## Problem
Image references that share a registry and namespace differ only at the
end, but table cells truncated them from the end:
- **Images (name column):** at 390 and 700px, three rows read
`ghcr.io/pulse-demo…`. From 390 to 900px the repository name was hidden
in 3–4 of 8 rows.
- **Swarm services (Image column):** from 390 to 700px, two of three
rows read `registry.…`.
- **Containers (Image column):** the repository name was hidden in up to
7 of 12 rows at 1024px, and in 3 rows even at 1440px.
So an update badge, a size or a service couldn't be traced to its image
without opening the row. Real fleets are worse: every
`lscr.io/linuxserver/*` image shares a 20-character prefix.
## Change
- **`splitDockerImageReference`** (in `dockerImageReference.ts`) splits
a reference just before its repository name. A digest never moves the
split, and a reference without a path stays whole.
- **`DockerImageReferenceText`** mirrors `KubernetesNameText`: the head
truncates, `name:tag` stays visible, and the full reference stays on the
title.
- **Where it's used:** the Images name cell (via `DockerResourceNameCell
imageReference`), the Containers Image column and the Services Image
column.
- **Copying:** head and tail are flex items, and a browser copy would
join them with a line break. So a selection inside one reference copies
without it; wider selections are left to the browser.
- **Unchanged:** sorting, search and the stored value.
This is a readability fix to existing columns, in the same class as
#2509, #2514 and #2516, so there is no demand-ledger entry.
## Verification
I ran a mock-mode private stack on main 800329cbd8, comparing before and
after at 12–18 widths from 390 to 1600px.
- **Visible text** (each character checked against its clipping box):
- Rows that read identically fall to 0 on every table at every width.
- Rows hiding the repository name fall to 0–1, and to 0 from 900px on
Images.
- **Clip sweep:**
- No column other than Image changes, and no metric sublabel is lost.
- Image cells with clipped text are never more numerous than on main at
any width, and fewer at most widths.
- Element counts rise only because a truncated reference now has two
clipping spans; the head truncates by design.
- **Clipboard:** a keyboard copy of one reference reads back
`ghcr.io/pulse-demo/backup-coordinator:2026.04` from
`navigator.clipboard`.
- **Interactions:** sort by Image, and row expand and collapse, work on
all three tables. No console errors.
- **Tests:**
- New tests: the split cases, a rendered Images-table test covering the
tail, the title and the copy scoping, and a source guardrail for the
three image cells.
- The new and updated tests fail against main's source.
- 544 test files pass across `src/features`, `src/hooks` and
`src/components/{shared,Workloads,Infrastructure}`.
The browser receipt is
`frontend-modern/browser-verification/a68fe9ea6a927e83.json`.
**Known residual:** when the head truncates completely there's no
leading ellipsis (as with `KubernetesNameText`). When it truncates
partly, the gap after its ellipsis can be up to one character wide.
KubernetesNameText renders a generated name as a truncating head and a
kept tail, and both are flex items. Flex items are blockified, so a
browser copy put a line break between them: copying a pod name gave
"cron-nightly-backfill\n-28918234", and pasting that into kubectl logs
broke. The same happened for the Pods table Node column and controller
names.
A copy whose selection sits inside one name now writes the name without
the break. Wider selections, such as a whole row, still copy as the
browser builds them.
Adding the Pulse Agent to a Proxmox cluster member meant generating a
token in the UI and pasting it into an interactive prompt on the host,
even though Pulse already knows the cluster and has an agent on its
siblings. The server-side cluster deploy that avoids this was still
shipped, but its wizard was deleted as an orphan in May, and it could
not reach members discovered by hostname: targets were parsed from the
API URL, so a name like "delly2" hit the agent's literal-IP guard and
failed preflight.
Deploy targets now resolve to a literal IP in the member's API URL or
its cluster endpoint's effective IP, matched on the member's own
connection, and members with no known address come back undeployable
with a reason. The Settings > Infrastructure install actions open one
dialog that checks, installs, and confirms each node is reporting,
falling back to the scoped installer when no sibling agent is
connected or a node cannot be reached.
The deploy routes are keyed on the cluster name, so they now refuse a
name shared by two connections instead of merging two sites, and job
creation admits one install per cluster at a time under a lock, which
lets the dialog keep a run going after it is closed.
The agent-lifecycle, api-contracts, frontend-primitives, and
storage-recovery contracts record the new install path, its address
rule, and the persisted, time-bounded admission check.
Supersedes #2507, which split this change across two commits; the
per-commit canonical completion guard needs the contract updates in the
same commit.
KubernetesNameText renders a generated name as a truncating head and a
kept tail, and both are flex items. Flex items are blockified, so a
browser copy put a line break between them: copying a pod name gave
"cron-nightly-backfill\n-28918234", and pasting that into kubectl logs
broke. The same happened for the Pods table Node column and controller
names.
A copy whose selection sits inside one name now writes the name without
the break. Wider selections, such as a whole row, still copy as the
browser builds them.
## Outcome
This batch makes guest identity persistence safer, keeps broadcast
resource rows and their History targets in one captured generation, and
makes completed-backup evidence accessible by keyboard and touch.
Existing help now explains container CPU scales and passive comparisons.
Native macOS guidance checks receive real GNU timeout without replacing
unrelated system utilities.
The work follows retained identity-write/shutdown failures, fleet CPU
and History consistency evidence, inaccessible backup disclosures and
confusion about container CPU readings. It improves those existing paths
without claiming field recovery or complete performance relief.
## Included reviewed candidates
The complete proposed range is
`800329cbd83c48fdb7402a18633189f2c8a2d6ff..9b057cb141221526a96fac05bdbdf8ed6e8625fc`:
- Guest metadata ordering/coalescing and safe retained shutdown:
`244d322dce` and `5e51069048`, integrated at `d7a305922a`.
