Skip to content
Back to skills

Plan Governance Vs Execution Boundary For Adversarial Review

ASecurity

Keep stale-approval/governance remediation out of execution-path pseudocode, TDD, files-to-change, and deliverable acceptance when hardening a GitHub issue plan under adversarial review.

  • 17 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added September 9, 2026
testinggogit

Security analysis

A100/100

Scanned September 9, 2026

npx -y skills add vamseeachanta/workspace-hub --skill plan-governance-vs-execution-boundary-for-adversarial-review --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Plan Governance Vs Execution Boundary For Adversarial Review?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Plan Governance Vs Execution Boundary For Adversarial Review
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/vamseeachanta-plan-governance-vs-execution-boundary-for-adversar/badge)](https://www.skillsdirectory.com/skills/vamseeachanta-plan-governance-vs-execution-boundary-for-adversar)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: plan-governance-vs-execution-boundary-for-adversarial-review
description: Keep stale-approval/governance remediation out of execution-path pseudocode, TDD, files-to-change, and deliverable acceptance when hardening a GitHub issue plan under adversarial review.
version: 1.0.0
category: workspace-hub-learned
tags: [planning, governance, adversarial-review, github, workspace-hub]
---

# Plan governance vs execution boundary for adversarial review

Use when a plan has stale approval-state drift (`status:plan-approved`, `.planning/plan-approved/*.md`, stale README row) and you are hardening the plan through repeated adversarial review waves.

## Core rule

Keep two concerns separate:
1. Deliverable/execution contract
2. Governance/review-state cleanup

If you mix them, adversarial reviewers will correctly flag the plan as internally inconsistent or unexecutable.

## What belongs in execution-path sections

These sections must talk only about the final deliverables and how implementation will validate them:
- `## Pseudocode` / `implement_with_tdd()`
- `validate_contract_docs()`
- `## Files to Change`
- `## TDD Test List`
- `## Acceptance Criteria`

Allowed content there:
- contract/checklist/doc requirements
- exact strings/thresholds/links the deliverables must contain
- test files to add/update
- target pytest commands
- discovery-surface updates like `docs/README.md`

Do NOT put these there:
- removing/relabeling GitHub `status:*` labels
- deleting/replacing `.planning/plan-approved/<issue>.md`
- review-artifact posting requirements
- tests that parse the plan text itself
- assertions about current branch/review/governance state
- conditional logic like "if fresh review still returns MAJOR..." inside implementation pseudocode

## What belongs in a separate governance section

Create a dedicated section such as:
- `## Governance Precondition Before Any Future Approval Reuse`

Put all stale-approval remediation there:
- stale live labels
- stale local approval markers
- review-artifact posting/state reconciliation
- README row hygiene as secondary to real approval gates

Explicitly label it as:
- a review/governance requirement
- not a TDD execution test

## If stale approval surfaces already exist

State the present-tense rule clearly:
- the live approval surfaces are already stale for this draft revision
- they must be removed, superseded, or revision-bound before the draft can be treated as approved again

If allowing revision-binding, require an immutable binding that names:
- exact approved plan file path
- exact approved plan git commit SHA or SHA256 hash
- exact approved review-artifact paths plus provider verdicts
- exact approval-storage surface used

## Evidence hardening pattern

When claiming approval drift, include concrete evidence in the plan itself:
- live label evidence from `gh issue view`
- local marker contents
- marker mtime
- current plan mtime
- latest provider verdicts

Avoid vague claims like "approval is stale" without the evidence mechanism.

## Common reviewer-triggering mistakes

Reviewers repeatedly flag these as MAJOR/MINOR:
- governance cleanup inside `implement_with_tdd()`
- governance assertions in `validate_contract_docs()`
- TDD rows expecting `plan text` as test input
- acceptance criteria that require review posting or label cleanup
- artifact-map table corruption during iterative edits
- stale review-summary text after the underlying issue was already fixed

## Safe structure

1. `Resource Intelligence Summary`
2. `Artifact Map`
3. `Deliverable`
4. `Pseudocode` — deliverables only
5. `Files to Change` — actual files only
6. `TDD Test List` — deliverable tests only
7. `Acceptance Criteria` — deliverable acceptance only
8. `Governance Precondition ...` — stale approval/review-state cleanup only
9. `Adversarial Review Summary`
10. `Risks and Open Questions`

## Verification checklist before rerun

Before sending the next adversarial review wave, verify:
- no governance items remain in Pseudocode/TDD/Acceptance
- no plan-text tests remain in the TDD list
- no pseudo-file/governance rows remain in Files to Change
- governance section contains all stale approval-surface cleanup
- review summary reflects the latest actual rerun, not older stale findings
- artifact map rows still render as valid markdown table rows

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.