Four defects, one shape: a write that returns 200 and lands nowhere.
- Unknown keys were dropped by a chain of `if (match) continue` with no `else`. The three prefixes
CommandTerminalWrapper actually writes — tmux, nvim, claude-code — were among them, so those panel
maps lived in the React Query cache only: every reload minted a fresh uuid and abandoned a running
pty. They now live in a `panel_state` bag on the dashboard row, and an unmatched key 400s.
- `ws-terminals-{id}: null` fell through to an upsert, writing NULL into a NOT NULL column on a live
dashboard and re-INSERTing a deleted one. Renaming a dashboard sends exactly that, paired with
`ws-layout-{id}: null`, so the old slug came back as a zombie row in the dashboards list.
- HostTerminalWrapper and CommandTerminalWrapper built their state key straight from `dashboardId`,
which is a workspace *key* (`ws-layout-<id>`), while TerminalWrapper stripped the prefix. The server
read the un-stripped form back as a dashboard id and created it. One rule now, in state-key.ts.
Verified against the live server: unknown key 400s, the three prefixes round-trip, a null on a live
dashboard is a no-op, and the rename sequence leaves no zombie.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The optimistic cache made a refused write invisible: the UI stayed correct until the next reload, at
which point the change was simply gone. That is tolerable for a pane size and not for a chat panel's
agent name, which is the address a peer agent is delivered to.
Roll back only if the cache still holds exactly what we wrote — writes to one key overlap freely (a
window resize fires one per group) and rolling back over a later successful write would turn one
failure into two.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
47 tests, the first under any Workspace path. These functions carry a panel's
identity now, so a regression in swapPanels is two agents exchanging names,
not a cosmetic glitch.
Writing them found one: setApp preserved config whenever appType was not null,
so changing a panel from chat to terminal handed the terminal the chat's
{agentName} to read as its own settings. The comment beside it already stated
the opposite intent. Not reachable through the UI today — the picker only
appears on an empty panel, so the only route out of an app is via null, which
does clear it — but setApp is exported and its signature permits the direct
swap. Now only a same-app set keeps the config.
Also pins two things as expectations rather than folklore: a move drops zoom
and fitContent (todo 5.3, to fail the day that is fixed), and a split
redistributes sibling sizes evenly (todo 5.5).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The re-rank the north star deferred, done now that the MVP is built and
running — so it is ranked against what the mechanism turned out to need.
The finding is that most of the list is not on this path. The mechanism is
server-side and a panel is a pointer to it, so a remount, a re-render or a
drag costs a replay, not a session. Section 5.2 and 5.3 are large downgrades;
5.3 was on the critical path when the north star was written and is disarmed
by resolving identity by name.
What is left is small and mostly one defect wearing four hats: a write that
silently does not land. Panel identity lives in the layout jsonb now, so the
swallowed persist catch, the dispatcher's missing else and the two debounce
lost-updates each become a panel that forgets which agent it is — invisibly,
for exactly as long as nobody is looking.
Also corrects two items the MVP made stale, and promotes layout-utils tests:
e588524 put agent identity inside those mutators and shipped them untested.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The server's uniqueness check is the one that matters, but its 409 currently never fires
(the constraint name is on err.cause, not in the DrizzleQueryError message), so a collision
came back as "Internal Server Error". The address book is already in hand here — checking
it first turns the common case into a sentence the human can act on.
Diagnosis of the server-side bug, with the patch, is in COMMS/agent-panels-split-2026-08-07.md;
that file belongs to another agent and is still uncommitted, so it is theirs to apply.
The panel remembers its agent's NAME in its own layout config, not the server row's panel
id: movePanel mints a fresh id on every drag, so an id-based lookup forgets the agent the
first time the dashboard is rearranged. The name travels with the panel contents; the row
is found by name and re-anchored to wherever the panel now is.
A named panel passes the row's sessionKey to useChat instead of letting the server mint a
throwaway uuid per connection. That is the whole of continuity: the same key comes back on
every load, resume-cursor replays the durable events under it, and the claude sidecar
resumes the same transcript from its write-through map even after the session was reaped.
Its cwd comes from the row too, because deliverToAgentPanel already runs an incoming
handoff there — otherwise the same agent would work in two directories depending on
whether the human or a peer spoke to it.
Only on real dashboards. The fixed screens keep anonymous chat panels exactly as before.
A panel can now hold an opaque config blob that the framework stores, moves and deletes
but never reads. It lives on the layout node for the same reason zoom does: the layout is
already persisted per panel and server-side, so a panel's configuration outlives the tab
and is deleted exactly when the panel is.
swapPanels and movePanel now exchange { appType, config } as one unit. They used to carry
only the app type, which would have silently reset a configured panel to defaults on a drag.
Apps read it through usePanelConfig(panelId); PanelSlot already passes panelId to every
registry app, so nothing else in the framework had to change.
docs/agent-coordination.md is the objective the workspace/panel work serves —
several agents on one dashboard handing work to each other instead of routing
every step through the human, with the human authoring the workflow at the top.
Written from the owner's own words during the 2026-08-07 conversation; where a
section records a decision, that decision is his.
It settles the questions that were blocking: sessions may be reaped and resumed
from the durable sessionKey→claudeSessionId map (no heartbeat), the address is
the human-assigned panel name rather than the panel id, roles are prompts rather
than features, and the protocol is turn-boundary-only by construction — which
routes around the mid-output restart failure instead of fixing it.
Cross-referenced from CLAUDE.md, workspace-panels.md and workspace-panel-todo.md
so it is findable from any of them. The todo is still ordered by defect severity
and now says so; the re-rank against the objective is deferred, not forgotten.
Also corrects two items dated 2026-08-08 to the day they were actually found.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A conversation is turn-based: your message, the work, the answer, repeat. The
moment you send the next message the tool calls and running commentary that
produced the last answer stop being what you are reading and start being what you
are scrolling past. So every turn but the live one collapses to three parts —
what you asked, one summary row, and what I concluded — with the summary naming
what you gave up ("5 tool calls · 2 messages · 1 failed") so you can tell whether
you want it back. Failures are counted on the summary rather than only inside,
because a red row you have to open to find is a red row you never find.
It is a pure derivation over the message list rather than state, which is what
makes a reload render identically to a live session: no wire format, no
persistence, no server change. Dividers and compaction seams split a fold instead
of disappearing into one, because "5 tool calls" hiding a /clear misreports what
happened to the conversation rather than to the work.
A turn that ends cleanly without saying anything gets a marker row. It happens
rarely and is disproportionately confusing — the composer re-enables and nothing
appears, which is indistinguishable from a turn that died. Deliberately a seam and
never prose: words in my voice that I did not write are a lie, and the next time it
happened you would not know which kind of row you were reading.
Fold-open state lives in the list, not the fold, because rows are virtualised and
state inside one would be thrown away when it scrolled past the overscan window.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Compaction was the one thing the harness does that emitted nothing at all while it
ran, and it can run for minutes — silence that reads as a hung turn, which costs a
server restart to discover it wasn't. The sidecar now reports both ends: the start
from the PreCompact hook, the finish from the compact_boundary message with the
token count, both durable so a reload or a reconnect still sees them.
Tool rows open themselves while they run and hold for five seconds after their
result, so the inputs are on screen at the moment the call is made rather than
after the fact. The clock lives outside React, keyed by tool call id: rows are
virtualised, so unmounting is not the call ending, and a fast call can render its
start and its result together — a row that only opens when it catches the pending
state never opens for exactly the quickest calls. A click outranks the clock for
as long as the row lives. A failure behaves identically and differs only in colour,
so it stays findable by scanning and nameable in conversation.