- Native guidance prerequisites: `13f6d10798` and `443faa53b6`,
integrated at `f7fca0e164`.
- Accessible existing Workloads backup details: `02e6a90374`, integrated
at `8661c1fc65`.
- Paired complete resource/History capture and faithful full-content
controls: `54de7bdca7`, `32f09dc13d`, `8b12e0ee1b`, `d206b03990` and
`9cc8523881`, integrated at `2787b4eabd`.
- Container CPU documentation: `0c283d2dd5`, integrated at `15e8c5ccf6`.
- Router-owned limiter joins and sealed late worker admission:
`5b560521e2`.
Reviewed identities and authors are unchanged. Current upstream Docker
link editing and platform layout repairs are incorporated by an additive
full-SHA merge; earlier upstream governance and Kubernetes layout
history are retained. The later upstream Docker image badge sizing
repair at `800329cbd83c48fdb7402a18633189f2c8a2d6ff` is also retained
after one successful bounded refresh and additive merge.
## Behaviour and safeguards
Identity polls merge only monitor-owned fields, preserve operator
links/tags/notes and OCI classification, and do not spin-retry
filesystem failures. Incomplete tenant shutdown retains ownership and
prevents directory removal. Router shutdown joins its own limiters
without resetting budgets or stopping another Router.
Broadcast resources and target mappings are detached together. Health
comparison controls use one evaluation instant while production still
samples real UTC at the existing boundary. Full-content, overlay/clear,
identity and observation-time controls remain.
Backup disclosures preserve hover and add keyboard/touch access to the
same read-only evidence. CPU guidance distinguishes host-capacity
normalisation, VM CPU counts, version-qualified rounding, missing
readings and actual alert delivery; it asks for ordinary observations,
not forced workload or threshold changes.
## Validation and remaining limits
- Provenance, retained ancestry, clean additive merge and whitespace
checks pass.
- Every unpublished non-merge commit passes the canonical
completion/history guard. Final-range browser receipt coverage, public
registry audit and contract audit pass.
- Accepted candidate evidence includes focused writer/API races, startup
observations, vet and actual builds; 363 connected backup frontend tests
with four desktop/touch browser cases; CPU help's 132 connected frontend
tests, lint/types/build/budgets, focused Go checks and four Docs browser
cases. Browser receipts cover unchanged final backup bytes and are not
rerun merely because the batch contains several candidates.
- Gate frontend proof at `9daeed13`: 246 tests in four connected files,
type-check, all lint checks, actual embed build and bundle budgets pass;
complete output hash matches and guest stop completes.
- Exact refreshed head `9b057cb1` passes 49 Docker native-table and
platform-layout guardrail tests (`source-l8diazgn`); complete output and
guest stop are verified. This covers the later width-only upstream
change without repeating the earlier full frontend checks.
- Gate backend attempt `source-pz6qwk2t` fails in guest readiness after
capacity waiting, client exit 2, with no command verdict and confirmed
guest stop. It is not a test pass and is not replayed. All backend Go
source, manifests and native workflow bytes are unchanged from reviewed
`5b560521` and its focused races/vet/real-build evidence; that evidence
remains usable without inventing an aggregate pass.
- The retained broad API ten-minute timeout, earlier drain/startup
failures, adverse allocation/frame costs and failed stable RSS retest
are not erased by these focused results. Required exact-head CI remains
authoritative for landing.
- This is not native guest thaw/recovery, PBS badge agreement, CPU/RSS
relief, notification delivery or release inclusion. Independent thaw,
fresh writes on every covered filesystem, liveness and only-prior-active
restoration remain necessary before resuming uncertain monitoring.
Related to #2199
Related to #1723
Related to #2439
Related to #2447
[How Pulse handles
triage](https://github.com/rcourtman/Pulse/blob/main/docs/AI_TRANSPARENCY.md)
Image references that share a registry and namespace differ only at
the end, but table cells truncated them from the end. At 390 and 700px
three Images rows read "ghcr.io/pulse-demo…", Swarm services read
"registry.…" from 390 to 700px, and the Containers Image column hid
the repository name up to 1440px, so an update badge or a size could
not be traced to its image without opening the row. Real fleets are
worse: every lscr.io/linuxserver image shares a 20-character prefix.
splitDockerImageReference splits before the repository name (a digest
never moves the split), and DockerImageReferenceText renders the head
as the part that truncates and keeps name:tag visible, mirroring
KubernetesNameText, with the full reference on the title. The Images
name cell, the Containers Image column and the Services Image column
use it. The two halves are flex items, which a browser copy would
join with a line break, so a selection inside one reference copies
without it. Sorting, search and the stored value are unchanged, and a
reference without a path renders as before.
## Problem
Two platform tables cut a status badge at the cell edge at tablet
widths.
- **Docker containers, Updates column.** Columns are sized per layout
mode with weights that renormalize over whichever optional columns show
(State, Restarts, Host, Engine). Between a 34rem container and the md
breakpoint, where the phone wrap rule stops, Updates got 54–76px. So
"Update" (74px) and "Current" (69px) were cut from 573 to 800px, and
ungrouped up to 1100px.
- **Kubernetes Nodes, Status column.** The tablet band split equally
(51px of room at 573px), and md gave Status 8% (53px at 768px). So
"Degraded" (64px) was cut.
## Change
**Docker containers** (`dockerContainerTableModel.ts`): Updates takes
room only from the CPU bar, which never shows a sublabel, and the
collapsed actions menu. Every other column keeps its exact width in
every layout mode and optional-column set, so the memory bar keeps its
used/total sublabel.
- The 74px states (Update, Current, Pinned, Failed) fit at each layout's
narrowest table width in every column set.
- The 100px states (Check failed, Updating..., Completed) fit in most
layouts, but can still clip at a layout's narrow end when several
optional columns show.
- Wide rows also gain the 108px their expanded lifecycle controls need.
- A new model test checks every layout and column set.
