--paired contrasts two arms running the same seeds seed by seed with the wilcoxon signed-rank test rather than treating them as two independent samples, reporting the per-seed differences, the sign, a12 within pairs and the seeds usable in one arm only. --question names the family holm corrected within, and the family size is recorded next to the p-values rather than left to the reader to reconstruct.
an obligation that never discharges is reported when the run ends, and timing it by the step that armed it recorded a liveness failure flushed at the budget as a violation found on step 1. the survival analysis now measures the detected step, falling back to the origin for campaigns written before the field existed, and says how many events that moved. the report gains the first and third quartiles beside the median.
the campaign tool and the sweep tools that drive it have to read a seed specification the same way, or a sweep records an intent that differs from what ran. parseSeeds becomes seedspec.Parse with no behaviour change.
reads a run directory's meta and every step, and refuses a step whose trace_version is not the current one: an older step stores no element depths, so its hierarchy decodes with a nil root and a structural hash over it is the empty string for every screen. Discover walks a tree for the directories holding both meta.json and trace.jsonl.
BundleSpec produces the goja bundle a run of a spec loaded, seeded as that run was. an offline replay has to load the same javascript, and the seed is one of the bundle's defines, so it is part of the bundle's identity.
an offline replay of a trace needs the name-to-index mapping the spec fixed at load, because a trace records extractor values by name, and needs each property's formula built over this verifier's own predicates so a rewritten formula observes exactly what the engine's evaluator does.
observation and apply now run under a timeout, so a driver that stops answering ends the step rather than the run. an undelivered gesture and a selector that matched nothing are recorded as their own skip reasons instead of counting toward the apply-failure streak, a failed observation is counted apart from a screen with nothing on it, and the summary names both. resolveCoordinates hands a point outside the viewport to the driver rather than dropping it: only the driver knows whether it can scroll that point back into reach. exceptions and navigations are collected per step and a Scroll goes to a driver's Scroller when it has one.
a step now carries trace_version, the platform log lines and uncaught errors behind state.logs and state.exceptions, the document-replacing navigations seen since the previous step, and observation_error naming why a device read produced no tree. version 0 is a step written before those fields existed, which is what separates a trace that cannot answer the question from a step that had nothing to report.
a page navigation replaces the runtime, so the seeded picker restarted the seed's stream at its first draw on every reload and a trace could not tell a reload from a generator repeating itself. the driver now drains the main-frame navigations it saw, reports the page's buffered uncaught errors so state.exceptions is the page's list on the goja host too, and carries the picker's draw position out of v8 and back in around each decision.
getBoundingClientRect keeps reporting elements the growing document pushed below the emulated viewport, and input coordinates are viewport-relative, so a click below the fold was hit-tested to the document root and the step read as an action that landed. every gesture now scrolls the point back in and reports ErrGestureUndelivered when nothing is under it; a selector that names no node reports ErrSelectorMatchedNothing rather than waiting. swipe dispatches a real touch stream instead of page-synthesized pointer events, scroll is a wheel so its distance is exact rather than a fling, and the second tap of a double tap carries click count 2 so dblclick actually fires.
the dump emitted a fixed standard attribute set, so a spec reading data-cents or data-account-id saw undefined on the goja host and nothing at all in the trace. it now keys every attribute by the name the markup writes, derived keys overwriting. checked and selected come from the dom property rather than whatever a component left on the object, which is also what the page-side element handle now reports, so a ticked box reads as ticked instead of reporting its starting state forever.
a Tree marshalled to json kept only the flat element array, so a stored tree decoded with a nil Root and resolved no selector. it now stores each element's pre-order depth and rebuilds Root from it, re-seating elements so Tree.Elements and &node.Element stay one pointer. a stored tree without depths keeps the old shape. a boolean field the producer sent as something other than a boolean now leaves the flag unset and increments UnreadableFlags rather than failing the whole dump.
an element's text is its whole subtree's text on web and on ios, so every ancestor of a matching element matched too, up to the root. a match a descendant also makes is now dropped, in internal/hierarchy, in the chrome xpath translation and in the page-side web runtime, so all three resolvers name the same element. a raw attribute now matches on a substring (exact for true/false) the way the docs describe, and tree-level FindBySelector considers the root, so ax.find("id:page") and ax.find({id: "page"}) agree.
OUT_OF_RANGE becomes ErrGestureUndelivered on tap, long press, double tap, swipe and the selector fallback; NOT_FOUND on TapSelector becomes ErrSelectorMatchedNothing. without this the runner reads either as a plain apply failure and counts it toward the failure streak.
a point outside the screen is refused with OUT_OF_RANGE, a selector that matches nothing with NOT_FOUND, and a key with no device-driver equivalent throws instead of pressing nothing. parseBounds also reads uiautomator's [left,top][right,bottom] form, which is what a device actually reports and which left every by-selector tap on a device resolving to nothing.
the companion now emits each node's depth, so the hierarchy mapper can find the containers that clip content reaching past their own frame and mark them scrollable:true, the same fact android reads off uiautomator and the web driver derives from overflow. a dump without depth makes every element a root and roots are never marked, so the legacy bridge reports no scroll rather than a guessed one.
the hierarchy reaches past the screen wherever a scroll container holds content below the fold, so an action derived from it can name a point no touch lands on. tap, double tap, long press and swipe now report ErrGestureUndelivered for such a point, the far edge exclusive because a touch at x == screenWidth arrives at screenWidth-1. resolveSelectorCenter reports ErrSelectorMatchedNothing rather than a bare error.
escape is a key a spec has real use for and no platform could send it. android maps it to KEYCODE_ESCAPE, the ios companion to HID usage 41 and the in-simulator runner to XCUIKeyboardKey.escape, and the Key union accepts it so it can be written at all.
ErrGestureUndelivered marks a coordinate gesture that reached no element and ErrSelectorMatchedNothing a selector that named nothing, so the runner can tell them apart from a device fault. Scroller lets a driver whose scroll is not a finger drag take Scroll separately from Swipe. ExceptionReporter and NavigationReporter carry an app's uncaught errors and document-replacing navigations to the runner.
delegation says to do installs, builds, test runs and greps in subagents and keep the main context for decisions. record-keeping says a finished task updates the files that describe its subject, writes down what was found, corrects old assumptions in place and verifies against the repository.
