Commit Graph
91 Commits
Author SHA1 Message Date
morluto abd36dd250 refactor(browser): parse Runtime and Debugger events once by registered schemas (#10742)
* refactor(browser): parse frame lifecycle events once by a registered event schema

CdpConnection admits each event through protocol/CdpEvents.ts, which parses
the five Page frame lifecycle events once into identity facts plus a frame
URL observation, like a frame-tree node. Screenshot, page capture, the
observation timeline, WebMCP, Electron inspection and the execution window
read those facts instead of re-parsing the raw payload; CdpFrameFacts no
longer carries the raw source record. Playwright CDP session subscriptions
name an event from Playwright's protocol table, and an architecture guard
counts the remaining raw event payload reads so they can only shrink.

* refactor(browser): parse Runtime and Debugger events once by registered schemas

Register Debugger.scriptParsed, Debugger.scriptFailedToParse,
Runtime.executionContextCreated, Runtime.executionContextDestroyed and
Runtime.consoleAPICalled. Script and context payloads carry both the
tolerant facts passive capture uses and the active runtime's strict
declaration, derived from one parse, plus the declaration's byte size for
the runtime retention budget. Page capture, Electron script capture, the
V8 Inspector provider and the runtime source inventory now read the
registered payload. The cdp-raw-event-params baseline drops from 20 to 7.

* refactor(browser): admit transport events through the registered-event owner

Remove the unused cdpFrameUrl wrapper; its fragment cases now exercise
observeCdpFrameUrl. Add a connection-level test that two subscribers share
one parsed payload and the raw params.

* fix(browser): report an unserializable event declaration's size instead of throwing

JSON.stringify can throw on a declaration JSON.parse accepted (nesting
deeper than the serializer's stack). The registry sized script and context
declarations inside the connection's WebSocket callback, so such an event
escaped as an uncaught exception. The size is now a Result; the runtime
source inventory raises the failure inside its own catch, as before.

* refactor(browser): parse Network events once by registered schemas (#10743)

* refactor(browser): parse Network events once by registered schemas

Register Network.requestWillBeSent, responseReceived, loadingFinished,
loadingFailed, webSocketCreated and webSocketFrameSent/Received. The
request payload carries the tolerant facts page capture and the timeline
read plus the active runtime's strict declaration. Page capture's network
and WebSocket handlers, the timeline's request tracking and the runtime
request inventory now read the registered payload. The
cdp-raw-event-params baseline drops from 7 to 5.

* refactor(browser): drop payload helpers the capture handlers no longer use

* refactor(browser): read every CDP event payload through the registry (#10745)

* refactor(browser): read every CDP event payload through the registry

Register Target.attachedToTarget, Target.detachedFromTarget,
Page.lifecycleEvent and WebMCP.toolsAdded/toolsRemoved, and carry a
requested navigation's reason and a commit's transition type in the
frame-event payload. The timeline no longer threads raw params into its
entries: each entry receives its timestamp, frame and loader identity
explicitly. WebMCP discovery, the service-worker policy blocker and the
target-event matcher read the registered payload. With no raw reads left,
the cdp-raw-event-params guard is enforced at zero instead of counted.

* refactor(browser): drop raw payload helpers the timeline and worker blocker no longer use

* refactor(browser): type Playwright CDP session events by the registry (#10751)

PlaywrightCdpSession.on subscribes only to registered events and hands
the listener the payload parsed by its registered schema, returning the
unsubscribe function. Private storage seeding reads registered payloads,
including the events of non-flat child sessions, and applies its strict
seeding policy to the parsed facts. Its local attach, detach, nested
message and pause schemas and the storage-only script/context schemas are
removed; Debugger.paused, Runtime.executionContextsCleared and
Target.receivedMessageFromTarget join the registry.

* test(browser): admit fake policy events through the event registry

The service-worker policy now reads event.registered, but the PlaywrightScenarioSession fake connection delivered raw events, so three detached-target tests never reached the policy. The fake now admits each event with receiveCdpEvent as CdpConnection does.

* fix(browser): fail the CDP session closed when an event listener throws (#10758)

* fix(browser): fail the CDP session closed when an event listener throws

An exception from a CDP event subscriber escaped CdpTransportCore's socket message callback uncaught (for example the service-worker blocker's protocol_error on a Target event without a sessionId). The transport now stops delivery, fails the connection as a protocol error whose cause is the listener's exception, rejects pending and later commands with it, and notifies protocol-failure subscribers. The Playwright CDP session relay applies the same rule to its event listeners.

* fix(browser): report a failed service-worker policy connection at the next policy check

The policy's event listener rejects a malformed Target event, which now fails its connection. With no policy command pending, nothing reported it and later iframe targets went unprotected. The policy records its connection's protocol failure and assertBlocked, run before each scenario step, rethrows it.

* fix(browser): keep cleanup commands available after a listener failure

A listener failure no longer marks the wire as broken. The transport still rejects pending commands, notifies protocol-failure subscribers, stops delivering events and refuses reuse, but it keeps reading replies and admits commands sent in cleanup grace. The service-worker policy can then remove its init scripts, restore the main realm, detach and release its socket. Wire-level protocol errors still refuse every later command. The PlaywrightScenarioSession policy fake gains onProtocolFailure, which the policy now subscribes to.

* refactor(browser): keep raw CDP event payloads out of the received event (#10764)

* refactor(browser): keep raw CDP event payloads out of the received event

CdpReceivedEvent and CdpTransportEvent no longer carry params: receiveCdpEvent hands the raw payload only to the registered schema, so a subscriber cannot read it by property access, destructuring or computed keys. This supersedes the event.params text guard in architecture-guards, which is removed.

* refactor(browser): read registered event payloads in the JavaScript verifiers

The runtime, storage and URL-userinfo verifiers subscribe to CdpConnection, which no longer hands subscribers raw params. They now read event.registered. A test scans every scripts/ onEvent callback for raw payload reads, since the TypeScript check does not cover them. The storage verifier's pause trace records a malformed pause instead of the unregistered reason member.

