The guard's `pgrep -f 'run-task\.sh'` matched the orchestrator's own stale
launching shell, whose command line contains that text, so it kept the sprite
awake for 17 hours after J-crashsafe finished. workers_running() now requires
argv[1] to be the runner script and argv[2] to be one of this wave's tasks.
Also commits the wave-7 sid/sentinel that were never added.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AQb3d68LqWD4EdND3C8jkD
The user's first scan on the wave-7 build crashed and never reproduced. Their
theory fits structurally — the Google Books Found path had never executed in
production before the key landed — but nine live lookups through the real
repository, including both of their ISBNs, threw nothing, so the parse/merge
code is exonerated by evidence rather than by argument. Written down so no
future worker "fixes" code that was measured working.
What the wave fixes is the defect found while looking: the lookup path catches
only IOException and the ViewModel catches nothing, so any other throwable kills
the process instead of reaching the user as "couldn't be reached" — and the app
has no crash capture at all, which is why one crash left no evidence.
Committed with an explicit pathspec: a wave is in flight (HAZARD #7).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PPpdG8VnRfS3KkisR3HUAE
Found while re-verifying the security posture against the public hostname now
that the server is internet-exposed. pb_hooks/main.pb.js covered bookcases,
shelves and books but never users, so an anonymous GET of the accounts
collection answered 200 with an empty array.
Nothing leaked: two real accounts exist and the declarative listRule filtered
both out, so no account, email or id was ever visible to an anonymous caller.
But "200 with []" is the exact wrong signal this hook exists to remove — some
clients read it as an allowed request — and the accounts collection is the last
place to leave it. Defensible while the server was localhost-only; not now.
Re-verified over the internet after restarting the service: books, shelves,
bookcases and users all 403 anonymous, self-registration 403, health 200.
Also re-verified that login still works, since this hook now runs on a
collection the app authenticates against: auth-with-password returns 200 with a
token, and an authenticated LIST of all four collections still returns 200. The
hook rejects unauthenticated list/search only, and auth-with-password is not a
list request.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PPpdG8VnRfS3KkisR3HUAE
The long-standing open question "where will the server actually live" is
answered for now: it stays on this sprite, bound to 0.0.0.0:8090 and published
over HTTPS by the sprite proxy. Several docs asserted the opposite — HANDOFF
said "127.0.0.1:8090, deliberately NOT internet-exposed (no --http-port, so the
sprite proxy can't reach it)", which is flatly wrong today.
The consequence is the part worth writing down: PocketBase's API rules are now
the only thing between this library and the internet. There is no NAT, no VPN,
no reverse proxy. So the anonymous-access curls stop being a formality, and they
have to run against the PUBLIC hostname — localhost cannot tell you what the
world can reach. Re-verified that way: books/shelves/bookcases LIST all 403,
self-registration 403, health 200.
One gap found while re-verifying, recorded but NOT fixed: users LIST answers 200
with an empty array instead of 403. Nothing is disclosed — two real accounts
exist and the listRule filters both out — but it is the same wrong-signal quirk
pb_hooks/main.pb.js exists to close, and that hook never listed the users
collection.
SPEC's offline-first rationale is amended rather than its rule: the reason is no
longer residential NAT but a sprite that suspends when idle and wakes on
request. The rule is unchanged and does not depend on which.
server/deploy/ still documents systemd/Docker/Tailscale on home hardware; it now
says up front that this is the intended end state, not what is running.
Also corrected, since it was adjacent and plainly false: README still claimed the
app had never run on a physical device. It has, since 2026-09-09. What is true is
that no *automated* test runs on a device — there is no emulator on this box.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PPpdG8VnRfS3KkisR3HUAE
The quota detector grepped `cat "$LOG" "$ERR"` for strings like "429" and "usage
limit". $LOG holds the worker's final JSON, including its .result prose — so a
worker whose TASK was about HTTP 429 finished successfully, wrote a report saying
so, and the runner read its own worker's words, logged "QUOTA hit", slept 600s
and was about to --resume a session that had already succeeded. That would have
burned quota redoing finished work with a fresh worker loose on a completed tree.
