GitHub
Docs Replay a session

Replay a Claude Code or Codex session step by step

Replay Claude Code and Codex sessions step by step: commands, output and diffs, from your local session history. Nothing is re-run.

coding.sh keeps a searchable history of your Claude Code, Codex and other coding agent sessions, and replays any of them one step at a time: each prompt, reply and tool call, with the command output and the diff each step made. It runs locally, from the agent's own records. Nothing is re-run.

Replay a session

The installer sets up capture hooks for the agents it finds. Import the sessions from before you installed once:

shell 
curl -fsSL https://coding.sh/install | sh
coding.sh ingest --backfill --since 30d

Find a session, then replay it:

shell 
coding.sh sessions --agent codex --since 7d     # or --agent claude-code
coding.sh search "rate limit"                   # full-text search over prompts, replies and tool calls
coding.sh replay <session>                      # the steps, and the files they touched
coding.sh replay <session> --step 12            # what step 12 did: command, exit code, output, diff
coding.sh replay <session> --step 12 --file src/app.ts   # that file as it was after step 12
coding.sh replay <session> --step 5 --to 12     # what steps 5 to 12 changed, per file

A session id can be a unique prefix of 4 or more characters, or latest. All options: coding.sh replay.

Replay it in the viewer

coding.sh ui opens the viewer on 127.0.0.1. On a session page, the Replay button (or R) plays the session back.

KeyDoes
SpacePlay and pause
Arrow keys, j, kStep
sChange the speed
EscLeave

A scrubber jumps anywhere, with marks on edits and failures. The pane shows the step, the files touched so far, and three tabs: this step's diff, a file as of this step, and what changed since an earlier step. The address ends in ?step=N, so a copied link opens that step.

See a timeline instead

coding.sh show <session> prints turns, tool calls, subagents and latency. A long session shows one line per phase (start, each resume, each compaction), and --all lists every event.

How files are rebuilt

File content comes from the agent's own tool calls. A full write or a full read sets a file, and edits (Edit, MultiEdit, str_replace, apply_patch hunks) then apply in order. A file a step created did not exist before that step.

Claude Code records each file before an Edit or Write, so its files are known from their first edit. A session recorded with coding.sh run -- <cmd> starts from a snapshot of the working tree.

What it does not do

  • It does not re-run commands or call the model. One exception: --fork sends a step's model call again with a changed AGENTS.md or another instructions file, for a session captured through the proxy (see Fork from a step).
  • A file edited before any step showed all of it reads as unknown, with the reason, until a later read or write. The step then shows the agent's own diff instead of a guess.
  • A shell command that changes files (sed -i, rm) changes them in the replay only when the agent recorded the change.

Questions

Which agents can I replay?

Any captured session. History import works for Claude Code, Codex, OpenCode, Cursor CLI, the Antigravity CLI, Gemini CLI, GitHub Copilot CLI, Factory Droid, Goose, Pi and T3 Code. Amp and commands wrapped with coding.sh run are recorded from install on.

Can my agent read an old session itself?

Yes. coding.sh init registers an MCP server with Claude Code, Codex, Cursor and OpenCode. Its session_replay tool returns the same as coding.sh replay: the steps, one step with step (and a file with file), or a range with from and to.

Can I watch a session while it runs?

coding.sh show latest --follow prints new events as the hooks write them. In the viewer, a running session's page adds new events every 4 seconds.

Where is the history stored?

In ~/.coding.sh: a SQLite index and a blob store for large payloads. Redaction of keys and tokens runs before anything is written.

Next