Shell logs move off the green-on-black pre onto the shared code surface, which is
the one block that had no copy button and the one you most often want to hand to
someone else.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
every multipart upload through /api/<sidecar>/* arrived corrupted. bodyParser ran on
proxy routes and called parseBody for multipart, so hono cached a FormData on the
request; when the proxy then asked for the bytes hono re-serialised them from that
cache with a NEW boundary, while the proxy still forwarded the ORIGINAL content-type
header. header and body disagreed and the far side rejected it with
"Multipart: Unexpected end of form".
bodyParser now skips prefixes owned by createSidecarProxy, which register themselves
so a new sidecar cannot forget. the proxy also forwards the body as a stream instead
of buffering it, which drops the second in-memory copy of every upload.
note the bug report proposed skipping multipart in bodyParser outright; that would
have broken /upload, /file-browser upload and /bug-report, which do read a multipart
body from ctx.get('body').
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Five Bash rows in a trace read `cd /home/…/platform && git status…`, `echo "=== server-side…`,
`echo "=== opencode handler…` — cut, every one of them, exactly where they started being useful. Two
things conspired. The summary showed the head of a compound command, which is usually scaffolding: a
`cd` into the repo, or an `echo` labelling output for a human. And both `slice()` and CSS `truncate`
drop the tail, which is where the identity lives — the filename that distinguishes ten Reads sharing
a directory, the target a command acts on.
So skip a leading `cd`/`echo` up to its `&&`, and pin the tail as its own non-shrinking span so the
head ellipsises and the cut lands in the middle at whatever width the panel is. Both are display
only: expanding the row, and the copy button, still give the command verbatim.
The right margin traded `done` for what the call found. Success was the loudest colour on the row
and reported the least interesting fact about it, once per row; a failure still earns its red. In its
place the count that used to cost an expand to learn — `no matches`, `12 lines`, `3 files`.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
soulseek had a zoom control wired into its own panel header, persisted
under its own screens/ key. move it into the framework so every panel has
it, and drop the soulseek-specific copy (its header was then identical to
the default, so that goes too, along with the orphan db row).
the factor lives on the LayoutPanel node rather than in its own
useDashboardState key: the layout is already persisted per panel, and a
separate key would seed a server row per panel on mount. absent at 1, so
an untouched panel adds nothing to the stored layout.
uses css zoom, not transform: scale. a transform repaints at a different
size without re-laying out, so the panel keeps its 100% geometry and
anything anchored or percentage-sized lands wrong — chat's composer made
that obvious. zoom scales used lengths instead: children reflow, h-full
still resolves to the panel, and rem-based tailwind text scales with it.
@container moves onto the zoomed element so container queries respond to
the effective width, the way they would in a genuinely narrower panel.
zoomable: false opts out the terminals (xterm measures its own cell grid)
and remote desktop (novnc does its own scaling and pointer mapping).
fixes chat's virtualiser under zoom: it measured bubbles with
getBoundingClientRect (rendered px) but positions them with translateY
(layout px), so at 70% every bubble was placed too early and they stacked.
new helpers/measure-zoomed divides the element's currentCSSZoom back out.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the framework has no tests, no error boundaries and a handful of known
defects that keep resurfacing mid-feature. write them down once, ranked,
so they can be picked off in the context of whatever is being built.
notable: the dashboards PATCH dispatcher silently drops any key family it
has no branch for, and three in use today (tmux, nvim, claude-code, all
from Terminal's statePrefix) match nothing — so that state never persists.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The composer already uploaded an image, split its data URL and put the bytes on the wire as
`images`. Nothing on the server read them. The `chat` ClientMessage had no such field, and the
prompt reached the sidecar as a bare string, so all the model ever saw was the client-generated
`[Attached image: …]` placeholder — a label describing a picture it was never shown.
The transport was never the obstacle: `query()` consumes an async iterable of user messages whose
`content` is an Anthropic `MessageParam`, and only `pushTurn` hardcoding a string kept it to text.
So `images` is threaded through the four hops that dropped it and turned into native image content
blocks at the end, renaming `mediaType` to the API's `media_type` at that last step.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The clone check keyed off `navigation.type`, which only reports `reload` for F5/Ctrl-R. Every other way
back into the app — Enter in the address bar, a link, re-opening the URL after the server was down — is
`navigate`, and threw away the name you typed.
Ask instead of guess: each tab holds an id beside its name, and a copy is a tab whose id is still held
by a live tab, which the original says over a BroadcastChannel. A refresh has nobody to answer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the cut-off notice is its own event now rather than an error: nothing is broken
and nothing is lost but the turn, so the row says what happened and offers the
one action that fixes it. the conversation is already durable — the claude
session id is written through to disk and passed back as resume: — so retry
just resends the prompt on a session the fresh agent picks up with full
context. read back out of the transcript, so a second window on the same
session can offer it too.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
restarting officer-agent takes every persistent session with it and nothing
downstream notices: the browser's socket is healthy, officer's subscription is
a bus filter, and there is simply never another event. the spinner ran forever
and a refresh didn't help, because the transcript has no ending to read.
keyed off the agent *registering*, not disconnecting — a disconnect fires on
every `pm2 restart officer`, when the turn is fine. a registration socket dies
with its process, so an agent appearing on it is a new one. covers the sitting
tab; the reconnect path covers the rest, with the client now sending its belief
that a turn is in flight and officer checking it against the agent over a new
claude:is-generating. the check fails toward alive.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
signin refused any non-owner arriving through the web or mobile platform
origin: "This account can only sign in through its app." so a member could
hold a gitea grant and still never reach a page — verified as a live 403
before this change.
that rule was correct while single-user was the invariant. the only
non-owner accounts were music-app accounts, and there was no way to say
"this person may use the platform, but only these parts of it", so keeping
them out entirely was the honest answer. capabilities say exactly that now,
per feature, at both doors and on every request.
so superAdminOnly is retired rather than patched. the web origin and the
platform app get no path scoping — what their caller may reach is decided by
their role, not by their Origin. per-app path scoping stays for the
single-feature apps (music, vault, tail), where it still means something.
note this WIDENS who may sign in: any Active account can now authenticate
through the browser. that is the intended product change, and it is only
safe because the capability backstop runs on every request behind it —
which is why it lands after that, not before.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
hover-revealed means most people never find it, and touch has no hover at all.
70% white on the dark block, brightening on hover; the pre reserves right
padding so a long first line scrolls up to the button instead of under it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the bubble's copy button copies the whole reply, which is the wrong unit when
the reply is prose ending in one command to run. fenced blocks get their own
button; inline code doesn't. text read from textContent at click time rather
than the markdown ast, trailing newline stripped so a pasted command doesn't
run itself. the positioned wrapper takes the vertical margin, or the pre's own
margin collapses through it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
pressing stop ended the turn with "Claude Code returned an error" — the agent
sdk reports interrupt() as an ordinary failed result, indistinguishable from a
real fault downstream. the sidecar now flags the session it interrupted and
rewrites that event to the existing durable 'stopped', which opencode already
emitted. escape stops the turn (bound to the chat subtree, not the document),
and the prompt comes back to the composer verbatim unless you've started typing
something else. history parity: claude files [Request interrupted by user] as a
user message, so the transcript reader maps those exact strings to the same
role instead of replaying them as something you typed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the server half of the previous commit, which belonged with it. the frontend
guard needs both lists: absence from `routes` cannot tell a route this
account lacks from a route no capability claims, so without this the guard
permits everything.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
useCapabilities is the frontend's view of the model and explicitly NOT its
enforcement — hiding a dock icon is a courtesy, the 403 in origin-validation
is the lock. so it fails OPEN: if the request errors the full dock renders.
a member clicking through to a 403 is a bad minute; an owner locked out of
their own platform by a transient network error is an incident, and the
server refuses what it should refuse either way.
the endpoint returns held routes AND denied routes, because absence from the
held list cannot distinguish a route this account lacks from one no
capability claims at all — `/`, the settings shell — and a guard that cannot
tell those apart either blanks the app or guards nothing. i wrote the first
version without the second list and it silently permitted everything.
