* Proofs build the PR's own server: warm workspaces for every review
The review guest carries GHC with the server's compiled dependencies, Rust and Go. The worker's new
prepare script builds each pull request's base offline in its workspace checkpoint, so a proof
compiles only the change: monoscope-app --server rebuilds the server after a Lucid page edit in about
90 seconds, where the base build took 23 minutes. Workspace builds and checks get 4 CPUs and 6 GB.
* Guest: the review account writes cabal's build log
* Server proofs: room to build, toolchains on PATH, no stray symlink, honest records
Check sides get 300 s so a proof can build the head; the guest links GHC, cabal, cargo and go into
the reset environment's PATH; the root node_modules link is ignored; checkpoint manifests record
their real VM size; the guest's dependency stage builds from cabal files alone; Go builds no longer
rewrite go.mod; the worker pins 4 agent, 2 check and 1 build slots.
* Guest: build the shared library's source with the server's dependencies
* Guest: the server's dependencies include the local CLI package
Nothing parsed the workflows (#859 broke Deploy with an expression no check read), and CI
regenerates monoscope.cabal before building, so a stale committed file (#858) passed every gate.
A second delivery of @monoscope prove it re-queued a completed review. Its fresh execution found
no saved session and the finished review's model calls, and failed as a lost session; every
retry then landed on a zero-slot node that re-deferred it as a new job with fresh odd-jobs
attempts, so it failed 54 times. The session guard now applies only to continuations, a lost
session ends the chain, and agent retries stop after three executions in one chain.
Publication recovery reuses the stored result and keeps retrying until it reconciles.
make ship now rebuilds the proof worker from the deployed production image (only its own
layers build) and restarts it on IMPACT_WORKER_HOST; make deploy-worker does it alone.
Slack, web conversations, routines and native review/fix jobs share the capability catalog, current authority checks, run ledger and operation receipts. Runs, charts and saved reports open inside the conversation inspector.
Consolidates duplicate regression suites and documentation while retaining the product contracts, native recovery evidence and migration baseline transition.
Validation: all eleven exact-input local checks pass, including 1,216 integration examples (25 pending VM cases), 2,105 doctests, 315 unit examples, 16 CLI examples, 1,041 UI tests and 172 browser tests (one skip). Separate native/KVM acceptance is recorded in the review. The exact Linux AMD64 image was built locally on this Mac with OrbStack/Rosetta.
* Share local CI dependency caches across checkouts
Each checkout path got its own cabal/npm/playwright volume set, so every
worktree and signoff clone leaked ~8GB of content-keyed caches; 35 sets had
piled up. Share those. dist-newstyle stays per checkout: sharing it was tried
and weeder then reported live functions as dead, because unchanged modules
lost their .hie files after another branch built into the volume.
* Note that local sign-off must stay fast
* Share dist-newstyle across checkouts; weeder repairs missing .hie files
After another branch built into a shared dist, GHC can leave a live module
compiled without its .hie, so weeder misses every reference it makes and
reports live functions dead. The weeder check now recompiles any tracked
module missing one (dropping cabal's build stamp, since a touch leaves the
source hash unchanged) instead of every checkout paying a cold build.
* Skip attested checks before starting local CI services
* Overlap integration and browser checks in local signoff
* Bring full local CI signoff under four minutes
* Keep browser checks responsive under full signoff load
* Run built test binaries during local signoff
* Trim redundant signoff work and stabilize browser readiness
* Fix RUM sessions, cache concurrency, and infrastructure drilldowns
* Exercise RUM guards against native cache and independent fixtures
* Preserve session list focus and scroll on live refresh
* Cache explicit empty RUM searches briefly
* Verify fresh empty RUM searches reuse cached results
* Fix RUM page links and preserve Explorer regex queries
* Preserve container identity through inventory and drilldowns
Keep namespace, cluster, node and typed scope through inventory grouping, drawer selection, charts and log links. Preserve time ranges and show recovery for ambiguous legacy links.
Regression reproduced collapsed rows and cross-container chart values before fixing. All 13 Containers examples pass on PostgreSQL and fresh native TimeFusion; browser guards pass at 390 and 1280px. Native chart tests use a bracketed explicit test pool because the existing test harness keeps its postgres-simple TF pool on PostgreSQL.
