* fix(logs): Clear all logs deletes and empties only what its confirmation showed (#389)
- /logs/files rows carry an identity string (size:mtime_ns:ino); the
workspace sends back the active and rotated rows its dialog was built
from. Under the files and handler locks every listed copy is re-checked
with _matches_review and the listed active log's inode; any difference
refuses with nothing changed (409 logs-changed). A copy that appeared
after the confirmation is kept.
- A second writer of the active log (our flushed stream is behind the
file's size, or our file is now a rotated copy) refuses first (409
logs-foreign-writer); each unlink is preceded by a fresh identity check
and the active log is emptied last.
- An active log with a hard link (st_nlink > 1) is never truncated (409
logs-hard-link), also on the file-logging-off path.
- The legacy admin page's POST without a body keeps clearing what is
there, with the foreign-writer and hard-link refusals.
- The browser fixture grows the active log through the handler's stream.
* fix(logs): accept pre-1970 identities, refuse duplicate rows, say the foreign-writer check is best effort (#389)
Review follow-ups: a negative mtime_ns no longer makes every workspace
clear 400; a name listed twice is a malformed list; the confirmed-row cap
is 10000; the hard-link refusal says how to recover; the hard-link and
foreign-append tests assert the kept bytes before the refusal code, so on
the old code they fail on lost data.
* test(browser): API restores stay unseen by the open page's Activity poll (#413)
Specs that restore through the API while a workspace page is open check
how the page handles a restore it did not see. The page's Activity store
reads the job list every 15 s; on a slow runner (CI run 40646112668, a
16 s restore) that read landed during the restore, the page followed the
job and, as designed (S9), showed the open page again when it finished,
closing the Create tag dialog the vocabulary spec was about to use.
unseenRestore holds the page's job-list reads until the restore has
answered, so they read it as already ended. Used by the vocabulary,
library editor and history specs.
* test(browser): arm the restore gate before the page loads and drain reads in flight (#413)
A job-list read the page sent before the restore could still be answered
while it ran. The gate now routes every job-list read from before the
first navigation and, before a restore, waits for the reads in flight to
be answered. The history and library editor conversions are pre-emptive:
same race, never seen failing.
* fix(review): anchor files on their saver in the scan and Needs review (#390)
The scan and Apply now resolve with the file's name, as the workspace
inspector does since #383, so the one saver whose filename_prefix names
the file anchors its branch. A file whose branch that anchor resolves is
never queued (from the inspection profile it offers no choice; a
default-ambiguous file asks saver_anchored) and its pending, undecided
items leave on the next scan. REVIEW_RULE_VERSION is folded into
inspection item versions so old items are re-evaluated. No match, several
matches or an unfinished walk anchor nothing. Apply refuses a choice the
anchor contradicts.
* fix(review): refuse a default choice the saver anchor contradicts (#390)
Apply re-resolves a default item with the file's name: when its saver
anchors the branch, no candidate is offered and the choice is refused.
The scan's anchor check skips bytes carrying a decision and remembers
unanchored file states per process, so a rescan does not inspect an
unchanged file again.
* fix(review): queue an anchored file with its saver's pair as the choice (#390)
Maintainer ruling: a file never falls through both the link and the
queue. When only the saver anchor resolves a file the default profile and
the parser floor leave unlinked, the scan queues it with the anchor's pair
as its one candidate (evidence anchor: saver), never linking it. Files
already linked or resolved by the default path are unchanged. Apply
accepts the anchored pair when the bytes now resolve to it, and an
anchored item is a choice the untraced settle keeps.
* fix(review): upgrade old default items of anchored files, tag anchored results (#390)
A pending default item recorded before the saver anchor no longer
shortcuts the rescan when only the anchor resolves the file: the
anchored item supersedes it. Apply accepts a resolved re-resolution only
when it is tagged as the saver anchor's, the anchored pair must carry a
positive text to choose, and limited anchored evidence still counts as a
choice. TESTING.md describes the ruling.
* fix(images): read a loading collection as loading, never as 0 of 0 images (#391)
A reset published {loading, loaded:0, total:0} and the caption painted it
as "0 of 0 images" until the first page answered, so a fast reader (the
inspector concept spec on WebKit) read an empty result the server never
returned. The count line now says "Loading images…" while a first page is
in flight; the spec waits for the answered count, and a held-listing spec
pins the ordering.
* test(thumbnails): make the no-JPEG-XL-encoder fixture authoritative (#371)
{available:false} stood in only the cached detection, and every settings
read probes a cached None again, finding the host's real encoder. The
fixture now stands in for the detector itself and reset restores it; a
unit test proves the route against a detector that finds an encoder.
* test(system): wait for the restore job's own end before its toast (#361)
The reviewed-restore specs waited for the "Database restore finished"
toast with the 5 s expect, in effect a deadline on the disk-bound publish.
They now wait for the job's progress stream to close, assert the job
finished, and only then check the toast with the default timeout.
