Repository navigation
v1.4.0
Added
-
Every release and every nightly carries the reporter as a zip (#835, from #408). Claude Desktop's Settings > Plugins takes only an archive, so
podium-reporter-<version>.zipis attached to the GitHub Release of eachvX.Y.Ztag (newrelease.yml), andpodium-reporter-<version>-nightly-<date>.zipto the nightly prerelease.node scripts/podium-reporter-zip.mjsbuilds it, the same way in CI and by hand, with.claude-plugin/plugin.jsonandhooks/hooks.jsonat its root and no tests. Releases are now published as the docs site and this archive (.claude/PROJECT.md,podium-release). -
How to install the reporter in the Claude desktop app on Windows 11 (#836, from #408).
docs/REPORTER.mdand the user guide's "Set it up" now say that the terminal commands do not work in the desktop app and that uploadingpodium-reporter-<version>.zipin Settings > Plugins is the only way, then walk through Node.js andPODIUM_NODE, the upload, the Settings fields and the/podiumcheck, and name what does not work on Windows. The README links to it. Documentation only. -
The word Shortcuts opens a list of every keyboard shortcut (#828, from #636). While the Shortcuts switch is on, "Shortcuts" in the Display settings row is a link, on the wall, the Board, the orchestrator page, Traffic, My reporters and History. It opens "Keyboard shortcuts":
\,[and]in one column andwandEscin the other, split by a thin line, with a note where a key does nothing ([and]on Traffic,won Traffic and My reporters). Under 480 px the columns stack. The X,Escand a click outside close it, and focus returns to the word. Off, the word stays a plain label. One list now feeds this modal and the switch's hint, so they cannot disagree. SPEC 8.1. Mockupmockup/podium-shortcuts-help.html.
Standards: the dotted underline now means two things: a help term (hover, help cursor, as before) and an action (a button, pointer cursor), the new.kbd-link; and a new pattern,.kbd-cols, two columns split by a 1px--borderpipe that stack under 480 px with a rule; approved by the maintainer on #828. -
/podiumsays why messaging is off, how the reporter signs in, and names a mistyped option (#825). AMessaging:line sayson, oroff, because …with the one cause (the session ended, thepeersoption is off, the session is not interactive, no token with auserNameor reporter config file, no report yet, or Podium's flag is off or the admin secret is in use), then thepeersoption as the mod read it, whether the session is interactive and the surface it started on, and what Podium last said. ASign-in:line says whether the reporter uses a plain token or a reporter config file, and which file. AnOption problem:line names a switch whose value is nottrueorfalse; the mod still does not read"true"as true, so a stray text never turns on a consent option. Thepodium_peers,podium_sendandpodium_request_accessrefusals give the same cause. No token or secret is shown. SPEC 5.3. -
End-to-end tests for every feature of the console, and an agent and a skill that keep them true (#782). A Cypress suite in
e2e/drives the built console in a real browser: the wall, the orchestrator page and its file diffs, the Board, History, Traffic, My reporters, sign-in, Respond, session recording and every Admin page, each with its empty, error, signed-out, no-access and flag-off states, live updates with no reload, reconnecting after the server stops, light and dark, and axe (with color-contrast and target-size) at 1440 and 400 px.npm run e2eruns it in containers only: Podium's own image as the server on a private network, and the pinnedcypress/includedimage as the runner, on Podman or Docker, with a fixed-data seed (scripts/e2e-seed.mjs) and nothing installed on your machine. A newE2Eworkflow is there to run it on GitHub, but it runs only on the release PR intomainand on a manual dispatch (#815); day to day it runs locally.docs/e2e/COVERAGE.mdmaps every feature to its specs and says what no end-to-end spec can reach;npm testfails when it names a spec, file, route or test id that no longer exists. Thepodium-e2e-updateskill and thepodium-e2e-keeperagent update the specs on any feature change, and the suite joins the release gate (.claude/PROJECT.md). The only app change is adata-testidon the page pill and on the connection pill (the latter with adata-state), which changes nothing a person sees. -
End-to-end specs for the Admin idle lock, the sign-in idle end and the lockout, with test-only timing variables (#791, from #782). These three states had times fixed in the code (10 minutes, 12 hours, 15 minutes), so no browser test could reach them.
