Autonomous Agent
The Kernax Autonomous Agent plans, reads codebases, modifies files, runs local terminal commands, and tracks task checklists.
Single Persistent Turn Timeline
Traditional AI chat tools create fragmented message logs with jarring layout shifts whenever an agent executes shell tools or edits files.
The Kernax Autonomous Agent introduces the Single Persistent Turn paradigm:
- Collapsible Reasoning Traces: The agent surfaces its underlying architectural plan, file discoveries, and decision trees inside collapsible reasoning chips before modifying a single line of code.
- Zero Layout Shifts: As terminal outputs stream or files change, the timeline maintains a stable scroll position. You never lose sight of your active editor buffer.
- Stateful Session Continuity: You can pause the agent mid-turn, provide steering guidance, or ask it to elaborate on a specific architectural decision without losing session state.
Parallel Tool Execution
Modern software engineering tasks require simultaneous multi-file analysis. Kernax equips its agent core with non-blocking, parallel tool dispatchers:
| Tool Primitive | Operation Mode | Safety Mechanism |
|---|---|---|
read_file_batch |
Concurrent read of up to 10 source files | Restricted to workspace root, client-side path validation |
ripgrep_search |
High-throughput regex and literal search | Respects .gitignore and .kernaxignore automatically |
apply_file_diff |
Precise hunk replacement via Tree-sitter | Generates pre-turn rollback checkpoint before applying changes |
run_terminal_command |
Spawns subprocess inside local shell | Intercepted by Hard-Lock Guard; destructive patterns blocked |
Interactive Diff & Review
When an agent edits one or several files, Kernax renders a synchronized diff viewer directly inside the timeline turn:
- Inline Unified / Split View: Easily switch between side-by-side split diffs and unified inline reviews.
- Granular Hunk Acceptance: Accept or reject specific chunks of code, or click Accept All to commit changes across all files.
- Syntax Highlighting & LSP Feedback: The diff engine runs your active language server in the background to confirm no syntax or type errors were introduced.
Task Plan (tasks.md) Tracking
For multi-step refactoring workflows (such as migrating an ORM or upgrading dependencies), Kernax maintains a hierarchical tasks.md checklist in memory or on disk:
# Task Plan: Stripe Webhook Migration
- [x] Create idempotency table in database schema
- [/] Refactor `webhooks/stripe.ts` to verify signatures via webhook secret
- [ ] Implement exponential backoff retry handler
- [ ] Run test suite: `npm test -- stripe.test.ts`
- [ ] Verify manual event emission in local environment
The agent dynamically marks tasks as completed ([x]) or in-progress ([/]). If a test runner fails, the agent self-corrects and updates the task plan automatically.
Browser State Verification
Kernax includes an integrated headless browser subagent capable of inspecting live web applications running on localhost:
The browser subagent navigates to your local preview URL, checks for console errors or broken DOM elements, and reports visual or functional defects back to the lead agent for autonomous resolution.