Repository navigation
vm: run VM replays side by side; keep installed boxes across replay-only changes - #112
Conversation
…nly changes Every guest forwarded ssh (and an unused http forward) from fixed host ports 10022/10080, so a second replay on the same host died at once and every remi proof queued behind the last one. Each guest now forwards ssh from a free port of its own (asking again if another process takes it first), and the http forward is gone. Runs side by side share the installed boxes, so an install (and the chroot box build a shelter run does) takes a lock on its cache entry: one run installs, the others wait and then use it. The box cache key hashed the whole of vm.py, so any change to the hop walk, shelters, relocate or prune reinstalled every box (10 to 20 minutes each) on the next run. It now hashes the source of build() and everything in vm.py it reaches by name (build_source), so only changes the install runs reinstall. Every run records the commit it ran (run.json "head", with "dirty"), so a run on a stale head shows as one. tools/test_vm_parallel.py proves the ports, the retry, the lock and the key by behaviour with a stand-in QEMU; the client job and `wend check` run it. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018rvHWMQYKBKpBz9S6j5fnj Signed-off-by: Claude <noreply@anthropic.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
github-actions
Bot
added
the
needs-claude-review
Runs side by side share the shim on 8085. Each VM run now holds a shared flock on $WEND_WORK/mirror.users while it lives; `vm.py shim-stop PID` takes it exclusively, refuses (exit 3) while any run holds it, and kills only a tools/mirror.py process. A run that starts during a stop waits for it, then starts a fresh shim from its own checkout. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018rvHWMQYKBKpBz9S6j5fnj Signed-off-by: Claude <noreply@anthropic.com>
|
@codex review The head is now 04522d2. It adds Generated by Claude Code |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 04522d27a9
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…ools/mirror.py (review) A run that finds base.qcow2 without taking the lock now finds its UEFI NVRAM beside it too. shim-stop now refuses a process whose mirror.py isn't in a tools/ directory with vm.py and sandbox.py beside it. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018rvHWMQYKBKpBz9S6j5fnj Signed-off-by: Claude <noreply@anthropic.com>
|
@codex review The head is now 39854c8. It fixes both P2 findings: install order, and the shim-stop path check. Generated by Claude Code |
|
Codex Review: Didn't find any major issues. 🚀 Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
…Rust 1.87) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018rvHWMQYKBKpBz9S6j5fnj Signed-off-by: Claude <noreply@anthropic.com>
|
@codex review The head is now 92d4c13, which merges main in. The conflicts were only in the test lists in client.yml and Generated by Claude Code |
|
Codex Review: Didn't find any major issues. Keep it up! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
…e fetched runs) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018rvHWMQYKBKpBz9S6j5fnj Signed-off-by: Claude <noreply@anthropic.com>
|
@codex review The head is now bd7e004, which merges main in again. The only conflict was the test list in Generated by Claude Code |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: bd7e004cdc
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
If the shim exits after the checks and its pid is reused, a signal by number would reach the newcomer. shim-stop now opens a pidfd first, reads /proc while the pidfd says that process is alive, signals through the pidfd and waits on it for the exit. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018rvHWMQYKBKpBz9S6j5fnj Signed-off-by: Claude <noreply@anthropic.com>
|
@codex review The head is now 3dcce73. shim-stop now signals through a pidfd, which fixes the pid-reuse finding. Generated by Claude Code |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3dcce738b1
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Runs with different work directories share the one shim on a port, so a lock under $WEND_WORK let shim-stop from another work directory kill a shim a run was using. Runs now hold /var/tmp/wend-shim-<port>.users for each adapter shim port, and shim-stop locks the port named on the shim's own command line. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018rvHWMQYKBKpBz9S6j5fnj Signed-off-by: Claude <noreply@anthropic.com>
|
@codex review The head is now 8f95d64. The shim lock is now keyed on the shim's port instead of Generated by Claude Code |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8f95d64d05
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
In sticky world-writable /var/tmp another account could unlink and recreate a lock file a run holds, so shim-stop would lock the newcomer and stop a shim in use. The locks now live in /var/tmp/wend-shim, which must be a real directory owned by root or the running user (or the user sudo ran as) and writable by nobody else, or it's refused. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018rvHWMQYKBKpBz9S6j5fnj Signed-off-by: Claude <noreply@anthropic.com>
|
@codex review The head is now f90d0bd. The shim locks now live in a protected directory, Generated by Claude Code |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f90d0bd648
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
As root in a checkout someone else owns, core.fsmonitor in its config is a command `git status` would run. head() now passes core.fsmonitor=false and --ignore-submodules. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018rvHWMQYKBKpBz9S6j5fnj Signed-off-by: Claude <noreply@anthropic.com>
|
@codex review The head is now 4df4573: Generated by Claude Code |
|
Claude review: SIGN-OFF on head 4df4573 An independent review of the full diff against main found no blocking findings. What it checked:
Tests:
Non-blocking, for follow-ups:
Generated by Claude Code |
|
Codex Review: Didn't find any major issues. 🎉 Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
…ve gaps) Signed-off-by: Claude <noreply@anthropic.com>
2 participants
Requested by Peter · project thread
What this adds
Before: every VM guest forwarded ssh and http from fixed host ports (10022, 10080), so a second
tools/vm.pyrun on remi died at once and every proof queued behind the last. The box cache key hashed all ofvm.py, so any change to the hop walk, shelters, relocate or prune reinstalled every box (10 to 20 minutes each) on the next run. A run'srun.jsondidn't say which commit it ran.After: each guest forwards ssh from a free port of its own (and asks again if another process takes the port first); the unused http forward is gone. Several replays share the installed boxes, and an install (plus the chroot box build a shelter run does) holds a lock on its cache entry, so one run installs and the others wait for it. The cache key hashes only
build()and what invm.pyit reaches by name (build_source). Every run recordshead(commit,dirty).How:
Guest.argv/Guest.startpick the port per start;Lockedis an flock beside the cache dir (sandbox.build_boxgets the same);build_sourcewalks the module's AST frombuild.Route: none (tools only). New fixes: none. Claims issue: none.
One-time cost: the key changes, so each box installs once more on its first run after this merges.
Model
anthropic:claude-opus-5-5, hand-written in a Claude Code session.
Evidence
tools/test_vm_parallel.py(in the client job andwend check), with a stand-in QEMU that binds its hostfwd port as QEMU does: two guests run at once on different ports, ssh targets the guest's own port, a taken port is retried, any other QEMU failure fails at once, a second run waits on an install lock, and the cache key ignores replay-only changes but changes for the kickstart, install,Guest.reboot, a constant the install uses, and the box's checks.wend checkpasses locally. The gate copies tools/ from base, so remi is the real test: one replay of an already-proven route, then two side by side.Self-review (AGENTS.md step 5)
git commit -s)🤖 Generated with Claude Code
https://claude.ai/code/session_018rvHWMQYKBKpBz9S6j5fnj
Generated by Claude Code