Seven items the earlier repair waves deliberately left alone, so that the
blocker fixes could land with a working safety net first.
CERTIFICATE LIFECYCLE (the audit's only "high"). iOS ships a
CertificateRotationScheduler; Android had no counterpart, so a device
certificate simply expired and needed a manual re-enroll. The decision is a
pure total function over five answers, with RE_ENROLL_REQUIRED checked first so
it outranks any backoff. Renewal is single-flight — a second trigger coalesces,
and a cancelled leader releases the slot so cancellation cannot wedge rotation
for the process lifetime. Failure leaves the prior identity fully live (the
existing atomic ping-pong flip already guaranteed that) and is published rather
than swallowed.
Two real defects surfaced while building it:
renew() decoded its 201 with EnrollResponseDto, whose deviceId is required —
but /device/:id/renew returns only {cert, caChain, notAfter}; only
/device/enroll echoes deviceId. Every silent renewal would have thrown
MalformedResponse. Fixed with a RenewResponseDto plus the deviceId from the
path segment we addressed.
And because no renew 201 carries renewAfter, a client that persisted it would
rotate exactly once and then report not-due until the cert died. That is why
renewAfter is re-derived from the live leaf as notBefore + 2/3·lifetime — the
control plane's own formula. This is a latent iOS defect that Android now
avoids rather than inherits.
/device/:id/recover was confirmed real (control-plane/src/api/renew.ts:443,
contract pinned by that suite's CP6e) and is now wired. It goes over a separate
NON-mTLS client that throws rather than falling back, because an expired client
certificate cannot authenticate its own recovery — the deadlock a previous
session already hit and recorded.
Asked explicitly about step-up: rotation is unaffected. stepUp appears nowhere
in control-plane/src; it lives only on the relay's WebSocket upgrade. On a
step-up host the SESSION is denied, not the rotation, so the certificate still
stays alive and that principal gap is orthogonal.
FOLLOW-UP QUEUE (W2). The queue frame decoded but the three routes managing it
were missing, so Android could see "N queued" and not queue anything. One
outcome union covers both guarded writes and distinguishes full / too-large /
disabled / session-gone / rate-limited, because they need different copy. 429 is
surfaced and never auto-retried — POST and DELETE share one 20/min bucket.
Queued text is raw shell input and travels verbatim, proven with control
characters and CJK.
UNREAD WATERMARK now persists, so the unread state survives process death —
which is exactly when it matters, since the product premise is walking away and
coming back. It stores a plain map rather than an UnreadLedger: the reducer
lives in :session-core, which :host-registry may not depend on, so the logic is
not restated anywhere.
COPY-OUT. Text selection was a recorded feature gap: the stock ActionMode's own
Copy handler dereferences the absent TerminalSession, so making it reachable
puts an NPE on the button — worse than no selection. The refusal was
re-verified from the bytecode, and the same disassembly showed the way out:
TerminalBuffer.getSelectedText touches only mLines/TerminalRow, no session and
no view. A first-party selection layer now builds on that, with a pure
row-major model so a backwards drag normalises correctly. Terminal content
reaches the clipboard only on an explicit user copy; OSC 52 stays declined.
HOST REMOVAL. PushRegistrar.unregisterHost had zero call sites, so a removed
host kept receiving pushes. Removal is now confirm-gated and cleans everything
keyed by host id — a half-removed host is worse than none.
/config/ui is finally honoured. It was implemented and never called, so
allowAutoMode was ignored and the plan gate offered "approve + auto" even where
the operator had disabled it. It fails CLOSED: an unfetchable config treats auto
as disabled, because the permissive default is the dangerous one.
Verified: ./gradlew test :app:assembleDebug :app:assembleDebugAndroidTest
:macrobenchmark:assembleBenchmark koverVerify -> BUILD SUCCESSFUL, 1150 JVM
tests, 0 failures (903 -> 1150). Instrumented suite re-run on emulator-5554
against live servers: 67 tests, 0 failures, 0 skipped — no device regression.
The plan marked A34 (instrumented E2E against a real host) and A35
(macrobenchmark) DONE. Neither had a single line of code — :app had no
androidTest source set at all, and `testInstrumentationRunner` named a
HiltTestRunner class that did not exist. They had been converted into checklist
bullets. This adds 67 instrumented tests and 3 macrobenchmarks, and every one
of them has actually been executed on an emulator against this repo's own
server — not merely compiled.
Why this layer had to exist: four crash-on-first-interaction defects shipped
behind a green 900-test suite, for one structural reason — no JVM test
instantiates an android.view.View, creates an InputConnection, dispatches a
MotionEvent or runs a layout pass. So the rig builds the REAL RemoteTerminalView
in a REAL Activity and drives it through the REAL framework entry points,
faking only the wire so assertions can be about BYTES rather than "did not
crash". That distinction matters: a no-crash-only test would have passed against
the silent-black-hole variant of the IME bug, which was arguably the worse half
of it.
The four crash regressions (28 tests):
keys — Enter is \r not \n, Ctrl+C is 0x03, Backspace 0x7f, and arrows in BOTH
DECCKM modes driven off the live emulator bit, so ESC[A vs ESC OA is
asserted rather than assumed; BACK falls through and emits nothing; soft
keyboard commitText/sendKeyEvent/deleteSurroundingText and a CJK
composition pass through verbatim.
alternate-screen scroll — every test feeds ESC[?1049h FIRST and asserts the
buffer is active before touching anything, because that is the branch that
used to crash. Real dispatchTouchEvent drags both directions, the DECCKM
form, the mouse-tracking branch emitting a wheel report and NO arrows
(branch order), the wheel's 3-rows-per-notch, and the fling proven not to
start by 800 ms of silence after ACTION_UP.
autofill — the framework's own autofill-structure gate is closed on both the
frame and the stock child, before and after focus.
resize — including the draw-race counterpart of the AIOOBE fixed in 538c8eb.
A34 E2E against a live server (15 tests): attach → attached → output round
trip, ring-buffer replay on reconnect, spawn failure via a second server
deliberately configured with a bad SHELL_PATH, a held gate resolved through
POST /hook/decision, and CswshDefenceE2eTest — the F9 bad-Origin rejection,
which the checklist calls the one non-skippable defence and which had never
been executed anywhere.
A35 (3 macrobenchmarks): startup, and pair → attach → type → approve.
Reaching a host from the emulator needs ALLOWED_ORIGINS=http://10.0.2.2:<port>,
because the server derives allowed origins from host NIC IPs and 10.0.2.2 is
the emulator's alias for the host, not an interface address. Tests take the
host as an instrumentation argument and SKIP with a clear message when it is
absent, so "passed" can never be confused with "never ran".
