abstract-gpt

Codex login and session management, usable as a CLI and as the optional /gpt/* category in abstract_toolserver. Requires Python 3.11+, Linux (for HTTP login locking), and an installed Codex CLI.

pip install -e /srv/pyit/dev/abstract_gpt
abstract-gpt oauth-status
abstract-gpt login                    # device URL + code; approve in your browser
abstract-gpt launch --                # new interactive conversation
abstract-gpt exec -- "Explain this repository"  # ephemeral noninteractive session

login --browser uses normal browser login. login --with-api-key reads a key from stdin via Codex itself. API key usage has separate API billing. No key is required for the ChatGPT subscription login flow.

Toolserver

Install in the toolserver's Python environment, then add alongside its Claude merge:

_merge_optional_toolset("abstract_gpt.toolset:get_toolset")

Restart the service to register these routes. They use the toolserver's existing authentication and execute as its OS user (vm_mgr on this host):

Use POST for mutations. Device login runs in a detached subprocess with a 15-minute timeout; a file lock prevents duplicate flows across gunicorn workers. Poll until authenticated, failed, expired, or interrupted. After service restart, an interrupted attempt can be started again. Failed login output remains in the private AG_ROOT/login-output.txt for local diagnostics, and is cleared at the next attempt. Raw login output and stored credentials are never returned by the HTTP tools. The device code is sensitive: use the existing authenticated toolserver connection.

Storage and sessions

• CODEX_HOME: existing Codex home, default ~/.codex. Login and launch use the same store so Codex manages token refresh. No credentials are copied between users.
• AG_ROOT: wrapper storage, default ~/.local/share/abstract_gpt, private mode 0700.
• AG_CODEX_BIN: optional executable path. Otherwise ~/.local/bin/codex, then PATH. The user-local binary takes priority to avoid selecting a Snap launcher inside a restricted systemd service.
• abstract-gpt init, state, save-template, restore, set-model MODEL, login-start, login-poll, and oauth-solution mirror their actions.
• launch starts a new conversation by default; explicit Codex resume flags still work. exec adds --ephemeral. Neither deletes existing sessions or changes sandbox/approval settings. Explicit model flags override the wrapper default.
• A config template can contain user-configured secrets; it stays private and is never returned through the API. It does not include auth.json or transcripts.

This package does not emulate Claude's durable token export, destructive reset, or automatic quota fallback. Codex owns its cached authentication and renewal. A service login applies to that service account; other OS users log in separately.

Toolserver tools inside Codex

Every abstract-gpt launch, abstract-gpt exec, MCT turn, and shared-console Codex app-server starts with the toolserver MCP bridge configured through Codex command-line options. No codex mcp add or ~/.codex/config.toml entry is needed. The bridge honors TOOLSERVER_URL and existing Hugpy operator credentials. The mcp install extra remains as a compatibility alias.

Automatic comms channel

abstract-gpt launch and abstract-gpt exec install a package-owned Codex SessionStart hook. At startup or resume it creates (or reuses) a toolserver channel and binds inbound messages to the Codex thread's native queue. The shareable URL is shown at session start and added to the model's context.

abstract-gpt comms          # show the newest bound endpoint
abstract-gpt set-comms off  # opt out (use `on` to enable again)

The wrapper passes Codex's hook-trust bypass only for launches it controls, so its installed hook can run without an extra confirmation step. This does not change Codex sandbox or tool approval policy.

Validation

PYTHONPATH=src python -m unittest discover -s tests -v

Tests use a fake Codex executable; they never consume model credits or change real credentials. Authentication behavior follows the official Codex documentation.

MCT chat

abstract-gpt mct
abstract-gpt mct /path/to/workspace --model YOUR_MODEL
abstract-gpt mct /path/to/workspace --resume
abstract-gpt mct /path/to/workspace --resume --dangerous

MCT uses the same pointer-exchange interaction as abstract-claude: each prompt and response lives under exchange/, with a multiline terminal input box, streaming activity, and an idle timeout (--timeout, default 1800 seconds, or MCT_TURN_IDLE_TIMEOUT). --plain selects line input.

While a turn runs, ordinary messages coalesce into the next turn, +message queues a separate turn, !message interrupts, and todo: message or ?message captures a task without calling the model. /new resets the conversation; /help lists all commands.

Default workspace: ~/.config/abstract_gpt/mct/repl (respects XDG_CONFIG_HOME). Every launch starts a fresh thread; --resume continues that workspace's saved thread. Codex credentials and configuration stay intact. --model overrides the wrapper's set-model default. Turns run with workspace-write access and no interactive approval prompts. --dangerous (also accepted as --dangerously-bypass-approvals-and-sandbox) gives Codex unsandboxed access and should only be used in an externally isolated environment. Workspace operator-guidance.md supplies additional instructions.

Prompt/response files, JSONL activity, and pending archive records remain on disk. Set MCT_LOCUS, TOOLSERVER_URL (or EXCHANGE_INGEST_URL) and TOOLSERVER_TOKEN to enable the same retryable toolserver archive as Claude MCT. This archive includes workspace exchanges, not unrelated conversations from the Codex login store.

The adapted MCT UI and archive retain their original notice in src/abstract_gpt/MCT-LICENSE.txt.