Repository navigation
✨ Show concurrent CI and Review lanes with evidence-based Fixing - #16384
Conversation
Signed-off-by: Andrew Anderson <andy@clubanderson.com>
|
[APPROVALNOTIFIER] This PR is NOT APPROVED This pull-request has been approved by: The full list of commands accepted by this bot can be found here. DetailsNeeds approval from an approver in each of these files:Approvers can indicate their approval by writing |
Open PR overlap checkBase branch: Conflicts are checked with every PR merged onto the current tip of
|
|
Hive status — updated in place; edits do not notify. Full history is in the Activity Monitor. clubanderson this PR is held at Failing checks:
Next: the hive does not push to your branch. Push a new head (or reply here); CI and the review swarm re-run automatically and the PR returns to the normal path. Review findings from Blocker is mechanical: the What I read and verified at head
Not a defect, just a note for the maintainer: the body lists four independent concerns (UI tracks, source-edit evidence, reviewer receipts, required-check bindings) in one Review findings from Read the 1. Related open question rather than a finding: when the manager's UID differs from the agent's (the 2. Coverage gate is red on this head (blocking). 3. Extra REST call per PR per poll (low, verify). What I read and found correct: the |
Signed-off-by: Andrew Anderson <andy@clubanderson.com>
There was a problem hiding this comment.
One red check on fc4bbc5 is caused by this diff and is a two-line fix; the rest of CI was still running when I looked, and I have not reviewed the 61-file runtime change itself — that needs a human reader.
relative links and anchors resolve fails because of this branch's client.go additions.
CI log: DRIFT src/docs/scale-envelope.md:37 src/pkg/github/client.go:5075-5085 -> … lacks anchor(s) near cited range: AgeMinutes.
- On
v5today,RankActionableIssuessits atsrc/pkg/github/client.go:5075andscale-envelope.md:37citesclient.go:5075-5085— in sync. - This PR adds +32/−11 to
client.goabove that function (not touching the doc), so on the merge resultpull/16384/mergethe function is atclient.go:5096(AgeMinutescompare at:5104) while the doc still says5075-5085. - The branch's own copy of the doc is older still (
5068-5078, vs.5092in the branch tree), which suggests it predates the last v5 top-up of that doc.
Fix: rebase on current v5 and run src/scripts/check-docs-citations.py --fix (re-anchors drifted citations by Go declaration), which should rewrite line 37 to the new range. I expect no other change is needed for that check.
I did not evaluate the CI/Review-lane rendering or the evidence-based Fixing logic; at 3.5k added lines across dashboard, engine and docs that deserves a maintainer read rather than a partial one from me.
— hive: agent=reviewer-2 backend=copilot model=claude-fable-5.1 copilot=1.0.88
Signed-off-by: Andrew Anderson <andy@clubanderson.com>
Signed-off-by: Andrew Anderson <andy@clubanderson.com>
There was a problem hiding this comment.
Read the new crux/agent/code_work.go end to end, the manager wiring in busy_visibility.go, the pkg/github check-run selection changes and pkg/review/dispatch.go. The github/review side looks correct to me; the crux "evidence-based Fixing" collector has one gap that makes it inert on the layout it targets, plus the red coverage gate.
1. verifySourceEdit cannot read a per-UID agent's worktree (medium, correctness) — src/crux/agent/code_work.go:148-153. scanCodeWork runs in the crux manager process and shells git -C <dir-of-edited-file> rev-parse --show-toplevel against the agent's checkout. On the per-UID interactive layout (authprobe.go:25-31, the default since #4596) that checkout is owned by the agent's UID while the manager runs as a different user, so git ≥ 2.35.2 refuses with detected dubious ownership in repository before it does anything; ok is false and the function returns nil for every edit. Net effect: agent.CodeWork is never set on per-UID hosts, trackerAssignedCodeWorker (stage_tracker_activity.go:221) never binds, and the row degrades to the "no evidence" state the PR describes as the fallback — i.e. the headline feature only works where manager and agent share a UID. code_work_test.go only exercises a same-user fixture, so this does not show up in tests.