An element carries find/findAll host functions, so json.Marshal refused the
whole value and the encoder answered nil. ChangedExtractors then emitted no
entry: no error, no warning, no value. Project the value the way the web host
already does (functions dropped, cycles and over-deep branches null, non-finite
numbers null) and turn whatever is still beyond JSON into an error the author
sees, rather than a missing extractor.
Attribute values match by substring, so a selector that named one element
where the candidate was built can name several in the tree it resolves
against, and the lookup sent every one of them to the first match. The host
gates blank an ambiguous tag at enumeration time; this closes the gap between
that moment and the action.
A bare-string target carries no coordinates, so the first match stays the
answer there rather than dropping an authored action.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Drives 40 real draws from a spec that taps each card, through the picker, the
serializer and DecodeAction, and asserts on the points the driver saw. Against
the shared-selector bug all 40 landed on the first card.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
ax.findAll stamped every result with the query selector, and resolveCoordinates
prefers the tree lookup over the element's own coordinates, so N sibling
candidates all executed on the first match. On folio's Home screen the fuzzer
could never open any account but the first.
The gate tests identity rather than cardinality: no node other than this one
answers to the rendered string, checked with the same lookup the runner runs.
A rendered object selector can resolve somewhere the query never matched, so
counting the query would call that unique.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
The V8 host names a web target by data-testid and TapSelector translates that
selector into a CSS attribute match, but the dump carried no such attribute
and no alias could supply one, since an alias only redirects to a key that
already holds the value. tree.Find was therefore always nil for exactly the
selectors examples/folio-web tags with.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
otherElementHoldsFocus answered true when FindNode returned nothing, so an
unresolvable target read as "another element holds focus". confirmFocus then
re-dumped, resolved nothing again, and errored unconditionally. Three of those
in a row abort the run.
Not knowing where the target is says nothing about where the text would land.
The guard's real case, a resolved target with focus outside its subtree, still
errors exactly as before.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
The handle hardcoded true, so every text node and container a spec reached
through state.ax claimed to be a tap target while the enumeration and the
hierarchy dump both resolved it through the tappable selector.
The parity test now compares the handle against the enumeration element by
element in a real browser, which is where the three answers have to agree.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
visibleLabel reads hintText first for an editable element. The dump never
emitted it, so an empty web input fell through text, description and
descendant text to its class name, and the model was shown an identifier no
user can read on exactly the fields a labelling experiment varies.
Same ladder as fieldHint in web-runtime.ts, so one field is named one way on
both hosts.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
A 2x2 of policy against labelling needs the runner to express both factors.
It could only express the policy, so half the factorial had to go through
--extra, where the manifest would not record what was actually run.
Rejected at parse rather than on dispatch: a sweep that finds the bad value
on run 1 of 40 has already spent a cell's worth of device time.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Brings in #73, which landed web fact-parity work overlapping this branch:
selector-tagged ax handles, aria-disabled in `enabled`, shadow-DOM traversal
and a `scrollable`/`editable` dump, plus a third folio property and two new
testTags on HomeScreen.
Six files conflicted. The option-carrying ones took the union of both sides'
fields, so `--label-source` and `--exit-on-violation` both reach the pipeline.
In web-runtime.ts both sides changed how an element is described: master gave
`elementHandle` a selector to tag handles with, this branch gave it raw
attribute names and a field hint. Both survive, and `enabled` now answers
through master's isEnabled while `editable` stays.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* feat(runner): stop the step loop at the first violation on request
* feat(testrun): report violations as a typed error under exit-on-violation
* feat(cli): add --exit-on-violation and exit 2 when it fires
* docs(cli): document --exit-on-violation, --max-steps, and exit codes
* fix(web): enumerate and query across shadow roots in both producers
* test(chrome): compare both producers on a shadow-dom parity page
* test(browser): drive a canvas-under-shadow-root fixture end to end
* fix(web): select the focused field inside a shadow root before typing
* fix(web): report the pathname as the screen when there is no hash route
* fix(web): settle on dom quiescence instead of returning at body ready
* feat(replay-ui): add data-testid hooks the dogfood spec drives
* feat(replay-ui): add the dogfood spec sanderling runs against the replay ui
* fix(replay-ui): scope the screenshot property to the named state panel
* chore(make): add per-platform sanderling build targets
* ci: add dispatch workflows for folio and the replay ui
* docs: describe the dispatch workflows and how to read a failure
* ci(folio): give the ios leg its jdk, android sdk, just, and a clean app start
* refactor(web): use max for the settle budget
* ci: pin calibrated seeds, skip the flaky ios reinstall, bound every job
* ci: authenticate and pin the buf setup step
the anonymous release download hit the shared runner ip rate limit and
failed the job with 'socket hang up' after three retries.
* docs: record that canvas apps need a dom proxy to be text-fuzzable
* fix(ios): bound lifecycle rpcs and claim the target device
a launch the simulator rejects sent the xctest session into a recovery
chain that answered minutes late or never, and the rpc had no deadline,
so the run hung with no trace and no error. also take a per-udid flock:
a second run's reinstall lands under the first's live automation session
and wedges it.
* docs(ci): correct the ios hang wording and note the device lock
* refactor(verifier): derive the lastAction shape from one field list
both hosts must show a spec the same lastAction. one ordered list now
feeds the goja object and the json the web host installs, so they
cannot drift.
* fix(web): install lastAction in the page before extractors read it
state.lastAction was hardcoded null on web, so every property reading it
was silently vacuous: a correct property passed without ever firing.
* fix(web): carry element identity on actions and fix findAll on paths
an action's target was coordinates only, so a property matching on which
element was acted upon could never fire. ax.findAll([a,b]) also returned
nothing on web.
* fix(chrome): wait out a route transition before sampling facts
the tree stays byte-identical and quiet across a cross-fade, so both the
quiet timer and the unchanged-tree escape called it settled mid-flight
and extractors read two screens at once.
* fix: bound the pre-run app launch
launch happens before the runner starts, so --duration never covered it
and a wedged driver hung with no trace and no error.
* fix(folio): read balances from merged cards and treat unreadable as unknown
compose for web merges the whole accountcard subtree, so the balance
child never exists there and every card parsed as 0. the property then
compared 0 to 0 and fired on any submit, which is a false positive
generator. unknown is now null and null is vacuously true.
* test(folio): cover merged-card parsing and unknown balances
* ci(folio): make web an expect-the-bug leg
the web runtime can observe the double submit now, so the health gate
understates it. seed 1 finds it at step 109, 3 runs out of 3.
