Skip to content

v1.5.3

Choose a tag to compare

@devsolux devsolux released this 17 Oct 08:20
· 43 commits to main since this release

1.5.3 — 2026-10-17

Operator-command hardening: setup --remove, concurrent runs, --host and
--dry-run now do what the operator asked, refuse what they cannot do and
say so. A second pass with the ECC Python, silent-failure, security and
test-coverage reviewers closed the gaps the first pass left (listed after the
first-pass items). No release pin moved since 1.5.2, so no card was due for a
recheck.

  • setup --remove no longer deletes a binary that another host still
    launches. Every host gets a row; one whose CLI is not run (not on PATH, or
    left out with --host) has its user-scope file read, and while it still
    registers the app at the managed binary, or cannot be read, the binary, pin
    and evidence stay. Before, --remove --host claude deleted the binary under
    a live Codex registration, and a host CLI missing from PATH left its
    ~/.claude.json or config.toml entry pointing at a deleted file, both
    reported ok.
  • A failing mcp remove is reported in that host's row and the run continues
    with the other hosts and apps; before, it aborted the whole removal without
    a report after earlier hosts were already unregistered. A failed binary
    delete or pin rewrite is reported the same way.
  • The pin now stays while any registration that may still launch the managed
    path is kept (for example one whose mcp get timed out), so a later
    --remove can still prove that entry is Storytold's and remove it. Before,
    the pin was dropped and that registration could never be removed by
    Storytold.
  • One changing run at a time: setup and setup --remove hold an OS advisory
    lock (setup.lock in the state directory) for the whole run, released by
    the OS when the process ends. The <lock>.writing marker guarded only the
    lock rewrite, so two runs could clear each other's staging or recovery
    marker.
  • An app without root flags and no --acknowledge-unconfined is skipped
    whatever its install path holds; an outdated pin or a foreign file there no
    longer stops the rest of the batch. The message names what the next run
    would meet.
  • A host named with --host whose CLI is not on PATH now fails the run
    (other named hosts are still registered); before, it was a warning with
    exit status 0.
  • --dry-run shares the real run's decision function and reports, per app,
    binary (keep, install, upgrade, skip, refuse with reason), a
    download only where one is needed, per-host registrations judged from
    the user-scope files, predicted_status and an overall would_pass.
    Before, it listed a download and a registration for every app, including
    ones the real run would skip or refuse. A work root that exists but is not a
    Storytold workspace is now refused before any directory is created.
  • Tests: nine new regression tests; the test runner strips
    CLAUDE_CONFIG_DIR and CODEX_HOME, and the CLI tests point
    CLAUDE_CONFIG_DIR at a temporary directory, so no test reads the
    operator's real host settings.

Second pass (ECC review of the first):

  • setup --remove keeps the binary and pin under a kept registration whose
    command the host CLI does not show (claude mcp get text without a
    Command: line, unparseable codex mcp get --json), and under one a local
    or project entry shadows in mcp get while the user-scope file still names
    the binary. Before deleting, it now searches both hosts' files for any
    registration that starts the binary: another server name, another spelling
    or symlink of the path, or a Claude Code local-scope entry, listed in
    references by host, scope and name only. An mcp remove that exits 0 is
    trusted only once the user-scope file no longer lists the entry.
  • A registration carrying settings mcp add never writes (tool lists,
    timeouts, a non-stdio type) is no longer treated as Storytold's: removal
    keeps it, and --replace refuses it instead of losing those settings on a
    failed add.
  • A failing evidence archive, or any unexpected error in one app, no longer
    ends setup --remove without a report: it becomes that app's row and the
    other apps still run. A real setup that stops at a failure lists the apps
    it never attempted as not_reached.
  • Refused before anything is created or downloaded: a leftover
    <lock>.writing marker (also by --remove, before any binary is deleted),
    a state or evidence directory that is not owner-only, and a bin directory
    that group or others can write, which used to be found only after a
    candidate was promoted and pinned.
  • --dry-run predicts those refusals, the Linux/Windows provenance gate (no
    bundle or no cosign without --trust-upstream-sums), a --replace that
    could not restore the old entry, and an unreadable user-scope file
    (registration_unknown); it exits 1 when would_pass is false, like
    --remove --dry-run. A real setup report without a recognized status no
    longer reads as success.
  • Host files never reach a report: unparseable or too deeply nested JSON or
    TOML (a RecursionError used to escape as a traceback) gives a fixed
    message instead of the parser's text, which could quote keys, and only text
    commands of at most 300 characters are shown. doctor does the same.
  • The run lock refuses a lock file that is a symlink, reparse point or not a
    regular file (Windows has no O_NOFOLLOW), names a filesystem without file
    locks instead of reporting "another setup run", always closes its file on
    Windows, and two first runs racing to create the state directory meet the
    lock instead of a raw FileExistsError.
  • Docs: operations says that claude mcp get, used by --remove, its dry
    run, --list and registration, starts the registered server for its health
    check, as a Claude Code session does.
  • Tests: thirteen more, 260 in all, including a run lock held by another
    process blocking setup and --remove, the Windows msvcrt branch through
    a fake module, every unrun_host branch, failed deletes and pin rewrites,
    recovery markers in a dry run and a Linux provenance preview; the
    killed-holder test has a watchdog and waits briefly for Windows to release
    the lock.