`can` and `canVisit` are memoised on the query data. a verb rebuilt every
render gets a new identity every render, which is how every playback report
in the jellyfin player was disabled for days; the dock filter puts one in a
useMemo dependency list, so it would have been the same bug.
the permissions screen is one role at a time, with an explicit save and a
dirty state, rather than a roles-by-capabilities grid — a grid invites
reading across rows, which is not a question anyone has, and makes revoking
gitea for every member one click among fifty. it also states plainly why
terminal, chat, files and the rest are absent, so their absence reads as a
decision rather than as a missing feature.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
sessionStorage is the per-tab store — separate per tab, survives reload and
navigation, dies with the tab. The route title is derived rather than assigned,
so navigating no longer wipes a name you typed.
Duplicating a tab clones sessionStorage, so a `navigate` that arrives already
holding a name is treated as a clone and drops it; fails soft.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
GET /api/user/capabilities is what the caller may reach, and every account
may ask — it is mounted on a core capability so an account granted almost
nothing can still find out what it has. the dock and route guards read it.
it is a courtesy, never enforcement: hiding an icon is not access control
and the 403 in origin-validation stays the lock.
GET/PUT /api/users/capabilities edit the policy, owner-gated. the write path
is where the registry's authority over capability keys is applied, which is
why the column has no CHECK: unknown keys and non-app kinds are refused
rather than stored for the resolver to drop on read.
the exception is `selfService`. useAuth calls PUT /api/users to change your
own name and avatar, and that route has always lived on the same router as
the owner-only account administration around it — so declaring /users an
admin capability locked every member out of their own profile. moving it to
/api/user would be tidier and would break every shipped mobile client, so
instead the registry says out loud that this one route is not what the
capability around it is. exact method and exact path, so it cannot widen:
verified that PUT /api/users passes while GET /api/users, PATCH
/api/users/:id/role, DELETE /api/users/:id and PUT /api/users/:id are all
still refused.
23 unit tests.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
replaces the account backstop. it was two hand-written lists — NON_OWNER_PATHS
confining every non-owner to /api/auth + /api/music, and NON_OWNER_WS_PROVIDERS
doing the same for sockets. they were not wrong, they were unscalable in one
specific way: an allow-list answers "which paths" but never "why", so onboarding
anyone who needed anything other than music meant editing an array in a
middleware file and hoping the socket half got edited too.
now both doors resolve against the registry, so they cannot disagree about what
a role holds. terminal, chat, task-runner, pipeline and desktop are refused by
being `execution` capabilities rather than by being absent from a list somebody
maintains.
fail-closed everywhere: an unknown capability key, a missing row, a database
error or a deleted user all deny. the grant cache is keyed on role and has an
explicit invalidation contract — unlike the one super-admin.ts refuses to have,
this one has exactly one writer and it lives beside the reader.
seeded Member → music at WRITE, which is precisely what the old path-based
backstop allowed. granting `read` would have been a silent downgrade that broke
playlists for the three live member accounts overnight.
verified against the live database and real accounts: 27 http/socket cases, the
read/write split (personal sub-paths writable at read, /music/scan not), cache
invalidation after a revoke, and the borrowed test account's role restored.
20 new unit tests; full suite 362 pass 0 fail.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
role_capabilities (role, capability, level). the subject of a grant is a
role and never a user — the owner's call, and it keeps "what can a Member
do" a question with an answer, which per-user rows would not. when one
person genuinely needs something different, that person needs a role.
a missing row means no access. nothing denies; absence denies. an empty
table is a freshly installed server where members reach nothing but their
own account, which is the right starting state.
the capability key is deliberately unconstrained: a CHECK listing the keys
would put the registry in two places and turn adding one into a schema
change. the api validates against the registry instead. what the database
does enforce is shape — a legal role, a legal level, one grant per pair, and
no rows for Super Admin, since the owner bypasses this table entirely and an
inert row that looks meaningful is worse than no row.
uniqueIndex not unique().on() per databases/CLAUDE.md. verified: pushed to a
scratch db twice, second push planned only the two known-harmless
pk_music_now_playing lines, so the declaration is stable. applied to
officer_dev without --force and without a prompt. all six constraint cases
behave — bogus role, bogus level, a Super Admin row and a duplicate pair are
each rejected; a legal grant and the same capability on another role are
accepted.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
permissions are capabilities, not routes. a capability is a feature — the
unit the owner grants, the dock filters on and the acl enforces — declaring
the api prefixes, websocket providers and screens it expands to.
four kinds. core is every account and is not grantable because it is not
deniable. app is the grantable surface. admin is the platform administering
itself. execution is never grantable at any level: terminal, chat, tasks,
files, desktop and browser all run as the owner's os user in the owner's
home, so granting one is co-ownership of the machine rather than a feature.
the part that matters is assertCapabilityTotality. the websocket hole fixed
in 2873948 was not a wrong rule — it was a door added without telling the
rule, because bun's route table matches /api/terminal/ws before the /api/*
catch-all that reaches hono's middleware. so the server now refuses to start
unless every mounted prefix and every user-facing socket maps to exactly one
capability. hono.ts mounts from a table and exports it, so the check reads
the real surface instead of a copy that can drift from it.
verified: passes against the live surface, and refuses all four ways — an
ungated router, an ungated socket, a claim on a deleted router, a claim on a
deleted socket.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The list collapses each chain to its newest link — the only one that can be
resumed — carrying the root's title and start time, the summed message count
and a part badge. The detail splices the chain's transcripts oldest-first with
a divider between parts, server-side, so the client's index-window pagination
needed no change.
The divider says "context cleared — nothing above this is in memory", because
the whole risk of merging is that the history reads as continuous when the
agent's context is not. Delete cascades the chain and the confirm says how many.
Supersedes the "continues X" line from the previous commit: there is nowhere to
link to once the parent is scrolled up above you.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
nothing reads it and nothing writes it any more, but a refresh re-requests the
address bar verbatim, so one left over from before the path-based groups sits
there indefinitely looking like it means something. on a bare /chat it still
says which group you wanted, so upgrade it to /chat/g/<path> rather than
dropping it — an old bookmark keeps working. anywhere else the session decides
its own directory, so it is just removed. other params untouched.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
/clear starts a new claude session and the list showed it as an unrelated
conversation. claude records no parent link anywhere — not in compactMetadata,
logicalParentUuid, summary.leafUuid, the per-session slug, or the live process
registry — so infer it: a cleared transcript opens with /clear, and /clear
happens inside one process, so the parent is the conversation in the same group
that was writing to disk at the instant this one began (4ms apart, measured).
matching is on per-minute activity rather than updatedAt, because resuming a
parent moves its end time past its child's birth and lost the link entirely for
two of the three cleared sessions here. the window is symmetric because clearing
makes claude summarise the conversation it is ending, so the parent's final
record can land after the child's first. ambiguity fails closed — the wrong
parent also renames the conversation.
read-only: nothing is written back to claude's store, and an explicit title
always wins.
also: cleared sessions were titled "<command-name>/clear</command-name>" because
claude does not set isMeta on slash commands; and the chat header was hardcoded
to undefined, so it read "New chat" above every conversation you opened.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the chat group moves from ?cwd= to a path suffix behind a g/ discriminator
(/chat/g/home/me/project), and a session url carries no group at all.