2. Security note for whichever fix you choose for (1) (code_work.go:194-199). If the answer is -c safe.directory=… (or the manager already shares the UID somewhere), the manager is then running git diff/git ls-files in a repository whose .git/config the agent controls, and those two commands honour repo-local core.fsmonitor (an arbitrary command) and hooks. Passing -c core.fsmonitor=false -c core.hooksPath=/dev/null on the command line (CLI -c overrides repo config) closes that without changing the verification. Not a defect in the current head — git's ownership check is what currently blocks the path — but it becomes one the moment (1) is fixed.
3. Coverage gate is red because of this file — the coverage job at this head reports crux/agent: 88.7% (floor 89%). The uncovered paths are all in code_work.go: apply_patch parsing (lines 110-131), the git@ → ssh:// rewrite (177-182), the ls-files --others new-file branch (199), and scanCodeWork (215+). A test for apply_patch with an *** Update File: block and one for an untracked new file would cover (3) and are also the two input shapes the fixture never exercises.
4. Minor — workDir is accepted by ObserveCodeWorkTranscript (code_work.go:37, runtime/code_work.go:14) and never read; the comment at line 121 explains relative paths are deliberately not resolved, so either drop the parameter or use it.
Verified, no action needed: the new checkRunIsNewer ordering (commit_ci.go:1160-1163, ID before started_at) does change the canonical selector used by EvaluateCommitCI/DCO, but check-run IDs are monotonic on github.com and commit_ci_order_test.go pins the transitivity property; checkRunsOnHead in EvaluateCommitCI is a no-op filter for ListCheckRunsForRef(sha) results. PendingReviewCompleted is a pure extraction of the previous loop body with identical semantics. The transcript scan cadence (transcriptLivenessScanInterval = 10s) is comfortably inside codeWorkFreshWindow (2m).
Summary: the dashboard/github half is sound; the crux collector needs (1) addressed and the coverage floor restored before this is mergeable. I have not approved or requested changes as a review event.
— hive: agent=reviewer backend=copilot model=claude-fable-5.1 copilot=1.0.88
Signed-off-by: Andrew Anderson <andy@clubanderson.com>
Signed-off-by: Andrew Anderson <andy@clubanderson.com>
There was a problem hiding this comment.
Partial review — the security-sensitive new code checks out; the bulk needs a human read and CI has not reported on 0e301fb yet.
What I verified (read at head 0e301fb):
changelog.d/security-activity-code-work-git.mdclaims the new Git inspection "cannot execute workspace hooks, filters or helpers". The code backs that:src/crux/pushbroker/workspace_observation.gorunsgit config --list -z --show-scopefirst and rejects the repo ifworkspaceConfigOffenders(an allowlist —pushbroker.go:175-216, any non-allowlistedlocal/worktreekey is an offender, socore.fsmonitor,core.hooksPath,filter.*,credential.*etc. all fail closed) returns anything.workspaceObservationEnvadditionally strips inheritedGIT_*, setsGIT_CONFIG_GLOBAL/GIT_CONFIG_SYSTEM/GIT_ATTR_NOSYSTEMto/dev/null, pinssafe.directoryto the discovered root only, and setsGIT_ALLOW_PROTOCOL=empty.hash-object --no-filtersis used so a clean filter can never run. Every command isrev-parse/remote get-url/ls-tree/ls-files/check-attr/hash-object— none invoke hooks. The post-read re-check of root/HEAD/origin/config andos.SameFile+ size + mtime stability binds the result to a stable file. Looks correct to me.src/crux/agent/code_work.go: onlytool.execution_completeevents withSuccess == true, empty error, and withincodeWorkFreshWindowcount; only absolute paths are attributed (sourceEditPaths), relative patch paths are deliberately dropped. Transcript read is bounded at 8 MiB; git calls at 2 s. The mtime window checkstarted <= mtime <= atbefore and after the git observation is sound.
What I did not read: the ~4k lines across pkg/dashboard/stage_tracker*, pkg/github/{commit_ci,pr_ci_activity,required_check_bindings,pr_batch_graphql}.go, automerge_sweep.go, review/dispatch.go and the tracker JS. Those carry the actual CI/Review-lane semantics and the automerge_sweep.go change is adjacent to merge authority; I have no finding against them, only no coverage. All CI checks on this head are still unreported (mergeStateStatus=BLOCKED), so there is no independent signal yet either.
Suggestion for the human picking this up: start with src/pkg/github/automerge_sweep.go and required_check_bindings.go — the PR body says native gates and merge authority are preserved, and those two files are where that claim would break.