* refactor(browser): track the timeline's main frame with the document tracker (#10779)

TimelineCapture kept its own mutable main frame id and replaced it on a root replacement. It now holds the CdpDocumentTracker that CdpTargetSession owns and observes the replacement document through it; every main-frame check (lifecycle, navigation, document request, redirect) reads the tracker's current frame. A characterization test pins that the replaced root's lifecycle events are no longer recorded.

* refactor(browser): select CDP events only by their registered name (#10829)

CdpReceivedEvent carried the wire method as a plain string, so a subscriber
could compare it with a misspelt or unregistered name and the check compiled
but never matched. Three owners selected Inspector.targetCrashed, an
unregistered event, that way.

Register Inspector.targetCrashed and drop the wire method from the received
event. Subscribers select events by registered.method, whose type is the
registered event names. The capture coverage start keeps reading the
wire domain, now the event's domain field.

* fix(browser): retire a policy child session that CDP reports gone

After a listener failure the connection stops delivering events, so an
iframe's Target.detachedFromTarget never retires its policy target. Cleanup
then sent Page.removeScriptToEvaluateOnNewDocument to the dead session,
kept the stale child, and every retry failed before detaching the root.

The service-worker policy owner now treats CDP's session-not-found
rejection as proof that the child is gone and retires it. The classifier
moves from CdpTargetSession to CdpCommandRejection, its owner, and the
target-session detach path uses the same one.

* refactor(browser): route target events through one session owner (#10835)

* refactor(browser): route target events through one session owner

Page capture, WebMCP discovery, screenshot capture and the runtime sources
each compared the envelope session ID themselves, while the observation
window and execution collection used cdpTargetEventMatches, which also
routes the root Target.detachedFromTarget that names the target's session.

onCdpTargetEvent in CdpTargetEvents.ts now subscribes every target consumer
with that one rule; the matcher is private and the per-subscriber
comparisons are removed. Electron inspection keeps its own filter while
its capture isolation awaits a maintainer decision.

* fix(browser): report WebMCP subframe tools whose document is unverified (#10839)

* fix(browser): report WebMCP subframe tools whose document is unverified

A committed subframe navigation kept the frame's tools silently when CDP
reported no loader at the same URL, and removed them on any URL change even
when the loader proved the same document.

Each frame now carries a CdpDocumentTracker, the session's document
identity owner. Only a proven change removes the frame's tools. An unknown
comparison keeps them visible, marks webmcp_tools attach-limited and names
the frames whose document identity could not be established.

* test(browser): prove an unverified live subframe limits WebMCP coverage

* fix(browser): key a WebMCP registration by its frame and tool name

A subframe tool kept across a committed navigation that changed the URL
but not the loader was stored under a key built from the old URL, so a
later toolsAdded for the same frame and name added a second entry.
Registrations are now keyed by frame and name; tool_key still records the
registering frame URL.

* test(browser): deliver fake CDP events through the transport's admission (#10840)

Event fakes each rebuilt the received event: CdpAgentHints typed its
listeners as raw wire events behind an unknown cast, and the Electron,
scenario-session and routing tests each wrapped receiveCdpEvent by hand.

fakeCdpEvents (tests/fixtures) is a CdpConnection.onEvent fake whose emit
admits each raw event with receiveCdpEvent and the wire byte count, so a
change to the received event shape fails at the fake instead of silently
not exercising subscribers.

* refactor(browser): own the selected page generation in one target document tracker (#10860)

* refactor(browser): own the selected page generation in one target document tracker

Page capture, screenshot, the observation window, the runtime session and
the execution window each built a document tracker or rule of their own
from root-frame navigation events. CdpTargetDocument now follows the target
session's root-frame events and the frame-tree replies its consumers read,
classifies root identity once, and reports each navigation to consumers
that decide only what it means for their evidence.

* fix(browser): report a lost scenario policy connection before the next step (#10864)

* fix(browser): report a lost scenario policy connection before the next step

The service-worker policy recorded only protocol failures. When its CDP
socket closed or its page session detached, later documents lost the
blocking script, but the scenario learned of it only when a frame check
found an unprotected realm. The owner now records the disconnect or
detach when CDP reports it; the next policy check and the next step's
action report it first.

* fix(browser): refuse an unregistered CDP command before it reaches the browser (#10868)

JavaScript verifiers can name any command, and the send owners refused an
unregistered one only when its reply arrived, after the browser had acted
on it. CdpConnection and the Playwright session now refuse the name before
sending, and the Playwright session refuses an unregistered event name
instead of subscribing to events no listener receives. The verifier text
scan stays as an early warning and documents what it cannot see.

* fix(browser): keep a proven root change and the loaderless selection's uncertainty

Two review findings on the target document owner. A root event naming
another frame with a malformed loader set `changed` but was announced as
unidentified, so the execution window failed with protocol_error and
dropped what it had collected; the owner now announces it as the proven
change it is. The runtime session compared each frame-tree reply with the
tracker's current document, which can learn a loader from an event, so a
loaderless selection followed by an event and a reply with that loader
compared as same; it now compares the reply with the selected document
through the owner, and reports target_changed as before.

* test(browser): refuse the next scenario step after the policy session detaches

* refactor(browser): read WebMCP document identity from the target document owner (#10884)

* refactor(browser): read WebMCP document identity from the target document owner

WebMCP discovery compared the first and last frame-tree roots itself and
kept a document tracker per subframe. CdpTargetDocument now follows the
selected root and each committed subframe for it: the final frame tree
feeds the owner, a root change proven by a navigation event during
discovery is target_changed as in page capture, and a subframe whose
committed document changed loses its tools. A subframe navigation without
a loader still keeps the tools and reports them as unverified.

* fix(browser): start a subframe first seen within its document with an unknown loader

A subframe absent from the initial frame tree and first seen through Page.navigatedWithinDocument had no document tracker, so a later commit started a fresh one and the tools it registered meanwhile lost the unverified-document limitation. The owner now starts such a subframe with an unknown loader.