**Kubernetes Nodes** (`KubernetesNodesTable.tsx`):
- The tablet band now sizes each column for its header, and Status for
Degraded, at a 34rem table.
- md moves 2.5 points from CPU (whose bar has no sublabel) to Status.
The visible total stays the same, so Node, Cluster, Roles, Memory and
Capacity keep their widths at every md width.
- Degraded, NotReady and Unknown fit from 573px. Unschedulable (89px)
fits from about 900px; main cut it at every width.
The guardrail test pins the new widths. Contract notes are in
`unified-resources.md` and `frontend-primitives.md`.
## Verification
Mock-mode private stack on main 800329cbd8, before against after, at
18–20 widths from 390 to 1600px:
| Table | Badge clipped (main) | Badge clipped (after) | Total clipped |
Columns worse | Sublabels lost |
|---|---|---|---|---|---|
| Docker containers, grouped | 573–900px | none | 89 → 53 | none | none
|
| Docker containers, ungrouped | 573–1100px | none | 195 → 132 | none |
none |
| Kubernetes Nodes | 573–800px | none | 85 → 78 | none | none |
- **Interactions:** sorting by Updates and by Status, and row expand and
collapse, work at 600, 768 and 1280px with no console errors.
- **Tests:** `src/features`, `src/hooks`, `src/components/shared` and
`src/components/Workloads` pass (489 files). The new and updated checks
fail against main's source.
**Known residuals:**
- "Review action" (112px, shown only while an update awaits review) can
still clip at tablet widths when the table is crowded.
- The Docker "RESTARTS" header was already clipped in the 480–768px
range on main and is unchanged.
The browser receipt is
`frontend-modern/browser-verification/6683727b493b9b45.json`.
The Docker container Updates column and the Kubernetes Nodes Status
column hold badges that cannot truncate, and both were cut at the
cell edge at tablet widths.
Docker containers size columns by layout mode with weights that
renormalize over whichever optional columns show (State, Restarts,
Host, Engine). Updates got 54-76px between a 34rem container and the
md breakpoint, where the phone wrap rule no longer applies, so Update
(74px) and Current (69px) were cut from 573 to 800px. The column now
takes its room only from the CPU bar, which never shows a sublabel,
and the collapsed actions menu, so every other column keeps its width
in every layout and column set and the memory bar keeps its used/total
sublabel. The 74px states fit in every layout and column set at the
layout's narrowest table. The 100px states (Check failed, Updating...,
Completed) fit in most layouts but can still clip at a layout's
narrow end when several optional columns show. Wide rows also gain the room their expanded
lifecycle controls need. A model test checks every layout and column
set.
Kubernetes Nodes Status shared the tablet band equally (51px of room
at 573px) and had 8 percent at md (53px at 768px), so Degraded (64px)
was cut. The band now sizes each column for its header and Degraded
at a 34rem table, and md moves 2.5 points from CPU to Status, keeping
the visible total, so Node, Cluster, Roles, Memory and Capacity keep
their widths everywhere. Degraded, NotReady and Unknown fit from
544px; Unschedulable (89px) fits from about 900px.
Incorporate the scoped-read upstream identity through the completed bounded refresh; preserve all reviewed commit identities and existing source evidence.
Change-source: pulse-maintainer
## Problem
On `/docker/images` the Update check badge was cut off at tablet widths.
Its longest labels ("Update available", "Managed by Pulse") need about
122px including cell padding.
- In the tablet band (a 34rem container below a 768px viewport) every
non-name column took an equal 17.5% share. That left Update check 100px
at 600px and 110px at 700px.
- The 16% md width left it 111px at 768px.
The badge was cut at the cell edge with a stray ellipsis dot, and the
"UPDATE CHECK" header clipped at 600px.
## Change
`DockerImagesTable.tsx` changes only column widths. The badge label is
unchanged.
| Column | Band (`bandWidth`) | md before → after |
|---|---|---|
| Image | 30 | 30 → 30 |
| Host | 17.5 | 18 → 17 |
| Used by | 17.5 | 24 → 23 |
| Size | 12 | 12 → 11 |
| Update check | 23 | 16 → 19 |
Host and Used by keep the band share they already had. Size gives up the
room in the band, because its widest values ("1023 MB", "9.99 GB",
49.6px) still fit its narrowest cell (53px). The md change takes one
point each from Host, Used by and Size.
`platformOverviewLayout.guardrails.test.ts` pins the new md and band
widths. Contract notes are in `unified-resources.md` (Docker images) and
`frontend-primitives.md` (sizing a badge column for its narrowest table
in each range).
## Verification
I ran a mock-mode private stack on main 8e68bab55a, which already
includes #2514's Used by span change.
- **Clip sweep:** 15 widths from 390 to 1440px, checking `scrollWidth >
clientWidth + 1` on every cell and its overflow-hidden descendants.
- The update badge clipped at 600, 640, 700 and 768px on main, and
nowhere after this change.
- Image, Host, Used by and Size have the same clip counts at every
width. Total clipped elements went from 65 to 61.
- **Label probe:** all six update labels cloned into the narrowest
Update check cell. On main, "Managed by Pulse" overflows from 600 to
800px and "Update available" at 600, 640, 700 and 768px. After this
change, every label fits at every width.
- **Headers:** no header clips at any width, including with a sort
indicator.
- **Interactions:** row expand and collapse works at 390, 600, 768 and
1440px. No console errors.
- **Tests:** `src/features/platformPage` and `src/features/docker` pass
(37 files, 653 tests). The updated guardrail test fails against the old
source.
The browser receipt is
`frontend-modern/browser-verification/7db5e194a586019d.json`.
The Images table's Update check badge needs about 122px with its cell
padding for its longest labels ("Update available", "Managed by
Pulse"). Between the phone projection and the md breakpoint every
non-name column took an equal 17.5 percent share, which left the
column 100px at a 600px viewport, and the 16 percent md width left it
111px at 768px. The badge was cut at the cell edge with a stray
ellipsis dot, and the "UPDATE CHECK" header clipped at 600px.
