Repository navigation
Conversation
4dc72aa to
c43cac8
Compare
Opus 5.5 Review (fork) — ✅ no blocking findingsReviewed |
Design Review (Opus 5.5, fork) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: the gateway-run strict-tier checkout reuses the reviewed Notes
[DESIGN-REVIEWED] 968b006 Description read[DESCRIPTION-READ] 38453586494464efa875fca4789fbc8903cd7eb5146f621d795c03be56105d45 That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
UX Review (Opus 5.5, fork) — ✅ PASSUX-level review of UX-Verdict: PASS The screenshots show every new control and state: the More actions trigger, Refresh, Release worktree, Creating…, the success toast and a refusal banner. Nothing blocks. No blind reader has seen them, because this lane runs without one. Pushing the branch to this repo runs that read. Notes
[UX-REVIEWED] 968b006 Description read[DESCRIPTION-READ] 7ecb02e3cb3200e45c50543f821814067991d31640a605b52cacd149a5ae22b0 That sha256 names everything this verdict read: the PR description, plus the 7 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
First Principles Review (Opus 5.5, fork) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks. Every item meets a named need. Two kinds of surface have only one caller. Notes
What this change shipsInventory (8 items) — 6 justifiedIntent: run the newest stable release as a Dev Fleet worktree without leaving the dashboard — ADDITION.
[FIRST-PRINCIPLES-REVIEWED] 968b006 Description read[DESCRIPTION-READ] 7c74ad0aa213a91bf1e923a700969f398a9c3bb534e27e832f113ce1ce650547 That sha256 names the PR description this verdict read. Recompute it with |
GPT 6.1 Review (fork) — ✅ no blocking findingsReviewed Review detailsNo findings. |
c43cac8 to
9cf097a
Compare
9cf097a to
d7edf2d
Compare
d7edf2d to
c98578d
Compare
|
First Principles PASS on Item 6, drop or gate the refresher's Item 4, |
|
Rebased: The only real red on |
|
Rebased:
|
|
Rebased: The rebase applied cleanly.
Push guard SAFE (single commit on base). |
|
span=204f62c484b8: fixed on
Why not "call Tests: (Re-posted from comment 6375946937 so each disposition record carries one finding; this verdict's items 6 and 8 have their own comments.) |
|
Item 6 (courtesy It was added as the GPT fix for span 9ead27a94181's sibling (an operator's (Split out of comment 6375946937 so each disposition record carries one finding.) |
|
Item 8 (
(Split out of comment 6375946937 so each disposition record carries one finding.) |
|
Rebased: This push re-runs the gates that failed on |
|
Rebased:
|
|
Design PASS on This is the same cost First Principles item 6 raises on this head. The lane's own read is "acceptable for now", and since |
|
First Principles PASS on The subtraction is well-formed: Create calls |
|
UX PASS on The notices render through the shared |
|
UX PASS on The point is right: in both cases the channel resolved and it is the checkout that couldn't be read. A dedicated key means a new string across all 12 catalogs plus the pseudolocale. I'll land it in the same follow-up as the notice copy above, so the error wording changes once rather than in two passes. |
|
UX PASS on The tooltip should name the way out ("remove this worktree from its detail panel, then Create the newest"). It's one more i18n key across the same 12 catalogs as the note above, so it goes into the same copy follow-up rather than this PR. |
iamwhatever
left a comment
There was a problem hiding this comment.
REQUEST_CHANGES at d622fc7.
Four follow-ups are marked 'accepted and deferred' with no GitHub issue: the three UX copy items on this head (raw git stderr in notices 06/07/11, a dedicated string for cases 12 and 14, a next step in the behind-badge tooltip) and the Design prune-abandonment policy for hand-detached trees. Please fix them here, or open one issue per item and link each in the PR description.
Two blockers remain after that, and both need people rather than code:
- A product decision. The maintainer's scope question (comment 6393308238: could a smaller 'create worktree at tag' action do, without the 60 s tag-mirror fetch on every install and the prune exemption?) has no answer, and the Design and First Principles dispositions on this head are 'needs-decision' pending it.
- A security review. The hermetic populate (worktree_ops_release.py#L510) is the fence for #12342. Together with the new POST /api/release-channel/create route, it decides what a checkout may execute, so a human security reviewer must sign it off.
Checks are green and every AI lane passes at this head.
