Module DD-15 — Command Code: The Taste-Learning Harness

Command Code: The Taste-Learning Harness

3,500+ stars. $5M seed. Closed-source TypeScript. The first harness that learns your coding taste from accept/reject signals and persists it as Markdown. A fundamentally new feedback model.

minutes

artifacts

sub-sections

Command Code's defining contribution is taste learning: the diff between generated and kept code becomes a preference signal, persisted as Markdown, that shapes every future output. This is architecturally novel — Hermes learns procedures, Command Code learns preferences. On Module 4's memory-tier framework, taste is a candidate sixth tier. The deep-dive's load-bearing claim: the taste loop is a genuine innovation, AND it introduces a compound-poisoning surface no other harness has — a poisoned preference shapes every future output, not just one entry.

Key Claims

Load-Bearing Claims

Diff as learning signal is a cleaner preference signal than thumbs-up/down or explicit feedback. The user's edit is the ground truth of what they wanted. A thumbs-down carries no information about what was wrong; explicit feedback requires a separate action the user would not otherwise perform. The diff is captured as a side-effect of editing already happening — no separate feedback action is required.

Taste is a candidate sixth memory tier: learned preferences derived from behavioral signal. Distinct from static memory (what you wrote, e.g. AGENTS.md) and episodic memory (what happened). The taste profile is what the system INFERRED from your accept/reject/edit behavior. Same compounding insight as Hermes's skills, applied to preferences instead of procedures.

Preference poisoning compounds in a way ordinary memory poisoning does not. Ordinary memory poisoning (Module 4.3) persists one bad entry. Preference poisoning persists one bad entry that then influences the GENERATION of every subsequent output. The same compounding property that makes taste learning valuable makes taste poisoning durable. As-built, no validation gate exists between inference and persistence.

The hosted taste model creates a preference-exfiltration surface. The profile syncs to api.commandcode.ai by design — your coding patterns leave your machine. For a proprietary codebase, the profile is a distilled representation of how you want code written, itself valuable IP. The closed-source obfuscation means you cannot audit what is actually transmitted.

After This Module

Explain Command Code's taste-learning loop end-to-end — signal capture, profile generation, Markdown persistence, server-side model, Git-style sharing — and distinguish verifiable mechanism from vendor characterization.

Articulate the distinction between static memory (what you wrote) and learned memory (what the system inferred), and place taste as a candidate sixth tier on Module 4's framework.

Analyze the two novel security surfaces — preference exfiltration and preference poisoning — and explain why both compound in ways ordinary memory poisoning does not.

Defend the 'diff as learning signal' design choice over thumbs-up/down and explicit feedback.

Score Command Code (31/60), explain the wins (Memory) and losses (Observability, Prompt), and judge whether the closed-source tradeoff is justified for a given deployment.

Artifacts

Teaching Document

~200 lines; the taste-learning loop (5-stage verifiable mechanism), static vs. learned memory (candidate 6th tier), the two novel security surfaces (preference exfiltration, preference poisoning), why diff is a better signal than thumbs-up/down, scoring (31/60), anti-patterns, key terms, references

READ

Diagrams

Mermaid diagrams — the diff-driven taste loop, static vs. learned memory composition, the compound-poisoning surface, the Module 4.3 defense gate

READ

Slide Deck

10 slides — reveal.js, dark theme, design-system teal; covers the thesis (learned preferences), the 5-stage loop, diff vs thumbs-up/down, static vs learned memory, the two novel surfaces, scoring, anti-patterns, the lab

READ

Teaching Script

Verbatim teaching transcript with [SLIDE N] cues, ~3,000 words spoken at ~140 wpm across 10 slide cues