agent_session

agent_session

A small, single-purpose Python library: uniform multi-turn session abstraction over CLI coding agents (Claude Code, Codex). One protocol, two interchangeable backends, zero opinions about what you do with the result.

It exists because higher-level frameworks (autor, aker, ...) all need the same primitive — "drive a claude -p or codex exec subprocess across multiple turns, capture file changes, return a canonical result" — and re-implementing it per project drifts.

from agent_session import (
    make_session,
    SANDBOX_WORKSPACE_WRITE, SANDBOX_READ_ONLY,
    BACKEND_CLAUDE, BACKEND_CODEX,
)

sess = make_session(
    cwd="/tmp/scratch",
    sandbox=SANDBOX_WORKSPACE_WRITE,
    backend=BACKEND_CLAUDE,         # explicit; or use env_var=...
    timeout_sec=600,
)

r = sess.send("Write hello.txt with 'hi' inside.")
assert r.ok
print(r.files_created)              # [PosixPath('hello.txt')]
print(r.final_message)
print(r.session_id)                 # uuid string

r2 = sess.send("Now append ' there'.")  # turn 2: same session_id, resumed

---

Public API

Everything below is importable from the top-level agent_session namespace. Anything inside agent_session._* modules is internal — do not import from there.

Session factory

Types

Constants

---

Sandbox semantics

Sandbox is enforced per-turn, not session-wide; resuming a session with a different sandbox is allowed (within each backend's own constraints).

workspace-write and danger-full-access look identical for claude because Claude Code does not currently expose a finer-grained native sandbox; consumers that need stricter confinement should rely on the subprocess cwd + filesystem permissions.

---

Backend selection

Three places control which backend make_session picks, in priority order:

1. backend="claude" / backend="codex" — explicit kwarg, no override.
2. env_var="MY_PROJ_BACKEND" — read that env var (case-insensitive, value must be in ALL_BACKENDS); raises AgentError on unknown values so typos are loud.
3. Random — uniform pick across ALL_BACKENDS via the supplied or default random.Random.

Consumers typically expose their own env var name:

sess = make_session(cwd=..., env_var="AUTOR_BACKEND")  # autor convention
sess = make_session(cwd=..., env_var="AKER_BACKEND")   # aker convention

---

Result reporting

Every backend computes files_created / files_modified by snapshotting mtimes + sizes of cwd.rglob('*') before and after the subprocess. Snapshot ignores nothing — if your subprocess writes Python bytecode into __pycache__/, that will show up. Filtering is the consumer's responsibility.

final_message is the agent's last assistant turn:

• claude: result field of claude -p --output-format json output
• codex: contents of the file pointed to by --output-last-message

If the subprocess times out, ok=False, exit_code=-1, and stderr contains the timeout note.

---

Prerequisites

• Python ≥ 3.9
• claude binary (Claude Code CLI 2.0+) on PATH if you want the claude backend
• codex binary (codex 0.120+) on PATH if you want the codex backend

You only need the binaries for backends you actually instantiate. The package imports lazily — make_session(..., backend="claude") does not import agent_session.codex and vice versa.

---

Install

git clone <this repo>
cd agent_session
pip install -e . --user

---

What this is NOT

• Not a graph framework — that lives in consumers (autor, aker).
• Not a prompt library — send takes a string, no templating.
• Not multi-process — sessions are not thread-safe; spawn one per worker if you parallelise upstream.
• Not deterministic — both claude and codex have non-zero temperature; this library does not change that.

---

Adding a new backend

1. Implement MyAgentSession somewhere in agent_session/<name>.py conforming to AgentSession protocol.
2. Add BACKEND_MYAGENT = "myagent" to agent_session/_common.py and append it to ALL_BACKENDS.
3. In agent_session/_backends.py:make_session, add a branch that imports your module lazily and constructs the session.
4. Optionally add a smoke test to tests/.