the real fix is not the spelling. a session's working directory was read back
off the query string to decide where the agent executes, so the address bar was
the authority on where code runs. a pasted or refreshed /chat/<id> arrives with
no ?cwd= at all, so a turn sent before the resolve landed ran in the default
general_chat_sessions dir instead of the project; and a hand-edited ?cwd= could
name a group the session doesn't belong to, with nothing to reconcile them.
loadClaudeSessionById already resolves a session's cwd from the id alone, so the
id is the only source of truth there. it now travels on SelectedSession.cwd,
which is what the composer reads. the url can no longer contradict it.
the vocabulary lives in apps/ChatHistory/chat-routes.ts so a link built in a
panel and one built in a screen cannot drift.
also: startAgentRun no longer returns a literal chatUrl — it returns cwd and
AgentRunnerModal builds the link, so the server holds no copy of the frontend
url shape. and the post-turn permalink strips a stale ?cwd= instead of carrying
it forward onto the new session's url.
walkthrough doc gains item 13.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
twelve items in click-through order, each with where to look and what the old
behaviour was — several are only visible if you know what was broken. states
plainly at the top that none of it has been rendered in a browser.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
read-aloud swallowed every error: the spinner stopped, the speaker icon came back, and a
TTS service that was simply down looked identical to a button that did nothing. Both the
synthesis failure and the playback failure now say so.
errorText moves to helpers — useClient rejects with a plain { status, message } object
rather than an Error, so the reflex instanceof check reports every API failure in the app
as "unknown error". Two callers now, and it is the wrong thing to reimplement per file.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
--duck-dark inverts to near-white in dark mode, so every place the chat used
it as a text or border colour was drawing light-on-light: the composer's own
text, the question prompts, the launcher textarea, the session detail bar, the
"send a message to start" placeholder. all of it moves to the semantic tokens,
along with the raw red-500s, which had no dark story at all.
the background task tray had seven labels below the 12px floor, including the
live log itself, and hardcoded green-600/red-600/amber-500 where the shared
tones exist. its detail panel was capped at a flat max-h-64 while the composer
it docks in is shrink-0 and the transcript above is flex-1 — so in a short
panel an open task could leave almost no conversation visible. capped against
the viewport too.
a failed background task drew the same CircleSlash as a stopped one: the two
outcomes you most need to tell apart were one glyph.
the attach menu offered four things and did two — "Text File" and "PDF" had no
onSelect at all. text files now inline into the composer as a fenced block
with the filename, size-capped and rejected if they turn out to be binary. PDF
is removed rather than faked: nothing in the platform extracts PDF text.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the list drew its own border per row on top of nothing, so every boundary
between two rows was a double hairline, and it emphasised two things per row
where the design language allows one. DataRow/DataList settle both.
loading and empty were the same grey sentence, which made a slow transcript
read look like an account with no history; they are now LoadingBlock and
EmptyBlock, and a failed read gets an ErrorBlock with the actual message
instead of rendering as "no sessions".
rename and delete swallowed their failures whole — useClient only raises a
dialog for 401, 403 and 5xx, and the likely error here is a 404 from a
transcript that vanished under you. both toast now, as does a deep link to a
session that cannot be read, which used to open an empty pane and say nothing.
the New Chat button was duck-teal filled with duck-yellow text: duck-teal is
a bright cyan in dark mode and duck-yellow has no dark override, so the pair
sat near 2:1 contrast in both themes.
active-row highlight now comes from the route rather than the selection
channel, so it is right on a deep link before any panel has published, and
deleting the open session navigates out of it instead of leaving a dead route.
deletes SessionBar and SessionContextMenu: the first was exported through two
barrels and imported nowhere, the second was never imported at all and typed
its session id as a number.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the most-looked-at panel in the chat was the worst rendered: an Edit showed
old_string and new_string as two flat monospace blobs, so the only question
you had — what changed — was the one thing you had to work out by eye.
adds a small LCS line diff rather than a dependency; jsdiff would ship to the
browser to do a textbook algorithm that is shorter than its own integration.
no line numbers, deliberately: Edit fragments carry no file position, and an
invented line number is worse than none.
the collapsed row now carries +n −m, so the size of an edit is legible without
expanding it, and copy yields the resulting text rather than a key: value dump.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the provider label was bg-duck-dark/80 text-white, and --duck-dark is
near-white in dark mode, so it was white on white for every session that
had already started. same class on the active provider tab.
the streaming bubble bailed on empty text, so the wait between send and
the first token — tens of seconds with thinking on — rendered nothing at
all. it now shows the bubble with pulsing dots.
also: tool status tones onto tone.ts instead of a third copy of the same
green/red/amber ternary, sub-12px labels up to text-xs, break-all to
break-words so shell commands stop breaking mid-identifier, and
break-words on the user bubble so a pasted url stays inside the pane.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the sessionKey to claude uuid map is what --resume needs to reattach a
conversation after the agent sidecar dies, so a 30s debounce put exactly
the wrong state behind a window. flushAndSave on SIGTERM covers a pm2
restart but not a crash or SIGKILL, which is the case resume exists for.
an equality guard keeps onSessionId from thrashing the file.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Commits, branches, cross-repo issues, notifications and organizations all move
onto DataRow, which is what makes them agree with each other — five lists that
were each hand-assembled from the same flex/gap/truncate parts now genuinely
share one row.
The repo header gets the same owner-muted / name-semibold split as the list row,
so the two screens agree about what a repository is called. Its description goes
from 12px to 14px, and the tab strip with it — those are read, not scanned, and
they were the smallest text on the busiest screen.
DataRow gains `href` for destinations outside the app. Notifications needed it:
Gitea's subject URL can usually be parsed back into an in-app route and
sometimes cannot, and the row should stay clickable either way rather than
becoming a dead div on the payloads that do not parse.
Empty states across the dashboard now say what the search actually covers.
"No open issues" was hiding that Gitea's cross-repo search only ever looks at
issues you created, are assigned, or are mentioned in — which is the difference
between a quiet week and a misconfigured token.
Typechecks clean and prettier is clean. Still not rendered in a browser.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gitea goes first because it is the first Officer surface anyone outside this
machine will see.
GiteaBits is now a thin adapter over components/Data rather than its own set of
styles: it translates Gitea's vocabulary into the five tones and gets out of the
way. Export names are unchanged so the ten call sites did not all have to move
at once.
The one judgement call worth flagging is CLOSED ISSUE -> neutral instead of a
colour of its own. Most issues in any list are closed, so giving them a tone
paints most of the screen and leaves the open ones no quieter than the rest;
neutral is what makes "open" findable, which is the only thing anyone scans an
issue list for. A closed PULL request keeps danger — that one was rejected, and
rejection is a real outcome. Merged is info, because a merged PR and an open one
were previously the same green.
Repositories and issues now build on DataRow. Repo rows put the owner prefix in
muted and only the name in medium, so one weight per row survives; issue rows
truncate rather than wrap so every row keeps one height and the list can be
scanned down its left edge. Timestamps are RelativeTime, so "2 months ago" now
carries the exact date as a hover title instead of losing it.
Swept the whole app for drift: emerald/amber/sky/red palette numbers to
success/warning/info/destructive, and every text-[10px] and text-[11px] to
text-xs. ring-black/5 and text-black/50 are gone — both were invisible in dark
mode. Nothing below 12px remains anywhere in Gitea.
Empty states say why they are empty now. "No open issues" on a repo with fifty
closed ones was technically true and useless.
Typechecks clean. Still not rendered in a browser — needs pm2 restart officer
and a hard refresh. text-xs is still 65 uses against 21 text-sm; the remaining
pass is per-case judgement about which of those are meta and which are content
that should be readable, and it is not a sweep.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
DuckSuite_Design_Language.md covers icons only — mascot geometry, lighting,
one-idea-per-icon, readable at 64px. It says nothing about type, density or
data, which is why the chrome looks considered and every data view does not.
This is the companion document plus the primitives that enforce it.
The diagnosis it is written against, from grepping two apps: 11 uses of
text-[10px], 7 of text-[11px], 52 of text-xs and 3 of text-base, with no rule
about which meant what; six radius values; hand-picked emerald/amber/red/purple
next to unused --success/--warning/--destructive tokens; text-black/50 and
ring-black/5, which are invisible in dark mode. None of that was a bad decision,
it was the absence of one thirty times over — so the fix is to remove the choice
rather than to have better taste.
docs/design-language-interface.md sets four type ranks with a 12px floor, one
focal point per row, five state tones, three radii and a spacing rhythm. The
principles are lifted from the icon language rather than invented, because
"one focal point, reads instantly, no unnecessary decorations, if it needs
explanation it is too complicated" is already the right rule for a dense list.
components/Data/ is how you spend that vocabulary: DataRow/DataList/RowMeta,
StatusPill/StatusIcon, LoadingBlock/ErrorBlock/EmptyBlock, RelativeTime. Rules
you have to remember are rules thirty views already broke, so the shape encodes
them — DataRow takes exactly one title and everything else is meta, RowMeta
puts separators between items so a trailing dot cannot appear, RelativeTime
carries the absolute timestamp as a hover title.
Adds --info (violet) as the fifth semantic tone, light and dark. "Merged" and
"in progress" are neither good news nor bad, and painting them with --success
makes a merged PR and an open one look like the same thing.
Entirely additive: nothing imports these yet, so no existing view changes and
there is no collision with the other agent working in this tree. Typechecks
clean. Not yet rendered in a browser.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The previous commit's explanation of retargetUrls was inverted, in the message and in the comment.
It said /user and /repos build from ROOT_URL and /contents from the public host. It is the other
way round: this instance's app.ini has ROOT_URL = https://gitea.pastilhas.dev/, which is what
/contents stamps, while /user and /repos echo the REQUEST host — loopback, because Officer dials
http://localhost:9004 directly instead of going through the reverse proxy.
Behaviour is unchanged and was already right: rewrite only a private/loopback URL, only onto a
public target. Which half of the response is wrong does not affect that rule, which is why the bug
did not show up in testing. But a comment that names the wrong cause is how the next change
re-breaks it, so this corrects the record.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Replicates what Gitea's own web UI offers, on top of the sidecar's /_api pass-through.
Repository browsing (tree, file view with the FileViewer's shiki renderer, README), commits,
branches, tags, releases, issues and pull requests both per-repo and cross-repo, notifications,
explore/search and organizations. Routes are /gitea/:section plus /gitea/repo/:owner/:name/:tab/:item,
all real Links with the URL as the source of truth — no selection channel.
Markdown is rendered client-side (react-markdown + remark-gfm + rehype-sanitize, rehype-raw
deliberately absent) rather than through the instance's /api/v1/markdown, because consuming that
means dangerouslySetInnerHTML and there is no DOMPurify in the tree with installs frozen. The cost
is Gitea's #123 and @mention cross-references; relative links and images are resolved instead.
The /markdown and /markup allow-list entries stay, so that door is open when a sanitiser lands.
retargetUrls rebases instance-minted URLs onto a browser-reachable origin, IN ONE DIRECTION ONLY.
This instance answers with two: /user and /repos build from its configured ROOT_URL
(http://localhost:9004), /contents from the public host. An unconditional rewrite onto the
connection URL therefore broke the second set to match the first, turning working https links into
dead loopback ones. Only a private/loopback URL is rewritten now, and only when the target is
itself public; when the connection URL is a dial address nothing is touched and the connection
screen says why avatars will not load.
Also carries the frontend half of the one-instance-many-tokens model: the connection form draws a
URL field only for the owner and sends no url key at all for anyone else, ServiceConnection.url is
string | null to match officerdb, and the rebase origin comes from the resolved instanceUrl rather
than connection.url, which is null for a member.
Not verified: no runtime pass since the last four changes, the issues and pull views have never
rendered a row (the instance has none), and the member path has never executed (one account).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the config cache is keyed by user, but a member's entry holds the owner's
base — resolved at read time, since a member's row stores no url. dropping
only the caller's entry was correct while each row was self-contained; once
rows inherit, the owner moving or disconnecting the instance left every
member cached against the previous host until the sidecar restarted.
also returns the resolved instanceUrl from connectionState. members not
being able to set it is the invariant; not seeing it never was and could
not be — every avatar_url the instance hands back is on that origin.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Gitea is the first service where the server is shared but the account is not.
The owner connects the instance; every other user supplies only their own access
token and sees their own repositories, notifications and issues.
The plumbing was already per-user — createSidecarProxy injects the authenticated
caller's id as X-Officer-User, the sidecar refuses a request without it, and
service_connections is UNIQUE (user_id, service). What was wrong is that url and
token lived in the same row and a save demanded both, so a member would have had
to type the instance URL. That is worse than inconvenient: a member who can name
the URL has a per-user SSRF hop behind a settings form, and the sidecar would
dutifully attach their token to it.
`url` is now nullable, and NULL means "inherit the instance". The owner's row
carries the URL and IS the instance; everyone else's row is a credential. A
member's URL is therefore not stored rather than merely hidden — which is what
makes "members never see the instance URL" a property of the schema instead of a
filter somebody has to remember on every response.
Resolution lives in one place (getResolvedServiceCredentials / getServiceInstanceUrl)
rather than in each sidecar, so there is a single answer to "where is this
service" and no sidecar can accidentally trust a member-supplied URL.
Rules, all enforced in the sidecar rather than the UI, because a form that hides
a field is a suggestion and these are rules:
owner PUT /_config { url, token }, as before
member PUT /_config { token } only; a url in the body is REJECTED, not
ignored — ignoring it would leave someone debugging a
screen quietly talking to a different server
member, no instance 409, "the server owner has not connected a Gitea
instance yet"
member GET /_config has no url to return
owner disconnects members keep their tokens and resolve to nothing; no
instance, no service
GET /_config also now answers `instanceConfigured` and `isOwner`, which is what
lets the UI tell "you have not connected yet" from "there is nothing here to
connect to" — different screens.
memos, slskd and transmission front a single daemon and always store their own
URL; they now treat a null as a malformed row rather than reaching for somebody
else's instance.
Verified on a scratch database: pushing twice adds no diff churn beyond the known
pk_music_now_playing pair, and the resolution behaves — member GET returns a null
url, member credentials resolve to the owner's base with the member's own token,
and deleting the owner's row leaves the member's token intact but resolving to
nothing. All four live rows have a url today, so the column change applies
without touching data.
Not reachable by a real member yet: the account backstop still confines
non-owners to /api/auth + /api/music. This works the moment the capability model
lands, and until then is testable only by minting a token.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The permission model being built reads the HTTP method to decide whether a
non-owner may make a call: safe methods are reads, everything else is a write.
That only works if the method tells the truth. These five read something and
returned it while announcing themselves as writes, so a member would have been
denied a read they are entitled to because of a habit in how the route was
declared.
/api/file-browser/video-info POST {url} -> GET ?url=
/api/file-browser/video-playlist POST {url} -> GET ?url=
/api/server-settings/ocr/models POST {url} -> GET ?url=
/api/transmission/_officer/port-test POST -> GET
/api/jellyfin/_config/:id/test POST|GET -> GET only
The last one already answered to both, which is worse than either: a method that
means nothing cannot be the thing authorisation reads.
Deliberately stops at five. A sweep of all 100 mutating routes found many more
reads wearing POST, and they are staying, for two reasons that are not going
away: some need a request body GET cannot carry (/stt takes multipart audio;
/tts, /ocr, /transcribe take payloads), and some carry a credential, where a
query string is the wrong place — access logs, shell history and Referer headers
all capture those, request bodies do not (/tts/voices takes an apiKey, the four
/test endpoints take connection secrets, /local-providers/probe takes auth).
So the method alone can never carry the permission model, and the registry will
need an explicit per-route classification regardless. Converting these five is
worth it because it is free; converting the rest would be a breaking change
across 117 mobile call sites that buys nothing.
Web callers updated in the same commit; the sidecar contract comments now match.
Mobile has exactly one caller to change — transmissionPortTest in
packages/core/src/services/transmission.ts — and no shim was added, because an
endpoint answering to both methods is the problem this commit exists to fix.
docs/api-method-changes-2026-08-06.md is the handoff for the mobile team: what
changed, the one line to edit, what deliberately did NOT change and why, and how
to verify.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A Member account could open a terminal. Demonstrated, not inferred: on 2026-08-06
a token for dingoshf@gmail.com (role Member) was refused GET /api/tasks with 403
and, in the same minute, opened /api/tasks/pipeline/ws with 101 Switching
Protocols.
The account backstop lives in originScopeMiddleware, which is a Hono middleware.
Websocket upgrades never reach Hono: Bun's route table in server.tsx matches
'/api/terminal/ws' and its siblings before the '/api/*' entry that hands off, so
NON_OWNER_PATHS was enforced on HTTP and nowhere else. upgradeWs verified the
token and the blacklist, then upgraded — it proved WHO was calling and never
asked what they could reach. None of the handlers behind it checked either;
terminal, chat and desktop have no authorisation code at all.
What was reachable with any valid token: a shell as the owner in the owner's
home, the agent with --dangerously-skip-permissions, arbitrary script execution
through task-runner and pipeline, and the owner's physical screen and keyboard
over the VNC mirror.
upgradeWs now refuses any provider a non-owner has no business on. cliamp and
cliamp-audio stay open to them — those are the socket half of /api/music, which
is what a music account exists for. An unrecognised provider denies.
NON_OWNER_WS_PROVIDERS is declared beside NON_OWNER_PATHS on purpose. They are
one rule at two doors, and the whole failure was that only one door had it;
splitting them across files is how that happens again.
The vault socket was already gated — it verifies isSuperAdmin in `open` and
closes 4001 — so upgradeWs was the only gap.
Not yet verified against the running server: platform TS does not hot-reload, so
the 101 above still reproduces until `pm2 restart officer`. The provider table
was checked in isolation: the five execution providers and an unknown name all
deny, cliamp and cliamp-audio allow.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The policy editor now opens read-only behind an Edit button, and edit mode brings up an assistant:
describe the change, get a complete revised document back, read a line diff of it, and only then press
Save. Nobody writes an ACL from memory — it is HuJSON in Tailscale's grammar keyed to names only the
server knows — and that, not typing speed, is what made this screen unusable.
The model never touches Headscale. It proposes text, the text lands in the editor's draft, and the
existing Save button is still the only thing that leaves the browser. A model that could write the ACL
directly is one that can partition the network the owner is connected through.
The sidecar calls officer-anthropic-proxy on loopback for one request with a timeout — no session, no
agent, nothing persisted. It sends the draft on screen plus the tailnet's vocabulary (user, node and tag
names) and no credentials of any kind. It does not validate the reply either: Headscale owns the only
parser that counts, same argument as policy.ts.
editing one server then another without closing in between reused the same
mounted form, which seeds its fields at mount — so the second server showed
the first one's name, url and ssh host. submit diffs that stale state
against the new server prop, so saving would have written them to it.
policy: /_officer/policy GET/PUT. the text goes up byte for byte and
headscale's verdict comes back verbatim — it owns the only parser that
resolves groups, tags and hosts, so a second one here would disagree with
what actually enforces. a file-backed policy is still served over GET and
only refuses on PUT, so writability cannot be read: the first save finds
out, and a refusal becomes a persistent read-only banner rather than a
rejection the owner would go hunting for a syntax error over.
nodes: move between users, and a tag editor for the setTags route that had
no ui. both sit together in the expanded card because both decide which
policy rules apply to a node, and a move says so before it happens.
Reads the per-server Officer Companion (COMMS/HEADSCALE_COMPANION_API.md): container health with its
own diagnosis, a log snapshot, a live SSE tail, and restart/stop/start.
The companion sits at ${server.url}/officer-api behind the same admin key we already store, so there
is nothing new to register — but only the sidecar can decrypt that key, so the sidecar proxies it and
the browser never talks to the companion directly. That also rules out EventSource for the stream
(no Authorization header), which is why the tail is fetch() + a hand-rolled SSE reader.
Two inversions the code is built around: /health is always HTTP 200 and must be read by verdict, and
an absent companion is a state to render rather than an error — an HTML 502 is nginx, a JSON 502 is
the companion reporting a failed docker op, and the admin API on the same domain is independent
either way.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every diagnostic in this app goes through headscale's API, which is exactly the
channel that is gone when you most need it — headscale crashed, the tailnet is
down, the logs are the only evidence. This adds the escape hatch: a per-server
SSH address and a Console section that opens a shell on that machine.
Deliberately thin. `headscale_servers.ssh_host` stores where to point ssh and
nothing else: no password, no key, no port. The console runs plain `ssh <host>`
in the same pty every other terminal panel uses, authenticating with whatever
~/.ssh on this box already knows. There is no credential here to protect and
this file must never grow one.
The address is NOT derived from the control-server URL and the form warns when
you type the same host into both — a console that resolves through the name
headscale serves goes down with it, which is the one thing it exists to survive.
It is also not validated on save, for the same reason: refusing to store the
escape hatch because the machine is unreachable is precisely backwards. Reaching
it is a separate, explicit Test connection button (BatchMode=yes, so a key that
needs a passphrase fails visibly instead of hanging on a prompt).
The host is validated to a conservative charset rather than quoted, because it
is typed into an interactive shell — rejecting `1.2.3.4; rm -rf /` while the
form is still open beats letting it survive to the shell as someone else's
problem. A jump host or an odd port belongs in ~/.ssh/config as a Host alias,
which the field accepts by name.
Also fixes a latent bug this would have hit immediately: TerminalView's
`initialInput` guard is scoped to a mount, so a remount typed the command again
into a live shell. A `replay` frame proves the session already ran it, so treat
it as sent. Harmless for `ls`; for the console it meant an ssh nested inside the
ssh you were already in.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Found while reading the app, not while fixing anything. The policy editor is the
only one that costs an ssh session today.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The chrome came back for a couple of seconds every few minutes during a film with
nobody touching anything. Only three things wake it, and none of them is periodic —
so the culprit is a stray pointermove. A mouse sitting on a desk still emits the odd
one-pixel event, a bumped table emits a few, and the browser synthesises a zero-delta
move of its own when the cursor style changes, which this player does every single
time it hides the cursor.
A move now has to travel eight manhattan pixels from wherever the pointer last
genuinely woke it. The anchor only advances on a real wake, so a slow deliberate
drift still accumulates past the threshold — it is jitter around a fixed point that
stops counting.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The auto-hide timer was there and correct; `position` was in its dependency list.
That advances about four times a second while a video plays, so the effect re-ran
and cleared the pending timeout every ~250ms and the 2.5s never elapsed. The chrome
could only hide while playback was stopped — which is exactly when the code
deliberately keeps it up — so in practice it never hid at all.
Not a fullscreen bug, though fullscreen is where a permanent scrubber is
impossible to ignore.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three things about the background-task strip, all the same complaint: the chip you
want is the one you cannot see.
Newest first, because the strip scrolls horizontally and the far end is off-screen —
chronological order put every task you just started exactly where you had to scroll to
reach it. Finished chips are capped at five, so a long session stops turning the tray
into a history; running ones are never counted or dropped, since watching them is the
whole point. And the default horizontal scrollbar is ~15px tall and laid out inside the
row, which is what made the chips read as squeezed — it is thin now, and the row and
chips have the padding back.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the gap limit was 1000, hardcoded, with the comment noting that raising it costs
node CPU and not correctness — and no way to raise it. it is the one thing about
a scan only the owner can know: how many addresses their old wallet handed out
and never had paid. RescanOptions threads from the POST body through the backend
and the source to utxos/scan, capped at 100k because past that the scan takes
longer than anyone waits. the card gets a "search depth" field beside the button,
blank meaning the default.
deleting a wallet with no seed took one unconfirmed request. the dialog asked for
the wallet's name and then threw the answer away, so the check existed only for
whoever went through the dialog — a node wallet still holds the credential, the
labels and the freezes. confirmName now travels with the request and the route
enforces it. a bodyless DELETE is told which field is missing rather than that
its JSON did not parse.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
four wallet defects, none urgent, all cheap:
selectCoins let an explicit coin-control pick override a freeze. naming an
outpoint now overrides only the confirmed-only default; frozen is absolute.
balances counted frozen coins in onchainConfirmed, so Send showed a figure a
max-value spend could not reach. Balances gains onchainFrozen — a component of
confirmed, not a deduction — filled at the route layer, because freezing is
Officer policy in Postgres and no backend can see it. The route only reads utxos
when something is actually frozen. Send subtracts it under "Spendable";
Overview lists it beside unconfirmed.
a rescan in flight upstream was invisible after a sidecar restart, and a second
POST would have queued behind it (scantxoutset is single-threaded node-wide).
adoptRescan polls an existing NBXplorer scan instead of starting one, and the
GET route falls back to it when local state is gone.
the per-variant scan deadline counted queue time, so a variant that sat behind
another wallet's scan timed out without ever having run. the deadline now
refreshes while the status reads Queued.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
two holes found auditing the wallet sidecar after the frozen-utxo fix.
the brute-force backoff lived inside UnlockSession.unlock alone, so /unlock capped
at five guesses a minute while export-seed — the one endpoint that returns the words
in the clear — took unlimited ones. every passphrase check now goes through the same
guard. verifyPassphrase rethrows LOCKED_OUT rather than folding it into `false`, so a
caller can tell "wrong" from "stop".
nextUnused advanced its mark on every issuance, paid or not, so a run of unpaid
addresses walked it past the end of the window the next scan covers; a payment there
would never be found again, and esplora has no rescan to go looking. sources now
declare how far past a scan's last index they can still see, and issuance clamps to
it — re-offering a virgin address rather than handing out one that could lose money.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
freezing a coin promised it would not be spent, and the promise only
held when the caller named outpoints explicitly. an ordinary send picked
its own inputs from a snapshot in which every utxo said frozen: false —
the chain has no idea what officer froze — so selectCoins, which has
always filtered on that flag, never saw one set. sendAll was the worst
case: "send everything" swept the frozen coin too.
the frozen list now travels with the request, set by the route and
overwritten if a caller supplies one. it can only ever restrict what is
spendable, so smuggling a value in gains nothing. the backend still
reads no officer table.
three tests pin it, including a control that sends successfully once the
coin is unfrozen — without it the other two would pass on a wallet that
could not spend at all.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the route typed its body as {name, defaultBip, config} and passed the
parsed object straight to updateWallet, which also accepted sealedSeed
for the passphrase change. a TypeScript annotation strips nothing at
runtime, so any authenticated caller could send a sealedSeed key and
overwrite the encrypted seed — no passphrase, no unlock. encryptSecret
encrypts nonsense happily, so the damage would have surfaced at the next
unlock, not at the write.
updateWallet can no longer touch the seed at all; resealing moves to
replaceSealedSeed, whose only caller has already proved knowledge of the
old passphrase. the route rebuilds its patch field by field as well, so
the next field added there cannot re-open it.
verified against the test wallet: the envelope is byte-identical after
the same request that previously would have replaced it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the card keyed off `sync.rescan` being null, which is what esplora
reports — and equally what nbxplorer reports before its first scan. so
the guard could not tell "unsupported" from "not yet run" and resolved
it the wrong way: every esplora wallet got a button that 501s.
canRescan carries the distinction explicitly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
registering an xpub only indexes it from that moment on, so an imported
seed with history read as a confident zero: every call succeeded, the
coins were simply absent. scantxoutset walks the node's current utxo set
directly and finds them regardless of when the account was registered.
runs all four script variants sequentially — the funds could be on any
one — and surfaces progress through the existing SyncState channel so
the balance says "scanning" rather than nothing. auto-fires on an
imported mnemonic only; a generated seed has no history to look for.
recovers spendable coins, not spent history.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Background tasks already had a row in the transcript, but a row scrolls away — and a
task started ten minutes ago is precisely the one you want to keep an eye on. The same
`role: 'task'` rows now also drive a tray docked above the chat input: a chip per task,
running ones pulsing, finished ones dismissable.
Clicking a chip opens what the task is actually doing right now. Nothing about that
crosses the wire between `task:started` and the notification, so it is read from the
file Claude Code streams the task into:
$TMPDIR/claude-<uid>/<project-slug>/<session-uuid>/tasks/<task-id>.output
For a backgrounded shell that file IS the log; for an agent it is a symlink to the
subagent's own transcript, which is ordinary session JSONL and so parses with the
reader we already had. Both kinds are therefore reachable from one directory.
Resolution is by task id alone, deliberately: the client learns a task id from
`task:started` and nothing else — officer's per-connection session key is not Claude's
session uuid, and the uuid only arrives with the turn result, long after the tray needs
to show the task. A task that has not written anything yet answers 200 `{kind:'pending'}`
rather than 404, because that is the ordinary first second of a task's life.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
it said fourteen; there are nineteen. photos, notify, caldav, memos and jellyfin have landed
since it was last written. also names the two mybiblepal entries that share this pm2 daemon
and belong to a different project, and points at `pm2 jlist` as the source of truth so the
next reader does not trust a list that goes stale every time a sidecar ships.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
nothing the player did ever reached jellyfin, so nothing ever appeared in continue watching
or in the currently-playing list. the proxy was fine — a progress POST through the sidecar
moves the resume point upstream and answers 204. the client was the problem.
useClient() rebuilds its verbs on every render, so `report` had a new identity every render,
so `sendReport` did, so the effect whose cleanup reports the final position re-ran on every
render — and that cleanup nulls planRef. planRef went null a few milliseconds after the
stream opened and every report after that returned early. the ten-second heartbeat never
fired either: its interval was cleared and restarted on every render, and timeupdate renders
about four times a second.
both are now held in refs, and the two effects have honest dependency lists. same treatment
for the negotiate mutation, which react query also rebuilds per render.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the sidecar now describes every audio and subtitle stream on the chosen source. a subtitle
jellyfin can deliver as an external file gets a /_jf webvtt path with the embedded ApiKey
stripped; one it cannot gets the reason instead, so the picker can show it disabled rather
than pretend it does not exist. colliding labels get their stream index — two tracks called
"English - ASS" are a coin flip otherwise.
quality is a named rung rather than a number. a bitrate the owner picks maps onto the rung
at or below it, which is what stops jellyfin's ResolutionNormalizer from inventing a
resolution from an unrounded number. auto stays absent so a copyable source still copies.
the control bar is ours in all three transports. progressive has no timeline the browser can
render, and an audio-track or quality change is a re-negotiation with the server rather than
something a <video> knows how to do — one bar for all three is what keeps those from being
three different players. the menus are hand-rolled because a radix dropdown portals to
document.body, which is outside the fullscreen element and would be invisible exactly when
the player is most likely to be used.
no scrubber thumbnails: they come from trickplay tiles and this server generates none.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
a live transcode has no length and no byte ranges, so the progressive mp4 the player
used could not be seeked — a scrub restarted ffmpeg at a new offset. jellyfin's own
HLS playlist is VOD and spans the whole runtime, so seeking it is a segment request.
hls.js is checked in rather than installed. installs are frozen so that adding a
package is a reviewed act, and a committed file also has no install-time hook, which
is the vector the 2026-08-04 npm worm used. provenance, hashes and the update recipe
are in vendor/README.md; the tarball sha512 matches the registry's published integrity.
the sidecar now overrides VideoBitrate and MaxWidth on the TranscodingUrl jellyfin
hands back, for the same reason progressivePath computes them: jellyfin resolves that
bitrate from MaxStreamingBitrate (~119 Mbit, the ceiling that exists to let a stream
copy through) and sets no width, which asks a CPU-only container to encode 4K.
safari is deliberately not given the m3u8 — segment URIs are relative and would not
carry the ?token=, and it cannot set an Authorization header the way hls.js can.
progressive stays as the fallback for any browser without MSE.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
a transcode url without `videoBitrate` is not "let the server choose".
jellyfin resolves the missing value to 0 and runs ResolutionNormalizer
over it, which maps a bitrate onto a resolution — 0 lands on the bottom
rung. verified here: a 3840x1600 hdr source came back through
scale=...min(max(iw,ih*a),416)... with -maxrate 0.
the number has to be picked per source, because the two cases pull
opposite ways: a cap is a reason for jellyfin to REFUSE a stream copy,
and the absence of one is what makes a cpu-only 4k encode hopeless. so
copyable h264 asks for a ceiling nothing hits and no width, and anything
being genuinely re-encoded asks for 1080p at 12 mbit.
verified both branches against the ffmpeg command jellyfin logs: h264 mkv
→ -codec:v:0 copy, hevc 10-bit hdr → libx264 at 1920 wide, tonemapped.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
adds the /jellyfin screen and its two panels (nav + view) on top of the
jellyfin sidecar, following the photos/invoiceshelf shape.
the transcode path is a progressive mp4 rather than hls, so no hls.js
dependency is added under the frozen-install rule. that stream has no
length and no byte ranges, so the player owns its own scrubber and seeks
by re-negotiating at a new startTimeTicks, tracking offset + currentTime;
a direct file keeps native controls. the hls url is still returned, so
switching later is a player change only.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
officer-jellyfin owns the whole Jellyfin contract: the instance URL, the access
token, the Jellyfin user it belongs to and the DeviceId its sessions are keyed
by. The platform side is a 17-line proxy holding no credentials.
Servers are a registry, not a single row — this machine runs four instances and
the owner switches between them. The password is never stored: it is traded once
for an access token through AuthenticateByName, and only that token is persisted,
encrypted.
Two doors. /_officer/* is a hand-written JSON façade for the things the browser
should not have to know — the user id in the path, the Fields lists that decide
whether a grid has posters, the PlaybackInfo negotiation. /_jf/* is a GET-only,
allow-listed byte pass-through for images, video, HLS and subtitles; it keeps
Jellyfin's own paths because a master playlist references its segments
relatively, so any renaming would mean rewriting m3u8 bodies.
TranscodingUrl arrives with api_key=<access token> in its query string and would
otherwise be handed straight to a video element. It is stripped before anything
is returned; the pass-through re-adds the credential as a header.
Video only — Officer's own player owns audio, so music collections are filtered
out of the library list.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
bunfig.toml sets [install] frozenLockfile = true, so `bun install` resolves from
bun.lock and nothing else — it fails rather than quietly picking up a newer
version, including a transitive one nobody chose.
Config rather than a documented habit, because a supply-chain compromise does not
wait for the one time somebody forgets a flag. On 2026-08-04 eleven cache
packages — keyv, flat-cache, file-entry-cache, cacheable-request, cache-manager,
the @cacheable/* scope and ecto — were published with a preinstall dropper that
harvested npm and GitHub tokens, AWS and Kubernetes credentials, SSH and PEM
keys, .env files and .claude/settings.json, then republished itself through any
npm token it found, reaching 434 further packages across 1381 versions. This
machine was unaffected only because nothing had installed since 2026-08-02.
Verified on bun 1.3.10 rather than assumed, and the first result was wrong: with
NO lockfile present, neither the flag nor the config refuses — bun simply creates
one, so an initial test made the setting look ignored. Against a lockfile that no
longer satisfies package.json, both exit 1 with "lockfile had changes, but
lockfile is frozen". The config is honoured; the earlier reading was a bad test.
`bun install` in this repo still reports 1058 installs, no changes.
Applied to monorepo-mobile too, where flat-cache, file-entry-cache and keyv are
present as eslint/got transitive deps — at versions old enough to be unaffected,
which was luck rather than design.
Documented in platform/CLAUDE.md and the workspace root, including the part that
matters most: do NOT add --no-frozen-lockfile to a script, a Dockerfile or CI to
make the error go away. The error means the lockfile and package.json disagree,
and an unexplained lockfile change in a diff is precisely the signal this exists
to produce.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
99.97% printed as "100.0%" beside a still-blue bar, which reads as a stuck
torrent. only a genuinely complete fraction shows a hundred now, as "100%".
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the album view already carries the transport, so the full-width dock was a
second bar costing the workspace a row. the host stays mounted (it owns the
engine) and only withholds its bar on /music; MusicMiniBar draws the scrubber
at the foot of the library panel, with play/pause and the lyrics toggle for
when you browse away from the album that's playing.
player-time now publishes duration alongside position so a scrubber outside
the host's tree can render without a 60hz state channel.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
they are indexed like any other folder — all six carry cover:true and serve a
jpeg — so the root view uses the same card as the folder grid instead of flat
gradient tiles. a root that holds tracks directly gets the hover play button
for free.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
same useGlobal MUSIC_PLAYER state, so on the album that is already loaded it
shows pause while it plays and resumes rather than restarting from track one.
any other album still starts from the top.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the sheet forces the dark theme tokens on its own subtree — they are CSS
variables scoped to a .dark ancestor, so a light theme would otherwise paint
near-black text on black. inactive synced lines drop the /50 and use plain
muted-foreground, the same grade as the artist under each track title.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the microphone is the same switch in two places — the play dock and the album
header — so its state moves to a channel seeded from localStorage. turning it
on splits the /music detail panel in two with a nested WorkspaceLayout: a fixed
layout, components keyed by panel id, no persistence and no registry entries.
the album view is handed to the left panel through context, so the split moves
the same element instead of remounting it and refetching the album.
playback position now reaches the pane through a module-level publisher rather
than props — it lives outside the player's subtree, and the feed ticks every
animation frame. the pane subscribes and re-renders only when the active line
changes, about once a line.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Ports the mobile app's lyric parser (packages/core/src/services/lyrics.ts) to the web player and
expands a sheet above the play dock. Synced .lrc lines highlight, auto-scroll and seek on click;
plain .txt scrolls by hand. The server side already served all of this — /api/music/lyrics and the
indexer's embedded-USLT extraction — so nothing changed behind the proxy.
The sheet lives in the dock rather than a /music panel because the dock is mounted app-wide: lyrics
follow the music onto every screen, and no saved workspace layout has to be migrated to see it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
pastilhasandClaude Opus 5