PODIUM_TEST_ADMIN_IDLE_MS,PODIUM_TEST_SIGN_IN_IDLE_MSandPODIUM_TEST_LOCK_MSshorten them for a test server, and only whenPODIUM_TEST_TIMINGS=1is set; a timing without the flag, or a bad value, stops the server at start, and when one is active the server logs aserver.test_timingswarning. Nothing a person sees changes and the defaults stay as they were. The three new specs (admin/idle-lock.cy.js,admin/lockout.cy.js,respond/sign-in-idle.cy.js) restart the throwaway server with them, and the gap is gone fromdocs/e2e/COVERAGE.md. SPEC 7.7, DECISIONS 50. -
The end-to-end suite no longer waits on a stale connection after a server restart, and covers the Comp pill (#810, from #782). A restart killed and recreated the server container; the proxy inside the Cypress container kept pooled connections to the old one, and on a GitHub runner the first request after each restart hung for 30 s. The server container now lives for the whole run and a restart replaces only its process (same address, emptied data), the script checks the new server from inside the Cypress container,
E2E_CPUSandE2E_MEMORYlimit the containers to try a slow machine, and the "folder is not writable" warning is gone. A new spec,wall/comp-pill.cy.js, checks the pill (#805), seeded with--compaction. -
A Comp pill on the wall card when Claude Code compacted a session on its own (#805, from #803). Since #803 an automatic compaction below 80% fill is a
compactevent, and nothing on the card said it happened. The card's Context cell now shows a small blue pill, a compress mark and the word Comp, after the blowout badge, and the orchestrator panel shows it after the Context readout. Its title reads "Claude Code compacted this session automatically N times". The Context cell's label row now wraps, so in the narrow cell the pill drops to its own line instead of being clipped. The count comes from the session's stored automaticcompactevents, so the report, the database and the reporter's version do not change; a reporter older than #803 sends those compactions as blowouts and shows no pill. The pill stays for the life of the session's events. SPEC 7.3, 7.4, 8.1.
Standards: a new component,.comp(the Comp pill,CompBadge): the.blowbadge's footprint in--delegand--deleg-text, with the tapered-lines icon, and the card's Context label row wraps; approved with the mockup on #805 (mockup/podium-compaction-pill.html). -
A performance audit of
development, with the scripts to repeat it (#682).docs/PERFORMANCE.mdrecords what the server, the console and the reporter cost at an idle, a small, a medium and a large fleet, and ranks fifteen findings, each filed as its own issue.scripts/podium-perf.mjsmeasures the server (CPU, memory, event-loop delay, stream bytes),scripts/podium-perf-browser.shan open tab, andscripts/podium-perf-mod.mjsthe reporter's own code. Each starts a throwaway server and stops it. No product code changed. -
Podium contexts: switch between Podium servers the way
kubectl config use-contextdoes (#606). A context is a name for one reporter config file (so one server and one person), kept in~/.podium/contexts.jsonwith a current one. The reporter choosesPODIUM_REPORTER_CONFIG, thenPODIUM_REPORTER_CONTEXT(one terminal), then the current context, then the newest file as before./podium contextandpodiumctl contextshare the verbslist,show,add <name> <file>,rename,removeanduse;addchecks the file exists and is a reporter config. A running session keeps the context it started with, even across a plugin reload, andusesays it applies from the next session, so two terminals can report to two servers at once. The status line shows[name]after the version and/podiumshowsContext:. A context that cannot be used shows[context unusable]in the status line andContext problem:in/podium, and the reporter uses the newest file.podiumctl contextalways writes~/.podium/contexts.json, where the reporter reads it. With no contexts nothing changes. SPEC 5.9, DECISIONS 46. -
SPEC 8.5 describes the History page (#792).
#/historyhad no section of its own: its day picker, table, paging, address, opened session, empty and error states, sign-in and narrow layout were spread over 7.1 and a few lines of 8.1. The new page 8.5 states them from the code and its tests, and 7.1 and 8.1 link to it. SPEC 8.1 now says an opened History session shows Resume and Fork, as the page does; it said Fork was not there. Documentation only.