* test(images): wait for the search's answered count in the concept comparison (#391)
searchOf took the first changed caption as the search's count, the same
read-before-the-response race as the inspector's tag chip.
* ci: pin black and flake8, let the bandit hook fail, cache npm in the JS job (#352)
The lint job installs black==26.10.1 and flake8==7.4.1 from
requirements-dev.txt (the versions CI resolved on main today), so a new
release cannot turn CI red with no code change.
The local bandit hook ran 'bash -c ... "$@"; true': it could never
fail, and bash bound the first staged file to $0 so it was never
scanned. It now runs 'bandit -lll -q' directly, failing on high
severity exactly like code-quality.yml.
The frontend modules job's setup-node caches npm keyed on
tools/vendor/package-lock.json.
* test(history): wait on the writer's state, not short wall-clock pauses (#352)
Four 0.08 s pauses before asserting that the worker had not probed yet
now wait until the idle worker has run a whole loop pass at the current
manual clock value (three clock reads), so a slow runner cannot end the
pause before the worker decides. A mutation that probes 20 ms early now
fails two tests deterministically.
The real-clock bounded close no longer asserts close(0.02) returned in
under 0.2 s wall time; it asserts the in-flight write is still blocked
and the worker alive, like the output writer test. The manual-clock
bound stays exact. The last 0.1 s non-blocking submit check uses the
1 s NON_BLOCKING_CEILING from #351.
* test(api): give each API fixture database its own temp folder (#346)
Every PromptManagerAPI sweeps its database's folder at startup for the
stages, uploads and safety-backup copies an interrupted restore left.
The API test fixtures created their database directly in the system
temp folder, so any other test process starting an API there (another
worktree's suite, a pre-push hook, an agent's targeted run) deleted
this process's in-flight restore files. The restore, epoch and
retirement tests then failed with 'staged restore changed', 'unable to
open database file' or 503 instead of 409. Busy machines fail because
they are running other suites, not because they are slow: the same
CPU load with a private TMPDIR passes.
private_temp_db() puts the fixture database in a new folder removed at
cleanup (with any safety backups a restore left beside it). A
regression test sweeps a stand-in shared folder in the middle of a
multipart restore, as another suite's startup would.
* test(history): count worker parks, hold saturation blocks until released (#352)
wait_for_worker_pass now counts the idle worker's parks on its wake
event (two parks from now = one whole pass begun after the call)
instead of assuming a fixed number of clock reads per pass, which a
future clock read would have made vacuous. The real-clock bounded
close keeps a loose upper bound (0.02 s + NON_BLOCKING_CEILING).
The two saturation tests (history and output writer) blocked the
worker with release.wait(3) while they queued 1023 more items. On a
loaded runner the fill took longer than 3 s, the worker unblocked on
its own and drained part of the queue (pending 868..1017 != 1024),
and cleanup then reported a leaked worker. The block now holds until
the test releases it; 60 s is only a hang guard, and cleanup always
sets the release. Under one core with 8 busy loops: 8/8 runs failed
before, 6/6 pass after.
* test(maintenance): wait on duplicate jobs with a disk-bound hang guard (#409)
On Windows py3.13 CI (PR #405) a one-file duplicate cleanup missed its
10 s WAIT. The test already waits on the job's own task; the 10 s bound
acted as a deadline on disk-bound work. In the same 50 s window the
following tests' setups took up to 9 s against 0.1-0.6 s around it
while the test bodies stayed fast, which is consistent with a
process-wide disk I/O stall on the runner rather than a lease, handle,
executor or job stall (the log has no thread dump to show more).
tests/disk_waits.py holds the one disk-bound hang guard (90 s per wait,
under the 120 s per-test watchdog), shared with RESTORE_SECONDS (#382).
Waits on find and cleanup job tasks in the duplicate-files tests use
it; gates and server-state polls keep WAIT. The keeper-validation test
turns the safety backup off, since it does not test backups.
* test(legacy): snapshot the routes and the settings response
Pins every registered route and the key paths of GET /prompt_manager/settings
while System S13 (#330) moves the legacy pages' tools to native System. Recorded
on the unchanged main; cutover (#334) removes it.
* test(legacy): prove each removed legacy control has native coverage
TESTING.md gains the "Legacy System UI removed by S13" table (no rows yet).
legacy-coverage.test.mjs fails when a removal cites a capability that is not
covered, copies another native test, lists an id or method twice, or leaves
the id or method on its page. legacy-system-links.spec.mjs holds the System
link table and guards the kept admin flows.
* refactor(admin): open System for Diagnostics and logs
The admin page's Diagnostics and View Logs buttons become links that open
System › Diagnostics & logs in a new tab. The legacy Diagnostics and System Logs
dialogs and their methods go; TESTING.md lists them against their covered
capabilities (rows B1–B7).