The table now declares tablet-band widths through the sortable
header's bandWidth (Update check 23, Size 12; Host and Used by keep
their 17.5) and moves 3 points of md width to Update check, one each
from Host, Used by and Size. The longest size values ("1023 MB",
"9.99 GB") still fit both narrower Size columns.
A clip sweep on a mock stack from 390 to 1440px (15 widths) found the
update badge clipped at 600, 640, 700 and 768px before and nowhere
after. A probe of all six labels against the narrowest Update check
cell found none clipping at any width, and no other column clipped
more often than before. The base already carries the Used by
block-truncate span from #2514, so Used by was measured as it now
renders.
Follow-up to rcourtman/Pulse#2477.
**Docker containers get the bulk editor**
The strongest demand for saved links (#1556, #624) came from Docker
users, but the editor only reached Proxmox and vSphere guests. The
Docker containers table header now has the same **Edit links** panel,
worded for containers.
- Rows are keyed through `toDiscoveryConfig`, the container drawer's own
metadata target, so bulk and drawer saves read and write the same
`app-container:<host>:name:<name>` record. This includes Docker running
inside Proxmox LXCs.
- The editor now takes a source. Workload tables pass live guests, so
the row model still loads only with the panel. Other tables pass
ready-made rows.
- The Docker table lazy-loads the trigger. Without that, Rollup merged
it into the chunk the Docker and workloads pages share, renamed that
chunk, and broke the `DrawerDiskListCard` bundle budget.
**Read-only tokens no longer see edits that would fail**
Metadata writes require `monitoring:write` for API-token sessions (the
backend's `EnsureScope`). Browser sessions always pass, because metadata
routes have no role check. A kiosk-style `monitoring:read` token saw
edit controls that every save refused with 403. The session store now
mirrors the gate from the security status `tokenScopes`
(`sessionCanWriteMonitoringMetadata()`):
- the bulk editor's **Edit links** button is hidden for such sessions;
- the shared web-interface field shows the saved URL with open and copy,
plus a one-line explanation, but no input, Save, Remove, or "Use this"
suggestion.
**Fewer refetches**
Saves now dispatch their metadata-changed events together after the
writes. `useUnifiedResources` coalesces onto one in-flight refetch, so a
bulk save costs two resource refetches whether it changes 2 links or 4.
Before this, it cost one per link.
**Verification**
- Private mock stack, Docker page and Proxmox page, at 1280 and 375px.
- Real token sessions with the admin bypass off: a `monitoring:read`
token sees no edit affordances, and a `monitoring:read+write` token
saves a container link.
- Receipt: `frontend-modern/browser-verification/e8946b639f30703c.json`.
- Production build: `WorkloadsSurface` 69.08 kB, `Docker` 28.56 kB,
`DrawerDiskListCard` unchanged; no violations.
- Full frontend lint, plus 841 test files (15,618 tests) across
features, hooks, components and stores.
Follow-up to #2509, which fixed the Kubernetes tables. The Docker,
Proxmox and vSphere tables still put summary values in a capped
`inline-block` span (`max-w-48 truncate`, `max-w-[18rem] truncate`, or
`truncate inline-block max-w-32`). These tables are `table-fixed` and
their `td` already truncates, so the cap ignored the column width:
- **Spill with no ellipsis.** Where the cap was wider than the cell, a
long value ran past the cell edge and was cut mid-glyph. At 700 wide: a
swarm service image by 225px, secret labels by 184px, image "Used by" by
168px, Mail Gateway domains by 71px. At 1440: volume mountpoints by
51px.
- **Off the row line.** An `inline-block` with hidden overflow aligns
its bottom edge to the baseline, so every one of these values sat
2–2.5px above the rest of the row.
A `block truncate` span fills the cell, ends in an ellipsis inside it,
and sits on the row line.
**Changed cells (19 spans in 11 files)**
- Docker: Configs (Created, Labels), Secrets (Created, Labels), Tasks
(Current), Volumes (Created, Mountpoint), Images (Used by), Networks
(Subnets), Services (Stack, Image, Update)
- Proxmox: backup coverage Node, restore-evidence Location and Details,
recoverable Location and Details, Mail Gateway drawer Top domains
- vSphere: Hosts vCenter
The Docker service **Update** cell still wraps on phones. From md it now
truncates inside its column instead of at a fixed 10rem cap.
The guardrail from #2509 (`keeps platform summary cells truncating
inside their own column`) now covers every table source that carried the
pattern. `ProxmoxCoverageTable.test.tsx` checks that a long node name
renders as the cell's only child, in a `block truncate` span, inside a
`table-fixed` table. Both tests fail against main's sources.
**Verification** (private mock stack, Playwright)
- **Live DOM, before switching anything.** At 700 and 1440, every one of
the 1,000 measured spans is the only child of its cell. Every table
computes `table-layout: fixed`, including the Mail Gateway drawer's Top
domains table.
- **After.** No span runs past its cell, all 134 truncated values end in
an ellipsis inside their own column, and row text sits on one line (0px
offset).
- **Clip sweep.** All eight affected routes (/docker/images,
/docker/networks, /docker/storage, /docker/swarm, /proxmox/backups/date,
/proxmox/backups/coverage, /proxmox/mail, /vmware/overview), 8 widths
from 390 to 1440. Clipped cells go from 931 to 928 out of 19,209
measured, and no column is worse at any width. The count barely moves
because a spilled value was already clipped by its cell. What changes is
a hard cut becoming an ellipsis.
- **Mock gap.** The mock Mail Gateway reports no domain stats, so the
drawer table was mounted by adding eight domains to the real
`/api/pmg/instances` response through a route intercept.
