Skip to content
UpIceLabsPublic

About

πŸ‘» The friendly ghost in your git. Your AI said done β€” Casper makes it prove it. Claim-evidence hooks + a verdict ledger for Claude Code.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Latest commit

Β 

History

29 Commits

Folders and files

Repository files navigation

Casper πŸ‘»

Your AI ghosted you with a "done." Casper caught it.

CodeRabbit reviews your code. Casper reviews your claims β€” every "done" needs evidence, every ruling lands in an append-only ledger, and the roadmap's liveness check asks the question no code reviewer can: is it actually live?

The friendly ghost in your git β€” it keeps the receipts.

Casper, the friendly ghost

CI hook tests license judgment skills collection release

Casper in action: an AI's "done" gets blocked, the verdict lands in the ledger, and calibration scores how often your agent's "done" actually held up β€” 1 in 4 here.

A commit says fix: done, all tests pass β€” but zero tests ran. That's your AI ghosting you: it claims it's finished and vanishes, leaving you the broken code. Casper is the friendly ghost that catches it β€” when a "done" has no proof, it blocks the commit and says:

πŸ‘» Boo β€” that ain't done yet.
   claim-evidence gate: this commit claims completion but has no evidence attached.

Casper is a curated set of Claude Code skills + agents built around one thesis: correction history is the asset, not the model. Its headline act is the judgment toolkit β€” 17 skills + 16 CI-tested, zero-LLM hooks that block unproven "done" claims, keep an append-only verdict ledger, and score how your confidence aged. Around it sits a browsable collection of the planning, verification, and tooling skills we actually run in production.

30-second quickstart

npx casper@latest              # interactive installer β€” mirrors install.sh, no clone needed

That one line is the fast path. Prefer a plugin? /plugin marketplace add ronniepinnell/casper then /plugin install casper. Or clone and run the script directly (install.sh is the floor everything else wraps):

git clone https://github.com/ronniepinnell/casper && cd casper

./install.sh --only refute              # one skill, into ./.claude/skills of this project
./install.sh                            # the whole judgment toolkit (17 skills)
./install.sh --all                      # toolkit + the entire collection (skills + agents)

npx casper@latest --all / --only <name> / --category <name> / --hooks / uninstall all forward straight to install.sh. Then, inside Claude Code: /refute the login fix works.

Non-invasive by design: install.sh copies only what you ask into .claude/ (--global for ~/.claude), writes a manifest, and ./uninstall.sh reverts byte-for-byte exactly what was installed. --dry-run shows the plan without touching anything. Hooks are separate and default-OFF (./install.sh --hooks).

Check an install with ./scripts/doctor.sh (a βœ“/βœ— diagnostic: judgment.json valid? hooks wired? ledger writable? which gates are on?). Brag about your receipts with ./scripts/share.sh β€” it reads the ledger and prints one copy-pasteable line (e.g. "πŸ‘» Casper caught a fake 'done' β€” 3rd this week. 2 gate overrides, 0 in 6 days.").

Add it to your existing repo

The hooks are the enforcement, and they ship default-OFF. Three lines turn them on, tuned to your stack:

cd your-project
npx casper@latest --init     # detects your stack, writes a tuned .claude/judgment.json, wires the hooks ON
casper doctor                # verify: judgment.json valid? hooks wired? ledger writable? which gates on?

casper --init writes judgment.json, live-blocks a fake done-claim, then ends with the mirror: your repo's own merged PRs graded EVIDENCED/UNEVIDENCED

That's it β€” from now on an evidence-free git commit -m "fix: done" is blocked in your repo, and every verdict lands in .claude/verdicts.log. --init asks at most 3 questions (confirm the protected paths for your stack, enable the dangerous-git guard, start the ledger); everything else is detected. Commit .claude/judgment.json so your whole team gets the same gates.

Casper works on three surfaces β€” use one or all:

Surface What it guards How to add
Local hooks (this repo) your commits + edits, before they happen npx casper@latest --init
Pull requests every PR, where teammates can't --no-verify past it add the refute-action GitHub Action (one uses: block)
Model context let any model query your correction history mid-task connect casper-ledger-mcp (claude mcp add)

Nothing is invasive: ./uninstall.sh reverts byte-for-byte, and --dry-run shows the plan first.

See each piece work

Every GIF below is real output β€” the hooks are the actual scripts run against a real repo; the ledger and calibration frames are real tool output.

In your editor β€” /refute breaks a claim inside Claude Code and logs the verdict. This is real claude -p output: it actually ran login('admin','wrong-password'), got True, and diagnosed a full auth bypass:

/refute in Claude Code: it executes the breaking input, finds any password authenticates, and rules REFUTED with evidence

The styled walkthrough of the same flow

/refute walkthrough

The zero-LLM hooks

git/CI, no model, no network.

claim-evidence β€” a "done" commit with no evidence is blocked, then passes once you attach proof.

claim-evidence

spec-citation β€” editing a protected migration halts until you cite the governing spec.

spec-citation

scope-creep β€” a "one-line fix" quietly touching too many files trips the tripwire.

scope-creep

The skills

procedure any model runs.