* docs(ci): explain why a submit tap landing on home is the bug
* fix(ios): read a StaticText's label as its text
AXValue was the only source for text, but a StaticText carries its
string in AXLabel, so nothing on screen had .text on ios: a spec reading
it saw everything on android and nothing here.
* docs(ci): correct the calibrated step ranges
* fix(folio): stop convicting on arithmetic float64 cannot hold
past 2^53 cents the gap between representable values is 128, so a real
1600-cent move reads back as something else and the equality is false
for a healthy submit as readily as a double one. also match parseCents:
a sign or an oversized amount is rejected, not read as an amount.
* test(folio): pin the safe-integer guard and its boundary
* docs: stop teaching the zero-default that caused a false alarm
* docs: write down the silent-vacuity failure modes
* feat(folio): tag the home total and the card transaction count
the total was the only untagged node on the screen, so the spec had to
sum cards and a clipped card broke the sum.
* fix(folio): read the app's own total and refuse contaminated windows
summing cards went null when one was clipped, and the null poisoned the
carrier for the rest of the run. the balance window also spanned every
transaction since the last home visit, so the property convicted on
deltas it could not attribute: the old web witness was 3.16x the typed
amount, not 2x.
* test(folio): pin the window rules and the count invariant
* fix(folio): never read a frame that shows two screens
android dumps a cross-fade with both screens in the tree. the route said
add-transaction while an unscoped find said home, so the oracle took a
half-rendered total as fresh and convicted on a tap that committed
nothing. one function now decides the route and returns null when the
frame is ambiguous.
* test(folio): cover transition frames, card readings and creation
* fix(folio): only disambiguate counts that came from merged text
the equal-length digit rule exists because web merges the card and an
account named -1 makes '12' ambiguous. a dedicated count node has
nothing to disambiguate, so applying it there threw away real evidence.
* ci(folio): pin the recalibrated seeds and drop android to a health gate
web 3 and ios 7 convict 3 runs out of 3 with an exactly 2x witness.
android convicts 2 in 5 because the same seed does not walk the same
trajectory there, so it proves the app runs instead.
* docs(ci): describe the two properties and why android cannot convict
* fix(android): wait out a route cross-fade before snapshotting
the dump could hold two screens at once, and the runner refuses to act
on such a tree, so a quarter of android steps applied no action and the
count varied per run: the same seed never walked the same trajectory.
the ios companion and the chrome driver already do this.
* ci(folio): let the android leg run far enough to see its conviction
* docs: only the repo owner merges
* ci(folio): a thrown predicate is not a conviction
exit 2 means the run recorded a violation, and a predicate that throws
is recorded as one too. so was newAccountBalanceIsZero, an unrelated
property in the same spec. the gate read the exit code and went green
with detection dead.
* ci: install idb-companion from its tap and stop interpolating inputs
idb-companion is not in homebrew-core, so the ios leg died before it
built anything. replay-ui expanded dispatch inputs into the shell.
* docs: correct the snippets and numbers that drifted from the code
* test(sidecar): pin that a slow read counts toward the stability streak
* fix(web): read the page's extractors only on steps that count
the page advances the spec's carriers when it evaluates, but the runner
applied the result only on non-transitional steps. a discarded step
moved the window forward anyway, so the next accepted pair bracketed two
transactions while counting one submit, and convicted a healthy app.
extractor errors now fail the run instead of leaving goja's values in
current against v8's in previous.
* fix(chrome): anchor the transition deadline when the dom goes quiet
it was anchored at script start, so a page that churned past the window
reached the check already expired and returned mid cross-fade. the
driver now publishes the idle timeout it needs, since the caller's 1s
could never spend the 800ms window.
* fix(web): fail on a partial extractor override
same mixed-producer hazard as the install error: some extractors hold
the page's value and the rest hold goja's, and a property comparing
across that split fires on a healthy app.
* docs: six of seven, the seventh is the stock property
* fix(folio): drop a name two cards answer to
homeTxnCountsOf keyed on the account name and let the last card win, so
two accounts the fuzzer named the same collapsed into one entry. a
reading that saw one Travel card and a later one that saw both then
subtracted two different accounts' counts, and
submitCommitsOneTransactionPerAction convicted a healthy app of
double-submitting. it is a gated property in folio-run.sh, so that reads
as "found the submit bug" over a card scrolling into view.
same rule createdAccountHasNonZeroBalance already applies: a name
nothing can attribute is no evidence. counted over every card, since an
unreadable twin spoils the identity too.
* perf(folio): read each frame once
every extractor asked routeOf, and routeOf does five ax.find calls. on
web each find walks the document and every shadow root beneath it, so
the spec cost 110 tree walks a step; homeCards was parsed four times
over. now 5 and once.
keyed on the identity of the state object because both hosts build a new
one per step and hand that one object to every getter, so it cannot
outlive its frame. holding the reference is what keeps that true rather
than likely.
* fix(web): keep an undefined reading's index through JSON
json has no undefined, so an extractor whose getter returned one had its
whole index dropped by JSON.stringify. that index then kept goja's
dump-derived value while its neighbours held the page's, and a property
comparing previous to current across the split fires on a healthy app.
folio has nine on(route, tag) extractors, so this was most extractors on
most steps.
each reading is wrapped in a {value} envelope: the drop now happens
inside the entry, and an absent value means the getter returned
undefined, which is what the goja host records for the same getter. a
json null would instead claim it returned null and x.current ===
undefined would answer differently on the two hosts.
* feat(verifier): report the registered extractor count
the web path needs it to check the page sent one reading per extractor.
* fix(runner): fail when the page reports fewer readings than extractors
the comment here already claimed a partial override was fatal. it was
not: the skipped check only catches indices outside the extractor list,
so a page reporting values for some extractors and not others left the
rest holding goja's reading of the dump with nothing said.
* test(browser): drive an undefined reading through the whole web path
four layers carry it: the page's envelope, the driver's unwrap, the
runner's count check and the verifier's decode. each has a unit test and
only a run proves they compose. goes red both ways, decoding an absent
value as null and dropping the envelope.
* fix(web): offer the aria roles a user activates
only role=button was in the tappable set, so link, checkbox, radio,
switch, tab, option, the menuitems and treeitem were invisible to the
enumeration however plain the control looked. the replay ui builds its
step rows as <li role="option">, and the spec dogfooding it had to
hand-write an action to reach them because no default verb could see a
single row.
both producers build the set from the same role list, since the parity
test compares them element by element.