- **Tests.** Docker, Proxmox, vSphere, platformPage and app-level
suites: 79 files, 1,549 tests pass.
- **Receipt.** The browser receipt is included.
The bulk web-link editor only reached Proxmox and vSphere guests, but the
strongest demand for saved links (#1556, #624) came from Docker users. The
Docker containers table now offers the same Edit links panel, keyed through
the container drawer's own metadata target so bulk and drawer saves share
one record. The trigger is lazy-loaded so it does not merge into the chunk
the Docker and workloads pages already share.
Metadata writes require monitoring:write for API-token sessions, so a
kiosk-style read-only token saw edit controls that every save refused. The
session store now mirrors that gate from the security status token scopes:
the bulk editor is hidden and the shared web-interface field shows the saved
URL without input, Save, Remove, or suggestion adoption. Browser sessions
are unaffected because the backend does not check roles on metadata writes.
Saves now dispatch their metadata-changed events together after the writes,
so a bulk save costs two resource refetches instead of one per link.
Follow-up to #2509, which fixed the Kubernetes tables. Docker, Proxmox
and vSphere summary cells still wrapped long values in a capped
inline-block span (max-w-48 truncate, max-w-[18rem] truncate, or
truncate inline-block). The cap ignored the column width of these
fixed-layout tables, so a long value ran past a narrower cell with no
ellipsis (a swarm service image by 225px at 700 wide, volume mountpoints
by 51px at 1440), and the inline-block sat about 2px above the row's
text line.
A block truncate span fills the cell, ends in an ellipsis inside it and
sits on the row line. The Docker service Update cell keeps its phone
wrap and truncates inside its column from md instead of at a 10rem cap.
The guardrail from #2509 now covers every platform table source that
carried the pattern, and the coverage table test pins the Node cell.
## Problem
The registry audit requires `require_explicit_path_policy_coverage:
true` on every subsystem. Under that setting,
`build_verification_requirements` never uses a rule's default
verification policy: a runtime file that no path policy matches becomes
a `missing-path-policy-coverage` gap. As a result, nothing ever consults
the rule-level `verification.test_prefixes` and
`verification.exact_files` lists. Their only readers were the audit's
format and path-existence checks, plus one test that pinned monitoring's
list as data.
Even so, 13 of the 16 rules carried them: 93 files and one prefix.
Authors kept adding proofs there as if they counted. relay-runtime alone
listed 20, and performance-and-scalability listed 30. This is the same
misleading-config problem as `allow_same_subsystem_tests`, which #2505
retired, one level up.
## What happens to the 94 entries
- 92 of the 93 files are already proof on a path policy: 77 on their own
subsystem's policies and 15 on another subsystem's.
- api-contracts' default prefix `frontend-modern/src/api/__tests__/` is
already on three of its own policies.
- The one orphan is `internal/api/router_bench_test.go`. It benchmarks
the api-contracts router, not any performance-and-scalability runtime
file, so no policy should take it. It stays an owned file of
performance-and-scalability.
Nothing was folded into a policy, because folding would loosen the
guard.
## Change
- `registry.json`: empties the 13 rules' default `test_prefixes` and
`exact_files`. The keys stay, since the schema requires them.
- `registry_audit.py`: requires both lists to be empty, and points
authors at the path policy whose runtime files the proof covers. This
replaces the per-entry format and path checks, which only ever validated
lists nothing reads.
- `CANONICAL_DEVELOPMENT_PROTOCOL.md`: states the empty-default rule. It
also drops the promise that same-subsystem tests count "when the
registry explicitly allows them", which the audit no longer permits.
- Audit test fixtures now model valid registries, with empty defaults
throughout. A new test covers the rejection.
**No verdict changes.** I compared `build_verification_requirements`
under the old and new registry. It returns identical requirements for
all 2,746 owned runtime paths, and for an uncovered path in every
subsystem.
## Verification
- `go test ./internal/repoctl`: passes.
- The Python suites (registry_audit, canonical_completion_guard,
status_audit, subsystem_lookup, staged_commit_shape_guard,
dev_runtime_governance, contract_audit, control_plane_audit): 427 tests
pass.
- New `test_audit_registry_payload_rejects_default_verification_proofs`
fails against the previous audit.
- `registry_audit.py --check` on the live registry: no errors.
- Open PR #2508 adds a path policy and owned files only, with no default
proofs, so it is unaffected.
Every subsystem must declare require_explicit_path_policy_coverage, so
build_verification_requirements never falls back to a rule's default
verification policy; an unmatched runtime file becomes a path-policy gap
instead. The default test_prefixes and exact_files were therefore never
consulted, yet 13 of 16 rules carried them (93 files and one prefix),
and authors kept adding proofs there as if they counted.
92 of the 93 files were already proof on some path policy, and
api-contracts' default prefix is on three of its own policies. The one
orphan, internal/api/router_bench_test.go, benchmarks the api-contracts
router rather than any performance-and-scalability runtime, so no policy
should take it; it stays an owned file only.
The registry drops the lists, registry_audit requires them empty and
points authors at the path policy instead, and the canonical protocol
says so. The protocol also stops promising same-subsystem tests "when
the registry explicitly allows them", which the audit no longer permits.
No verdict changes: build_verification_requirements returns identical
requirements for all 2746 owned runtime paths, and for an uncovered
path, under the old and new registry.
Capture all endpoint limiters under their Router owner and seal late worker
admission. Concurrent shutdown joins each cleanup worker without resetting
request budgets or stopping another Router. Integration fixtures use the real
shutdown lifecycle before monitor teardown.
Contract-Neutral: Router worker ownership and concurrent shutdown only; HTTP payloads, limits and windows, auth, tenant selection, revocation and persistence contracts are unchanged.
Change-source: pulse-maintainer
Accept exact PI commit 0c283d2dd5 with retained focused source and Docs browser proof. Preserve conditional runtime-host examples and passive comparison safeguards; this is documentation maintenance, not native collection or alert-delivery acceptance.