Two independent defences, because either alone suffices:
- the blob is now stderr plus the result text ONLY when is_error is true,
falling back to the whole log when it isn't valid JSON (a hard crash writes
no JSON, and has no prose to be confused by).
- the success check runs BEFORE the quota and session-vanished checks. A
finished worker is finished regardless of what strings its output contains.
Regression-checked against the real logs/I-gbkey.json that triggered this: the
old logic matches the quota pattern, the new logic yields SUCCESS.
Recorded as HAZARD #9 in docs/HANDOFF.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The user obtained a restricted Google Books key. Keyless requests 429 for every
caller on the internet — all anonymous traffic bills to one shared Google Cloud
project whose daily quota is permanently exhausted — so the documented fallback
has never once answered. Because MetadataRepository.combine turns "a source
failed, none found" into Unavailable, that standing failure meant every Open
Library hiccup reached the user as "couldn't be reached". The app has been
effectively single-sourced since it was written.
Build plumbing reads GOOGLE_BOOKS_API_KEY from local.properties (gitignored),
falling back to the environment and then to empty. A blank key is a supported
state: a fresh clone still builds a working app that falls back to the keyless
endpoint, rather than failing to build.
GoogleBooksClient appends the key only when non-blank, building the URL with
HttpUrl.Builder in a pure requestUrl() so it is testable without a socket. The
key is scrubbed from SourceResult.Failed.reason before that string can reach the
scan sheet — it is rendered to the user and is our only diagnostic channel from a
real phone, and some okhttp/JDK IOExceptions embed the full request URL in their
message. Defensive, not a response to an observed leak.
Resolves the RATE_LIMITED decision parked in RetryPolicy's KDoc: a keyed 429 is
the short per-user rate limit and gets exactly one retry, honouring Retry-After
capped at 2s. A keyless 429 is still the dead daily quota and is still never
retried.
Verified against the live API, not only offline: both ISBNs that failed on the
phone (9781883937386, 9781883937676) plus a control return HTTP 200, in the
percent-encoded URL shape HttpUrl actually produces. Both books are in Google
Books, so the restored fallback now covers precisely the Open Library TLS-reset
failure that broke those scans.
189 unit tests (was 172), 0 failures; Paparazzi unchanged; release APK builds.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J7WHnTx2Cso4VV245WDAJY
The ghost bookcase was an inset bug, not a data bug. LocationsScreen's list
branch dropped the Scaffold's innerPadding while its empty-state branch applied
it, so the first bookcase row rendered under the top app bar and was invisible.
Both of the user's bookcases were always real; they just could not see the first
one, so they made a second. Every other screen was checked for the same class of
bug — Locations was the only one.
Metadata lookup is now designed against a measurement rather than a guess
(docs/METADATA-SOURCES.md § "Measured again 2026-09-09"):
- Google Books keyless is dead for everyone, permanently. The user's
residential-IP test returned a quota error naming a shared anonymous PROJECT,
not an IP, so the earlier "your phone may well get answers" guess is wrong and
is now marked CORRECTED in place. Because combine() turns any Failed-with-no-
Found into Unavailable, that standing failure meant every Open Library hiccup
surfaced as "one or more sources couldn't be reached". The API key is
deliberately deferred by the user; this commit leaves the source broken.
- Our own timeouts were manufacturing failures. Over 30 live requests, 13%
failed — all fast TLS resets under 2.5s — while successes ran to a median of
4.3s and a max of 22.0s. Two of 26 successes exceeded the old 12s callTimeout,
so ~8% of lookups that were about to work were cancelled and reported as
unreachable. Timeouts are now 25s/20s/20s.
That asymmetry (cheap failures, expensive successes) is what RetryPolicy encodes.
It retries TRANSPORT and SERVER_ERROR with a 250ms/750ms jittered backoff, and
deliberately does not retry TIMEOUT (the budget is already spent) or RATE_LIMITED
(hammering a quota is how an intermittent block becomes a permanent one — this
project's IP has already been refused outright once during research).