* fix(browser): refuse a verifier's unregistered CDP event name at subscription (#10885)

JavaScript verifiers subscribed to every event with CdpConnection.onEvent
and compared registered.method with a string literal. A misspelt or
unregistered name never matched and the verifier waited until its timeout.
CdpConnection.onRegisteredEvent now takes one event name, refuses an
unregistered one before subscribing (the same registry check the Playwright
session applies), and hands the listener only that event's registered
parse. The verifiers subscribe only through it, and the evaluation test
rejects a raw onEvent subscription in scripts.

* fix(browser): keep subframe documents by frame and loader without their URLs
2026-11-15 19:54:17 +08:00
morluto d22da738d5 refactor(adb): report a pulled file's digest failure from its own type (#10848)
* fix(pcap): report a refused tool launch as an unavailable tool

When a capture read's tool launch was refused with EPERM, pcap reported it
as unreadable tool output (AnalysisProtocolError) instead of an unavailable
tool. Pcap also never told "not executable" apart from "missing".

Add classifyLaunchFailure in src/process as the one owner of which spawn
errnos mean missing (ENOENT, ENOTDIR) and which mean present but not
executable (EACCES, EPERM, ENOEXEC, EISDIR). Pcap, apktool and gdb used three
different errno lists; all three now use it.

* refactor(process): classify owned command outcomes through one owner

Add ProviderOutcome, read once from the supervisor's facts, and classifyProviderOutcome, the one mapping from an outcome to a typed error. Migrate pcap and the shared artifact producer command failure (TShark, sigrok). Add the rea/provider-failure-classifier lint rule with a shrink-only baseline of per-site (cause: unknown) => AnalysisError functions.

* refactor(process): classify Java and adb launch refusals with the shared owner (#10598)

The Ghidra and JADX Java probes and adb's version probe now take
missing-versus-not-executable from classifyLaunchFailure instead of their
own errno checks. ELOOP and ENAMETOOLONG join the owner's missing set, so
adb keeps its existing classification and both Java probes now treat a
non-directory path component as a missing runtime.

* fix(pcap): preserve command launch attribution

* test(pcap): expect an unknown failing stage when a limited capture tool exits unsuccessfully

* refactor(process): classify owned command outcomes through one owner

Add ProviderOutcome, read once from the supervisor's facts, and classifyProviderOutcome, the one mapping from an outcome to a typed error. Migrate pcap and the shared artifact producer command failure (TShark, sigrok). Add the rea/provider-failure-classifier lint rule with a shrink-only baseline of per-site (cause: unknown) => AnalysisError functions.

* test(lint): cover baselined files in the provider-failure classifier rule

* fix(apktool): report the decoder JVM's heap exhaustion as memory, not a target refusal

The supervisor now watches for a managed runtime's out-of-memory report while output streams (JVM java.lang.OutOfMemoryError, V8 'JavaScript heap out of memory'), so a retention budget cannot hide it. ProviderOutcome gains a runtime-exhausted kind and classifyProviderOutcome maps it once. Apktool decode opts in and moves to the shared classifier; Ghidra's single-purpose output token becomes the JVM runtime option and its area matching reuses the shared table.

* fix(process): keep the runtime's out-of-memory report on the memory error it supports

* fix(pcap): attribute a limited capture's exit through the prlimit wrapper

* fix(process): keep a shortened out-of-memory line marked as shortened

* fix(process): identify the refused executable for every launcher shape

Name the executable on errnos Node throws synchronously from spawn
(ENOEXEC, ENOTDIR, ...) and read the Windows native launcher's
executable-launch constraint, so pcap attributes those refusals to the
selected tool or limiter instead of an unknown stage.

* test(pcap): prove synchronous launch naming with ENOTDIR on every POSIX host

* fix(apktool): advise the JVM limit that governs the reported memory area

* fix(ghidra): read the supervisor's parsed OOM area instead of reparsing lines

* fix(sqlite,capture): read V8 heap exhaustion from streamed output, not retained stderr (#10724)

The SQLite worker and the Node capture decoders now declare runtime: "v8", so the supervisor observes V8's heap report as output streams. Both classifiers read the runtime-exhausted outcome instead of matching retained stderr, so a report beyond the retention budget, or one whose process then overflowed its output budget, is still reported as memory exhaustion.

* refactor(evm): classify EVMole worker outcomes through the shared provider classifier (#10716)

* refactor(evm): classify EVMole worker outcomes through the shared provider classifier

The EVMole worker's cancellation, deadline, output budget and CPU/file-size signals now go through classifyProviderOutcome; only the prlimit launch refusal and the reserved EFBIG exit remain in its dialect. The worker is opted into V8 exhaustion scanning so a heap report names its fixed 256 MiB heap. providerOutcomeOf now reads a launch refusal only from a spawn errno, so a private-storage or carrier read errno is no longer reported as a limiter launch failure.

* fix(evm): keep a stream failure's exit status and update Linux worker expectations

* refactor(process): read launch refusals through the launcher identity owner

* refactor(wasm): classify WABT command outcomes through the shared provider classifier (#10727)

* refactor(wasm): classify WABT command outcomes through the shared provider classifier

WABT's cancellation, command deadline, output budget, and inherited-limit signals now go through classifyProviderOutcome; its dialect keeps only the configuration-probe and launch-refusal unavailability and the unclassified-exit output error. A tool ended by SIGXCPU/SIGXFSZ or the memory limit is now a resource constraint instead of an output error, and a refused launch in the producer phase is reported as unavailable. The shared timeout context gains the analysis setting that selected the deadline, and the output-limit context may name its exceeded budget.

* fix(wasm): keep setup guidance when a WABT version probe's output cannot be collected

* refactor(providers): classify raw provider failures through one owner

* refactor(process): build classified cancellations with cancelledBy and the carried caller reason

* chore(lint): count runOwned's close-failure mapper from main in the classifier baseline

* fix(process): carry the runtime's exhaustion report on every memory error that claims it

SQLite and historical-capture memory errors recognised a streamed V8 heap report but dropped it, so a report outside retained output reached callers without its source line or omission flags. One projection, runtimeExhaustionReport, now owns that shape for the shared classifier and both consumers. Also lowers the EVMole and WABT cancellation-constructor baselines.