Change-source: pulse-maintainer
Follow-up to #2506. The Kubernetes Configuration and Deployments tables
still clipped names and values between phone and desktop widths.
- **Tablet band widths.** Between the phone projection and the md column
widths (a 34rem container below a 768px viewport) these tables split
their width evenly, which cut names to a few letters. Headers can now
declare a band percent through `bandWidth` /
`platformTableBandWidthStyle`, and both tables give names the most room.
- **Scope shows the namespace.** With one cluster in view, every scope
cell repeated the cluster name ("Production EU/ser…"). The cell now
shows the namespace (`Cluster` for cluster-scoped objects) and keeps the
full scope on the title and for screen readers. Several clusters keep
the combined label.
- **Deployments.** Desired appears only on wide tables, because Ready
already reads `2/2`. Ready keeps its denominator until Desired is shown.
On a phone, Age takes Ready's spare width so `18h ago` fits.
- **Summary cells.** Values sat in a capped `inline-block` span that
ignored the column width. Long values spilled past narrow cells with no
ellipsis and sat about 4px above the row. A `block truncate` span fills
the cell, ends in an ellipsis inside it, and lines up with the rest of
the row.
Verification:
- Clip sweep of every Kubernetes tab at 600–1440px, compared with main:
clipped cells drop from 566 to 356, and no column is worse at any width.
- Deployment, Resource and Namespace clips on Workloads and
Configuration drop from 25 to 0, and Scope clips from 136 to 7.
- Row text alignment is 0px spread on every tab.
- The browser receipt is included.
The five predeclared diagnostic cases retain a real 1m versus 2m health-age counterexample between separate wall-clock evaluations. Give the independent content oracle and production projection the same private health clock without removing any JSON field; pin minute-boundary age advancement and original LastSeen. Production samples real UTC at the unchanged post-read evaluation point.
Contract-Neutral: Internal health-clock test seam preserves the real UTC evaluation point, public resource/history fields, source precedence, agent identity/lifecycle, tenant/revocation and safety contracts; no wire or permission delta.
Change-source: pulse-maintainer
Document host-capacity normalisation, version-qualified rounding and the difference between collected readings and delivered alerts. Keep comparisons on ordinary workload observations rather than forced load, threshold changes or diagnostic reruns.
Contract-Neutral: Existing container CPU semantics and safety guidance only; no runtime, API, threshold, permission or lifecycle contract changes.
Change-source: pulse-maintainer
Preserve the exact Web candidate and its own-parent browser receipt. Retained connected source checks and desktop/touch evidence cover the existing read-only disclosures without changing collection or backup operations.
Change-source: pulse-maintainer
Between the phone projection and the md column widths (a 34rem container
below a 768px viewport) the Configuration and Deployments tables split
their width evenly, so names were cut to a few letters beside short
counts. Headers can now declare a percent width for that band through
bandWidth / platformTableBandWidthStyle, and both tables size names
first.
With one cluster in view every scope cell repeated the cluster name and
crowded out the namespace that tells rows apart, so scope cells show the
namespace alone and keep the full cluster/namespace scope on the title
and for assistive technology. Several clusters keep the combined label.
Deployments Desired returns only with wide tables, since Ready already
reads ready over desired, and Ready keeps its denominator until Desired
is visible. On a phone Age takes Ready's spare width so "18h ago" fits.
Summary cells used a capped inline-block span that ignored the column
width: long values spilled past narrow cells with no ellipsis and sat
above the row's text line. A block span fills the cell, ellipsizes there
and aligns with its neighbours.
The new complete-content oracle exposed a typed-view discrepancy. Keep that failed result and report the first differing JSON field, rather than accept an undifferentiated mismatch or normalise away evidence. Only the synthetic fixture failure message changes; equality assertions and production source are unchanged.
Change-source: pulse-maintainer
Adding the Pulse Agent to a Proxmox cluster member meant generating a
token in the UI and pasting it into an interactive prompt on the host,
even though Pulse already knows the cluster and has an agent on its
siblings. The server-side cluster deploy that avoids this was still
shipped, but its wizard was deleted as an orphan in May, and it could
not reach members discovered by hostname: targets were parsed from the
API URL, so a name like "delly2" hit the agent's literal-IP guard and
failed preflight.
Deploy targets now resolve to a literal IP in the member's API URL or
its cluster endpoint's effective IP, matched on the member's own
connection, and members with no known address come back undeployable
with a reason. The Settings > Infrastructure install actions open one
dialog that checks, installs, and confirms each node is reporting,
falling back to the scoped installer when no sibling agent is
connected or a node cannot be reached.
The deploy routes are keyed on the cluster name, so they now refuse a
name shared by two connections instead of merging two sites, and job
creation admits one install per cluster at a time under a lock, which
lets the dialog keep a run going after it is closed.
The agent-lifecycle, api-contracts, frontend-primitives, and
storage-recovery contracts record the new install path, its address
rule, and the persisted, time-bounded admission check.
Use native read-only disclosure buttons for both backup entry points, with the shared Dialog preserving named completion evidence, focus and guest row action ownership. Keep existing freshness and uncertainty semantics, collection and backup operations unchanged. Include connected regression tests, the subsystem contract and inspected desktop/phone browser verification.
Change-source: pulse-maintainer
The full current-state regression now requires a critical alert to reach the same freshly projected row. Continuity controls compare a healthy VM, critical transition and clear instead of using an already-offline host as a vacuous health control.
Change-source: pulse-maintainer
The independent full-state oracle retains the parent unified view watermark. The paired benchmark preprocesses active health alerts once per pass, like the exact parent, rather than adding per-row work to the baseline. Production projection and source fields are unchanged.