SourceResult.Failed now carries a FailureKind alongside its human reason, and the
reason names the specific failure ("tls connection reset, 3 attempts") instead of
a generic "network error". That string was already threaded to the UI and dropped
on the floor; LookupFailedSheet now renders it. It is the only diagnostic channel
we have from a real phone, so nothing may parse it.
Also from the same feedback round:
- Grouped ModalBottomSheet shelf picker, replacing two near-duplicate flat
dropdowns that listed every bookcase x shelf pair. Sections per bookcase,
empty bookcases say so, and the most recently used shelf is pinned on top.
- The recent shelf persists across sessions (SettingsStore.LAST_SHELF_ID) and is
cleared on sign-out. It is offered, never pre-selected: the user weighed that
and chose one tap over the risk of silently mis-shelving a book.
- Locations dialogs and the manual-ISBN dialog auto-focus their first field.
- The library filter menu offers "Add a bookcase to enable filtering" instead of
a lone "All books" that is already the active state and cannot be changed.
- The Locations button is Material Symbols' "shelves" (a bookcase) instead of
Warehouse (a barn). material-icons-extended 1.7.8 has no bookcase glyph.
- The scan sheet drops "you can lower the book" — the ISBN echo already says it.
assembleDebug exit 0; testDebugUnitTest 172 tests, 1 skipped, 0 failures (was
138); verifyPaparazziDebug exit 0; assembleRelease exit 0, signed with the real
release key; zero "always 'false'" warnings on a --rerun-tasks rebuild.
Three soft spots are recorded in docs/HANDOFF.md and are NOT verified: the
ghost-bookcase Paparazzi snapshot renders a lookalike of the screen rather than
the screen, the auto-focus calls swallow their own failure and no emulator exists
here, and the picker opens as a sheet stacked on the save sheet.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LSnVqWdiQNEcPFRq1hGZAi
The wave-guard was killed after its 11:13 renewal and never wrote
logs/WAVE5-DONE, despite the worker succeeding at 11:21. That makes this
file's own first-command heuristic actively misleading — "no sentinel +
no processes -> workers were KILLED" would have thrown away a completed,
verified wave. Recorded the reliable signals instead: the size of
logs/<name>.json and the tail of logs/<name>.state.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Explicit pathspec only — a worker is writing to the tree right now
(hazard #7: `git add -A` mid-wave has swept half-finished source into
commits twice on this project).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Records the coil3 StateFlow trap (a Kotlin warning, not an error, that
silently blanked every book cover) so a future session greps for it, and
the two Open Library cover defects behind it.
METADATA-SOURCES.md answers the user's research question about
alternative lookup sources: measurements over a 60-ISBN sample drawn
from a third-party catalogue, the options with costs, and a
recommendation to fix diagnosis before buying coverage. Nothing there is
implemented — the user asked to be consulted first.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CDcPottghJXEvfYKqFM7zf
Wave 4 is verified complete. Also records the two mechanisms that stopped F3
from reporting (the 600s background-task ceiling in `claude -p`, then the quota
wait) so the next worker launch does not repeat it, and flags the one thing
wave 4 owed and never delivered: nobody has looked at the fourteen new
screenshots.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014TyzeWmdTqi7U85iYNGy7P
The wave-3 loss was caused by sprite auto-suspend, not nohup process-group
semantics. /.sprite/llm.txt: 'When idle, sprites pause automatically. Services and
sessions keep sprites alive.' Detached processes are on neither list, so setsid is
necessary but not sufficient.
tasks/wave-guard.sh holds a /v1/tasks lease (max 3600s, renewal is DELETE+POST since
re-POST returns 409), renews every 15 min while workers run, writes logs/WAVE<N>-DONE,
then releases the lease so the sprite can suspend rather than idle hot.
Also documents hazard #6 (pgrep -f / pkill -f matching the orchestrator's own shell)
and adds the wave-3 worker prompts and shared screen contract.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Spriteandclaude