PLUGIN·preview

Find drift between what you asked, what you planned, what was claimed, and what shipped.

A Claude Code plugin for finding drift between what you asked, what was planned, what was claimed, and what actually shipped — sourced from the session transcripts already on your disk.

Install the plugin ↗INSTALL ↓READ THE LAUNCH POST →

withagents.dev/orbit

SKILLS

COMMANDS

HOOKS

EVIDENCE RANKS

FAKES SHIPPED

“Plans are evidence, not history. Assistant claims are claims, not proof. Codebase validation is required before completion is trusted.”

Orbit is a Claude Code plugin that mines existing local session transcripts under ~/.claude/projects/ — current project, filtered project sets, or all projects — and compares user intent against plan evidence, assistant claims, tool evidence, and current codebase truth.

Inventory: 4 skills, 8 slash commands, 2 hooks, 1 Python audit engine. Outputs a static dashboard, evidence JSON, gap-analysis markdown, intent-timeline mermaid, and an append-only intent ledger — all under .claude/audits/.

Iron rule: plans are evidence, not historical truth. Assistant claims are claims, not proof. Codebase validation is required before completion is trusted.

01. ARCHITECTURE

How it's built

MODULE 00

Mine

Walks ~/.claude/projects/**/*.jsonl in time order, normalizes events through a fixed key allowlist, extracts user intents and assistant claims and tool evidence — never duplicates capture.

MODULE 01

Compare

Joins intent against plan items, claims, tool evidence, codebase truth, and validation results. Honors a 12-level evidence ranking — later user instruction beats earlier plan.

MODULE 02

Visualize

Emits gap-analysis.md, evidence.json, intent-gap-flow.mmd, intent-timeline.mmd, plan-execution-matrix.csv, and a hand-written dashboard.html in flat black with lime + blue accents.

02. FEATURES

Mines what already exists

Claude Code already stores sessions locally under ~/.claude/projects/. Orbit reads them in order, never duplicates capture, and never requires Docker or a Python package install.

12-level evidence ranking

Later user instruction > correction > direct session evidence > tool-use > tool-result > codebase state > validation > earlier instruction > written plan > assistant claim > inference > metadata. Honored across every analyzer.

Cross-project review windows

--days N | --since YYYY-MM-DD | --from/--to. Repeatable --project-filter and --all-projects flags surface drift across project families, not just within one repo.

Refusal when proof is missing

Plan omission is not proof never-requested. Missing evidence is marked uncertainty, not absence — gap analyzer never silently downgrades to PASS.

03. QUICK START

# install via Claude Code marketplace
claude plugin marketplace add krzemienski/orbit
claude plugin install orbit@orbit

# 3-day default audit on the current project
/orbit:audit-gaps
# or explicit window
/orbit:review-window --days 7

# Cross-project drift across two project families
/orbit:review-window --project-filter orbit --project-filter blog-series --days 30

# Append today's intents to the durable ledger
/orbit:rebuild-ledger

04. INSTALL

01claude plugin marketplace add krzemienski/orbit
02claude plugin install orbit@orbit
03Run /orbit:audit-gaps to execute the default 3-day audit on the current project
04Inspect outputs at .claude/audits/latest/dashboard.html, gap-analysis.md, evidence.json
05Run /orbit:rebuild-ledger to append today's user instructions to the durable intent ledger