sdd-converge

name	sdd-converge
description	Drive systemic alignment toward spec target after targeted corrections.
risk	high

sdd-converge

Purpose

Assess and drive the workspace toward systemic alignment with the governance spec. Operates at the strategic level — evaluates the overall delta between current state and target spec, produces a convergence plan, and may invoke sdd-correct iteratively for individual violations within that plan.

When to use

• After one or more sdd-correct passes, residual delta remains
• A pattern of drift is detected across multiple areas (not a single violation)
• Spec alignment score is below threshold
• diagnose or review-architecture returned a systemic divergence

Required protocol

1. Load .sdd/agent-instructions.md
2. Confirm skill is registered in .sdd/skills/registry.json as sdd-converge
3. Query current governance state: sdd ask "current alignment delta"
4. Assess the full delta: what diverges from the spec target?
5. Score alignment: estimate alignment_score (0.0–1.0)
6. Build convergence_plan: ordered list of corrections to reach alignment
7. For each item in plan: invoke sdd-correct or flag for human review
8. Re-assess after each correction: update alignment_score
9. Validate when done: sdd governance validate
10. Return alignment_score, convergence_plan, delta_report, next_corrigir_targets

Allowed CLI

• sdd ask
• sdd governance validate
• sdd runtime status --force
• sdd doctor run

Output format

Return YAML with:

• alignment_score: float 0.0–1.0 (1.0 = fully aligned)
• convergence_plan: ordered list of corrections with priority and target
• delta_report: summary of gaps between current state and spec
• next_corrigir_targets: remaining violations to be addressed by sdd-correct
• policy_result: PASS | WARN | FAIL | BLOCKED

Non-compliance

• Do not declare convergence without running sdd governance validate
• Do not make structural changes without review-architecture gate (strict mode)
• Do not violate mandates in the process of converging toward the spec
• Do not skip alignment_score in the output — it is required for traceability
• Do not invoke sdd-converge for a single isolated violation — use sdd-correct instead