Skip to content

Releases

watchglass 2026.11.11.3

Choose a tag to compare

@smprather smprather released this 11 Nov 19:56

Windows registration keeps its success

Register-ScheduledTask was followed by a best-effort Remove-Item of the Startup shortcut. When the shortcut did not exist (the usual case on GitHub's Windows runner), that last command failed even with -ErrorAction SilentlyContinue, and powershell.exe turned it into exit 1 with empty output. Registration itself had succeeded.

register_script and unregister_script now end with exit 0; a genuine registration error still stops the script first under $ErrorActionPreference = 'Stop'. The real windows-integration CI job is green again, including the Task Scheduler probe, ConPTY typing and the password-only sftp resume test.

Evidence

  • CI run 43090480905: all six jobs success, including windows-integration.
  • uv run pytest -q -> 978 passed, 17 skipped.
  • uv run ruff check src tests and uv run ty check src clean.
  • uv lock --check, uv build, and uvx twine check dist/watchglass-2026.11.11.3* PASSED.
  • watchglass --version reports 2026.11.11.3.

watchglass 2026.11.11.2

Choose a tag to compare

@smprather smprather released this 11 Nov 19:41

A push says how far and why it stopped

status and the TUI now classify the last failed attempt instead of leaving a bare FAILED: dropped mid-transfer, unreachable, refused overwrite, auth rejected, timed out, or failed. It appears in the status detail line and in the TUI NOTE (alongside a running transfer's file name), so a mid-upload drop is no longer indistinguishable from a slow one.

On a narrow terminal NOTE now stays: FAIL/OBS go below 80 columns, and COND/NEXT below 68.

A push under password_mode = "conpty" now streams scp's console progress meter into activity.json, so the existing percentage, rate and ETA line works for uploads as well as pulls: pushing NAME 41% (1.5 GiB/3.8 GiB), 3.4 MB/s, 12m left. Without a console meter scp prints no progress, and the line stays pushing NAME (size) with the header's elapsed time.

Evidence

  • uv run pytest -q -> 978 passed, 17 skipped.
  • uv run ruff check src tests and uv run ty check src clean.
  • uv lock --check, uv build, and uvx twine check dist/watchglass-2026.11.11.2* PASSED.
  • watchglass --version reports 2026.11.11.2.

watchglass 2026.11.11.1

Choose a tag to compare

@smprather smprather released this 11 Nov 19:22

Waiting is not online, and a conpty resume authenticates

watchglass status and the TUI now share one state rule. A watch whose last poll answered unknown reads waiting, not online, and its reason is shown: status prints waiting: <reason> below the row, and the TUI fills NOTE with waiting: <reason>. The other labels keep their precedence: offline, parked, polling/transferring, idle, then online.

With password_mode = "conpty" and resume, sftp now runs interactively instead of with -b: batch mode suppresses the password prompt conpty exists to answer, so the resumed upload never authenticated and ended with a bare Connection closed. The reget/reput/put command is typed at the sftp> prompt and followed by bye. Interactive sftp exits 0 after a failed command, so its console is checked for sftp's error lines; a run that never shows the password prompt now says so explicitly.

Evidence

  • uv run pytest -q -> 967 passed, 17 skipped.
  • uv run ruff check src tests and uv run ty check src clean.
  • uv lock --check, uv build, and uvx twine check dist/watchglass-2026.11.11.1* PASSED.
  • watchglass --version reports 2026.11.11.1.

watchglass 2026.11.11

Choose a tag to compare

@smprather smprather released this 11 Nov 18:13

An honest no-op, a reachable SSH server

A pull that finds an identical file already in local_dir no longer logs pulled NAME (3.8 GiB) in 9.4s — a throughput the link cannot reach, because no bytes moved. It logs NAME already present, verified (size and sha256 match); nothing transferred, and the pre-transfer pulling line is skipped entirely. pull() exposes the no-op through an on_skip callback.

push gains reachable_probe = "tcp" | "banner" (default tcp). With "banner", the pre-flight probe reads up to 255 bytes within reachable_timeout and requires an SSH-2.0- / SSH-1. prefix. A VPN or tunnel that accepts TCP but never sends an SSH banner now raises Unknown — files wait, no failure accounting, no ERROR record with a traceback — instead of starting scp and dying during banner exchange. ctx.transfer.ssh_probe exposes the same probe to check hooks.

Evidence

  • uv run pytest -q -> 959 passed, 16 skipped.
  • uv run ruff check src tests and uv run ty check src clean.
  • uv lock --check, uv build, and uvx twine check dist/* PASSED.
  • watchglass --version reports 2026.11.11.

watchglass 2026.11.9 — a double-click pokes a watch

Choose a tag to compare

@smprather smprather released this 09 Nov 17:10

Poke a watch with a double-click

watchglass tui already pokes the selected watch with p; a double-click on a watch row now asks for the same thing. Only a cell counts: the blank fill past the last column carries the row's index but is marked out of bounds, so clicking there does nothing, and a single click still only selects.

Both routes share one code path, so the refusals are unchanged: no poke is written without a running service, and a write error is reported instead of silently lost.

Docs, the tui help and the README now name the mouse route.

