Repository navigation
Releases
Monitor 2.1.7 - operator screens, a stronger server, sentinels that tell the truth and a library of sounds
Version 2.1.7 brings together everything planned for 2.1.6 and 2.1.7. By theme: screens and kiosks (reload on a new version, a Fleet screen, density, palette, contrast, overscan, print and save of a report); phone and PWA (install as an app, compact view, QR pairing, palette and shortcuts); real-time and recovery (webhook queue, WebSocket without a public address, GitHub status, heartbeat); scale and observability (histograms, allowance by bucket, OpenTelemetry, workspaces, cost and carbon); the command line and the AI (formats, watch, doctor, proposals with approval, alerts history); the Sentinels, with honest reads and no false findings; and the release chain (immutable releases, a zizmor job, an updater that keeps copies and a log).
Added
-
A wall screen or a kiosk that stays on for days reloads itself when the monitor is updated or restarted: the new route
/api/version(never cached, with the same access as the other reads) answers{"version", "started"}, and the wall screens and the kiosk profile ask it everyscreens.version_check_secondsseconds (300 by default; 0 = never; the least is 30). When the version or the start time is not the one the page loaded, it reloads itself, so a TV never keeps an old page. It reloads only after a valid answer that differs: a monitor that does not answer never makes it reload in a loop. The guides of the screens, the kiosk and the API say so in the eight languages. -
A result that is too old is called stale (
ci.stale_result_hours, 0 = off by default; 72 is a good value): a workflow whose last result is older than that many hours is markedstale_resultin the lists of workflows of/api/dataand shows the label stale in the table of the CI sentinel. It is information only: a stale result is neither red nor green, changes nothing in the score, and a repository whose runs could not be read this time is not marked. It reuses the rule of the aging of the reds (redrule.py). The guide of the sentinel says so in the eight languages. -
A countdown to the 60 days after which GitHub switches off the scheduled workflows (
sentinel.schedule_countdown_after_days, 45 by default, 1 to 59): for a public repository with a scheduled workflow whose last push is older than that, the hygiene panel of the sentinel adds an information row with the days left. It uses only the date of the last push that the sentinel already reads (no new call to GitHub), says not evaluable when that date is not known, ignores private repositories and is never red. The guide of the sentinel says so in the eight languages. -
The updater that runs without the monitor can show the copies, read the log and discard a copy: its page in the browser has the buttons Copies and Process log and a Discard copy button on each copy (it asks first, and twice for the copy the way back needs), its text menu has the options 7 (copies), 8 (log of the process) and 9 (discard a copy) and its window (Tkinter) has the same three actions. It is what the Updates menu of the page and the commands
copies,discardandlogalready did; only the copy chosen is deleted, never the installation or the data, and the act is written in the history. The guide of the update names the three doors. -
Every error of the read API carries
help(/api/v1/): next toerror,codeandmessage, one sentence says the next step and names the guide to read, in the language of?lang=(a code without one has nohelp, and a client that ignores the field loses nothing). It is the same idea as thehelpof the errors of the command line; the shape of the contract only gains an optional key, so no client breaks. The guide of the read API says so in the eight languages. -
The conditional reads are noted after every cycle (I-155): the line the monitor writes to its log for each cycle ends with
etag=<answered 304>/<reads> no_etag=<answers without ETag>when the cycle made conditional reads,/api/quotakeeps the figures of the last cycles in the way of savingetag(recent_cycles, memory only) anddoctoraddslast cycle 6/8to its line of the ETag. A proxy that strips the header, or a cycle that spares nothing, shows in the cycle itself and not only in the total since the start. -
A fork can say how far behind its original it is (
fleet.fork_behind, off by default;fleet.fork_behind_per_call, default 5): the Repositories table gets a column for the forks with the number of commits the original has that the fork lacks (the default branch of the fork against the default branch of its original, one read of the fork and one comparison). At mostfleet.fork_behind_per_callforks are read a round, two rounds are at least 30 seconds apart and the answer livesfleet.meta_ttl_hourshours. A fork that cannot be read (no access, the original is gone, the branch is missing) is not marked, neither green nor red: the column says "not evaluable". Only the repositories the sentinel knows to be forks are asked about, andGET /api/meta-columnscarries the answer asfork_behind. -
Reusable workflows can stay out of the failure lists, and orphan workflows are listed as information (
ci.reusable_out_of_rankingsandci.orphan_workflows, both off by default): with the first, the old runs of a workflow whose file only hasworkflow_callleave the red list of A2 and the rankings of duration and of flaky workflows, and the sentinel says how many workflows it left out. With the second, the workflows that GitHub still lists and whose file is no longer in the default branch appear in A2 as information, never red, and a repository whose workflow files were not read is counted as not evaluable instead of being judged. Both use the list of workflows and the workflow files the sentinel already reads, so they add no call. -
The runners and their jobs can be grouped by label or by runner group (
runners.group_by:noneby default,labelorgroup): the runners ofGET /api/runnersare grouped by their first label that is not a generic one (groups), and the jobs ofGET /api/runners/modelby that label or by the runner group GitHub gives the job (grouped), with the jobs, the ones waiting and the oldest wait of each group. Nothing new is read: it groups what the monitor already holds. The list of runners of GitHub does not say the group of a runner, so there the grouping by group is shown as not available instead of being guessed. -
Density, less movement and a palette for colour blindness for the whole page (
display.density,display.reduce_motion,display.palette;?density=,?palette=for one window):densetightens the spacing and the type of every tab, not only of Now;reduce_motion: onstops every animation and transition whatever the system says; thecolorblindpalette (blue, vermilion, yellow and a lighter blue, one set for each mode) keeps the four states apart for the three common kinds of colour blindness, and the state is always written in words and shape too. -
A contrast gate (
tools/check_contrast.py, run by the CI): every text on every surface of the dark and light modes and of the sixteen themes reaches 4.5:1, the colour of a state reaches 3:1, the colour-blind palette is checked with a simulation of the three kinds of colour blindness, and a new dimming of text withopacityfails the build. -
The guides and the next step on the command line:
monitor_cli docs listnames the guides of the installed version with their title anddocs show NAMEprints one in the language of--lang(the English text when a guide is not translated); it reads the installation, so it needs no running monitor. Every error of the command now carries ahelp: one sentence with the next step, often naming the guide to read; in text it is the last line (help: ...), with--jsonit is the fieldhelpnext toerror. -
The page does not jump while you read (
display.live_throttle_s, 30 by default, 0 turns it off): while the pointer or the keyboard focus is inside the tab you are reading, the periodic refresh of the data waits and is applied when you leave, never more often than that many seconds; a refresh you ask for is applied at once. -
What is new after an update (
display.whats_new, on): a notice that never interrupts offers the entries of the changelog of the versions you skipped, in your language, read from the installed version (GET /api/whats-new); what you have seen is kept in the browser before anything is shown, and the palette opens it at any time. -
A glossary and the limits of the analysis (
GET /api/glossary;report_export.limits, on): the palette opens a window with what each figure means exactly (red on main, success rate, streak, duration, wait in the queue, recent failure rate), written with the real parameters of the calculation and of your settings, the attacks the rules of the linter look for, six places to learn more and the limits of the analysis, which the saved and printed reports now end with. -
Branches to watch, repository by repository, and a count of what is left out (
ci.branches_by_repo, entriesowner/repo=pattern, empty by default): a maintenance branch such asrelease-*can be watched in one repository without watching it in all (ci.extra_brancheskeeps adding to every repository). The failures on branches that nobody asked to watch no longer vanish: the Now tab says how many there are (they still do not count as red) and lists the repository, workflow and branch of each. -
The header filter says what it covers, the selector of repositories says what is happening, and an empty or failed page says what to do: a note lists the tabs that follow the filter and the ones that are always global; the selector of repositories shows loading, failed with a button to try again, or nothing yet; the error bar has a button to try again, and a page with ...