* test(legacy): find a removed modal id passed to showModal and settle the links spec
The removal check also finds an id passed as a string to a ...Modal() call and no
longer reads data-id= as id=. The no-request test waits for the page's own
start-up reads to finish, the sort check compares as the server sorts, and the
kept-flow test checks the added prompt shows in the list.
* fix(restore): a restore cut off by shutdown waits for its worker thread (#360)
A cancelled restore handler or job no longer removes the stage or the
upload, or ends its exclusive registry admission, while the executor
thread still stages or publishes. The cancel waits for the worker; a
publish that finished is adopted before the cancel goes on, and a stage
finished after the cancel is removed.
* fix(threads): a thread or pool refused at interpreter exit is quiet and leaves nothing pending (#375)
During interpreter exit, threading._shutdown shuts every thread pool
before the package teardown runs while our daemon threads keep working.
map_parallel now runs the items a refused pool did not start on the
calling thread, schedule() drops the path and returns False when its
worker cannot start, and the preview queue skips the file. Each
refusal is logged once at DEBUG without a traceback; other start
failures behave as before.
* fix(restore): keep the cancel and the adoption when a stopped restore's publish ends (#360)
A publish that fails after the cancel now ends in CancelledError (after
its prune) instead of a 500, and the restored library is adopted to its
end even when a second cancel arrives. The blocking calls go through
_run_in_executor again, as the executor contract test expects.
Exit-window refusals: the comments state what was measured (a pool is
refused during threading._shutdown; Thread.start only once the
interpreter is finalizing), and the tests assert the one DEBUG line.
Review fixes: a 409 source-changed reload moves focus to the heading as a
success does; a selection the listing no longer holds says none could be
found instead of sending nothing; a Tags & categories Browse opens a
plain Images visit filtered by the category (remembered and with the
Appearance defaults) rather than a saved-view visit.
D140 (Set category writes images' own categories), D146 (the chip's visible
origin) beside D141, D142 and D145; D143 and D144 in the vocabulary
section. D8, D21 (the D20/D21 side panel) and D39 amended; the D8, the
"all N matching" and the Native vocabulary UI paragraphs name the Prompt
library editor and the legacy admin "Set Category" for prompt
categories; the #332 section lists the interface and its comparisons.
The vocabulary fixture seeds own image categories on request
(image_categories: Landscapes on two images, Portraits on one image only).
mapVocabularyConcept maps D143 ("N positive prompts · M images", image-only
rows before Manage category), a model checked against the concept page with
those rows added at runtime, and mapVocabularyDialog takes the preview
sentence, so the Manage category preview is compared with D9 and D144's
images. The default fixture and mapping are unchanged.
Replace the #259 prompt-category tests with CUJ 1, 2, 3, 8 and 10, the
suggestions datalist, the 500-image limit, the source-changed path, D142
and coded refusals; Images never posts /prompt_manager/bulk/category.
D145: with an own category set on the baseline's newest output, the
Recorded run viewer's chip is compared with the concept viewer's category
chip laid out with the D146 origin and the own category at runtime; D21
maps its hidden "Image category: " prefix. Unmapped, the two differ only
in the text, the prefix, the width and span.chip-source.
With one image's own category set through the real route, the toolbar's
#category-filter options are compared with D141 mapped (the image category
sorted in with the Library's, D36), and the Set category dialog against the
concept's simpleInput laid out as recorded at runtime: D140 (empty field,
datalist#workspace-input-suggestions of the toolbar's categories, Use prompt
category between Cancel and Apply) and D8 (Apply). Unmapped, the dialog
differs only in those parts.
D146: the inspector comparison measures each chip's span.chip-source and
lays the concept's own chip out with the recorded rule and origin added at
runtime (withChipSource), in the side panel and the viewer's details. The
D39 recipe case asserts the visible " · prompt"; the mutation check
removes the origin.
* fix(system): Tagging selects show their whole chosen option at 390 px (#398)
The concept's phone cap (120px) cut Existing tags ("Replace after review").
#tag-existing now never shrinks for its row text and keeps its natural width
at phone width, beside the text. The comparison expects selectsFit for every
select, and a new test walks every option of both selects at 390 and 320 px
in both themes. TESTING.md D132 records the decision.
* fix(system): bound Existing tags by its row at phone width (#398 review)
The selection bar's Set category writes the selected images' own
category (linked or not, up to 500, or all N matching up to 500), with
the toolbar categories as suggestions; Use prompt category removes it
after a confirmation that states the count. It never writes a prompt
(#259's prompt action is retired from Images; the Library editor and
the legacy admin Set Category still change prompt categories). The
toolbar lists the listed images' own categories beside the prompt and
recipe ones. The inspector chip shows its origin: "· this image",
"· prompt" or "· recipe". Tags & categories counts images and opens
Images for a category only images use.
Review fixes: on Windows the own-category map keys each file with
normalize_image_path (abspath drops trailing dots and spaces), and the
plan test also covers the category's narrowing clause.