Evidence

  • uv run pytest -q -> 954 passed, 16 skipped; ruff check src tests and ty check src clean.
  • uv lock --check, uv build, and twine check PASSED.
  • A real terminal in tmux: a raw SGR double-click wrote runtime/poke/a and showed poke requested: a; a single click and a header double-click wrote nothing.
  • watchglass --version reports 2026.11.9.

watchglass 2026.11.3.2 — one autostart mechanism on Windows

Choose a tag to compare

@smprather smprather released this 03 Nov 21:44

One autostart mechanism on Windows

watchglass install no longer creates a Startup-folder shortcut beside a Task Scheduler task that is already registered. A failed-looking registration now asks Task Scheduler whether the task exists: if it does, the task is kept, any stale shortcut is removed, and the service is started; only a task that is really absent falls back to the shortcut. If the task state cannot be read, install fails rather than risking two services racing at logon (#32).

Diagnostics now name what PowerShell actually said — stderr first, then stdout, then the exit code — so (no details) is gone. watchglass uninstall already removes both mechanisms.

Evidence

  • uv run pytest -q -> 949 passed, 13 skipped; ruff check src tests and ty check src clean.
  • uv lock --check, uv build, and twine check PASSED.
  • watchglass --version reports 2026.11.3.2.

watchglass 2026.11.3.1 — a transfer says where it writes

Choose a tag to compare

@smprather smprather released this 03 Nov 21:08

Where the bytes are going

watchglass status and the TUI now name the file a transfer is writing. A pull shows its .NAME.part, so progress can be checked with ls or Get-Item without guessing the staging path; a push shows its local source and remote destination, which it never reported before (#31).

The TUI header appends the path after the rate and ETA and shortens the home directory to ~. status prints the full path on its own line under the in-flight line. activity_text keeps its existing short form, so consumers of that line see no change.

Evidence

  • uv run pytest -q -> 944 passed, 13 skipped; ruff check src tests and ty check src clean.
  • uv lock --check, uv build, and twine check PASSED.
  • watchglass --version reports 2026.11.3.1 from the editable install.

watchglass 2026.11.3 — a pty's failures, from the console

Choose a tag to compare

@smprather smprather released this 03 Nov 16:46

conpty failures, from the console

password_mode = "conpty" runs scp/sftp under a pty that merges the command's stderr into stdout, so every failure classifier has to read the merged console.

A rejected credential was classified from stderr only, so it looked transient and the watch retried the login until max_failures instead of parking on max_auth_failures. It now parks on the first rejection, as intended (#29).

The same console text now feeds the remaining classifiers: a -p transfer whose bytes arrived but whose times could not be set is the usual warning instead of a failure, and a vanished source on a fresh pull or an sftp reget raises TransferMissing instead of being retried as an outage (#30). scp's own lines are also recognised under the Windows executable prefix the pty shows (C:\Program Files\OpenSSH\scp.EXE:); a remote shell's /usr/bin/scp: lookalike is not.

Anything a run printed to stderr still behaves exactly as before.

Evidence

  • uv run pytest -q -> 940 passed, 13 skipped; ruff check src tests and ty check src clean.
  • uv lock --check, uv build, and twine check PASSED.
  • watchglass --version reports 2026.11.3 from the editable install.

watchglass 2026.11.2 — poke a watch

Choose a tag to compare

@smprather smprather released this 02 Nov 17:05

Poke a watch

watchglass poke NAME asks the running service to poll one watch now instead of waiting for its interval. It writes one request file per watch under the runtime directory; the service consumes it within a second, so it works on every OS without a socket or signal.

A poke overrides the interval, the failure backoff, and an ordinary offline marker (the forced poll is a trial). It refuses enabled = false, a user watchglass disable, and a permanent credential rejection: retrying those cannot succeed. The service logs watch.poked or watch.poke_ignored with the reason.

In watchglass tui, p pokes the selected watch after checking that the service is running; a notification confirms the request. The TUI reference no longer calls it strictly read-only.

Evidence

  • uv run pytest -q -> 926 passed, 13 skipped; ruff check src tests and ty check src clean.
  • uv lock --check, uv build, and twine check PASSED.
  • CLI and PTY TUI smoke runs forced a real off-interval poll (1.01s from request to poll).

watchglass 2026.10.31.1 — keep a grown resume temp

Choose a tag to compare

@smprather smprather released this 31 Oct 20:05

Keep a grown resume temp

A resumable mirror upload can read past its pre-upload snapshot when the writer appends while put/reput runs. The resulting temp is longer than that snapshot, so the whole-file hash check no longer matches it — and 2026.10.31 deleted the temp, forcing the next poll to upload the file from zero again, exactly the churn resume exists to avoid.

The temp is now kept. The next poll hashes it against the then-current local prefix and:

  • continues it with reput when it is a shorter prefix,
  • finalizes it with os.replace when it is already the complete current file,
  • discards it and sends whole only when it is not a prefix (the writer rewrote or truncated it).

The prefix check, not the pre-upload snapshot, is the authority for reput, so keeping the temp is safe. A test covers an upload that appends bytes and a temp that matches the grown file; the next poll finalizes it without re-uploading.

resume still needs protocol = "sftp", direction = "push", Python 3.6+ and key authentication on the destination; settle remains the cheaper option for a live folder.

Evidence

  • uv run pytest -q → 909 passed, 13 skipped; ruff check src tests and ty check src clean.
  • Wheel and sdist built; twine check PASSED.