agenttrace: local TUI observability for AI coding agent sessions

[luoyuctl]

I built agenttrace as a local TUI observability dashboard for AI coding agent sessions. It parses Claude Code, Codex CLI, Gemini CLI, Aider, Cursor exports, Hermes Agent, OpenCode, OpenClaw, Kimi CLI, and Copilot-style logs.

What it is trying to make visible:

• token and cost burn
• tool failures and retry loops
• hanging gaps and latency
• shallow reasoning / redacted thinking
• per-session health score
• session diffs
• JSON output and CI health gates

Install:

curl -sL https://raw.githubusercontent.com/luoyuctl/agenttrace/master/install.sh | sh
agenttrace --demo

Or:

brew install luoyuctl/tap/agenttrace

I would especially like feedback on:

• parser support for other coding agents
• which metrics are actually useful in daily agent debugging
• whether the TUI surfaces the right first-screen information
• workflows where CI health gates would catch real regressions

Release: https://github.com/luoyuctl/agenttrace/releases/tag/v0.3.7
Demo/playbook: https://github.com/luoyuctl/agenttrace/blob/master/docs/demo-playbook.md

[luoyuctl]

Update for new readers: the current release and maintained onboarding links have moved forward since this discussion was opened.

• Latest release: https://github.com/luoyuctl/agenttrace/releases/latest
• Sample HTML report: https://luoyuctl.github.io/agenttrace/demo-report.html
• Demo playbook: https://github.com/luoyuctl/agenttrace/blob/master/docs/demo-playbook.md
• README Quick Start: https://github.com/luoyuctl/agenttrace#-quick-start

The feedback questions above still stand: parser support, useful daily debugging metrics, first-screen TUI signal, and CI health gate workflows.

[yeaight7]

Thank you for the contribution at awesome-ai-devtools. Your request has been approved and agenttrace is already displayed at categories Terminal Agents and Agent Observability

[luoyuctl]

Update for current readers: agenttrace is now at v0.4.6.

Current release: https://github.com/luoyuctl/agenttrace/releases/tag/v0.4.6

Current README-level support includes Claude Code, Codex CLI, Gemini CLI, Qwen Code, Cline, Aider, Cursor exports, Hermes Agent, OpenCode, OpenClaw, Pi, Oh My Pi, Kimi CLI, Copilot-style logs, and generic JSON/JSONL traces.

Active install paths remain:

curl -sL https://raw.githubusercontent.com/luoyuctl/agenttrace/master/install.sh | sh
brew install luoyuctl/tap/agenttrace
go install github.com/luoyuctl/agenttrace/cmd/agenttrace@latest

The npm package is still unpublished, so npm install -g agenttrace is not an active install path yet.

Feedback is still most useful around parser coverage, first-screen triage, which cost/token/time metrics help in daily debugging, and where JSON/Markdown/HTML reports or CI health gates would catch real regressions.

[armorer-labs]

For daily debugging, the first-screen metric I would want is less of a generic health score and more of a small incident timeline:

• last successful user-visible milestone
• longest idle gap and what preceded it
• repeated tool failure loop, grouped by tool + normalized error
• files/tools touched since the last checkpoint
• first point where cost/token burn diverged from the session baseline

For CI gates, the strongest use case is probably regression testing agent workflows rather than judging one-off sessions. A JSON report that can answer "did this scripted agent task become slower, more expensive, more failure-prone, or touch a broader file/tool surface than last week?" would be very useful.

One extra dimension worth considering is tool authority/safety metadata. For coding agents, not all tool calls are equal: read-only file access, test runs, package installs, shell execution, git writes, and external network calls should probably be separate counters. That makes the report useful not just for productivity debugging, but also for reviewing whether an agent session crossed a higher-risk boundary.