The category routes also accept the Images listing filter of "Select
all N matching" (#392) instead of explicit items, resolved now and
capped at 500 images like an explicit selection.
Under a category the link read keeps only links whose raw recipe or
positive category is the value (Python still decides the inherited
one), the links of images whose own category is the value are read by
path, and plain reads skip the inherited-category work. The own-category
map is keyed by normalized path. TESTING.md records D141, D142 and D145
and amends D13, D36 and D39; the filter budget seeds 2,000 own
categories.
The inspector's library facts gain category_scope "image": an image's
own category overrides the inherited one for any link status (linked,
ambiguous, none). Images reads it for output sources only; History reads
it once per inspection inside its snapshot, keyed by the same file key
from the symlink-resolved output folder. A failed read hides the chip
rather than show a possibly wrong value. The chip's hidden prefix names
the origin ("Image category: "); the visible origin text comes in PR 3.
Review fixes: a source matches by its configured spelling too, so a
detected output folder that is a symlink keys its files (routes and
duplicate cleanup). File keys are computed in the executor. A path no
source holds is a plain 403, a lone surrogate a 400. The install repairs
a table that lacks its category index instead of refusing to open.
An image's own category wins; without one it inherits its link's: the
recipe's category when set, else the positive prompt's (MC4, one rule
with the inspector chip). Ambiguous and unlinked images inherit nothing,
so an unlinked image with its own category is now listed. The link read
no longer filters by category in SQL; the page filter, select all N
matching and saved views compare the effective value. A failed
own-category read answers the existing listing-failed 500.
Tags & categories lists every category value with image_count, the
stored image_categories rows holding it. Values only images hold follow
every prompt-backed value (first_pid at SQLite's INTEGER maximum), so
keyset cursors cross the boundary. /library/options keeps its own query
and never lists image-only values.
Manage category counts image rows (impact.images), treats an image-held
target as existing, renames image rows in the same transaction, allows
renaming an image-only value, and hashes the sorted image rowids into
the impact token so a change between preview and apply is refused with
vocabulary-conflict. Every image-category statement reads
idx_image_categories_category.
PUT /prompt_manager/images/category sets the own category of up to 500
listed files, DELETE removes it (listed or not), and
GET /prompt_manager/images/categories lists the values whose file is
listed. Targets are the Images listing identity of an output source,
checked by media_target (shared with Collections). The routes never
touch prompts, recipes or links; POST /prompt_manager/bulk/category is
unchanged.
The kept copy's own category wins; a kept copy without one receives the
removed copy's, keyed by the deepest output source holding each file.
The removed copy's row is never deleted, so a give-back or a failed
delete loses nothing. The job log reports how many were kept.
One new table, image_categories, and its category index, created empty
by a forward-only migration that reads and writes only its own objects.
Startup checks the structure; restore also checks the values. The store
sets, clears and reads own categories by file key (the deepest output
source holding the file and the path below it), and copies a removed
duplicate's category to a kept copy that has none.
* fix(scan): read sampler-less graphs from the one wired text, retire the heuristic (#252)
A graph where no node takes a positive or negative input fell back to the
pre-redesign _extract_readable_prompt heuristic: the last node with any string
text input won, and with none it read stale UI workflow widgets. It stored LoRA
syntax, display-node reports, install notes, model file names and numbers.
The scan now asks sampler_less_prompt: only the executed API graph counts; a
candidate is a node with a text input that another node consumes; blank and
numeric texts are never prompts; this repository's own positive nodes are
authoritative; one distinct text links, several (or an unreadable wired text
that no candidate builds) stay unlinked and try the recorded-run attribution.
The heuristic and its helpers are removed; A1111 files read the parsed
positive prompt directly.
* fix(scan): refuse loader side outputs, seeded templates and keep A1111 text (#252)
Review findings: an unreadable text link counts as built from a candidate
only through that node's first output, so a LoRA loader's trigger-word output
is unknown, not its text. A node with a seed input holds a template, never the
prompt that ran. A sampler-less graph the rule cannot decide still links the
A1111 positive a metadata saver wrote beside it, as the old scan did. Only
plain numbers are numeric; words such as nan or inf are prompts.
* fix(scan): linear plain-number check for sampler-less texts (#252)
The numeric pattern had overlapping quantifiers that backtrack quadratically
on a long digit run ending in a non-digit; it now has none.
* fix(tagging): Tag output images tags each image, not its prompt (#379, PR 1)
The native "Tag output images" job now runs the image-tags loop over every
still image the output folders list, linked or not: each image gets its own
model tags (at most "Suggestions per image") in its image tag set, and its
bytes are marked auto-tagged in a new media_image_autotags table. No prompt
tag or prompt text is written.
- database/autotag_schema.py: one forward-only WITHOUT ROWID table, optional
in old backups, refused without the annotation group.
- add_image_tags/replace_image_tags take mark_autotagged; autotag_state and
autotagged_locators read the marks through the one current-record rule.
