Claude Handoff

claude-handoff

[claude-handoff: the noisy transcript flows through chf and becomes a clean handoff.md and standing project memory]

[PyPI] [Python] [CI] [Downloads] [License: MIT]

Turn any Claude Code session — even a crashed one — into a clean handoff.md another AI can continue from. And give Claude Code permanent project memory, distilled from your own history.

chf

That's it. Your latest session becomes handoff.md: the conversation without the noise, the files that changed, the commands that ran — opening with instructions to the receiving assistant, so you can paste it straight into Gemini, GPT, claude.ai, or a fresh Claude Code session with zero extra prompting.

[chf -o clipboard in action — five seconds from session to paste-ready handoff]

Claude Code stores every session locally as JSONL (~/.claude/projects/…/*.jsonl), full of tool calls, tool results, thinking blocks and system reminders. Existing exporters dump all of that into markdown. claude-handoff instead produces a handoff document — and, since it can read your entire history, a project memory brief too.

• Zero dependencies. Stdlib only, Python 3.9+. A nine-module package — also shipped as a generated single-file script you can curl and audit.
• Deterministic by default. No API call, no cost, works offline.
• --llm when you want a real summary. Claude, OpenAI or Gemini via your own API key — or --llm claude-cli, which runs your locally-installed Claude Code CLI on your existing Pro/Max plan: no API key at all.
• Noise-free. Drops tool results, thinking blocks, system reminders, subagent chatter, slash-command envelopes. Keeps user intent, assistant answers, files modified, commands run — including the files and commands of subagents (agent-*.jsonl), whose full transcripts stay behind --include-sidechains.
• Project memory. chf --brief distills a project's ENTIRE session history into one living brief — what this is, where things stand, decisions, fixes, conventions, and an ordered resume plan of open threads, every claim citing its session; --install-brief-hook injects it into every new Claude Code session, so Claude starts already knowing the project.
• Safe to paste. Secret-looking strings (API keys, tokens, password=…) are redacted from every output — the handoff you paste into a web chat is egress too. --anonymize goes further for public sharing.

Prerequisites

No third-party Python packages, ever — everything runs on the standard library.

Install

pipx install claude-handoff        # or: pip install claude-handoff

brew install Vasilispapg/tap/claude-handoff   # Homebrew

# or just grab the generated single-file build — stdlib-only, auditable:curl -O https://raw.githubusercontent.com/Vasilispapg/claude-handoff/main/single/claude_handoff.pypython3 claude_handoff.py --list

Installing the package gives you two identical commands: claude-handoff and the short alias chf. Tab completion:

eval "$(claude-handoff --completions zsh)"    # bash works too

60 seconds: pick your situation

A session crashed, hit the usage limit, or you closed the terminal:

chf -o clipboard

…then paste into claude.ai, ChatGPT, Gemini — or a fresh claude session. Works on any old session; nothing needed to be installed before the crash.

Moving work from Claude Code to another model:

chf --fit 32k -o clipboard         # sized to the receiver's context window

"Which session was it where we talked about CORS?"

chf --list --grep "CORS"           # every match, with a 🔍 context previewchf --grep "CORS"                  # or export the newest match directly

Give Claude Code permanent memory of this project:

chf --brief --llm claude-cli       # distill ALL sessions → one cited briefchf --install-brief-hook           # every new session starts knowing it

A real summary instead of the transcript (goal / decisions / state / next):

chf --llm claude-cli               # your Claude Code login — no API key

A claude.ai or ChatGPT web chat instead of a terminal session:

chf conversations.json --list      # each app's data export works as inputchf conversations.json --name "webhook bug"

Project memory (--brief)

Claude Code forgets everything between sessions — but the whole history is on your disk. chf --brief reads every session of the current project and writes one memory document to ~/.claude/briefs/<project>.md:

• a factual session timeline + most-touched files (deterministic, free);
• with --llm, a distilled memory that opens with what this is (product, stack, current state) and where things stand (done / in flight / not started), then decisions with their why, fixed bugs, conventions — and open threads as an ordered resume plan, each with its concrete next action and a [in flight] / [blocked] / [not started] tag. Every bullet cites the session id it came from (chf --name <id> opens the source).

[chf --brief in action — the whole project history distilled into cited memory]

Per-session notes are cached, so refreshing after new sessions only pays for the new ones — and a monster session (beyond ~120k chars) is map-reduced inside the note, so the memory path never truncates: nothing is silently dropped, at any size.

You can also curate what feeds the memory: --exclude <id> leaves a session out (a duplicate, an experiment — bare --exclude opens a numbered picker that remembers: the stored exclusions arrive pre-selected with ✗ and typed numbers toggle, so you edit the set instead of re-picking it), and --keep first:2,last:20 windows a huge history to the founding sessions plus a sliding recent window. Both are sticky — stored in the brief's stamp, so hooks and later refreshes keep honoring them until you change them (--exclude none, --keep all). And chf --brief -o clipboard ships the current brief — distillation included — straight to the clipboard, one paste away from handing your project memory to another model.

The brief also goes beyond the project store: chf --brief --grep X -o - distills a thematic memory (only the sessions that talked about X), and chf conversations.json --brief -o brief.md builds standing memory from a claude.ai or ChatGPT export — every conversation, cited by its id. Both are exports by design (explicit -o only) so they never overwrite the standing brief.

Running graphify? The two tools compose in both directions, zero config: chf --brief -o graphify files the current memory into raw/ — one evolving project-memory.md that the next /graphify --update links into the code's knowledge graph — and once graphify-out/graph.json exists, every brief gains a free ## Code map (communities, hub concepts, cross-community bridges:, labeled flows:) so new sessions know the code's structure, not just its history. The in-project copies (raw/project-memory.md, optional root BRIEF.md) refresh on every brief write — strictly refresh-only, hooks never create files. The full loop lives in docs/GUIDE.md.

chf --install-brief-hook

installs two hooks: SessionStart injects the brief as context (Claude starts already knowing your project — re-injected after /compact too), SessionEnd auto-refreshes the factual part for free. No LLM ever runs from a hook; the distilled part refreshes only when you say so. The brief carries a freshness stamp, and when newer sessions or commits exist the file and the injection don't just warn — they list them (session titles, commit subjects), so a fresh session sees what changed, not merely that something did. Fully local; redaction applies as everywhere.

→ Step-by-step mechanics, the honest cost table, and a full day-with-it walkthrough: docs/GUIDE.md.

Make it automatic

chf --install-hook                 # SessionEnd + PreCompact → handoff to ~/.claude/handoffs/chf --install-brief-hook           # SessionStart/End + PreCompact → project memory (above)chf --install-skill                # /claude-handoff skill → Claude Code drives chf correctly

PreCompact matters: right before Claude Code compacts a long session's context, both hooks snapshot state — the handoff preserves detail that compaction is about to squeeze away, and the brief skeleton stays fresh mid-session.

Both edit ~/.claude/settings.json non-destructively, are idempotent, and have matching --uninstall-* flags. Hook failures never break the host session, and hooks never trigger LLM calls or create files on their own.

The third one is for Claude itself: --install-skill puts a /claude-handoff skill into ~/.claude/skills/ (plus its trigger in ~/.claude/CLAUDE.md), so Claude Code knows the tool's grammar instead of guessing — which session bare chf picks from inside a live session, claude-cli (subscription) vs claude (API key), --fit for token budgets, no interactive pickers from an agent shell. Idempotent, respects a hand-written section, and --uninstall-skill removes both pieces without touching anything else in either file.

What the output looks like

# Conversation handoff
> To the receiving assistant: … you are taking over …
## Session- Project: /home/you/myapp (branch main)- When: 2026-08-20 09:00 → 09:04- Activity: 2 user messages, 2 assistant replies, 4 tool calls
## Files created / modified- auth.py
## Commands run- python -m pytest tests/test_auth.py -q
_🤖 2 subagent(s) contributed to the work above (--include-sidechains for their transcripts)._
## Conversation
_Condensed digest, every turn capped — verbatim messages with --full, a real summary with --llm._
- **🧑 User:** the login breaks on unicode passwords…- **🤖 Assistant:** Found it — ascii encoding. Changed to utf-8, tests pass. _[4 tool calls]_- 🔔 Agent "Update the docs" finished

Every turn is there, condensed to its lead — it's a summarize tool, not an exporter. --full switches to classic verbatim messages, and --llm replaces the digest with a real summary that read everything.