* test(browser): tap a role-based control end to end
every control on the page is an <li role="option">, the shape the
replay ui gives its step rows, and the spec carries no action of its
own: the property firing is the evidence the default enumeration offered
a tap on one.
* fix(web): read aria-disabled as disabled
the enabled fact came off the disabled property, which only real form
controls have. it reads undefined on the role-based controls the
tappable set now covers, so every one of them looked enabled however
plainly it was marked otherwise, and the fuzzer would spend actions on
inert ones.
both producers answer the same two ways, and the parity fixture carries
a disabled row so the comparison covers it: reverting one side alone
names the element and the fact.
* docs(replay-ui): the enumeration reaches step rows now
the comment said role="option" is not in the tappable selector set,
which stopped being true a few commits ago. selectAStep stays, for the
reason the tab weight below it stays: one row among the page's clickable
elements is a thin chance, and both step-facing properties go vacuous on
a run that never selects one.
* test(runner): bound the last-action test by steps, not wall clock
100ms of wall clock against an assertion that two steps ran fatals under
load with "the web path never installed it", which reads as a
regression. every sibling test in the package uses a long duration and
MaxSteps.
* ci: run the kotlin tests in make test
RouteTransitionTest and the stability poll cover the android settle and
nothing in ci ran them. :sidecar:test needs no android sdk, checked by
running it with ANDROID_HOME pointed at nothing.
* fix(sidecar): measure the stability streak as observed quiet
parameterising pollUntilStable also moved the clock to the start of the
read that opened a run of identical snapshots, so a read's own duration
counted as quiet. the pre-existing caller polls a real uiautomator dump:
at 400ms a read, 750ms of required quiet became 250ms of observed quiet
and the poll settled in two reads instead of four.
the parameters stay, the semantics go back.
* test(sidecar): pin the transition cap by driving it
it asserted 1500 >= 700 + 300, two constants, which can only fail if
someone edits a constant. it now drives awaitSettledTree against a fade
that lands after 700ms and asserts it hands back the settled tree before
the cap. cut the cap to 1000 and it goes red.
* ci: pin buf-setup-action to a commit
it takes a token now, so a floating tag is a token handed to whatever
that tag moves to. note v1 there is a branch, not a tag, so the ref
lookup that resolves it is matching-refs/heads/v1.
* ci: declare least-privilege permissions
none of the three declared any, so each got the repository default.
release.yml and docs.yml already do this. all three only check out,
build, test and upload artifacts.
* ci: fail fast when a server never comes up
the readiness loops fell through silently after 30 tries, so a server
that never started surfaced as an opaque driver failure minutes later.
each now says what did not answer and on which port.
* ci(folio): a missing trace is not a verdict
with no trace the android gate ran its grep against ./trace.jsonl and
reported "never reached AddTransactionScreen, so it never got past
login", which is not what happened. the web and ios branches had the
same misdiagnosis on exit 0.
same class, one line up: the classifier's own failure was swallowed, so
with the evidence reader dead the gate printed a healthy run and exited
0.
* ci(replay-ui): skip a run directory with no trace
the summarise step is if: always(), and under github's bash -eo pipefail
an unmatched glob stays literal, the redirect fails, pipefail carries it
into the assignment and -e kills the step. so a failed fuzz run went red
twice, once for the real reason.
At one model call per step the model arm takes 359 seconds where the seeded arm
takes 47, so a second-based deadline reported violations that were the arm's
speed rather than the application's behaviour. The three cross-arm reachability
properties now bound by steps, derived at the measured 6.383 steps per second.
The two auth-transition properties keep seconds: a user waits through those
regardless of which policy is driving.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Same 300-observation trace at two cadences: a 300 second bound holds for the
seeded arm and violates for the model arm eight observations before the
predicate fires, while a step bound holds for both. Green before and after,
because the step unit already worked; this pins the property rather than
fixing it.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
reduce decremented StepBound into the residual, so the trace reported the
remaining window rather than the authored one: a within(1915, "steps") showed
up as 1875 after 40 steps, and the replay UI renders that string verbatim. The
duration case was fixed when bounded windows were made to serialize their
resolved deadline; the step case was not, and withinFor's comment claimed
otherwise.
The window is now immutable and the closing observation is resolved once, which
mirrors Deadline exactly. A step counts observations the evaluator reduced,
not steps the runner executed, because a skipped step gave the property no
chance to discharge and transitional-step rate is itself policy-dependent.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
The reaping test gave the wedged script one second to install its TERM trap,
so a loaded machine signalled it first and the test failed for a reason it does
not test. It now waits for the script to say the trap exists, then cancels.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
A host asleep mid-run tested nothing, and charging that sleep to an arm reports
it slower for a reason unrelated to the arm. The legend also claimed wall clock
while the number was monotonic. Campaigns written before the split are still
read through the old field name so their run hours do not silently zero.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Duration came from the monotonic clock, which does not advance while a host
sleeps: one calibration run under-reported by about 15 minutes. A run now
carries monotonic_millis for how long it worked and wall_clock_millis for how
much time passed, which is what makes a sleep visible at all.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Measured over 717 InputText steps: nothing was focused before the tap 23.8
percent of the time, the target already held focus 60.4 percent, and a
different element held it 15.8 percent. Silent corruption is only reachable
from that third class, and all four real rejections observed came from it.
Gating there keeps every rejection, skips 84.2 percent of the extra hierarchy
reads, and recovers about 8 percent of Android run time. The pre-tap and
post-tap conditions are now the same predicate stated once.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
CommandContext kills outright, so a run stopped by --run-timeout never ran its
own shutdown and left a sidecar holding a port and a quarter gigabyte,
reparented to init and deaf to SIGTERM. The timeout exists for unattended
hosts, which is exactly where nobody is watching to reap what it leaves.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
InputText tapped its target, slept, then typed. Android and web both inject
into whatever holds focus, so a tap that missed sent the whole string somewhere
else and nothing reported it. On an emulator with a floating keyboard panel
parked over the password field, the tap pressed the keyboard's emoji key and
every step appended the password to the email instead, forever, because the
setup leaf is guarded on the password being empty.
