Claude Session Handoff

Claude Session Handoff

Recover the state of a prior Claude Code session, then confirm what actually exists in the current workspace before continuing.

Workflow

1. Find the right session

Default to the current project path unless the user gives a specific JSONL file.

For the common handoff case, do not search every Claude transcript first. Bias toward the current repo, the most recent top-level session, and recent activity.

Run the helper script:

python3 scripts/claude_session_handoff.py --project-path "$PWD"

Useful variants:

python3 scripts/claude_session_handoff.py --project-path "$PWD" --list
python3 scripts/claude_session_handoff.py --project-path "$PWD" --date YYYY-MM-DD
python3 scripts/claude_session_handoff.py --session /path/to/session.jsonl
python3 scripts/claude_session_handoff.py --recent 10
python3 scripts/claude_session_handoff.py --project-path "$PWD" --date YYYY-MM-DD --contains "anchor phrase" --from-match

Use these defaults:

• Start with the current repo and the newest top-level session.
• If the user implies "today", resolve that to the exact local date and pass --date YYYY-MM-DD.
• Read the last meaningful turns first. Do not read the entire transcript by default.
• Use --list when multiple Claude sessions exist for the same project and you want the candidate set before choosing one.
• Use --recent when the repo moved and the stored cwd no longer matches the current path.

2. Read only the meaningful turns

The helper script already filters the noisy parts, but keep these rules in mind:

• Ignore subagents/ files unless the user explicitly wants delegated-session context.
• Ignore local-command pseudo-messages such as `, , and `.
• Ignore assistant messages that are only [thinking] or other non-user-visible placeholders.
• Treat images as context markers, not text that needs verbatim transcription.

Default to the most recent meaningful turns only. Large transcripts should be narrowed before you read deeply.

If the session is large and the user wants a specific earlier point, ask for one anchor only:

• a distinctive phrase
• a short quoted line
• an approximate time

Prefer phrase matching over line numbers. Use --contains "anchor phrase" --from-match to jump to the relevant point without reading the full transcript.

3. Reconstruct the handoff

Summarize the session in this order:

1. The user's original goal
2. The options, decisions, or constraints already discussed
3. What actually happened in the session
4. What did not happen yet
5. The exact next step

Call out failures explicitly. If the session ended on an API error, tool error, overload, or interrupted command, say so instead of implying the work completed.

4. Verify the workspace separately

A Claude session log tells you conversation state, not code state. Confirm the actual workspace before continuing:

• Check whether the repo exists and whether it is a git repository
• Inspect the file tree with ls or rg --files
• Run git status --short when git is present
• Read the touched files or run targeted verification when the session claims code was written

If the workspace is empty or unchanged, say that explicitly. Do not infer implementation from discussion alone.

For interrupted-session recovery, assume the last discussed change may not have landed until you verify it in the workspace.

5. Report back concisely

Your handoff should usually include:

• The selected session file
• Whether the prior session was exploratory or implementation-focused
• Whether any code or files exist in the current workspace
• The unresolved next step

If you used an anchor phrase, also include the matched timestamp and make clear that the summary starts from that point rather than from the beginning of the session.

Resources

scripts/

• claude_session_handoff.py: Find relevant Claude Code session files, filter meaningful user and assistant turns, and print a compact handoff-oriented summary

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: buddyh
• Source: buddyh/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.