Common commands

chf                                # latest session → handoff.mdchf -i                             # numbered picker; "1,3" or "2-4" merges severalchf --list                         # what sessions do I have? (title · first prompt)chf --list --format json           # the same, machine-readablechf --name "login bug"             # newest session whose title/prompt matcheschf "login bug"                    # same — a non-path argument is a name searchchf --grep "CORS"                  # newest session that *talked about* CORSchf --grep CORS --grep auth        # …that talked about BOTH (AND)chf a.jsonl b.jsonl                # several paths → ONE merged handoffchf --project myrepo               # latest session of a specific projectchf path/to/session.jsonl -o -     # explicit file → stdoutchf -o clipboard                   # straight to the clipboard — go paste itchf --full                         # verbatim messages instead of the digestchf --last 5                       # only the last 5 user turnschf --since 2h                     # only the last 2 hours of the sessionchf --fit 32k                      # sized to fit a 32k-token contextchf --include-tools                # keep collapsed per-tool-call detailchf --include-sidechains           # append full subagent transcriptschf --anonymize                    # public-safe: ~ paths, no emails/IPs/usernamechf --project myrepo --merge       # whole project in ONE handoff, oldest → newestchf --format json -o session.json  # machine-readable handoff
# LLM summaries (goal / decisions / current state / next steps).# --llm composes with EVERYTHING above — picker, --name/--grep/--project,# --merge, --last/--since, clipboard, JSON… (only --fit stays# deterministic-only, by design):chf --llm claude-cli               # your Claude Code login — no API keychf --llm ollama                   # local model — fully offlinechf --llm claude                   # Anthropic API   (ANTHROPIC_API_KEY)chf --llm openai --model gpt-4o    # OpenAI API      (OPENAI_API_KEY)chf --llm gemini --with-transcript # Google API      (GEMINI_API_KEY)chf --llm claude-cli --focus "emphasize the API decisions"chf --grep CORS --llm claude-cli   # find the session, then summarize itchf -i --merge --llm claude-cli -o clipboard   # pick several → ONE summary, pastedchf --since 2h --llm claude-cli    # summarize just the last two hours
# project memory:chf --brief                        # free factual brief (timeline + files)chf --brief --llm claude-cli       # + distilled decisions/fixes/conventionschf --brief --exclude              # numbered picker: edit the sticky exclusions (✗ pre-selected, numbers toggle)chf --brief --keep first:2,last:20 # bound a huge history: founding + recentchf --brief --keep since:30d       # …or by time — the window slideschf --brief --grep auth -o -       # thematic memory: only the auth sessionschf conversations.json --brief -o brief.md   # memory from a claude.ai/ChatGPT export

Where does it look? Sessions live in Claude Code's global store (~/.claude/projects), so you can run chf from anywhere. If your current directory is a project (or a subfolder of one), it scopes to that project's sessions; a parent "master folder" scopes to every project under it; --any ignores the directory entirely. Auto-selection skips nearly-empty sessions (like the stub claude /login leaves behind) so "latest" means your latest real conversation — an explicit path, --name or -i always wins.

Big sessions. Transcripts beyond one pass (~400k chars) are summarized map-reduce style: notes per chunk, then one synthesis — nothing is silently dropped, and finished chunks are cached in ~/.cache/claude-handoff so an interrupted run resumes for free. Chunks run 4-way parallel on API providers; claude-cli and ollama stay sequential by design. In a terminal you get a live progress bar:

[█████████░░░░░░░░░░░░░░░] 3/9 chunks | 4m12s elapsed | ~8m left | summarizing part 4/8 (199,867 chars)…

Sessions with API usage data also get a Tokens line in the header, and every run reports the output's ≈token size.

Privacy & zero-trust

• Nothing is sent anywhere unless you pass --llm — deterministic mode is fully offline.
• Redaction is on for every output, not just LLM traffic: secret-shaped strings (API keys, tokens, JWTs, password=…) are stripped from the handoff itself, hook files, and MCP replies — a pasted document is egress too. --no-redact opts out per run (and is deliberately not allowed in the config file).
• --anonymize additionally collapses your home directory to ~ and replaces emails, IPv4s and your username with placeholders — for pasting into public issues and forums.
• --llm claude-cli and --llm ollama keep everything inside accounts and machines you already control.
• Prompt-injection defense: transcripts routinely embed untrusted text (web pages in tool results, pasted READMEs). Every prompt that consumes a transcript, the handoff preamble, and the brief injection wrapper all frame that content as data, not instructions — pinned by tests. A mitigation, not a proof; the parser itself never executes anything.

Config (optional)

Put defaults you always use in ~/.config/claude-handoff/config.json (CLI flags always win; CLAUDE_HANDOFF_CONFIG overrides the path):

{ "llm": "claude-cli", "fit": "32k", "include_tools": true }

Allowed keys: llm, model, fit, output, include_tools, include_sidechains, max_chars, anonymize, focus. Security switches (no_redact) are deliberately not configurable — weakening redaction must be an explicit per-run choice. A broken config warns and is ignored, never fatal.

Environment variables

claude-cli needs no variable — it shells out to your installed Claude Code CLI, billed to your Pro/Max plan (run claude once to log in).

MCP server

Any MCP client (Claude Desktop, Claude Code, …) can pull handoffs directly:

claude mcp add claude-handoff -- claude-handoff --mcp

Tools: list_sessions (what's on this machine) and handoff (build the document for a session by name/project/path; pass anonymize for a shareable version). Deterministic by default — an MCP client can only trigger LLM summaries when you start the server with --allow-llm.

Troubleshooting

claude-handoff: command not found after pip install pip puts scripts in a user bin dir that may not be on PATH. Use pipx install claude-handoff or brew — both manage PATH — or add ~/.local/bin (Linux) / ~/Library/Python/3.x/bin (macOS) to your PATH.