The hierarchy is re-read after the tap and the target, or something in its
subtree, must hold focus. Platforms whose hierarchy carries no focused
attribute skip the read, so they pay nothing.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
The handle fallback read only text, which is textContent and therefore always
empty for an input, so the model could not tell the amount field from the note
field. It now mirrors visibleLabel's ladder rather than introducing a second
naming scheme.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
attrs was spread from element.dataset, whose DOMStringMap keys are camelCase,
so a spec reading attrs["data-cents"] the way every native host reports it read
undefined. In folio-web that left ledgerTxnCount and ledgerBalance permanently
zero: someTransactionExists could never be satisfied, balanceMatchesTransaction
Delta could never fire, and totalBalanceMatchesAccounts compared 0 to 0 and
passed vacuously. Three properties reported nothing because the harness was
blind, not because the application was correct.
The handle also fills hintText and editable now, so an authored InputText on
web names its field the way the same action names it on Android instead of
rendering as Type "12.34" into "".
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
The two edge-case typing leaves become the typing builtin at their combined
weight: that text is deliberately not domain-specific, so naming the field and
leaving the text to the policy is the designed path, and it keeps the seeded
arm on the corpus while the model writes its own.
Total weight is unchanged at 165, so every surviving branch keeps its share and
submitTxn stays at 9.70 percent.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Sampling inside an authored leaf is refused under the model policy now, because
the draw collapses to its first item there. Each sampled leaf offers one action
per value instead.
Lists are short, three rather than five, because the two form leaves also carry
their submit and the seeded picker splits a leaf's probability across the
actions it returns. The doubleTaps path that reaches the planted defect is
unchanged at 5.88 percent, since no root or defaults weight moved.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Setup runs through the picker with the rng under both policies, so a generator
there is legitimate and must keep working. Interleaving enumeration and setup
catches the flag leaking out of the model's walk.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
integers, strings, emails and edgeCaseText read the same rng from() does, so
under the model policy an authored InputText typed the same value on every
step while the seeded arm varied it. That is a silently different experiment,
not just a silently different action space.
Single-valued spans are exempt, because both policies then get the same value:
between(7,7), a zero-length string, and a one-entry corpus. length(4,4) is
still refused, since the length is pinned but each character is drawn from 62.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Candidates returns an error now. The refusal is thrown at the draw and wrapped
with the source of the leaf that made it, since generate() cannot know which
leaf it is inside. Only that marked refusal is fatal: this walk calls every
leaf on every step, so promoting the rest would kill model runs the seeded arm
survives.
Authored actions on a disabled target are no longer dropped from the model's
candidate list. The seeded picker executes whatever the leaf authored, and a
control the application forgot to re-enable is exactly where boundary defects
live, so a policy that cannot attempt it cannot find them.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
from().generate() draws from the picker's rng, which exists only inside
walkActions. The model policy enumerates authored leaves outside that walk, so
the sampler silently yielded its first item on every step: measured over 30
draws the seeded arm reached three targets in roughly equal proportion and the
model was offered only the first. The two policies had different action spaces
and nothing said so.
A single-item sampler short-circuits before the rng, so both policies get the
same value and it is not refused.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
The native desc rule accepts the label or the label at the head of an iOS
merged label; both web translators compared the whole string, so the same
selector matched natively and missed on web. The live-page parity test
caught it.
Same rule and the same message as the native side: a key no element can
carry throws instead of matching nothing. The accepted list is one list,
committed as a fixture both suites assert, so a spec cannot be accepted
by one runtime and rejected by the other.
An empty match is indistinguishable from a screen with no such element,
so a mistyped key generates no action for the whole run and the campaign
finishes clean having explored nothing. The goja boundary now throws,
naming the key and the accepted list.
The object form fell through to the raw attribute map, which carries no
id or desc key on any platform, so {id: "save"} matched nothing while
"id:save" matched. The repo's own web spec uses the object form thirty
times. Both forms now resolve through one switch.
Adds the accepted-key list and UnknownSelectorKeys with it, since the
same silence hides any mistyped key. A key some element carries is always
accepted, so raw driver attributes stay reachable.
Selector matching is written once per runtime: internal/hierarchy over the
dump, web-runtime.ts over the DOM. Nothing made the two agree, and a
selector that resolves on one and not the other is silent, since an empty
match yields no action and the run still passes.
Binding it at module scope bound it to whenever the module was first
imported, so a test file that imported the runtime before setting
SANDERLING_SEED froze the seed at zero for every file after it. The
bundler still replaces the expression with a literal.
The DOM has no package prefix, so the native rule reduces to [id^=]. Both
prefix kinds now go through the one key table, which drops the separate
descPrefix branch that string and object selectors each carried.
idPrefix: is id: with starts-with in place of equality, so a list whose
rows are named <role>_<record id> is reachable by the durable half. The
Android package prefix is skipped the same way id: skips it.
Routing both prefix kinds through matchAttr also makes the object form
work: {descPrefix: ...} matched nothing on the native side while the web
runtime honoured it.
Compares the recorded driver calls across 13 authored shapes. The builtin path
had a parity guard and the authored path had none, which is why it drifted on
almost every verb.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
serializeAction sent the container's own point as both endpoints, so an
authored Scroll({in, direction}) reached the driver as a drag from a point to
itself and did nothing, on the seeded arm. The wire now carries the selector
and leaves the drag to the runner, which sizes it from the container's bounds
and has always had tested support for it that nothing could produce.
No rng runs in the serializer, which lowers an already-drawn action, so the
draw stream does not move. Builtin scrolls compute both endpoints and their
bytes are unchanged.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
The authored descriptor path had no parity guard and diverged from the wire
format on almost every verb. A Wait lost its duration and was skipped as a
zero-duration wait. A Scroll lost its endpoints and its 250ms. A target that
resolved to nothing became a tap at the origin, a phantom focus tap, or a swipe
to (0,0) instead of being dropped.
An authored target object with no x property panicked the whole run at
candidate enumeration: ToInteger was called on a nil goja.Value. A target on
the screen origin is still kept, so the drop rule cannot swallow it.
Builtins were never affected. They serialize through the same path the seeded
arm uses, which the existing policy parity test covers.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Defects per thousand actions counted every step, including steps that chose
nothing and steps whose action was never dispatched. The inflation is
policy-dependent, so it does not cancel between arms: on the fixture campaign
the model arm's yield was reported at 60.3 per thousand against a true 120.7,
because half its steps did nothing.