* Preserve RUM session search URLs and service scope
* Address cache review and verify CPU charts through the selected backend
* Revalidate stale RUM session lists inside their swap target
* Expose container identity in narrow table rows
Show full pod, namespace, cluster and node labels beneath container names when the corresponding columns collapse. Wrap UUIDs while preserving the existing desktop columns.
Retained e2e regression covers five same-name identities, including replicas on the same node. Missing namespace and Pod checks failed before the fixes; normal deferred navigation, full metadata visibility and viewport fit now pass at 390 and 1280px. All 13 focused Containers examples pass; formatter, Hlint and review loops are clean.
* Count distinct RUM sessions across the full time window
* Use exposed Servant API parser in container regression
* Preserve native JSON predicates and invalidate stale compiled filters
* Keep loaded RUM summary values visible during live refresh
* Keep Kubernetes container drilldowns within their cluster
* Keep error lifecycle fixture clocks and job stages ordered
* Keep session form windows aligned with the time picker
* Keep container inventory focused on finding entities
* fix(rum): refresh panels with the current picker window
* fix(rum): replace initial panel requests when the window changes
* test(kql): round-trip escaped quoted regex patterns
* Show singleton observations in line charts
* Preserve inventory facets when changing filters
* Keep inventory search accessible on mobile
* Compute RUM web vitals from observation populations
Share scalar and explicit-histogram populations across field, trend and exact URL views. Preserve cumulative baselines, coverage, scalar atoms and metric siblings when capping pages.
Validated: native handler 9/0, PostgreSQL RUM 28/0, formatting and HLint, two review rounds. Styled browser proof remains pending after capture harness timeouts; local doctests failed before examples because the monoscope package was not registered. Combined CI must verify codecs and integrated behavior.
* Keep primary container CPU visible with mobile filters
* Fix mixed vital atoms above finite histogram bounds
* Keep vital trend axes and coverage accurate
* Accept optional starts for delta vital populations
* Keep collector container CPU visible on mobile
* Preserve explicit RUM windows when since is absent
* Reuse RUM cache entries for blank window fields
* Match container summary to its selected time slice
* Use singular labels for individual vital observations
* Reuse RUM populations across equivalent explicit window requests
* Keep RUM vital observations and coverage visible on phones
* Keep partial vital trend buckets inside the selected axis
* test(rum): check singleton point within short time axis
Extend the existing axis browser guard with a fixed 30-second window and actual plotted point containment, scrolling the lazy chart into view. Browser proof remains pending authoritative CI: isolated WAI attempts failed before the existing 1H axis check, so they do not reproduce clipping. The b4 handler regression remains verified.
* Refresh container inventory when its time window changes
* fix(time-picker): apply automatic refresh menu choices
Use native querySelector in the existing hyperscript click handler; keep the hx-live proxy bindings unchanged. The browser regression reproduces Off retaining interval 15000, then verifies default, Off, 30s and 15s states and selection at mobile and desktop widths after the fix. Focused browser: 1 passed (4.0s); formatting and lint passed. Full CI remains pending parent verification.
* Load RUM panels before application globals initialize
* Refresh infrastructure inventory when its time window changes
* Keep page vital measurements readable on mobile
* Keep infrastructure refresh hyperscript covered by parser checks
* Preserve refreshed chart registry through HTMX settling
* Preserve captured error identity in regression lifecycle fixture
* Keep initial RUM requests alive during automatic ticks
* Await infrastructure fixture visibility through shared snapshot cache
* Preserve pending RUM updates and remove redundant prewarming
* Preserve ownership of replacement session searches
* Avoid a separate epoch read for noncumulative RUM vitals
* Scope RUM query-count guard to its handler
* Load selected session details independently of the list
* Send RUM summary widgets as SQL so TimeFusion serves them from rollups
The four summary stats and the activity chart now send the exact
browserScope, pageViewPredicate and errorPredicate SQL that TimeFusion's
rum_* rollup measures declare. KQL's lowering (startswith as a regex,
exception.type != null as a JSON search) could never match them. The chart
becomes one aggregate unpivoted to two series; a page view that errored
now counts in both.
* RUM improvements
* Auto-format code with fourmolu
* Simplify RUM population queries and type query failures
* Harden RUM vital SQL and consolidate regex parsing
* improve signoff
* Consolidate repeated RUM, infrastructure and test logic
- Derive SQL encoding for VitalCoverage/VitalReadKind instead of 12 label bindings
- panelSwapAttrs owns the panel URL and hx-sync; share timeWindowVals_ with Components
- One kqlFilter plus node/cluster fields for container, host and Kubernetes pivots
- linkParams/renderedWidget test helpers; shared histogram and panel-loading builders
* Cache npm downloads and e2e node_modules in local CI
* Auto-format code with fourmolu
* RUM sessions list: paint the newest slice first and bound the follow-up scans
A cold session list over a wide window scanned the whole window three times in sequence
(core, enrichment by id, recording cross-lookup); the panel took 22-97s on a busy project.
Wide cold lists now paint the newest three hours (1.7s) and revalidate the full range
through the stale-serve path; the follow-up scans are bounded to the rows already found;
the recording cross-lookup runs only for text searches.
* RUM: morph live-tick panel refreshes, keep the pulse widgets across revalidation, one empty state for missing Web Vitals
Every panel replaced itself with outerHTML on the 15s tick, and a stale pulse panel re-rendered
its four self-fetching widgets on revalidation — a 2-10s spinner flash on first load. Rendered
panels now morph; a populated pulse skips revalidation. A project that reports no Web Vitals
gets one explanation on the Performance tab instead of five No-data rows and two empty panels.
* Infrastructure snapshot: reduce to the newest datapoint per series before the blob lookups
The pivot evaluated ~20 resource/attributes JSON lookups on every datapoint in the 15-minute
slice (49.9K rows on the demo project) although only the newest row per series survives. A
first window over series_id keeps those survivors, so the lookups run over ~750 rows:
projection 1.37s -> 40ms, cold infra pages 2.4-3.5s -> 0.85-0.92s.
* Charts: update a live echarts instance in place instead of dispose+init; stop the 401 reload loop
A morphed panel keeps its chart element, but every re-init disposed the instance and drew it
again, which blanked all five Web Vitals charts on each 15s tick and cleared the stat tiles'
values while their first fetch was in flight. chartWidget now updates a live instance in place
(hx-morph-skip on the slot), the server-rendered stat value stays until the response replaces
it, and a 401 from /chart_data is verified once before reloading, at most once a minute,
instead of reloading the page on every occurrence.
* RUM: local times, a second signal on rated numbers, family hues off the status hues, widen action on the empty vitals state
Top pages and browser errors printed UTC beside a header showing the viewer's zone; they render
through local-time now. P75 cells carried their rating by colour only and get a dot plus a
screen-reader label. Audience chips used the warning, success and error hues for Linux, Android
and Firefox. The sessions workspace gives the last-page column room to show the distinguishing
tail. The empty Web Vitals state's primary action now does what its copy advises.
* RUM: never cache the newest-3h slice; clamp span pads to the window; release charts a morph left detached
A cached slice would be served as the final list whenever the full revalidation failed. The
recording pad could reach outside the selected range. Charts kept alive across a morph were
never disposed when the fragment that held them left with a tab switch.
* RUM: spanScope doctest builds its scope from the module's own imports
* RUM: show the refreshing hint at the top of a provisional session list
* e2e: a refreshed vital chart keeps its instance on its element
* Infrastructure pages: denser host list, honest drawer empty state, one live region per table, status color only for status
Integration chips no longer wrap into 65px rows or clip the last column, and the empty
Configuration column appears only when a host reports a provider or region. The host drawer's
metrics empty state reloads the drawer at one hour instead of navigating to the host map, and its
copy states the 15-minute freshness slice. Group-by and resource controls drop their duplicate
count live region. Brand and success fills used as decoration (drawer header, count pill, Ready
badges) become plain text with a dot; the one Not ready badge stands alone. Pod names and
workload badges truncate at the end. CPU renders as cores to three decimals on every list.