Changed
-
The console says where to turn on a reporter option in the Claude desktop app, not only in the terminal (#837, from #408). My reporters, Admin > Reporters > Messaging, the admin-secret fix in Admin > Reporters, the recording picker and the Activity section now name both routes:
/plugin configure podium-reporter@eengelking-marketplace(the full key, where two of them had left out@eengelking-marketplace) and Settings > Plugins > Podium-Reporter > Settings. A session the reporter callsnot-interactiveis no longer told it "runs withclaude -p": the console cannot tell aclaude -prun from a Desktop session no window has opened yet, so its note says it is either. The user guide,docs/MESSAGING.md, a gotcha, SPEC 6.8 and 8.1 and the e2e map say the same. No standard is changed here:docs/standards/shows the old My reporters note until the release-prep PR. -
The README screenshot shows the console as it is today (#827).
docs/images/podium-chaos.pngis retaken from the current build at 1914 px wide, on a throwaway server fed by the simulator with 12 sessions, chaos on and nested agents, so it shows only simulated data. The alt text is updated to match. -
The
E2Eworkflow runs on the self-hosted runners (#815, from #782). All three jobs use the sameruns-onasci.yml([self-hosted, Linux, X64],ubuntu-latestfor a fork PR), and the e2e job takes a host-wide concurrency group so two runs never share the runners' Podman host, because the cleanup step removes every e2e container on it. It now also runs on a pull request intomain, the release PR, and still not on a PR intodevelopment. -
Decision 49 points to decision 50, the log guide lists two more server warnings, and CONTRIBUTING.md names the test-only timing variables (#819, from #791). Decision 49's "Left out" bullet ends with "Decision 50 adds it (#791)." and nothing else in the entry changed.
docs/user-guide/07-logs.mdgains rows forserver.weak_tokenandserver.csp.CONTRIBUTING.mdsays thePODIUM_TEST_*variables are for the end-to-end suite only. SPEC 8.1 still defers the lockout and sign-in-ended texts to the mockup and does not quote them; that is on purpose. -
.claude/PROJECT.mdnow opens with the storage rule (#817). A "Read This First" section and a new Do Not say that every database query goes throughStoragefromopenStorage, and that direct SQLite use (node:sqlite,sqliteHandle,activityHandle, raw SQL outside the adapter) is not allowed, because PostgreSQL, MariaDB and other databases are planned as adapters. It points to the storage page and DECISIONS 48, and namesneutral.test.ts, which already enforces it. No code changed. -
The SQLite adapter's two queries that used its JSON functions now run on any SQL database (#808).
countAutoCompactionsreads the session'scompactevents and counts the automatic ones in code, and the Activity history's bulk delete binds its ids in batches of 200 instead of passing a JSON array tojson_each. No schema change, no migration, same counts, same rows left, same speed. A new test fails if a query inserver/src/storage/sqlite/usesjson_extract,json_each,json_object,json_group_arrayor->>, so a second database has fewer queries to port (#314). One difference: a compact event whose stored data is not valid JSON is now skipped instead of making the count throw. -
scripts/promise-misuse.mjsfinds a store call withoutawait(#809). It loads a tsconfig with the TypeScript API and reports a Promise used in a condition, under!,&&,||or??, in a comparison, or left floating.// promise-ok: <why>on the line or the one before allows a deliberate case, andnpm testruns it overserver/andscripts/, tests included. Tooling only; no behavior changed. -
The server reaches its database through async storage interfaces, with SQLite as the only adapter (#313). Every store call is now a promise,
PODIUM_DBstill names a SQLite file and is still the default, andactivity.sqlitekeeps its own connection.PODIUM_DATA_DIRcan now name the folder for the TLS CA and other server files; unset, it is the database file's folder, as before. Nothing changes in the console, the API or the data on disk, and a differential run of the old and new builds over the same data found no difference. A few writes land a few ticks later than before, such as the log lines for a respond or recording event. Open asks, sign-ins and lockouts still live in one process's memory. No second database is added. DECISIONS 48, anddocs/architecture/02-components/19-storage.mdfor the rules a contributor must follow. -