Change-source: pulse-maintainer
Resolve canonical source mappings once per complete projection while preserving detached resource content and target selection after host coalescing. Bind resources and targets to the same locked generation, including absence and continuity overlays, and avoid another owned outer-slice copy. No timestamp cache or native CPU/RSS claim. Retain cold, clean and dirty full-content controls and paired complete-projection measurements.
Contract-Neutral: Internal bulk-read and copy-ownership repair preserves existing public resource/history fields, source precedence, agent identity/lifecycle, tenant/revocation and safety contracts; no wire or permission delta.
Change-source: pulse-maintainer
Kubernetes names clipped badly between about 600 and 900px wide. At
700px node names read "Pr…" and deployment names two letters.
**Why:** between a 544px table container and a 768px viewport, platform
tables show their phone-hidden columns again, but `md:w-[…]` widths are
viewport media queries and don't apply yet. So every visible column took
an equal share. A seven-column Kubernetes table left its name 83px, with
the chevron and status dot inside it. From 768px the Kubernetes name
widths (14–19%) still cut common node, service and event names.
**Fix:**
- The shared table shell keeps the phone 30% name anchor in that band.
Tables sized by a weighted colgroup are unaffected.
- Kubernetes name columns get room for typical names at 768px. The Nodes
columns are rebalanced so Capacity fits at 1100px and wider.
- Generated names (pods, a pod's node, controllers) truncate their head
and keep the trailing segment that tells siblings apart:
`checkout-a…-c7z2p`, `prod…-02`. The tail never outgrows its cell.
Deployment names keep plain end truncation.
**Measured** with a clip sweep over every Kubernetes tab at 600–1440px:
- Identity-column clips fall from 284 to 186. Most of the remainder are
pod heads truncated by design.
- The Nodes table has no name clipping at any width.
- In the band, TrueNAS Storage and a vSphere resource table go from 14
name clips to 0, with no new clips on other platforms.
## Problem
`check_staged_contracts` (and `canonical_commit_shape_errors`) passed
the impacted entry from `infer_impacted_subsystems` as the `rule`
argument of `staged_verification_files_for_requirement`. The
`allow_same_subsystem_tests` branch matches through
`subsystem_matches_path(rule, path)`, which reads `owned_prefixes` and
`owned_files`, and the entry carried neither. So since the v6 import
(778a2577b6) a same-subsystem test never satisfied a requirement.
Meanwhile `subsystem_lookup.py`, which passes the real registry rule,
and the guard's own BLOCKED text ("same-subsystem test/spec files are
also accepted") both said it would.
## Decision: keep the strictness, don't honour the flag
Honouring the 12 registry flags would have loosened the guard far beyond
what the flags appear to have been meant for:
- All 16 rules set `require_explicit_path_policy_coverage`
(`registry_audit` makes it mandatory). The default policy is never
consulted, so the 5 rule-level flags (alerts, cloud-paid, monitoring,
relay-runtime, unified-resources) are dead either way.
- On the 7 path policies, the flag accepts any test the subsystem owns,
not just tests near the policy's code. Proof files each policy would
newly accept on `main`:
| policy | accepted today | newly accepted |
|---|---|---|
| ai-runtime/canonical-mutation-registry | 4 | 455 (any `internal/ai`,
AI Chat or `internal/agentcapabilities` test) |
| ai-runtime/patrol-qualification | 27 | 433 |
| unified-resources/product-trust-actions-frontend | 16 | 128 (Go tests
in `internal/unifiedresources`) |
| agent-lifecycle/action-runner-runtime | 4 | 119 (mostly
`internal/hostagent`) |
| agent-lifecycle/collector-lifecycle-installer-transport | 7 | 118 |
| agent-lifecycle/agent-privilege-helper-runtime | 8 | 115 |
| security-privacy/audit-log-runtime-store | 10 | 7 (`pkg/audit` tests)
|
Six of the seven already list their own directory in `test_prefixes`, so
the flag only added reach across the subsystem. That reach would also
grow whenever the subsystem gained ownership, without anyone editing the
policy.
## Change
- `registry.json`: sets the 12 flags to `false`. This is what the guards
have actually enforced.
- `registry_audit.py`: rejects `true` at both the rule level and the
path-policy level, and tells the author to list proofs in
`test_prefixes` or `exact_files`.
- `canonical_completion_guard.py`: the impacted entry now carries the
rule's `owned_prefixes` and `owned_files`. The completion guard, the
commit-shape guard, `subsystem_lookup.py` and the BLOCKED message now
read the flag the same way.
**Enforcement does not loosen.** No live requirement accepts a proof it
rejected before. The visible change is in `subsystem_lookup.py`: it
stops reporting matches the guard refused. For example, a new
`internal/hostagent/*_test.go` used to claim three agent-lifecycle
policies and now claims none.
`security-privacy/audit-log-runtime-store` is the one policy where the
flag was the only directory-level proof route. It keeps its 12 exact
files. If `pkg/audit` tests should count, the explicit form is
`"test_prefixes": ["pkg/audit/"]`.
## Verification
- `python3 -m unittest canonical_completion_guard_test
registry_audit_test staged_commit_shape_guard_test subsystem_lookup_test
status_audit_test dev_runtime_governance_test`: 400 tests pass.
- New
`test_check_staged_contracts_reads_same_subsystem_tests_from_ownership`
runs `check_staged_contracts` end to end with a synthetic flagged
policy. Against the unfixed guard it fails (`1 != 0`).
- New
`test_audit_registry_payload_rejects_same_subsystem_tests_wildcard`
covers both rejection sites.
- `registry_audit.py --check` on the live registry: no errors.
Between a 544px table container and a 768px viewport, platform tables
show their phone-hidden columns again, but md:w widths are viewport media
queries and do not apply yet, so every visible column took an equal share.
A seven-column Kubernetes table left its name 83px, with the chevron and
status dot inside it, so node names read "Pr..." and deployment names two
letters. From 768px the Kubernetes name widths (14 to 19 percent) still
cut common node, service and event names.