* RUM: rated cells keep their coverage tooltip; pin that a failed revalidation never serves the 3h slice as final
* e2e: the host drawer's one-hour retry is a button
* Charts: one sweep for registrations whose element is gone, on after:swap and after:settle
An outerMorph settles without firing after:swap, so charts kept alive across a morph and
then removed with the fragment were never released; the two sweeps are now one.
* RUM: refreshing hint on the Overview's recent sessions too; the session list fits a phone
The Overview's recent-sessions panel is served from the same newest-3h slice and now says so.
At phone width the list dropped its Duration column behind a 28rem minimum; the minimum applies
from sm up and the Activity counts hide below it.
* Speed up local CI signoff with cache reuse and serialized runners
* Keep RUM vital reads and session cache behavior correct
* Auto-format code with fourmolu
* Handle empty CI mounts and align host layout assertion
* Infrastructure: drawer retry reuses hostDetailUrl, empty state names the real freshness slice, wrapper divs folded
* Monitors: drop an unused import that fails -Werror
* RUM replay workspace: fact row in the header for every session, selected row scrolls into view, duration column fits a phone
The header showed only avatar, name and id for a session with a recording; duration, page
views, events, errors, last page and service now sit in one fact row for every session. A
shared link's selected row sat thousands of pixels down an unscrolled list; it scrolls into
view once on init and stays put on live ticks.
* Session replay player: fit the frame to the recording's aspect ratio, one-row compact toolbar
Full-width embeds used a fixed 550px frame, so a wide recording sat between black bands (280px at
1440, 367px on a phone). The frame now follows the recording's aspect ratio, clamped to the
viewport. Below 760px of player width the toolbar keeps one row: icon buttons with labels in
tooltips, and the identity no longer truncates.
* RUM: Web Vitals coverage notes say what the export did and what fixes it
* RUM: the Overview's recent sessions are the Sessions tab's list, one cache entry for both
The Overview ran its own span and recording queries under separate cache keys, so visiting both
tabs scanned the window twice. Both now read the unfiltered session search, which already merges
recordings; the two Overview-only query variants are gone.
* RUM panel cache: serve stale entries for twelve hours instead of one
The first visit of the morning paid a cold 24h scan because last night's entries were past the
one-hour horizon. Entries now serve stale and revalidate for twelve hours; the on-write prune
keeps the same bound.
* Infrastructure images: registry names stay on one line and tag pills wrap instead of clipping
* RUM: trim three comments the review pass found restating the code
* local-time: re-localize after a morph swap
A morph patches the element's text back to the server's UTC fallback without touching its
attributes or reconnecting it, so every live tick reverted the page's localized timestamps.
* RUM Overview: one pulse scan serves presence, the session count and the P75 tile
The presence probe was a LIMIT 1 while the four stat tiles each fetched their own number, so
the tiles spun 2–10s after the panel and never shared its cache. The probe is now one aggregate
over the window; the session and P75 tiles render with the panel (and are cached with it), and
the two sparkline tiles keep their own fetch, whose series is the same request as their number.
* RUM pulse: no row means no telemetry; the overview test pins the served tile values
* RUM Overview: all four headline tiles render from the pulse scan
A timeseries stat tile now shows a server-supplied value until its sparkline response brings
the stats, so page views and browser errors arrive with the panel like sessions and P75.
* Log explorer sessions: one grouped scan plus a page-bounded context read
The Sessions viz referenced a wide CTE four times (DataFusion inlines each
reference) with body::text in the projection: 54s on the demo project over
24h. Now one scan per session key with only rollup-servable aggregates
(any per-group distinct count forces the raw store: 7s vs 0.15s), sort,
paging and the header summary in Haskell, then landing url / user agent /
first error / services for the page only, bounded to the page's own span.
The header's service count is a GROUP BY service read on the first page.
traceCount is gone from SessionRow; the [+N] chip counts events, which is
what the expand endpoint returns, and grows with its count. sessionSortColumn
was left with only its doctest and is deleted.
Demo project, 24h: 54s -> 7-28s (scan cold vs warm), load-more 7.6s, 3h 2s.
