[CamiloSierraH]

Why

PR #56 collects what each Keeper member reports about itself (role, Raft log position, snapshot size and settings), and the pre-pass draws its HC-3.18–3.20 findings from it. Only the skill could see those findings. A customer opening dashboard.html saw nothing of them: not the member that fell behind the leader's retained log, and not the 2 GiB snapshot sent unchunked.

Builds on #56 (merged). It reads the keeper/ files #56 introduced. Main has been merged in, so the diff is now only this PR's own 7 files.

What changed

The dashboard's Keeper section gains a Keeper members table, one row per member:

member · ensemble · state · role · version · Raft log first → last · committed / leader · latest snapshot · chunk size · snapshot format · on disk (snapshots / log)

• Marks (⛔ / ⚠, the reason on hover) follow the pre-pass's checks:
  • not serving, or probe failed (HC-3.18);
  • behind the leader's retained log, needs a full snapshot (HC-3.19, only against exactly one observed leader in the same ensemble; with none, or two mid-election, there is no verdict). The row turns red;
  • an unchunked snapshot over 2 GiB (critical) or over 1 GiB (warning) (HC-3.20);
  • snapshot format below 7, judged against the member's own version (HC-3.20).
• A banner appears when no member of an ensemble in the bundle is serving. It says that members outside the bundle may still hold a quorum, because the bundle does not prove voting membership. Apache ZooKeeper's "This ZooKeeper instance is not currently serving requests" counts like Keeper's wording.
• Its own visibility. The table is shown whenever keeper/ is in the bundle, even without metric_log. It is absent, not empty, without the files: older bundles, cloud by default, -keeper-mntr off. The existing metric-driven parts of the section keep their behaviour.

internal/keeper.ReadMembers ports the pre-pass's rules to Go. A truncated reply is not evidence, members are grouped by ensemble, and the leader is required for the snapshot verdict. Versions drop mntr's build suffix.

The rules now exist twice, Python and Go, so a test keeps them together. It runs the pre-pass and ReadMembers on the same fixtures (incident, leaderless, growing snapshot during a rolling upgrade, no leader, two ensembles) and requires the same verdict per member. On the incident fixture it compares 8 verdicts, and changing only the Go side fails it.

Verification

• Reader tests: the incident shape, a healthy ensemble with no marks, no leader means no verdict, a truncated reply, two ensembles, a member without the ensemble line, the format per member's version, the version suffix and Leaderless. Each rule fails its test when reverted.
• Template and page data: the panel has its own switch and escapes every file value. It adds nothing to the page without keeper/.
• Lab, real run: a healthy 3-member ensemble shows three members serving and no marks, only ℹ for format 6, the default on 26.7. On the incident-shaped demo bundle, the lagging member's row is red with "needs a full snapshot", and the 2.24 GiB unchunked snapshot is marked on the leader and the follower.
• Review round: ZooKeeper's not-serving wording (collector, both readers), two leaders mid-election and the banner's caveat. The consistency test gains a ZooKeeper and a two-leader fixture; both caught the readers disagreeing before the fix.
• gofmt, go vet (also GOOS=windows) and make test are green: 305 passed.

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

Unresolved classification and leader-selection errors can produce misleading health verdicts, and the outage banner overstates incomplete evidence.

3 open issues

• Inconsistent leader selection between dashboard and pre-pass · New
• Leaderless status ignores uncollected quorum members · New
• Unrecognized ZooKeeper not-serving response masks outages · New

What changed in this PR

Adds a Keeper members panel to the diagnostic dashboard, building on #56’s collection of member-level Raft and snapshot information.

Changes:

• Displays member status, log positions, snapshot settings, and health marks independently of hourly metrics.
• Adds Go reader logic and Python/Go consistency tests.
• Expands Keeper collection and documents the diagnostic checks.

File	Description
skills/​clickhouse-diagnostic/​scripts/​inspect_bundle.py	Adds Keeper parsing and health checks.
skills/​clickhouse-diagnostic/​references/​running-the-tool.md	Updates Keeper collection guidance.
skills/​clickhouse-diagnostic/​references/​known-patterns.md	Documents snapshot-transfer failure patterns.
skills/​clickhouse-diagnostic/​references/​health-checks.md	Defines HC-3.18–3.20.
skills/​clickhouse-diagnostic/​references/​file-guide.md	Explains member evidence and dashboard marks.
skills/​clickhouse-diagnostic/​references/​bundle-layout.md	Documents expanded Keeper files.
README.md	Describes collection and the new panel.
internal/​keeper/​members.go	Parses members and computes verdicts.
internal/​keeper/​members_test.go	Tests member parsing and verdicts.
internal/​keeper/​fourletter.go	Expands parallel collection and gov filtering.
internal/​keeper/​fourletter_test.go	Tests collector behavior and isolation.
internal/​dashboard/​generator.go	Adds panel data and rendering.
internal/​dashboard/​alerts_panel_test.go	Tests panel structure and data.
cmd/​prepass_keeper_test.go	Tests Python Keeper checks.
cmd/​main.go	Updates collection messaging.
cmd/​keeper_consistency_test.go	Compares Python and Go verdicts.

_{Review effort: Balanced}