Verified independently of the authoring agent, on emulator-5554 (webterm AVD,
android-35, arm64, software GPU), with both servers live:
Starting 67 tests on webterm(AVD) - 15
Finished 67 tests on webterm(AVD) - 15
tests="67" failures="0" skipped="0"
JVM gate unaffected: ./gradlew test :app:assembleDebug
:app:assembleDebugAndroidTest :macrobenchmark:assembleBenchmark koverVerify ->
BUILD SUCCESSFUL, 903 tests, 0 failures.
The instrumented suite was run on a device for the first time and the app died
on the first test:
ArrayIndexOutOfBoundsException: length=31; index=31
at com.termux.terminal.TerminalRow.getStyle
at com.termux.view.TerminalRenderer.render
at com.termux.view.TerminalView.onDraw
PROGRESS_ANDROID.md had classified exactly this as "Accepted (not a defect):
steady-state append runs concurrently with the UI-thread onDraw ... torn read
self-corrects next frame". Both halves of that were wrong.
It is not cosmetic. TerminalRenderer.render caches its column bound once from
TerminalEmulator.mColumns and then indexes TerminalRow.getStyle(column) into a
raw long[] sized when that row was allocated. TerminalEmulator.resize publishes
the new mColumns BEFORE resizeScreen() reallocates the rows, so a draw landing
in that window reads the new bound against old-width rows. The first
out-of-range column is exactly oldColumns, which is why the exception was
always length == index — the consistency was a deterministic signature of a
grow, not a rare interleaving. A shrink is harmless, which is probably why
"self-corrects" looked plausible. TerminalBuffer.resize is non-atomic in
several more ways, and TerminalRow.setChar swapping mText for a larger array is
the same hazard on the append path.
Nor does it match upstream. Upstream's background reader only fills a
ByteQueue; TerminalSession$MainThreadHandler is what calls append, and
TerminalView.updateSize resizes — both on the main thread. Upstream has no
concurrent reader at all. There is also no lock to lean on: monitorenter
appears nowhere in TerminalEmulator, TerminalBuffer, TerminalRow or
TerminalRenderer.
The race was dormant until this session. updateSize had no call sites, so
mColumns never changed after bind and there was never a width mismatch to tear
on. Making resize work is what made this reachable — the same lesson as the
rest of this pass: the green suite could not see any of it because no JVM test
instantiates a View.
Fix: every emulator mutation now runs on the render thread, from inside the
same single confined consumer. §6.2 is intact — still one writer draining one
ordered channel, suspending across the main hop, so a resize and an append
still cannot interleave and submission order still holds. They are now also
serialised against onDraw, because one Handler work item cannot run inside
another. Resize still reaches the wire and the §6.4 forced re-assert still
bypasses the dedup.
Chunking becomes load-bearing rather than incidental: each 4 KiB slice is its
own main-thread work item, so a multi-MB ring replay interleaves with frames
and input instead of blocking behind one long call — which is what §6.2's
no-ANR requirement actually needs, rather than "runs off the main thread".
Gating the child's dispatchDraw was considered and rejected: it would have to
cover appends too, leaving a choice between blanking the terminal and blocking
the UI thread on a full reflow. A read/write lock was rejected because append
calls back out through TerminalOutput (title, bell, DA/DSR reply), so holding a
write lock across it would bury a deadlock invariant in app-level callbacks.
The KDoc that made the false claims now states the mechanism with offsets, and
RemoteTerminalHostView records why dispatchDraw is deliberately not gated.
Verified: the device crash is reproduced as a JVM test first (red), so this
regression is now guarded without a device. ./gradlew test :app:assembleDebug
koverVerify -> 903 tests, 0 failures. On the emulator the previously-crashing
test now passes and 5 of 6 alternate-screen scroll tests pass.
Four gaps that made "ship it" and "test it on hardware" impossible.
RELEASE SIGNING. assembleRelease was silently producing
app-release-unsigned.apk because there was no signingConfigs block at all —
the failure mode where you only notice at install time. Credentials now
resolve from keystore.properties, then local.properties, then
WEBTERM_RELEASE_* env; nothing is created or committed, and
keystore.properties.example is the template.
The degradation is ENFORCED, not documented. The signingConfig is only
created when credentials actually resolve — an empty config is precisely what
makes AGP emit an unsigned artifact quietly — and a guard on packageRelease
throws with an actionable message, distinguishing "not configured" from
"credentials found but the keystore file is missing: <path>". It hooks the
PACKAGE task on purpose: R8 has already run and reported by then, so a
developer without a keystore still gets the full shrink result rather than an
early abort. Verified: minifyReleaseWithR8 completes, packageRelease fails
loudly, and no unsigned APK is written.
VERSIONING. Was still the scaffold 1 / "0.1.0". Now 1 / "0.1.0-alpha01",
where the suffix is a factual claim about device verification rather than
decoration: alpha means device QA is essentially unstarted, beta means the
A34/A35 checklist blocks pass on hardware, and plain 0.1.0 means the whole
checklist is ticked. versionCode stays 1 because no artifact has ever left
the build machine.
MINIFICATION. release had isMinifyEnabled = false and an empty rules file —
so nothing had ever exercised R8 on this app, and the first release build
would have been the experiment. Now minify + resource shrinking, with a
deliberately short rules file: the actual artifacts were checked, and
kotlinx.serialization, Tink, Hilt, FCM, OkHttp, Compose and CameraX all ship
their own consumer rules, so none are re-declared. The termux packages are
kept whole because the JitPack artifacts ship no rules and :terminal-view
binds to non-API internals — it writes the public TerminalView.mEmulator
field and calls TerminalRenderer.getFontWidth()/getFontLineSpacing(). Getting
that wrong yields a build that crashes only in production.
lintVitalRelease also passes now, the two pre-existing lint ERRORS having
been fixed earlier this session.
ANDROID TEST ENABLEMENT. :app had no androidTest source set and no
instrumentation runner, which is the mechanical reason A34/A35 have no code
at all. The runner (Hilt-aware) and dependencies are wired, and a
:macrobenchmark module exists. The tests themselves are a separate change.
Verified: ./gradlew test :app:assembleDebug :app:assembleDebugAndroidTest
:app:assembleBenchmark :macrobenchmark:assembleBenchmark
:app:lintVitalRelease koverVerify -> BUILD SUCCESSFUL; 901 JVM tests, 0
failures on a forced re-execution with test-results deleted and
--no-build-cache.
This file's cross-review was explicitly skipped when it was written —
PROGRESS_ANDROID.md says so — and it had no production call sites until this
session, so nothing had ever reviewed OR executed it. Wiring it to the home
screen finally made an adversarial concurrency review worthwhile. It confirmed
the hard parts are right (the check-then-insert dedup critical section,
withPermit's exactly-once release with the fallback outside it, the
NonCancellable finally, and rethrowing CancellationException) and then found
five real defects. Every premise was re-verified from the termux v0.118.0
bytecode before fixing.
1. Thumbnails drew DUPLICATED ROWS. TerminalBuffer.externalToInternalRow is
`(mScreenFirstRow + row) % mTotalRows` and mTotalRows is the constructor's
transcriptRows, so passing TERMINAL_TRANSCRIPT_ROWS_MIN (100) while rows
went up to 200 aliased external rows 100..199 onto 0..99 — and left the
emulator's scroll state invalid during append(). A pure ThumbnailGrid now
derives transcriptRows = max(rows, MIN).
2. A legitimately-sized session was LMK bait. cols/rows are server-supplied
and the protocol validates resize only to [1,1000], so a 1000x1000 session
forced 2800x2800 ARGB_8888 = 31.4 MB, twice over for two permits, every
poll tick. Rather than clamp the grid — which would CROP a real screen —
the cell size is now derived from the target, shrinking the font's own cell
by an exact integer fraction with the aspect preserved. Any geometry in
1..1000 now yields at most 480x480 = 900 KiB; a real 161x50 goes from
3.1 MB to 322 KiB.
3. Two stalled fetches wedged the feature PERMANENTLY. There is no OkHttp
callTimeout (the per-socket read timeout is reset by every trickled byte)
and the render is deliberately uncancellable by any caller, so two
slow-trickle previews held both permits until close() and no thumbnail
rendered again. A withTimeout now sits inside withPermit, budgeting the
work rather than the permit wait. The subtlety that makes this correct:
TimeoutCancellationException IS a CancellationException, so it is caught
BEFORE that branch and routed to the placeholder — otherwise a slow host
would still kill the pipeline instead of degrading.
4. The in-flight entry was leakable, contradicting a comment that claimed it
never was. `scope.async` on an already-cancelled scope never runs its body,
so the finally never ran: a post-close() request left a dead Deferred that
every later request awaited forever, and inFlight grew unbounded. Now
guarded inside the mutex, and the comment names the one path that genuinely
cannot release (where the map dies with the pipeline anyway).
5. Obsolete renders did guaranteed wasted work. lastOutputAt advances on every
PTY byte and the list polls every 5s, so an active session mints a new key
each tick; the row cancelled the old caller but the old render ran a full
GET plus a full raster whose bitmap was cached under a key nothing would
request again — occupying both permits ahead of the key actually on screen,
and growing without bound once render latency exceeded arrival rate. A
newer key now cancels the older in-flight render for that session.
Also made two false statements true rather than leaving them: retainOnly now
prunes under the same monitor publish() takes, so a render landing after the
prune cannot resurrect a killed session's bitmap; and PreviewCap's KDoc no
longer claims to prevent an oversized-body OOM (the body is already buffered
and parsed by then) nor to "mirror the server's cap" (the server defaults to
24 KiB, not 256 KiB, and an operator can raise it unbounded).
RED was observed for all four behavioural fixes before implementing.
Verified: ./gradlew test :app:testDebugUnitTest :app:assembleDebug koverVerify
-> BUILD SUCCESSFUL, 901 JVM tests, 0 failures (893 -> 901). The agent
distrusted a FROM-CACHE green, forced a real execution, and confirmed the code
reached the packaged APK by scanning its dex.
Four agents' worth of :app wiring. Everything here existed as tested code with
zero production call sites, or as a screen rendering a shape the server stopped
sending months ago.
Dead code that shipped as features. All three were fully implemented and
unit-tested while being reachable from nothing:
- Session-list preview thumbnails. SessionsHome called SessionListScreen
without the `thumbnails` argument, so the `= null` default won and every
row rendered a placeholder forever — on the home session chooser, whose
entire purpose is seeing what each session is doing.
- Quick-reply chips and their palette editor.
- PointerContextMenu (large-screen secondary click).
ThumbnailPipeline's formal cross-review had been explicitly skipped when it was
written — PROGRESS_ANDROID.md said so in as many words, and it is concurrency
code, so nothing had ever reviewed OR executed it. It has now been reviewed and
kept: the LruCache key, the fair Semaphore(2), the Mutex-guarded in-flight
dedup and the NonCancellable in-flight release are all sound, and a retained
map entry would have poisoned a key forever.
The v0.6 project-detail git panel. Android rendered the pre-w6 shape: one bare
dirty dot, a bare hash+subject commit row, and a worktree heading that renamed
itself to "Branch" at n=1 — the exact behaviour G5 removed. Now the sync band,
unpushed marking with a single upstream boundary, and "Worktrees (n)". The rule
that drives it is enforced, not documented: exactly one state may render green,
and no-upstream is not that state.
Remote approvals are no longer blind. status.preview reaches GateBanner and
PlanGateSheet, so the user can see the command or diff they are approving
instead of a tool name. Absent preview stays an ordinary gate, not an error.
The WEBTERM_TOKEN gate is no longer inert. NetworkModule was calling
OkHttpClientFactory.create(identityProvider) with no cookieJar, so the shared
client ran on CookieJar.NO_COOKIES and none of the cookie plumbing did
anything. It now installs the real AuthCookieJar — one client, so the cookie
rides the WS upgrade too — backed by the Tink-sealed store, and binds
InMemoryAuthCookieStore if a cipher cannot be constructed rather than ever
persisting a shell credential in the clear. The persister runs on an IO scope
because OkHttp calls a CookieJar synchronously and must never block on
DataStore or Tink.
And the user-facing half that made the gate unusable: a token-gated host
answers the pairing probe's unauthenticated GET with 401, which the frozen
taxonomy could only report as "this port is running something else" — so such
a host was impossible to pair AND the copy was misleading. Pairing now routes
to a token prompt and resumes through the same cert-gate choke point. The
token is a parameter, never a field and never part of any UI state.
Docs corrected rather than left aspirational: the plan asserted a cleartext
allowlist "permitted only for private LAN/Tailscale CIDRs", which the platform
cannot express at all — the network-security-config format has no netmask or
prefix syntax. PROGRESS_ANDROID.md now opens with what "COMPLETE" actually
meant, and PROGRESS_LOG.md records the whole pass including what is still
undone: A34/A35 have no code, and device QA remains ~0%.