A runs.jsonl without the count is refused by name and line rather than read as
zero actions, which would report every per-action rate wrongly. The report also
carries steps beside actions now, so the gap is visible rather than folded into
a denominator.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
A step where the policy declined has no action, and a step whose action was
never dispatched did nothing. Both were being counted as actions by everything
downstream.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Descriptions can now repeat after candidates dedup by what they execute. The
guard is index-anchored, so this pins that a repeated string cannot make it
misfire in either direction.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
applyAction could return nil without calling the driver, so the trace showed an
action that looked executed and acted on nothing. Six paths did it: a tap,
double-tap or long-press whose coordinates do not resolve and which carries no
selector, a long-press whose selector is stale, an empty key press, and a
zero-duration wait. It now reports whether it dispatched, and the runner records
the reason and clears lastAction so the verifier never attributes the next state
to an action that did not run.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
The dedup key was the rendered description, which embeds the label, so two
distinct controls sharing a visible label collapsed to one entry and the
survivor carried the first one's action. The second control was not mislabelled,
it was absent from the candidate list, so no policy could reach it. Two
scrollable containers collapsed the same way, leaving the second unscrollable.
The key is now the executable Action struct itself plus whether the model
supplies the typed text, so a new Action field cannot silently fall out of it.
Descriptions may now repeat; the numbering disambiguates and the echo guard is
index-anchored, not description-anchored.
This also makes the label source a pure observation-channel change. It was not
one before: the label fed the dedup key, so the two arms of the labelling
factor enumerated different-sized candidate lists, in both directions depending
on the screen.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Unknown values are rejected at parse time rather than falling back to the
default, matching the generator check: a campaign that completes with the wrong
arm and a correct-looking output directory is worse than one that fails.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Recorded for seeded runs too, unlike model and instructions. Without it the two
seeded cells are indistinguishable in the artifact and the manipulation check
cannot be grouped.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Candidates takes the label source as an argument rather than storing it, which
is what keeps the asymmetry structural: the seeded picker selects by index and
never calls Candidates, so the mode cannot reach it. That asymmetry is
load-bearing, because it makes the two seeded cells of the factorial a
manipulation check with identical draw streams.
The identifier ladder deliberately has no text rung. A fallback that reached
for text would silently turn one arm back into the other.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* fix(replay-ui): drop @font-face rules for fonts that were never shipped
* fix(replay-ui): drop unresolvable JetBrains Mono from --font-mono stack
* chore(replay-ui): remove vestigial empty public/fonts dir
Steps to first violation with clean runs right-censored at the budget, since
per-run yield is a binary at 11 to 45 percent and separating two arms on it
would need roughly 80 runs per arm. Kaplan-Meier, log-rank, Wilcoxon rank-sum
with Vargha-Delaney A12, Holm within each family.
A hand-rolled log-rank that is subtly wrong is a silent-wrong-number generator
and would be believed, so every statistic is validated against a published
worked example with the source named in the test: R survdiff on aml, Freireich
6-MP, Hollander and Wolfe 1973 for the rank sum, printed p.adjust output for
Holm. Two could not be: the k>2 log-rank, guarded by calibration instead, and
the tie-corrected variance, checked against an exact permutation variance.
Failed and timed-out runs are excluded as missing data and counted by reason,
never treated as censored observations, which would bias the result.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
A step could carry a next_action that the foreground guard or an apply error
stopped from running, and nothing said so. An executed-action count read off
trace.jsonl included actions that acted on nothing.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
The strict echo-skip left only a logger.Warn, so a step the guard discarded was
indistinguishable in the trace from a picker that legitimately declined. Any
yield or actions-per-hour figure computed from model traces mixed the two.
Every path that ends a step without a model-chosen action now records its own
outcome.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
It lags the runner's current step whenever a transitional tree caused an
observation to be skipped, which is exactly when the model is shown an older
screen than the step it is choosing for.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
llm-calls.jsonl carries the prompts as sent, the candidate list as the model
saw it, the screenshot reference, the raw response, tokens, latency and how the
step ended. It sits beside trace.jsonl rather than inside it because every trace
line already carries a full hierarchy and both the replay server and the
campaign summarizer scan all of them; folding prompts in would grow the lines
those readers parse for data neither reads.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
An LLM-in-the-loop evaluation has to report tokens per action and cost per
defect, and the client discarded both counters. Served model is recorded
separately from the requested one because a router can substitute a
differently-priced variant.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* feat(cli): add --max-steps for step-bounded runs
runner.Options.MaxSteps already worked but was unreachable from the command
line. A step budget is what makes two generators comparable: one making a
model call per step and one drawing from a PRNG are not comparable per second.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* feat(trace): record arm membership and host in meta.json
meta.json recorded the seed but not which picker ran, how it was configured,
what budget it was given, or which machine produced it. A directory of runs
cannot be attributed to an experiment cell without those, which makes any
factorial computed from such a directory unanalysable after the fact.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* feat(cli): add --arm and populate run meta from it
Model and instructions are recorded only when the LLM picker is the one that
will actually run, so a spec declaring generator = llm() that is run under the
seeded picker does not label its trace with a model it never called.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* feat(campaign): sweep seeds for one experiment cell
campaign.json lists the seeds a sweep intended to run and is written before
the first run, so a host that dropped runs shows up as missing seeds rather
than as a smaller sample. Seed 0 is rejected: sanderling test reads it as
"derive a seed from the clock", which is why conformance/gates.sh controls
nothing today.
Each run contributes one runs.jsonl line carrying steps to first violation by
origin step, the step that armed the failed obligation, so the survival
analysis never reopens a trace.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* fix(runner): no silent generator fallback, and llm on web
--generator llm against a spec declaring no generator = llm(...) logged a
warning and ran the seeded picker. For a comparison campaign that is silent
arm corruption: the run completes, the directory looks correct, and the wrong
policy drove it. It is now fatal.
pickSources also returned the V8 source for both action and extractor on web
before it looked at the generator, so the llm policy was unreachable there.
The two axes are now independent: the driver picks the extractor source, the
flag picks the action source, and llmSource composes with either because the
runner populates the candidate list and screenshot on every platform.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* fix(chrome): make the hierarchy dump agree with the web runtime
Three facts differed between the dump the goja host reads and the DOM the V8
host reads, so the two enumerated different candidates on one page.
scrollable was never emitted, and worker.go reads exactly that attribute while
targets.ts requires it for scrolls, so the goja host could not offer a single
web scroll. clickable tested el.onclick, which React assigns to its root
container for event delegation, making the whole viewport a tap target here and
in no other enumeration. Both now resolve through the selector sets in
pkg/spec/src/web-runtime.ts.
