claude

deep-investigation

Skill + three subagents that answer a why-question by enumerating every hypothesis first, then falsifying branches until one mechanism accounts for the whole effect.

2 on GitHubby galilei2050v0.3.0AGPL-3.0updated 4 days ago
/

Plugin deep-investigation from marketplace belay in galilei2050/belay. Commands read from the repo's manifest — 4 days ago.

What it installs

Skills (1)

  • investigate

    Explains a measured effect whose cause is genuinely unknown — a metric that moved, an incident nobody can account for, two sources reporting different numbers for the same thing, a regression that survived the obvious fix. Enumerates every candidate cause as a hypothesis tree before checking any of

Subagents (3)

  • evidence-collector

    Retrieves one narrow slice of fact for exactly one investigation hypothesis from sources reachable from this machine — code, files, logs, git history, databases, internal services and CLIs. Returns the exact command and its raw output, never an interpretation. Use when an investigation needs a fact

  • hypothesis-falsifier

    Adversary for a single investigation hypothesis that already has supporting evidence — verified or partial. Attacks it seven ways (alternative cause for the same evidence, direction and timing, effect size, circularity, counter-case, selection bias, overlap with other verified leaves), gathering its

  • web-researcher

    Establishes facts that live outside this system for exactly one investigation hypothesis — known mechanisms, published benchmarks, vendor and provider incidents, status-page history, policy or platform changes, upstream release notes. Every fact returns with a source URL and a publication date, and

View source on GitHub

More from galilei2050/belay

All 13 plugins →

review-panel

claude

A panel of eight read-only reviewer subagents, dispatched by a PreToolUse hook over every commit of 64+ changed lines.

8 agents · hooks

2

usable-ui

claude

Skill deciding UI wording, control, placement and states, plus five read-only UI reviewers, supplied by a PreToolUse hook while UI is written and once it is committed.

1 skill · 1 command · 5 agents · hooks

2

pr-flow

claude

Skill that writes PR descriptions a reviewer can act on, a CI script whose one backgroundable call carries a branch from a CI verdict through the merge to the deploy that ships it — GitHub Actions or Google Cloud Build — while printing only the failing log, a PreToolUse hook that denies that wait in the foreground so it never freezes the session, another that denies a PR body with no comparison table or diagram, one that narrates the diff, or a title only the code can decode, plus a PostToolUse hook that says what the branch still owes — a push, a PR, a refreshed body, a green CI, a live change — right after each commit and push.

1 skill · 2 commands · hooks

2

agent-knowledge

claude

Keeps the knowledge an agent loads lean and current, starting with CLAUDE.md. A PostToolUse gate holds every CLAUDE.md to one skeleton: an auto-stamped `Updated:` date, mandatory Business decisions and Technical decisions sections, a 200-line limit, no dead paths, no decisions duplicated from a parent file. Ships the skill that writes to that skeleton and an audit CLI.

1 skill · hooks

2

second-opinion

claude

Skill and script that put a question to Gemini through Vertex AI — a model that has not read this session — and bring the answer back as an argument to check, not an authority to follow.

1 skill · 1 command

2

acl-hook

claude

PreToolUse Bash gate enforcing a project ACL. Vendored from clarity-auto-care.

hooks

2

Maintain deep-investigation?

This page updates itself from your repo. Add the badge to your README, or get it Featured to pin it above the organic results with a followed link to your site.

claude & codex plugins13 plugins