muxover/deeptrace

v0.3.0MIT

Evidence-based debugging for AI agents: map the project, run it, trace what actually executes, then report findings tied to file:line.

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Codex plugin (plugin.json, .agents/plugins/marketplace.json): codex plugin marketplace add muxover/deeptrace.
Cursor plugin (.cursor-plugin/) for Customize → From GitHub Repository.
Install for any skills-aware agent with npx skills add muxover/deeptrace.
Release notes for every version and a release workflow that runs CI, then creates the GitHub release from them.
A test that keeps the version identical across every plugin manifest and the changelog.

The Claude Code marketplace is named deeptrace; install with deeptrace@deeptrace.
Node, JavaScript, and TypeScript are profiled with node --cpu-prof, which covers process.exit(), ES modules with top-level await, and tsx.
CI uses the current action versions and can be called from the release workflow.

Claude Code plugin and marketplace manifests: install with /plugin marketplace add muxover/deeptrace.
trace-http.py --repeat N --concurrency C fires a request in parallel and summarizes statuses, latency percentiles, and distinct response bodies, for live race and idempotency checks.
trace-http.py shows the body of error responses, and adds --show-headers and --insecure.
trace.py labels every exception as caught (with where), uncaught, or exit, and reports the program's exit code.
trace.py --returns records return values, -m traces a module like python -m, and --exclude trims the scope.
trace-ui.py --screenshot, plus source locations on console messages and stacks on page errors.
trace-go.py prints call counts and goroutines seen, and adds --stack, --max-lines, --output, and program args after --.
trace-rust.py prints the flamegraph's hot frames as text.
run.py passes args after -- to the command, trims output to the last --max-lines lines, keeps full output with --log, and ends with a summary.
recon.py finds entry points from manifests and from real main functions.
The skill adds a verify-the-fix step: rerun the evidence after a change.
A regression suite, with end-to-end runs against a real browser and Delve in CI.

Script paths in the skill resolved against the project instead of the skill folder; they now use ${CLAUDE_SKILL_DIR}.
trace-ui.py network timings were meaningless (an epoch timestamp was subtracted from a relative value).
trace-ui.py crashed when the page failed to load.
trace-go.py traced every function by default, which broke the run; it now scopes to the project's module.
trace-rust.py passed flags placed after the project path to the program, and --unit-test ran the binary instead of the tests in the fallback.
run.py ran yarn and bun projects with npm, ran make test when no such target existed, and left child processes running after a timeout.
trace.py also traced sibling directories whose names start with the project's name (/proj matched /proj2), and did not restore sys.argv/sys.path.
recon.py counted every DEBUG as a BUG marker.

The DeepTrace skill: six analysis levels, execution simulation, and a fixed report format with severities and confidence.
Security, performance, API, and UI domain lenses.
Toolkit: recon.py, run.py, and tracers for Python, Node, Go, Rust, live HTTP services, and browser UIs.
Five worked examples.