Verified: ./gradlew test :app:assembleDebug :app:testDebugUnitTest koverVerify
-> BUILD SUCCESSFUL, 893 JVM tests, 0 failures (baseline was 617). The :app
run was re-measured with outputs deleted and --no-build-cache, and the APK
re-packaged, so Hilt graph validity is an observation rather than an
up-to-date claim.
The server shipped the project-detail git panel in 8fe1f52/f711db9/042c4cd,
all after the last Android commit, so the client silently ignored every new
field (they are additive and optional, so nothing crashed — it just rendered
the pre-w6 shape: one bare dirty dot and no sync information at all).
Added: SyncState + ProjectDetail.sync/dirtyCount, ProjectInfo.dirtyCount,
ProjectSessionRef.cwd, GitLogResult.upstream, CommitLogEntry.unpushed, plus
GET /projects/worktree/state (read) and POST /projects/git/fetch (write).
Two rules from the server's own commit message are encoded as behaviour
rather than left to each call site to remember:
`SyncState.isFullySynced(now)` is the only sanctioned way to ask whether
the green all-clear may be rendered, and it answers false for anything
unknown. ahead/behind compare against @{u}, a locally cached ref that only
a fetch moves, so `behind = 0` from a stale FETCH_HEAD proves nothing. Most
importantly NO UPSTREAM leaves both counts undefined — which is not zero —
and that is the normal state of a fresh worktree branch. The server calls
that fall-through the easiest bug to ship here; there is a test per way it
could regress, including a fetch timestamp in the future, which is clock
skew rather than evidence that anything was checked.
`GitLogResult.upstreamBoundaryIndex` finds the LAST unpushed commit rather
than assuming the unpushed ones come first. git log is date-ordered, so
merging an older branch interleaves unpushed commits below pushed ones, and
the row-count shortcut fails in the dangerous direction — it would label an
unpushed commit as pushed. It returns null, not -1, when no boundary may be
drawn, so "no upstream" cannot be mistaken for "boundary before everything".
Origin policy is asserted in both directions: worktree/state carries no
Origin, git/fetch carries a byte-equal one, and a test also asserts the fetch
body contains no remote or refspec — the remote is derived server-side
precisely so a client cannot aim a fetch at an arbitrary URL. Fetch has its
own rate-limit bucket, so a 429 is surfaced rather than retried, and a failed
fetch leaves lastFetchMs unknown so the UI keeps saying "stale" instead of
pretending it refreshed.
Verified: :api-client:test + koverVerify green; 22 new tests.
Both had shipped on the server for a while and Android silently discarded
them, which is worse than not supporting them — the UI looked correct while
missing information the user needed.
W2 `queue`: `ServerMessageType` had no member for it, so `decodeServer`
returned null and `SessionEngine.onFrame` counted it as undecodable and
dropped it. The "N queued" badge could therefore never appear on Android,
on any device, ever. Adding the branch made the compiler flag the
non-exhaustive `when` in the engine, which is the same gap seen from the
other side. A negative depth drops the frame rather than feeding nonsense
to a badge; the next broadcast carries the real value.
W1 `status.preview`: not modelled at all, so a remote approval was BLIND —
the gate could show that `Bash` wanted to run, but not what it wanted to
run, and the same for an Edit's diff. Approving blind defeats the point of
the remote gate, which exists so whoever is holding the phone can make an
informed call. `ApprovalPreview` now models both variants and
`GateState.preview` carries it to the UI.
Two decode rules are load-bearing and pinned by tests:
- A broken preview must never cost the FRAME. `pending` is what makes the
gate appear at all, so trading a missing preview for a missing gate
would be a bad bargain. Every malformed shape — null, non-object, no
kind, unknown kind, command without text, diff without a file — degrades
the preview to null and still yields the Status.
- Unknown values are kept, not coerced. A diff line kind or file status
this client has never seen renders as itself; dropping a line we cannot
classify would hide part of what is being approved. Only a diff file
missing its paths is dropped, because then the preview cannot be trusted
to describe the right file.
A sustained pending refreshes the preview without minting a new epoch: the
server's bounded render can arrive a frame after the gate rises, and a
decision tapped in between has to stay valid.
The preview's diff subset lives in :wire-protocol rather than reusing the
:app diff model — :session-core may not depend on :app, and new wire types
belong in the frozen contract module (CLAUDE.md).
Verified: :wire-protocol:test :session-core:test + both koverVerify green;
16 new tests (10 codec, 6 engine-level).
The client built to an APK and 617 JVM tests passed, but nothing had ever
run on hardware, and three defects made it unusable the moment it did. All
three were invisible to the JVM suite because none of those tests
instantiate an Android View.
1. Every key press crashed. `RemoteTerminalView` bound the forked emulator
but never installed a `TerminalViewClient`, and the stock Termux view
dereferences `mClient` with no null guard on the first line of the paths
a user actually hits (`onKeyDown` offset 82, `onCreateInputConnection`
offset 0, `onKeyUp`, `onKeyPreIme`). Installing any client is not enough:
returning false continues into `mTermSession.write()` at offset 150, and
`mTermSession` is null forever because `TerminalSession` is final and
forking a process is what this app must not do. So the client consumes
the key itself and `KeyRouting` has no defer-to-stock branch at all —
the crash is unrepresentable, not merely avoided. Termux's `KeyHandler`
stays the authority for the key tables, so DECCKM still emits `ESC O A`.
2. The PTY never learned the real size. Nothing called
`RemoteTerminalSession.updateSize`, and the stock fallback is dead
(`TerminalView.updateSize` early-returns without a session), so every
full-screen TUI rendered into 80x24. `TerminalResizeDriver` now drives
it from real cell metrics — read through the public `getFontWidth()` /
`getFontLineSpacing()` accessors rather than reflection, which would
break silently under R8, or a re-measured Paint, which would drift from
what the renderer actually draws (plan risk R5).
3. Bare-LAN connections were impossible. `network_security_config.xml` was
a self-labelled STUB denying all cleartext while `HostEndpoint` accepts
http and derives ws://. The format has no CIDR syntax, so the allowlist
its own comment promised cannot be written; cleartext is permitted in
base-config, the tunnel domain keeps a TLS block, trust anchors are
pinned to system-only, and the guard moves to the §5.4 pairing tiers.
