🧲 Magg Release v1.3.0

Changes

Robustness fixes: config data loss, unmounting, hung backends, CLI and mbro bugs (#21)

Date: Sat Sep 26 09:28:09 2026 -0700

• Add backend_init_timeout setting

• Stop config saves from silently deleting user data

• A config.json that fails to parse is no longer overwritten by the next
  save; the save is refused and the file left as-is. An empty file is
  still treated as new.
• Server entries that fail validation, and top-level settings like
  self_prefix, are written back unchanged instead of being dropped.
• The reloader now reads the file with the same loader as
  ConfigManager, so its cached config includes kits and settings. With
  auto-reload on (the default), the first save from a running server
  used to erase the kits section.
• ServerConfig, KitConfig, AuthConfig and BearerAuthConfig are plain
  models. As BaseSettings with no env prefix they read unrelated env
  vars like PREFIX (set by Termux), URI, ENV and KEY_PATH, which made
  valid servers fail to load and then get dropped on save.
• Config and auth.json writes go through a temp file and rename, keeping
  the existing file's mode and following symlinks.
• The watchdog handler also reacts to files created or renamed over the
  config, which is how many editors and tools save.
• A disabled server without a command or uri no longer blocks every
  subsequent reload.

• Fix auth CLI crashes and harden server search

Auth:

• magg auth init creates missing parent dirs (e.g. no ~/.ssh) and
  reports "key already exists" as an error instead of a traceback.
• The private key file is created 0600 rather than chmod'ed after write.
• --key-path is saved to auth.json; previously only a custom issuer or
  audience was, so auth status/token couldn't find the new key.
• auth private-key --export quotes the key so the output can be eval'd.
• auth status no longer prints the MAGG_PRIVATE_KEY line twice.

Search:

• A GitHub repo with a null description, or an npm package with a null
  author, no longer throws away all results from that source.
• Sources are searched concurrently, with a 15s session timeout instead
  of aiohttp's 300s default, and proxy env vars are honored.
• The ranking bonus for Glama results now matches their source name.

• Make documented kit and reload settings real; fix docs that drifted

• Kits in a kit.d next to the config file are found even when the config
  is set with MAGG_CONFIG_PATH or --config, as the docs described.
• MAGG_RELOAD_USE_WATCHDOG was documented but ignored; false now forces
  polling, which helps where file notifications don't arrive (network
  filesystems, Docker Desktop bind mounts).
• Docs: --env-pass/--env-set are global options (magg --env-pass serve);
  kit discovery uses MAGG_PATH, not the nonexistent MAGG_KITD_PATH; an
  omitted prefix means no prefix, not the server name; MAGG_PREFIX_SEP
  only applies to Magg's own tools; magg_check's timeout default is 2.5;
  the kit example is valid JSON; document MAGG_BACKEND_INIT_TIMEOUT and
  how reload treats invalid files.

• Report invalid MAGG_* settings without a traceback

A bad value such as MAGG_AUTO_RELOAD=maybe or MAGG_SELF_PREFIX=my_prefix
used to print a pydantic traceback on every command, including
--version. The CLI now prints one line per invalid field naming the
environment variable to fix, and exits 1. python -m magg goes through the
same path.

• Fix mbro argument parsing, error handling, and connection bugs

• call/prompt keep a JSON object argument as raw text (quoted or not),
  and scripts join JSON objects that span lines until braces balance
• key=value values are decoded as JSON with a raw-string fallback,
  keeping string-typed schema params as strings (no more 007 -> 7 or
  ValueError on 10-20); prompt accepts key=value
• Non-interactive runs (-n, -X, -) exit 1 when any command fails;
  per-command errors are contained and later commands keep running,
  with --fail-fast to stop at the first failure; quit stops scripts
• status takes args and checks for a connection; None descriptions no
  longer crash tools/prompts/search filtering or JSON search output
• script run/dump with a missing script report an error instead of
  exiting; resource/prompt tracebacks only print with -v
• Connection errors carry the real reason (including duplicate names);
  connects time out after --timeout seconds (default 30)
• Only bare URLs get /mcp/ appended; /mcp and /sse paths are kept
• mbro - runs non-interactively; --repl without arepl returns cleanly
• help lists status and script instead of the nonexistent refresh;
  unknown tools get "Did you mean" suggestions
• Remove unused JsonArgParser

• Align mbro docs with actual commands and options

Replace nonexistent --connect/--call-tool flags and mbro scripts,
document disconnect <name>, script subcommands, script discovery
paths, JSON/key=value value handling, exit status, --fail-fast,
--timeout, URL defaults, and fix the REPL example.

