ATimedOutComparisonIsRecordedAndLiveUpdatesGoOn polled until the server's
settings reached the local copy and then asserted that the link's error was
cleared. The engine writes the local copy first and stores the link (error
cleared, versions agreed) afterwards, so a poll landing between the two writes
saw the old error: 1 failure in 60 runs idle, 11 in 60 under CPU load.
The agent tests now wait for SyncService.Changed, which the engine raises once
a comparison stored every link and state (the moment the Server screen reloads),
and assert on that settled state. The same fix applies to the proxy refresh
test (the cache is complete before the link and the job: it waits for the job,
whose row showed Transferring until then) and the retry test (the agent stores
the snapshot before it clears the last error and raises Changed).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Union of the Tooltips resources; sample data keeps the sync server and the
simulated cards; the import test names App.Infrastructure.Formatting fully now
that a test namespace of the same name exists.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
- Conflicts are settled against the versions the user was shown; a later change on
either side refuses the choice and compares again. Title, tags, rating and notes
changed differently on both sides are a conflict too (no silent tag union).
- Links are written under the engine's gate and forgotten only after a 404; linked
local copies are read by id; unlinked copies with the same originals are linked
by content.
- A failure on one sequence is recorded on its link; the others are still compared.
- Downloads hash unverified local copies first and restore originals missing here.
- Background proxy downloads give way to the user's requests; "Keep proxies offline"
refreshes the caches of sequences opened with proxies, checks free space and backs
off after a failure.
- Renders travel on request: Send exports (chunked, hash-verified upload keeping the
render id; new POST /sequences/{id}/renders endpoints) and Get exports.
- Server screen: stable row buttons, mockup column grid, Tab through rows, conflict
rows first on the card colour, localised conflict detail, Save both confirmation,
Pause metered uploads, culture switch, Library one-word sync glyphs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
- The rebuild gate adopts 400 sequences of 250 real photos (one fifth of the
§1.1 library, same frames per sequence) and runs on SQLite and PostgreSQL.
- The rebuilt index is compared with the original in every column of every
sequence and frame (only UpdatedAt, the rebuild time, is left out), and the
first differing row is named.
- Synthetic frames are seeded by the sequence's position, so every run builds
the same library; the Library open gate says it runs on a new connection pool
with a warm OS cache instead of calling it cold.
Measured locally (PostgreSQL 16): Library open 259 ms SQLite / 181 ms PostgreSQL,
filters <= 6.4 ms; rebuild of 100,000 frames 1.8 s SQLite / 4.7 s PostgreSQL
(budget 120 s).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
"1 sequenze su 3" read wrong in Italian: a single failure now reports its
own line (title and reason); the count is kept for two or more. A worker
whose job could not be finished while the engine stops logs it instead of
faulting, so disposing the engine never throws.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
BRIEF §14.7: the server falls back to managed formats when no node is
available and says so in the job log. An MP4 job waits for render nodes at
least the node timeout, then as long as an online node can encode its codec
at its size; after that the server's engine takes it and writes each output
it cannot encode as MJPEG in AVI at the same size and frame rate, with a job
log line. A job the server started resumes there. Selective acceptance is
asynchronous for that; when node availability cannot be read the job keeps
waiting.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
A render reserves the output name it picks, so the same preset queued twice
no longer overwrites one output with the other; render rows take ids from
their job and entry, so a resumed job recognises only its own outputs and
never deletes another render's file. A light pass records its status under
the sequence lock and cannot clear a deep pass's. Renders and Auto value
changes made by the server engine publish sequenceChanged events.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
A node registering again gives back the job its previous run abandoned, so
a restart within the heartbeat timeout no longer leaves it running forever.
Frame downloads retry dropped connections from the partial file; a 404 at
registration waits for the next poll; render remarks reach the job log; a
render the server engine started stays with it. POST /jobs splits a batch
into one job per preset, and a node fails a batch it cannot send back
instead of reporting it done. Renders honour Settings > Rendering: hardware
encoders only where the preference allows them, also for what a node offers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Accepts that throws now claims the job so it fails with its reason; claim
and finish errors are logged and the worker carries on. Other kinds keep
their atomic claim next to selective ones, and a lost race retries at once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Moving them to Jobs.Analysis.* deleted keys that the integration branch
translates into every culture. The advisor resource test now covers the
Advisor.Log.* job lines, as on the integration branch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
The first original of an upload now creates the server sequence through its
session key (the archive ingest files API-created empty sequences nowhere), and
settings sent after an upload are retried when the server's own analysis edits
the new sequence meanwhile.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Category perf, skipped unless TITANO_PERF=1. A synthetic library of 2,000
sequences and 500,000 frames (SQLite, and PostgreSQL when TITANO_TEST_PG is
set) must open the Library in under 1 s cold and answer every filter, search,
sort and next page in under 200 ms; a 100,000-frame archive adopted from real
files must rebuild from its sidecars, with and without the index snapshot,
into the identical index within the scaled 2-minute budget. perf.yml runs
them weekly and on demand against PostgreSQL 17 and writes the timings to the
job summary.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
- JobEngine: ISelectiveJobHandler lets a host decline jobs of a kind it runs;
declined jobs stay queued for other claimers without holding up the rest.
