feat: eversale-cli v2.1.218 — Complete Z.AI Integration + Windows Native Assistant Orchestrator Architecture

[codegen-sh]

✅ Status: FULLY OPERATIONAL — End-to-End Verified

Complete analysis and surgical patching of eversale-cli v2.1.218 to run locally with personal Z.AI API keys. The LLM pipeline is fully working:

[GPU LLM] Initialized with endpoint: https://api.z.ai/api/coding/paas/v4
[GPU LLM] Response: 65 chars, 536 tokens, 3958ms  ← SUCCESS ✅

---

🚀 Quick Start (3 env vars, that's it)

export OPENAI_API_KEY="your-z-ai-api-key"
export OPENAI_BASE_URL="https://api.z.ai/api/coding/paas/v4"
export OPENAI_MODEL="glm-5"
eversale "Research fluxy-bot"

---

🔧 Patched Files (5 files, 15 surgical changes)

File	Patches	Purpose

---

💻 View my work • 👤 Initiated by @Zeeeepa • About Codegen

[sourcery-ai]

Sorry, we are unable to review this pull request

The GitHub API does not allow us to fetch diffs exceeding 300 files, and this pull request has 769

[cubic-dev-ai]

40 issues found across 769 files

Note: This PR contains a large number of files. cubic only reviews up to 75 files per PR, so some files may not have been reviewed.

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="eversale/engine/ace/playbook.py">

<violation number="1" location="eversale/engine/ace/playbook.py:99">
P2: On load failure, `_index` is left stale while `self.strategies` is reset, which can return outdated strategies.</violation>

<violation number="2" location="eversale/engine/ace/playbook.py:153">
P1: `query_strategies()` drops page-specific matches when they are fewer than `limit` instead of combining them with general matches.</violation>

<violation number="3" location="eversale/engine/ace/playbook.py:213">
P1: `add_strategy()` does not update the page-type index, so newly added page-specific strategies cannot be found by page-specific queries.</violation>
</file>

<file name="eversale/engine/ace/injector.py">

<violation number="1" location="eversale/engine/ace/injector.py:140">
P1: Using `split("Available tools:")` can truncate the prompt when the marker appears multiple times; only the first two segments are preserved.</violation>
</file>

<file name="eversale/engine/agent/BACKEND_MIGRATION_GUIDE.md">

<violation number="1" location="eversale/engine/agent/BACKEND_MIGRATION_GUIDE.md:214">
P2: The documented `async with await create_backend(...)` pattern can crash when backend creation fails, because the `None` check is executed too late.</violation>

<violation number="2" location="eversale/engine/agent/BACKEND_MIGRATION_GUIDE.md:294">
P3: The mock-testing example uses `NavigationResult` and `ElementRef` without importing them.</violation>
</file>

<file name="eversale/engine/ace/reflector.py">

<violation number="1" location="eversale/engine/ace/reflector.py:239">
P2: Recovery detection treats `'failed'` responses as success, generating incorrect learning patterns.</violation>

<violation number="2" location="eversale/engine/ace/reflector.py:386">
P2: `is_quality_strategy` can crash on `None` input due to unsafe `len()` and slicing in the debug log path.</violation>
</file>

<file name="eversale/engine/ace/playbook.yaml">

<violation number="1" location="eversale/engine/ace/playbook.yaml:7">
P2: Marker values are mojibake (`✓`/`✗`) instead of `✓`/`✗`, so marker-based strategy classification will fail.</violation>
</file>

<file name="eversale/engine/agent/HISTORY_PRUNER_CONFLICT_ANALYSIS.md">

<violation number="1" location="eversale/engine/agent/HISTORY_PRUNER_CONFLICT_ANALYSIS.md:139">
P2: This document states `_compact_context()` has no call sites, but the method is invoked in multiple places in `brain_enhanced_v2.py`.</violation>

<violation number="2" location="eversale/engine/agent/HISTORY_PRUNER_CONFLICT_ANALYSIS.md:165">
P3: The claim that `ConversationContext.add_message()` is never called anywhere is incorrect; it is used inside `ui_tars_patterns.py`.</violation>
</file>

<file name="eversale/engine/agent/FAST_MODE_README.md">

<violation number="1" location="eversale/engine/agent/FAST_MODE_README.md:5">
P3: The headline speedup claim overstates performance relative to the benchmarks shown later in the same document. Align the claim with measured results.</violation>

<violation number="2" location="eversale/engine/agent/FAST_MODE_README.md:58">
P2: The configuration path is hardcoded to a local dev filesystem path, which is not portable for end users. Document the canonical user config location instead.</violation>
</file>

<file name="eversale/engine/agent/CAPTCHA_CONFIDENCE_SCORING.md">

<violation number="1" location="eversale/engine/agent/CAPTCHA_CONFIDENCE_SCORING.md:32">
P2: Documentation says the default text validation model is `qwen2.5:7b-instruct`, but the implementation default is `0000/ui-tars-1.5-7b:latest`.</violation>

<violation number="2" location="eversale/engine/agent/CAPTCHA_CONFIDENCE_SCORING.md:356">
P2: The documented test command points to a non-existent file (`test_captcha_confidence.py`), so the verification steps fail as written.</violation>

<violation number="3" location="eversale/engine/agent/CAPTCHA_CONFIDENCE_SCORING.md:450">
P3: Use repository-relative paths in docs; the absolute `/mnt/c/ev29/...` path is machine-specific and not portable.</violation>
</file>

<file name="eversale/README.md">

<violation number="1" location="eversale/README.md:69">
P2: README lists an unsupported `eversale login <email>` command; this flow is not implemented by the CLI parser and will mislead users.</violation>

<violation number="2" location="eversale/README.md:151">
P2: README troubleshooting suggests `eversale setup`, but the CLI only supports `eversale --setup`.</violation>
</file>

<file name="eversale/engine/agent/ACCESSIBILITY_ELEMENT_FINDER_QUICKREF.md">

<violation number="1" location="eversale/engine/agent/ACCESSIBILITY_ELEMENT_FINDER_QUICKREF.md:160">
P2: The integration example dereferences `ref.ref` without checking for `None`, which can raise at runtime when no element is matched.</violation>
</file>

<file name="eversale/engine/agent/BROWSER_BACKEND_README.md">

<violation number="1" location="eversale/engine/agent/BROWSER_BACKEND_README.md:191">
P2: The context-manager example can crash because `create_backend()` may return `None`, and `async with` is executed before the null check.</violation>

<violation number="2" location="eversale/engine/agent/BROWSER_BACKEND_README.md:417">
P2: The testing snippet uses a `SimpleAgent` API that does not exist (`backend=` constructor arg and `execute_task`).</violation>
</file>

<file name="eversale/engine/RUN_SIMPLE_QUICKSTART.md">

<violation number="1" location="eversale/engine/RUN_SIMPLE_QUICKSTART.md:271">
P2: The quickstart uses a hard-coded local dev config path (`/mnt/c/ev29/...`) that is not portable; document the user config path (`~/.eversale/engine/config/config.yaml`) instead.</violation>
</file>

<file name="eversale/engine/LINKEDIN_FALLBACK_FIX.md">

<violation number="1" location="eversale/engine/LINKEDIN_FALLBACK_FIX.md:126">
P2: The documentation incorrectly states DuckDuckGo search has no rate limits; this can lead to missing retry/throttling logic and brittle behavior.</violation>
</file>

<file name="eversale/engine/agent/ASYNC_IMPLEMENTATION_SUMMARY.md">

<violation number="1" location="eversale/engine/agent/ASYNC_IMPLEMENTATION_SUMMARY.md:5">
P2: The documentation points to an incorrect absolute path for `skill_library.py`; use the repository-relative path to avoid misleading references.</violation>
</file>

<file name="eversale/engine/agent/FAST_MODE_IMPLEMENTATION_CHECKLIST.md">

<violation number="1" location="eversale/engine/agent/FAST_MODE_IMPLEMENTATION_CHECKLIST.md:324">
P2: The checklist has a configuration contradiction: it says fast mode is default ON and also says it is opt-in. Align the wording to one behavior to avoid deployment confusion.</violation>

<violation number="2" location="eversale/engine/agent/FAST_MODE_IMPLEMENTATION_CHECKLIST.md:395">
P3: Use repository-relative paths instead of machine-specific absolute paths so the checklist is portable across environments.</violation>
</file>