Monitor 2.1.5 - every run read for the score and the reports, a command line with output formats, watch and completions, a kiosk profile for a wall screen and two new guides
-
The score, the reports, the mirror and the analytics can read every run, not only the newest 600 (
score.source,recent600by default,ledgeron request): on an account that reads thousands of runs in a cycle the page keeps the newest 600, and the score was made from them. Withledgerthey are made from every run the cycle read, counted once in the ledger, which now also keeps hourly totals (successes and failures per hour) that survive its ceiling of 20000 runs; a window the ceiling could not hold whole says so (data_quality.truncated) and gives the totals the hours hold (data_quality.aggregate). With the default the snapshot says when the cycle read more than the score saw (data_quality.cut,data_quality.runs_read, the reasonrecent_cutand a line on the page). The default becomesledgerin the next version. A ledger and the snapshots of earlier versions are read as they are. No call to GitHub is added; it is in the guide of the score, the settings help, the guides and the skills. -
Every command of the command line takes
--format text|table|plain|json|ndjson:tablefits the columns to the width of the terminal and counts wide Japanese and Chinese characters as two cells,plainis one tab-separated line per record forgrepandawk, andndjsonis one JSON object per line.textis what every command wrote before and stays the default (cli.default_formatchanges it);--jsonis unchanged. -
watch status|alertsrepeats the read of the command (--interval,--once,--no-clear): it reads the monitor on this computer, never GitHub, keeps a floor under the interval (cli.watch.min_interval_seconds, 2 by default), retries a monitor that cannot be reached with a longer pause each time and ends with exit code 0 on Ctrl-C. -
watch --exit-on-changeends with the new exit code 6 when the set of alerts or the figures of the status change: a clock that ticks is not a change, and without the flag the code 6 never appears. The codes 0 to 5 are unchanged and a test keeps the table of the guide and the code in agreement. -
completions bash|zsh|fish|powershellprints the script that completes the commands, words and options, written from the real parser so it cannot fall behind (a test compares both). It only prints: it never edits a file of yours. -
A kiosk profile for a TV or a wall, off by default (
kiosk.enabled):monitor_cli.py kiosk ops|team|release|quietprints the address of a preset, and the page, only when the setting is on and the address haskiosk=1, makes the type larger (kiosk.font_scale_percent, 150 by default), can use a high contrast (kiosk.high_contrast) and can turn through the tabs of the preset (kiosk.rotate_views,kiosk.rotate_seconds). -
Two new guides in eight languages,
CLI_OPERATOR.mdandKIOSK.md: the command line for the operator (formats,watch,completions, exit codes andjqexamples that a test runs) and the kiosk; the guide of the command line and the one of the screens point to them. -
The wall screens take a repository filter (
?repo=WORD,WORD):/screen/overviewand/screen/nowkeep only the repositories whose name holds one of the words, as the page, the kiosk and the tab screens already did. A repository whose name the screen hides is never matched, so the filter cannot be used to find out what a hidden name contains; the figures (fleet, red, queue, events, alerts) follow the rows that are kept, the score and the map of groups of the whole fleet are not shown while a filter is on, and/screen/sentinelsis not filtered. The Screens menu offers the parameter for every wall screen and the window names the filter next to its number. GuideSCREENS.mdin the eight languages.
Changed
-
In the Docker Compose file the monitor believes one address, the one of Caddy, and not the whole network: with the profile
proxy, Caddy now has a fixed address (MONITOR_PROXY_IP, 172.29.77.250 by default, insideMONITOR_SUBNET) andMONITOR_TRUSTED_PROXIESnames that address. Another container of the network that skips Caddy can no longer choose its own address withX-Forwarded-For; the Docker check proves it with a container that lies at every try. If you already named the whole network inMONITOR_TRUSTED_PROXIES, put the address of Caddy there instead. -
The example file of the configuration now carries every setting, and the page of Settings shows a default for all of them: the AI settings (the model, the limits and the provider of each of the five tasks, the fallback chain, the budgets, the private addresses, the record of prompts, the hidden words and patterns), the read pool (which source answers each kind of read first and how many calls each keeps free), the three folders, the days after which a silence expires, the application id of the GitHub App and the private addresses of the notifications. Each value is what the code does when the setting is absent, a test compares them, and the five per-task time limits that said 30 seconds now say 0, which is what they are: the task uses the general limit. The three settings that are computed (the state of the signing of the webhook and of Slack and Discord) are shown but never written.
