completion results with verifiable evidence

Statistic

Projects generated: Countless

Monthly cost: LLM & electric bills

Non-refundable cost: My soul

Revenue generated so far: 0 and still counting zero

Special Messages

Don't star this repo. It will just get me excited to drag you into the jobless community

Any issue not related to self-learning, submit at https://github.com/Zoo-Code-Org/Zoo-Code/issues as they know more than me (no cap)

Changelog

v3.58.8 — Bulk Mode API Config: Assign Profiles to All Modes at Once

Backend:

• ProviderSettingsManager.setModeApiConfigs() — bulk-replaces the entire mode→config map in one atomic write under lock
• ProviderSettingsManager.removeModeConfig() — removes a single mode's config assignment (used on mode delete)
• Config ID validation — setModeApiConfigs verifies every config ID exists in apiConfigs before saving, rejects stale/deleted IDs with a clear error

Webview:

• "bulkModeApiConfig" message type + handler in webviewMessageHandler.ts
• Error path calls vscode.window.showErrorMessage() (native toast) instead of ignored webview state message
• Error path re-pushes state so optimistic UI rolls back on failure
• Auto-populate: new custom modes receive the first available config ID automatically

UI:

• "Set for all modes" button — assigns the currently selected API profile to all modes at once
• "Clear default overrides" — removes only entries matching the default profile, preserving actual per-mode overrides
• Optimistic UI — optimisticModeApiConfigs local state updates instantly, syncs back when backend state arrives
• Disabled guard — "Set for all modes" is greyed out when no API profile is selected, with tooltip explaining why
• Per-mode override indicators — shows assigned profile name or "(default)" for each mode, with X button to clear individual override
• setModeApiConfigs setter added to ExtensionStateContext

Cleanup:

• Orphan prevention — deleteCustomMode handler calls removeModeConfig() to clean up stale entries in modeApiConfigs

v3.58.7 — Codebase Mapping: Full Implementation & Hardening

Complete implementation of all 17 stubs across 6 files in @zoo-code/codebase-mapping, plus 30+ bug fixes across 4 audit rounds.

All stubs implemented:

• Serializer: Mermaid, Graphviz, ASCII, HTML, Markdown format generators — full data-driven output with dead code sections, HTML escaping, edge styles
• TokenCompressor: L0 summary, L1 signatures, L2 declarations, line-based diff — all 4 LODs now produce real content
• SecurityLayer: checkComplianceBoundaries() — glob-to-regex matching against gitignore, rooignore, custom deny/allow patterns
• DocGenerator: Template-based JSDoc/TSDoc generation with kind detection, existing doc extraction, stale doc detection
• CodebaseMappingService: getDelta() (added/deleted/modified), getConfigLinks() (13 config types), getImplicitFlows() (event/middleware/callback), getGitMetadata() (git log + blame)
• SymbolExtractor: resolveReferences() — cross-references imports against local declarations

Performance (6 fixes):

• getImplicitFlows() O(n³) → O(n) with reference index map
• getConfigLinks() O(n²) → O(n) with cached directory index
• getSymbol() O(n) → O(1) with _symbolByName index
• getSymbols(kind) O(n) → O(1) with _symbolsByKind index
• serialize() caches format-independent payload — no recompute on format switch
• getDocUpdates() sequential await → batched Promise.all at concurrency 50

Memory leaks (8 fixes):

• _cachedRawContent cleared on reset, dispose, scan, incremental update
• _previousFileHashes cleared on resetScanState and dispose
• _symbolByName/_symbolsByKind cleared on dispose
• _cachedSerializationData/_cachedSerializationPayload/_cachedDirIndex/_cachedRefIndex invalidated on all lifecycle events

Data integrity (5 fixes):

• getDocUpdates() no longer overwrites real docs with TODO stubs — only generates for symbols without existing docs
• _symbolByName changed from Map<name, ExtractedSymbol> to Map<name, ExtractedSymbol[]> — duplicate names no longer silently overwritten
• getConfigLinks() reads raw file content for JSON parsing, not secret-masked AST text
• getDelta() snapshots hashes even on failed scan — delta detection always has a baseline
• getDocUpdates() stale detection bounded by limit * 2 — watchDocFiles flag now has real behavior

Dead code removed (5 fixes):

• buildSerializationData() unused _graph parameter removed
• resolveReferences() no-op replaced with real implementation
• watchDocFiles config flag wired as stale-check cap
• ast-parser.ts initialize() JSDoc explains it's intentionally a no-op
• import("./types.js").ParseResult/DependencyEdge/CacheStats type imports cleaned up

v3.57.3 — Memory Bank: Settings Toggle, Auto-Init, .gitignore, File Truncation

Settings UI toggle — disableMemoryBank experiment now appears in the Experimental Settings panel under "Memory" category. Turn off to remove memory bank from system prompt and hide the update_memory_bank tool. i18n entries added to English locale.

Auto-initialization — exists() now auto-creates template files on first check if the memory-bank directory doesn't exist. No need to manually run zoo-code.initMemoryBank anymore.

.gitignore management — initialize() adds memory-bank/ entry to .gitignore automatically. Creates .gitignore if it doesn't exist. Prevents session-specific state from being committed to version control.

File truncation — Append-only files (decisionLog.md, progress.md) are automatically truncated when they exceed 100KB. Keeps the header and the most recent entries. Old entries are replaced with a notice. Prevents unbounded growth that wastes context window.

Workspace switch cleanup — MemoryBankManager.resetAllInstances() called on workspace change via ClineProvider.updateCodeIndexStatusSubscription(). Stale cached instances are cleared when switching projects.

v3.57.2 — Memory Bank: 4 formatting and blind-spot fixes

New feature inspired by roo-code-memory-bank methodology — structured markdown files that preserve project context across sessions.

MemoryBankManager service (src/services/memory-bank/)

• Creates memory-bank/ directory at project root with 5 template files: productContext.md, activeContext.md, decisionLog.md, systemPatterns.md, progress.md
• Append-mode files (decisionLog, progress) timestamp new entries
• updateFile() handles both append and full-replace modes

Session-start context injection

• getMemoryBankSection() added to system prompt generation
• Reads all 5 files at every task start and injects into the LLM prompt
• Agent sees project goals, decisions, and progress immediately without being told

update_memory_bank native tool

• Agent calls it anytime: arch decisions → decisionLog.md, new task → progress.md, focus change → activeContext.md, pattern discovered → systemPatterns.md
• Always available to all modes
• Append mode by default (except productContext/activeContext)

VS Code commands

• zoo-code.initMemoryBank — creates 5 template files with placeholder sections
• zoo-code.openMemoryBank — reveals memory-bank/ in the file explorer

v3.56.3 — Codebase Mapping Reliability & UX Overhaul

Complete rewrite of codebase mapping status system across 8 files.

Real scan status — Previously always showed green "Ready" with zero stats.

• Service: _scanStatus (idle→scanning→completed) properly mapped to UI (idle→scanning→ready→error)
• All 3 status senders (ClineProvider._sendCodebaseMappingStatus, requestCodebaseMappingStatus, refreshCodebaseMap) now read real status instead of hardcoding "ready"
• Empty graph (null currentGraph) no longer falsely reported as "ready" with zeros

Live progress indicator — No feedback during long scans.

• _filesScanned / _totalFilesToScan counters emit batched progress every 50 files
• Progress bar in popover (CSS, shows percentage when totalFileCount known)
• Live edge count (_accumulatedEdges) during scan instead of 0 until graph build
• Scan message shows Scanning... 234/1500 files, 890 edges

Precise stats — Files: 0 Edges: 0 Cache hit rate: 0.0%.

• scanWorkspace() checks cache before parse (getAST()/getSymbols() first) so cache hits accumulate on re-scan
• Cache hit rate now correctly >0% on subsequent scans after file saves
• totalFileCount?: number added to CodebaseMappingStatus type, populated in all status payloads

Concurrency & safety — Scan corruption, stuck status, handler leaks.

• _scanInProgress guard prevents overlapping scans (save watcher + folder change + manual refresh)
• _pendingRescan flag: if a save/delete/folder-change arrives during scan, it's queued and re-triggered in finally
• Full try/catch/finally: critical disk errors don't leave status stuck on "scanning"
• offEvent() added to service + interface; progress handlers properly unregistered on re-subscribe and dispose
• Dispose resets _scanStatus, _scanInProgress, _pendingRescan
• refreshCodebaseMap handler wrapped in try/catch — no silent hangs on scan failure

Performance — Double filesystem walk removed.

• _rootFileCache caches discoverFiles() results from pre-count pass for reuse in scan loop
• Saves ~50K stat() calls per scan on large monorepos

Install compatibility — "Invalid extension detected".

• engines.node changed from exact "20.20.2" to range ">=20.20.2" — VS Code 1.124 snap bundles different Node patch version

v3.56.2 — Initial codebase mapping fixes + VSIX packaging

• Wire codebase mapping service to webview status updates
• Fix defaultNS: "common" for webview dashboard i18n
• Various self-improving and orchestrator bug fixes

---

FAQ

Q: What is your day job?

A: Jobless

Q: What is your night job?

A: Sleep

Q: Ooo... Can I buy you coffee?

A: No. I have insomnia.

Q: Can I...?

A: This is end of conversation.