- "Skip already tagged" skips images whose current bytes are marked, counted
in a stat-only pre-pass; hand tags do not count.
- The review counts images and estimates the time (WD14 on CPU or CUDA).
- System > Auto-tagging review rows and copy, D134 and the Auto-tagging
contract follow.
Part of #379
* fix(tagging): end the batch job after its cleanup and report pre-pass progress (#379)
The job read as ended while it still held the model gate, so a quick
restart could be refused with model-busy. Pre-pass chunks now publish
progress, and a cancel between chunks keeps exact counts.
Part of #379
* fix(tagging): measured GPU rate, and an ended image-tags job shows no time left (#379)
The live check measured WD14 on CUDA at 0.167 s per image (100 images in
16.7 s with the model load), so the review estimate uses 0.17 s. A finished
or cancelled Auto-tag/Add tags job kept "Processing file N of N · about X
min left"; it now ends with "<verb> <state> · N files processed".
Part of #379
* fix(tagging): legacy pages tag each image, not its prompt (#379, PR 2)
The admin and gallery batch starts the shared Tag output images job
("Tag output images (legacy page)") and relays its progress; closing the
window cancels it. Review asks per image with Skip first and applies by
image and sha256; apply refuses a prompt_id body. The Tag image dialog
marks an image auto-tagged only after Generate suggestions on that
capture. A second legacy batch (also one that wins the start race)
answers JSON 409, and both pages now show that sentence.
* test(tagging): gallery review units apply by image; record each progress change
The gallery review unit tests still pinned the removed prompt_id apply.
The legacy batch spec read the progress text once per observer callback,
so a phase frame and the first image frame landing in one task lost the
phase (WebKit); it now records every mutation record.
The workspace Images page selects like a file manager in Grid, List and
Masonry, across every loaded batch: Shift+click ranges from the anchor,
Ctrl/Cmd+click toggles a card without opening it, Ctrl/Cmd+A selects the
loaded images, Shift+arrows extend, Escape clears, and a touch long-press
starts a selection mode in which taps toggle. A plain click still opens or
inspects.
Once every loaded image is checked and more match, the bar offers
"Select all N matching". Add tags, Auto-tag selected and Set category then
send the listing filter instead of N identities: the image-tags job takes
`filter` in place of `targets` and /bulk/category takes `images_filter`,
both resolved by the listing's own admission code (up to 10,000 images, a
Source bound by its root_scope). Export metadata waits for loaded images.
Recorded as deviation D135; the concept is unchanged.
* fix(tagging): Images tag search goes through the tag index (#379, PR 3)
The Images Tags filter and text search no longer stream every image tag
row per request. image_tag_ids reads the names that tag at least one image
(exact under Tags, case-folded substring for q) and image_tag_locators
reads their files by tag_id through idx_media_image_tags_tag. Results are
unchanged: a differential test compares the lookup with the old stream,
and a trace test checks that no statement reads media_image_tags except
by tag id. The search contract in TESTING.md and a browser test (CUJ 8)
follow.
Closes#379
* test(search): pin prompt-only tags and multi-chunk tag lookups (#379)
Reads the tag names and the locators in the same annotation read helper.
* fix(tagging): Tag output images tags each image, not its prompt (#379, PR 1)
The native "Tag output images" job now runs the image-tags loop over every
still image the output folders list, linked or not: each image gets its own
model tags (at most "Suggestions per image") in its image tag set, and its
bytes are marked auto-tagged in a new media_image_autotags table. No prompt
tag or prompt text is written.
- database/autotag_schema.py: one forward-only WITHOUT ROWID table, optional
in old backups, refused without the annotation group.
- add_image_tags/replace_image_tags take mark_autotagged; autotag_state and
autotagged_locators read the marks through the one current-record rule.
- "Skip already tagged" skips images whose current bytes are marked, counted
in a stat-only pre-pass; hand tags do not count.
- The review counts images and estimates the time (WD14 on CPU or CUDA).
- System > Auto-tagging review rows and copy, D134 and the Auto-tagging
contract follow.
Part of #379
* fix(tagging): end the batch job after its cleanup and report pre-pass progress (#379)
The job read as ended while it still held the model gate, so a quick
restart could be refused with model-busy. Pre-pass chunks now publish
progress, and a cancel between chunks keeps exact counts.
Part of #379
* fix(tagging): measured GPU rate, and an ended image-tags job shows no time left (#379)
The live check measured WD14 on CUDA at 0.167 s per image (100 images in
16.7 s with the model load), so the review estimate uses 0.17 s. A finished
or cancelled Auto-tag/Add tags job kept "Processing file N of N · about X
min left"; it now ends with "<verb> <state> · N files processed".
Part of #379
* test(browser): reload the admin page only after the in-place edit saved
The gallery-to-admin workflow pressed Enter and reloaded as soon as the
edited text showed; the page saves after showing it, so a reload that won
the race abandoned the PUT (WebKit, run 40564511000). The test now waits
for the save's answer.
Part of #379
* test(tagging): the batch coverage row names the renamed #379 test
* fix(logs): declare the compressed length so a log download finishes
The gzip middleware compressed a text body but kept the handler's declared
Content-Length (the uncompressed size), so a browser waited for bytes that
never came: downloading any log over 1 KB hung, in the workspace and on the
admin page. The middleware now drops a declared length before compressing.
Part of #329
* feat(system): show Diagnostics & logs natively and retire the hand-off panel
System › Diagnostics & logs is native (D109, D110): the concept's
Diagnostics surface with the live database, source, Viewer.js and renderer
lines, Run checks (the shared Health report), the diagnostic report, Inspect
and the confirmed test link; and the Logs surface with the minimum-level
viewer, Lines 1-500, statistics, the log download, the runtime level and
console setting saved by Save preferences, the log files with their
downloads, Review log cleanup (the rotated_logs maintenance plan, no backup
box, Listed files) and a confirmed Clear all logs.
D94's hand-off panel, SECTION_HANDOFFS and the undelivered runtime adapter
are gone. The browser server logs into its own temporary root.
Part of #329
* test(system): compare Diagnostics & logs with the concept
System › Diagnostics & logs and its log cleanup review are compared with the
concept (app.js 694–695 and maintenancePlan("rotated logs")) in both themes at
1600x1100 and 390x844, with D109 and D110 (and D100 for the review) as the only
mapped differences, a 320-wide select check and a mutation check.
The fixture gains a concept seed mode (the concept's four log lines, a 1.4 MB
active log with three rotated copies, console logging off), and the plan
dialog's measureReview and mapPlanConcept move unchanged into
concept-inventory.mjs for Maintenance and this review.
The comparison found that the Database line read a report key that does not
exist (database_info.size_bytes) and so always showed "0 bytes"; it now reads
database_size_bytes, the key get_database_info() returns.
Part of #329
* test(system): prove every legacy System capability has an owner
TESTING.md gains "Legacy System capability coverage": every System capability
of the legacy admin and gallery pages with its anchor on main, owner, status
(covered, kept in legacy or approved removal), native control and the exact
title of the test that proves it.
tests/vendor/legacy-coverage.test.mjs (with tools/vendor/legacy-coverage.mjs)
fails on a row with no owner, an unknown status, a malformed native control, a
control on a row that is not covered, or a test title its file does not
declare; a title only in a comment does not count. workspace-system.spec's
"every covered legacy System capability has its native control" opens each
covered row's section and finds the control by role and accessible name.
Versus the design's table: the anchors are re-checked against main, "Run
health checks" is on Maintenance (D123), not Overview, Diagnostics run is
proven by "Run checks opens the shared Health report", and the export rows
repeat their test title instead of "(same as above)".
Part of #329
* fix(system): confirm Clear all logs on a fresh file list and record coverage
Review fixes for the Diagnostics & logs section:
- Clear all logs reads the log files again before it confirms, so the
dialog names the copies it deletes now (a list it cannot read gives the
general sentence, never an old count); Refresh reads the files too.
- Runtime values that cannot be read stay disabled and say so.
- Create test link holds its button until the write answered, so a second
press cannot write a second record.
- The Lines field names its range.
- The browser server builds its logger in its own root, so no log file is
ever opened in another folder; the fixture closes every directory scan
and replaces a link instead of writing through it.
- system.mjs drops the unused section fallback; frontend coverage floors
recorded with coverage:update (system-diagnostics.mjs 97.5% lines).
Part of #329
* fix(system): read the database size the report sends
The Database line read database_info.size_bytes, a key the report does
not have (get_database_info answers database_size_bytes), so it always
read "Readable · 0 bytes · N images". The browser test now requires a
non-zero size. Found by PR 3's concept comparison.
Part of #329
* test(system): require a running test and a control for every covered capability
Review fixes for the legacy coverage proof:
- A title declared only with test.skip, .fixme or .fail no longer proves a
row (.only and .slow still do).
- Every covered row names its native control, except the rows in
CONTROL_EXEMPTIONS, each with its reason (today only the gallery per-image
auto-tag row, owned by the Images selection bar). The browser test asserts
that the covered rows without a control are exactly the exemptions.
- JavaScript comments are stripped across the whole source, wherever a block
comment starts, skipping strings, template literals and regular
expressions. A Python def test_ counts only at function or method
indentation and outside comments and docstrings.
- A control name several rows share ("Review plan", "Review options",
"Refresh") is looked up inside the row the table names (· in "<row>").
- The diagnostics comparison records each select's chosen option, so a wrong
default level is caught.
Part of #329
* fix(logs): declare the compressed length so a log download finishes
The gzip middleware compressed a text body but kept the handler's declared
Content-Length (the uncompressed size), so a browser waited for bytes that
never came: downloading any log over 1 KB hung, in the workspace and on the
admin page. The middleware now drops a declared length before compressing.
Part of #329
* feat(system): show Diagnostics & logs natively and retire the hand-off panel
System › Diagnostics & logs is native (D109, D110): the concept's
Diagnostics surface with the live database, source, Viewer.js and renderer
lines, Run checks (the shared Health report), the diagnostic report, Inspect
and the confirmed test link; and the Logs surface with the minimum-level
viewer, Lines 1-500, statistics, the log download, the runtime level and
console setting saved by Save preferences, the log files with their
downloads, Review log cleanup (the rotated_logs maintenance plan, no backup
box, Listed files) and a confirmed Clear all logs.
D94's hand-off panel, SECTION_HANDOFFS and the undelivered runtime adapter
are gone. The browser server logs into its own temporary root.
Part of #329
* fix(system): confirm Clear all logs on a fresh file list and record coverage
Review fixes for the Diagnostics & logs section:
- Clear all logs reads the log files again before it confirms, so the
dialog names the copies it deletes now (a list it cannot read gives the
general sentence, never an old count); Refresh reads the files too.
- Runtime values that cannot be read stay disabled and say so.
- Create test link holds its button until the write answered, so a second
press cannot write a second record.
- The Lines field names its range.
- The browser server builds its logger in its own root, so no log file is
ever opened in another folder; the fixture closes every directory scan
and replaces a link instead of writing through it.
- system.mjs drops the unused section fallback; frontend coverage floors
recorded with coverage:update (system-diagnostics.mjs 97.5% lines).
Part of #329
* fix(system): read the database size the report sends
The Database line read database_info.size_bytes, a key the report does
not have (get_database_info answers database_size_bytes), so it always
read "Readable · 0 bytes · N images". The browser test now requires a
non-zero size. Found by PR 3's concept comparison.
Part of #329
* test(browser): seed resets on a scratch copy and publish them in one commit (#384)
The concept and plain resets made hundreds of committed transactions on the
live database; on a slow CI disk they queued behind WAL checkpoint syncs and
took 30+ s. Each reset now copies the live database into a scratch database
(backup API, no syncs), runs the unchanged seeder against it through an API
proxy, and replaces the rows of the tables the seed changed in one live
transaction. A commit by another live connection meanwhile fails the reset.
tests/browser/reset_equivalence.py proves every reset kind leaves the same
rows, files and response as the old path (PM_TEST_SEED_ON_LIVE=1).
* test(browser): empty media folders by a rename, delete them in the background (#384)
With the database commits gone, throttled resets still stalled in rmtree:
unlinking the last fixture's files waits for their pending writeback. Each
reseed now moves the old output and input folders into <root>/reset-trash
and makes empty ones; a background thread deletes the trash, and the
server's cleanup stops it and removes the rest before the temp root goes.
* fix(test): recheck the live data version under the publish's write lock (#384)
A writer whose transaction was open while the seed ran committed before
BEGIN IMMEDIATE returned and its rows were overwritten. The publish now
rechecks data_version under the lock, and restores foreign_keys and
detaches the scratch database on every error path.
* test(browser): keep the reset scratch database in /dev/shm when there is one (#384)
Throttled runs still waited on the runner's disk while creating and
rewriting the scratch database, whose pages were under writeback.
* test(browser): keep the scratch on disk when /dev/shm is short, delete in place after cleanup (#384)
The largest scratch copy is about 8 MB; Docker's default /dev/shm is 64 MB
and shared with the browsers. A reset that empties a folder after the
server's cleanup closed its trash deletes in place instead of starting a
thread nobody stops.
* fix(test): drop the gallery caches only after the reset's rows are published (#384)
The seeders drop the Images listing at their end. Run on the scratch copy,
that drop came before the publish, so a background walk in between could
cache per-file values built from the live database's old rows: CI's
inspector tag chip found 0 of 0 images. The proxy now records the drop
and the seeder makes it after the publish, as the old reset did.
* fix(test): remove /dev/shm reset scratch left by killed browser servers (#384)
A server killed outright never runs its cleanup, and tmpfs keeps the
scratch folder in memory until reboot: the Python suite's server tests
left one per killed server. Scratch folders now carry the owner's pid,
and each new one first removes those whose process is gone (POSIX only).
* fix(inspector): pass text through unknown nodes instead of refusing them (#381)
The inspector refused any node class missing from the semantics table, so a
PromptManagerPair behind an rgthree Any Switch showed "Prompt uses a node the
inspector does not support". The inspection resolver (Python and browser) now
reads through unknown nodes: a text input is the text, positive/negative route
by output slot, a conditioning input is followed alone, otherwise every linked
input except core type inputs (pass_through.skip_inputs). PromptManager nodes
are authoritative; several different texts become one candidate each
("Several prompts reach the sampler"). A seeded text node with no conditioning
input is generated at run time, and a recorded PromptManager execution is not
checked against text that passed through an unknown node.
