consult-decide

name	consult-decide
description	Planning-only dual-blind-proposer for decision-between-options questions. Reuses sdd-debate's state machine without entering the SDD pipeline. Outputs a consensus / hybrid / divergence verdict in plain prose.
allowed-tools	["Task","Read","Bash"]
auto-activate	false
version	1.0.0
when	[{"command":"/consult"},{"mode":"decision"}]

consult-decide — Dual-Blind-Proposer for Option Decisions (v7.1 Ola 19)

Purpose: When the user asks /consult to choose between two (or three) options — tech stack, architectural pattern, library — fan out two blind proposers, each defending one option, then converge mechanically via the existing sdd-debate state machine. No SDD change, no spec, no tasks — pure planning.

Difference from sdd-debate: That skill runs INSIDE the SDD pipeline post-sdd-design. This skill runs STANDALONE from /consult. They share _state_machine.py but write to different engram topic keys and never persist a design artifact.

---

Inputs

Field	Type	Required	Description
question	string	yes	The raw user question (Spanish or English)
force_debate	bool	no	True when --debate flag is present. Skips keyword auto-detect.
domain_hint	general | tech	no	Default: detected from question keywords

---

Auto-trigger

Skill is invoked from /consult when the question matches any of (case-insensitive, en + es): vs, or, o, versus, elegir entre, decidir entre, comparar, cuál es mejor, which is better, compare. Explicit --debate flag overrides detection and forces invocation.

---

Workflow

Phase 1 — Parse options

Split the question on the separator keyword. Trim. Accept 2–3 options.

Example: "PostgreSQL or MongoDB for this feature?" → ["PostgreSQL", "MongoDB"].

If parsing yields <2 options, abort and tell the user the question wasn't structured as a choice.

Phase 2 — Pick proposer agent

Default: agent007:product-expert (general decisions).

If the question contains stack-level keywords (database, postgres, mongo, redis, typeorm, microservice, monorepo, nestjs, kafka, queue, cache, framework, language, typescript, python, go, rust): use agent007:backend-db-expert.

Phase 3 — Fan-out (blind parallel)

Spawn two Task invocations in the SAME orchestrator turn. Each receives:

You are running in adversarial peer-review mode.
You will NOT see the other proposer's reasoning.
DEFEND option <X>. Argue FOR it. Find ≥3 concrete reasons it beats <Y>.
Be specific (cite trade-offs, scale, team, ecosystem) — not generic.
Emit a SubagentResponseV1 envelope with a markdown body under "## Findings".
Each finding is a bullet: `- <area>: <claim>`.

Persist bodies to engram:

• A → consult/decide/<timestamp>/<option-a>
• B → consult/decide/<timestamp>/<option-b>

Phase 4 — Converge via state machine

import sys
sys.path.insert(0, ".claude/skills/sdd-debate")
from _state_machine import run_debate

result = run_debate(
    change_name=f"consult-decide-{timestamp}",
    proposal_a_path=body_a,   # raw markdown — module accepts strings
    proposal_b_path=body_b,
    tier="medium",            # always 1 round, early-exit on convergence
)

The state machine handles hash-merge, stance resolution, verdict derivation. Skill does NOT modify it.

Phase 5 — Render verdict in prose

Verdict	Output template
consensus	"Both proposers converged on . Reasons: ."
hybrid	"Each option wins in different dimensions. Use for ; use for ."
divergence	"Proposers maintained opposing positions. Disagreements: . Decide with the user."

The skill returns prose, not JSON. The orchestrator pastes it directly into the /consult reply.

---

Example

/consult ¿PostgreSQL o MongoDB para este feature de 100 usuarios?
→ Mode: consult-decide (matched "o")
→ Options: ["PostgreSQL", "MongoDB"]
→ Proposer agent: backend-db-expert
→ Fan-out: 2 parallel Task calls
→ State machine: consensus (PostgreSQL — relational fit, ACID, small scale)
→ Reply: "Both proposers converged on PostgreSQL..."

---

Token budget

Two subagent spawns + 1 state-machine pass. No second round (medium tier). ~1.2× the cost of a single deep /consult --deep call. Early-exit when both findings hash-match.

---

Determinism contract

• Same question + same proposer bodies → same verdict (state machine is content-addressed).
• Skill itself adds NO probabilistic branching beyond the proposer LLM calls.
• Topic keys include timestamp to avoid collision across consultations.

---

Friction with reused modules

• _state_machine.py has NO if __name__ == "__main__" guard and exposes run_debate as the public entrypoint. Import path requires sys.path.insert(0, …) because the parent directory uses a hyphen (sdd-debate) which Python cannot import as a package without aliasing. Not refactored per Ola 19 constraint.
• The module accepts raw markdown strings (not just file paths) — see _load_proposal() — so this skill can pass subagent bodies directly without writing temp files.

---

See also

• ../sdd-debate/SKILL.md — pipeline twin (runs inside SDD)
• ../sdd-debate/_state_machine.py — shared convergence engine
• consult-critique/SKILL.md — adversarial review counterpart