Adversarial review then found three more, and one it got wrong:
- Swipe-scrolling inside an alternate-screen TUI still crashed. The touch
path bypasses `TerminalViewClient` entirely: `doScroll` turns a scroll
into `handleKeyCode` whenever the alternate buffer is active, and that
dereferences `mTermSession`. Claude Code is an alternate-screen TUI, so
this was a crash during normal use on the app's main screen.
`TerminalScrollGesture` claims the vertical drag first and reproduces
all three `doScroll` branches, closing the fling runnable and
`onGenericMotionEvent` with it.
- `autofill()` dereferences `mTermSession` unguarded while the view
advertises itself autofillable, so any password manager crashed it.
The subtree is excluded from the autofill structure.
- The review asked for stock text selection to be restored. It cannot be:
the ActionMode's own Copy and Paste handlers dereference the absent
session, so a reachable selection UI has an NPE on its Copy button.
Long press stays consumed and copy-out is recorded as a feature gap.
Also: the WEBTERM_TOKEN cookie is carried on REST and the WS upgrade and
sealed with Tink AEAD under an AndroidKeyStore key (fail-closed — a seal
failure persists nothing rather than falling back to plaintext), and
`allowBackup=false` keeps a 30-day shell credential off Google Drive.
Verified: ./gradlew test :app:assembleDebug -> 757 JVM tests, 0 failures.
Nothing here is device-verified yet; that is the next step.
margin-left:auto was correct for the old single-line row, where it pushed the
path to the far right of a flex row. The two-line card stacks branch over path
in a column, so the same declaration shoved the path to the opposite edge from
the branch it labels. Removed, with a note saying why it must not come back.
Same shape as the .proj-wt-row duplicate earlier in this series: an element
reused in a new layout kept a positioning declaration that only made sense in
the old one. Caught by screenshot; the DOM assertions could not see it.
Two things, both invisible in code review and obvious on screen.
1. Eight controls in the project detail had class names but no CSS rule at
all, so they rendered as raw user-agent widgets — white input boxes and grey
buttons sitting inside a dark panel: View Diff, the new-branch input, Create
Worktree, and the five fan-out fields. Gave them the skin the panel already
established, with two button roles: filled accent for the control that
commits a form (matching Fetch), outline for one that only reveals or
toggles (matching Open). Layout stays in each element's existing rule; this
block sets only the shared skin, so no property is declared twice.
2. The w6 CSS referenced --fg, --fg-dim, --fg-faint, --bg-sunk and --mono,
none of which exist in :root. Every one of those 19 references was silently
falling through to its hardcoded fallback, which happened to match the
default amber theme — so it looked correct while being wired to nothing.
Under any other theme the sync band would have kept the default colours
while the rest of the app changed. Rewired to the real tokens: --text,
--text-dim, --text-faint, --surface-1, --on-accent, and the mono stack the
file already uses.
The fallback values are what hid this: a var() fallback turns a typo into a
silent success. Worth remembering when reaching for one.
Verified: tsc clean, npm test green (78 files / 2161, e2e 27), bundle rebuilt.
Two layout defects the screenshots caught that the DOM tests could not.
1. `.proj-wt-row` was declared twice — my two-line-card rule earlier in the
file, the original flex rule later. Same specificity, so the later copy won
and the Open button sat against the text instead of at the card's right
edge. Folded the grid into the ORIGINAL rule and left a note where the
duplicate was. One rule per selector; a second declaration of a layout
property is a silent override waiting to happen.
2. The unpushed count rendered as its own line under "Recent commits". The
design has it ON the heading, because it qualifies that heading rather than
making a separate statement. renderGitLog now takes the heading element and
writes the badge there, clearing any previous badge first so a re-render
cannot stack duplicates.
Both were invisible to the existing assertions: they checked that the elements
exist and carry the right text, which was true in each case. Added a test for
the no-duplicate-badge path; the right-edge alignment is a pure CSS outcome and
is verified by screenshot, not by the suite.
Closes the three deviations between the shipped git panel and
docs/mockups/project-detail-git.html: the four-cell sync band with its
explanatory footnotes, the dirty count on the title line, and worktree rows as
two-line cards with a working Open button. Adds structural assertions so
layout drift is caught by the suite, not by eye.
The shipped panel was functionally right — every number matched git — but it
was not the approved design. I implemented from the plan's semantics (which
state is green, when to flag stale) and let the mock's layout drift, and
nothing caught it: the tests asserted behaviour, and no assertion covered
structure. Three deviations, now closed against docs/mockups/project-detail-git.html.
1. The sync band was compressed into one row of chips. Restored to the four
labelled cells: Upstream / To push / To pull / Fetch, with display-size
numerals and a footnote under each count. The footnotes are load-bearing,
not decoration — "Last fetched 19d ago — this number cannot be trusted" is
what tells a reader WHY the zero is suspect. Compressed to
"fetched 19d ago", the causal link was left for the reader to infer, and
not having to infer it is the entire point of the panel. The unverified
state is a chip on the count again, so the doubt attaches to the digit.
2. The working-tree count sat inside the band. Moved to the title line beside
the branch, where the design puts it: the band is about the remote, and the
dirty count is not. Reads "● 3 uncommitted" instead of a bare dot; the dot
remains as the fallback when a server sends no count.
3. Worktree rows were the old single line with the path right-aligned against
the chips. Rebuilt as the two-line card the design calls for — identity and
state on top, path underneath — plus the Open button, which was specified
and simply missing, so a worktree row could be read but not entered. Open
reuses the detail view pointed at the worktree path: a linked worktree is a
project directory, so this needs no new concept and no new route.
Also added the unpushed count beside the commit-list heading, so the number is
readable before the eye walks the rows hunting for rails.
Tests now assert structure, not just behaviour: cell captions exist, the stale
footnote says why, counts render at display size, the path is outside the chip
row, and Open fires with its worktree. That is the gap that let the drift
through, so it is the gap that is now covered.
Verified: tsc and build clean; npm test green (unit 78 files / 2159, e2e 27).
Adds the sync band (upstream, ↑ to push, ↓ to pull, staleness), the unpushed
boundary in the commit list, a reworked worktree panel with per-row state and
session counts, and a read-only POST /projects/git/fetch so the numbers can be
refreshed without a pull.
The governing rule throughout: only one state may look reassuring — even with
upstream and a fetch inside the hour. behind is derived from a locally cached
ref, so a stale zero is a guess and is marked as one; no upstream renders as
'no upstream', never as synced.
Also fixes the test suite's long-standing flakiness (5-11 shifting failures per
run) so the feature could actually be gated on it.