"No sessions found under ~/.claude/projects" You're on a machine (or user) that hasn't run Claude Code, or your store lives elsewhere — point CLAUDE_HOME at it. Inside a project folder the tool scopes to that project; pass --any to search everything.

It picked the wrong session "Latest" skips nearly-empty stubs but is still just the newest file. Use -i (picker), --name "part of the title", or --grep "something said".

--llm claude-cli fails or asks to authenticate Run claude once and log in (/login). It works even when invoked from inside a Claude Code session — inherited CLAUDE* env vars are scrubbed so the nested CLI authenticates like a fresh one.

--llm claude-cli fails with "Credit balance is too low" An ANTHROPIC_API_KEY exported in your shell was winning over your Pro/Max login and billing an (empty) Console account. Since 0.18.0 the key is scrubbed from the spawned CLI automatically — upgrade if you see this on an older version, or unset ANTHROPIC_API_KEY for the run.

"Set ANTHROPIC_API_KEY … to use --llm claude" API providers need a key in the environment — see the table above. No key at all? Use --llm claude-cli (subscription) or --llm ollama (local).

--fit refuses to combine with --llm / --max-chars --fit sizes the deterministic output on its own. If you didn't type it, your config file probably sets fit — override with an explicit --max-chars removed, or drop the key.

The brief injection warns "sessions newer than this brief exist" That's the freshness stamp doing its job: run chf --brief --llm claude-cli to re-distill (cached — only new sessions are paid for). The factual part refreshes itself if the SessionEnd hook is installed.

Something silently did nothing? Tolerant-by-design paths (corrupt JSONL lines, unreadable files, cache trouble) never crash the run — add --debug (or CLAUDE_HANDOFF_DEBUG=1) to see exactly what was skipped and why. Hooks always report their errors on stderr while still exiting 0.

Garbled characters on Windows Set PYTHONUTF8=1 (the CI runs the whole suite that way).

Full flag reference

Roadmap

• Gemini exports as input (Google Takeout ships HTML only — bring a real, redacted export to build against)
• Session chains: auto-detect /compact-continued sessions and offer to merge the lineage (--follow)

PRs welcome.

How it compares

This space isn't empty — it's fragmented. Pick the tool that matches your situation:

• Exporters — claude-conversation-extractor, claude-code-log, claude-code-transcripts, claude-to-markdown — turn transcripts into readable Markdown/HTML, tool noise included, no handoff framing.
• Cross-CLI session movers — cli-continues (npm i -g continues) reads 16 coding CLIs' native session stores (Claude Code included) and injects a context doc into another terminal tool. Excellent for Claude Code → Codex/Cursor/Gemini CLI; but it can't target web chats, does no LLM summarization, and needs Node 22.5+.
• In-session handoff skills/plugins — thepushkarp/handoff, claude-session-handoff, claude-code-handoff — great if you remember to run them before the session ends; the model writes the summary using your session's context, and the output targets the next Claude session.
• Browser extensions — Handoff, LLM Context Bridge, ContextSwitch — transfer web chats between ChatGPT/Claude/Gemini; they can't see Claude Code sessions.

claude-handoff is the post-hoc, paste-anywhere corner of this map: it works on the JSONL after the fact — old sessions, crashed sessions, sessions that hit the usage limit — needs nothing installed in advance, costs zero tokens by default, can write a real summary when you ask for one (--llm), and produces a document any receiving model can pick up, including claude.ai, ChatGPT and Gemini in the browser or on your phone. And with --brief, it's the only one that turns that history into standing project memory.

Development

git clone https://github.com/Vasilispapg/claude-handoff && cd claude-handoffpython3 -m unittest discover -s tests -v      # the whole suite (no deps needed)python3 -m claude_handoff tests/fixtures/agent_session.jsonl -o -   # smoke runpython3 scripts/build_single.py --check       # single-file build is freshuvx ruff check claude_handoff scripts tests   # lint (config in pyproject)

Runtime code lives in the claude_handoff/ package; single/claude_handoff.py is generated — rebuild it with python3 scripts/build_single.py after any package change (CI fails when it's stale). New parser behavior starts with a redacted fixture in tests/fixtures/ — see CONTRIBUTING.md and AGENTS.md (instructions and invariants for both human and AI contributors).

Learn more

docs/GUIDE.md — a day with claude-handoff: walkthrough, how --brief works step by step, honest cost table, cheatsheet · INDEX.md — file map · docs/DEVELOPMENT.md — architecture, JSONL schema notes, design decisions · AGENTS.md — contributor guide for AI coding agents · CONTRIBUTING.md · CHANGELOG.md

License

MIT

mcp-name: io.github.Vasilispapg/claude-handoff