-
The walk through the page also tells a monitor that fails: besides script and console errors, an answer with a 5xx status to a call of the page is now a problem of the walk (with its address, without the query), and the walk can be made at the size of a desktop and of a phone (
--size desktop mobile).
Added
-
Every run that is going is counted, workflow by workflow (
fleet.show_all_active, on by default): the health of a repository hasactive(shown as Going now in the repository dialog) with how many runs of each workflow run and wait and the newest run of each branch, so two runs of the same workflow that go at once (releases in parallel, one run per branch) are not shown as one; the rows of the fleet carryrunningandqueued. Off: a workflow shows only its newest run. -
What the plan of the account does not have is told apart from a refusal, and it is not counted against the sentinel (
sentinels.security.plan_na, on by default): when GitHub answers that a check needs another plan (branch protection or code scanning of a private repository on the free plan) the Security sentinel shows it asN/A (plan), a refusal asN/A (no access)and a feature that is off asN/A (off)(unavailable_whyin the status, a line under the sub-panels with a tooltip). The pairs of the first kind are left out of the readable count and of the coverage of the score and counted apart (sources_na_plan); none of the three is a finding and none is a clean bill of health, so a plan with none of what weighs givesunknownand no security part of the score. Off: they count as sources that could not be read, as before. -
Actions of another private organisation that the login cannot see can be left out of the supply chain findings (
sentinels.supply.ignore_uses, empty by default): namesowner/repo(with*for any) whose unpinned or moving references are not reported on a guess; they are not hidden either, because the status counts them apart (uses_left_out). -
When an allowance runs out, the read is made with another source and, if all are spent, the last good copy is shown with its age (
quota.failover, off by default, withquota.failover_max_age_minutes): a read that answers 429 or "rate limit exceeded" is made once more with another source that can read (your login, the GitHub App, the second account) and the one that ran dry counts as spent until its next free check; when every source is spent the last good copy of that read is shown with its age instead of an error, never older than the limit asked for. A limit is never taken for a refusal of permission, a write never fails over, and the status of the quota says how many times it acted. -
The daily trend can mark the days a workflow file changed (
sentinel.trend_workflow_marks, off by default, withsentinel.trend_workflow_marks_per_tick,sentinel.trend_workflow_marks_hoursandsentinel.trend_workflow_marks_days): a square on the chart on the days a commit touched a file of.github/workflowsin a repository, next to the marks of the releases, so a jump in duration or in failures has its explanation beside it. It costs one request per repository, a few repositories a round and each again only after the hours asked; only the days the daily history already has are marked, and the marks are ina3.trend.workflowsofGET /api/sentinel. -
The workflow linter can read a public repository with no credential (
lint.anonymous_public, off by default;lint_workflows.py --remote OWNER/REPO --anon): to look at the workflows of a third party the files come fromraw.githubusercontent.com, which spends none of the allowance of your token, and the listing of the folder is an anonymous request; a file that did not change is not read again, a file over 256 KiB is skipped, no header with a secret is sent, an air-gapped monitor never reads and a private or missing repository says "not found" and never "clean". -
Every tab says what the wall screens do with it: the decision of the wall screens is now written tab by tab (
TAB_DECISIONSinscreens.py): Now is on Right now; Alerts, Charts and Fleet are on the overview; every sentinel tab is on Sentinels; and the tabs that are not shown say why (people,controls,readingortext). The guide of the screens lists the table in the eight languages, and a test fails when a tab is added without its decision or a guide lacks a row. -
The demo with data now fills every tab that the plain demo leaves empty:
tools/demo_monitor.py --richadds, besides the findings, the daily trend, the comparison, the book of failures and one scheduled report it already had, a saved report o...
Monitor 2.1.3 - help on every setting, a tab for Yesterday vs today, a harder Updates menu, panels that no longer overlap and guards for the guides and the tabs
Changed
- The pictures of the README are taken at 1080p and show the newest tabs:
tools/make_screenshots.pyphotographs every page at 1920x1080 by default (--viewport WIDTHxHEIGHTchanges it; the 4K, ultrawide, portrait, phone and wall-screen pictures keep their own size),tools/demo_monitor.py --richgives the demo findings for the Delivery and Supply chain sentinels, a daily trend and a comparison of periods, and the README (in the eight languages) shows the Delivery, Supply chain, Daily trend, Compare periods and Updates pages. - Yesterday vs today has a menu entry of its own: the comparison with the last saved report of the Sentinel moved out of the summary into its own tab (address
#sentinel-yesterday, next to the Daily trend), so the report of the Sentinel takes the whole page; the entry follows the CI sentinel, has its help, its text in the eight languages and the same look as the other tabs of the Sentinel.
Security
- The Updates menu is harder to abuse: five refused or failed attempts to update or go back inside ten minutes pause the buttons for ten minutes (the answer is
cooldown, with its text in the eight languages), the status says why an update is blocked (blocked_reason,cooldown_seconds), the progress file of the updater is read only when it is a plain file of a reasonable size (a link, a huge file or a truncated or hostile one counts as no update running), the log of the updater is closed once the process has started, and the workflow linter that feeds the Supply chain and Delivery sentinels cuts a file at 1 MiB, 20000 lines and 4000 characters a line, so an endless line or a million lines cannot hold the cycle. Tests refuse a foreign Host or Origin on every route of the menu.
Added
- A tooltip on every setting: pointing at a setting, giving it the keyboard focus or pressing and holding on a touch screen shows what it does, its range or values and its default, what happens when a switch is turned on and off, and whether it applies at once or after a restart; the settings that cost quota or add a page (the comparison of periods, the check for updates, the wall screens) also say what you will see and what it costs. The help text of every setting is checked in the eight languages and an entry for a setting that does not exist is refused.
- Two guards so that nothing grows alone: the guides must have the same headings and blocks of code in the eight languages, and a single register of the tabs of the page checks for each one its button, its name in the eight languages, its help topic, its line in the guide of the AIs, its row in the table of signals and its place in the lite edition; a new tab that is not registered, or a missing piece without a version that closes it, fails the test.