`npm test` was failing 5-11 tests per run with the set shifting between runs,
which made it useless as a gate. Two independent causes, neither of them a
product defect.
1. Fixtures had outgrown vitest's 5 s default. The git ones spawn 6-12
sequential `git` processes (init/config/commit/clone/push); the real-server
ones boot a server and a shell. Alone they finish easily; under a full
parallel run they do not, and the failure reads `Test timed out in 5000ms`.
Gave those describes an explicit 30 s ceiling, and the real-PTY waits a
named PTY_WAIT_MS. The deliberate short bounds in the "must be rejected"
handshake tests are left alone — there a timeout IS the assertion.
(The same rebudgeting also touched the test files in the preceding commit.)
2. test/integration/server.test.ts cannot share the machine with the rest of
the suite. It asserts on real prompt output within seconds, which is not
achievable while ~8 workers saturate the box: it passed alone (27/27,
repeatedly) and flaked in the full run. Raising the numbers further only
moved the flake around, so this is fixed as a scheduling problem — `npm
test` now runs two passes, `test:unit` (everything else, parallel) then
`test:e2e` (that file on its own). `vitest run` still runs everything at
once for anyone who wants that.
Also bounded the srv.close() in H1's finally. Unbounded, it swallowed whatever
really went wrong in the body — the inner waits threw, control jumped to the
finally, close() blocked, and the case reported a bare timeout instead of its
own diagnosis. The root-causing above only became possible after making that
failure legible.
Note on the `[needs real PTY (sandbox-off)]` labels: those cases were NOT
skipping here. PTY_AVAILABLE is true on this machine, so they were running and
failing on timing, not being gated out by a sandbox.
Verified: npm test green end to end across repeated runs (unit 78 files /
2147 tests, e2e 27).
The page carried one bare `●` for "dirty" and nothing else, so "do I have
commits I haven't pushed" and "which worktree am I in" still meant dropping
into a terminal. Design mock: docs/mockups/project-detail-git.html; plan and
task breakdown (G1-G7): docs/plans/w6-project-git-panel.md.
One rule drives the whole feature: ahead/behind compare against `@{u}`, a
LOCALLY CACHED remote ref that only a fetch moves. This repo was the live
example while building — `↑9` true, `↓0` false, because FETCH_HEAD had not
moved in 19 days. So:
- `ahead` needs only local refs and is never flagged.
- `behind` is flagged `stale` once FETCH_HEAD is older than an hour.
- Exactly ONE state may render green: ↑0 ↓0 AND a fresh fetch. Green means
"I checked, ignore this"; getting it wrong is lying to the user.
- No upstream (the normal state of a fresh worktree branch) leaves ahead and
behind undefined — it renders an explicit `no upstream`, never the green
path. That fall-through is the easiest bug to ship here.
What landed:
G1 SyncState (upstream/ahead/behind/lastFetchMs/detached) + ProjectDetail.sync
and .dirtyCount. All additive and optional — the Android and iOS clients
decode these shapes. The ahead/behind helper already existed for the list
view; buildProjectDetail had simply never called it.
Fixes a pre-existing bug on the way: readBranch read <repo>/.git/HEAD
directly, so it returned nothing inside a LINKED worktree, where .git is a
file. resolveGitDirs now resolves both the per-worktree gitdir (HEAD) and
the shared common dir (FETCH_HEAD).
G2 POST /projects/git/fetch. Same discipline as push: the remote is derived
server-side and no remote or refspec is ever read from the body, so a
client cannot aim it at an arbitrary URL. Touches refs/remotes only — no
working tree, no index, no merge; it is not a pull. Own rate-limit bucket
so refreshes cannot eat the budget a real push needs. On failure
lastFetchMs is left alone, so the UI keeps saying "stale" instead of
pretending it refreshed.
G3 makeSyncBand replaces the bare dot: upstream name, ↑n, ↓n, stale flag,
dirty count, Fetch button (disabled on a detached HEAD).
G4 The commit list marks unpushed commits and draws the upstream boundary
once, after the last of them. Marking is server-side from `rev-list`,
deliberately NOT "the first N rows": `git log` is date-ordered, so merging
an older branch interleaves unpushed commits BELOW pushed ones, and that
shortcut fails in the dangerous direction — calling an unpushed commit
pushed. A regression test builds exactly that backdated-merge shape.
G5 The worktree section is always "Worktrees (n)" (it used to rename itself
to "Branch" at n=1) and the current row carries its own state chips.
G6 Cost control. The plan called for a .git-mtime cache; that was dropped
during implementation because a fingerprint over HEAD/index/reflog does
NOT move when push updates a remote-tracking ref — the cached `ahead`
would still claim "9 to push" right after a successful push, which is the
exact lie the feature exists to prevent. Replaced with three measures that
cannot go stale: in-flight coalescing (N devices watching one repo cost
one probe, entry dropped as it settles, nothing cached across time),
skipping the re-render when the payload is byte-identical (this also stops
the 5 s re-mount of the commit log, two more git spawns per tick), and
pausing the timer while the document is hidden.
G7 Per-worktree state via GET /projects/worktree/state, kept narrower than
/projects/detail so N rows do not pay for worktree listing and CLAUDE.md
reads nothing renders. Needed an unplanned prerequisite: ProjectSessionRef
carried no cwd, so sessions could not be attributed to a worktree. Added
it, plus countSessionsByWorktree, which matches DEEPEST-first because
.claude/worktrees/<name> lives INSIDE the main checkout and prefix
matching would count every worktree session against the parent repo too.
Out of scope, unchanged: no reset, no checkout, no clean, no rebase, no
force-push. stage/commit/push stay exactly as they were.
Verified: tsc and build clean; 46 new tests.
`worktree.baseRef: head` is the setting CLAUDE.md's session workflow depends on:
without it EnterWorktree defaults to `fresh` and branches from `origin/main`,
which trails `develop` by ~75 commits. Shared config, so it belongs in the repo —
personal overrides still go in the ignored `.claude/settings.local.json`.
Sessions that change files now work in their own git worktree and merge back to
`develop`, instead of developing directly in the main checkout. Documents the
tool's actual behaviour rather than the assumed one: `EnterWorktree` prefixes the
branch as `worktree-<name>`, branches from the last commit (so uncommitted edits
do not carry over), and `ExitWorktree({action: "remove"})` deletes the branch
with the directory — hence the merge must happen before the cleanup.
Base ref is pinned to `head` in `.claude/settings.json` because `develop` runs
well ahead of `origin/main`, and the default `fresh` would branch from the stale
release trunk.
Also ignores `.claude/worktrees/`, which lives on disk per session and is never
committed.
Closes the three leftovers called out after the renewal-deadlock fix: a source
file the blanket `dist/` ignore rule kept out of every checkout, tunnel logs
that named no host (including all 6380 warnings during the outage), and the
missing phone-track counterpart to /recover.
1. `.gitignore` swallowed a source file. `agent/src/dist/buildBinary.ts` is the
packaging config, not build output, but the blanket `dist/` rule meant it was
never committed — a fresh clone could neither typecheck `agent/src/index.ts`
nor import it from the committed `agent/test/buildBinary.test.ts`. Re-included
the DIRECTORY first (git does not descend into an excluded one, so un-ignoring
just the file would not have worked) and committed the file. Verified build
output under `agent/dist/`, `dist/`, `public/build/` is still ignored.
2. Tunnel logs named no host. `pair` learned hostId/subdomain from the enroll
response and threw them away, so the long-running `run` process logged
`{"subdomain":null,"hostId":null}` — including all 6380 warnings during the
8-day outage, at exactly the moment you want to know which host. New
`config/hostRecord.ts` persists them at enrol; `resolveHostIdentity` resolves
config > record > the subdomain embedded in the leaf's SPIFFE SAN, so hosts
enrolled before the record existed get an identifier back without re-pairing.
3. The phone track had no recovery path. Added `POST /device/:id/recover`,
mirroring the host route: cert in the body, device-CA path validation, SPIFFE
parse, `notBefore` never graced, and the full registry check (active + same
account + `:id` matches the cert + same key) — only `notAfter` is relaxed.
Verified: agent 300/300, control-plane 296/296, tsc clean on both.
NOTE: the iOS/Android clients are not wired to call /device/:id/recover yet —
the server capability exists, the client-side trigger does not.
The tunnel had been down 8 days because /renew is mTLS-authenticated by the very
leaf it renews, so a lapsed leaf could never renew itself. Recovery now runs over
plain HTTPS with the expired cert in the body (CSR self-signature still proves key
possession), bounded by a 30-day grace; past that the agent stops retrying and
names the re-pair. Verified live end-to-end: an 8-day-expired leaf self-healed and
the tunnel returned HTTP 200 with no human action.
They were local worktree conveniences (symlinks into the main checkout so vitest
could resolve the file: workspace deps), swept in by a broad `git add agent`.
Correction to the previous commit. Its recovery vhost could not work: nginx
will not forward an expired client certificate in ANY `ssl_verify_client` mode.
`optional` answers a bare `400 The SSL certificate error` before any location
runs, and `optional_no_ca` only tolerates CHAIN failures — see nginx's
`ngx_ssl_verify_error_optional()`, which covers self-signed / unknown-issuer /
unverifiable-leaf and NOT `X509_V_ERR_CERT_HAS_EXPIRED`. Verified live against
the deployed :8472 server, which rejected the real expired leaf.
So recovery drops mTLS instead of trying to bend it:
- agent: `buildTlsOptions` and the mTLS renew transport go back to being
strictly fail-closed on expiry — the relaxation is gone from the TLS layer
entirely. The rotator now decides per attempt: valid → mTLS `/renew`,
expired-inside-grace → plain `/recover`, expired-beyond-grace → terminal
`onExhausted` with no request issued at all.
- new `recoverCert` POSTs `{cert, csr}` with no client certificate. Possession
of the private key is still proven: the CSR is self-signed by it and the host
signer already enforces CSR PoP plus `CSR key == registered key`, so a
replayed (public) cert without the key yields at most a certificate the
attacker cannot authenticate with.
- control-plane: `/renew` is strict again (grace 0, matching the terminator).
The grace lives on the new `POST /recover`, which reads the cert from the BODY
and ignores the `x-client-cert` header, then runs the identical trust
pipeline: X.509 path validation to the frp-client-CA anchors, SPIFFE parse,
`notBefore` (never graced), registry `active` + account match. Revocation
still bites.
- deploy: no new vhost, no new SNI, no DNS. One `location = /recover` merged
into the existing enroll vhost, documented in
`deploy/nginx/enroll-recover-location.md` with the nginx source citation.
The load-bearing new test is "a self-signed cert with a FORGED SPIFFE SAN is
refused → 401": nginx no longer validates the chain on this path, so that
assertion is what keeps `/recover` from being a cert vending machine.
Verified: agent 289/289, control-plane 290/290, tsc clean on both.
`POST /renew` is authenticated by mTLS with the very leaf it renews, so once
that leaf lapsed the host could never renew it and the tunnel stayed down until
an operator re-paired by hand. Production hit exactly this: the Mac slept
through its 8h renewal window, the 24h leaf expired, and the agent then logged
`client certificate has expired; renew before dialling` 6380 times over 8 days
without recovering.
Three layers independently refused an expired leaf, so all three had to move:
- agent `buildTlsOptions` gains an opt-in `expiredGraceMs`. Absent/0 keeps the
historical fail-closed behaviour, and the TUNNEL dial never passes it — only
the renew transport does. Past the window it throws the new
`CertExpiredBeyondGraceError`.
- agent rotator routes an already-expired leaf to a separate recovery endpoint
and treats "beyond grace" as TERMINAL: report once via `onExhausted`, stop
retrying, and name the fix (re-pair) instead of spamming warnings forever.
- control-plane `assertPresentedCertTrusted` grants a bounded grace on
`notAfter` only. Chain validation, SPIFFE identity, `notBefore`, and the
registry `active`/account checks are all unchanged, so revocation still bites.
- new `deploy/nginx/recover-mtls.conf` (:8472). The strict enroll vhost cannot
host this: under `ssl_verify_client optional` nginx answers a bare 400 as soon
as a presented cert fails verification, so an expired leaf never reaches the
location — and the directive is server-level, not per-location.
Grace defaults to 30 days on both ends and is configurable (`RECOVER_URL`,
`expiredRenewGraceMs`). The honest trade is recorded in the code: a stale stolen
leaf stays reusable for the window, which widens an existing exposure (an
unexpired stolen leaf already renews indefinitely) rather than opening a new one.
Verified: agent 296/296, control-plane 286/286, tsc clean on both.