<file name="eversale/engine/agent/HISTORY_PRUNER_INDEX.md">

<violation number="1" location="eversale/engine/agent/HISTORY_PRUNER_INDEX.md:5">
P2: Use repository-relative paths instead of local absolute `/mnt/c/...` paths; current links are non-portable and will break for other developers.</violation>
</file>

<file name="eversale/engine/agent/BROWSER_IMPROVEMENTS_QUICKREF.md">

<violation number="1" location="eversale/engine/agent/BROWSER_IMPROVEMENTS_QUICKREF.md:158">
P2: Initialize `browser` before the `try` block; otherwise the `finally` guard can raise `UnboundLocalError` when browser creation fails.</violation>
</file>

<file name="eversale/engine/agent/HEALTH_DASHBOARD_README.md">

<violation number="1" location="eversale/engine/agent/HEALTH_DASHBOARD_README.md:226">
P2: The example function uses `iteration` but does not accept or define it, so the snippet will fail with `NameError`.</violation>

<violation number="2" location="eversale/engine/agent/HEALTH_DASHBOARD_README.md:369">
P2: This snippet uses `await` in non-async example code, which is invalid syntax for users copying it directly.</violation>
</file>

<file name="eversale/engine/agent/ASYNC_IMPLEMENTATION_INDEX.md">

<violation number="1" location="eversale/engine/agent/ASYNC_IMPLEMENTATION_INDEX.md:170">
P2: `portalocker` is labeled optional even though the same document says Windows requires it, which is contradictory and can break setup on Windows.</violation>
</file>

<file name="eversale/engine/agent/A11Y_QUICK_START.md">

<violation number="1" location="eversale/engine/agent/A11Y_QUICK_START.md:399">
P2: The example uses top-level `await`, which will fail in normal Python scripts. Use `asyncio.run(...)` for a runnable quick-start snippet.</violation>
</file>

<file name="eversale/engine/agent/AUTO_OPTIMIZE_INTEGRATION.md">

<violation number="1" location="eversale/engine/agent/AUTO_OPTIMIZE_INTEGRATION.md:126">
P2: The SimpleAgent example uses `async def __init__`, which is invalid in Python. This will mislead users integrating the example and will not work at runtime. Use a synchronous `__init__` plus an async initializer method instead.</violation>
</file>

<file name="eversale/engine/RUN_SIMPLE_INDEX.md">

<violation number="1" location="eversale/engine/RUN_SIMPLE_INDEX.md:294">
P3: Documentation uses environment-specific absolute filesystem paths; switch to repo-relative paths so links remain valid across machines.</violation>
</file>

<file name="eversale/engine/agent/IMPLEMENTATION_COMPLETE.md">

<violation number="1" location="eversale/engine/agent/IMPLEMENTATION_COMPLETE.md:9">
P3: Replace machine-specific absolute `/mnt/c/...` paths with repository-relative paths to keep documentation portable across environments.</violation>
</file>

<file name="eversale/engine/agent/CACHING_IMPLEMENTATION.md">

<violation number="1" location="eversale/engine/agent/CACHING_IMPLEMENTATION.md:233">
P3: The test snippet is missing `import time`, so it fails at runtime with `NameError` when calling `time.time()`.</violation>
</file>

<file name="eversale/engine/agent/ARCHITECTURE_DIAGRAM.txt">

<violation number="1" location="eversale/engine/agent/ARCHITECTURE_DIAGRAM.txt:221">
P3: The “No Cache (TTL=0)” section documents behavior that is actually infinite caching, not no caching; this can lead to stale-data misconfiguration.</violation>
</file>

<file name="eversale/engine/agent/A11Y_BROWSER_COMPLETE_API.md">

<violation number="1" location="eversale/engine/agent/A11Y_BROWSER_COMPLETE_API.md:146">
P3: Core Navigation method count is incorrect versus the listed methods, which makes the API reference inconsistent.</violation>
</file>

<file name="eversale/engine/agent/EXTRACTION_HELPERS_README.md">

<violation number="1" location="eversale/engine/agent/EXTRACTION_HELPERS_README.md:401">
P2: Documentation examples encourage raw `page.click`/`page.fill` in agent workflows, which bypasses required engine wrappers and can disable auto-recovery behavior.</violation>
</file>