The local skills the
eengelkingplugin now covers are gone (#801).podium-worktreeandpodium-ui-verifyare deleted, andpodium-pr-workflowandpodium-releasekeep only what is Podium's own: themod/checks, the session-link patterns, the version-file table, the atlas, docs-site and standards checks.scripts/page-shot.mjsand its test stay in this repo, becausescripts/podium-perf-browser.shand the docs call it by path and the plugin's copy sits in a versioned folder. -
.claude/PROJECT.mdnow says what and why, and the how moved into skills (#788). The worktree steps, the PR checks and the trace check, the post-PR check watch, closing issues after a merge, the release sequence, the version-file table, hotfixes, the release check and the page-shot steps are in four new project skills (podium-worktree,podium-pr-workflow,podium-release,podium-ui-verify), and the doc-page rules joinedpodium-docs-site.PROJECT.mdkeeps the project's purpose, every Do Not and override, and a Skills table that names the skill to load for each action, so a session no longer carries the procedures it does not need..claude/CLAUDE.mdis the current standard from theagentrepo (skill names carry the/eengelking:prefix, issues are closed after the merge, a hotfix ondevelopmentwhen it matchesmain), with this repo'sdocs/architecture/anddocs/spec/paths kept, and thePROJECT.mdoverrides the standard now states itself are gone. -
The Files tab loads every changed file's diff in the background, so a diff opens with no wait (#631). While the owner has a session's page open, the console asks for each file's diff, first-listed first and two at a time, and asks again when the file is edited again. The reporter marks each file with the git hash of its content (
rev), so an edit that leaves the line counts alone is seen, and has a newfileDiffPreloadoption, on by default whenfileDiffis on; set it to false to read a diff only when the file is opened. More of a session's source code leaves the machine, redacted, than before: a diff nobody opened is now read. The diffs live in the page's memory and are thrown away when the page is left; the server still keeps none. The admin log records a load asfile_diff_loadedand an open asfile_diff_opened. SPEC 6.4 and 7.7 now say "while the page is open" where they said "on request". The History page still shows no diffs for an archived session until they are stored (#632). -
A gotcha for an Activity section that asks you to turn on
activityFeedwhen it is on. A session reads its reporter options once, when the plugin loads, so one that started before the option was turned on keeps saying it is off until/reload-pluginsis run in it. Documentation only. -
.claude/CLAUDE.mdis now the standard baseline, and everything specific to Podium is in.claude/PROJECT.md(#772).CLAUDE.mdmatches the baseline policy file used across projects, apart from namingdocs/architecture/anddocs/spec/where the baseline names single files, and it loadsPROJECT.mdwith an@PROJECT.mdimport.PROJECT.mdholds the overview, the Podium Do Nots, the worktree rule, theCloses #Noverride, the project skills, the UI standards, the release steps and the PR trace, atlas and release checks. The rules are unchanged; they moved. Docs and comments that cited a moved section now citePROJECT.md. -
The single-key shortcuts are off until you turn them on (#636).
\,[,]andwworked anywhere outside a text field and could not be switched off, so speech input or a stray key could collapse a pane or start wall mode (WCAG 2.1.4). They are now off by default. A Shortcuts switch in the bar under the top bar (beside Refresh every and Theme, and beside Theme on History) turns them on, and the choice is kept in that browser, so each browser needs it once. While it is off the hover tips and the Wall mode button's title no longer name their keys, and every button andEscstill work. If you use these keys today, turn Shortcuts on once.
Standards: the Shortcuts field in the Display settings row (a.fieldholding the.swswitch, between Refresh every and Theme); approved with the mockup on #780 (mockup/podium-shortcuts-off.html, Option A, off by default).