* fix(process): keep the provider's recovery for a native-thread exhaustion report

The shared classifier replaced the provider-supplied recovery with generic capacity advice whenever the JVM reported it could not create a native thread, dropping Apktool's launcher guidance and flipping retryable to true. The provider's recovery now wins; the capacity advice remains the fallback for providers that supply none.

* refactor(pwntools): classify pwntools failures through the shared provider classifier (#10782)

* refactor(pwntools): classify pwntools failures through the shared provider classifier

pwntools' per-site classifier is replaced by pwntoolsFailureDialect for classifyProviderFailure. Command outcomes go through classifyProviderOutcome, launch refusals through launcherRefusalOf instead of a 'spawn ' syscall prefix check, the input limit through the file dialect's fileSize, and the cleanup report through commandCleanup. The shared read cancellation now carries the reader's partial observation, and resource-signal reports accept provider detail.

* fix(pwntools): report a failed command's output once and expect resource_constraint for overflow

The shared outcome owner already attaches the command's output as capturedOutput, so the pwntools adapter fallback no longer copies the same text into diagnostics.stdout, diagnostics.stderr and diagnostics.captured_output; diagnostics keep the exit facts. Update the overflow expectations the dialect migration missed (layout provider cleanup and over-budget rows, pwntools cleanup) to resource_constraint.

* refactor(wasm): leave WABT cancellation to the shared owners

WabtFailures matched any Error named AbortError as a cancellation. No WABT path produces one: reads, path checks and owned commands report cancellation as ArtifactReaderFailure or OwnedCommandFailure, which classifyProviderFailure and classifyProviderOutcome already classify with the caller's reason. The provider's two signal checks now build the cancellation with cancelledBy. Lowers the AnalysisCancelledError constructor baseline for both WABT files to zero.

* fix(wasm): report a cancelled snapshot write as the caller's cancellation

The private snapshot write passed the caller's signal to Node's writeFile, whose abort envelope reached the dialect's unclassified output failure once the name-based AbortError guess was removed. The write now turns the envelope carrying its own signal's reason into cancelledBy at the producer.

* refactor(firmware): classify firmware failures through the shared provider classifier (#10810)

Firmware producers now report their own deadline expiry: the start check, the output inventory and launcher admission throw the phase timeout where they observe the abort, and a read the phase signal cancelled reports it through the dialect's interruption. The classifier no longer rereads a cancellation's reason as a timeout. A new readCleanup hook keeps firmware's cleanup report shape.

* refactor(captures): classify TShark, sigrok and pcap failures through the shared provider classifier

TsharkCaptureProvider and SigrokRecordingProvider replace #failure with a dialect, and PcapProvider replaces analysisFailure and the tool-identity classifier. ArtifactAnalysisResources exposes its command context and a readCleanupFailure composer so the readCleanup hook keeps each provider's cleanup report shape.

* refactor(jeb): classify JEB failures through the shared provider classifier

JEB's raw MCP failures go through classifyProviderFailure with a JEB dialect;
a typed attribution step reports a shared phase's failure under the REA
operation. Each MCP phase deadline now reports its own interruption when it
aborts, deciding a cancellation's owner from the caller's own signal instead
of comparing abort reasons. A refusal carrying the JVM's OutOfMemoryError
report becomes a resource constraint naming JEB's Java heap.

* fix(jeb): advise the JVM limit that governs the reported memory area

* refactor(ghidra): classify Ghidra failures through the shared provider classifier

PSP extension inspection reports through classifyProviderFailure with a
Ghidra dialect. The startup deadline decides at abort time whether the shared
startup owner closed it, so Ghidra and Hopper no longer compare the abort
reason with the owner's sentinel, and Ghidra's startup no longer compares a
producer cancellation's reason with the deadline's. Three helpers that only
construct one error kind now name that kind instead of AnalysisError.

* refactor(hopper): classify original-source read failures through the shared provider classifier (#10838)

* test(ghidra): pin the aborted startup deadline over a producer cancellation

* refactor(kuna): classify Kuna failures through the shared provider classifier

Session operations, the workspace copy and the project export report raw
failures through classifyProviderFailure with Kuna dialects; the copy and
export keep their cleanup report shapes through the readCleanup hook.

* refactor(adb): report a pulled file's digest failure from its own type

* refactor(ghidra): classify Ghidra failures through the shared provider classifier (#10836)

* refactor(ghidra): classify Ghidra failures through the shared provider classifier

PSP extension inspection reports through classifyProviderFailure with a
Ghidra dialect. The startup deadline decides at abort time whether the shared
startup owner closed it, so Ghidra and Hopper no longer compare the abort
reason with the owner's sentinel, and Ghidra's startup no longer compares a
producer cancellation's reason with the deadline's. Three helpers that only
construct one error kind now name that kind instead of AnalysisError.

* refactor(hopper): classify original-source read failures through the shared provider classifier (#10838)

* test(ghidra): pin the aborted startup deadline over a producer cancellation

* refactor(kuna): classify Kuna failures through the shared provider classifier (#10843)

Session operations, the workspace copy and the project export report raw
failures through classifyProviderFailure with Kuna dialects; the copy and
export keep their cleanup report shapes through the readCleanup hook.

* chore(guards): shrink the cancellation construction baseline after merging the base
2026-11-15 19:53:32 +08:00
morluto 428e006f66 fix(filesystem): close file handles only through the owned-handle owners (#10578)
* fix(filesystem): close file handles only through the owned-handle owners

Three sites still closed a fs/promises FileHandle in a finally block, where a
failed close replaced the operation's own failure or result: the agentenv
runner head read, the Linux cgroup kernel-file read and the JavaScript
analysis transfer writer. They now use withOwnedFileHandle, which retains the
operation outcome alongside a failed close.

A lint rule, rea/no-raw-file-handle-close, refuses close() on a handle opened
from fs/promises or declared as FileHandle, and fs close/closeSync on a
descriptor, in src outside src/filesystem/Owned*.