JobContext.EnqueueFollowUpAsync queues follow-up work on the same engine.
- Server renders the jobs whose every output a managed encoder produces
(ManagedRenderJobHandler); MP4 jobs stay for render nodes. Program.cs
registers ProcessingModule next to ArchiveModule. Server renders ignore a
client's export folder.
- Light analysis is queued after imports and after scans that indexed new
sequences, so the advisor's Auto values exist without user action.
- Analysis jobs log a failing sequence and go on; the job then fails and a
retry analyses only the failed ones. The advised frame rate follows the
user's fit-to duration.
- ProcessingJobs.RenderEach splits a batch into one job per preset for
servers; render nodes say which batch presets they did not render, fetch
each content once and stop on a changed certificate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
An engine with no handlers (or a filter it could not serve) claimed every
job, so a host without a render handler would fail render-node work.
With several workers, jobs carry node ids such as pc#1, but crash
recovery only matched the bare node id and left them running forever.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
ArchiveService facade over indexer (scan, adopt, rebuild), ArchiveIngest
(IArchiveIngest with per-sequence group commit, leases and a write-ahead
journal), verifier, maintenance (duplicates, orphans, cleanup), mass rename
with resumable journal, thumbnails/proxies and card import; job handlers for
scan, rebuild, verify, import, thumbs, proxies and maintenance with checkpoints.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
SQLite steps ran synchronously on the calling thread, so a repository
awaited on the WPF UI thread would freeze the window for the length of
the query. Blocking calls now continue on the thread pool.
A cancel that landed before sqlite3_step started the statement was lost
(SQLite clears the interrupt when no statement is running) and the query
ran to completion. Each connection now has a cancel flag polled by a
progress handler; it is cleared outside every interruptible operation so
a stale cancel cannot stop a later statement.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
The test timed from before CancelAfter, so a thread-pool timer firing
late on a busy 4-core Windows runner (parallel tests block pool threads
in synchronous SQLite steps) counted as interrupt latency. A dedicated
thread now records the moment it cancels.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
JobEngine claims queued jobs atomically, persists progress at most every
checkpoint interval, stops a handler when its job is cancelled or paused
elsewhere and requeues jobs a crash left running.
Also installs the SQLite import resolver under a lock so a concurrent
first caller cannot reach a P/Invoke before the resolver is in place.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
- SQLite restore holds every pool slot instead of the write lock first,
removing a lock-order deadlock with sessions that begin to write.
- PostgreSQL sessions always return their connection on dispose, a
rollback while disposing never masks the original exception, and a
busy session keeps its transaction until it is discarded.
- The pool disposes a connection whose health check the caller cancels.
- Listener commands keep their lock until the server answers and never
tear a write with the caller's token; closing and pending no longer race.
- Tag replacement locks the sequence and the tag row; unused-tag cleanup
locks candidates and re-checks them with a fresh snapshot; render
counts and stats refreshes lock the sequence row on PostgreSQL.
- Sequence upserts keep rendered_count, which renders maintain.
- AuthenticationOk after an unfinished SCRAM exchange is rejected.
- Oversized read buffers are released; failed backups leave no .tmp;
a zero idle lifetime is rejected.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Pin bytea_output=hex and standard_conforming_strings=on in the startup
packet instead of tracking ParameterStatus values nobody read, so the
legacy bytea escape decoder goes away. A PostgreSQL connection now
refuses a new command while a row stream is still open, and the pool
discards such a connection. Drop unused SQLite declarations.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Shared abstract suites run against in-memory/temp-file SQLite and, with
TITANO_TEST_PG, a private PostgreSQL schema per test: migration
idempotency and version refusal, keyset paging with ties for every sort,
statistics, exclusive job claiming under 8 concurrent claimers, and the
100k-frame bulk insert under 15 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Shared session tests run on SQLite and PostgreSQL (types, nulls,
placeholders, transactions, constraint errors, COPY/bulk escaping,
cancellation). PostgreSQL tests cover SCRAM/MD5/cleartext/trust auth,
TLS, result decoding, bytea, LISTEN/NOTIFY, error mapping, pooling and
broken-connection recovery, and the statement cache. In-memory SQLite
now pools shared-cache connections and retries SQLITE_LOCKED.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC