OPENCODE_TURNS=serve picks the new engine; unset keeps the subprocess, which is the default
and stays the default until this has been lived with. A bad evening should cost one restart,
not a revert. Claude is a different sidecar and is untouched.
Verified end to end through the real chat socket:
session:init -> tool:start(bash) -> tool:result -> assistant:delta x3 -> assistant:text
-> result, cost in=304 out=73
Those deltas are the first token streaming an opencode turn has ever produced in officer.
Stop is now an INTERRUPT: the turn ends and the session survives — verified by sending a
second prompt to the same session afterwards and getting an answer, which killing a
subprocess could never do.
Reads the LIVE global stream rather than the durable per-session one, because it is a strict
superset — same tool.called, tool.success, step.ended, text.ended, plus the deltas that are
the whole point. Global means one socket carries every session, so everything filters on
sessionID; one subscription is shared for the process rather than one per turn.
A turn ends on step.ended with finish != tool-calls. tool-calls is a step boundary MID-turn,
and treating it as terminal would cut every tool-using conversation in half.
delivery is stated explicitly as queue because it DEFAULTS to steer, which injects into a
running turn — wrong for an ordinary send, where two quick messages would merge into one.
Wiring steer to the button that means it is phase C.
What phase B does not do: read the durable stream. The sidecar still commits every event to
chat_session_events as it arrives, so durability is unchanged, but recovering a turn this
process never saw needs the ?after= cursor and is its own change.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The mapping half of the serve migration, written and pinned before anything depends on it,
so the switch-over is not also the moment the parsing turns out to be wrong. Nothing routes
through this — turns are still opencode run subprocesses, and the claude path is untouched.
The finding that matters: the serve publishes each turn TWICE, and reading the wrong one
makes it look like it cannot stream at all.
/api/session/{id}/event?after= durable, per session, replayable, durable.seq on every
event, whole values only, NO deltas
/api/event live, GLOBAL, ephemeral, carries text.delta and
tool.input.delta, no cursor
Same turn: 13 events durable, 21 live, the difference being 3 text.delta and 5
tool.input.delta. I probed the per-session one first and nearly recorded "no streaming" as
a fact — it would have removed the main reason to migrate. The split maps exactly onto what
officer already does for claude: durable to chat_session_events, live to UI deltas. The cost
is that the live stream is global, so a consumer must filter on sessionID.
tool:start is emitted on tool.called, not tool.input.started, because only tool.called has
the resolved input object — the input arrives as JSON fragments ({"comman) and a tool row
rendered with half-parsed arguments is worse than one that appears a moment later.
step.ended with finish tool-calls is a step boundary MID-turn, not the end of the turn, so
nothing terminal is emitted for it. Treating it as the end would cut every tool-using
conversation in half.
Fixtures are verbatim captures from 1.18.16. Replaying both real streams through the mapper
reconstructs the turn identically from each, with the reassembled deltas exactly equal to
the committed text and identical cost, and zero unrecognised events.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Bun.spawn throws on a missing or non-executable binary rather than resolving to a failed
process, and that throw escaped runOpenCodeTurn entirely — past the bookkeeping, out of the
sidecar command handler, with no opencode:event ever emitted. The browser sat on a spinner
nothing could end, because the code that ends turns had not been reached.
A wrong OPENCODE_BIN is the ordinary way to get there, so the message names the path it
tried: that is the difference between a fix and a debugging session.
Also records that messageCount is not a gap. SessionList renders an OpenCode badge in place
of the count for those rows, so the hardcoded 0 never reaches a screen, and computing a real
one would cost an HTTP call per listed session — the session record has no count field — to
populate something nothing shows.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
opencode keeps credentials in two unrelated places. The CLI, opencode run and the legacy
/session surface read auth.json. The newer /api surface — the one with steer, queue,
interrupt and a resumable per-session stream — reads its own integration store and knows
nothing about that file.
With none connected it does not fail. It falls back to what needs no credential, the free
tier, and a request for a paid model is never executed: prompt accepted, admitted, prompted,
then no step, no error, no message, forever. That silence cost most of an afternoon and would
cost it again on every new machine — alpha included.
So the sidecar does it, rather than depending on someone having run a curl. Best-effort and
never blocking: turns go through opencode run, which reads auth.json and does not care.
Retried, because /api/health answers before the integration store is ready — the first
version of this shipped without a retry and failed on its very first real boot with a 500,
while the identical request succeeded seconds later. Only 5xx retries; a 4xx means the
request is wrong and repeating it just prints the same complaint six times.
Verified by deleting the credential, restarting, and running sonnet on the new pipeline with
no manual step.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Follow-up to the /vaultwarden mount: the suffix is superfluous if officer can tell a
bitwarden client apart, and it can.
Most of vaultwarden surface does not collide at all — /identity, /notifications, /icons and
/events belong to it and to nothing here, so those are served at the root by path alone, no
sniffing. Only /api collides (vaultwarden has /api/settings/domains, officer has
/api/settings), and there the client says who it is: every bitwarden client stamps
Bitwarden-Client-Name, older ones Device-Type.
Trusting a client header is fine because this is ROUTING, not authentication — the worst a
forged one achieves is reaching vaultwarden, which then demands its own credential exactly
as it would have. Nothing is authorised by it.
Registered before /api so it wins for a bitwarden client, and narrow enough that an ordinary
officer request never matches. Verified: /identity reaches the proxy, /api/sync with the
header diverts, /api/chat/models without it still answers 401 from officer, and the SPA is
untouched. /vaultwarden still works for anything that prefers an explicit path.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
So the bitwarden browser extension can point here and the separate public vaultwarden
hostname can be taken down.
/api/vault cannot serve it: that router requires an officer session and REPLACES the caller
Authorization header with a server-held vaultwarden token. Right for our own clients — the
device then holds no vault credential — and impossible for a third-party client that gets
its own token from /identity/connect/token and has nowhere to put a platform JWT.
So a separate mount rather than a mode of that router: blending them would put an
unauthenticated branch inside the authenticated path. This one forwards Authorization
untouched and rewrites nothing.
Leaving it open is not a new exposure — everything here was already reachable at the
vaultwarden URL it replaces, behind the same master password, and officer cannot add a check
it has no credential for. It is also going behind tailscale.
Temporary. The end state is our own extension reusing @officer/vault, which already runs as
a plain JS bundle outside react native (the iOS autofill extension hosts it in
JavaScriptCore), against the /api/vault/session/login broker — then nothing addresses
vaultwarden directly and this mount is deleted rather than adjusted.
Needed its own entry in server.tsx: only listed paths reach hono and the rest fall through
to the SPA, so without it the endpoint answered 200 with the react shell — a missing route
that looks like a working one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Regression from my own B7 change, reported within the hour: turns collapsing to "turn
completed without output" and coming back only on refresh.
B7 stopped resume-cursor defaulting an unidentified session to claude-code. Correct for the
durable cut-off row, wrong for adoption: useChat sends model only if modelRef.current is
set, so a reconnect without one is routine, not exotic. Declining to adopt left the socket
unbound to the live session, so the running turn output went nowhere — and a refresh looked
like a fix because it rebuilds from the durable log.
Adoption is about DELIVERY and must be generous; only the durable write needs certainty. So
adopt on the default again, mark it as an assumption, and skip the cut-off check on it.
That keeps B7 fixed — no false "agent went away" written against an opencode session — with
no unbound sockets.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Same shape as the claude side, which has never shown a live row without a name.
OpenCode titles a session from the conversation and does it well, but asynchronously — so
for the whole time a turn is RUNNING, which is exactly what /chat/live shows, the session is
still called "New session - <ISO>". Its own title wins the moment it exists; until then the
row falls back to the prompt that started the session.
Kept per sessionKey, first turn only, so it stays the name of the conversation rather than
following whatever was asked most recently. Dropped with the session id it sits beside.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
OpenCode titles a session from the conversation, but asynchronously — a finished turn of
ours ended up called "Single color in oc-red2.png", better than anything we would generate.
Until then the session is literally named "New session - 2026-08-10T15:44:17.178Z".
That window is exactly when a session is most visible: /chat/live shows turns that are
RUNNING, so the placeholder is what the panel catches, and a live row was being labelled
with a timestamp string.
Recognise it and treat it as untitled, so the good name arrives on its own. Passing --title
on the run was the other option and is worse: it fixes the transient case by permanently
replacing opencode own title with a truncated prompt, degrading it where it lasts longest.
A pattern match rather than startsWith, because a genuine title is allowed to begin with
those words. Two defects the tests caught while writing them: a whitespace-only title was
not treated as unnamed, and the mapping let undefined through where a string was required.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
B4 properly. The composer gate was the honest stopgap; this is the fix. opencode run takes
attachments with --file, so images work on the subprocess path we already use — the parity
doc had them down as phase 4, behind the serve migration, and they were not.
The bug was one omission: handleOpenCodeChat`s msg type had no images field, so the browser
sent them, the bubble rendered them, and they stopped at that signature. Nothing reported a
loss anywhere.
Attachments are paths, not inline data, so the sidecar spills each image to a temp file for
the length of the turn and removes it in settle — the same place every other per-turn
resource is released, so a killed or superseded turn cleans up too.
The load-bearing detail is `--` before the prompt: --file is an array option, so without the
separator the prompt is eaten as another filename and the turn dies with "File not found:"
followed by the entire message. Confirmed against the binary, and pinned by a test that
records argv from a stub.
list-models now reports each model own capability instead of a hardcoded false — opencode
publishes capabilities.input.image per model and nothing had ever read it. Defaults to false,
so a model that does not declare it keeps the affordance hidden.
Verified end to end: a red png sent over the chat socket to opencode/claude-sonnet-4-6 came
back "Red".
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Email installs end to end now, which was the point of picking it as tier one: no container, no external
wiring, so the machinery is exercised without the provisioning half.
Verified against the running system, not asserted:
POST /api/app-store/email/install -> {"status":"installed","completed":["preflight","schema","process"]}
row -> email mode=config status=installed enabled=true
pm2 -> officer-email online
second install -> all three steps skipped, process not restarted
disable -> stopped
The server boots with the new router, which is the real test of the capability entry: totality.ts throws
before serve() if a mounted router has none, so booting IS the check passing.
pm2.ts shells out rather than importing pm2 as a library. PM2 is already the supervisor and the
ecosystem file is already the definition of how each process runs; a second thing in charge of that
means two supervisors disagreeing. It also means an owner can undo anything the app store did with a
command they already know. The one fact that matters: `pm2 start <name>` fails for a process PM2 has
never seen, so a first install starts from the ecosystem file with --only, and everything after goes by
name. Callers cannot know which case they are in, so startProcess decides.
Disable stops rather than deletes: a stopped process still shows in `pm2 list`, which is the honest
picture. Deleting would make a disabled sidecar indistinguishable from one never installed.
beginInstall returns the existing row instead of replacing it — that is what makes a retry a resume
rather than a re-provision — and clears lastError on the way in, so a UI never shows a stale failure
beside a working service.
The container half of enable/disable/uninstall is deliberately absent rather than stubbed silently: a
disable that leaves Immich running is a different thing from one that stops it, and the difference is
memory on the user's machine.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Implements the half of the template contract that faces the platform: answers go in as environment and
never as prompts, and results come back as OFFICER_RESULT_<KEY>= lines on stdout.
A line protocol rather than JSON because the same stream is the user's live log — it goes to a terminal
panel while the install runs. A script that must emit clean JSON cannot also narrate, and one that emits
both needs a framing convention anyway. This mirrors the @@officer:progress@@ sentinel the job runner
already uses, with the same rule: marker lines are plucked out, everything else passes through.
parseResults is pure and tested against the realistic near-misses: a line that MENTIONS the prefix
without starting with it, an empty value (Transmission with no RPC auth returns exactly that, and blank
is a real answer), a value containing `=` (splitting on every one would truncate a credential), and a
prefix with no assignment (a script bug — skipped rather than stored as a blank key).
Verified end to end against a real script: environment reaches it, stderr is forwarded (docker compose
writes its progress there, so dropping it would hide most of what a user watches), OFFICER_NONINTERACTIVE
is set so a script that would block fails loudly instead of hanging behind a web form, and a non-zero
exit is reported with the tail.
Notes an artifact rather than hiding it: the two streams are pumped concurrently, so the error tail can
interleave differently from real time. The live log is correctly ordered; only the summary can read out
of order. Serialising the pumps would make a script that writes heavily to one stream block on the other.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Install spans a container start, a health wait, an upstream API call and a process start. Any can fail,
and one of them — a token only a human can mint — is EXPECTED to stop the run. A straight-line function
has two bad options there: unwind everything, or leave a half-installed service that neither works nor
uninstalls, which is the state users cannot get out of.
So each step is named, completion is persisted, and running install again resumes. planSteps is a pure
function of (entry, mode) and the effects are injected, which makes ordering, resume, blocking and
failure testable with no Docker, Postgres, PM2 or Immich in sight. 15 tests cover exactly the behaviour
that only appears when something goes wrong.
Two rules are enforced by the plan rather than remembered at call sites: 'existing' never provisions, so
pointing at an instance the user already runs cannot start a container; and the members step is omitted
entirely for a service with no user concept, so a Transmission install does not report a step that did
nothing — which reads as a silent failure to anyone debugging a member's access.
`blocked` is a first-class outcome, not an error. For Immich the container is up and healthy and only
its own UI can mint a key; calling that a failure would make a normal install look broken and invite the
user to tear down a working container. The blocking step is deliberately NOT recorded as complete, so a
resume re-runs the step the human just answered.
Results feed forward — provision discovers the URL that connect writes down two steps later — over a
copy of the caller's values, so a failure halfway cannot rewrite what an earlier attempt achieved.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The owner installs, but a server may already have members, and a member added next month needs the same
work. So the unit is (service × member) reachable from two triggers — install a service, provision
existing members; add a member, provision installed services — rather than a loop inside the installer.
Only handling the first works on day one and rots.
No new table. A member is provisioned exactly when they hold a service_connections row: their own
credential, url NULL, inheriting the instance from the owner's. That schema anticipated this before this
existed, and a second record of the same fact would only be able to disagree with the first.
Three outcomes, declared per catalogue entry so the installer never special-cases a service. `accounts`
is fully transparent. `none` is a single-tenant daemon with nothing to do — filtered before the
provisioning loop so callers can tell "nothing to do" from "did nothing", which look identical at a call
site and matter when someone is asking why a member cannot see a feature.
`invite` is not a weaker `accounts`, it is the correct outcome: Vaultwarden derives its encryption key
from the master password, so a credential we could mint would mean a vault we could read. Transparent
right up to where being transparent would be a defect.
The per-service work is an interface implemented beside each sidecar rather than a switch in core — a
central function growing a case per service is what would stop any of this shipping from its own
repository. Implementations must be idempotent, since both triggers can fire for the same pair and a
duplicate account upstream is not ours to undo. Deprovision is optional and defaults to leaving the
upstream account alone: deleting an Immich user deletes their photos.
Written assuming the vault's multi-user adaptation has landed. Today /api/vault is owner-only by an
explicit ownerGate, so a member is refused before Vaultwarden is reached — verified, and out of scope.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Each provisionable service gets a directory holding a compose template and a setup.sh. Deliberately the
shape a sidecar needs once it lives in its own repository: metadata, compose, setup script, schema.
The contract (templates/README.md): answers come from the ENVIRONMENT, so the web form fills them in and
a person on a VPS is prompted only for what is missing, and only on a TTY — one script for both, not two
code paths. Idempotent, writes only inside its own directory, streams progress on stdout (the installer
pipes it to a terminal panel), and returns results as OFFICER_RESULT_<KEY>= lines so nothing has to
scrape a log.
House conventions throughout: relative bind mounts so data sits beside the compose file rather than
hiding behind `docker volume inspect`, containers running as the installing user so downloads are not
root-owned, loopback-only ports unless the service's whole job is inbound connections, and no external
networks — the owner's own composes attach to an `nginx` network that a fresh VPS does not have.
Transmission verified end to end on this machine, on non-conflicting ports, then torn down: renders,
starts, waits, reports. Its health check accepts 409 because Transmission rejects the first request by
design — only-200 would have waited out the full timeout against a working daemon. Re-run produced
exactly one container, and files landed owned by the user rather than root.
Vaultwarden covers the case where we GENERATE the credential rather than asking for one. An existing
token is reused, never rotated, because rotating during a resumed install would lock the owner out of
the admin page. The Argon2 hash has its `$` doubled or compose interpolation mangles it. The token is
not returned to the platform at all — the vault sidecar proxies the Bitwarden protocol and never needs
it, and a secret we do not hold is one we cannot leak.
Corrects the design doc, which assumed provisioning always knows the connection. Three shapes: we set
the credential, we generate it, or a human must mint it in the service's UI afterwards (Immich, Jellyfin,
Memos). The third makes "provisioned and running but not yet connected" a real state rather than a
failure.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An install that discovers a missing dependency halfway through has already made a directory, possibly
started a container and written a row, and then has to unwind — leaving the user with something that
neither works nor uninstalls. A 30ms check first is worth most of that.
Verified while writing this: nothing in scripts/ installs Docker, and nothing checks for it.
setup-dockers.sh invokes `docker compose` with no preflight, so a fresh host without Docker fails
partway through setup with a bare "command not found". Recorded in the design doc rather than fixed
here — the intended fix is a setup.sh per sidecar, which is also what a sidecar needs once it ships from
its own repository.
`docker compose version` is the probe, not `docker --version`: the latter passes with a dead daemon,
which is the failure people actually hit. "Not installed" and "daemon unreachable" are reported
separately because the remedies differ.
Checked per MODE, not per entry. A host without Docker can still install Photos by pointing at an Immich
somewhere else; refusing the whole entry is the over-strict check that makes people work around the
installer instead of using it.
Dropped `requires: 'docker'` from the catalogue type. Needing Docker is exactly "this entry can
provision", which `modes` already says, so declaring it twice invites the two to disagree. Derived by
needsDocker instead, and a test asserts the derivation matches every entry.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The install layout a machine should have, seasoned owner or not:
~/officerdev/
platform/ the app
data/ DATA_PATH
dockers/ services the app store provisioned
capabilities/ the file-based item store
One root, everything under it. OFFICER_ROOT derives from DATA_PATH rather than being a second variable
that has to agree with the first.
Deliberately not `~/dockers`, where a seasoned user already keeps their own estate — 47 services on this
machine. That separation buys two things. Containers the app store created are distinguishable from the
user's own structurally, rather than by a naming convention we would have to enforce and they could
break. And we never reason about someone else's compose files: the store does not scan, adopt or modify
anything outside its own directory.
That also simplifies "I already have one of these" — it is answered by the user giving a URL, never by
us finding a directory and guessing whose it is. An earlier draft had the installer adopting existing
directories, which meant reading, and potentially writing over, services Officer did not create.
This development machine predates the convention and derives an ugly-but-correct path, since the project
sits inside ~/dockers/officer.dev. Still isolated, still one root. New installs get the clean shape.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
First slice, on a worktree branch so none of it touches the tree the live server runs from.
`sidecar_installs` — server-level, no userId, because a sidecar is one process serving the machine.
That is the line that keeps the model coherent for several users: installed is server-level and
owner-only, configured is per user in service_connections. A member can use Gitea without being able to
install it or point it somewhere else.
`installed` and `enabled` are separate because they answer different questions, which is what gives the
reversible middle ground: disable stops the process and keeps container, config, schema and data.
`completedSteps` makes install resumable rather than merely retryable — the failure mode being designed
against is a half-installed service that neither works nor uninstalls.
The catalogue is data, not code: no functions, no compile-time coupling, because the same shape has to
arrive as JSON from marketplace.officer.dev later. Its test pins it to the real estate — it offers
exactly the processes the light profile excludes, names processes that exist, and claims capabilities
that exist. That last check earned itself immediately: it caught `vault` (no capability at all — it is
EXEMPT because Bitwarden clients carry a Vaultwarden bearer, not a platform JWT) and `notify` (which
does have one, where I had written null).
Docker templates follow the convention already in use across 47 services in ~/dockers: a directory per
service, compose inside, relative bind mounts so data sits beside it, USER_UID/USER_GID as the owner.
An existing directory is evidence of an existing install and must be adopted, never overwritten.
Records what Phase 0 must not foreclose: a remote marketplace, sidecars moving to their own
repositories, and third-party plugins — including the note that catalogue.test.ts pins Phase 0's
invariant rather than the design's, since that relationship inverts once sidecars leave this repo.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The previous commit migrated only when SQLite was completely empty, on the assumption that a non-empty
file is an authoritative one. A real account disproved that within minutes of it landing.
The older Gmail backfill had already written SOME keys into SQLite — last_sync_at and the uidvalidity
set — and never the imap_lastuid ones. So the file was non-empty and half-migrated at the same time,
all-or-nothing skipped the migration, and nine imap_lastuid keys stayed only in Postgres. A missing
lastuid makes the next sync refetch that folder from UID 1: on the mailbox this was found on, 18,755
messages and 6.9 GB.
Now merged per key with the file always winning a conflict. That keeps the property all-or-nothing was
protecting — a restored older emails.db still overrides a newer Postgres row for every key it has, so it
cannot be advanced past mail it does not contain — and adds the keys the file never had, which are
exactly the ones whose absence is expensive.
Verified against the live account: all 22 Postgres keys present afterwards, imap_lastuid:INBOX restored
to 208407, and last_sync_at left at the file's older value, so it re-checks a fortnight rather than
skipping it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The messages were in emails.db and the position — last_sync_at, and per-folder uidvalidity/lastuid —
was a jsonb column on email_accounts in Postgres. Two stores for one fact, with an edge that only shows
up when you try to move a mailbox to another machine.
The expensive part of an email account is the first sync: hours of IMAP for a large mailbox, which is
exactly why "copy emails.db to the new server" is the obvious way to bring one across. With the position
in Postgres that silently does not work — the new server's column is empty, !last_sync_at says first
sync, and the whole mailbox downloads again on top of the one just restored.
The other direction is quieter and worse. Restore an OLDER emails.db while Postgres holds a NEWER
position and the sidecar skips every message between the two, permanently, because nothing looks below
lastuid again. Re-syncing is slow; skipping mail is data loss nobody notices.
Not a new idea — the Gmail path already read SQLite and fell back to Postgres, backfilling so the
fallback was taken once. Only the IMAP path had not followed. This extracts that pattern so both use one
copy, and unifies the isFirstSync fork in accounts.ts, which is how the two drifted apart to begin with.
The file wins over Postgres, always, and only migrates when it holds nothing at all. Topping up a
partial position from Postgres would reintroduce precisely the divergence this removes.
email_accounts.sync_meta is kept and marked legacy rather than dropped: it is the one-time backfill
source for every account created before this, and dropping it would strand any that has not synced
since. Nothing writes to it now.
11 tests on the migration, aimed at both expensive failures — migrating when we should not, and failing
to migrate an account that predates the change.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The serve has a second, newer API surface nobody here had looked at, and it publishes
exactly what the parity doc calls impossible under stdin ignore: delivery steer and queue
on POST /prompt, an interrupt that does not tear down, and a per-session event stream with
an after cursor — the durable-replay machinery officer hand-built for claude, as a
primitive. That would have made migrating obvious.
It does not execute. A prompt is accepted with an admittedSeq, stored, emits
prompt.admitted and prompted, and then never steps. Ruled out separately: the model, the
permissions (build is *:allow, no pending requests), the per-request location (the surface
is location-scoped via header or a deepObject query, supplied everywhere, no change), and a
config gate. The legacy POST /session/id/message?directory= generates fine in 17s, so the
serve itself works — only the new pipeline is inert. session.next.* is the tell.
And not a version problem, which is the part everything here had backwards: this Mac runs
1.18.11 and alpha runs 1.17.9, measured. The dead pipeline was tested on the NEWER binary.
The original "this server runs 1.17.9" meant alpha and was copied to a machine where it was
false; corrected in runner.ts and the test.
So building against it now would produce code that looks finished and does nothing, which
is the failure mode this project keeps rediscovering. One request reopens the question
after any upgrade, and the doc names it.
Also de-flakes the lifecycle tests: they spawn real processes, and a fixed sleep(750) went
red once on a machine busy running these probes. Presence assertions poll now.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
They were permanently unnamed, and the two halves needed to name them already existed on
officer: the sidecar reports its own sessionKey because that is all it has, while the ses_ id
arrives separately over opencode:session and is recorded in opencode/state.ts. Nothing joined
them. /chat/live joins them now, so no protocol or sidecar change — widening
LiveOpenCodeSession would have meant sending the sidecar a fact it told officer in the first
place.
One list call names every row rather than one transcript load each, and it is skipped when
nothing is running or no id has been reported, so an idle Live panel never touches the serve.
Verified against a real turn, which also showed the design working as intended: the first
poll has no id yet and shows nothing, the next shows title and cwd. That window is real and
short, and showing nothing beats showing a key the user has never seen.
Worth knowing: opencode titles its own sessions "New session - <ISO timestamp>", so the row
is located but not meaningfully named. That is genuinely its title, not a bug here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
B8, both halves. They share index.ts, so they share a commit.
In-flight turns: `opencode run` is spawned, not supervised, so pm2 restart officer-opencode
left every turn ALIVE — reparented, still spending tokens, still writing files as the agent,
with the only reader of its stdout gone. The transcript stopped mid-tool-call, which reads
as the agent hanging.
stopAllOpenCodeTurns kills them and settles each synchronously, because the caller is about
to process.exit and nothing waiting on proc.exited would ever run. Settling writes a reason,
so a reload after a restart explains itself instead of trailing off. Turns are stopped BEFORE
the connection is destroyed — that write travels over it — and the flush is bounded, since
losing the explanation is bad but hanging the restart is worse.
Stale serves: the sweep read /proc, so it was a no-op on macOS and orphaned serves piled up,
one per unclean exit, each holding a port. Added a pidfile sweep alongside it. A pid we wrote
ourselves needs no cwd guard to prove it is ours, which is the part ps cannot answer portably
(macOS would need lsof), and a serve started by hand is never in the file.
The guard checks command AND subcommand: matching the word serve anywhere in the line would
sweep a running turn whose prompt merely mentioned it. Fixtures are real ps output from both
machines, not invented. Split into serve-sweep.ts because index.ts spawns a serve at module
scope, so a test importing it would start one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
B7. `msg.model || DEFAULT_MODEL` declared every session without an explicit model to be
claude-code, and the parity doc recorded only the visible half of what that cost.
The durable false cut-off is real: endTurnIfAgentIsGone asked the claude sidecar about a
key it had never held, was told false, and wrote "the agent went away" into a turn that
was running fine. It survives reload, because surviving reload is what that row is for.
The same default also handed the session to adoptOrphanedSession as a claude one, which
subscribes it to that sidecar bus and pins session.model — so an opencode turn output
never arrived, and stopping it called killClaude on a key that sidecar never had. A stop
button that silently does nothing.
decideResume makes both rules explicit: the server record beats the client claim, and an
unknown harness stays unknown — no adoption, no cut-off check, just the replay. Silence
is the safe failure when the wrong answer is written durably.
DEFAULT_MODEL stays in handleAttach and is now commented as to why: that path reached its
sessionId by asking the claude sidecar to resolve a claudeSessionId, so only claude could
have answered.
First test in api/chat, which had none. websocket.ts has no seam to drive the handler
through, so the decision is extracted and tested; the wiring around it is not covered.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two comments describing behaviour the code does not have.
LiveOpenCodeSession had been inserted between LiveClaudeSession and its docblock, so a
comment about isGenerating, pendingTasks and the idle GC read as documentation for the
OpenCode type — where it is contradicted by the correct comment directly beneath it.
Moved below, and it now states that it carries no ses_ id.
That absence is the point: /chat/live claimed title and cwd come from the session store
"so a turn whose id has not been reported yet shows unnamed". Nothing is looked up, and
there is no id here to look one up with. They are null permanently, not until-known.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A replaced turn is killed but dies asynchronously, so its proc.exited fired long after
the replacement was registered under the same sessionKey — and then ran the whole
completion path against it: emitted "OpenCode exited with code 143", which the sidecar
commits to chat_session_events so a false failure became permanent history, then deleted
the replacement from `running`. That blinded the new Live panel, made the stop button a
no-op and orphaned a process nothing could reach.
Mark the handle before killing it, retire it silently, and identity-check the delete —
a superseded turn does not own that key any more.
Two leaks in the same family, found while fixing it. An early return would not have been
enough: both watchdogs call finish, so the armed 10-minute hardTimer would have fired an
error at whichever turn held the key by then. And handleLine had no `done` guard, so
stdout still draining from the killed process was emitted under the replacement key.
Reproduced before fixing. The lifecycle tests need no real opencode — RunnerConfig.bin
takes a shell script that sleeps. The control test pins that an ordinary non-zero exit
still reports an error, so the guard cannot overreach.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The Live panel asked `claude:list` and nothing else, so a running OpenCode turn was invisible — the panel
claimed to show what the agent is doing and silently omitted half of it.
Adds `opencode:list` / `opencode:sessions` and merges both harnesses in `/chat/live`, asked in parallel,
each failing toward empty so one sidecar being down contributes nothing rather than breaking the panel.
The OpenCode row is deliberately thinner than the Claude one rather than faked into parity:
isGenerating always true — a subprocess exists only while it generates, so there is no "merely open"
pendingTasks always 0 — `opencode run` has no background-task concept; reporting a number would
suggest a capability that does not exist
title / cwd null — the session store is keyed on the `ses_…` id the runner reports, not on
our sessionKey, so an unreported turn shows unnamed rather than guessed
This is the incremental option from docs/opencode-serve-path.md — enumeration without moving turns onto
the serve, so it buys the Live panel with no warm sessions, no SSE loop and no lifetime questions.
NOT verified end to end: no OpenCode turn was running to enumerate, so the verb is wired and typechecked
but has never returned a non-empty list. See COMMS/BLOCKERS.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Path names: the serve's cwd is DATA_PATH/opencode_server, not opencode-sidecar. Two comments said
otherwise and would send the next reader to a directory that does not exist.
Version pin: the comment claimed "verified live against 1.17.9" as though that were a property of the
code. It is a property of whichever binary is installed, and this project already runs two — 1.17.9 here,
1.18.11 on the other machine. Says so now, and points at the test as the thing that actually enforces it.
Tests, the first on the OpenCode path. `runner.ts`'s NDJSON → ChatEvent mapping was described as pure and
untested; it was untested but not pure — it lived inside `handleLine` as a closure over `emit`, the
accumulated cost and a reported-session flag, so it could not be called without spawning a binary.
Extracted as `mapRunLine`, genuinely pure: line in, {sessionId, events, costDelta} out. The two concerns
that span lines stay with the caller, because they are not properties of a line — emitting the session id
exactly once, and accumulating cost across steps. Behaviour is unchanged.
11 tests over what the mapping forwards, what it drops and what it must not turn into NaN. The last one
matters: a missing `cost` on a step_finish would otherwise propagate NaN into the turn total.
Phase 1 is complete: dead code deleted (previous commit), comments corrected, tests added.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phase 1 item 7, done in the order the parity doc asks: read the dead design into a note, then delete it.
`docs/opencode-serve-path.md` records what `event-mapper.ts` and the SSE half of `client.ts` did, and
what a rebuild would want back from them — the delta model and tool-state transitions, which are exactly
parity Phase 3's token streaming rather than new work.
Deleted: `event-mapper.ts` entirely, and `subscribe`/the shared `GET /event` SSE loop, `createSession`,
`postMessage`, `abort` from the client, plus `isServerHealthy` from server-manager. All had no callers.
`client.ts` goes 200-odd lines to 99. What stays is the REST reads the chat list and transcript use:
listSessions, getSession, getMessages, deleteSession, renameSession.
While in there, Phase 2's blocking question turned out to be cheap to settle, so it is answered rather
than left open. The review asked whether the serve can take a per-request directory, since without one a
serve-based turn path would reintroduce the single-directory coupling that shelved this work:
POST /session?directory=/tmp/oc-phase2-probe -> directory: "/tmp/oc-phase2-probe" honoured
POST /session with directory in the BODY -> directory: "<serve cwd>" ignored
It is a query parameter on every /session* route. So the coupling is gone on both architectures and the
blocker is cleared. The note does NOT start the migration: which of the three options to take is a
product call, and it lays them out rather than presuming one.
The first probe put `directory` in the body and appeared to prove the opposite. Recorded in the note,
because it is the obvious way to test this and it gives a confident wrong answer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The sidecar seeded an AGENTS.md into the serve's project root telling the agent to read its working
directory from the "Working directory for this session" line in its system prompt. The run path sends no
system prompt, so there was no such line and the instruction had been inert since turns moved off the
serve to `opencode run --dir`.
What it was standing in for, `--dir` does properly — tested rather than assumed
(docs/opencode-phase0-review.md): `--dir` anchors the agent's own file operations, not just the process
cwd, and the anchor survives a multi-step turn with a write in the middle. Nothing replaces it.
The generated file is removed from disk too, not only from the code that wrote it; leaving it would have
kept feeding standing instructions to every session while looking, in the source, as though it were gone.
Also corrects the claim that all OpenCode sessions live in one server's project. True when turns
inherited the serve's directory, false now: one serve lists 7 sessions across several directories, which
is why the cwd filter has to read `directory` rather than assume a single one.
docs/opencode-phase0-review.md, item 2 — Phase 1.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three defects and one honest removal. Each was reproduced before being changed, as the doc asks.
B4 — images were offered and silently discarded. Every OpenCode model advertised `images: true`, the
composer gates on that flag, the bubble rendered the attachment, and `handleOpenCodeChat`'s message type
has no `images` field, so it never left officer. Flipped to false: 61 OpenCode models now decline, the
three Claude ones still accept. Plumbing them through OpenCodeRunParams stays Phase 4; advertising a
capability that does not exist is the part worth fixing today.
B5 — every OpenCode turn overwrote the previous turn's subscription handle without detaching it, so the
old session-scoped listener stayed attached and delivery doubled, tripled, and so on for any termination
that is not result/error/stopped. Deliberately NOT the Claude guard: Claude keeps one persistent session
and skips re-subscribing, while OpenCode spawns a fresh `opencode run` per turn, so a new subscription
each time is correct — detaching the old one is what was missing.
B6 — the sessionKey → `ses_…` map had no writer of deletions, so it grew for the process lifetime and a
reused key resumed a stale OpenCode session. Cleared in `deleteSession` only, never in `releaseSession`:
releasing means "let go, leave it running", and a returning browser must find the same `ses_…` again.
Phase 0 item 1 — the thinking toggle is removed rather than fixed. `thinking` is accepted on the wire
and forwarded by neither channel, so the control changed its own label and nothing else. Out of scope
for both harnesses by decision. The inert plumbing beneath it is left for a follow-up that touches the
socket contract; the props stay accepted-and-unread so no call site had to change.
Phase 0 is complete: B1, B2, B3 landed earlier; B4, B5, B6 and the selector here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two defects, one cause: the session type declared two fields opencode 1.17.9 does not return.
`GET /session` returns `directory` at the top level. There is no `location` object and no `metadata`.
Re-verified by reading the live server rather than the type.
So `metadata.officer.cwd` was compared against `undefined` for every session and the list filter matched
nothing — and since `cwdOf` substitutes a default when no `?cwd=` is given, the `!cwd` escape never fired
either. There was no configuration in which an OpenCode session appeared in /chat. Confirmed against the
running server: 7 sessions present, 0 returned, and the `OpenCode` badge in SessionList was unreachable
code. Now 1 of 7 is listed under the default chat dir, the other 6 correctly filtered to their own
directories.
And `location?.directory ?? ''` was likewise always '', so resuming a session reported no cwd and
relocated the conversation to the default chat dir — which matters because OpenCode rebuilds its
working-directory system prompt every turn. Detail now reads the session's own record via a new
`getSession`, alongside the transcript.
`officerMeta` and the `metadata` tag are gone rather than fixed: the only writer of that tag
(`client.createSession`) has no callers, because the sidecar creates sessions with `opencode run --dir`.
Tagging would have been a second source of truth for something `directory` already answers.
docs/opencode-parity.md B1 and B3. Its suggested fix — derive from `location.directory` — was written
against a field that does not exist; the doc asked for the version to be checked first, and this is why.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An upload over a few MB came back as a 400 with an empty body, no message anywhere, and nothing logged
by Immich, the sidecar or officer. It was a race, not a size limit. Immich judges an asset from its
first few KB and rejects immediately, then closes; both our hops were still writing the body; Node
treats the leftover bytes as a protocol violation and replaces the application's answer with a bodyless
`400 Bad Request` + `Connection: close`. The real message never reached the wire.
Measured before the change: streamed lost the message 1/4 at 8 MB and 4/4 at 32 MB — probability rising
with size, which is why small photos usually worked and a phone's video never did.
Both hops needed it. Fixing only the sidecar took 32 MB from 4/4 failing to 2/4, because the platform
proxy was losing it one hop up.
Bounded at 512 MB, above which the body streams exactly as before. That ceiling is not a refusal and is
deliberately not a 413: a file Immich ACCEPTS is read to the end and never races, so a 4 GB video is
unaffected. All that is given up above the cap is the error message on a file that was going to be
rejected anyway. A first attempt refused over-cap uploads outright and would have broken the working
4 GB case to improve diagnosis of the doomed one.
`bufferRequestBody` is opt-in and off by default: the vault and wallet proxies must keep streaming so a
passphrase or macaroon never lands in the platform's heap.
Also adds the proxy error logging that made this findable at all — status and two byte counts from
headers, never the bodies. `responseBytes: "unknown"` is what exposed the stripped response.
Verified live at 32/256 MB (buffered) and 640 MB (streamed, passes through).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every live row read 'Starting…' and linked nowhere, because the title lookup searched for a transcript
named after the session KEY. It isn't. The key is officer's handle for a conversation; the transcript is
named after Claude's own session id, and the mapping between them exists only inside the agent
(setClaudeSession/getClaudeSession). I assumed the two were the same and never checked — confirmed wrong
by looking for the ids from the officer log under ~/.claude/projects and finding nothing.
claude:list now reports claudeSessionId beside the key, the route resolves titles by that, and rows link
to it. Null means the first turn has not reported one yet, which is a genuinely unwritten conversation
and stays unlinked.
Needs the agent sidecar restarted to take effect — the new field comes from there.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The titles were looked up client-side against the sessions query already in cache, which only covers the
group being browsed — so anything running in another directory rendered as a truncated uuid, which is
most of them.
Resolved server-side now. The agent reports keys and nothing else, so liveSessionTitle finds the
transcript by scanning the project slugs, takes the cwd off its own first entry, and hands that to
claudeSessionContext — the same path the list uses, so the two agree on naming, /clear chains merged
included, rather than offering a second opinion.
A null title means no transcript has been written yet. That row says 'Starting…' and is deliberately not
a link: pointing at a session you cannot open yet is worse than plainly not being a link.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
resolveOwner already retried forever — but only when the query SUCCEEDED and returned nothing, which is
a fresh install waiting on bootstrap. A query that THREW escaped the function, rejected the top-level
await and exited the process, into exactly the PM2 restart loop its own comment says it exists to avoid.
So any Postgres restart (57P03 'the database system is starting up') or moment of unavailability killed
every live agent session on the machine and spun the sidecar until the database answered.
That is what took a session down on 2026-08-10, and why this process showed 468 restarts against 0 for
every peer that starts without needing the database.
The loop now catches as well as checks. Still retries forever, matching the case beside it: a database
coming back is a matter of time, and an agent that gave up would need a human to notice.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Officer's session records are in memory and die with `pm2 restart officer`, while the agent is a PM2
peer and keeps generating. `adoptOrphanedSession` rebuilds a binding — but only when a browser
reconnects to a session *by id*, which you can only do if you already knew the id. So a session that
survived a restart was invisible, and nothing could answer "what is running right now".
`claude:list` returns each live session with `isGenerating` and `pendingTasks` — the same two fields the
agent's own `armIdle` consults before collecting a session, so a caller can tell "busy" from "merely
open" the way it does. Surfaced as `GET /chat/live`, which sits beside `/chat/sessions`: those are
transcripts on disk, these are the ones with a process behind them.
`getActiveSessionKeys` is replaced rather than joined. It returned bare keys, could not distinguish a
session mid-turn from one merely open, and had never been called by anything.
`listLiveClaudeSessions` fails toward EMPTY, where `isClaudeGenerating` beside it fails toward alive.
The asymmetry is deliberate: not knowing there means leaving a spinner up, and not knowing here would
mean inventing sessions.
Step 2 of docs/chat-session-lifetime.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Officer's hour-long idle timer was doing two unrelated jobs: collecting its own in-memory binding, which
is its business, and terminating the agent, which is the sidecar's. It could not do the first without
the second, because `unsub` was a closure reachable only through `kill`.
So a browser that went away killed a live agent an hour later — including one the sidecar had
deliberately protected. The sidecar already refuses to collect a session that is mid-turn or holding
background tasks: `task:started` disarms its idle GC, and `armIdle` re-checks and re-arms rather than
firing once. Officer had no view of any of that. A laptop running out of battery overnight took a
`run_in_background` job with it for no reason.
`detach` now sits beside `kill` on both streaming handles, and `_sidecarUnsub` — declared and called for
a long time, never once assigned — is populated at all three sites. `releaseSession` unsubscribes and
forgets the record without killing; the idle timer points at it. `deleteSession` is unchanged, so an
explicit disconnect still ends the session.
The third assignment site was not in the plan: `adoptOrphanedSession` sets `_claudeKill` but nothing
else, so an adopted session that later idled out would have dropped its record while the listener stayed
subscribed — a leak of one per adopt-then-leave.
No double subscription: releasing unsubscribes first, so a returning browser either adopts with a fresh
listener or starts a first turn with none behind it.
Step 1 of docs/chat-session-lifetime.md. Step 2 (a list verb, so running sessions can be found after a
restart) is still open.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Platform chat ran at 200K while the same `claude` in a terminal got Opus 5's
full 1M. Nothing to do with the model, the account or compaction tuning — the
CLI gates 1M on `provider === 'firstParty' && Fp()`, and `Fp()` is satisfied
only when ANTHROPIC_BASE_URL is unset or its host is api.anthropic.com. We
point it at 127.0.0.1:5051 so the agent's traffic goes through the OAuth
proxy, which fails that host check and silently drops the window to 200K.
_CLAUDE_CODE_ASSUME_FIRST_PARTY_BASE_URL is the CLI's own escape hatch for a
first-party passthrough, and the assertion holds: the proxy forwards verbatim
to api.anthropic.com and already preserves anthropic-beta.
Read from the CLI binary (2.1.223), not inferred: claude-opus-5 carries
context:{window:1e6, native_1m:true}, so the model half of the gate always
passed. Only the hostname was wrong.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
both forwarders buffered the whole request body into an ArrayBuffer before
re-sending it to Immich. with maxRequestBodySize at 4GB that put a phone's
video upload in the sidecar's heap for a hop that never reads the bytes.
callUpstream now sets duplex: 'half' so a stream is a legal body, matching
what createSidecarProxy already does on the platform side. the four JSON
callers are unaffected.
the platform proxy forwards no content-length, so the body already reached
us chunked; this extends that one hop to Immich. verified against the live
instance (3.1.0): bulk-upload-check round-trips a streamed body and returns
the right verdict.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Locked assets are gated on elevation, and elevation lives on a session — an API key
has no `auth.session`, so no key permission or allow-list entry can reach them. The
gate sits inside the generic owner-access check, so a locked asset's thumbnail and
original are covered too, not just its listings.
So `/_locked` mints a session at unlock, holds it in memory for the elevation window,
and closes it on lock, on idle, or when the active immich account changes. Nothing new
is written to photos_config: the pin and the password are never at rest, and a full
compromise of officer's database still does not open the folder. The cost is that
unlocking asks for the immich password as well as the pin.
`auth/*` stays refused wholesale in routes.ts. The four auth routes this needs are
reached through named endpoints that each do one thing, and the elevated forward
carries three resources rather than the main allow-list.
Not yet exercised at runtime — the sidecar has not run this code.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
there were four doors, not two. the ws upgrade in server.tsx and the
vault notifications socket each verified the jwt themselves, so a key
that worked against /api would have 401'd on cliamp — signed in and can
play audio would have been two different questions for the music app.
both now call resolveAuthToken. verified: owner key upgrades cliamp
(101), bogus key 401, member key 403 on terminal exactly as their jwt
is.
reset-password and verify-token deliberately keep verify() — they read
a purpose-scoped reset token and a key must not be spendable as one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
a user can mint a long-lived key for an app or a device instead of
carrying a 30-day session, so multiple logins on the mobile apps are
per-device revocable rather than one shared token.
identity was being decided independently in userMiddleware and
originScopeMiddleware, each verifying the token itself. teaching only
one of them a new credential format is how those two stop agreeing, so
both now call resolveAuthToken and neither knows what a bearer string
is. verified: a member's key returns the same status as their jwt on
every route tried, 403s included.
a key carries its holder's full authority — not an escalation, it
equals what the password could already do. scoping wants a scopes
column, not a change here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Extracting resolveNotifyUser into its own module immediately caught a hole
in the fix from the previous commit: a header that was present but
unparseable fell through to the body, so a browser could send junk in the
header, name any user in the body and win.
PRESENCE of X-Officer-User is the signal, not its validity — a malformed
header means a proxied request went wrong, and falling through hands the
decision back to the caller we just declined to trust.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A turn could stop producing events and stay `isGenerating` indefinitely. Nothing
covered it: the idle timer answers the opposite question — how long a session with
NO turn in flight may sit before collection — and every client showed a spinner
with no timeout of its own, so a wedged turn presented as a chat that was still
thinking.
On 2026-08-08 one ran for seventeen minutes inside an auto-compaction, reached
over the socket to an iPad, and was indistinguishable there from a dead app. The
compaction is silent by design (the PreCompact hook is the only announcement, and
the code's own comment allows 2.5 minutes), so there was nothing to distinguish it
from.
A stall watchdog now rides every emitted event: any sign of life pushes the
deadline back, and expiry ends the turn the way a real failure would — isGenerating
off, idle re-armed, and an `error` the client can render. The agent process is
deliberately left alive, since it may still be working and the next turn resumes
it; what this guarantees is that the client is TOLD, which is the part that was
missing.
The budgets are generous rather than tight — ten minutes of silence normally,
twenty while compacting, re-armed from the PreCompact hook because that hook fires
as the long silence begins and the deadline the turn is holding was sized for
ordinary work. Killing a turn that was about to succeed is worse than the hang this
prevents.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
On a Mac, Claude Code stores its credentials in the login Keychain and never
writes ~/.claude/.credentials.json — the only file this proxy knew how to read.
The workaround was to copy the Keychain blob into that file by hand, which is a
snapshot: a refresh ROTATES the refresh token and revokes the previous one, so
the two stores were not redundant copies but competitors, and whichever
refreshed second got `401 OAuth access token has been revoked`.
That is not hypothetical. On 2026-08-08 it took out every chat turn from the
iPad for six hours while the terminal CLI beside it worked fine — the harness
spawned, retried for three minutes and wrote the 401 into the transcript, which
from the app looks like an agent that simply never answers.
So on darwin the Keychain is the authority and the file is a mirror, holding the
same token rather than a different rotation of it. Everywhere else — every Linux
server — the file is still the authority and nothing changes. Detection is
process.platform, and a machine with no `security` binary or no such item falls
through to the file rather than failing.
Three recoveries, cheapest first:
- a watchdog checks every 30 minutes and refreshes when under an hour remains.
It checks rather than refreshing on a blind schedule because each refresh
rotates the token, so a needless one is another chance for the stores to
disagree.
- an upstream 401 now RE-READS before refreshing. When a token has genuinely
been revoked the machine usually already holds a good one, because Claude Code
refreshed it into the Keychain minutes ago; spending our own refresh token
there is what caused the divergence in the first place.
- only if nobody else has moved do we refresh ourselves.
The Keychain write goes through argv, which is the only non-interactive form
`security` offers, and matches on the service AND account pair — the account is
read off the existing item rather than assumed, or the update would silently
create a second entry instead of replacing the one Claude Code reads.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
CLAUDE.md asserted "single-user is a hard invariant, not a stage" while
users held six rows and role_capabilities held grants. Every doc that
repeated it is corrected here, in prose and in the code comments that
carried the same claim.
The accurate statement is narrower: one owner who bypasses every check,
other accounts holding only what their role is granted, and a set of
capabilities — terminal, chat, files, tasks, items, desktop, browser — that
are structurally ungrantable because they execute as the owner's OS user.
TODO.md gains a Multi-user section for what the read turned up: no way to
create a second account, dashboards.id colliding across users, authorize.ts
untested, pty/vault/opencode taking no identity, Radicale still owner_only.
claude-sidecar-isolation.md's open question is answered rather than left
open — the per-email spawn model is dead weight, because chat is an
execution capability and no second account can ever reach it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
reset-password accepted any valid signed jwt as a reset token, including a
30-day session token — its sibling verify-token.ts already gated on
purpose === 'reset-password' and this handler did not. forgot-password mints
that claim, so the gate costs the legitimate flow nothing.
notify's DELETE /_officer/devices/:token deleted by token with no user
predicate: a token is the address of a device, not a secret, so any account
holding the notify capability could deregister another's device.
deletePushDevice now takes an optional userId — the route passes it, the
APNs/FCM dead-token paths deliberately do not.
POST /_officer/notify let a request body's userId override the
proxy-injected X-Officer-User. The header now wins where present, which is
what separates a signed-in browser from a loopback producer that has no
session to speak from.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Committing work that was left uncommitted in the shared tree. I did not write it;
I reviewed it in full, verified it against the running system, and am landing it at
the owner's explicit request because no one currently owns it.
This REPAIRS master. `useAgentPanel.ts` shipped in dbe585f and calls
`/chat/agent-panels`, but `registerAgentPanelRoutes` existed only in the working
tree — so on master as pushed, every one of those calls 404s. The feature has been
half-landed since that commit.
What it is. A panel on a dashboard can be given a name ("frontend", "code-reviewer").
Naming it mints two things: a `sessionKey`, which is the panel's permanent continuity
(it keys the sidecar's on-disk resume map and `chat_session_events`, so the same panel
reopens the same Claude session), and a `handoffToken`, a bearer credential scoped to
exactly one verb. The agent in that panel is then addressable by name, and can pass
work to a peer on the same dashboard over `/api/agent-handoff`.
Three doors, deliberately separate:
- `/chat/agent-panels` (browser, session-authed) — name / list / rename / forget.
Mounted on the chat router rather than given its own prefix: these routes create
and name Claude sessions, which is authority `chat` already grants. A second
top-level mount would have meant a second capability entry claiming the same
thing under a different name.
- `/api/agent-handoff` (agent, token-authed) — peers and send. Unprotected by the
session middleware and exempted in `capabilities/totality.ts` with its reasoning
written down, because the caller is a subprocess with a token, not a browser with
a cookie.
- The transcript stays where transcripts live. DELETE forgets the address and the
panel's claim on the session; it does not touch ~/.claude/projects.
Security, as verified rather than assumed:
- The sender is derived from the token, never from the request body — there is no
`from` field on the wire, so it cannot be forged.
- Every lookup is scoped to the token's `userId` AND `dashboardId`, so an agent can
only see and reach peers on its own dashboard.
- `toAgentPanelView` strips `handoffToken` and `userId`, and it is the only shape
the browser routes return. Confirmed by reading every return path.
- Live-tested: a real token on `GET /api/agent-handoff/peers` returns 200 with
correctly scoped peers; a bogus one returns 401.
Two judgement calls in the code worth knowing about, both already commented at their
site: the introduction turn inlines the handoff token into a runnable curl (a
single-owner MVP trade), and `agent_panels` carries no FK to `dashboards.id` because
that primary key is mid-rework to a composite.
Schema uses `uniqueIndex` throughout, never `unique().on(...)` — the rule that exists
because drizzle-kit mis-diffs named composite unique constraints and re-creates them,
which is what wiped seven tables on 2026-08-03.
NO `bun db:push` IS NEEDED. `agent_panels` is already live in Postgres with 6 rows;
the schema file is catching up to a database that already has it.
Verified: `bunx tsgo` clean, `bun test` 538 pass / 0 fail across 35 files.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The harness delivers a finished background task to the agent by writing it as the user's next
message, so Claude's file holds a raw <task-notification> envelope as a user turn. Live it never
shows, because the same event travels separately as task:notification — it appeared only when a
refresh rebuilt the conversation from the file, as a bubble on the owner's side he never typed.
Same defect as INTERRUPTION_MARKERS and the same fix. Anchored to the start of the message so
quoting one inside a real message stays yours. Also skipped when picking a session's title, where
it is no more a title than a slash command is.
Verified against a live transcript: 38 user bubbles before, 32 after, the 6 removed being exactly
the notifications.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
pastilhasandClaude Opus 5