The dump also rooted at body while collectTargets walks querySelectorAll("*"),
so the goja host never saw html, where page-level scrolling lives. It now roots
at documentElement and skips the head subtree, which is all zero-bounds and
would otherwise carry script and title text into the trace.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* fix(conformance): give the gate reproducible seeds
SEED defaulted to 0 and sanderling test reads --seed 0 as "derive a seed from
the clock", so the tunable controlled nothing and a gate failure could not be
re-run. SEEDS now takes one explicit non-zero seed per run, recorded in the
results table so a failing row names its stream.
The five runs stay on five different streams: a gate that scored one path five
times would catch less than one that scores five.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* fix(chrome): emit editable as a plain boolean
editable was emitted as `isEditable || null`, and an absent field sends
internal/hierarchy into the native fallback, which reads any class name
containing "EditText" as an Android text widget. On web that is just a CSS
class, so a page styling a div with it was editable to the goja host and not to
the web runtime, and the model policy could be offered typing into a div.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* fix(spec): leave the head subtree out of the web target walk
collectTargets walked querySelectorAll("*") while the hierarchy dump skips head,
so the two hosts enumerated different element sets on every page with a <head>.
No candidate changes: builtinCandidates pushes only for targets acceptsTarget
admits, and head elements have no positive bounds, so the list the draw ranges
over is untouched. What changes is that targetIndex now means the same thing on
both hosts.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* test(chrome): compare the facts both hosts derive from one DOM
The existing parity harness hand-authors the facts on both sides, so it proves
that given identical facts both hosts select identical candidates, and says
nothing about the two code paths that derive those facts from a real page. Four
divergences lived in that blind spot and it passed throughout.
This drives one real page and compares clickable, enabled, editable, scrollable
and positiveBounds element by element, plus the element sets themselves, which
is what catches a host that omits html or includes head. Reverting any of the
four fixes makes it fail naming the element and the fact.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* chore(make): run the browser packages one at a time
Both launch Chrome and launching two at once has failed with "Launch: context
canceled".
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* style: remove every em-dash and en-dash
Eighteen occurrences across fourteen files. Each sentence was repunctuated to
suit what the dash was doing rather than swapped for a hyphen, which produces
comma splices. The minus sign in folio-web's ledger is a minus sign and stays.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* fix(chrome): honor the caller context in Launch
Launch and clearState ran against d.tabCtx, so a target that accepts the
connection and never answers wedged the process past its own --duration and
through SIGTERM, needing SIGKILL. Unattended that is a campaign worker lost for
the rest of the sweep with no diagnostic.
The browser is still allocated against d.tabCtx first, because chromedp starts
Chrome under whichever context calls Run first and allocating under a caller
deadline would kill the browser when Launch returns. Everything after
allocation goes through runCtx.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* fix(sidecarassets): publish the extracted jar through a rename
Extract wrote a 96 MB jar with a plain WriteFile into a temp path every
sanderling process on the host shares. On a cold host several concurrent
workers all miss the checksum and all write the same path, and O_TRUNC lets one
spawn a JVM against another's half-written archive. A fresh experiment host is
exactly a cold host.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* feat(campaign): kill a run that outlives --run-timeout
A wedged run holds its worker for the rest of the sweep, and on an unattended
host nothing else will send it a signal. Defaults to three times --duration and
must exceed it. A killed run is recorded as timed_out rather than as a generic
failure, so the analysis can tell a lost cell from a real crash.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* style(test): gofmt browser_test.go
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
* fix(ltl): give every thunk a construction identity
Two distinct unnamed predicates both described as "Thunk(...)", so obligation
collapse merged their residuals and could drop a live violation. Identity is
assigned at construction and the fields are unexported, so a thunk cannot be
built without one.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(ltl): reduce a thrown-predicate residual instead of panicking
The verifier substitutes an ErrorFormula for the residual of a property whose
predicate threw, and that residual is fed back in on the next step. reduce had
no case for it, so the run crashed. It re-reports the same failure now.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(ltl): make a bounded always the dual of a bounded eventually
G<=n(f) and not F<=n(not f) disagreed on traces where the inner was still
pending when the window closed, so nnf's negation normal form was not semantics
preserving. Both sides now range over the observations at which their inner can
definitely resolve: the eventually keeps a pending inner as a disjunct, and the
always discharges vacuously at window close.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(ltl): arm a one-shot root once per run
A root that carries its own horizon is one obligation for the whole run, not one
per observation. Re-instantiating a top-level eventually monitored G F<=n(p)
instead of F<=n(p) and left one live obligation per step behind; a bounded
always restarted its window every step and never closed.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(verifier): stop wrapping a top-level eventually in always
`eventually(p).within(300, "seconds")` as a property meant "within 300 seconds
of every step", which spawned an obligation per step with its own resolved
deadline. A 553-step run carried 553 of them and serialized a 75 KB residual.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(ltl): serialize the resolved deadline of a bounded window
Two obligations spawned at different steps from one duration-bounded formula
differ only in the deadline the evaluator resolved for them, so they serialized
identically and the trace erased a distinction the evaluator makes. The authored
window stays in amount/unit; the resolved deadline rides alongside.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(verifier): split a witness's origin step from its detection step
A deferred obligation spans two steps: the one that armed it and the one whose
reduction failed. They were conflated under one index, so the extractor snapshot
(which is the detecting step's state) was reported against the origin step.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(runner): record a witness's detection step in the trace
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* feat(replay-ui): show the step a violation was detected at
The witness evidence is the detecting step's state, so say which step that is
and let a reader jump to it.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(verifier): record the extractor state the predicates actually read
On the web path extractor bodies are evaluated in V8 and injected here, but only
the goja value was replaced. The trace diff and the violation witness therefore
described a state no property ever saw.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* refactor(spec): one candidate producer over one target-eligibility rule
Both hosts routed verbs themselves and both policies enumerated their own
actions, and all four drifted. Web sent `swipes` to scrollable containers only,
so swipe-to-dismiss on a list row was reachable on native and unreachable on
web; the model policy folded gestures its own way and could not reach what the
seeded picker drew.
