agent-handover

Session handover engine for AI coding agents: checkpointed, resumable, backend-agnostic persistent memory.

AI coding agents (Claude Code, Codex, OpenCode, Cline, ...) are stateless between sessions. Every new session starts with re-explaining the project, the decisions already made, and what was left half-done. agent-handover is the small piece of infrastructure that fixes this: at the end of a session the agent writes a structured handover; at the start of the next one, it resumes from it — even if the previous handover crashed halfway.

Extracted from a personal "AI Team OS" that has run daily handovers across multiple machines and agents since 2025.

The problem

• Context loss: the next session doesn't know what the last one decided.
• Half-written state: a handover that dies mid-run silently corrupts memory. The next session boots from a state that is partly updated — worse than stale.
• Remote memory is fragile: vector DBs and hosted notebooks go down. If your agent's memory has no local fallback, your agent has no memory.

Design

┌─ session ends ──────────────────────────────────────────────┐
│  HandoverEngine(steps, checkpoint, pause_file)              │
│    step 1: collect session facts          ── checkpointed   │
│    step 2: MemoryStore.write(layer=1...)  ── checkpointed   │
│    step 3: MemoryStore.write(layer=2...)  ── checkpointed   │
│    step 4: GitBackend.publish(...)        ── checkpointed   │
└──────────────────────────────────────────────────────────────┘
┌─ next session starts ───────────────────────────────────────┐
│  $ agent-handover check    # 0=clean  1=RESUME  2=completed │
│  read layer2/current-state.md  →  agent has context again   │
└──────────────────────────────────────────────────────────────┘

Three memory layers, all plain Markdown (git-friendly, diff-able, readable by humans and any agent that can read a file):

Three rules learned in production:

1. Every step is checkpointed. An interrupted handover is detected (agent-handover check → exit 1) and resumed without re-running done steps.
2. A PAUSE file is an absolute kill-switch. Agents that write to your repos need a brake a human can pull with touch PAUSE.
3. The filesystem is the source of truth. Remote layers (NotebookLM, vector stores) are optional accelerators; the local copy is always enough to boot.

Install

pip install agent-handover   # or: pip install -e . from a clone

Usage

from pathlib import Path
from agent_handover import Checkpoint, HandoverEngine, MemoryStore, Step, GitBackend

store = MemoryStore("memory/")
cp = Checkpoint(".agent-handover/checkpoint.json")
git = GitBackend(".", paths=["memory/"])

engine = HandoverEngine(
    steps=[
        Step("session_note", lambda: store.write(1, notes, tag="refactor")),
        Step("current_state", lambda: store.write(2, snapshot)),
        Step("publish", lambda: git.publish("handover: session sync")),
    ],
    checkpoint=cp,
    pause_file=Path.home() / ".agent_handover" / "PAUSE",
)
engine.run()  # resumes pending steps automatically after a crash

In your agent's bootstrap (CLAUDE.md / AGENTS.md):

agent-handover check   # exit 1 → finish the interrupted handover first

Use with Codex CLI

A first-class adapter for the OpenAI Codex CLI. Install the bootstrap block into your repo's AGENTS.md (idempotent, preserves your other rules) and write a handover at the end of each session:

from agent_handover.adapters.codex import install_agents_md, build_codex_handover_engine

install_agents_md("AGENTS.md")            # Codex runs `agent-handover check` on start

engine = build_codex_handover_engine(     # one call for the end-of-session hook
    session_note="reviewed PR #42; released v0.2.0",
    current_state="parser refactor merged; next: TOML step config",
)
engine.run()                              # writes memory/, publishes, checkpointed

See examples/codex-cli/ for a runnable end-of-session script and the AGENTS.md block.

Status & roadmap

• checkpointed engine, 3-layer Markdown store, git backend, pause guardrail
• adapter: Codex CLI (AGENTS.md bootstrap + end-of-session handover)
• handover quality scoring (was the note actually useful next session?)
• adapters: OpenCode, Cline, NotebookLM, sqlite-vec
• agent-handover run with declarative step config (TOML)

Issues and PRs welcome — especially reports from other agent stacks (Codex CLI, Cline, OpenCode).

License

MIT