/refute β€” try to break the claim before believing it.

/refute

/gate β€” no plan without a numeric abort condition; watch it trip.

/gate

/verdict β€” every ruling is one grep-able line; "show every gate we overrode" is a query.

/verdict

/calibrate β€” score how your confidence aged: 3 of 4 "done"s broke.

/calibrate

The judgment toolkit

The flagship. 17 skills that turn "trust me, it's done" into a checkable record.

Command One-liner
/refute Try to break the claim before believing it β€” CONFIRMED / REFUTED / UNVERIFIED
/door Reversible? pick fast. Irreversible? enumerate the lock-in first
/gate No plan without a numeric abort condition
/drift Spec vs code-as-built β€” which one is lying?
/altitude Fix at the cause's layer, not the symptom's
/premortem It already failed β€” write the incident report first
/think Forced thinking moves: invert, second-order, base-rate, analogy, flip, decompose
/verdict Append-only judgment ledger β€” "every gate we overrode" is a grep
/calibrate Score how your past confidence aged; corrections become mechanisms
/escalate Queue the hard call, ship the rest; burn to batch-adjudicate
/precedent Grep prior rulings: FOLLOW or explicitly DISTINGUISH
/sweep Massive audit fan-out β†’ adversarial verification β†’ graded synthesis
/judgment The map + router: given a situation, names the one tool that fits
/merge-train Merge only mechanically-proven-safe PRs; every held PR names its missing fact
/audit-skills Measure your skill library (tokens, bloat, overlap); UNVERIFIED over guessed
/backfill Retro-grade merged PRs: which done-claims shipped with evidence, which shipped unproven

First 60 seconds β€” grade the dones you already shipped:

python3 scripts/backfill.py --since 2026-01-01          # EVIDENCED / UNEVIDENCED per PR
python3 scripts/backfill.py --badge badge.json          # shields endpoint: "evidenced dones: 87%"

Publish badge.json (gh-pages/gist) and embed https://img.shields.io/endpoint?url=<public-url> β€” the score keeps you honest in public.

backfill grades a repo's merged PRs EVIDENCED/UNEVIDENCED and writes the badge JSON β€” this repo scored 50%

The wider suite (refute-action, ledger MCP, badge, report): docs/SUITE.md.

Start here β€” 5 tools: /refute, /gate, /verdict, /door, /calibrate.

Backed by 16 zero-LLM hooks (claim-evidence, spec-citation, scope-creep, dangerous-git, and three telemetry/guard hooks), each with a block-case AND pass-case regression test β€” 24 assertions, run in CI (hooks/judgment/test.sh). Every skill appends a one-line, grep-able verdict to .claude/verdicts.log:

grep 'GATE.*OVERRIDE' .claude/verdicts.log   # every gate we overrode, with who signed it

Full deep-dive β€” thesis, operating loop, hook wiring, the ledger badge, weekly digest β€” lives in MANUAL.md.

The full collection

54 authored units β€” 41 skills + 13 agents

(origin: authored, CI-checked provenance), organized into six categories. Each category page has a table β€” unit, what it does, when to call, and a one-line install. Install one skill, a whole category, or everything.

Category What's inside Units
Planning & Lifecycle Initiative β†’ milestone β†’ epic β†’ task cadence + orchestration 11
Verification & Audit Completion reality-checks, spec/rules conformance, bug hunts, over-engineering review 18
Docs & Context Keep docs, AI-context files, and project records in sync 4
Research & Analysis Generate and pressure-test ideas, names, designs, decisions 6
Meta & Tooling Scaffolding, session control, import/sync, dev-tool integrations 13
Communication Hand work off cleanly β€” to the next session or another agent 2
./install.sh --category verification-and-audit   # a whole category
./install.sh --only spec-audit,find-bugs         # pick individual units by name

Docs

  • MANUAL.md β€” the one deep doc: thesis, operating loop, authoring guide, skill-vs-hook decision tree, troubleshooting.
  • CONTRIBUTING.md β€” how to add a unit; the collection flow.
  • CHANGELOG.md β€” release history (Keep a Changelog format).
  • ROADMAP.md Β· LAUNCH.md

License

MIT Β© 2026 Ronnie Pinnell. All skills and agents are original authored work (origin: authored frontmatter, CI-checked provenance).


Built by the team behind the autonomous software factory where these procedures run in production. Casper is and stays free; if you want the hosted verdict ledger, team calibration dashboards, or CI enforcement at org scale, that's where they live.


Casper is the free, friendly ghost β€” it watches your commits while you're away. Its bigger sibling runs the whole factory while you're gone. πŸ‘»

About

πŸ‘» The friendly ghost in your git. Your AI said done β€” Casper makes it prove it. Claim-evidence hooks + a verdict ledger for Claude Code.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages