A dev chart failed with 'Chart result column count does not match its decoder' on the buffered
/chart_data route, which never retried a bad connection; only the stream route did. The retry now
lives in fetchMetricsDataWithProgress, so both share it, and a row wider than its decoder raises
StreamDesynced naming the columns that arrived, so the next occurrence states its cause.
* Floating AI assistant: page-aware sessions that act, navigate and reload the page
Cmd+K or typing a letter opens the palette; Enter asks the assistant, which runs in the
background as a web conversation bound to the page it was asked from. The dock shows
sessions as bubbles, the selected transcript with live steps, and visits pages the agent
opens (open_page) or reloads the current page after a write.
* Dock polish: transcript-style progress, launcher with shortcut, entry motion; snapshot goldens
* Dock: OpenSession record, single-meaning navigate, inline bubble classes
* Dock panel reveals and focuses itself when a session first opens
* e2e: palette leads with Ask AI and hands the request to the dock
* Dock: catalog write policy, indexed session query, concurrent loads, 2s poll
* Dock: keep the changes toggle across sends; follow-up turn binds its page; lint
* hpack: list the page-context migration
* Dock: a log-view answer opens the explorer; prompt says only open_page changes the screen
* e2e: wait for the palette's lazy items before moving the selection
* Dock requests run on the receiving server; the queued job is a deferred fallback
* Browser tools: read_page, click and fill run in the user's tab; the run pauses until it answers
* Charts saved without a height get three rows; dashboard tools describe their body
* Dock reloads the current page by a whole swap; morphing a dashboard onto itself left it empty
* Stats saved without a height get two rows
* Assistant prompt lists a monitor's own page
* Page snapshot skips empty headings
* Assistant prompt: on-page parts of a request use the page tools; no unproven claims
* Document why the page tools carry a read-only policy
* Widget lifecycle spec: sizeless charts are three rows tall
* e2e: retry the palette's ArrowDown across the lazy items landing
* One prompt field for the thread and dock composers; tool names from the catalog
* Stored widgets get distinct ids; same-titled widgets no longer share one identity
* Widgets aggregating duration show time units; one inference for every widget
* A resize never redraws a chart before its data lands; echarts threw on the null source
* Dock: one let for the page change
* Distill: inline single-use tool list, merge height arms, dock spec helpers; hlint
Results and chat answers saved before runs pinned their charts re-queried the
live window, so a run's "11 ERROR events" sat under a stat reading 0 or 133.
Stored widgets now pin to the answer's time the way new runs do.
A run's tables re-queried the default window ending now, so a fixed Oct 31 hour
showed 45, 64, 0, 1 and 72 on successive loads of a live project instead of 145.
Browser-drawn charts already honoured timeFrom/timeTo; server-rendered tables,
top lists and stats now do too.
A run's tables showed "Widen the time range or loosen the dashboard filters" and
a "Show last 24 hours" button that changed the page, not the run's window.
Integration flakes that read back nothing after an ingest were undiagnosable:
the retry's `make ci-down` removed the services and their logs with them.
Docs, comments, formatting, workflow and deployment config, and test-only changes
are fully shown by the diff; recording them only replays the changed lines.
A lone legend entry only repeats the query's alias (count_) under a title that
already names the series; the chart takes the space back. Value-table legends stay.
Email charts were 900x300 exports shrunk into a 600px column, ~4px text on a
phone. An email profile renders 600x240 with Slack's 14px labels, on the
email's own surfaces and without alert annotations.
A report's weekly chart ended each side with "15:33:20": the window's boundary
ticks fall mid-day. Over two days or more the daily ticks mark the window.
A run's Visualizations re-parsed the raw answer, so its charts re-queried today.
They now pin to the run's start like the stored answer does.
Same-named containers listed pod, namespace, node and cluster on four lines even
when most were shared; rows now show only the fields that tell them apart.
* Container rows spell out pod, namespace, cluster and node only when another listed container shares the name
* A running routine's facts say when its run started instead of repeating the Running pill
* Routine pages read more clearly: latest-result heading wraps on phones and says later runs had no findings, secondary actions look like buttons, failure counts never clip, run summaries leave colour to the status pill
* Single-use run error tone moves beside its caller; latest-result separator matches the others
* Routines: status strip, informative run history, latest result, table list with actions menu
* Reports become AI routines: weekly on, daily paused; dated report documents, highlight charts, email delivery; remove legacy reports
* Wide tables: one-line headers with hints, compact container rows; routine documents panel polish
* Routines: run now, pause and resume from the routine page
* Tests: report routines specs, goldens for routines, wide tables and bot copy
* Routines on phones; docs and e2e for reports as routines
* Report runs: digest only on the report path, charts pinned to the report window
* Routines: choose in-app, email or Slack delivery; show delivery in the status
* Report routines: Observe policy, paid plans only
* fourmolu
* Routines: plain policy labels, clean wrapping facts, menu semantics, clearer intro
* Spec: run now starts an active report and leaves a paused one paused
* Routines: report view with a chart beside each highlight; one-line ask preview; plain policy names; compact stat widgets
* Routine actions disable while in flight
* Routine menu: start-aligned, divider before Delete; report hook matches run completions exhaustively
* Resume returns to the routine, like pause and run now
* Simplify: local times via local-time, shared run summary and chart image, Email only on reports, lighter report fetch, drop dead IssueSummary
* hlint
* Reports deliver to Slack by code like email; report templates install read-only
* Reports on real data: agents query both periods (a 1h default made weekly reports blind), exact period split, ms latency charts, enforced materiality; dev report preview
* hlint
* Routine page leads with the latest result; list shows runs that need attention; containers fit at 1280px
* Routines list keeps runs from deleted routines reachable; containers golden
* Reports deliver through the team fan-out alerts use (same recipients and channels, Discord too); one report window definition
* Routine actions stay top-right when the status line wraps
* Report routines run on Free too; a run keeps only the document versions it wrote
* Routine catalog cards lead with the title; schedule and reporting rule share one line
* Tables keep the first column's header on phones unless rows are selectable; routine ask reads as a quote; issue goldens for one-line headers
* Dashboard list golden: no bulk-action padding without selectable rows
* Report email preview charts latency in ms, as real reports do
* A routine's kept result pins its charts to the run; last-run outcome as a dot; phone routine page leads with the result; container rows show the workload, not a truncated image
* Report email separates highlights with a hairline
* Drop unused import
* Phone routine list shows the last outcome; PNG charts drop one-series legends; dark email button keeps contrast; workload badge without namespace; no one-item 'Show all'; category chips wrap
* Issue triage reports its outcome in plain words; a run that merged nothing has no findings
* Containers show Pod and Namespace columns from 1800px, so every column fits beside the filter rail
* Latest result shows when it was written
* Routine status line leaves the cadence to the header's schedule control for editors
* Needs attention: one row per run whose newest attempt failed, five shown; untitled routines' runs no longer read as deleted
* Containers drop columns no container reports (Pod, Namespace, Restarts, Ready under Docker); the name takes the width
* Containers spec pins which columns show for a mixed and a Docker-only inventory; golden tracks the container column again
* Run details: local time like the run list, attempt only on retries, a scheduled run's request no longer repeats its instructions
* Event-triggered routines get the routine composer too: no autofocus, 'Ask about these results'
* Routine health: recent failure count on the routine page, failed runs name their routine, Active reads as a dot
* Runs list skips the summary of the result already shown above
* Runs list calls an attempt a retry only after a failed one; a new push reads 'run N'
* A routine with no runs says so before its settings
* Issue autofix shows the fix runs it started and its real last run instead of 'Not run yet'
* Report charts leave a duration the agent already converted unlabelled, instead of calling 4.6 ms 4.6 ns
* Fix runs name the autofix routine that started them
* A routine's latest result reads as a record: no chat avatar gutter
* Review fixes: a path filter no longer drops a chart's ns unit; one showMore_ for runs, documents and attention lists; runFailed predicate
* Images: long tags truncate inside their column; drop the always-unavailable Security column (the drawer still explains SBOM coverage)
* Charts whose query is prose are dropped, not drawn empty; latest result notes newer quiet runs; Installed rows count recent failures; 'Result shown above'
* Routines list on phones: editors' status column gives way to the last-run outcome; failed runs put the routine name on its own line
* Phone routines list: last run and recent failure count on their own line
* Routine status line keeps 'Last run' on one line on phones
* Routines list failure count says it covers the last N runs
* A routine's last agent round can still save its document
Scheduled triage runs spent every round reading issues, then reached the
final round with no tools and reported that the day's document could not be
written. Since Nov 2 no routine kept a daily document. The final round now
offers save_artifact alone and honours that call; the round after offers none.
* Autofix run titles mark where a long issue title was cut
* Routine delivery menu posts a plain form field; disclosure class has no stray space
* Routine runs resolve their owning routine once; email delivery only for reports; undeliverable reports and retired jobs are logged
* One report item type and producer for email, chat and preview; one report template table; shared failed-runs note
* Routines list actions post without nested forms; the last agent round runs only the offered write; email delivery for a non-report routine is refused, not ignored
* The last agent round's unoffered calls leave its message too; row menu items never submit the table form
* nonEmpty for the kept tool calls
* Run history filter stays indexable; one report-template lookup; report charts use the pinned widget window; conversation titles mark their cut
* Remove report-only helpers weeder flagged; routines e2e follows Documents nav and unfolded runs; closing details returns focus to a routine's draft
* Phone composer shows its extras for a draft, not on focus, so tapping away doesn't shift the page; e2e matches dated document links
* Overview's country map works: server spans place by their peer, our own spans record the forwarded client, page loads count, top lists rank by count and say when empty
* Container rows keep their pod, namespace, cluster and node where those columns are hidden, so duplicates stay distinguishable
The hourly job moves every unedited catalog routine (and its opening message) to its template's current version and logs how many moved; provisioning no longer gates upgrades on repository analysis. Production's triage routine stayed on v1 after a successful analysis.
Production showed both routines idle: the hourly sweep ran on app replicas
without a sandbox ("Agent sandbox is not configured on this host"), and
triage could not page 1,562+ open issues inside one run. One longJob rule now
places jobs on the agent queue for seeding and enqueueing alike. Triage v2
compares new issues only against open issues of the same service and type,
and installed defaults still on their template's text follow newer versions.
* AI routines UI: render agent markdown, group PR threads under their routine, resizable sidebar
- Markdown MMark rejects (review reports with tables, images, <sub>) falls back to
sanitized GFM instead of printing raw source; textOnlyHtml is shared with the wiki.
- Query-less tables render their inline dataset rows (blank "Review query" widgets).
- PR-review threads leave the Conversations list; they live under their routine.
- Routine page: settings collapse into one disclosure; runs list is compact, titled
per subject (or attempt within one), shows errors, and folds older runs.
- Run inspector: icon-only close, separated metadata row, pretty-printed evidence,
link to the run's own thread.
- Sidebar edge drags to resize (persisted), double-click collapses.
- Lost-session failure says what to do instead of naming storage internals.
* Sidebar marks the PR routine active on its threads; test PR threads stay off the conversation list
* Impact review: the service-events chart reads its evidence window, not the page's range
* Monoscope-view review captions use Markdown, not <sub>; older runs keep their disclosure across live refresh
* PR threads lead with their result; blockquoted findings lose speech quotes; resizer keyboard steps don't lag the width transition
* AI views share one chevron disclosure for settings, run instructions and evidence
* Distill: one renderTable clause for inline and fetched rows; read-only policy text reuses policyLabel
* 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
A finished fix queues PublishFix on the agent queue. Personal GitHub tokens are tried first and attach the video and diagram; otherwise the App installation opens the draft PR linking Monoscope-signed media (new per-asset signed route). Fix-job failures now log their cause, and a failed autofix issue is retried a day later.
* install.sh: accept a pinned version without the v prefix
Release tags are v-prefixed, so `install.sh 0.6.31` built a /download/0.6.31/ URL that 404s.
* Re-run review
* Re-run review
* Live Tail: keep the toolbar on one line at desktop widths
The query editor's 288px flex-basis pushed the toolbar's basis widths past ~1047px, so Clear
wrapped onto a row of its own beneath the filters. A 256px basis fits; the editor still grows
to fill the remaining space.
* Live Tail toolbar: let the query editor shrink to 12rem so Pause and Clear stay on one row at 1280px
* Errors seen only from automated browsers get their own issues tab
The browser SDK marks webdriver sessions with OTel's user_agent.synthetic.type.
A pattern stays synthetic until a real occurrence arrives; synthetic issues are
kept out of the Inbox and every other tab and listed under Synthetic, and their
alerts say so.
* Serve browser SDK 0.13.3, which marks automated browsers as synthetic
* land: an unproven check blocks the merge
`make ci-status | grep -q` under pipefail read grep's early exit (make's SIGPIPE)
as "nothing pending", so #863 merged with e2e unproven.
* Synthetic issue filter probes a partial index
The bool_and subquery added 20-100 ms to every Inbox count on prod. Framework
issues group many patterns and are never synthetic.
* Migration 0279 was taken on master; this one is 0280
#862 changed only TypeScript, yet its image build recompiled 133 modules for
26 minutes: the cache's last writer predated #859, so every module #859 touched
looked changed. Only builds containing master write back now.
Install-time webhooks return no message timestamp, and a webhook post carries no metadata for the timestamp search,
so every root sent that way turned threadless and its replies ("Resolved by ...") landed in the channel alone.
The bot now posts whenever it can; the webhook is only the fallback when Slack refuses the bot (not_in_channel).
* Guests read package registries, so a proof or build can install a missing toolchain; frozen-check log assertion follows the 24 MB cap
- every guest gets GET/HEAD-only HTTPS to npm, PyPI, crates and rustup, the Go proxy, Hackage and ghcup, Maven and RubyGems, with a fixed User-Agent; GitHub is excluded except ghcup's metadata
- guests get udhcpc, which gondolin's init needs to bring up the network
- the agent is told to install missing tools with user-space installers; PROOF.md says not to build this repository cold
* Fresh nodes start warm from the lineage's archived build; workspace builds record a duration by outcome; plain proof clips crop to the rows they print
* fourmolu 0.19 and hlint
* Agent reach: fetch_url reads public docs through the host; list_issues and list_monitors read the review's mapped services
* fourmolu, hlint, docs for the new agent tools
* fourmolu
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 query may have only one summarize: the SQL compiler cannot nest them, so a second one is a positioned parse error.
- Monitors: every write path (form, widget, API create/PUT/PATCH, monitors-as-code) builds through monitorFromInput, and compileAlertSql is the one alert-SQL compiler (evaluation included).
- "Any results" means a count of at least 1 (migration 0279 fixes stored threshold-0 monitors).
- Below monitors read the lowest group and measure.
- 12h windows parse as 12 hours.
- The explorer editor reopens on the saved condition and window.
- Dashboard template variables are filled ("unset = all").
- Edits keep trigger history; an omitted active keeps a paused monitor paused; a query that does not compile is refused.
- One error predicate (isErrorSql/isErrorKql, agreement pinned against the evaluator), used by sessions, the errors flag, reports, bots, impact review and Issues.
- Uptime and cron writes need edit access, like the Prometheus targets that share their table.
- Monitor page chart follows vizFor, and its preview chart reads the monitor's own scope (Widget.scope, pinned_scope).
- Shared nativeDialog_/dialogTrigger_ and resourceList_ components, used for the Prometheus, review and remove-repository dialogs and the uptime/cron/Prometheus lists (status no longer colour-only).
- drawer_ closes on the server's closeModal, replacing four inline close scripts.
- Billing tables on Table via a gridClasses config field.
- parseSeverity replaces three log-level lists.
- Query builder reads a query that opens with summarize.
Seen by the recurrence watch: "Failed to load …/echarts.min.js" was an
unhandled rejection the browser SDK filed, and the cached rejection meant no
chart on that page tried again. The load is now forgotten on failure and
reported through logFailure.
* Deploys wait for the Mac's image; image-follow builds master's head; CI image bakes ffmpeg/graphviz; dashboard edits refuse missing dashboards, tabs and widgets
- release.sh: the deploy's image step waits up to wait_minutes for the image, and stops early once master moves on
- make image-follow: builds only master's newest head, cancelling a superseded build
- Dockerfile.deps bakes ffmpeg/graphviz; ci.sh's fallback install is bounded
- query-preview test polls for CodeMirror's time-sliced highlighting
- EditRefusal: DashboardGone/TabGone/WidgetGone join EditReadOnly in the route types
- CLI: disambiguate EventsSearchOpts record updates
* log-list test: poll for the load-newer reveal scroll, which lands on a later frame under load
* image-follow: an image already built for the merged tree counts
* hpack: list migration 0278 in monoscope.cabal
* Proof comments: issues found while testing, many claims per recording, a player bar, and a rating link
Recordings can prove several planned claims (covers), a review can check up to twelve, and anything
a recording shows broken that the PR does not claim becomes a caution callout. The preview's
timeline brackets the clip with its start and end times; the footer asks whether the proof got it
right through signed Yes and No links.
Also: deploys restart the proof worker only once no review is running; a saved proof setup is
dropped when .monoscope/PROOF.md changed since it was recorded; Live Tail keeps three-digit
milliseconds; plan_proofs is tested through the native host.
* Rating asks before it records; review follow-ups
A rating link only asks, and the button POSTs, so mail scanners and unfurlers that open every link
cannot answer. An issue must cite its recording; a retry clears an old rating; the rating question
appears only with recordings. A failed image pull fails the deploy again, and a second ship reuses
the pending restart. A truncated GitHub comparison and a failed one are both logged and drop the
saved setup. The player bar's right-hand time is the video's length.
Each failed only while sign-off compiled the image alongside: the drawer's
Add widget was clicked before htmx powered its form (no PUT was sent), two
multi-dashboard specs ran past their timeouts, and a history restore outlived
the default 5s.
- Every response carries X-Monoscope-Build (the frontend bundle the server
serves) and the page carries it in a meta tag; an htmx response naming a
different bundle reloads the tab once. A tab left open for months on SDK
0.11.6 kept filing errors that read as regressions of fixed code.
- scripts/land.sh / make land PR=<n>: sign off, wait for the tree's image,
merge only while master is still the base, else rebase and repeat.
Seen by the recurrence watch: "[Worker] Error: the log worker failed to load"
with no deploy nearby, so the stale-chunk reload did not cover it, and waiting
requests sat until the 2-minute timeout. The list now fetches on the main
thread, re-running what was waiting. The page builder the worker, the early
preload and the fallback share is one toPage.
Seen by the 24h recurrence watch: an empty body on the world geojson or the
echarts theme JSON ("Unexpected end of JSON input") was console.error'd and
filed as an issue. Both recover — the map loader retries on the next render
and charts keep echarts' default theme.
fetchJson passed any body with an error field through for the caller to route,
and a 401 answers {error: "unauthenticated"}: the log list showed that in the
query box instead of asking the reader to refresh.
The server links only profiled objects; cabal's vanilla pass exists to run
Template Haskell splices. Building it at -O0 (ghc-prof-options keeps the
profiled pass at -O1) cut it ~2.9x per module in a cold build. Dropping it via
-fexternal-interpreter is ruled out by the AutoInstrument plugin. New cache id,
since the build config changed.
A prove-it on #826 reused the progress comment an earlier review of the same head had updated, so
its answer landed far above the ask. A new ask now clears the run's progress comment.
rum.yaml is read_only, so it always renders the shipped YAML and edits go through
Duplicate. Migration 0277 keeps the one RUM row that froze its own layout as an
editable "(Customized)" copy and relinks the original, as 0266 did for the others.
* Handled browser failures stop becoming issues; fix three e2e flakes
- One fetch-failures module (ShownFailure, logFailure, fetchJson) for the log
list, its worker, charts and dashboard variables; the worker's fetch copy is gone.
- Chart HTTP errors and unreadable bodies, and variable-option 5xx/unreadable
responses, are shown to the reader and warned, not console.error'd.
- A log worker that fails to load (a tab open across a deploy) reloads once.
- Vendor browser SDK 0.13.2: transport failures are breadcrumbs, and error spans
carry exception.escaped, which the record extractor now reads as handled.
- query-preview forces its parse so a busy page never shows it unhighlighted.
- routines e2e: wait for the install/save redirects; drop 1s loading-bar budgets.
* Local test targets install the impact-review worker's packages, as CI does
2026-11-11 01:39:12 +01:00
219 changed files with 5704 additions and 4480 deletions