* Log explorer sessions: a live poll skips the header and its service read
The client's 15s newer-poll sends cursor=… without a window; the sessions
route ignored it, so every tick re-ran the header summary and the service
read on top of the aggregation. The route now takes cursor and a poll gets
rows only; fetchSessions returns Maybe SessionSummary gated in one place.
Poll 1.4s -> ~1.0s on the demo project. Also from the second review pass:
uniqueServices pinned in Test 11.1, an orphaned Default deriving dropped,
the row copy via RecordWildCards.
* Log explorer sessions: run the header's service read under the grouped scan
On projects whose sessions span the whole day the page context read is a
window scan too, so the view was three raw TimeFusion scans in sequence
(7s + 5s + 10s on the main project). The service read is independent of
the scan, so it forks alongside it in a ki scope: 24h view 31.6s -> 16-19s.
* log-list: a live tick during the initial load no longer re-runs it
With the sessions viz on a big project the initial load took 24s, so the
15s tick fired before any row existed; without a cursor it issued a second
full aggregation (15s more TimeFusion time) for nothing. A tick while the
initial or refresh load is in flight is dropped.
* Log explorer sessions: retry transient TimeFusion errors like the data read does
Two concurrent raw scans from one request is the burst shape that trips
TimeFusion's transient resource errors, and the sessions handler was the
one explorer read without retryTransientEff around it.
* Apply the branch review: typed regex lowering, versioned RUM cache keys, explicit vital coverage rank, keyboard-reachable session actions
hs-distill / hs-evasion-review / hs-lob-review over the branch, applied:
- Expr: CaseMode reaches the jsonpath lowering as a type, not a string op;
compileRegex lives in evalExpr; seedColumns is one modifyIORef'.
- RUM: PulseQuery + queryVersion/rumCacheDbKey replace per-constructor
key suffixes in the handler; VitalEstimate is a record with optional
bounds; VitalCoverage priority is an explicit rank, not spelling order;
interpTimefusion folds the pool routing into one call.
- Widget: hx-morph-skip is static markup; a served stand-in lives in its
own servedValue field with a null source so the sparkline still loads.
- Components: the table live-tick listener fires on after:settle (an
outerMorph never fires after:swap); timeWindowVals_ uses window.params().
- Explorer sessions: SessionsRead sum type; the [+N] chip and Replay run on
click so Enter works; the list header no longer prints a page-local event
count as the total; histogram bars carry aria-labels; durations roll into
min/h/d; URLs cap at 40ch.
- Tests: deferredBody/shellHtml shared in TestUtils; RUM spec fixtures
deduped; doctests for routeKql, VitalEstimate, getDurationNSMS.
* Auto-format code with fourmolu
* Utils: integer literals in the getDurationNSMS doctest
* Shells read the query string inline again; a served stat value wins for any tile type
Two e2e contracts the review pass broke: deferred shells must resolve before
the application script arrives (window.params is not there yet), and the
Sessions / P75 tiles are plain stats whose served value the type-selected
fallback skipped, leaving the tiles blank.
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
CI's image build takes ~71 minutes; the same image built natively on an amd64
box takes ~36 cold, and far less warm. Waiting on a remote queue for the slow
half of a deploy is what makes "push it and see what CI says" tempting, which is
how untested work reaches master.
`make ship` does all four steps here: checks (attesting each pass), build and
push the image, push the commit, deploy to CapRover. The order is deliberate —
checks first because everything after is expensive, image before push so CI's
probe skips its own build, push before deploy because production must run a
commit that exists on origin. It refuses to deploy a commit origin/master does
not contain, and refuses when any deploy-path check lacks an attestation for
this exact tree. weeder and hlint failures do not block it: they gate pull
requests, not deploys.
`make builder-setup` points buildx at a native amd64 host over SSH, capped so a
-j48 GHC build cannot starve whatever else that machine runs. Without one the
build falls back to the emulated local one, and a stale builder name falls back
too rather than failing a deploy. The build now shares CI's registry cache in
both directions, so a laptop build no longer starts cold.