* fix(lint): distinguish owned handle class fields

* fix(filesystem): retain diagnostic file cleanup owners

* test(hopper): retry after confirmed application exit

* fix(filesystem): settle doctor peers, sample cgroup outside admission, and bind lint this

Desktop doctor waits for every client inspection before reporting a retained close failure. Ghidra memory cgroup sampling no longer passes the resource scope's admission gate, so a pending owner cannot mask the provider failure; failed closes are still retained for session cleanup. The raw FileHandle close rule resolves this through arrow functions only, treats static blocks as static, and keeps #name distinct from constructor parameter properties.

* fix(ghidra): join in-flight cgroup samples before session cleanup

ArtifactResourceScope.track runs optional work outside the admission gate while cleanup drains still join it, so a cancelled request's final retry cannot finish before a sample retains a failed close.

* fix(filesystem): join work tracked during a cleanup drain and report doctor file cleanup as typed errors

A cleanup retry now repeats until no tracked work is active and every retained owner has had one attempt, so diagnostic work admitted while an earlier owner is releasing is joined. Doctor reports a retained runner or configuration file close as a ProviderCleanupError bound to its lease instead of rethrowing a raw file failure, so the error document keeps the retained file.

* fix(desktop): read the agentenv runner head through the artifact file owner

The runner read retained a failed close under one lease while the reported failure carried another, so a successful cleanup retry left the doctor error reporting an already-closed file. withArtifactFileHandle binds one lease to both.
2026-11-15 18:08:56 +08:00
morluto 5bf66f7500 refactor(domain): one owner for the non-array record predicate (#10588)
* refactor(domain): one owner for the non-array record predicate

Thirty-four source files each declared isRecord as a non-null object that is
not an array, and six more spelled the same predicate as isObject,
isSchemaRecord or isStorageRecord. They now import isRecord from
src/domain/isRecord.ts, whose overloads keep JsonValue narrowing for the
desktop and environment readers. DshCordisPatch's predicate also excludes
Cordis expressions, so it is renamed isPatchRecord.

architecture-guards refuses a new local copy under any name. The seven files
that still hold one are listed and must leave the list when their copy goes.

* fix(guards): count record predicate copies precisely

* test(hopper): retry after confirmed application exit

* fix: reject stale record predicate exceptions
2026-11-15 17:05:54 +08:00
morluto 6a4caa065b refactor(providers): classify raw provider failures through one owner (#10757)
* fix(pcap): report a refused tool launch as an unavailable tool

When a capture read's tool launch was refused with EPERM, pcap reported it
as unreadable tool output (AnalysisProtocolError) instead of an unavailable
tool. Pcap also never told "not executable" apart from "missing".

Add classifyLaunchFailure in src/process as the one owner of which spawn
errnos mean missing (ENOENT, ENOTDIR) and which mean present but not
executable (EACCES, EPERM, ENOEXEC, EISDIR). Pcap, apktool and gdb used three
different errno lists; all three now use it.

* refactor(process): classify owned command outcomes through one owner

Add ProviderOutcome, read once from the supervisor's facts, and classifyProviderOutcome, the one mapping from an outcome to a typed error. Migrate pcap and the shared artifact producer command failure (TShark, sigrok). Add the rea/provider-failure-classifier lint rule with a shrink-only baseline of per-site (cause: unknown) => AnalysisError functions.

* refactor(process): classify Java and adb launch refusals with the shared owner (#10598)

The Ghidra and JADX Java probes and adb's version probe now take
missing-versus-not-executable from classifyLaunchFailure instead of their
own errno checks. ELOOP and ENAMETOOLONG join the owner's missing set, so
adb keeps its existing classification and both Java probes now treat a
non-directory path component as a missing runtime.

* fix(pcap): preserve command launch attribution

* test(pcap): expect an unknown failing stage when a limited capture tool exits unsuccessfully

* refactor(process): classify owned command outcomes through one owner

Add ProviderOutcome, read once from the supervisor's facts, and classifyProviderOutcome, the one mapping from an outcome to a typed error. Migrate pcap and the shared artifact producer command failure (TShark, sigrok). Add the rea/provider-failure-classifier lint rule with a shrink-only baseline of per-site (cause: unknown) => AnalysisError functions.

* test(lint): cover baselined files in the provider-failure classifier rule

* fix(apktool): report the decoder JVM's heap exhaustion as memory, not a target refusal

The supervisor now watches for a managed runtime's out-of-memory report while output streams (JVM java.lang.OutOfMemoryError, V8 'JavaScript heap out of memory'), so a retention budget cannot hide it. ProviderOutcome gains a runtime-exhausted kind and classifyProviderOutcome maps it once. Apktool decode opts in and moves to the shared classifier; Ghidra's single-purpose output token becomes the JVM runtime option and its area matching reuses the shared table.

* fix(process): keep the runtime's out-of-memory report on the memory error it supports

* fix(pcap): attribute a limited capture's exit through the prlimit wrapper

* fix(process): keep a shortened out-of-memory line marked as shortened

* fix(process): identify the refused executable for every launcher shape

Name the executable on errnos Node throws synchronously from spawn
(ENOEXEC, ENOTDIR, ...) and read the Windows native launcher's
executable-launch constraint, so pcap attributes those refusals to the
selected tool or limiter instead of an unknown stage.

* test(pcap): prove synchronous launch naming with ENOTDIR on every POSIX host

* fix(apktool): advise the JVM limit that governs the reported memory area

* fix(ghidra): read the supervisor's parsed OOM area instead of reparsing lines

* fix(sqlite,capture): read V8 heap exhaustion from streamed output, not retained stderr (#10724)

The SQLite worker and the Node capture decoders now declare runtime: "v8", so the supervisor observes V8's heap report as output streams. Both classifiers read the runtime-exhausted outcome instead of matching retained stderr, so a report beyond the retention budget, or one whose process then overflowed its output budget, is still reported as memory exhaustion.

