Roadmap

Each entry says what to use in the meantime. Something you need that isn't here? Open a feature request.

Agents with no adapter

Cline CLI, Kiro CLI, Qwen Code and Crush have no adapter, so coding.sh doesn't read their hooks or transcripts. Until one lands, record a run with coding.sh run --. It keeps the command as a session with its duration, its exit code and what changed in the repo:

coding.sh run --agent my-agent -- ./my-agent --task 42

To see the agent's model calls as well, point its API base URL at the proxy, if the agent lets you set one. To ask for an adapter, open an agent support issue.

Adapters not run live

Nobody has run these adapters on a live install yet:

• Cursor CLI. It needs a Cursor login or CURSOR_API_KEY.
• Gemini CLI, older installs. Gemini CLI moved to Antigravity CLI; this adapter reads sessions from older installs.
• Amp. It needs an ampcode.com login.

If you use one of them, run coding.sh doctor after a session and tell us what it shows in an agent support issue. A short "it works" helps as much as a bug report.

Capture on Windows

On Windows, the installer has been checked, and so has capture from OpenCode, Pi, Goose, Factory Droid, GitHub Copilot CLI and Aider. Capture from Claude Code, Codex CLI, Antigravity CLI, Cursor CLI, Amp and T3 Code has not been checked on Windows. Under WSL, coding.sh works as it does on Linux.

Repo tasks from hook sessions

A session recorded by hooks becomes a repo task from the commit that was checked out when it started. Changes that were not committed at that point are missing, and coding.sh eval new says so. For the exact starting state, record the run with coding.sh run --, which snapshots the whole working tree. See Turn failed runs into evals.

One bench command and coding.sh try

One bench command over evals, experiments and agent comparisons, and coding.sh try for a first run. Today coding.sh eval, coding.sh exp and coding.sh compare do this work, and coding.sh bench runs benchmarks on fresh cases.