Repository navigation
fix: web preview reachable through the host port forward - #198
Conversation
…oject The managed AGENTS.md zone now carries standing web-preview guidance in every versioned project, even with no pinned skills: a dev server must bind 0.0.0.0 inside the guest to be reachable through the host port forward, and the host side of the forward binds 127.0.0.1 only, so the guest-side bind exposes nothing on the network. Local-only selections strip the zone entirely instead of leaving a pin-less remnant. The local-only startup preflight now blocks only when the guest zone still lists pinned skills; the standing guidance alone is not evidence that versioned skills survived the switch. Init always owns the zone, which also makes a skills-less Albert Code import idempotent: the second import re-renders an identical zone instead of reporting a repair. Part of #153. 👾 Generated with [Letta Code](https://letta.com) Model: lc-qwencloud/deepseek-v4.1-flash Co-Authored-By: Letta Code <noreply@letta.com>
`just-code check` now probes every port the project explicitly mapped, from the host side of the forward, in both isolation modes. Any HTTP answer counts as reachable; no answer prints the guest-side fix (bind 0.0.0.0 inside the guest) instead of the wrong LAN-exposure warning the agent used to give. The probe is print-only and runs even when the OpenCode health endpoint is down, because preview reachability does not depend on backend health. The legacy implicit 3000-3010 range stays unprobed: those are defaults, not project choices. Part of #153. 👾 Generated with [Letta Code](https://letta.com) Model: lc-qwencloud/deepseek-v4.1-flash Co-Authored-By: Letta Code <noreply@letta.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. |
The managed AGENTS.md zone now exists even without selected skills and carries the guest-side 0.0.0.0 binding guidance; the local-only refusal now depends on leftover skill pins, not the zone itself. `just-code check` reports reachability of explicitly configured ports in both isolation modes. Part of #153. 👾 Generated with [Letta Code](https://letta.com) Model: lc-qwencloud/deepseek-v4.1-flash Co-Authored-By: Letta Code <noreply@letta.com>
![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: 6e8fe67eeb
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".
Address the Codex review findings on the standing managed zone and the check port probes: - The standing web-preview guidance describes the Microsandbox host port forward. On tart and agent-vm it gave agents an ineffective verification command and an incorrect exposure assurance, so MergeManagedInstructions now renders it for Microsandbox only; other runtimes get the pinned-skill list alone and no managed section at all without pins. - The configured-port probes in `check` hit 127.0.0.1 on the host, which only means something on Microsandbox: tart connects to the VM IP directly and agent-vm forwards dynamically, so a probe there could report an unrelated local service as a preview forward. The probe now runs only on Microsandbox. - The probe client no longer follows redirects: a redirect answer is reachability, while following it could send a diagnostic request outside loopback or misreport a dead redirect target as a dead forward. Docs updated to state the Microsandbox scope of both behaviors. 👾 Generated with [Letta Code](https://letta.com) Model: lc-qwencloud/deepseek-v4.1-flash Co-Authored-By: Letta Code <noreply@letta.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d23888513a
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".
managedInstructionRuntime parsed the stored manifest runtime with the host-OS-gated parseRuntimeName, so a tart or agent-vm manifest read on Windows rendered as Microsandbox and skill update, rollback, and recovery rewrote its managed section with the port-forward guidance the scoping excludes. Stored names now resolve without host compatibility checks; those checks belong to runtime execution, not instruction rendering. The wizard validated AGENTS.md even for a pin-less tart or agent-vm init whose merge changes nothing, so a symlinked AGENTS.md blocked an init that touches nothing. The plan now claims and validates AGENTS.md only when pins, the Microsandbox standing guidance, a local-only strip, or a stale managed zone require changing it. Also platform-gate the two tart tests that failed the Windows CI legs: the dispatcher test points the platform-gated sweep at darwin to keep the tart path under test everywhere, and the wizard test skips where tart is not selectable, matching the existing mounting test. 👾 Generated with [Letta Code](https://letta.com) Model: letta/auto Co-Authored-By: Letta Code <noreply@letta.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 18c559dbd0
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".
Port mappings changed in the manifest only apply when the Microsandbox is recreated, so probing the desired values could report an unrelated service on a newly configured host port as a reachable preview, stay silent about a removed mapping the sandbox still forwards, and diagnose a missing guest bind for a forward that does not exist yet. The probe now reads the running instance persisted network port list (the same authority ServerPortForwarded uses) and reports three states: mappings both configured and applied are probed as before, configured mappings the sandbox does not forward are reported as pending recreate instead of being probed, and applied mappings no longer configured are reported as still forwarded until recreation. The OpenCode backend server forward is excluded: ValidatePortMappings keeps DefaultPort out of user mappings entirely. When the applied list cannot be read the probes are skipped rather than falling back to the manifest view. 👾 Generated with [Letta Code](https://letta.com) Model: letta/auto Co-Authored-By: Letta Code <noreply@letta.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d61f498b0d
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".
