{
  "id": 9880234,
  "title": "I built a tiny CLI to see how my coding agent actually worked",
  "url": "https://urgent.news/2026/09/26/i-built-a-tiny-cli-to-see-how-my-coding-agent-actually-worked",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-26T01:44:11.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/jay_stride/i-built-a-tiny-cli-to-see-how-my-coding-agent-actually-worked-5co8"
  },
  "original_language": "en",
  "account": "A developer created a simple command-line interface (CLI) to examine the operations of their coding agent, called Agent Pigeon. This tool reads the local session history of Claude Code and Codex and compiles it into a brief Flight report. The goal was to provide a concise summary of the editing, checking, and retrying process behind a long coding session, instead of just the final result.\n\nThe developer wanted a single command, \"npx agent-pigeon flight,\" that could transform a lengthy coding-agent session into a quick-to-read report. The report displayed essential information such as the provider and session span, edit counts, verification checks, and whether a previously recognized failure had later passed. The final line indicated the most recent verification state.\n\nAgent Pigeon also allows users to open the original session when more context is required. However, if certain data is missing, such as Read tool calls, the CLI will display \"READ: N/A.\" A \"READ: 0\" indicates that the parser could not reliably count the reads.\n\nThe developer experimented with live warnings and interventions, but encountered challenges due to incomplete event history. These gaps made it difficult to determine whether a check was skipped or if the log had failed to capture it. Consequently, they decided not to treat these gaps as reasons for interruption.\n\nThe public CLI operates in read-only mode, meaning it reads completed sessions without making any changes. When the signal is incomplete, observation can be more useful than interjecting. Comparing sessions using \"npx agent-pigeon compare session-a session-b\" allows users to side-by-side compare the observed facts from two sessions, without any model scores or declared winner.\n\nHowever, Agent Pigeon has limitations. Verification detection is heuristic, and custom scripts may be missed. Additionally, READ attribution can be unavailable, particularly when the agent is embedded in broader command orchestration. The FAIL→PASS sequence shows the progression, but it cannot pinpoint which edit led to the pass.\n\nThe session span runs from the first to the last recorded event, including idle time. EDIT counts file events, not lines changed or the size of the diff. To try the CLI, users can run \"npx agent-pigeon@latest flight.\" The source code is available on GitHub, and the developer encourages feedback on what the tool may miss or misread, especially around edge cases or different provider formats.",
  "summary": "Agent Pigeon reads local Claude Code and Codex session history after a run and turns it into a short Flight report. I built it because a final answer rarely shows the path that led there: the edits, checks, and retries in between. From a long session to one report npx agent-pigeon flight I wanted one command that could turn a long coding-agent session into something I could scan in a few seconds.…",
  "key_points": [
    "Developer created CLI called Agent Pigeon to examine coding agent's operations.",
    "CLI generates Flight report with edit counts, verification checks, and final verification state.",
    "Public CLI operates in read-only mode, providing insights into session without making changes."
  ],
  "editors_take": "This development gives users of coding agents a transparent view of the editing and verification process, allowing them to understand and compare the operations behind a coding session.",
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}