Architecture Remediation Planner

Plan an architecture change

Turn an accepted target into a safe, observable, dependency-aware roadmap without editing code or reopening the decision.

Load and classify the source chain

Read ../../resources/references/review-contract.md and ../../resources/references/remediation-contract.md completely. Load one accepted Architecture Decision and validate its source chain.

Choose exactly one planning mode:

• Remediation: bind the accepted Decision to a verified Review. Include only

confirmed unresolved Findings, their exact fingerprints, and their Knowledge IDs. Exclude candidate, rejected, needs-evidence, and resolved Findings.

• Greenfield: bind the accepted Decision to its approved Design Brief and

target architecture. Use Plan 1.3. Do not invent a Review or Finding to make the legacy remediation shape fit: omit finding_ids and finding_bindings. Bind the Brief and Decision bytes. Across the Plan, cover every target runtime and deployment unit, data-ownership record, interface, trust boundary, technology ID, critical flow, and constraint. Keep operational controls explicit in deployment, observability, rollback, stop, and acceptance fields.

Recheck a remediation Finding if its fingerprint, evidence commit, contract, or owning boundary changed. Stop when the Decision is proposed, stale, rejected, superseded, or no longer satisfies the selected source mode.

Plan the work

1. Group remediation Findings by violated invariant and owning boundary. For

Greenfield, group work by target unit, critical flow, dependency, or surviving required constraint. Do not create fake findings.

1. Build a dependency graph covering prerequisites, external dependencies,

sequencing conflicts, authority decisions, and work that will collapse after an earlier slice.

1. Preserve the accepted option and all surviving constraint assessments. Do

not reopen technology, architecture, or preference scoring.

1. For each item, record the source binding, owner, desired invariant or target

outcome, affected unit/flow/constraint, recommended option, alternatives, do-nothing consequence where applicable, effort and uncertainty, change risk, governed flag, dependencies, ordered slices, protection, observability, rollback, stop conditions, and measurable acceptance criteria.

1. Sequence by risk reduction, dependency, and reversibility: evidence and

safety net; compatibility seam; reversible internal change; data or contract migration; consumer rollout; old-path removal after measured acceptance.

For every Greenfield item, name the Brief question and Decision target it implements. Map each slice to at least one runtime unit and one critical flow or explicit constraint assessment. The union of item bindings must cover the full accepted target. Keep data ownership, interface compatibility, trust-boundary controls, and operations visible in the acceptance criteria.

Do not prescribe a shared service because two projects look similar. Require stable shared semantics, aligned lifecycle, explicit ownership, and acceptable coupling. Mark persisted data, public contracts, authorization, production infrastructure, deployment, and destructive effects as governed/high-risk.

Output and validation

Write:

• .architecture/reviews/-remediation.yaml;
• .architecture/reviews/-remediation-plan.md.

Use .architecture-portfolio/reviews/ for portfolio work. Use Plan schema 1.2 for readable remediation plans and Plan schema 1.3 for new Greenfield plans. The machine-readable plan must bind the accepted Decision, and then either the verified Review plus every Finding fingerprint or the Design Brief plus the target architecture. Use only Knowledge IDs present in the Decision selection.

Validate the remediation chain:

python3 ../../resources/scripts/architecture_tool.py validate-plan \
   --decision  --project  \
  --review 

Validate the Greenfield chain:

python3 ../../resources/scripts/architecture_tool.py validate-plan \
   --decision  --project  \
  --design-brief 

Leave completion_evidence empty until evidence exists. A completed item must bind repository-relative evidence files and exact SHA-256 values for every declared evidence type. End with the smallest safe first slice and its acceptance evidence; do not execute it unless the user separately requests implementation.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: qingye-lab
• Source: qingye-lab/hengmu
• License: MIT
• Homepage: https://qingye-lab.github.io/hengmu/

Install and usage instructions live in the source repository linked above.