Fixed
- Folding panels of the Sentinel tabs no longer paint one over another: when the panels that stay folded filled the box, the one panel unfolded on its own was squeezed to a few pixels and its body was drawn over the panels below it. It now keeps a minimum height and the box scrolls instead, for all the Sentinel tabs. The browser test unfolds every panel of every Sentinel tab one by one, all together, by hand and in turns, wide and narrow, and compares their rectangles.
Monitor 2.1.2 - the updater as a menu, comparison of periods, tabs for delivery, supply chain and the daily trend, and shared files under the lock of the system
Fixed
- The unfolded panel of a Sentinel tab reaches the bottom of its box: in the CI, collaboration and security tabs the one panel that is unfolded was cut at 46% of the height of the window and left an empty strip below it; it now takes the rest of the box and scrolls inside when its content is taller. The rule is measured in a headless browser in the eight languages, wide and narrow (
tools/check_unfolded.py). - Pressing a counter of the header no longer leaves Now, the alerts and the other menus cut down until the page is reloaded: the counters of red workflows and of unsigned builds jump to the sentinel together with a filter of repositories, and that filter stayed on for every tab; leaving the tab they jumped to now takes it away (the chip of the filter still clears it by hand), and the browser test presses every counter, goes back to Now and walks through all the tabs looking for a cut list, an empty tab or an error of the page.
- Starting the monitor from the updater no longer opens a console window for every git and gh call: the process started after an update had no console at all, so each command it ran opened a visible one. It now starts with a hidden console that its children share, and so do the processes of the account supervisor; a test reads every place that starts a process and refuses the detached flag.
Added
- Comparison of periods (setting
compare.enabled, off by default): this period against the one before (7 or 30 days or any range, the same dates a year ago or a baseline of your own) with a filter for workdays or weekends in the time zone of the monitor. It sets the share of failed runs, the median duration, the time to resolve alerts, the mean score and its coverage, each with its change and its direction; a figure without enough data or without a baseline saysinsufficient_dataorbaseline_unavailableinstead of showing a 0 or a 100. It is read only, made from the ledger and the snapshots of the score, and asks nothing of GitHub. While it is on, the ledger of the score keeps two periods ofcompare.default_window_days. Guide COMPARE.md. - What changed between the two periods: the repositories that got worse or better, the workflows that started or stopped failing and the alerts opened, in
GET /api/compare(withformat=ndjsonandformat=csv), inmonitor_cli.py compare(exit 4 when a figure cannot be compared, 5 with--fail-onwhen one got worse) and in a section of the Charts tab. - A short comparison in the scheduled report (setting
compare.include_in_report, off by default): one line per figure against the period before, in both formats and the eight languages. - The updater is now a menu of the monitor (System > Updates, Complete edition): the page shows the installed version and the one the check found (setting
updates.check: with it off, or the monitor offline, nothing is asked of GitHub), lists the published versions with their notes when you press "Look for versions", and updates to the latest or to a chosen version after a confirmation that says what will happen: the SHA-256 of the zip against SHA256SUMS, the copy made first (with its folder), the migration of the configuration and the restart. The update runs in a process of its own: the monitor stops by itself, the update is applied, the monitor starts again and the page reconnects; if the new version does not start, the previous one is put back and started. "Go back to the version before" does the same in reverse. Only an administrator at the computer where the monitor runs can apply; one update at a time; every step goes to the log of actions; it is not offered in the Lite edition, in a container or under a service manager. The routes are/api/updates/status,progress,history,check,plan,applyandrollback; the stand-alonetools/monitor_updater.pystays as the way back when the monitor does not start and shares the same code. - The updater has two new options:
--handoverwaits for the monitor that started it to stop, applies the update and starts the monitor again (a failed start brings the previous version back and starts it), and--status-filewrites the progress to a JSON file that the page reads, also after the restart. - The Delivery and Supply chain sentinels have a tab of their own: like the CI, collaboration and security ones, each has its entry in the Sentinel group of the menu (
#sentinel-deliveryand#sentinel-supply), a panel with the open findings of each repository, the state of its cycle and what could not be read, a badge with the errors, its help topic and its texts in the eight languages. - The daily trend has a menu entry of its own and fits its frame: the small curves of the summary of the sentinel moved to their own tab (
#sentinel-trend, opened from a button of the summary, which keeps only its own panels), and every curve and figure now adapts to the width of its frame instead of overflowing or being cut, on a phone and on the wall screens too. - A prompt to set up the monitor with an AI (
docs/AI_SETUP_PROMPT.md, in the eight languages): a text to paste into an AI assistant that installs the monitor from the zip after checking its SHA-256, asks one question at a time about the account, the repositories, the sentinels, the quota, the notifications, the screens and the AI, writes only settings that exist in the template, keeps every secret out of the chat and the file, leaves everything that changes things switched off, and ends by explaining how the installation will look and how to update or remove it. It travels in the package and a test checks that each setting it names exists.
Changed
- The files that several processes write are guarded by the lock of the operating system: the log of actions, the state and the feed of the AI channel, the notification ledger, the dispatcher, the lessons, the schedule, the report state and the storage of the accounts take a lock that the system lets go when its owner dies, so a process that was killed can no longer leave one of them locked until the marker file gets old; the old marker files, if any, are inert.
- The architecture picture shows the five sentinels, the comparison of periods and the Updates menu.
Monitor 2.1.1 - the supply chain and delivery sentinels and a stricter release gate
Fixed
- The release gate accepts a green run by the count unittest itself gave: the record of a complete green run carries the number of tests that unittest says it ran (the lines of the report can be fewer when a test prints in the middle of its line), and the gate now compares that number with the total instead of the number of lines it could read, so a complete run is no longer refused for a few uncounted lines.
- A clock set back no longer keeps the webhook coverage live: an event marked ahead of the clock is not believed, so the indicator goes back to waiting instead of showing live until the clock catches up, and the savings that depend on a live webhook are not kept on by it.