— crew-pr-reviewer (r128)
|
Pushed Four deferred follow-ups, one issue each (
Also on this head, from the local pre-push review lanes:
The two blockers that need people, unchanged and still open:
Local on |
|
Replying to @chenmingwei23's question:
Thanks for asking before reviewing. Your smaller version is what this push is. The user. A Dev Fleet operator answering a "broken in stable" report, who needs The background tag mirror: removed. This version adds no timed fetch, and the Prune merged: unchanged. No prune rule or preselection changes, and no exemption The base-branch fetch: unchanged. The action's fetch is separate from it, so a Size. One action. 31 files, +3,001 lines, of which tests are 1,562 (1,459 backend, Fixed on |
|
self-added: yes
|
|
self-added: yes
|
|
self-added: yes
|
|
Pushed
Left as is: the rounds view reported SEAM LIMIT, so a retrospective ran this round. It also proposed smaller replacements for the second Local evidence on
|
Add an owner-only gateway action that creates release-channel-stable as a sibling worktree detached at the newest published vX.Y.Z, so an operator can pod it or make it live without finding the tag by hand. The gateway asks the configured remote for the newest stable tag and the commit it advertises, fetches only that tag's objects (it writes no ref and no FETCH_HEAD), and runs git worktree add --detach at the strict sandbox tier through handlers/worktree.py's runner, with hooks turned off and a declared content filter refused. The local steps run with ~/.aws, ~/.ssh and the vault bind-masked wherever an OS sandbox backend runs them. The network steps run at the stricter of standard and the install's agent.sandbox tier. The action adds no background tag fetch, no pinned row and no prune-rule change: the tree is an ordinary detached fleet row. The commit verify, the checkout and the HEAD readback pass --no-replace-objects, so a refs/replace graft cannot swap the checked-out tree for another while HEAD reads the advertised oid. A release is a whole vX.Y.Z of ASCII digits with no leading zero; every other v* name the remote lists is ignored, so none of them can refuse the action or outrank the newest release. A partial clone is refused before the fetch, because its checkout would fetch the release's files itself, outside the action's one bounded fetch. The toolbar keeps Prune merged as its one button beside a labelled More actions menu; Refresh moves into that menu next to the new Release worktree item, so the row holds two controls.
|
self-added: yes
|
|
Pushed |
|
Superseded by #33391 -- moved to a same-repository branch because Sensitive Change Review refuses fork PRs that touch |
Problem / Motivation
Goal: From Dev Fleet, check out the newest published stable release as a worktree you can pod or make live, without leaving the dashboard.
You cannot click through a shipped release from Dev Fleet. The fleet lists feature worktrees, and every one is unreleased code. To run what
stableactually shipped you leave the dashboard, find the current tag by hand, and make the checkout yourself. Dev Fleet manages git checkouts, and a git checkout has no release channel.Why it matters
A bug report says "broken in stable". Answering it means running stable. Today that
is manual work outside the tool that exists to manage checkouts, so it mostly does
not happen, and a fix gets tested against
maininstead of against the build thereporter is on.
Not a goal
nightlytags nothing, so its row would bemain;insidertags resolve, but ordering-insider.Nagainst-rc.Nneeds a rule this repository does not state.$KIROCREW_HOME/channel.What changed (motivation → approach → change)
Dev Fleet's toolbar gains a labelled More actions menu beside Prune merged, which stays the toolbar's one button; the menu's trigger is the second control (
max-two-buttons-per-rowcounts it). Refresh moves into that menu, next to the new Release worktree item. It createsrelease-channel-stable: a real worktree detached at the newest publishedvX.Y.Zrelease. Pods and Make Live already drive any fleet row, so a shipped build boots the same way a feature branch does.The approach is resolve-when-asked, not a standing mirror. The gateway asks the configured remote (
git ls-remote) for the newest stable tag and the commit it advertises, fetches that one tag's objects, and checks it out detached. The action adds no background fetch, no pinned-row UI and no private common dir.dev_fleet/release_worktree.pyvX.Y.Zof ASCII digits with no leading zero and ignoring every otherv*name the remote lists, confirms its advertised commit, refuses a partial clone (whose checkout would fetch the release's files itself), fetches that one tag, then runsgit worktree add --detachfrom the gateway at the strict sandbox tier, with--no-replace-objectson the commit verify, the checkout and the HEAD readback so arefs/replacegraft cannot swap the checked-out tree; refuses a name already registered at another path; a request cancelled mid-checkout settles before it unwindsdev_fleet/runtime.py_run_cmdgains two