Skill Name
code-review (reference addition)
Programming Language
PowerShell
Purpose
The code-review skill selects review depth and ranks hotspots by feel: depth-tiers says "use Tier 3 for high-risk changes" without defining risk, and file risk classification is a three-bucket path heuristic. As agent-authored change volume grows, review depth is the scarce resource and its allocation has no quantitative basis.
Two decades of just-in-time defect prediction (Mockus & Weiss; Kamei et al.; Meta's production Diff Risk Score; CodeScene's delta analysis) established which change signals predict defects — size, diffusion, change entropy, hotspot overlap, missed co-change — and all are computable with git commands the reviewing agent already runs.
This adds that model as a focused reference the Tier 0 context bootstrap applies when drafting the change brief, so depth-tier recommendations and hotspot rankings carry evidence. It informs the existing human-confirmed scoping step — advisory only, never a gate — without creating a competing review workflow.
Prerequisites
None — documentation-only reference file (no executable scripts). Signals derive from git history already available to the skill.
Requirements
Definition of done:
Out of scope — follow-up sub-issues if this lands: script-backed change-risk skill (bash + PowerShell parity, Pester, 80% coverage) emitting the profile as JSON; calibration guidance and eval cases against repo history; schema'd evidence emission via the persisted emission-record contract. Placement question for maintainers at that point: coding-standards (adjacent to its consumer) or experimental (incubation).
Change risk hve core proposal.md
Skill Name
code-review (reference addition)
Programming Language
PowerShell
Purpose
The code-review skill selects review depth and ranks hotspots by feel: depth-tiers says "use Tier 3 for high-risk changes" without defining risk, and file risk classification is a three-bucket path heuristic. As agent-authored change volume grows, review depth is the scarce resource and its allocation has no quantitative basis.
Two decades of just-in-time defect prediction (Mockus & Weiss; Kamei et al.; Meta's production Diff Risk Score; CodeScene's delta analysis) established which change signals predict defects — size, diffusion, change entropy, hotspot overlap, missed co-change — and all are computable with git commands the reviewing agent already runs.
This adds that model as a focused reference the Tier 0 context bootstrap applies when drafting the change brief, so depth-tier recommendations and hotspot rankings carry evidence. It informs the existing human-confirmed scoping step — advisory only, never a gate — without creating a competing review workflow.
Prerequisites
None — documentation-only reference file (no executable scripts). Signals derive from git history already available to the skill.
Requirements
Definition of done:
references/change-risk-model.md: signal taxonomy (likelihood / severity / detectability / recoverability), a scoring rubric producing a named-signal profile with evidence (never a single opaque score), and confidence guidance for shallow-history repos.context-bootstrap.mdupdated: profile drafted at Tier 0 alongside the change brief.depth-tiers.mdupdated: tier recommendation cites profile evidence.Out of scope — follow-up sub-issues if this lands: script-backed
change-riskskill (bash + PowerShell parity, Pester, 80% coverage) emitting the profile as JSON; calibration guidance and eval cases against repo history; schema'd evidence emission via the persisted emission-record contract. Placement question for maintainers at that point:coding-standards(adjacent to its consumer) orexperimental(incubation).Change risk hve core proposal.md