Added
-
make_release.py --progresssays how far the suites are: with the flag, both the ordinary suite and the suite of the clean install print one line every 250 tests, so the gate no longer stays silent for the better part of an hour. -
The Supply chain sentinel (Complete only, off by default): the fourth sentinel reads, without writing anything to GitHub, the workflow files of each repository through the API (no clone; a file whose SHA did not change is not read again) and raises one alert per repository and signal:
actions.supply.unpinned(auses:not pinned to a full commit),actions.supply.mutable_ref(it follows@mainor@master),actions.supply.permissions_wide(write-all),actions.supply.permissions_missing(nopermissions:block) andactions.supply.pr_target_checkout(pull_request_targetwith a checkout of the pull request). The findings come from the workflow linter (no rules of its own, the rule that looks for secrets stays off so the value of a secret is never read), each signal has its own switch and level, a repository whose workflows cannot be read is listed underunavailableand never counted as clean, and a restart or the first look announces nothing. The cost is bounded bymax_repos_per_tick,max_calls_per_tickandmin_rate_remaining. New:sentinels.supply.*settings in the Sentinel sub-tab,GET /api/sentinel/supply,alerts --sentinel supply, the metricssupply_*, anddocs/SENTINEL_SUPPLY.mdin the 8 languages. -
The Delivery sentinel (Complete only, off by default): the fifth sentinel reads, without writing anything to GitHub, the latest releases, the latest tags, the CHANGELOG and the version files you name of each repository through the API (no clone) and raises one alert per repository and signal:
release.unreleased_pending(entries wait under[Unreleased]while the latest published release is older thanunreleased_days),release.tag_without_release(a version tag newer than the latest release has none of its own),release.release_without_notes(the latest release has an empty description),release.version_mismatch(the version files you list and the CHANGELOG do not say the same version),release.ecosystem_drift(a repository pins an older version than the latest release of the provider of a shared contract, read from the pin tables of the readiness settings) andrelease.assets_missing(the latest release lacks a file for one of theexpected_assetspatterns). Each signal has its own switch and level, a signal that is off is not read, a repository that cannot be read is listed underunavailableand never counted as clean, a restart or the first look announces nothing, and the score does not use it. The cost is bounded bymax_repos_per_tick,max_calls_per_tickandmin_rate_remaining. New:sentinels.delivery.*settings in the Sentinel sub-tab, the eventsdelivery_newanddelivery_fixed(family releases, off in every default list),GET /api/sentinel/delivery,alerts --sentinel delivery, the metricsdelivery_*, anddocs/SENTINEL_DELIVERY.mdin the 8 languages.
Changed
- The screenshots of the README show the current version: every picture of the page, the wall screens and the console tools was taken again with invented data. New ones show the menu in titled groups, the sentinels of CI and collaboration, the fleet, the reports, the whole page in another theme and the wall screens with the logo of an organization, in other themes and turning by itself; the About panel is taller so that the operation and backup card is visible.
- An address that asks for no first-run wizard also hides the one-time notice about the budget of the AI, so that a picture taken from a clean browser shows the page at rest.
- The doctor says in red that the incoming webhook has no secret: with the receiver on and no shared secret set, every delivery is refused with
401and the doctor now reports a failure instead of a warning; the webhook guide says that the secret is the only barrier. - The release gate compares the content of each file of the zip with the commit: besides the names and a clean tree, every member is compared with the file the commit versions (line endings aside), so a file that differs from the commit cannot travel.
Monitor 2.1.0 - wall screens with theme, logo and rotation, operation and backup, and the updater
Fixed
- The webhook through the supervisor of the accounts answered 400 to every delivery: the process of an account, when the supervisor forwards a delivery to it, could not tell who sent it because a standard module was used without being imported, so a valid signature, a wrong one, a missing one, a repeated delivery and a large body all got the same generic answer. The sender is read again as before (a valid key of the supervisor gives the address it reports, any other falls back to the peer), and a test now looks through the whole package for a standard module used in a function without being imported.
- The status label of a detail panel stays on one line: in the Sentinel tab the label (for example the one that says something must be done) in the heading of each panel of Detail by layer was squeezed to one letter per line, in every language and especially on a phone. It no longer shrinks or breaks inside a word, and a test in a real browser measures it in the eight languages, wide and narrow.
- The container guide tells how to open the first session and which image tags exist: it asked to open the address with the access token, which the monitor has refused by default since the access token stopped travelling in the address. It now says to send the token with a POST or the access header (or to turn the address on for one visit and off again), and the tags it quotes are those of this version.
Added
- A complete backup of the installation, verified and optionally sealed (
backup create):monitor_cli.py backup createwrites one package (format 2) with the dump of every account, the cleaned settings and lists of the installation and a manifest with the size and the SHA-256 of every member; it is verified before it is put in place and a package of the first format still verifies and restores. Tokens, keys, passwords, the value of any*_envvariable, the webhook secret, the sessions and the content of the secret store never travel: the settings go through a white list and then through a scrub of anything that looks like a secret. Withops.backup.encrypt(or--encrypt) the package is sealed with AES-256-GCM using a key from the variable named byops.backup.key_env; it needs the optional librarycryptographyand, without it or without a key of 12 characters or more, the copy is refused with a clear message and nothing is written.backup verify(with--deepit restores into a temporary folder),backup listandbackup statuscomplete it. - Restore by account or into a folder, with a checklist (
backup restore): a restore verifies the package first, is refused while a monitor answers on its port and into a storage that already holds accounts unless--force, merges the settings (a key the package does not carry is never removed), can restore one account alone (--account) or an empty folder (--to-dir), and writes a report with what is left to do (the login of GitHub, the settings left out, the restart, the doctor). - A scheduled copy, a status and a card in the About panel (
ops.backup.enabled): off by default, the monitor makes a backup everyops.backup.interval_hoursin a thread of its own, keeps the newestops.backup.keepand records the result; a failure is written down and never stops the monitor.GET /api/ops/statusandPOST /api/ops/backup(an administrator, from this computer) and the Operation card of the About panel show and run it.deploy/github-actions-monitor-backup.servicewith its timer make the same copy under systemd, anddeploy/compose.yamlkeeps the copies in a volume of their own. The guidedocs/OPS.mdexplains it all in eight languages. - The doctor looks at the operation and exits with codes that do not change (
doctor --strict): the Complete edition checks that the data, logs and backup folders really accept a file, that a running monitor and its supervisor are healthy, the webhook secret by length only, the integrity of every SQLite database read only and with a time limit (quick_checkby default), the age of the last copy, the free disk and/metrics. It exits0when all is well,1when a check failed,3for warnings with--strict(orops.doctor.strict_exit) and4when it could not run. The lite edition does not run or load any of it. - A purge on demand that does not exist until you switch it on (
ops.purge.enabled):POST /api/ops/purgeand the button of the About panel run the retention purge now, are a dry run unless asked, and are refused without a good backup newer thanops.purge.require_backup_hours; the line goes to the audit of the retention and to the log of actions. - A backup package of one account (
backup create --account): it holds that account alone, with no user, no session and none of the settings or lists of the installation;backup verifychecks that it holds that account alone andbackup restorerestores it without touching the other accounts, the users or the files of the destination, and is refused only when that account already exists there and there is no--force. - The whole product wears the theme chosen in Settings, and each wall screen can have its own: the theme (
ui.theme: auto, light, dark or one of the sixteen) is also kept by the monitor, so the wall screens and the sign-in page wear it. The screens inherit it by default (screens.default_themeis nowinherit), andscreens.default_theme,screens.theme_overview,screens.theme_nowandscreens.theme_sentinelsfix another one for all of them or for one. No wall screen writes a colour of its own any more: they use the palettes of the page, so the sixteen themes work there too. - A logo of the organization on the wall screens: uploaded in Settings (PNG, JPEG, WebP or SVG, at most 1 MB; the type is checked by the first bytes of the file and an SVG with scripts, event handlers or links to other files is refused) and kept in the data folder, never in the configuration.