A host now reports facts about every element and never decides which verb may
act on it: targets.ts acceptsTarget owns that for both. pick.ts builtinCandidates
is the single enumeration, and the model policy reads it through
__sanderlingEnumerateBuiltin__ instead of reimplementing it in Go.
Gesture verbs change with it: scrolls stay vertical over scrollable containers,
swipes go free-form in all four directions from any element with real bounds.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(runner): name a builtin scroll by its drag origin
A builtin gesture carries endpoints and no selector, so every scroll rendered as
"Scroll down " in the prompt's recent-action memory and two scrollable regions
were indistinguishable.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(chrome): clear storage over cdp instead of scripting an opaque origin
Launch runs while the tab is still on about:blank, whose opaque origin denies
storage access, so localStorage.clear() threw SecurityError and every web run
died at launch. Storage.clearDataForOrigin needs no navigation. The exception
helper lands here because "Uncaught" is what hid this for so long.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(chrome): enable the swiftshader webgl fallback
Headless Chrome runs with --disable-gpu, and without this flag it refuses the
software WebGL backend: getContext returns null, so a canvas-rendered app paints
nothing and every screenshot is identical black.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* fix(web): resolve testTag through data-testid or id
Compose Multiplatform emits its testTag into the element id, which the native
table already accepts via the resource-id alias. The two web selector tables
were the only place that rejected it.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* test(spec): type-check the spec api as part of make test
The fake runtime in api.test.ts did not return a chainable handle from extract,
so the file had not type-checked since named() was added. Wiring the check into
make test stops it drifting again.
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* docs(manual): one-shot eventually and the gesture verbs
Claude-Session: https://claude.ai/code/session_01Fj4wJUikdABuMQEETwW55J
* feat(spec): add llm() action-backend marker
* feat(spec): make llm marker inert on the JS picker
* feat(spec): expose __sanderlingSampleInput__ corpus draw
* feat(openrouter): minimal chat-completions client
* test(openrouter): cover request shape, parse, and errors
* feat(verifier): thread screenshot + capture corpus sampler
* feat(verifier): LLM accessors — candidates, config, sampler
* test(verifier): cover AllCandidates, LLMConfig, SampleInput
* feat(trace): record action Source and LLMReasoning
* feat(runner): thread step screenshot into PushSnapshot
* feat(runner): llmSource selects actions via OpenRouter
* feat(runner): wire llmSource selection and trace stamping
* test(runner): cover llmSource selection, mapping, downscale
* docs(folio): add llm action-backend example spec
* docs(folio): document the LLM action backend run
* feat(llmclient): support OPENAI_API_KEY, openrouter wins
* refactor(runner): rename openrouter package to llmclient
* docs: both api keys, example model gpt-5.4-nano
* docs: add pr style rules to claude.md
* fix(runner): explain action kinds in llm prompt to stop swipe loops
* feat(trace): record llm ranked list and chosen rank
* feat(runner): stamp llm ranked list and chosen rank on trace
* fix(runner): tap by selector to survive layout shift after observe
* revert(runner): drop selector-first tap; broke path/testTag selectors
* feat(spec): llm() accepts optional instructions
* feat(verifier): read llm instructions off config
* feat(runner): append spec instructions to llm system prompt
* docs(folio): describe app in llm spec instructions
* feat(bundler): map generator export to globalThis.generator
* feat(verifier): read llm config off globalThis.generator
* feat(runner): gate llm source on --generator flag
* feat(cmd): add --generator llm|seeded flag
* test: cover --generator flag parsing and pickSources gating
* feat(verifier): enumerate llm candidates by walking actionsRoot
collect-walk the weighted action tree: recurse weighted branches
accumulating selection probability, call authored leaves once for
concrete actions, enumerate builtins per element. label controls by
visible text (borrowing descendant text), fold gestures into directional
scrolls over scrollable containers, drop disabled, dedup descriptions.
* test(verifier): cover candidate enumeration walk
* feat(verifier): add SetupAction to walk setup without the seeded root
* test(verifier): cover SetupAction setup-only precedence
* refactor(llmclient): make JSONSchema.Schema raw json for pinned field order
* feat(trace): record llm choice number and chosen_action echo
* feat(runner): llm picks one number from weighted candidates
drop the seeded-root call for a setup-only precedence path, render a
numbered weighted candidate list, pin a reasoning-first choice schema,
strict-skip when chosen_action does not echo the numbered entry, and let
the model supply typed values (corpus fallback when empty).
* test(runner): cover choice schema, strict-skip, and setup precedence
* refactor(verifier): drop the superseded AllCandidates enumeration
* feat(folio): drive spec.ts under --generator llm; drop spec-llm.ts
* fix(verifier): label editable fields by hint, not the typed value
an editable field's own text is its transient content; prefer the hint
so the field is named by purpose and the label stays stable.
* test(runner): cover weight-suffixed echo and stripWeightSuffix
* fix(runner): accept chosen_action echo that carries the weight suffix
real runs showed the model copies the whole numbered line including the
trailing (w34) weight annotation, so strict-skip rejected ~91% of picks
and the llm was paralyzed. strip the weight suffix before comparing. also
nudge the prompt to stress-test repeated submissions (idempotency).
* fix(verifier): skip llm enumeration on cross-fade frames
a navhost mid-transition carries >1 route *Screen in a collapsed
coordinate space; acting on it taps garbage (soft keyboard). real runs
showed the llm acting on 44% of steps being such frames. skip them so the
llm re-observes a settled frame next step.
* feat(folio): show current balance on the add-transaction screen
renders the account's balance (testTag TxnCurrentBalance) below the
account name, above the credit/debit toggle, so before/after screenshots
carry comparison data.
* fix(replay): derive device space from screen extent, not first node
the first positive-bounds element is often a short status-bar node
(320x24 on android); using it gave a 320/24 aspect ratio that squashed
the screenshot overlay into a grey horizontal band. use the max extent
across elements (like the runner's screenBounds) instead.
* fix(folio): show balance as a compact one-line label
per review: one line, account-name-sized, e.g. "Balance: $0.00"
instead of a large balance card.
* fix(folio): move balance into the header, one compact line under the account name
* fix(replay): attribute deferred violations to the causing step, not detection
* fix(replay): show a step's own violations in both panels, no next-step bleed
* refactor(hierarchy): one Tree.Transitional, drop the duplicated cross-fade check
* chore: ignore .playwright-mcp scratch output
* docs: document the llm generator and --generator flag
* docs(spec): correct the llm() comment; config reads off globalThis.generator
* docs: add pr description rules
pj