— hive: agent=reviewer-2 backend=copilot model=claude-fable-5.1 copilot=1.0.88
Current-head native review handoff (not an approval)Target head: Actual full-test run 41589851329 passed all 23 Go test shards and the test aggregate. The new Git-observation security regressions have actual PASS receipts. Coverage is still blocking: job 125765329501 reports The actual current-head partial review is recognized: its security inspection is not a complete configured review or aggregate approval. Please continue the genuinely assigned native dispatch and collect any missing correctness, security, intent-alignment, style and docs-currency perspectives with each real reviewer's own matching head/report/dispatch receipt through the existing review relay and collector. If matching reports or assignments already exist, collect/continue them rather than duplicate or redispatch them. Preserve any genuine human gate. This is one factual handoff, not a verdict, assignment, approval-label request or gate bypass. Do not reuse the older |
Signed-off-by: Andrew Anderson <andy@clubanderson.com>
There was a problem hiding this comment.
Partial review — this is 71 files / +4581 lines, and I read only the new Git-inspection path in full, since it carries the security- changelog fragment. I did not read the stage_tracker*/tracker-row*.js dashboard changes or the pkg/github changes this kick; those still need a human (or a later pass). Please read the below as "the security module checks out", not "the PR checks out".
src/crux/pushbroker/workspace_observation.go — no defect found. The hardening claim in changelog.d/security-activity-code-work-git.md is supported by the code:
- Only read-only subcommands run (
config --list,rev-parse,remote get-url,ls-tree,ls-files,hash-object --no-filters,check-attr), all with--literal-pathspecs(:62-63). - Repo-scoped config is gated by the broker's existing
workspaceConfigOffendersallowlist (:65-68, defined inpushbroker.go:175) before any other git command runs, andworkspaceConfigOverrides(hooksPath=/dev/null, fsmonitor=false, credential.helper reset —pushbroker.go:84-89) are applied at command scope on every invocation (:154-160). - Env strips
GIT_*and the push-token var, setsGIT_CONFIG_NOSYSTEM/SYSTEM/GLOBAL=/dev/null,core.attributesFile=/dev/null,safe.directory=<discovered root>(not*),GIT_NO_REPLACE_OBJECTS=1,GIT_ALLOW_PROTOCOL=(:151-162). - Output is bounded at 64 KiB (
:206-212), and the result is re-bound to toplevel/HEAD/origin/config plusos.SameFile+size+mtime of the observed file before returning (:128-146), so a TOCTOU swap after the allowlist read fails closed.
src/crux/agent/code_work.go — the transcript collector only attributes absolute paths from explicit edit/create/apply_patch tool calls that completed with Success=true and no error within the 2-minute window (:68-78, :92-137), and verifySourceEdit requires the file's mtime to fall inside [call start, completion] (:141-144). workDir is intentionally unused per the comment at :124-125. Nothing to flag.
Suggestion (not a defect): given the size, splitting the crux module from the dashboard/github changes would let each land on its own review.
— hive: agent=reviewer-2 backend=copilot model=claude-fable-5.1 copilot=1.0.88
|
Current-head review handoff: The published partial reviews explicitly cover only the repository-verification path, not the complete dashboard and CI-observation changes. Please complete the five configured perspectives on this exact head, including parallel tracker rendering, actual source-work attribution, required check/App and legacy-status observations, reviewer completion receipts, and native stage/count preservation. Submit actual head-bound reports through the existing reviewer relay so canonical collection can evaluate them. If a full pass requires a maintainer decision, record that scope explicitly rather than carrying an older partial review as approval. This is a factual author-side collection request, not approval, a verdict artifact, a request to weaken gates, or a request to rerun CI. |
|
merged by Hive trusted-author auto-merge: author @clubanderson holds role owner and GitHub permission admin on repo; CI green at be5e3de. |
|
Thank you for your contribution! Your PR has been merged. We'd love to hear how your experience was: share feedback |
There was a problem hiding this comment.
Read the diff against the v5 merge-base plus the callers (busy_visibility.go, status_builder.go, stage_tracker_activity.go, merge_block_reason.go). This looks correct to me; nothing blocking.
What I verified:
src/crux/pushbroker/workspace_observation.go— the observer runs only read commands under--literal-pathspecs, rejects any config key outside the existingworkspaceConfigKeyAllowedallowlist before anything else, scrubsGIT_*/push-token env, pinscore.hooksPath/core.fsmonitor/credential.helperviaGIT_CONFIG_*overrides, useshash-object --no-filters, caps output at 64 KiB, and re-reads toplevel/HEAD/origin/config plus re-stats the file before returning.Lstat+IsRegular()means a symlinked path from the transcript is dropped (code_work.go:159). A 2 s context (codeWorkGitTimeout) bounds a hostileinclude.path. I could not find a way for an agent-supplied path to reach a hook, filter or helper.- Cross-kick staleness:
status_builder.go:1232-1238dropsCodeWorkwhoseKickAtno longer equalsproc.LastKick, andtrackerAssignedCodeWorker(stage_tracker_activity.go:229-243) additionally requires head/repo match, a current agent-caused audit action on the same PR, and bothAt/ObservedAtinside the fresh window. The "Fixing" signal cannot be satisfied by an assignment or an administrative write alone, which matches the body. commit_ci.go—checkRunIsNeweris now a single lexicographic order (ID first), so selection is transitive;latestCheckRunsByNameAndAppstill yields one run per (name, app, head). The GraphQL batch now supplies realdatabaseId/app/oidinstead of the synthetici+1IDs, so both paths feed the same selector.review/dispatch.go—PendingReviewCompletedis a pure extraction of the oldremoveCompletedReviewsbody; same predicate.
Two low notes, neither a blocker:
src/crux/agent/code_work.go:37—workDiris accepted byObserveCodeWorkTranscript(and theruntimewrapper) but never read; relative paths are intentionally dropped atcode_work.go:108/133. Either use it to anchor relative paths or drop the parameter so the signature doesn't promise something it doesn't do.runtime.ObserveCodeWorkTranscriptalso has no non-test caller yet.src/pkg/github/protection_facts.go:100-120— theGetCombinedStatusfetch now fires for any PR where a required context has no check run on the head (previously only whenmergeable_state=blockedand nothing proved check-run reporting). On repos whose required set includes a legacy status (e.g. atidecontext) that is one extra paginated call per PR per enrichment. Also, when that read fails,missingstays empty andMissingRequiredChecksis now[]rather than the previously-listed names; since the only consumer is the explanation string inmerge_block_reason.go:101, this is display-only ("unknown" rather than "not reported"), and it matches the stated intent — just flagging that the block reason gets quieter on a transient status-API error.
— hive: agent=reviewer backend=copilot model=claude-fable-5.1 copilot=1.0.88
… as chevrons (#16464) Restore content-driven chevron height (one line of caption -> one-line chevron, two lines -> two, never more) and centred captions, which #16454 replaced with a fixed two-line grid row and a flush-left stretched copy box. Drop the boxed-card opt-out that #16384 gave the concurrent CI / Review lanes so they share the stage chevron clip-path, padding and height, with lane state expressed through fill and text colour. Geometry probe and CSS-contract tests now fail on the previous behaviour; density fixture gains a one-line held row. Signed-off-by: Andrew Anderson <andy@clubanderson.com>
The chevron-strip grid rule from #16384, `.ham-strip:not(.ham-head-cols)`, also matched the title line: `ham-head-cols` is a class on the strip's row, not on the strip. The title line's lead / PR / cap cells landed in the first three chevron-wide grid columns, and cap pills wider than one column (`train · joins behind #N next pass`, `stage ↓`) right-aligned out of the cap cell and drew over the PR title and base pill. Spell the exclusion on the row so only the chevron strips and the column header are grids; the title line goes back to its flex sizing. Claude-Session: https://claude.ai/code/session_01BQYnPGFJx81c8JXR3o1sg1 Signed-off-by: Danathar <Danathar@users.noreply.github.com> Co-authored-by: Danathar <Danathar@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Changes
Evidence and remaining gates
Independent precommit review cleared exact tree f13f52acbf025538a92cf61e59a92097814f9585 after correcting all six findings. The implementation owner reports 173 passing standalone Node tests and regenerated assets. Actual Go build/tests/lint/coverage and canonical guards are pending remote CI under repository policy; source review is not a CI verdict. Embedded extraction has separately disclosed baseline failures and extraction gaps and is not represented as a full passing suite.
No production deployment or integration-execution activation is included.