screens.logo_overview,screens.logo_nowandscreens.logo_sentinelschoose the screens (only the overview by default),screens.logo_cornerthe corner (top-right by default) andscreens.logo_max_widthandscreens.logo_max_heightthe largest size (160 x 64 pixels by default). The picture is fitted without changing its shape and has a place of its own, so it never covers the content. - One wall screen that rotates and jumps to a new red (
screens.mode= rotate): instead of three separate screens, one window turns through overview, right now and sentinels everyscreens.rotate_every_secondsseconds (20 by default, 5 to 300). A new red jumps to the screen that shows it (Right now for a workflow that failed, Sentinels for a panel in red) and staysscreens.alert_hold_secondsseconds (30 by default, 5 to 600) or until the red is gone, whichever comes first, and then goes on turning; a red already shown does not make it jump again until it changes or comes back.?jump=offand?rotate=0calm one window. The default stays three separate screens. - The overview wall screen shows the map of the fleet by groups: under the four tiles of the fleet, one bar for each group of the Fleet view (the same grouping: ecosystem, label or state, from
fleet.group_by) with how many of its repositories are red, need attention, are in order or have no data. It carries only the name of the group and numbers, never the name of a repository, and follows the theme and the language of the screen. Settingscreens.group_map(on by default). - A failing "Dependabot Updates" is an alert of its own, and the activity of Copilot is information: the failed run of the dependency updates that GitHub makes for itself becomes an alert with medium severity, one per repository (
ci.dependabot_failed, switchci.dependabot_alerts, on by default), and a failed Copilot workflow can become a notice (ci.copilot_activity, switchci.copilot_alerts, off by default). Neither is ever a red of the repository, both can be silenced like any alert. - The beep of this computer has the kinds of event of the page, and the live view that comes back has its own sound: with
sound.server_beepthe monitor now sounds one pattern for each kind (a new red, a lost connection, a warning, a solved one, a security alert, a stuck run, a drop of the score, a notice, a report and a lesson) by the same switches and quiet hours as the page, in one burst of at most three. A new sound, four rising notes, tells that the connection is back after a loss, in the page, in the wall screens and on this computer. Settingsound.on_live_recovered(on by default). - An updater that moves an installation to any published version and migrates its configuration (
monitor_updater.py): one file with only the standard library, published next to every release (with its SHA-256) and also inside the product (tools/monitor_updater.py) and as the commandmonitor_cli.py update. It detects the installation, lists every version published on GitHub, lets you choose the target, shows a dry run of what changes in your configuration, applies it with a copy first, the SHA-256 checked, a safe unpacking, your own launcher lines kept and an automatic return if anything fails, and keeps a history. It opens as a page in the browser, a window (Tkinter) or a text menu. The registry of migrations (assets/migrations/, built bytools/make_migrations.py) covers every release from 1.8.1. Seedocs/UPDATE.md.
Changed
tools/backup_storage.pywrites the package of the product: its backups are now the package ofbackup create(format 2, the data of the accounts with their settings cleaned of secrets) and it takes--account; the backups of the first format still verify and restore with it.- The release gate says WHICH test failed, keeps the whole output and can accept a green run it has already seen: a failed suite (the ordinary one or the one of the extrac...
Monitor 2.0.9 - security hardening, menu groups, sounds per event and read-only chat commands
Security
- The access token no longer travels in the address by default (
listen.query_accessis now off): in the network mode the?access=address opened the session, but the token stayed in the history, the logs and theReferer. The setting is now off unless you turn it on (the default, the example file and the Settings help say so); the session is opened byPOST /api/access(the token in the body) or by the headerX-Monitor-Access. An installation that wrotelisten.query_accessitself keeps its value; one that relied on the old default and opens the monitor from an address with?access=has to turn the setting on, or usePOST /api/access. The environment variableMONITOR_QUERY_ACCESS=1turns it on for a service. - A silence by pattern of a repository no longer silences what is not a workflow: a pattern with only a repository (
alerts.silences) kept quiet the notice of a new security alert (secret scanning, CodeQL, Dependabot), of a required check in red, of a pending signature and of a release notice of that repository, and the Alerts tab did not show them as silenced. A silence now covers only the red, fixed, stuck and failed notices of a workflow, the channels and the tab share the same list, and the help says so. - The cleaner of text that leaves the monitor catches more forms of a real log and does not touch what is not a secret: Windows paths written with doubled bars (how an error message or JSON writes them), names with
_PASS,docker login -p,sshpass -p,curl -u,.netrc,.pgpass,_auth=andX-Auth, passwords in Spanish, tokens of Hugging Face, Docker Hub, Stripe and Google OAuth, Telegram tokens without the word bot, passwords with an at sign in an address, session identifiers in a query, ntfy topics, hidden e-mail addresses and names that only exist in a private network are now cleaned, and a secret that only shows once the text is plain is cleaned even when the text also has a decoy. The counters, states and names that look like a secret (max_tokens,secret_scanning: disabled,cache-key, a repository called secret-santa, a version of four numbers) are no longer replaced, and two keys that look the same once cleaned are both kept. - A header that carries a secret never follows a redirect: the token of an ntfy topic on a server that is not this computer and the local hook no longer send it on to the address a redirect points to; the redirect ends as an error.
- What a webhook delivery may say is checked: the status, the conclusion and the dates must have the shapes GitHub uses, and the names of the workflow, the branch, the job and the steps are kept in one line and cleaned like any text that leaves the monitor. A date that is not a date can no longer stop a cycle.
- A provider on the internet only enters the chain of fallbacks if you allow it: the text of a failed build is no longer sent to a cloud provider written in
ai.fallbackunless its name is inai.fallback_cloud_allowed(empty by default). Every time the chain goes from this computer to the internet it is written in the log of actions and a notice is sent. - An https address of an AI provider on a private address needs its own switch (
ai.allow_private_targets, off by default): an address that points to a private, local-network or metadata address is refused, as is one with a user or an at sign before the host. If you use a server of your own network, turn the switch on. - A webhook payload has to name its owner, and its link has to point to its repository: a delivery without the full name of the repository is no longer believed, and the link of the run that goes to the notices, the live list and the stream is kept only when it is under that owner and repository.
Added
logs.failed_step.max_per_minutelimits how many different runs the log of the failed step reads from GitHub in a minute (10 by default): asking for many runs one after another no longer spends your allowance on all of them.- Silences by pattern can expire, and a signal lists the ones that need a decision (
alerts.silences_expire_days, 0 = never by default): with a number, a silence stops working that many days after its review date; the reminders already made are kept in the state file, so a restart does not repeat them; and the signal of the silences lists the ones past their date, expired or without a date. - The reds that aged into inactive are listed in the view of what is happening now: a red that nothing of the events that decide has touched for the number of days of the age (30 by default) still leaves the red list, but now a line says how many there are and a list of them (repository, workflow and date of the failure) opens from it, so a failure never disappears without a trace.
- The inactive workflows are listed in the view of what is happening now as well: a workflow that failed and has not run for the days of ci.inactive_days (90 by default) no longer counts as red, and a second line says how many there are and opens their list, next to the one for the aged reds.
- The page says once that the budget of the AI changed: the first time this version is opened a bar explains that the daily limit and the limits of a model now also count the active notice; "Understood" hides it and the page remembers it.
- The menu on the left has titled groups, and Team and Sentinel are split into one entry each: five groups (Watch, Sentinels, Team, Reports and learning, System; they fold and are remembered) classify every tab; Team becomes Team health (
#team-health), Fleet (#fleet) and Scheduled reports (#team-reports), and the Sentinel tab becomes a Summary (#sentinel) plus one entry for the CI sentinel (#sentinel-ci, with the panels A1 to A13), the Collab sentinel (#sentinel-collab) and the Security sentinel (#sentinel-security). The entry of a sentinel is in the menu only while that sentinel is on (sentinel.enabled,sentinels.collab.enabled,sentinels.security.enabled); its address still opens a panel that says it is off and opens its switch in Settings; each badge counts only the alerts of its own sentinel. The old addresses (#team,#sent,#team/fleet,?view=team) still open the right panel, each tab opens its own help topic, andui.screen1_tabslists the new tabs (fleet,trep,sci,scol,ssec). - One sound for each kind of event, with a theme, switches per kind and quiet hours: a new red, a serious or security alert, a stuck build or queue, a drop of the score, a lost connection, a warning, something solved, a new report, a new lesson and a plain notice each have their own pattern of tones, made with the Web Audio API (no audio files, nothing leaves the browser), so you can tell by ear what happened. The settings
sound.on_warning,sound.on_resolved,sound.on_report,sound.on_lesson,sound.on_sentinel,sound.on_stuck,sound.on_score_dropandsound.on_infojoin the two that existed (all on by default),sound.themechoosesclassicorsofttones, andsound.quiet_enabledwithsound.quiet_from_hourandsound.quiet_to_hourset quiet hours (off by default). In Settings, Interface, a button per kind plays it with the volume and theme in use. The wall screens use the same sounds (screens.soundand?sound=onstill decide whether a window sounds).docs/SOUNDS.mdhas the table of sounds in eight languages. - Read only commands from Slack and Discord (setting
chatops.enabled, off by default): people in an allowed channel can ask the monitor forhelp,status,alerts,red,queueorscoreand get a short answer that only they see, with no mention and, whenchatops.linksis on (the default), no link but the address of a repository next to a name shown as it is.chatops.show_repo_namesisallby default (private repositories included): the names reach every channel and user your lists allow, which is why the lists are required and must all match;publicandnonehide them. The list of commands is closed (chatops.allowed_wordscan only narrow it) and every command only reads what the monitor already holds: none writes on GitHub, runs an action, changes a setting or reads a secret.POST /api/chatops/slackchecks the signaturev0of Slack (HMAC-SHA256 of the time and the body, at most five minutes old, the secret only in the environment variable named bychatops.slack.signing_secret_env);POST /api/chatops/discordchecks the Ed25519 signature of the interaction with the public key named bychatops.discord.public_key_env; both answer the challenge of the platform, never accept a signature twice, refuse a body over 16 KB and count the requests per address, user, channel and minute. The listschatops.allowed_teams,chatops.allowed_channelsandchatops.allowed_usersare required (with none, no command runs), every outside text is cleaned, cut and escaped, and each command leaves a line in the log of actions with the ids of the chat and never the text. A block in the Webhook tab of Settings, a guide in eight languages (docs/CHATOPS.md) and tests with the vectors of RFC 8032.
Changed
- The line of a cascade in the hook is rebuilt field by field and the name of a workflow it hands to an AI is cut to 40 characters, like the alerts of their own: a name written by anybody with write access can no longer carry a sentence through the list of the followers.