• Fix server robustness bugs: unmount, hung backends, signals, read-only, auth

• Unmount on fastmcp 3: track the provider each mount adds and remove it
  from mcp.providers, so disable/remove/update actually drop or replace a
  backend's tools (previously a no-op that kept stale providers).
• Hung backends: pass backend_init_timeout to backend clients and resolve
  Magg's own tools/resources/prompts locally first, so one unresponsive
  backend no longer blocks list_tools or Magg's tools.
• Suppress backend stderr via the stdio transport's log_file; drop the
  stdio_client monkeypatch that fastmcp bypassed.
• Per-server env now overrides --env-pass/--env-set.
• magg_add_server: accept uri (AnyUrl -> str), require command or uri,
  report the mount failure cause, roll back the mount if saving fails.
• magg_check: use MountedServer.client (it crashed with any mount) and
  iterate over a snapshot of mounted servers.
• Read-only mode is checked up front in add/remove/enable/disable, kit
  load/unload, and non-report check actions.
• Refuse to serve HTTP/hybrid when a configured private key can't be
  loaded instead of silently running without auth.
• Runner: use loop signal handlers so SIGTERM wakes an idle stdio loop,
  bound the wait on the cancelled server task, exit via the signal once
  cleanup is done, force exit on a second signal, and restore SIG_DFL.
• Proxy tool: path is required for 'info' (fix description).

• Allow check remount/unmount in read-only mode; fix Python 3.12.3 CLI parsing

• magg_check's remount and unmount actions only affect the running
  process, so they stay available in read-only mode; disable still
  refuses since it writes the config.
• On Python 3.12.3 (Ubuntu 24.04's system Python), argparse treated
  magg server update --env as an ambiguous abbreviation of the global
  --env-pass/--env-set options. Disable abbreviations on the top-level
  parser.
• magg_add_server's prefix description matches actual behavior: no
  prefix by default.

• Make new tests check behavior instead of internals

• Backend stderr: assert what actually reaches stderr (via capfd) with
  suppression on and off, instead of checking a transport attribute.
• Reload: rename a real file over the watched config and check the
  reload fires; check Magg's own atomic save doesn't trigger a reload
  while the next outside edit still does; check polling mode without
  watchdog by observing a detected change.
• Placeholder and unparseable-file reload tests go through
  reload_config() rather than private helpers.
• Hung backend test fails with a timeout instead of hanging on
  regression, and no longer inspects private client attributes.
• --key-path test checks that auth token works afterwards, which was
  the user-facing failure.
• Search concurrency is proven with a barrier instead of timing.
• Settings: a bad value no longer skips later valid settings.
• Remove tests that couldn't fail: a tool description string check, and
  a timeout-0 check that fastmcp already guarantees.

• Test that mbro --timeout 0 disables the connect timeout

fastmcp treats an init_timeout of 0 as disabled, so passing it through
already matches the documented behavior. The test fails if 0 ever turns
into an immediate timeout.

• Bump version to 1.3.0

• Add a version bump script and drop the server.json tests

scripts/bump_version.py bumps the version with uv version (which
updates pyproject.toml and re-locks uv.lock) and copies it into
server.json. With --check it only verifies that server.json matches
pyproject.toml, for CI to run instead of the tests.

The server.json tests in test_registry_search.py checked release
metadata rather than code behavior, so they're removed.

• Enhance CI with version consistency check job

Added version consistency check for server.json and pyproject.toml.

Merge pull request #18 from sitbon/claude/magg-features-cli-r92ff7

Date: Sun Aug 2 13:56:21 2026 -0700

Gate PR Docker image publishing behind push-image label; add GHCR cleanup

Gate PR image publishing behind push-image label; add GHCR cleanup

Date: Sun Aug 2 20:26:12 2026 +0000

• PR runs now build and test images without publishing. Adding the
  'push-image' label to a same-repo PR publishes the pr-NN / pr-NN-dev
  tags (the 'labeled' event re-triggers the workflow). Fork PRs never
  push - their GITHUB_TOKEN cannot write packages, which previously
  would have failed the job outright.
• Set retention-days: 1 on the dev-image tar artifacts, which are only
  consumed by the next job in the same run.
• Add a weekly ghcr-cleanup workflow that deletes ephemeral pr-* tags
  and untagged manifests older than a week, complementing the post-push
  cleanup-untagged job (which never runs on PR events). Version and
  branch tags are never touched, so pinned deployments are unaffected.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_011JFaNeMxJCxXhQ3J9RQACR

Document label-gated PR image publishing in readme

Date: Sun Aug 2 20:23:31 2026 +0000

Co-Authored-By: Claude Fable 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_011JFaNeMxJCxXhQ3J9RQACR

Installation

uv add magg==1.3.0