review-agent-session — Claude Code skill for session reviews
The review-agent-session Claude Code skill examines recorded agent sessions, including prompts, tool calls, errors, and file changes. Ask it to trace a failed run, repeated detours, or session costs.
MIT ★ 149 updated 2026-11-16
npx skills add jackulau/coding.sh --skill review-agent-session -a claude-codeWhat review-agent-session does
The skill reads saved agent sessions through the coding.sh MCP server and explains what happened: prompts, tool calls, errors, changed files, and tests. It can also report on subagents, token use, and cost.
It can open the current session, find earlier ones by agent, repository, time, or failure flag, and search for words used in a run. Long timelines can be inspected by turn, while file replay shows a file at a selected step. Session records must be stored on the machine, and the coding.sh MCP server must be running. The skill reads data without changing it.
What it helps with
- locate the turn where an agent run failed
- find an earlier session by repository or words used in it
- trace repeated actions and derail points
- inspect a file as it appeared at a particular step
- check session token use and cost
Example prompts for the agent
- Explain where the agent went wrong in the current session.
- Find the earlier run in this repository and show why it kept repeating actions.
- Report this session's cost and show
src/app.tsat the selected step.
How to install review-agent-session
1Skills CLI (Vercel)
npx skills add jackulau/coding.sh --skill review-agent-session -a claude-code2Manually: put the files in place
~/.claude/skills/review-agent-session/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/review-agent-session/
1Skills CLI (Vercel)
npx skills add jackulau/coding.sh --skill review-agent-session -a codex2Manually: put the files in place
.agents/skills/review-agent-session/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add jackulau/coding.sh --skill review-agent-session -a cursor2Manually: put the files in place
.cursor/skills/review-agent-session/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install jackulau/coding.sh packages/coding.sh/skills/review-agent-session2Manually: put the files in place
.github/skills/review-agent-session/SKILL.md1Skills CLI (Vercel)
npx skills add jackulau/coding.sh --skill review-agent-session -a gemini-cli2Platform CLI
gemini skills install https://github.com/jackulau/coding.sh --path packages/coding.sh/skills/review-agent-sessionManually: put the files in place
.devin/skills/review-agent-session/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about review-agent-session
How do I install review-agent-session in Claude Code?
Run npx skills add jackulau/coding.sh --skill review-agent-session -a claude-code. The skill file is placed at ~/.claude/skills/review-agent-session/SKILL.md.
What does review-agent-session do?
It reads recorded agent sessions and explains the prompts, tool calls, errors, file changes, tests, subagents, token use, and cost.
What does review-agent-session need to run?
It needs session records stored on the machine and the coding.sh MCP server. The skill uses read tools and leaves the recorded data unchanged.
Source text
SKILL.md · MIT
--- name: review-agent-session description: "Reads a recorded agent session from coding.sh and explains what happened in it: the prompts, tool calls, failures, file changes, tests, subagents, tokens and cost. Use when the user asks what an agent did, why a run failed, went in circles or cost so much, wants a past session found, or asks you to review your own current session." license: MIT --- # Review an agent session Uses the coding.sh MCP server's read tools. They read sessions recorded on this machine and change nothing. ## Find the session - This session: call `session_current`. - A past one: call `sessions_list`. Narrow it with `since`, `agent`, `repo`, or `flagged: true` for sessions that went wrong. - By something said or run in it: call `search` with `query` (the words), then open the session a hit names. ## Read it 1. Call `session_get` with `sessionId`. Read `session` first: turns, tool calls and errors, tokens, cost, `flags` and `derails`. 2. The timeline is paged. Pass `turn` for one turn, `cursor` for the next page, or `tail: true` for the end. Do not page through a long session from the start. Pick the turn from `turns`, the per-turn index: one with `toolErrors`, or the one running at a derail's `at`. A turn's `at` is milliseconds after `session.startedAt`. 3. A flag names a failure pattern for the whole session. A derail names the point where a long run went wrong, with a `detail`. Their meanings are in `references/flags.md`. Read it when a name is not clear. 4. For a session that spawned subagents, call `subagents` for the tree, or `session_tree` for cost and flags on every node. Open a child with `session_get` and its session id. 5. To see a file as it stood at one step, call `session_replay` with `sessionId` for the numbered step list, then again with `step` and `file`. ## Report - Lead with the answer to what the user asked: what failed, where, and what the agent did next. - Quote the tool call or error that shows it, with its turn number, so the user can check. - Give tokens and cost from `session` when the user asks about cost. For totals over days, call `usage`. - Keep causes you can see apart from guesses. A flag says what pattern occurred, not why. ## When data is missing A result with `pending` was answered before every capture was indexed: call again shortly. If a session the user expects is not there, call `status`. It reports each agent's last capture and the last capture error.
© jackulau · MIT · jackulau/coding.sh