A launchd/Finder-launched server has no LANG/LC_* at all, so tmux — which
decides per client whether the terminal is UTF-8 capable purely from those
variables — fell back to its non-UTF-8 mode and rewrote the byte stream
server-side, before it ever reached xterm:
- every wide character became `_` (中文测试 → ________)
- `✓` / `⏺` / emoji became `_`
- `═║╔╗` was downgraded to DEC Special Graphics (ESC ( 0), so rounded
corners vanished and boxes rendered as loose horizontal lines
The `-l` login shell cannot fix this: tmux is the PTY's root process and has
already made its decision by the time the shell sources ~/.zprofile.
withUtf8Locale() fills a locale in on the spawn env, and only when needed —
an existing UTF-8 LC_ALL / LC_CTYPE / LANG (zh_CN.UTF-8, ja_JP.UTF-8, …) is
left untouched; only a missing or non-UTF-8 charset is replaced.
Existing tmux sessions do not need recreating: re-attaching with a UTF-8
client restores correct output (already-scrolled `_` stays mangled).
- wire the enroll flow into the app UI (EnrollmentScreen + ViewModel + Hilt DI +
host-menu "自动获取证书"), mirroring iOS — Android previously only had manual
.p12 import; the enroll library was built but unreachable.
- renew is now mTLS-only ({csr}-only body, no Authorization header) matching the
/device/:id/renew contract (the enroll bearer is minutes-lived → silent
rotation would have thrown weeks later).
- enroll refreshes the identity-repository cache so a mid-session-enrolled cert
is presented on the next mTLS handshake without a process restart.
gradle :app:assembleDebug + api-client/client-tls-android unit tests + koverVerify
green. On-device QA (keygen/enroll/present) is the operator's step.
launchd has no default log destination (unlike systemd's journald), so a unit
that fails to start (bare node, missing env) was silent — which hid the EX_CONFIG
+ loadConfig failures during this deploy. Route both units' stdout+stderr to
<stateDir>/{base-app,agent}.log. 281 tests pass.
Two real-deploy renew bugs (found running the live /renew):
- the /renew mTLS request pinned the enroll caChain as the server CA → TLS
'unable to get local issuer certificate' (the LE-fronted CP is publicly
trusted). Verify the server against SYSTEM roots (drop ca), keep the client
cert + rejectUnauthorized:true. (TlsClientOptions.ca now optional.)
- renewCert parsed {cert, caChain:string}, but the CP returns cert=base64(DER)
+ caChain=base64(DER)[]; normalize to PEM (shared certs/pem.ts, reused by
native enroll). Verified live: cert rotated 13:41→next-day, frpc restarted,
tunnel stayed up. 281 tests pass.
nginx has no njs deployed, so the mTLS terminator forwards the verified client
cert via $ssl_client_escaped_cert (URL-encoded PEM). headerPresentedCert now
normalizes base64-DER, PEM, and escaped-PEM alike to raw DER. 281 tests pass.
Fixes found deploying `pair --install` against the live control-plane:
- native enroll tolerates hostContentSecret:null and converts the base64-DER
cert + caChain[] the CP returns into PEM for frpc/keystore
- launchd/systemd units use an absolute node (process.execPath) + a PATH with
/usr/local/bin (bare `node` died with EX_CONFIG 78)
- the agent unit now carries ENROLL_URL/RELAY_URL/STATE_DIR/LOCAL_TARGET_URL
(`run` loadConfig()-validates them; missing → supervisor exited 1)
281 tests pass. Follow-up: add unit tests for the native-enroll tolerance +
agent-unit env (only install node-path is covered so far).
Wire the prod boot to issue frp-client + device leaves from the existing on-disk
P-256 CAs: NATIVE_* env vars, a file-backed KmsResolver (loads the PEM key,
validates prime256v1, never logs key material), buildFileBackedNativeCas threaded
into buildControlPlane via a nativeCas override. Fail-closed on partial NATIVE_*.
281 tests pass.
When the always-on tunnel daemon already serves :3000, the app reuses it
(close() is a no-op) rather than spawning a second server on another port, so
the GUI and the tunnel share one base app.
Wire the SecureEnclave enroll library into a real flow (login->bearer->CSR->
/device/enroll->keychain identity), presented on the existing mTLS path; add a
rotation scheduler. Atomic keychain replace (add-before-delete); renew body is
{csr}-only; renewal-failing surfaced in the UI. ClientTLS 48 tests pass.
Wire createCertRotator/renewCert into superviseNative so the frp-client cert
renews silently at ~2/3 TTL over mTLS /renew and frpc restarts onto the fresh
leaf; failures retry with backoff, never crash the supervisor. mTLS transport
has a 15s timeout + 64KB response cap. 281 tests pass.
Unblocks the phone-enrollment track: an operator-password login mints a
short-lived device:enroll capability token that POST /device/enroll requires.
Constant-time (SHA-256 fixed-length) compare, per-client rate-limit, fail-closed
when unset. 260 tests pass.
Execution plan grounded in current VPS reality: pre-flight checks, Track A
(host auto-enroll via the agent) and Track B (phone auto-enroll), risks, and
the honest one-bootstrap-tap constraint. R1 (CA material) verified present on VPS.
An OPTIONAL shared token so the app can be used off-LAN (via the relay/tunnel) more
safely than "anyone who reaches the port gets a shell". Sits IN FRONT OF the existing
Origin/CSRF model — never replacing it — and is fully inert when unset.
- src/http/auth.ts (new, pure): constantTimeEqual hashes both inputs to sha256 (32
bytes) then crypto.timingSafeEqual — no `===`, no length side-channel, never throws.
parseCookieHeader / cookieIsAuthed / buildSetCookie / isAuthEnabled.
- WEBTERM_TOKEN in config: 16–512 URL/cookie-safe chars or the server refuses to start.
- GET /?token=<t> or POST /auth (rate-limited 10/min) validates → sets
HttpOnly; SameSite=Strict; Secure-when-https cookie; public/login.html (no <script>).
- When enabled: the WS handshake (AFTER the Origin check, before handleUpgrade) + a
central authGate over all remote HTTP require the cookie. Open: /login, /auth.
Loopback bypass scoped to /hook* ONLY (tighter than the plan — gates the local
browser too).
- Unset ⇒ isAuthEnabled false ⇒ pure passthrough (LAN zero-config unchanged).
Honest tradeoff (in code + CLAUDE.md + login page): a bar-raiser, NOT a TLS/Tailscale
substitute — on bare ws:// the token is cleartext + replayable; only hardens the
TLS-terminated relay path. Verified: typecheck + build:web clean, 2118 pass at
--test-timeout=30000 (disabled-mode regression proven; only the known tmux flake red);
auth.ts 100% line coverage.