Two review findings on the previous round, both edge states of the new conditional behavior: `ports remove` of the final entry persists portsConfigured with an empty list, and the probe returned before reading the applied mappings, so the sandbox kept forwarding the removed mapping until recreate while check printed nothing. The early return now keys on portsConfigured alone: an explicitly emptied mapping list still counts as configured, and the stale-forward report covers it. A local-only init on a pin-less tart or agent-vm project forced the AGENTS.md claim even when the file has no managed zone, so a symlinked AGENTS.md still blocked an init whose strip would change nothing. The local-only flag no longer forces the claim by itself: the zone-presence check decides, so a zone-less file is left untouched while a stale zone is still stripped. 👾 Generated with [Letta Code](https://letta.com) Model: letta/auto Co-Authored-By: Letta Code <noreply@letta.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a365020c6a
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".
A local-only plan never merges managed instructions, so pins and the Microsandbox runtime must not force the AGENTS.md claim on their own: with no managed zone the strip changes nothing, and a symlinked AGENTS.md was still rejected for a local-only Microsandbox init or one carrying a local skill selection. The claim predicate now requires a non-local-only plan for those forcing conditions; local-only plans claim only when the zone-presence check finds a stale managed zone to strip. 👾 Generated with [Letta Code](https://letta.com) Model: letta/auto Co-Authored-By: Letta Code <noreply@letta.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f5d06556a0
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".
…ning A plan leaves AGENTS.md unclaimed when it would change nothing, so a managed zone appearing between the interactive Plan and Apply steps was never re-read: the manifest committed as local-only while the newly added versioned pins survived in AGENTS.md. Apply now rechecks the unclaimed path and, when a zone has appeared, claims it and strips or refreshes the zone on current content — the same drift adaptation the claimed path already performs. A non-regular AGENTS.md holding the appeared zone aborts the apply rather than writing through it. 👾 Generated with [Letta Code](https://letta.com) Model: letta/auto Co-Authored-By: Letta Code <noreply@letta.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 51665d259e
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".
Both zone-presence checks read AGENTS.md before inspecting its type, so a FIFO or a symlink to an unbounded device such as /dev/zero could block indefinitely or exhaust memory instead of being refused. The new managedZoneAt helper resolves the entry on the real filesystem first — following symlinks so a link to a regular zone-less file stays untouched — and reads only a regular target. A missing or dangling path still reads as no zone; a non-regular or unreadable entry fails closed into the existing refusal. Covered for both the planning probe and the apply-time recheck by unix-only tests. 👾 Generated with [Letta Code](https://letta.com) Model: letta/auto Co-Authored-By: Letta Code <noreply@letta.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f8aa5630dd
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".
managedZoneAt validated one filesystem object and then re-opened the path, so an entry swapped in between the Stat and the read could still be a FIFO or a symlink to /dev/zero and block or balloon the probe. It now opens the path once with O_NONBLOCK, inspects the opened handle, and reads only a bounded regular file; a larger file fails closed into a claim instead of being decided from a truncated read. Port probing ran one 2s-timeout probe per mapping with no total bound, so a manifest with many dead forwards could stall just-code check for minutes. The probe loop now runs under a shared 10s budget: each probe keeps its own shorter timeout, and mappings beyond the budget are reported as not probed rather than silently skipped. 👾 Generated with [Letta Code](https://letta.com) Model: letta/auto Co-Authored-By: Letta Code <noreply@letta.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e85152c5e9
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".
The oversize fail-closed from the previous round re-broke the zone-less-symlink property (a symlink to a large regular file was refused instead of left untouched) and only moved the full read one step later, into Plan and Apply. The probe now scans the opened file line by line for the begin or end marker — either marker claims the file, matching hasManagedSkillZone, where a balanced pair is a zone and any other combination fails closed — in bounded chunks. A line longer than the buffer cannot be a marker and is skipped whole, so file size no longer bounds memory, and an oversized zone-less file is not claimed at all. 👾 Generated with [Letta Code](https://letta.com) Model: letta/auto Co-Authored-By: Letta Code <noreply@letta.com>
|
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". |
1 participant
Problème
Dans la VM, l'agent lance les serveurs de développement sur
127.0.0.1: le port forwardé depuis l'hôte ne répond jamais, la prévisualisation web est inaccessible. En plus, quand l'utilisateur demande un bind0.0.0.0, l'agent le refuse en invoquant une fausse exposition réseau — alors que côté hôte le forward n'écoute que sur127.0.0.1et n'expose rien sur le réseau local.Closes #153
Changements
Instructions gérées toujours présentes (
fix(skills)) :AGENTS.mdporte désormais un texte permanent qui énonce les faits : bind0.0.0.0requis dans l'invité pour être joignable via le forward, forward hôte limité à127.0.0.1, une requête lancée depuis l'invité ne prouve rien,just-code checkvérifie depuis l'hôte ;initpossède toujours la zone, ce qui rend un import Albert Code sans skills idempotent (le second import re-rend une zone identique au lieu de signaler une réparation).Sonde de ports dans
check(feat(check)) :just-code checkinterroge chaque port explicitement configuré par le projet, depuis l'hôte, dans les deux modes d'isolation ;0.0.0.0), sans le faux avertissement d'exposition réseau ;check, et s'exécute même si le backend OpenCode est en panne ;Tests
internal/justcode: zone permanente sans épingles (init vide, clearing, import idempotent), préflight sensible aux épingles réelles, sonde joignable / non joignable / plage legacy silencieuse ;go test ./...verte.👾 Generated with Letta Code