- The integrated AI cleans its answers more and believes less of what a provider says about itself: links of any active scheme, HTML tags and terminal codes are removed from an answer; the tokens a provider says it spent are believed up to ten million (a mistake no longer shuts the AI for the day); the table of data per model keeps at most 500 rows, the ones with figures typed by hand first, and refuses a provider label that is not a host name or a preset.
- **
ai.base_urlrefuses an address with a u...
Monitor 2.0.8 - a safer webhook secret, check runs and a guide to set the webhook up
Security
- The shared secret of the incoming webhook only travels in the environment (
webhook.secret_env,MONITOR_WEBHOOK_SECRETby default): it was a text inactions_monitor.json, where a backup, a copied configuration or a shared screen could leak it. The receiver now reads it from the environment variable named bywebhook.secret_env(or from the file named by that variable plus_FILE, the way a container hands over a secret), the monitor takes it out of its own environment once it has read it so the programs it starts do not inherit it, and a secret written in the file is ignored. The old way can be kept for a migration withwebhook.allow_config_secret(file only, off by default) and the Live view warns while it is used. The Settings tab shows the name of the variable and whether the secret is defined, never the secret.
Added
- The webhook also takes
check_runevents, and the Live view says how much of the time the events cover the state:check_runjoins the closed list of events (workflow_run,workflow_job,check_run,push,release,check_suiteandping); like the others it only marks the repository for a read and keeps no text. The Live view and/api/live-status(newcoverageandsecret_source) say whether the state is covered by events (the last one is newer thansavings.webhook_window_minutes), has had no event for a while (a quiet repository or a lost tunnel) or waits for the first one. The polling stays the fallback, so a delivery that never arrived is caught by the next read. - A guide to set the webhook up in GitHub, in eight languages:
docs/WEBHOOK.mdlists, step by step, how to make the secret, how to give it to the monitor in PowerShell, a shell, a service or a container, what to fill in on the webhook page of GitHub (address, content type, secret, events) and what each answer (404,401,413,400,202and429) means.
Changed
- The paid runners of another company that the model recognises by default include
codspeed(runners.model.thirdparty_labels): public repositories use it for their benchmarks, and its jobs were counted as a runner nobody knew. A list you wrote yourself is not touched.
Fixed
- The documents no longer say that macOS runs in the CI: the README and its seven translations said that the tests run in the CI on three systems and that Linux and macOS are tested in the CI; since the CI tests on Ubuntu and Windows only (macOS will be taken up again later), they now say so.
- A workflow whose file was deleted is not one of the workflows you can enable: GitHub keeps the history of a deleted workflow in the list with an empty path (six of the 96 public repositories studied have one), and the hygiene check offered it as a workflow disabled by hand. It now stays out of the lists that need a file.
- A fork that was never pushed is not "recently pushed": a fork inherits the push date of its upstream when it is made (693 of the 722 forks of an account measured have it before their own creation), so the sentinel took every new fork for the most active repository of the account and read it as a hot one. The push date of a fork is now ignored while it is not newer than its creation.
Monitor 2.0.7 - a read API for clients, the original repository of each fork and workflows switched off by hand
Added
- A read API under
/api/v1/for clients (settingapi.v1_enabled, off by default): every read route is also served under the prefix (the same answer and roles, plus the headerX-API-Version: 1), and three routes are new:/api/v1/me(who is asking: the mode, the user, the role and the accounts of the team session),/api/v1/repos(the watched repositories, one line each, in pages with a signed cursor,fields=andprofile=compact, gzip when the client accepts it) and/api/v1/openapi.json. The new routes answer errors in one envelope (error, a stablecodesuch asapi.bad_cursorand amessagein the language asked). Everything is read only, from what the monitor already holds. The shape is frozen in the new sectionv1_endpointsof the contract (python tools/check_api_contract.py --v1); the limits are the settingsapi.default_limit,api.max_limit,api.cursor_ttl_secondsandapi.gzip_min_bytes. Texts in eight languages, the guide of the API updated in all of them and a test with the real server. - A workflow the owner switched off by hand is no longer a red (setting
ci.disabled_apart, on by default): a fork copies the nightly job of the project it follows, the job fails for lack of the secrets, the owner switches the workflow off, and its last run stayed "failure" for ever (two accounts of the study showed it). Now a workflow in the statedisabled_manuallyleaves the list of reds and is counted apart in the panel ("switched off by hand"). A workflow GitHub switched off for inactivity is not touched, because that is not a decision of the owner. Text in eight languages and a test. - The original repository of each fork can be followed (setting
upstream.enabled, off by default): for an account that collects forks the useful question is what happens in the projects it follows, not whether its own CI is green. With the switch on, the monitor reads the original of each fork (the status of the checks of the last commit of its default branch and its latest release) and shows them in the hygiene panel of the Sentinel; withupstream.notifyit also gives an information notice when the original is red or published a release. The cost is bounded by settings in the configuration file:upstream.requests_per_hour(6),upstream.forks_per_request(8),upstream.refresh_hours(12),upstream.news_days(7) andupstream.repos(the forks to follow; empty means all). One request reads several forks, the least recently read go first, and nothing is written anywhere. Texts in eight languages and a test with the cases of the study.
Monitor 2.0.6 - three wall screens, drawn and off by default
Added
- The three wall screens are drawn: Overview (score ring and letter, fleet lamps, 24-hour and 7-day area charts, a heat map by repository and hour, what has been red the longest, runners and minutes, the latest events and the allowance as a gauge), Right now (running runs with progress, the queue, runs waiting for approval, recent failures, merges and releases, and a live line of the last two hours) and Sentinels (one card for each sentinel with its level, open findings, readable or not and the last sweep, an animated radar and the alerts by severity). The page draws every chart itself (no library), counters and rings animate, a new failure flashes once and old ones calm down (
screens.blink_calm_minutes), nothing moves for who asked for less motion, a monitor that does not answer dims the page and is tried again after longer and longer waits, everything shifts a few pixels every minute against burn-in (screens.burn_in_protection), the page can dim at night (screens.night_dim) and one window can turn through the three pages with?rotate=30. The Settings tab opens them with a link, and a browser test checks the three pages in 1080p, 4K and turned, in the eight languages, with and without less motion.