Windows: Full-mode manifest still gates open/open-link/last to macos+linux after #31 revert #43

Activity

1. backnotprop commented on Sep 8, 2026

   

   backnotprop

   on Sep 8, 2026

   Contributor

   More actions

   Thanks, and yes, we want this. Two corrections and the shape we'll merge.

   1. Flipping the three actions alone doesn't work: open and last launch the doc pane, which would still be gated off, so every review fails to open.
   2. A second doc pane can't exist. Herdr rejects duplicate pane ids even with disjoint platforms, and a PowerShell wrapper in the middle is out: it puts a process between Herdr and the TUI, and on 0.8.2 the plugin root arrives as an extended-length \\?\ path that PowerShell mishandles.

   Accepted design: a separate Windows install variant, windows-full/herdr-plugin.toml, requiring Herdr 0.9.0, where panes resolve relative programs against the plugin root. Its doc pane is direct argv: ["./bin/plannotator-tui.exe", "herdr", "pane"], no wrapper. Actions use the same direct argv. The root manifest stays as is; Windows on 0.8.2 keeps Lite.

   The acceptance checklist (manifest parity, the fetch contract, and what the Windows CI runner must prove) is going into the repo shortly and I'll link it here. If you want to take the PR against it, say so. You have Windows hardware and we don't, so your live results would also count as the native qualification.
2. backnotprop

   added

   enhancementNew feature or requestNew feature or request

   on Sep 8, 2026
3. backnotprop

   added a commit that references this issue on Sep 9, 2026

   docs: add the windows full acceptance checklist (#44)

   ...

   8ae5510
4. backnotprop commented on Sep 9, 2026

   

   backnotprop

   on Sep 9, 2026

   Contributor

   More actions

   The acceptance checklist is in the repo now: https://github.com/plannotator/herdr-annotate/blob/main/docs/windows-full-acceptance.md. That's the target for the windows-full/ variant if you take it.
5. timo-tigges commented on Sep 17, 2026

   

   timo-tigges

   on Sep 17, 2026

   Contributor

   More actions

   Native Windows results, since they're the blocker on hardware you don't have. Measured against herdr 0.9.0 stable (protocol 22), Windows 11 Pro 10.0.22621, PROCESSOR_ARCHITECTURE=AMD64, plugin at 7c8f5a1. Method: throwaway plugin link plugins, one pane per case, opened with herdr plugin pane open --entrypoint <id> --cwd <dir>; outcome read from pane.exit in herdr-server.log.

   TL;DR: the accepted design works. ./bin/plannotator-tui.exe as a direct-argv doc pane resolves correctly from a review-folder cwd on 0.9.0.

   Reading the exit codes

   Outcome	Meaning
   plugin_pane_open_failed	CreateProcessW failed synchronously, no pid — and the error names the exact path herdr passed, which is what makes resolution observable
   exit 0 / 1	image started and ran (1 = herdr-annotate editor refusing without a pending annotation)
   exit 4294901760 (0xFFFF0000)	image started and then exited abnormally — the log shows pane.spawned outcome="ok" pid=… first. Not a spawn failure.

   1. Panes resolve relative programs against the plugin root

   ["./bin/herdr-annotate.exe", "editor"], plugin root C:\Users\<user>\hp, opened with --cwd C:/projects/terminal-setup → exit 1. Found and ran, from a cwd containing no bin/.

   A deliberately missing relative program makes the resolution explicit:

   CreateProcessW `"\\?\C:\Users\<user>\hp\bin\nope.exe editor"`
     in cwd `Some("C:/projects/terminal-setup")` failed: os error 2
   

   Resolved against the plugin root; cwd passed separately. Not source-dependent — this was a linked plugin, not a GitHub install.

   2. Resolved paths arrive extended-length, and that is what kills a PowerShell wrapper

   Only relative spellings get rewritten. An absolute missing program reports CreateProcessW "C:\Users\<user>\hp\bin\nope2.exe" — no prefix.

   Isolating it on one binary (a copy of powershell.exe staged at <root>/bin/probe.exe), every run with --cwd outside the root:

   Program as written in the manifest	Exit
   C:\Users\<user>\hp\bin\probe.exe	0
   \\?\C:\Users\<user>\hp\bin\probe.exe	4294901760
   ./bin/probe.exe	4294901760
   .\bin\probe.exe	4294901760
   bin/probe.exe	4294901760
   bin\probe.exe	4294901760
   ./bin/probe.exe, cwd = plugin root	4294901760

   Rows 1 and 2 differ only by a prefix typed in by hand. So the failure is the extended-length image path — not relative resolution as such, and not the pane cwd, since row 7 fails where a cwd-relative lookup would have succeeded. Not path length: identical at a 36-character binary path and a 140-character one. Native binaries are unaffected.

   So the checklist's "no launcher process between Herdr and the TUI" is stronger than a design preference on Windows — a PowerShell-wrapped pane cannot stay up. Good news for windows-full, since plannotator-tui is native.

   3. A correction to the \\?\ note in this thread

   On 0.9.0 the extended-length form reaches the program argument, but HERDR_PLUGIN_ROOT read inside a pane arrives in normal form (C:\Users\<user>\probe). So code rebuilding a path from the env var and code reading the image path see different things. Worth pinning, given the checklist's "do not blindly remove \\?\" row.

   What this does not show

   • Both come from the CLI's plugin.pane.open. A pane opened by the plugin itself over the socket is not exercised.
   • plannotator-tui was never run — this is "the pane can start", not rendering, q, or teardown.
   • ARM64 untested; this machine is AMD64.

   Taking the PR

   I'd like to, against docs/windows-full-acceptance.md. Two questions before I start:

   1. Onboarding sections 7/8/9 are referenced by §1, §3 and §6 but I can't find that document in the repo. Is it published somewhere, or could it be added?
   2. §2's subdirectory staging. fetch-plannotator-tui.ps1 does Set-Location (Join-Path $PSScriptRoot "..") and writes ./bin, so called from windows-full/ it stages to root bin/ — which §2 explicitly refuses. Do you want a destination parameter on the shared fetcher, or a windows-full/scripts/ wrapper that supplies it? Happy either way; it changes which regression tests I add.

   I can cover §1, §2 and the AMD64 half of §4 on this hardware. §3's GitHub Windows runner I can attempt, but flagging early that §4's native ARM64 row is not closeable here.
6. backnotprop

   added 8 commits that reference this issue on Sep 20, 2026

   docs: document the windows-full install

   ...

   acc00b2

   test: restore HERDR_ANNOTATE_BIN after the pane run

   ...

   a97de83

   test: initialize the unc cleanup variables before the try

   ...

   0c55d85

   test: drop the meaningless exit status check around the pane case

   ...

   3ea3768

   test: assert the windows-full build scripts exist

   ...

   3047eba

   fix: treat an empty PLANNOTATOR_TUI_BIN as absent

   ...

   240fb22

   fix: exit zero explicitly after a successful fetch

   ...

   0cfdfb2

   feat: add the windows-full manifest variant (#49)

   ...

   7480133
7. backnotprop commented on Oct 15, 2026

   

   backnotprop

   on Oct 15, 2026

   Contributor

   More actions

   Fixed by #49 and #58: Windows now has its own full-mode manifest with open, open-link and last. Thanks for the report.
8. backnotprop

   closed this as completedon Oct 15, 2026