continuous-learning ยท diff
git:20261025.d69cffe to git:20261025.ea2b348
51 added, 22 removed. Audit A to A.
---
name: continuous-learning
description: Continuous conversational learning protocol (CLP) skill. Persists user corrections and project conventions to .agents/learned_conventions.md to prevent repeating mistakes.
---
# Continuous Conversational Learning (`/continuous-learning`)
> **Philosophy**: *Mistakes happen once; repeating them is unacceptable. Absorb developer corrections immediately, record invariants to disk, and enforce them forever.*
- This skill governs the Conversational Learning Protocol (CLP). It turns user corrections, preferred patterns, and architecture rules into persisted project invariants.
+ This skill governs the Conversational Learning Protocol (CLP). It turns user corrections, preferred patterns, architecture rules, and repo idioms into persisted project invariants that survive across agent contexts, sub-agent invocations, and session restarts.
---
## 1. The Continuous Learning Cycle
```mermaid
- flowchart LR
- A["User Correction / Constraint"] --> B["Acknowledge & Isolate Invariant"]
- B --> C["Persist to .agents/learned_conventions.md"]
- C --> D["Pre-Flight Inspection on Every Prompt"]
- D --> E["Zero Repeated Mistakes"]
+ flowchart TD
+ A["Developer Feedback or Correction"] --> B["1. Signal Detection & Invariant Extraction"]
+ B --> C["2. Conflict Check against Existing Rules"]
+ C --> D["3. Write to .agents/learned_conventions.md"]
+ D --> E["4. Pre-Flight Verification on Every Future Task"]
+ E --> F["Zero Repeated Mistakes & Constant Alignment"]
```
- ### Rule 1: Immediate Memory Persistence
- - When a user corrects a pattern (e.g. *"Use Tailwind v4 syntax"*, *"Never use Axios, use native fetch"*, *"Our DB IDs use prefix usr_"*), invoke `record_learning.py` immediately.
- - Never rely solely on in-context memory. Persist to disk so future sessions and sub-agents inherit the rules.
+ ### 1.1 Detection Signals: When to Record Invariants
+ Trigger the learning protocol automatically whenever:
+ 1. **Direct Correction**: The user corrects an implementation choice (e.g. *"Use Tailwind v4 instead of v3"*, *"Never use Axios, use native fetch"*).
+ 2. **Architecture Enforcement**: The user clarifies an organizational rule (e.g. *"All models must reside in src/domain/models"*).
+ 3. **Naming & Schema Conventions**: The user defines an entity convention (e.g. *"All table IDs must use ULID prefixes like usr_..."*).
+ 4. **Environment Quirks**: A unique workaround is discovered during build or test execution (e.g. *"On macOS ARM64, invoke python3 with specific DYLD path"*).
- ### Rule 2: Invariant Structure
- Every learned rule is stored with standard categorization and timestamps:
- `- **[CATEGORY-001] (YYYY-MM-DD)**: <Concrete actionable invariant>`
- - `UI`: Styling, tokens, icon preferences, font scales.
- - `ARCH`: Framework choices, package prohibitions, layer boundaries.
- - `API`: Response envelope formats, status code handling.
- - `DATA`: Database migration tools, ID schemas, index requirements.
- - `SEC`: Specific auth protocols, disallowed patterns.
+ ---
+ ## 2. Invariant Taxonomy & Schema
+
+ Every learned rule is stored in `.agents/learned_conventions.md` using strict semantic prefixes:
+
+ | Category Code | Domain | Scope & Example |
+ | :--- | :--- | :--- |
+ | `[UI-xxx]` | Design & Styling | Palette restrictions, typography scales, icon library preference. |
+ | `[ARCH-xxx]` | System Architecture | Layer boundaries, dependency inversion, package prohibitions. |
+ | `[API-xxx]` | API Contracts | Error payload envelope, idempotency headers, versioning schemes. |
+ | `[DATA-xxx]` | Persistence & DB | Migration tooling, soft delete policies, indexing rules. |
+ | `[SEC-xxx]` | Security & Auth | Token storage mechanisms, allowed JWT algorithms, CORS whitelist. |
+ | `[TOOL-xxx]` | Tooling & Build | Custom test runner flags, linter rules, compiler arguments. |
+
+ ### Canonical Entry Format:
+ ```markdown
+ - **[ARCH-004] (2026-10-25)**: Standardize all HTTP network calls on Node 18+ native `fetch`. Do not install `axios` or `request`.
+ ```
+
---
- ## 2. Learning Recorder CLI
+ ## 3. Pre-Flight Enforcement Workflow
- Record and query learned conventions:
+ Before generating or modifying any code in the repository:
+ 1. Read `.agents/learned_conventions.md` if present in the workspace root or parent directory.
+ 2. Cross-reference the proposed solution against all recorded invariants.
+ 3. If a conflict is detected between standard practices and a project-specific learned convention, the **project learned convention takes absolute precedence**.
+ 4. Confirm compliance silently in the thought process before delivering the code.
+
+ ---
+
+ ## 4. Learning Recorder CLI
+
+ Record and query learned conventions automatically via script:
+
```bash
- # Record an invariant
- python3 skills/continuous-learning/scripts/record_learning.py --category ARCH --text "Use native fetch over axios in all new modules"
+ # Record an architectural convention
+ python3 skills/continuous-learning/scripts/record_learning.py --category ARCH --text "Standardize on native fetch; do not install axios"
- # List all recorded conventions
+ # Record a UI convention
+ python3 skills/continuous-learning/scripts/record_learning.py --category UI --text "Use SF Pro font stack with -0.02em tracking for headings"
+
+ # List and audit all recorded conventions
python3 skills/continuous-learning/scripts/record_learning.py --list
```