Fixed
-
One session that does not take part in messaging no longer switches a whole machine to Off (#838, from #408). The server kept one
peersvalue per machine and the latest report won, so aclaude -prun (which reportsfalse) beside an interactive session made My reporters flip between On and "Off on the machine" and left the machine out ofpodium_peers. A machine now reads On while any of its live sessions reportspeers: true, and Off only when none does. A machine with no live session keeps its last value. DECISIONS 52. SPEC 5.9. -
The reporter finds its config file on Windows (#834). The mod read the home folder from
HOMEalone, which Windows normally does not set, so it found noreporter-config-*.pem, said nothing, and sent over a plain token. It now takes the home fromHOME, elseUSERPROFILE, elseHOMEDRIVEplusHOMEPATH, so the folder on Windows is%USERPROFILE%\.podium, the same name as elsewhere (DECISIONS 51)./podiumshows aHome folder:line with the folder and the variable it came from, or says that no home was found. Node has no Unix sockets on Windows and the engine refuses a named-pipe address, so there the mod writes each request as a file in.podium\run\<session>\and the helper answers on its stdout (it needs Node 20.17 or later); macOS and Linux keep the socket. The search fornodeaccepts a Windows path such asC:\Program Files\nodejs\node.exe(it took only paths starting with/) and also looks in Program Files, the per-user installer folder and Volta.REPORTER.mdnow says a Windows 11 install needs Node.js and why the Windows mailbox exists. The user guide gains a Windows 11 set-up section and troubleshooting rows (nonode, a config file that is not found, a config whose# server:line sayslocalhost, a stalled helper), and two gotchas cover theHTTP 502from alocalhostconfig and theEACCESon a.sockpath. macOS and Linux behave as before. Whether the whole TLS connection works on a Windows 11 machine is checked by hand and recorded on #834. SPEC 5.9. -
A Claude Desktop session that starts before its client attaches now takes part in orchestrator messaging (#408). Desktop on Windows starts a session with no surface, which the reporter read as a
claude -prun: it reportedpeers: false, sopodium_peers,podium_sendandpodium_request_accessnever appeared, and Respond was refused. Now the session counts as interactive as soon as a client shows on the roster (on a flush or on attach), and stays so for the session, across a latersession.startand a reload. A real-prun, which never has a client, stays out.docs/REPORTER.mdnow says the mod runs in the desktop app on macOS and Windows. The console's wording for turningpeerson in Desktop is not changed here. -
A choice answered in the terminal leaves the Respond modal as soon as the session runs again, and its card is no longer a blank circle (#829). The "No longer waiting" card for a question, plan or tool prompt that was answered in the terminal stayed until the session finished its turn and went idle, up to 10 minutes. The server now forgets the closed ask on the first report that shows the session running after it closed (spec 7.1 and the responding page say so). The card's body also showed only a small circle and the letter "A" at the right edge, because the top bar's
.infobutton rule matched the "Answered in the terminal" frame (.out.info) and shrank it to a 24px circle. That rule now names the button. -
Seven accessibility findings from the end-to-end axe check are fixed (#794, #795, #796, #797, #798, #799, #800, from #782). Five pieces of text were below the 4.5:1 contrast axe asks for: the modified-files count on a session card in the light theme, a Board card's wait line in the light theme, Traffic's sub-lines and muted text at 400 px, a request's hint on My reporters in the dark theme, and the failed-agents button on the orchestrator page. The count now reads
--warn-textlike the wait line, and three token values are stronger:--warn-textin the light Podium theme,--mutedin both, and--crit-textin both. The Recording button and the Files rows were named by anaria-labelthat did not hold the words on them, so someone saying what they see did not reach them. Both are now named by their own words with the action in front ("Recording Record this session", "Open diff of" and the file name). TheKNOWNentries for all seven are gone from the axe specs, so a new violation of any of them fails the suite.