keyword-only opt-ins, a sandboxed hook-name scan and a byte-capped incremental read; both default to today's behaviour and only the release action's network steps pass themdev_fleet/gateway_routes.pyPOST /api/apps/dev-fleet/create-release-worktree, owner-only, every app token refused;{ok, path, name, version, ref, oid}, or a refusal: 503 when a step is refused for lack of an OS sandbox backend, 409 otherwisesecurity_posture.pyrelease_worktree.pywith the other Dev Fleet modules that redact text for the app's own replies and audit recordswebsite/src/apps+website/src/i18n/localesDevFleetPage.tsx, five strings in all twelve locales), and the gateway-route note (devFleetApi.ts)docsdev-fleet.mdspec sectionwebsite/capture+website/scriptscapture-23344-release-worktree.mjsdrives the isolateddevfleet-release-worktreeentry (the realDevFleetPageover afetchstub, fixture data only);website/.jscpd.jsonexempts that entry from duplicate detection against its Dev Fleet sibling, by fileWhy the gateway and the strict tier:
git worktree addreads the repository's own.git/configand followsinclude.path, so a hostile config could name a content-filter driver. The checkout and the other local steps, the worktree-list and filter probes included, therefore run from the gateway at the strict tier (~/.aws,~/.sshand the vault bind-masked wherever an OS sandbox backend runs them), reusinghandlers/worktree.py's filter probe and runner. The two network steps (ls-remote, the one-tag fetch) run at the stricter ofstandardand this install's effectiveagent.sandboxtier, so no git child the action runs is less confined than an agent's own commands; on acc/strictinstall they clamp up to that tier. The route's main-checkout discovery, which runs before the action, is outside that statement. On a host with no backend that permits unsandboxed execution (agent.sandbox_allow_unsandboxed_execin effect, its default on Windows only, and no governancesandbox.min_levelfloor aboveoff), every step runs its git unsandboxed with a credential-scrubbed environment, as agent spawns there do. Neither comparison covers an agent session whose isolation Kiro Crew leaves to the Kiro CLI's own sandbox (macOS with that sandbox switched on, or the official Kiro CLI on Windows). The cost: credentials are withheld abovestandard, so a private remote cannot be fetched there, and the action refuses with a worded reason that names the tier. A public remote and the defaultstandardinstall are unaffected.On those network steps the scan that lists config-defined hook names, so each is disabled by name, runs inside the sandbox at the step's tier rather than as a gateway git. Each step's own git output is read incrementally under a 4 MiB cap, past which the child is killed and the step refuses, so a hostile advertisement is not buffered whole in the gateway. The fetch also resets several repo-config keys git would otherwise execute (
core.sshCommand,credential.helper, gpg) as defense in depth; the boundary is the tier, not that list. Provenance is the whole.git, as Make Live and sync already trust it, except that the local object steps ignorerefs/replace.Backwards compatibility
Compatible: one gateway route, one toolbar menu and its item are added; Prune merged keeps its place and Refresh moves from a toolbar button into the More actions menu (same action, same label), the
/fleetpayload and theworktreesshape are unchanged, and_run_cmd's two new parameters are keyword-only with defaults that keep every existing caller on its current path.Moves one control: Refresh is no longer a toolbar button but the first item of the More actions menu. This is deliberate, because
max-two-buttons-per-rowcounts the menu trigger and Prune merged is the row's primary action; Refresh becomes two clicks, and the fleet already re-reads after every action. Otherwise the diff adds a module, a route, a menu, catalog strings and tests, and narrows no existing contract. The prune rules are unchanged; no exemption is added.Tests
test/test_dev_fleet_release_worktree.py(new, 69): newest-by-semver resolution and the advertised-commit confirm; a tag outside the ASCII release shape (a leading zero, a non-ASCII digit in two scripts, a ten-digit part) never outranks the tip, a tag that is no release refuses nothing however long or many, and a release tag is matched whole; a promisor remote in force, read the way git reads a boolean, is refused aspartial_clonebefore the fetch; the refusals (unpublished, a retracted tag, provenance mismatch, a declared checkout filter, an existing destination, a name registered at another path or left registered at the reserved path, a worktree list that cannot be read, a local step whose git cannot complete, a racing destination claim); after a failure past the claim, a failedworktree addreleases the claim with anrmdir, which removes only an empty directory, and leaves another writer's tree alone; a tree that did not land detached is removed by git, and one git cannot remove is left with the command that clears it; a timed-outworktree addremoves nothing and names the command that clears the tree once the checkout stops; a cancel mid-checkout still runs the readback or the unwind, holding the lock until then; a missing sandbox answerssandbox_unavailableon bothls-remotereads, the fetch, and the worktree-list and filter probes; every network step engages both_run_cmdopt-ins, and the worktree-list probe runs strict through_run_git; the network steps run atstandardby default, the shippedautoincluded, and clamp up on astrictinstall, and a private remote unfetchable there is refused with a worded reason; real-git arms (a plantedfilter.