`make tf-image` builds TimeFusion for this architecture, which is the only
reason integration-tests could not run on a Mac; ci.sh detects the result with
no exported variables. Its image is distroless, so the compose healthcheck could
never pass — readiness is now probed on the pgwire port from the runner, which
also proves the thing the check needs rather than that a container exists.
CI's deploy job asks `ci.sh deployed <sha>` and skips a commit already deployed
from a machine, so this cannot restart production twice with the same image.
Deploy credentials are env-only, via the per-app deploy token rather than the
admin password.
bash reads a script incrementally by file offset. The container ran ci.sh
straight off the bind mount, so editing scripts/ci/ci.sh while a run was in
flight made the running shell resume at a stale offset and die with
scripts/ci/ci.sh: line 605: syntax error near unexpected token `;;'
The check itself had already passed and been attested; only the wrapper
died, which is precisely the "bookkeeping must not fail a green check"
failure this repo has hit twice now. And a `make ci` lasts tens of minutes,
so editing the script meanwhile is a normal thing to do, not a mistake.
Copy it into the container and exec the copy.
With tests cached the deploy is 4m22s, of which 3m52s is the image build — the
whole critical path. The image is not fingerprinted like the other checks
because it produces an artifact rather than a verdict, but it does not need to
be: it is tagged with the commit SHA, so the registry is already its cache and
the SHA is already its identity.
build-image now probes the registry and skips when the tag exists. Nothing
changes for a new commit; it collapses the repeat cases to ~30 seconds — a
re-run of a failed or cancelled deploy, a workflow_dispatch of an existing
commit, a revert to an already-built SHA. Skipping re-points :latest at the
reused image rather than leaving it on an older commit, since docker-compose.yml
pulls :latest.
The same probe lets a bigger machine pre-build the image (`make deploy-image`,
linux/amd64 — Rosetta on Apple Silicon, not QEMU). That moves the artifact that
runs in production outside CI, and unlike an attestation it cannot be verified
afterwards, because a Haskell build is not bit-reproducible. So the build
refuses a dirty tree, and the deploy job prints `built by: …` every time —
never fatally, since a provenance lookup must not be why a gated build fails to
reach production.
Docs: CONTRIBUTING gains the section a contributor needs, including the two
caveats that would otherwise surprise them — the first `make ci` is a cold
build, and publishing a result needs push access, so from a fork it is fast
local feedback but CI re-runs the checks itself. README links docs/local-ci.md.
CI re-runs work that has already been done: a PR run proves a tree, the squash
merge keeps that exact tree, and master builds and tests all of it again. On
4-vCPU runners that is ~20 minutes per push, and none of that capacity is
available to a developer whose laptop is faster than the runner.
ci/checks.tsv is now the one definition of what CI is; the workflows and
scripts/ci/ci.sh both read it. Each check is fingerprinted over the git blob
hashes of its declared inputs, and a pass publishes an attestation as a ref:
refs/ci-attest/v1/<check>/<fingerprint>/<platform>/<capabilities>/<date>
A gate job resolves those with one ls-remote and no object fetches, and skips
any check already proven for the tree it is about to test — whether by an
earlier run, a re-run, or `make ci` on someone's machine. Producing one needs
push access, the same trust boundary as pushing code; forks can neither create
nor be affected by them.
Reuse is deliberately conservative because this gates a production deploy.
Checks declare the capabilities they need and attestations record the ones the
environment had, so a run that fell back to a stub service can never satisfy a
check that needs the real one. Fingerprints are pinned from the gate, so steps
that rewrite the tree (hpack, npm ci, hlint --refactor) cannot desync what gets
attested from what was looked up. The deploy condition is spelled out rather
than implied by `needs`: build-and-test may be skipped, but only when the gate
says every Haskell check was already proven.
Also switches the minio service to bitnamilegacy/minio. The bitnami/minio
repository was emptied in the Aug-2025 catalog change and now has no tags at
all; CI has only been working off a runner-side image cache, and an eviction
would have taken the integration suite down with nothing in the diff to blame.
Verified end to end: all six checks ran through ci.sh and published
attestations, and a clean checkout on macOS/arm64 then resolved CI's
linux-amd64 attestations and skipped the entire Haskell suite.
See docs/local-ci.md.