Standards: the token values--warn-text(light Podium theme, now defined intokens.css:#8a5a00),--muted(light#626159, dark#a3a19b) and--crit-text(light#a82020, dark#f58080), and.git .dirtyreading--warn-text; approved with the mockup on #794 to #798 (mockup/podium-contrast-fixes.html), with dark--mutedraised from the mockup's#97958fto#a3a19bafter axe measured Traffic's selected row at 4.16:1, approved in chat. The Recording cell's button and the Files row are named by their content, and the standards pages for those samples change with it. -
An automatic compaction at low context fill no longer shows as a blowout (#803, from #687). Two sessions that Claude Code compacted at 22% and 25% of a 1M window, while they waited on an agent, each counted a blowout, so the card, the orchestrator page and the Blowouts KPI said "context window exceeded" for a context a quarter full. The reporter now counts an
autocompaction as a blowout only at 80% full or more, or when the fill is unknown; below that it logs acompactevent ("auto-compacted at 22%"). Awindow-exceededstop and aserver-compactionstop still count, and an auto compaction within 2 seconds of one is still the same blowout. The compaction and blowout events now carryidleMs, how long the loop was quiet, to confirm or rule out an idle timeout as the cause. Blowouts already stored are not corrected. The new gotcha "Blowouts 1x on a card that is nowhere near full" explains it. Ships in the reporter at the next release. SPEC 5.3. -
At 400 px the top bar's edge tab no longer sits on the bar's last button (#787, from #636). The tab is 24 px tall and reaches 12 px up into the bar, but the bar had 8 px of bottom padding, so on the wrapped second row it overlapped "Sign in" (or "History" when signed in) by 4 px and touched the line below. Below the 1040 px stacking width the bar now has 12 px of bottom padding. The bar at wide widths and the tab's size are unchanged.
Standards: the top bar's bottom padding is 12 px below 1040 px, 8 px above; approved with the mockup on #787 (mockup/podium-top-bar-edge-tab-gap.html). -
The reporter picks the newest config file by the epoch in its name, and reads only that file (#607). Before, every session start read every
reporter-config-*.pemin~/.podium/in full, private keys included, to compare theirissued-atlines. Now the mod orders the names by their epoch and reads one file at a time, newest first, so the key of a file it does not choose is never read. SPEC 5.9 says "by the epoch in its name" where it said "by itsissued-atheader line"; a file renamed so the two disagree is ranked by its name. -
The wall card's "Copy session id" is a real button beside the card, and big enough to hit (#759). It was a
span role="button"inside the card's own button, a control within a control that some screen readers skip or read oddly, about 19 px tall, and named "Copy session id 3f9a1c2e" while it read "Copied". It now sits next to the card's button, as the pin does, is at least 24 by 24 px, comes right after the card in the tab order, and is named by what it says. The card looks the same, but the path line is about 5 px taller. -
Every page says where it is, to a screen reader (#636, the first part of the accessibility pass). The page title now follows the route: "Podium board", "Podium traffic", "Podium my reporters", " - Podium" for an open orchestrator and "Replay of - Podium" for a replay (History and Admin already had theirs). The wall, the board and History each have one
h1that is hidden from view and read by a screen reader ("Fleet", "Board", "History"); an open orchestrator and a replay keep their own. The row under the top bar that holds the refresh and theme controls is now a landmark named "Display settings", on Admin too. Nothing changes on screen. -
Focus no longer falls to the top of the page after a sign-in (#636). After you signed in from the top bar, or unlocked Admin, the form or button you used was replaced and keyboard focus dropped to the page, so the next Tab started again at the logo. Focus now moves to "Signed in as ..." in the top bar, or to the Admin page's heading. Neither draws a focus ring, since neither is a control.
-
The Agents tree's sort buttons are no longer announced as a table row with nothing around it (#636). The header carried
role="row"with no table, so a screen reader met a row that belonged to nothing, and the cells it held were plain boxes. The header is now a group named "Sort the agents". The agent rows below it were never table rows and are unchanged. -
A control's name now holds the words it shows (#636). Voice control and screen readers match on what is on the button, and several names did not contain it. The orchestrator page's Session cell is now "Session 3f9a1c2e, copy the full id", and its Resume and Fork cells read "Resume Command" and "Fork Command" as two words. A card on the board is named by its own text, so a step that shows its live wording ("Porting endpoints") is no longer named by its plain one, and a waiting card includes its reason; its state follows as one hidden word. A Traffic line's name starts with the same words as its pill, such as "1 message, waiting".
-
A password manager can fill the sign-in forms (#636). The Admin unlock form and the Sign in dialog told the browser not to remember or fill the name and token (
autocomplete="off"), so a person who keeps their token in a password manager had to type or paste it by hand each time (WCAG 3.3.8). The name is nowusernameand the tokencurrent-password. The Admin secret is a token like any other, and Admin still locks after 10 minutes without activity. -
Three pieces of text are easier to read in every theme but Solarized (#636). The "need attention" count on the wall was amber on a light page (1.7:1 to 3.5:1 against the 4.5:1 WCAG asks for); it is now the amber mixed toward the ink. "no plan reported" on the selected rail row, and the "orchestrator" label on an agent row, were the muted grey on a highlight (3.8:1 to 4.5:1); both are now the secondary ink. No token changed. Solarized's own pairs (its secondary ink on its raised surface) are still just under 4.5:1 and wait for a decision about whether a theme may adjust its published colors.
