Claude calls Edit with file instead of file_path. The tool rejects it. Claude retries with the right parameter, burning tokens and time. Tomorrow it makes the same mistake. Next week, 23 more times. Across your team, hundreds of wasted retries — all from the same mismatch.
dp captures every failed tool call, surfaces the patterns, and shows you exactly what to fix — or what to build next.
Think of it like desire paths on a campus — worn trails through the grass where people actually walk. You don't fight the path. You pave it.
# 1. Hook into Claude Code (one-time setup)
$ dp init --source claude-code
✓ Configured Claude Code hooks
PostToolUseFailure → dp record --source claude-code
# 2. Use Claude Code normally — failures get recorded in the background
# 3. A week later: what's been happening?
$ dp paths --top 5
RANK PATTERN COUNT FIRST_SEEN LAST_SEEN
1 search_files 47 2026-01-15 2026-02-08
2 Edit:file 23 2026-01-18 2026-02-07
3 execute_command 18 2026-01-20 2026-02-08
4 write_file 12 2026-01-22 2026-02-06
5 list_directory 8 2026-02-01 2026-02-05
# 4. "search_files" keeps failing — what real tool is it closest to?
$ dp similar search_files
CANDIDATE SCORE
Grep 0.82
Glob 0.64
# 5. Map the hallucination to the real tool
$ dp alias search_files Grep
✓ Alias: search_files → Grep
# 6. Or better — 47 sessions tried "search_files". That's a feature request.
# Build an MCP tool called search_files that wraps Grep.
# dp just told you where to pour the concrete.47 failures, one pattern, one fix. The AI is telling you what the tool should have been called. That's a desire path.
🔍 Pattern Detection — Aggregate failures into ranked paths. See what's breaking most, not just what broke last.
🧠 Smart Suggestions — Levenshtein-powered similarity engine finds the real tool name behind every hallucination. CamelCase and underscore-aware.
🔗 Alias System — Annotate patterns with what they should map to (search_files → Grep). Aliases appear in dp paths output and short-circuit dp similar. Use them to build your rosetta stone, then act on it — add instructions to CLAUDE.md, build MCP tool wrappers, or lobby for better tool names upstream.
🔌 Plugin Architecture — Extensible source plugins. Claude Code ships built-in. Write your own in ~50 lines of Go.
📊 Full Telemetry — Track all tool calls (not just failures) with --track-all. Success rates, usage patterns, session timelines.
💾 Zero-Config Storage — Embedded SQLite, pure Go, no CGo. Just works. Single file at ~/.dp/desires.db.
📤 Export Anything — Dump raw data as JSONL or CSV. Pipe through jq for ad-hoc analysis, feed into dashboards, or back up with dp export > backup.jsonl. Supports filtering by date and data type (failures vs all invocations).
🖥️ Beautiful Output — TTY-aware tables with bold headers. --json everywhere for scripting.
⚡ Async & Silent — Hook execution is async. dp never slows down your AI assistant.
🏗️ Cross-Platform — Linux, macOS, Windows. amd64 and arm64. Single binary, zero dependencies.
60 seconds from install to insights:
# Install
go install github.com/scbrown/desire-path/cmd/dp@latest
# Hook into your AI tool
dp init --source claude-code
# (Use Claude Code normally for a while...)
# What's failing?
dp paths
# Deep-dive a pattern
dp inspect read_file
# Fix it with an alias
dp alias read_file Read
# See your aliases
dp aliasesgo install github.com/scbrown/desire-path/cmd/dp@latestgit clone https://github.com/scbrown/desire-path.git
cd desire-path
make installPre-built binaries for Linux, macOS, and Windows available on the Releases page.
| Command | Description |
|---|---|
dp record |
Record a failed tool call from stdin JSON |
dp ingest |
Ingest tool call data via a source plugin |
dp init |
Set up automatic recording from an AI tool |
| Command | Description |
|---|---|
dp list |
List recent desires with filtering |
dp paths |
Show aggregated patterns ranked by frequency |
dp inspect |
Deep-dive a specific pattern with histograms |
dp stats |
Summary statistics and activity overview |
dp export |
Export raw data as JSON or CSV |
| Command | Description |
|---|---|
dp similar |
Find similar known tools via string similarity |
dp alias |
Create or update a tool name mapping |
dp aliases |
List all configured aliases |
| Command | Description |
|---|---|
dp config |
View or modify dp settings |
📖 Every command supports
--jsonfor machine-readable output and--helpfor details.
# Failures only (default)
dp init --source claude-code
# Everything — failures AND successes
dp init --source claude-code --track-allHooks into Claude Code's PostToolUseFailure (and optionally PostToolUse) events. Async execution, zero impact on your workflow.
| Tool | Status |
|---|---|
| Gemini CLI | Planned |
| Cursor | Planned |
| Kiro CLI | Planned |
| OpenCode | Planned |
🔌 Want to add your tool? The plugin interface is ~50 lines of Go. See Writing a Source Plugin.
graph LR
A[AI Tool] -->|hook payload| B[Source Plugin]
B -->|Extract| C[Universal Fields]
C -->|Ingest| D[SQLite]
D -->|Query| E[dp list / paths / stats]
E -->|Analyze| F[dp similar / inspect]
F -->|Fix| G[dp alias]
style A fill:#e1bee7,stroke:#7b1fa2,color:#000
style B fill:#bbdefb,stroke:#1565c0,color:#000
style C fill:#c8e6c9,stroke:#2e7d32,color:#000
style D fill:#fff9c4,stroke:#f9a825,color:#000
style E fill:#ffccbc,stroke:#d84315,color:#000
style F fill:#b3e5fc,stroke:#0277bd,color:#000
style G fill:#dcedc8,stroke:#558b2f,color:#000
Data flow: Hook fires → source plugin parses the payload → universal fields extracted → stored in SQLite → query, analyze, and fix with the CLI.
# See all settings
dp config
# Change database location
dp config db_path /path/to/desires.db
# Default to JSON output
dp config default_format json
# Customize known tools for suggestions
dp config known_tools Read,Write,Edit,Bash,Glob,Grep,MyCustomToolConfig lives at ~/.dp/config.toml. See the Configuration Reference for all options.
Full documentation available at scbrown.github.io/desire-path:
- Introduction — The what and why
- Getting Started — Zero to insights in 5 minutes
- Concepts — Desires, paths, aliases, invocations
- Command Reference — Every command, every flag
- Integrations — Claude Code setup, plugin authoring
- Architecture — Data model, storage, plugin system
# Build the docs locally (requires mdbook)
make docs
# Serve with live reload
make docs-serveContributions welcome! The plugin system is specifically designed for community extensions.
Quick wins:
- Add a source plugin for your favorite AI tool
- Report desire paths you've discovered (meta!)
- Improve documentation
MIT — do what you want with it.
Every failed tool call is a feature request from the future.