<file name="eversale/engine/RUN_SIMPLE_INTEGRATION.md">

<violation number="1" location="eversale/engine/RUN_SIMPLE_INTEGRATION.md:461">
P3: Nested triple-backtick fences in the README example break Markdown rendering; use a different outer fence length (e.g., four backticks) or escape inner fences.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P1: add_strategy() does not update the page-type index, so newly added page-specific strategies cannot be found by page-specific queries.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At eversale/engine/ace/playbook.py, line 213:

<comment>`add_strategy()` does not update the page-type index, so newly added page-specific strategies cannot be found by page-specific queries.</comment>

<file context>
@@ -0,0 +1,264 @@
+        key = f"{domain}:{action_type}"
+        if key not in self._index:
+            self._index[key] = []
+        self._index[key].append(new_strategy)
+
+        logger.info(f"Added strategy for {domain}/{action_type}: {strategy[:50]}...")
</file context>

[cubic-dev-ai]

P1: query_strategies() drops page-specific matches when they are fewer than limit instead of combining them with general matches.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At eversale/engine/ace/playbook.py, line 153:

<comment>`query_strategies()` drops page-specific matches when they are fewer than `limit` instead of combining them with general matches.</comment>

<file context>
@@ -0,0 +1,264 @@
+                # Otherwise fall through to get general strategies too
+
+        # Get general strategies for this domain+action
+        key = f"{domain}:{action_type}"
+        candidates = self._index.get(key, [])
+
</file context>

[cubic-dev-ai]

P1: Using split("Available tools:") can truncate the prompt when the marker appears multiple times; only the first two segments are preserved.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At eversale/engine/ace/injector.py, line 140:

<comment>Using `split("Available tools:")` can truncate the prompt when the marker appears multiple times; only the first two segments are preserved.</comment>

<file context>
@@ -0,0 +1,233 @@
+
+        # Find appropriate injection point (before tool list or at end)
+        if "Available tools:" in system_prompt:
+            parts = system_prompt.split("Available tools:")
+            return f"{parts[0]}\n\n{injection}\n\nAvailable tools:{parts[1]}"
+        else:
</file context>

[cubic-dev-ai]

P2: The documented async with await create_backend(...) pattern can crash when backend creation fails, because the None check is executed too late.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At eversale/engine/agent/BACKEND_MIGRATION_GUIDE.md, line 214:

<comment>The documented `async with await create_backend(...)` pattern can crash when backend creation fails, because the `None` check is executed too late.</comment>

<file context>
@@ -0,0 +1,544 @@
+async def run_task(task: str, backend_type='auto'):
+    from browser_backend import create_backend
+
+    async with await create_backend(backend_type) as backend:
+        if not backend:
+            return {"error": "Failed to create backend"}
</file context>

[cubic-dev-ai]

P2: is_quality_strategy can crash on None input due to unsafe len() and slicing in the debug log path.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At eversale/engine/ace/reflector.py, line 386:

<comment>`is_quality_strategy` can crash on `None` input due to unsafe `len()` and slicing in the debug log path.</comment>

<file context>
@@ -0,0 +1,411 @@
+        """
+        min_length = self.MIN_STRATEGY_LENGTH_TRAINING if self.training_mode else self.MIN_STRATEGY_LENGTH
+
+        if not strategy_text or len(strategy_text) < min_length:
+            logger.debug(f"Strategy too short ({len(strategy_text)} chars): {strategy_text[:30]}...")
+            return False
</file context>

[cubic-dev-ai]

P3: Use repository-relative paths in docs; the absolute /mnt/c/ev29/... path is machine-specific and not portable.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At eversale/engine/agent/CAPTCHA_CONFIDENCE_SCORING.md, line 450:

<comment>Use repository-relative paths in docs; the absolute `/mnt/c/ev29/...` path is machine-specific and not portable.</comment>

<file context>
@@ -0,0 +1,454 @@
+
+## Related Files
+
+- `/mnt/c/ev29/agent/captcha_solver.py` - Main implementation
+- `/mnt/c/ev29/test_captcha_confidence.py` - Test script
+- `~/.eversale/captcha_metrics.jsonl` - Metrics log
</file context>

Suggested change

	- `/mnt/c/ev29/agent/captcha_solver.py` - Main implementation
	- `eversale/engine/agent/captcha_solver.py` - Main implementation

[cubic-dev-ai]

P3: The claim that ConversationContext.add_message() is never called anywhere is incorrect; it is used inside ui_tars_patterns.py.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At eversale/engine/agent/HISTORY_PRUNER_CONFLICT_ANALYSIS.md, line 165:

<comment>The claim that `ConversationContext.add_message()` is never called anywhere is incorrect; it is used inside `ui_tars_patterns.py`.</comment>

<file context>
@@ -0,0 +1,369 @@
+
+✗ UI-TARS ConversationContext
+- Instantiated in line 1023-1024
+- `add_message()` never called anywhere
+- Acts as dead code / unused import
+
</file context>

Suggested change

	- `add_message()` never called anywhere
	- `add_message()` is called inside `ui_tars_patterns.py`, though not from `brain_enhanced_v2.py`

[cubic-dev-ai]

P3: The mock-testing example uses NavigationResult and ElementRef without importing them.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At eversale/engine/agent/BACKEND_MIGRATION_GUIDE.md, line 294:

<comment>The mock-testing example uses `NavigationResult` and `ElementRef` without importing them.</comment>

<file context>
@@ -0,0 +1,544 @@
+### After - Easy Mocking
+
+```python
+from browser_backend import BrowserBackend, SnapshotResult, InteractionResult
+
+class MockBackend(BrowserBackend):
</file context>

Suggested change

	from browser_backend import BrowserBackend, SnapshotResult, InteractionResult
	from browser_backend import BrowserBackend, SnapshotResult, InteractionResult, NavigationResult, ElementRef

[cubic-dev-ai]

P3: Nested triple-backtick fences in the README example break Markdown rendering; use a different outer fence length (e.g., four backticks) or escape inner fences.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At eversale/engine/RUN_SIMPLE_INTEGRATION.md, line 461:

<comment>Nested triple-backtick fences in the README example break Markdown rendering; use a different outer fence length (e.g., four backticks) or escape inner fences.</comment>

<file context>
@@ -0,0 +1,602 @@
+
+### Example README Update
+
+```markdown
+## Quick Start
+
</file context>

Suggested change

	```markdown
	## Quick Start
	### Simple Tasks (Fast)
	```bash
	eversale "Search Google for AI news"
	eversale --headless "Navigate to github.com"
	```
	Uses `run_simple.py` - Fast, reliable, accessibility-first approach.
	### Complex Workflows
	```bash
	eversale --ultimate "Multi-step complex workflow"
	```
	Uses `run_ultimate.py` - Advanced recovery, memory, multi-agent support.
	````markdown
	## Quick Start
	### Simple Tasks (Fast)
	```bash
	eversale "Search Google for AI news"
	eversale --headless "Navigate to github.com"

Uses run_simple.py - Fast, reliable, accessibility-first approach.

Complex Workflows

eversale --ultimate "Multi-step complex workflow"

Uses run_ultimate.py - Advanced recovery, memory, multi-agent support.

```

<a href="https://www.cubic.dev/action/fix/violation/78461715-33ab-4aac-b130-94745933557a" target="_blank" rel="noopener noreferrer" data-no-image-dialog="true">
  <picture>
    <source media="(prefers-color-scheme: dark)" srcset="https://cubic.dev/buttons/fix-with-cubic-dark.svg">
    <source media="(prefers-color-scheme: light)" srcset="https://cubic.dev/buttons/fix-with-cubic-light.svg">
    <img alt="Fix with Cubic" src="https://cubic.dev/buttons/fix-with-cubic-dark.svg">
  </picture>
</a>

[cubic-dev-ai]

P3: The headline speedup claim overstates performance relative to the benchmarks shown later in the same document. Align the claim with measured results.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At eversale/engine/agent/FAST_MODE_README.md, line 5:

<comment>The headline speedup claim overstates performance relative to the benchmarks shown later in the same document. Align the claim with measured results.</comment>

<file context>
@@ -0,0 +1,469 @@
+
+## Overview
+
+Fast Mode is a high-performance execution layer that bypasses LLM planning for simple browser actions, achieving **15-30x speedup** over traditional LLM-based execution.
+
+### Performance Comparison
</file context>