-
Each Admin section says which one it is in the tab title (#636). Every section kept the title "Podium admin", so a person with several tabs open, or a screen reader announcing the page, could not tell People from Settings (WCAG 2.4.2). Once Admin is open the title is now "Settings - Podium admin", "People - Podium admin" and so on, and follows the section as you move between them. The lock screen keeps "Podium admin".
-
The Activity panel's times and "go to" hints are readable on a gone row and on the selected row (#636). A row whose session had gone was dimmed with
opacity, which took its time and hint to 1.4:1 to 3.2:1 against the 4.5:1 WCAG asks for, and on the selected row the muted grey was 3.8:1 to 3.9:1 on its highlight in the default dark and Solarized themes. A gone row is now dimmed by the secondary ink instead of by opacity (a failed row keeps its own color), and the selected row's time and hint use the secondary ink. No token changed. The selected row in Solarized is still just under 4.5:1 (4.47:1 light, 3.85:1 dark) and waits with the rest of Solarized. -
The "Notice turned off" confirmation waits while you are reading it (#636). It cleared itself after 9 seconds whatever you were doing, and it holds the link to Admin > Settings, so someone moving slowly, or reaching it by keyboard, could lose it before they got there (WCAG 2.2.1). It now stays while the pointer is over it or focus is inside it, and the 9 seconds start again once both have left. The wait, the look and the X are unchanged.
-
The edge tab is big enough to hit (#636). The collapse control on the top bar, the Kanban KPI rows, both side panes and the Plan / Files fold was 16 px on its short side, under the 24 px WCAG 2.5.8 asks for. It is now 32 x 24 on a horizontal edge and 24 x 32 on a vertical one, wherever it appears. It straddles its edge 12 px instead of 8, so at 400 px, where the top bar wraps, the top tab overlaps the bottom few pixels of the bar's last button; the button stays reachable.
Standards: the edge tab (.edge-tab) is 32 x 24 on a horizontal edge and 24 x 32 on a vertical one, up from 32 x 16 and 16 x 32 (#626), and straddles 12 px instead of 8; approved with the mockup on #779 (mockup/podium-edge-tab-target-size.html, Option A). -
A permission prompt reaches the card on a Team or Enterprise login, and the console says sooner that it cannot be answered from Podium (#742). Claude Code's guard
cc-plugin-sec-defaultskips the mod'sclassic.PermissionRequestthere, so the prompt never showed. The reporter now records it fromtool.checkresolvingask(waiting on a permission, with the tool and its input) once, even if the hook arrives late, and a session whose turn ended with no classic hook having run is marked as one that cannot be answered before its first prompt, not only after it. Atool.checkhook cannot answer for the person, so answering stays in the terminal; plan feedback and questions still work. See SPEC 5.3 and the gotcha. Reporter only; the console and the server do not change. -
A session waiting on a permission prompt shows as waiting on a Team or Enterprise login, even with responding off, and no longer stays waiting when the prompt was answered within two seconds (#741). Claude Code skips
classic.Notificationthere. The wait now comes from the sametool.checkfallback, which no longer depends on the responding option. Other notification types have no other source and were never used. Reporter only. -
Documented: the card cannot show a session's background tasks and scheduled wake-ups on a Team or Enterprise login (#740). The engine offers no call that lists them apart from the
classic.Stophook the guard skips, so a session idle on background work reads as waiting for you. A test pins what the card shows. See the gotcha anddocs/REPORTER.md. -
A session in the Claude desktop app is named after the title the app shows for it (#595). Every desktop session in one folder showed that folder's name, because the app's titles never reached the
session_titlethe reporter reads from theclassic.SessionStartandclassic.UserPromptSubmithooks. On a Team or Enterprise login, or a machine with managed settings, Claude Code's own guard skips everyclassic.*hook of a user-installed mod, so no title ever arrives. The reporter now also reads thenameof the session's entry inclaude agents --jsonon the first git poll and about every 30 seconds after, and uses it when no hook has given a title; a hook's title and/podium namestill win (SPEC 6.6). The same skip also stops the otherclassic.*hooks, so on those logins background tasks and crons (classic.Stop), notification signals and answering permission prompts from Podium are still not reported; the gotcha "My session's name is the folder, and background tasks, notifications and permission answers are missing" says so. -
