agent-sessions

Share coding-agent sessions with your team through the git repo you already work in. One folder, .claude/sessions/, holds handoffs from Claude Code, Codex CLI, OpenCode and (experimentally) Gemini CLI; anyone on any agent can pull them back into their own chat.

/sessions:push "Reconcile fabric tables"        Claude Code, end of a session
$agent-sessions pull reconcile                  a teammate in Codex, next morning

Coding agents keep sessions in private, undocumented, short-lived local stores. This tool turns one into three committed files and gives teammates one command to load them.

What a pushed session looks like

.claude/sessions/2026-09-22_reconcile-fabric-tables_prajwal-p_7dcecef6/
├── summary.md      handoff the author's agent wrote at push time: goal, outcome, decisions and why,
│                   files changed, commands and gotchas, open questions, a paste-ready handoff prompt (≤120 lines)
├── transcript.md   the conversation as markdown along the active path (rewinds and compaction handled),
│                   tool calls collapsed, secrets redacted, file reads and query results omitted
└── meta.json       title, author, agent and version, branch, dates, files touched (subagents included), tags, outcome, tokens

Pull reads summary.md first and loads a transcript only when asked, so a teammate gets the reasoning without burning their context window.

Install

Requirements: git and Python 3.9+ (python3, python, or the Windows py launcher). No Python packages.

Then, in each repo you want to share sessions in: /sessions:init in Claude Code, or $agent-sessions init in Codex, or agent-sessions init in a terminal. It fixes a blanket .claude/ ignore rule so only .claude/sessions/ and .claude/settings.json are tracked, drops a README and .gitattributes, registers the Claude plugin in .claude/settings.json, and commits exactly those files. Add --vendor-skill to also copy the portable skill into .agents/skills/ so every teammate's agent discovers it on clone.

Using it

Push flow, whichever agent runs it:

1. locate finds this session in the agent's store, checks the repo (ignore rules, detached HEAD, merges in progress, missing upstream) and prints an outline of the conversation.
2. The agent writes summary.md from the template, using the outline plus its own memory, through write-summary (agents' file tools tend to refuse .claude/).
3. export renders and redacts the transcript, writes meta.json, renames the folder to the final title, and prints a redaction report.
4. You see a short preview and confirm once. --yes skips it.
5. commit --push builds the commit on a temporary index inside .git and pushes to your current branch. Your staged changes, working tree and any in-progress merge are untouched. --branch pushes to a new remote branch instead.

Pushing the same session again updates the same folder and keeps everything after the first push.

Pull flags: --n N how many summaries (default 3), --full <id> one full transcript, --all-branches sessions on unmerged branches, --mine your own.

How the live session is found

No agent tells a shell command which session it belongs to, so locate never guesses:

If more than one session still matches, locate prints CANDIDATES: and the skill asks you which one, then re-runs with --agent and --session-id. Codex fork and compaction stubs are skipped.

Where each agent keeps sessions

What is and is not redacted

GUIDs are never redacted. Placeholders like ${DB_PASSWORD} and <your-key> are left alone. If a high-confidence pattern still matches after redaction, export deletes the transcript and refuses. Treat a pushed transcript the way you treat a chat log: read the preview before you say yes.

Sandboxes and permissions

Codex's default sandbox blocks network access, so commit --push may fail; approve the escalation when Codex offers it, or run the printed MANUAL command in your own terminal. The temporary index lives inside .git, which stays writable. In OpenCode, allow the skill's script in opencode.json permissions to avoid a prompt per step.

Layout of this repo

agent_sessions/                   the package (stdlib only): cli.py, core.py, model.py, adapters/, templates/
scripts/sessions.py, sessions.sh  entry point + shim (finds Python, always exits 0, prints STATUS)
scripts/setup.sh, setup.ps1       teammate bootstrap: --claude, --skills, --repo <path>
skills/{push,pull,init}/          the Claude Code plugin skills
agents-skills/agent-sessions/     the portable Agent-Skills skill, with a bundled copy of the package
tools/sync_bundle.py              keeps that copy identical (`--check` in CI)
tests/                            python3 -m unittest discover -s tests
.claude-plugin/                   plugin.json (name: sessions) and marketplace.json (name: agent-sessions)

Test the Claude plugin from a checkout with claude --plugin-dir /path/to/agent-sessions. Test the portable skill by copying agents-skills/agent-sessions into ~/.agents/skills/.

Contributing an adapter

An adapter is one file in agent_sessions/adapters/ that turns a native store into Session events with canonical tool kinds (shell, read, edit, write, search, web, agent, mcp, ask). The renderer, result policy, redaction and files-touched logic never see agent names. Start from gemini.py (small) or codex.py (complete). A real session sample from your agent, with secrets removed, is the most useful contribution.

License

MIT