The shared table shell now keeps the phone 30 percent name anchor in that
band. Tables sized by a weighted colgroup are unaffected. Kubernetes name
columns get room for typical names at 768px, with the Nodes columns
rebalanced so Capacity fits at 1100px and wider. Generated names (pods,
the node a pod runs on, controllers) truncate their head and keep the
trailing segment that tells siblings apart, and that tail never outgrows
its cell.
staged_verification_files_for_requirement() matches
allow_same_subsystem_tests against a subsystem's
owned_prefixes/owned_files, but both commit guards passed it the
impacted entry, which carried neither. Since the v6 import (778a2577b6)
a same-subsystem test never satisfied a requirement, while
subsystem_lookup.py (which passes the registry rule) and the guard's own
BLOCKED message ("same-subsystem test/spec files are also accepted")
said it would.
Honouring the 12 flags would have loosened the guard well past their
apparent intent. Every rule requires explicit path-policy coverage, so
the five rule-level flags are never consulted, and on the seven path
policies the flag reaches every test the subsystem owns: a
mutation-registry change would accept any internal/ai or AI Chat test
(~455) instead of its 4 listed proofs, and a frontend Actions change any
Go test under internal/unifiedresources. Six of the seven already list
their own directory in test_prefixes.
So the strictness stays and the tools now agree on it: the registry sets
the 12 flags to false (what has always been enforced), registry_audit
rejects true and points at test_prefixes/exact_files, and the impacted
entry carries the rule's ownership so the completion and commit-shape
guards read the flag the same way subsystem_lookup does.
No requirement accepts a proof it rejected before. subsystem_lookup
stops reporting matches the guard refused: a new internal/hostagent test
no longer claims three agent-lifecycle policies.
TestSSEBroadcaster_AddClientWritesCurrentStatusBeforeReturning failed
under -race in CI on unrelated frontend PRs (2479, 2485). The test
broadcast a stage and connected straight away. When the broadcast loop
picked that message up after the client had registered, it queued the
stage for the new client too, so the client's writer goroutine wrote
the httptest recorder while the test read it without a lock.
AddClient itself is correct: it writes the preamble and current status
before returning, and a stage racing the connect can reach the client
twice, which the update modal already ignores. The test now lets the
broadcast reach an earlier client before connecting, so nothing else
writes the recorder and the status in the body can only come from the
synchronous snapshot write.
Reproduced by injecting the CI schedule through -overlay (broadcast
loop held until the client registers, caller paused as AddClient
returns): the old test failed 199/200 runs with DATA RACE, the new one
0/200. A mutant that moves the snapshot write onto the writer goroutine
still fails the new test 100/100.
Follow-up to #2502. CI now compares contract text against the commit
being checked, but the guard still decided which contracts that commit
owed using the subsystem registry and contract index from `HEAD`. In CI,
HEAD is the branch tip (or `main` after a merge), so a later commit
could still change an earlier commit's verdict:
- **False block.** A later commit adds a `Shared Boundaries` reference
to a file. An earlier commit that changed that file was complete when
written, but is now told it is missing the referencing contract.
- **False pass.** A later commit drops ownership of a file from the
registry. An earlier commit that changed that file without its contract
now passes.
The pre-commit hook reads both from the commit's own index, so CI and
the hook could still disagree.
**Change**
- `repo_file_io.read_repo_text` / `load_repo_json` and
`subsystem_contracts.tracked_contract_paths` / `tracked_contract_files`
/ `load_contract_index` accept a `revision`, alongside the existing
working-tree and index reads.
- A revision read uses only that commit's tree. It fails closed and
never falls back to the disk.
- In diff-base mode, the guard reads the registry and contract index
from the evaluated commit. So does the Prettier-only inference that
`--commit` enables.
- Pre-commit mode still reads the index, unchanged.
**Tests**
- `EvaluatedCommitGovernanceSourceTest` runs the real guard against a
temporary repo with its own registry and contracts, and covers both
cases above. Both fail against current `main`.
- New revision-read tests for `repo_file_io` and `subsystem_contracts`.
**Replay.** I replayed CI's per-commit loop (history registry,
Contract-Neutral trailers, `--diff-base <c>^ --commit <c>`) over the
last 120 non-merge commits on `main`, with the current and the new
guard. Verdicts are identical, and every historical registry loaded. The
PR #2494 commits still pass with HEAD at that PR's tip.
The maintainer preflight and coordinator already pass `--commit`
(rcourtman/pulse-dev-infra#704), so they pick this up without changes.
The comment in AddClient said a broadcast racing a connect is either in
the snapshot or queued behind it. It can be both, so the client then
receives that status twice, possibly with an older in-flight stage in
between. Trusting the "either" is what led the connect-snapshot test to
assume nothing writes the recorder after AddClient returns, which made
it flaky under -race. Say what actually happens and why it is harmless:
stage progress only rises and the update modal ignores backwards stages.
Contract-Neutral: Comment-only correction in internal/updates/sse.go; no runtime or contract change.
The completion guard's CI mode now compares contract text against the
evaluated commit (#2502), but it still decided which contracts a commit
owed from HEAD's registry and contract index. CI walks a branch with HEAD
at the tip (or at main after a merge), so a later commit could change an
earlier commit's verdict. If a later commit added a Shared Boundaries
reference to a file, an earlier commit that was complete when written got
blocked. If a later commit dropped ownership of a file, an earlier commit
that changed it without its contract passed. The pre-commit hook read
both from the commit's own index, so CI and the hook could still
disagree.
repo_file_io and subsystem_contracts can now read from a revision as well
as the working tree or index. A revision read uses only that commit's
tree and fails closed rather than falling back to the disk. In diff-base
mode the guard reads the registry and contract index from the evaluated
commit, as does the Prettier-only inference that --commit enables.
Replaying CI's per-commit loop over the last 120 commits on main gives
the same verdicts as before.
rcourtman