Agent Skills: Review Recent Sessions

Review Recent Sessions

Review multiple recent sessions from the current project directory to identify cross-session patterns.

Do not use nested subagents. This workflow may dispatch first-level reviewer and synthesis agents. Those agents must read the provided files directly and must not dispatch additional subagents.

Prerequisites

• The ed3d-extending-claude plugin must be installed.
• The ed3d-session-reflection plugin must be installed (provides the conversation-reviewer agent and reduce-transcript.py script).
• The current session's transcript path must be available (to determine the project directory).

Invocation

The user may invoke this as:

• /review-recent-sessions — review last 5 sessions
• /review-recent-sessions 10 — review last 10 sessions

Steps

1. Find the project's session directory

Use the current session's transcript path to determine the project directory. The transcript path looks like:

~/.claude/projects/-Users-ed-Development-.../SESSION_ID.jsonl

The directory containing it is the project's session directory.

If you cannot determine the project directory, ask the user.

2. List recent sessions

Find the most recent JSONL files in the project directory, sorted by modification time, limited to the requested count (default 5).

ls -t "<project_session_dir>"/*.jsonl | head -<count>

Exclude the current session's transcript (the user doesn't want to review the review session itself).

If fewer than 2 sessions are found, tell the user there aren't enough sessions to do a cross-session review and suggest using /review-session instead.

3. Reduce all transcripts

Create a working directory:

mkdir -p /tmp/session-review-batch

For each session, run the reduction script:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/reduce-transcript.py" "<session.jsonl>" "/tmp/session-review-batch/reduced-<N>.txt"

This can be done in a single bash command with a loop.

4. Dispatch parallel reviewers

For each reduced transcript, dispatch a conversation-reviewer agent in the background:

Transcript path: /tmp/session-review-batch/reduced-N.txt Write your findings to: /tmp/session-review-batch/findings-N.md

Read the transcript, analyze it, and write your findings following your output format. Do not dispatch or invoke any subagents. </parameter> </invoke>

Dispatch ALL reviewers in a single message to maximize parallelism. Tell the user you've dispatched N reviewers and are waiting for results.

5. Synthesize findings

Once all reviewers complete, dispatch a general-purpose Sonnet agent to synthesize:

Read all findings files in /tmp/session-review-batch/findings-*.md

Produce a synthesis that identifies:

1. Recurring patterns — issues that appear across multiple sessions. These are the highest-value findings because they represent systematic problems.

2. Progression — is the user getting better or worse at prompting over time? Is the agent handling certain tasks better or worse?

3. Highest-impact recommendations — across all sessions, which recommendations would have the biggest effect? Prioritize:

   • CLAUDE.md changes (things the user keeps correcting)
   • Hooks (behaviors that should be enforced automatically)
   • Skills/workflows (multi-step processes that keep being done manually)

4. Session-specific highlights — any single-session finding that's particularly noteworthy even if it didn't recur.

Write your synthesis to /tmp/session-review-batch/synthesis.md

Format as Markdown. Be specific — reference which sessions showed which patterns. Be concise — this is a summary, not a repetition of individual findings. Do not dispatch or invoke any subagents. </parameter> </invoke>

6. Present synthesis

Read /tmp/session-review-batch/synthesis.md and present the full synthesis to the user.

If any individual session findings are particularly interesting, mention that the user can find per-session details in /tmp/session-review-batch/findings-N.md.