A failed TLS handshake on the reporter port says why and where from, and a symlink can no longer redirect the reporter's config write (#672). The server log named only
socket hang up; it now gives Node's error code and the short OpenSSL reason (ERR_SSL_PEER_DID_NOT_RETURN_A_CERTIFICATE: peer did not return a certificate), words repeats alike so the text log folds them, and keeps the source address as a field. The reporter helper creates its temporary file withwx, and rewrites a config file that is a symlink at the file it points to, so the link stays a link. A refused client certificate that reads as a reset is written up in the gotchas. A 403 from the connection helper (ROUTE_NOT_ALLOWED) is no longer retried every few seconds like an outage; a 403 from the server is handled as before.config unreadableon the status line now carries the reason, such as a config holding more than one person certificate. Pressing Escape on the People confirmation row now closes it, and Escape or Cancel on it and on the Rotate CA and Retire old CA confirmations returns focus to the button that opened them. -
A session the Claude desktop app keeps open no longer leaves the wall because another session is on screen (#594). The desktop app attaches a client only to the session in view, so every other open session read as "nobody attached" and left the wall a minute later behind "background sessions hidden". The reporter now counts a session as attached once a desktop client has been on it, until it ends. Deleting a session from the app's sidebar ends it, and its card leaves the wall as before; a terminal session with nobody on it is still a background session.
-
Four tests that timed out at 5,000 ms on CI runners no longer do (#680, #681). The activity history "compact in steps" test ran about 175 one-step vacuums on a file database, so a slow disk took it past the limit; it now runs a dozen steps over a smaller file and still crosses several delete chunks. The side, history and Admin delete-people tests render a large tree and are slow under coverage, so each has a 30 s limit that matches the work. No test code outside those four tests and no production code changed.
Changed
-
npm testnow runs an accessibility check on the console's pages and the published standards pages (#636).frontend/src/fleet/a11y.axe.test.tsxruns axe-core over the wall, an orchestrator's page, the board, Traffic, My reporters, History, the Admin lock screen and every standards page butcomponents.html, and fails on any violation that is not listed as known with its reason. jsdom has no layout, so contrast and target size are left to the manual pass. -
The server prints the reporter port and the CA backup reminder at startup, and the "renewal due" badge follows the reporter's rule (#672).
podium: made a CA in …; back the folder upandPodium HTTPS listening on … for reporters (CA …, expires …)print at the default log level, sopodiumctl runandnpm startsay where reporters connect. A person's certificate shows renewal due when 30 days or half its lifetime remain, whichever is shorter, as the reporter renews it; the Admin People list carries the certificate's start date for that, and the lifetime hint in Admin > Settings says the same. Documentation: a gotcha for the reset a refused certificate causes, decision 42 words what stays on plain HTTP, and MESSAGING.md says messaging works over a certificate. -
The docs say what happens to Claude desktop app sessions after the app restarts (#737). The app restores only the session on screen and leaves the others dormant, and the reporter runs inside each session, so nothing reports for a dormant session. SPEC 7.5, the user guide and the desktop gotcha now state that the on-screen session is revived under its old id, the others return when opened (possibly under a new id), and Podium keeps no card for a session that is not reporting. The lifecycle timers do not change. Documentation only.
-
A blowout event records the context window and how full it was (#735). A usage reset showed as a blowout on a card whose context bar read 25% (#687), and the event sequence at a reset could not be told from a real overflow. The reporter's
blowevent now carrieswindowandpercentbeside the kind and the tokens before and after, for a window-exceeded stop, a server compaction and an automatic compaction. Nothing is counted differently yet: the fix that stops a reset counting waits for the next one to show which path fires. Reporter only; the console and the server do not change.