INSPECTION_RESOLVER_VERSION goes to 2; ingestion keeps the default profile.
A legacy parity gate records every committed graph whose inspection positive
matches the legacy parser's; no graph may leave it.
* fix(inspector): keep recorded pair text, string-only seed rule, role inputs and unique candidates (#381)
Review of the pass-through: a walk through an unknown node never overrides a
PromptManagerPair's execution record (any outcome but work-limit is
unverifiable while the record is checked); the run-time seed rule reads only
string outputs; a node with positive/negative inputs follows only the role of
the requested slot, nothing when it is unlinked; one text offered twice is one
candidate.
* fix(inspector): anchor on the saver that wrote the file and resolve stages on shared links (#381)
An inspected file's name now anchors the inspection on the one saver whose
filename_prefix (literal or traced) makes the file's stem <prefix>_<digits>[_];
the samplers in that saver's image ancestry are the candidates. No match,
several matches or an unreadable walk anchor nothing, so the file stays
"several possible prompts". In a latent walk, an earlier sampler reading the
same conditioning links as the later one is the same prompt, not an
unattributed stage. Python and the browser resolver behave the same; the
legacy parity gate states each graph that differs from legacy on purpose.
* fix(inspector): skip only plain first stages on shared links; an unfinished prefix scan anchors nothing (#381)
Review of the saver anchor: a same-links earlier stage is skipped only when it
is not a detailer, sets no prompt override and continues no other sampler, so
an earlier prompt or override behind it is never hidden. A work limit while
tracing a saver prefix ends the match scan without an anchor instead of
counting that saver as no match.
* test(inspector): cover every plain-first-stage outcome on both resolvers (#381)
The frontend coverage gate found local-resolver-stages.mjs branches at 92.44%
(< 93). Each way an earlier stage stops being a plain first stage (no latent,
literal, missing or cyclic latent, a sampler-shaped node, a detailer) is now a
resolver test and an anchor-first-stage-* parity vector.
* ci: run the required workflows for the merge queue
main no longer requires a branch to be up to date before it merges, so a
PR no longer re-runs CI after every other merge. The Tests, Code Quality
and Viewer.js workflows also run on merge_group, so the merge queue can
test each PR against the latest main once, in order. The tracked
protection record now says strict: false.
* test(history): give waits that span a database restore their own ceiling
The merge queue's Windows py3.12 job failed
test_restore_or_generation_rotation_between_transactions_discards_evidence
(restore=True): the restore it waits for, a full verify/migrate/backup/
replace cycle, took over the 10 s BLOCK_SECONDS on a busy runner. The two
waits that span a restore now use RESTORE_SECONDS (90 s), which stays under
the suite's 120 s hang watchdog, so a real hang still fails with its thread
dump. Behavioral waits keep BLOCK_SECONDS.
* test(history): drain the writer before every bounded close or dispose
The merge queue's Windows py3.12 job failed
test_saturated_start_cannot_capture_outputs_or_retain_previous_roots:
dispose(2) ran while the writer still worked through 1,024 queued
commands on a busy runner. Like #370, the test now drains first; the 2 s
bound is not what it tests. Four more closes that follow a released
block get the same drain, so the next slow runner can't fail them one at
a time.
* ci(comfyui): run the real-host job in the pinned Playwright image
The real ComfyUI compatibility job installed Chromium with
`playwright install --with-deps`, which pulls the browser's system
packages from Ubuntu's apt mirror on every run. Three runs (#349 run
40367579020, #378, #382 job 122287230985: "Fetched 19.7 MB in 12min 23s
(26.5 kB/s)") spent 11-12 minutes there and were cancelled at the
24-minute limit after all 89 tests had passed.
The job now runs inside mcr.microsoft.com/playwright:v1.64.1-noble, the
same image as the frontend coverage and page workflow jobs, which ships
Chromium and its libraries, so the job has no apt step at all. The
whole job moves rather than only the browser part: the Playwright
suites launch the managed ComfyUI host themselves (PM_TEST_PYTHON), so
splitting the browser off would mean running the host on the runner and
the browser in a container across a network boundary, a larger change
than moving one job whose host is plain Python and pip wheels.
Python follows the other image jobs: the image's own Python 3.12 with a
user-space venv bootstrapped from a checksum-pinned pip wheel (toolcache
setup-python builds do not run in the container). A step fails the job
when the image Python differs from pythonVersion in
tools/vendor/comfyui-runtime.json, so the pinned host stays honest;
torch, torchvision and ComfyUI's requirements install from the same
pins as before. pip and npm download caches replace setup-python's and
setup-node's.
tests/vendor/browser-image.test.mjs now checks this job like the other
browser jobs: image tag equals @playwright/test in the lockfile, no
browser install or apt, no setup-python/setup-node, checksum-pinned pip,
Node 24, and the runtime Python check. The 24-minute limit is unchanged.
Vito Sansevero