/nexus

Intent Clarification (absorbed from Cipher)

Purpose: Interpretation method for ambiguous user requests. Read when: The request is vague and routing depends on resolving intent first. See also: After resolving intent, use § Routing Decision Output below to present the chosen chain and material alternatives.

Methodology for decoding ambiguous user intent before agent routing. Previously a standalone agent (Cipher), now integrated as a Nexus capability.

The Three Laws

1. No Interpretation Without Context: git log → .agents/PROJECT.md → conversation history. Context reveals intent. Words are noise.
2. Ambiguity is Sin, Over-Questioning Also Sin: Context clear → Proceed. 2+ valid paths → Ask. Safe default → Proceed. Don't block flow.
3. Never Hide Assumptions: Always state "I interpreted this as..." "I'm assuming that..." Hidden assumptions are time bombs.

Process: GATHER → READ → DECIDE → OUTPUT

Tone & Scope Interpretation Patterns

Precedence rule (tone vs GATE): frustrated tone changes how much is asked, never whether something is asked. When GATE fires (context_confidence < 0.60 OR 2+ valid interpretations — see routing-matrix.md § Classify Flow), a frustrated tone COMPRESSES the ask to exactly ONE focused clarifying question (tightest possible framing, safest-default options pre-filled) — it never authorizes proceeding with zero questions on a sub-floor classification. "Use safest default, proceed" applies only when GATE has not fired (confidence ≥ floor, single valid interpretation); it is not a tone-based override of the confidence floor itself.

Scope Detection

CIPHER_GATE Integration

The following rules apply as an internal Nexus capability (previously the Cipher integration protocol):

• Trigger: context_confidence < 0.60, multiple valid interpretations, or missing critical context
• Auto-clarification: Nexus attempts to resolve using gathered context
• Single question: If still ambiguous, ask ONE focused question with options
• Re-evaluate: integrate the answer, re-type the dimensions, and assign a new evidence band; never add a fixed score bonus

When the blocker is missing context (not just ambiguous wording), run the Context Sufficiency Gate (_common/CONTEXT_SUFFICIENCY.md): inventory which context dimensions the outcome needs (GOAL / SCOPE / ACCEPTANCE / INPUTS / CONSTRAINTS / ENVIRONMENT / INTERFACES / AUDIENCE / PRIOR_ART / RISK), retrieve the inferable ones first (Law 1 — git → PROJECT.md → conversation), and batch only the irreducibly-missing blocking dimensions into a single option-driven question. This makes the one allowed question comprehensive and targeted ("give me X, Y, Z") rather than a vague re-ask. The "one question max" rule (Law 2) means one AskUserQuestion turn — it may batch up to 4 dimensions.

Uncertainty Typing — deciding which question the one question is

The confidence band (confidence-scoring.md) answers whether the available evidence clears a routing threshold. It does not answer what is still unresolved — and those are different questions with different remedies. Two requests can both fall below the gate and need opposite treatment: one needs a target named, the other needs a priority chosen. Since Law 2 allows only one question, typing the uncertainty first is what makes that question land.

Type the gap on these six dimensions before asking. Each maps to a distinct remedy, and each has a characteristic way of being asked badly:

Rules.

1. Do not average the six into one number. A run that is certain about five dimensions and blind on one is blocked on that dimension. The evidence band gates whether to ask; the type decides what to ask.
2. Authority is not averaged. Unresolved Authority never auto-proceeds regardless of the evidence band, and never resolves by inference from a broad request. It is the one type where the answer must come from the user (Q23, and the SKILL.md Ask First triggers it feeds).
3. A forming Goal is not a defect. When the user is still deciding, an early binary question freezes the intent prematurely — offer the axis and let them explore. This is the one case where "proceed with a reversible draft" (Q24 tier degradation) usually beats asking at all.
4. Type before batching. The Context Sufficiency Gate batches up to 4 missing dimensions into one question; typing tells you which 4 are worth the user's attention.

Overloaded-Anchor REDIRECT (recipe ambiguity)

Some English anchors map cleanly to one recipe; these do not — the same word fits 2+ recipes, so a bare keyword match would mis-route. When the input's main anchor is one of these AND context (Law 1) does not already disambiguate, run the classify REDIRECT as a single option-driven question (Law 2) before selecting a Recipe. Pick the option from the family axis in SKILL.md ### Recipe Families.

Option-count discipline: a single REDIRECT question presents at most 4 options. When a row's candidate set is larger (e.g. the audit/review/check row), first eliminate candidates that context (Law 1) already rules out; if more than 4 still remain, ask a two-tier question — one broad-category pick first (e.g. "code quality / security / compliance / other"), then at most one narrow follow-up within the chosen category. This mirrors the Context Sufficiency Gate's ≤4-dimension batching bound; never present a 6-7-branch single question.

After redirecting, state the interpretation per Law 3 ("I routed this to <recipe> because …").

Routing Decision Output

Keep the explanation proportional to the ambiguity resolved:

routing_decision:
  intent: "<resolved goal>"
  route: "<recipe, task type, or direct specialist>"
  why: "<one discriminating reason>"
  assumptions: ["<only assumptions that affect the result>"]
  alternatives: ["<only materially valid alternatives and why they lost>"]

Omit empty assumptions and alternatives. Do not repeat the routing matrix, emit confidence arithmetic, or invent rejected options merely to make the explanation look complete.

Anti-Patterns

Learning Loop

Record interpretation patterns in .agents/nexus.md:

vocabulary_corrections:
  - phrase: "[User's phrase]"
    means: "[Actual meaning in this project]"
    context: "[When this applies]"