<name>.smudgeis refused and never runs, a--filter=blob:noneclone is refused before the fetch and the remedy its refusal names clears it, apost-checkouthook does not run, a clean repo lands detached on the tag's peeled commit, a plantedcore.sshCommandis neutralized, a plantedurl.<evil>.insteadOfredirects resolution, the readback's removal of its own tree clears even a locked record, a credential-shaped filter name is redacted from the refusal, arefs/replacegraft over the release commit does not change the checked-out tree, and a tag-mapping fetch refspec writes no local tag orFETCH_HEAD); and a parametrized sweep across prune/pruneTags, a tag-mapping or mirrorremote.<name>.fetch,tagOpt,fetch.bundleURIand submodule recursion, proving that under each of them the fetch writes, updates or deletes no ref and runs no fetch inside a submodule.test/test_dev_fleet_run_cmd_opt_ins.py(new, 14): the opt-ins are keyword-only and default to today's path; the default path keeps the unsandboxed scan andcommunicate(); only the release action's network runner passes them (a source scan ofsrc/kiro_crew); each opt-in, engaged, takes its branch; the capped reader kills a child flooding stdout or stderr, including one still writing when the cap trips, and keeps stderr's last line.website/src/apps/dev-fleet/DevFleetPageCoverage.test.tsx(updated): the two post-action refresh cases open the menu and pick Refresh.test/test_dev_fleet_gateway_routes.py(updated): the route is owner-only, refuses every app token, answers 503 for a missing sandbox and 409 for any other refusal, and the registration test lists the new route.website/src/apps/dev-fleet/DevFleetPage.releaseWorktree.test.tsx(new, 4): Prune merged is the one toolbar button and Refresh and Release worktree are items in the menu, its accessible name its visible label and its hint the title; success names the tree and version; the route's refusal reaches the page; the trigger is busy and the reopened item disabled while the request runs.Determinism (testing-conventions § Proving a determinism fix): the three backend files are green 20/20 at
-n0(167 passed) and 10/10 at-n 4 --dist loadgroup --max-worker-restart=2withtest_dev_fleet_app_guards.pyandtest_dev_fleet_repo_accessor.pyas co-tenants (246 passed, 1 skipped), and in a shuffled order (seed 25971, 167 ids). The new frontend test is green 10/10, and the Dev Fleet frontend suite passes shuffled (seed 8317, 212 tests). Static ratchets versus the merge-base (semgrep, jscpd, the flake-pattern ratchet and the website restyle ratchet) report no new findings.Manual verification
The four states were captured from the real
DevFleetPagewithfetchstubbed at the network seam, fixture data only, by the committed harnesswebsite/scripts/capture-23344-release-worktree.mjs— see Screenshots; each scene asserts its state before the shot, and the success scene waits for the new row the fleet re-read returns. The tier and provenance statements inrelease_worktree.pywere checked by real-git probes: git rejects a content-tampered loose object (hash mismatch), and aurl.*.insteadOfredirect is honored, which is why the module states the trust boundary as the whole.git; arefs/replacegraft is now ignored by the local object steps (--no-replace-objects).Screenshots / video
Dev Fleet toolbar, dark theme, attached to this description (nothing is committed). The harness that produced them,
website/scripts/capture-23344-release-worktree.mjs, drives the isolatedwebsite/capture/devfleet-release-worktree.{html,tsx}entry, so a reviewer can regenerate them.The menu open — Prune merged is the one button; Refresh and Release worktree are the menu's items:
In flight — the trigger spins, and the reopened item reads "Creating…", disabled:
Success — the toast names the worktree and version, and the re-read fleet shows the new
release-channel-stablerow:Refusal — the server's reason in the action-error banner:
Related Issues
Closes #23344
#23344's proposed background mirror, pinned fleet row and hermetic private common dir were replaced by resolution when the action runs, after @chenmingwei23's review; the simpler action solves the same problem. #12342 (the TOCTOU race in the earlier hermetic populate) is deliberately not closed here: this version has no populate step, so the race has no counterpart. #26550, #26551, #26552 and #26553, filed against the earlier design, become moot under this version.
Checklist
docs/system-specs/modules/dev-fleet.md)