why · git:20261108.ac1df7f · 2026-11-08 · sha256 3b2e5cbe879b9766
why git:20261108.ac1df7fA
Immutable. This exact content never changes and is served at /api/v1/blob/3b2e5cbe879b9766.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_c4zlqkhpemoxb6ze/3b2e5cbe879b9766154dba6c03ad36a2457fc8ed9823931439fd0b3b4061c1ed.json
Then /plugin install why@markdownregistry-c4zlqkhpemoxb6ze in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
---
name: why
description: "Investigate why code is the way it is from evidence — human rulings verbatim, git archaeology, linked issues, incident records — every claim labeled evidence, inference, or guess. For 'why was this built this way'."
aliases: []
metadata:
credits:
investigation-epistemics: pstack
author: Lauren Tan
url: "https://github.com/cursor/plugins/tree/main/pstack"
port: pstack-claude
port-author: Michael Denyer
port-url: "https://github.com/michael-denyer/pstack-claude"
---
# Why
Answer "why is it like this?" from evidence, not vibes. The deliverable is a
cited answer, not a fix; a fix that falls out of the answer goes through its
own lane.
## When to use
- "why was this built this way", "why does this exist"
- "are we sure about Z" before betting work on Z
- "should we do X or Y" where the answer lives in history, not preference
## Sources, in order of authority
1. **Human rulings.** A decision a person made is quoted verbatim with
attribution (commit message, issue comment, PR review, chat log). Never
paraphrase a ruling into your own words; label agent interpretations as
interpretations, and missing originals as missing. A copy of a ruling is
not a new decision.
2. **Code archaeology.** `git log --follow`, `git log -S`, blame on the
deciding lines, and the diff that introduced them. Name the commit and
what problem its message claims to solve.
3. **Linked records.** Issues and PRs reachable from the code (references in
commit messages, `gh issue list --state all --search`, PR review threads) — acceptance
criteria and rejected alternatives live here.
4. **Incident and observability sources.** When configured, query them
through their own tools (Sentry, Datadog, Databricks, dashboards,
incident postmortems) or MCP servers (Slack, Linear, Notion). Cite the
alert, thread, ticket, or postmortem. If a source needs credentials you
do not have, say so and move on.
## Method
1. State the question and what would count as an answer.
2. Fan out source gathering to a read-only subagent when the history is long
(`oh-my-claudecode:explore`); keep synthesis in the lead.
3. Reconcile: where sources conflict, the more authoritative source wins, and
the conflict itself is reported.
4. Write the answer with three labeled tiers — **evidence** (cited),
**inference** (your reasoning from evidence, marked), **guess** (marked).
End with the one action the answer most supports, or "no action".
## Rules
- Every claim about a past decision names its source in the same sentence.
- Absence of evidence is reported as "no record found", never as "no reason".
- The answer never edits code. If the user asks to act on it, hand off to
`execute` or `debug`.