autonomous-coding-loop

Files (3)

• SKILL.md
• loop_specification.md · references/
• verify_loop.py · scripts/

SKILL.md (first lines)

# Autonomous Coding & Self-Correction Loop (Antigravity Core Workflow)

This skill mandates and guides the **Autonomous Coding & Self-Correction Loop** for Antigravity in all coding and debugging sessions.

---

## 1. Golden Invariant: Never Deliver Unverified Code

Antigravity must **NEVER** guess, assume, or deliver code without proving correctness through the 6-phase closed-loop cycle:

```mermaid
flowchart LR
    A[Phase 1: Spec & Invariants] --> B[Phase 2: Test Synthesis]
    B --> C[Phase 3: Code Actuation]
    C --> D[Phase 4: Multi-Tier Verification]
    D --> E{All Tests Pass?}
    E -- Yes --> F[Phase 6: Convergence & Proof]
    E -- No --> G[Phase 5: Diagnostic & Repair]
    G --> C
```

---

## 2. Mandatory 6-Phase Execution Protocol

Whenever user tasks involve writing, editing, or fixing code, you **MUST** execute the following phases:

### Phase 1: Problem Decomposition & Invariant Grounding
1. Identify functional requirements, non-functional performance targets, and boundary constraints.
2. Establish explicit **Invariant Rules** $\mathcal{I}$ that must remain invariant throughout execution.
3. If the task is non-trivial, activate Planning Mode and construct [`implementation_plan.md`](./references/loop_specification.md).

### Phase 2: Multi-Tier Verification & Test Bench Synthesis (Test-First / CEGIS)
Synthesize verification probes *before* or *in conjunction with* code changes:
- **Level 1: Static AST & Typing**: Type hints, linters, import validation.
- **Level 2: Core Invariant Tests**: Functional happy paths and mathematical correctness.
- **Level 3: Boundary & Edge Tests**: $0, -1, \text{None}$, empty collections, max buffer sizes.
- **Level 4: Concurrency & Safety Probes**: Thread synchronization, race conditions, memory leaks.
- **Level 5: Rollback & Transactional Invariants**: Fault-injection and atomic state resets.

### Phase 3: High-Precision Code Actuation
- Use targeted modification tools: [`replace_file_content`](#) or [`multi_replace_file_content`](#) for surgical edits.
- Maintain comment integrity and adhere to existing codebase architectural idioms.
- Avoid massive full-file rewrites unless creating brand new modules.

### Phase 4: Sandboxed Execution & Telemetry Capture
- Run unit test suites, linters, and verification scripts via `run_command`.
- Intercept and isolate standard outputs, errors, exit codes, and structured tracebacks.

### Phase 5: Seven-Tier Failure Classification & Anti-Regression Repair
If any verification probe fails, classify the failure using the **Failure Taxonomy Matrix**:

| Category | Diagnostic Signal | Actionable Repair Strategy |
| :--- | :--- | :--- |
| **`SYNTAX_ERROR`** | `SyntaxError`, indentation, token mismatch | Re-parse AST, fix syntax, balance delimiters |
| **`TYPE_ERROR`** | `TypeError`, `AttributeError`, `NameError` | Inspect interfaces, add type narrowing & guards |
| **`RUNTIME_ASSERTION`** | `AssertionError`, state mismatch | Trace state transitions, audit return transformations |
| **`EDGE_CASE_BREACH`** | Boundary condition failures | Add defensive clauses for zero/null/empty/negative inputs |
| **`TIMEOUT_INFINITE_LOOP`** | Process hang, deadlock | Add loop invariants, recursion limits, non-blocking locks |
| **`SEMANTIC_LOGIC_ERROR`** | Incorrect output despite valid syntax | Step through algorithm invariants and formula logic |
| **`ENVIRONMENT_OR_IO`** | Missing packages, permissions | Check environment, install dependencies, resolve paths |

#### Anti-Regression Invariant:
- Maintain a checkpoint of the best passing state.
- After applying each patch, run the entire test suite.
- If a patch introduces a regression (fails a previously passing test), immediately revert the patch and formulate an alternative hypothesis.

### Phase 6: Formal Convergence & Walkthrough Synthesis
- Convergence is achieved when **100% of test cases pass with 0 regressions**.
- Document changes, validation results, and telemetry in [`walkthrough.md`](./references/loop_specification.md).

---

## 3. Quick Reference Scripts & Tools

- Run loop verification harness:
  ```powershell
  python .agents/skills/autonomous-coding-loop/scripts/verify_loop.py
  ```
- Detailed mathematical specification:
  [references/loop_specification.md](references/loop_specification.md)

From nguyenhungtran18/skill-and-tool-tracker, MIT.

More from nguyenhungtran18/skill-and-tool-tracker