No product decisions

No Product Decisions — XIIGen

ABSOLUTE constraint — priority 2 in governance chain. Cannot be overridden.

Priority: ABSOLUTE — second only to agent-constitution

---

What This Skill Does

Prevents Codex from making product decisions that belong to Luba. Every DNA threshold, every scoring weight, every quality gate criterion is a product decision. Codex executes — it does not decide.

---

The Core Rule

Codex may not change behavior that affects engine output without Luba's explicit approval.

Product decisions include, but are not limited to:

• DNA pattern thresholds (e.g., compliance score cutoffs in quality-scorer.ts)
• Quality scoring weights (test_quality, spec_adherence, etc.)
• AF station activation logic (which archetypes trigger which stations)
• BFA severity classification
• Task type contract requirements (which fields are mandatory)
• Flow promotion criteria (what score constitutes "ready")
• Model selection logic (when to use Codex vs GPT-4o vs Gemini)

---

The XIIGen Analog of FIX-27

What happened with FIX-27 (translated to XIIGen context):

A developer changed the DNA compliance threshold in quality-scorer.ts from 0.7 to 0.8 to "improve quality." The change broke backward compatibility with FLOW-01 through FLOW-05 — six flows that previously passed were now marked as FAILURE overnight, with no warning and no rollback path. The engine was generating the same output it always had; the scorer's criteria had changed beneath it.

The pattern:

• Symptom: Flows that worked yesterday fail today
• Root cause: A "small" threshold/weight change that seemed like a bugfix
• Why it happened: Developer confused "fixing a bug in the scorer" with "deciding what score is acceptable"
• The rule that emerged: Scoring behavior is product behavior. Fix the code, not the criteria.

---

What Codex CAN Do (Without Approval)

• Fix a bug where the scorer returns NaN (that's a code defect)
• Fix a bug where a DNA check crashes with a null pointer (code defect)
• Fix a bug where AF-9 calculates quality score using the wrong task type's contract (code defect)
• Add new test coverage for existing behavior
• Refactor code that doesn't change observable output

What Codex CANNOT Do (Requires Luba Approval)

• Change any numeric threshold that affects pass/fail outcomes
• Change scoring weight distributions
• Add or remove required fields from engine contracts
• Change which AF stations run for a given archetype
• Change which DNA patterns apply to which generated code types
• Modify BFA conflict rules
• Add new quality layers without approval

---

Baseline Regression Check

Before any change to guardrails/, learning/, or engine-contracts/, run:

# Record current output shapes for existing flows
grep -r "StationOutput\|DataProcessResult\|qualityScore\|dnaCompliance" \
  server/src/guardrails/ server/src/learning/ | head -50

# Check which flows will be affected
grep -r "FLOW-0[1-9]\|FLOW-[12][0-9]\|FLOW-3[01]" \
  server/src/engine-contracts/ server/src/factories/

If the change touches anything that affects the above output shapes: STOP and escalate.

---

Decision Log Requirement

Every time Codex is tempted to change a threshold and does not (because this rule applies), log it:

AVOIDED-PRODUCT-DECISION: [date]
File: [file]
What I almost changed: [description]
Why I didn't: no-product-decisions rule
What Luba should decide: [the actual question]

Write this to DECISIONS.md so Luba can review and decide explicitly.

---

STOP Test + External-Tool Boundary (G02 universal addition from llm_mvp_core)

The one-sentence STOP test

Before choosing between two approaches, ask: "Would this require editing DECISIONS.md / DECISIONS-LOCKED.md?" If yes, it is a product decision — ⛔ STOP and escalate. STOP when the choice affects user-facing behavior, a public API contract, architecture, or a locked decision. In this TypeScript monorepo "public API" means a NestJS controller / DTO contract, a React route / props contract, or a FastAPI RAG-sidecar endpoint — changing the shape of any of these is a product decision, not a refactor.

External-AI tool-approval boundary (R7)

Agents must NOT use external AI tools or services that the task does not require. If such a tool is genuinely needed, the agent must request Luba's explicit permission BEFORE use. Using an external AI tool/service without permission is a hard violation. This boundary is absolute and sits at the same priority as the no-product-decisions rule itself: convenience is never authority.

Log a refusal the same way as an avoided product decision:

AVOIDED-EXTERNAL-TOOL: [date]
Tool/service I was tempted to use: [name]
Why the task did not require it: [reason]
What I did instead: [in-repo path] OR "requested Luba approval"

File	Read When
rules/no-product-decisions.md	Language rules for identifying product decisions
rules/baseline-regression-check.md	Protocol for regression baseline before guardrail changes
rules/session-completeness-rule.md	What "done" means under this constraint