* refactor(evm): classify EVMole worker outcomes through the shared provider classifier (#10716)

* refactor(evm): classify EVMole worker outcomes through the shared provider classifier

The EVMole worker's cancellation, deadline, output budget and CPU/file-size signals now go through classifyProviderOutcome; only the prlimit launch refusal and the reserved EFBIG exit remain in its dialect. The worker is opted into V8 exhaustion scanning so a heap report names its fixed 256 MiB heap. providerOutcomeOf now reads a launch refusal only from a spawn errno, so a private-storage or carrier read errno is no longer reported as a limiter launch failure.

* fix(evm): keep a stream failure's exit status and update Linux worker expectations

* refactor(process): read launch refusals through the launcher identity owner

* refactor(wasm): classify WABT command outcomes through the shared provider classifier (#10727)

* refactor(wasm): classify WABT command outcomes through the shared provider classifier

WABT's cancellation, command deadline, output budget, and inherited-limit signals now go through classifyProviderOutcome; its dialect keeps only the configuration-probe and launch-refusal unavailability and the unclassified-exit output error. A tool ended by SIGXCPU/SIGXFSZ or the memory limit is now a resource constraint instead of an output error, and a refused launch in the producer phase is reported as unavailable. The shared timeout context gains the analysis setting that selected the deadline, and the output-limit context may name its exceeded budget.

* fix(wasm): keep setup guidance when a WABT version probe's output cannot be collected

* refactor(providers): classify raw provider failures through one owner

* refactor(process): build classified cancellations with cancelledBy and the carried caller reason

* chore(lint): count runOwned's close-failure mapper from main in the classifier baseline

* fix(process): carry the runtime's exhaustion report on every memory error that claims it

SQLite and historical-capture memory errors recognised a streamed V8 heap report but dropped it, so a report outside retained output reached callers without its source line or omission flags. One projection, runtimeExhaustionReport, now owns that shape for the shared classifier and both consumers. Also lowers the EVMole and WABT cancellation-constructor baselines.

* fix(process): keep the provider's recovery for a native-thread exhaustion report

The shared classifier replaced the provider-supplied recovery with generic capacity advice whenever the JVM reported it could not create a native thread, dropping Apktool's launcher guidance and flipping retryable to true. The provider's recovery now wins; the capacity advice remains the fallback for providers that supply none.

* refactor(pwntools): classify pwntools failures through the shared provider classifier (#10782)

* refactor(pwntools): classify pwntools failures through the shared provider classifier

pwntools' per-site classifier is replaced by pwntoolsFailureDialect for classifyProviderFailure. Command outcomes go through classifyProviderOutcome, launch refusals through launcherRefusalOf instead of a 'spawn ' syscall prefix check, the input limit through the file dialect's fileSize, and the cleanup report through commandCleanup. The shared read cancellation now carries the reader's partial observation, and resource-signal reports accept provider detail.

* fix(pwntools): report a failed command's output once and expect resource_constraint for overflow

The shared outcome owner already attaches the command's output as capturedOutput, so the pwntools adapter fallback no longer copies the same text into diagnostics.stdout, diagnostics.stderr and diagnostics.captured_output; diagnostics keep the exit facts. Update the overflow expectations the dialect migration missed (layout provider cleanup and over-budget rows, pwntools cleanup) to resource_constraint.
2026-11-15 16:28:42 +08:00
morluto 254bb21857 refactor(browser): send only CDP commands with registered result schemas (#10666)
* refactor(browser): send only CDP commands with registered result schemas

* refactor(browser): open Playwright CDP sessions only through the registered-command owner

* fix(browser): keep the restore obligation for an applied override with a malformed reply

A schema-rejected success reply means the browser applied the Emulation
override, so the attached environment owner records the reset obligation
before reporting the protocol error. Only a rejected or unanswered command
leaves no obligation.

* refactor(browser): open verifier-script CDP sessions through the registered-command owner (#10710)

* fix(browser): register the CDP commands JavaScript verifiers send

The verifier scripts send Browser.close, Page.createIsolatedWorld, Page.navigate, Page.reload, Target.activateTarget, Target.closeTarget and Target.createTarget through CdpConnection, which now admits every reply through its command's result schema. Register those schemas, refuse an unregistered run-time name as a typed malformed reply instead of a TypeError, and check script send literals against the registry.

* fix(browser): keep Target.closeTarget's success member for the verifiers
2026-11-15 14:49:33 +08:00
morluto 1f7884b5b9 refactor(process): classify owned command outcomes through one owner (#10628)
* fix(pcap): report a refused tool launch as an unavailable tool

When a capture read's tool launch was refused with EPERM, pcap reported it
as unreadable tool output (AnalysisProtocolError) instead of an unavailable
tool. Pcap also never told "not executable" apart from "missing".

Add classifyLaunchFailure in src/process as the one owner of which spawn
errnos mean missing (ENOENT, ENOTDIR) and which mean present but not
executable (EACCES, EPERM, ENOEXEC, EISDIR). Pcap, apktool and gdb used three
different errno lists; all three now use it.

* refactor(process): classify owned command outcomes through one owner

Add ProviderOutcome, read once from the supervisor's facts, and classifyProviderOutcome, the one mapping from an outcome to a typed error. Migrate pcap and the shared artifact producer command failure (TShark, sigrok). Add the rea/provider-failure-classifier lint rule with a shrink-only baseline of per-site (cause: unknown) => AnalysisError functions.

* refactor(process): classify Java and adb launch refusals with the shared owner (#10598)

The Ghidra and JADX Java probes and adb's version probe now take
missing-versus-not-executable from classifyLaunchFailure instead of their
own errno checks. ELOOP and ENAMETOOLONG join the owner's missing set, so
adb keeps its existing classification and both Java probes now treat a
non-directory path component as a missing runtime.

* fix(pcap): preserve command launch attribution

* test(pcap): expect an unknown failing stage when a limited capture tool exits unsuccessfully

* refactor(process): classify owned command outcomes through one owner

Add ProviderOutcome, read once from the supervisor's facts, and classifyProviderOutcome, the one mapping from an outcome to a typed error. Migrate pcap and the shared artifact producer command failure (TShark, sigrok). Add the rea/provider-failure-classifier lint rule with a shrink-only baseline of per-site (cause: unknown) => AnalysisError functions.

* test(lint): cover baselined files in the provider-failure classifier rule

* fix(pcap): attribute a limited capture's exit through the prlimit wrapper

* fix(process): identify the refused executable for every launcher shape

Name the executable on errnos Node throws synchronously from spawn
(ENOEXEC, ENOTDIR, ...) and read the Windows native launcher's
executable-launch constraint, so pcap attributes those refusals to the
selected tool or limiter instead of an unknown stage.

* test(pcap): prove synchronous launch naming with ENOTDIR on every POSIX host

* refactor(process): build classified cancellations with cancelledBy and the carried caller reason

* chore(lint): count runOwned's close-failure mapper from main in the classifier baseline
2026-11-15 13:45:46 +08:00
morluto 8c052bdec5 test(architecture): ratchet top-level layer direction and mutual imports (#10584)
* test(architecture): ratchet top-level layer direction and mutual imports

Declare a layer for every top-level src/ directory and root source file
(domain, contracts, foundation, providers, application, composition,
adapters) in scripts/lib/module-layers.mjs. verify:module-boundaries now also
builds the production import graph between those units and compares it with
a counted baseline: 45 unit pairs import a higher layer (248 imports) and 23
unit pairs import each other. A new upward pair, more imports on a recorded
pair, a new mutual import, or an undeclared unit fails; shrinkage is
reported so the baseline can be lowered.

* fix(architecture): enforce layers for isolated units
2026-11-15 08:37:04 +08:00
morluto fd555be8b4 refactor(errors): type 14 baselined resource refusals as fixed ceilings or measured exhaustion (#10639)
* refactor(errors): type 14 baselined resource refusals as fixed ceilings or measured exhaustion

Fixed ceilings (NTFS journal file, web-bundle worker transfer, Playwright trace parser, lipo field retention, exactly representable file sizes) now name their allowance; observed heap or result-capacity exhaustion and occupied admission declare measuredExhaustion. The typed-resource-refusal baseline drops from 24 sites in 19 files to 10 in 10.

* fix(browser): name the trace parser allowance the record actually exhausted

* test(ntfs): compare the oversized-acquisition recovery with its owning formatter
2026-11-14 19:00:38 -05:00
morluto e53212655b refactor(contracts): derive the public tool catalog from one contract-group list (#10564)
* refactor(contracts): derive the public tool catalog from one contract-group list

toolContracts.ts and toolFamilies.ts each imported and listed the same
contract groups. toolContracts.ts now owns one ordered group object;
TOOL_CONTRACTS is its flattened values in the same catalog order, and
toolFamilies.ts reads TOOL_CONTRACT_GROUPS from it instead of re-importing
every group. The static catalog-source lint rule learns the
Object.values(groups).flat() composition so positional-contract enforcement
still discovers every group without executing modules.

* fix(lint): reject computed tool catalog groups
2026-11-15 04:07:17 +08:00
morluto e83e3fa8a0 feat(errors): type refusals at the producer with a total recovery (#10577)
* feat(errors): type refusals at the producer with a total recovery

A resource refusal without authored text fell back to resource-kind advice:
a fixed parser ceiling reported as memory told the caller to add host
memory, an upstream IDA reply ceiling pointed at REA's MCP transport, and a
fixed Kuna text ceiling reported as capacity was marked retryable. An
ambiguous selector was reported as invalid input that pointed at the schema.

Add AnalysisRefusal with three variants built from facts producers already
own: fixed_ceiling (FixedAllowance, now with limit, unit and narrowing),
adjustable_budget (AnalysisSettingLimit) and ambiguous_selector (new
AnalysisInputError option). Each variant is attributed to the error's
operation and projected under details.refusal. Presentation is a total
switch over the variants, and authored text still takes precedence.

Migrate the first producers. GraphQL nesting, Playwright trace input, the
Kuna function text and the IDA upstream reply become fixed ceilings. jadx
overloads and Kuna ambiguous names become ambiguous selectors, replacing the
two Kuna-specific ambiguity texts. Seven occupied admissions declare
measuredExhaustion. The rea/typed-resource-refusal lint rule rejects new
untyped resource constraints, and the remaining 24 are counted in a baseline
that can only shrink.

* fix(errors): honor refusal recovery ownership
2026-11-14 14:36:14 -05:00
gensokyo-devandgensokyo-dev 55c9e8e546 test: close the wait-deadline review gaps
- doctorCancellation measured from the signal, after the probe had
  started, so a probe stopped by its own deadline could still pass. The
  fake probe now records its start, and the check runs from there.
- rea/project-wait-deadline missed literal computed `timeout` keys and
  spread options. It now reports both, as well as computed keys it
  cannot read.
- The testing guide notes that a long-running test's expect.poll keeps
  its project budget, while waitFor follows the tag's deadline.

Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
2026-11-14 19:28:38 +08:00
gensokyo-devandgensokyo-dev 07e4b700ac test: let the deadline policy own test waits
Vitest's expect.poll and vi.waitFor stop after one second by default, a
host-speed guess that the project's deadline policy does not see. Under
hosted-CI load, kunaProjectConcurrency failed that way twice on unrelated
PRs. Another 256 waits relied on the default, and 97 passed literals
picked to fit some host.

Waits now take as long as their test may run, less the reporting reserve
that children already keep. Each Vitest project sets expect.poll's
default, and the new waitFor in tests/support/testWait.ts replaces
vi.waitFor. Fixture readiness waits (awaitAcknowledgement and the provider
process waits) default to the same deadline, and their host-speed literal
arguments are removed. One remainingTestDeadline owns the arithmetic for
both children and waits.

Ten waits whose short timeout was the real check now assert it directly
(Kuna version probe, Cutter discovery, doctor and Inspector cancellation,
apktool fixture teardown, JADX disconnect): settlement before the
production deadline that would otherwise also satisfy the wait, or a
deadline pinned beyond every test deadline.

The rea/project-wait-deadline lint rule rejects vi.waitFor and explicit
expect.poll timeouts. The testing guide and the CI tests audit skill now
document the policy.

Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
2026-11-14 19:28:38 +08:00
morluto b138f52e12 refactor: prune internal seams and preserve configuration ownership (#10430)
* refactor: prune internal seams and preserve configuration ownership

* fix(setup): keep YAML aliases and inherited registration settings on refresh

- Reuse an unchanged YAML collection only while every alias in it still
  resolves after the edit; otherwise rebuild it so the alias becomes its
  value. A sibling anchor replaced or moved in the same refresh no longer
  leaves an unresolved alias in Hermes or Goose output.
- Keep a connection selector whose existing value already chooses stdio
  (for example an inherited Hermes transport: stdio) instead of superseding
  it; remote selectors are still replaced. Explicit stdio selectors are
  therefore no longer rewritten (Kimi expectation updated).
- When the whole Hermes registration is inherited through a merge,
  override it with that merge and write only the changed fields.

* fix(setup): edit registration fields whose YAML key is an alias

A key written as an alias to a string scalar (for example *field with
field: &field env) was treated as a source-only key, so any change to
its value was refused in Hermes and Goose. Resolve such keys to their
string, keep the aliased spelling while its anchor survives the edit,
and write the plain key when the same edit rebuilds the anchor.

* test(process): bound cleanup-boundary waits and always release gates

Both cancellation tests awaited a bare started promise and released
their gates only on success, so a missed boundary waited for the outer
deadline and could leave cleanup parked. Use awaitAcknowledgement and
release and join the cleanup in finally.

* fix(setup): repair only aliases an edit leaves unresolved

Replace the pre-edit alias survival guess, which expanded aliases to
sibling anchors that survive unchanged, with an exact repair: record
what each alias in the replaced value resolves to, write the edit, then
replace only aliases left without a preceding anchor by that value.

- Keep an aliased key's source position, so its anchor stays ahead of
  later aliases.
- Treat an aliased key as its string only when no other key in the map
  resolves to the same string; an ambiguous one stays guarded.
- Override an inherited Hermes registration only when the requested
  entry keeps every inherited field, so an explicit replacement can drop
  remote selectors.

* fix(setup): repair aliases whose binding an edit changes, keeping comments

An alias whose own anchor was removed could resolve to an earlier anchor
with the same name and silently change value. Record the node each alias
resolved to and repair any alias that no longer resolves to that node.
Carry the alias's comments and spacing onto the repaired value.

* test(setup): split the Hermes anchor and merge refresh cases

* fix(setup): refuse removing an overridden key its merge would restore

The merge-retention check skipped keys the mapping also overrides, so
dropping such a key (for example a remote url during an explicit stdio
replacement) let the retained merge restore it. Check removal before
skipping overridden keys, so the edit is refused with the existing
message instead of producing a registration that cannot be applied.
2026-11-14 04:08:37 -05:00
87743d67f5 test: check the deadlines Vitest applies to collected tests
Probes in a source and a built project read each collected test's
deadline, including test, suite, and module long-running tags, so a
change in project inheritance or tag resolution fails directly. Both
the policy and the probes read one set of base deadlines.

Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-14 05:27:07 +08:00
morluto cdc4541ef7 refactor: remove redundant seams and preserve cleanup ownership (#10375)
* refactor: remove redundant seams and preserve cleanup ownership

* fix: validate PE resource extents before budget checks
2026-11-13 10:11:06 -05:00
morluto bbe2bcc979 refactor: remove legacy seams and consolidate resource ownership (#10349) 2026-11-13 04:45:16 -05:00
morluto 7f920465c5 fix(ci): align verification with current lifecycle contracts 2026-11-11 15:10:51 -05:00
N0zoM1z0andgensokyo-dev 91b11f62c9 fix(settings): preserve syntax guard contracts across stack integration
Retain existing merge resolutions for update modes, source-scan failures and external-reader fixtures while replaying the settings layer on the merged inventory. Preserve parent provenance, candidate counts and presentation-test resource bounds.

Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
2026-11-11 19:48:33 +08:00
morluto 3769fe8077 fix(ci): reject mutable setting spreads 2026-11-11 19:39:06 +08:00
morluto 021b5de719 fix(ci): parse runtime setting seeds as syntax 2026-11-11 19:39:06 +08:00
morluto 6e2e5a0aca fix: preserve numeric limit evidence and uncertainty 2026-11-11 19:39:06 +08:00
morluto 694e8c4f7b fix(ci): classify direct numeric value expressions 2026-11-11 19:39:06 +08:00
morluto bc01caea10 fix(ci): remove stale temporary limit exceptions 2026-11-11 19:39:06 +08:00
Cursor Agent 646daa1d29 fix(limits): classify parsed property reads and ignore template prose
A numeric parser now reads property and element accesses in its argument, so Number(process.env.REA_ARCHIVE_ENTRY_LIMIT) is a producer without a schema key. Template discovery keeps generated assignments and no longer treats a diagnostic sentence as a read.
2026-11-11 19:39:06 +08:00
Cursor Agent ead6cc83d2 fix(limits): record construction reads and schema-owned numeric names
A constructor copy is not a per-call read. Numeric limits also come from a number schema or a numeric parse, and the inventory no longer lists path and provider constants.
2026-11-11 19:39:06 +08:00
Cursor Agent 575a6c486e fix(limits): read destructured environment keys and keep the pre-install guard dependency-free
Object bindings, in-checks, and object keys that a spread can copy are producers. The architecture guard skips the TypeScript parse until typescript6 is installed, and the inventory links stay inside the published site.
2026-11-11 19:39:06 +08:00
Cursor Agent b1f8a9bba8 fix(limits): discover numeric limit reads from the syntax tree
Property access, element access, and helper-call arguments are producers. Comments and array literals are excluded by the tree, so a later argument such as read(environment, "REA_NEW_MAX_BYTES") stays classified.
2026-11-11 19:39:06 +08:00
Cursor Agent 7fe465505f fix(limits): keep bracket-access reads in numeric-limit discovery
A quoted name after an identifier, call, or optional chain is a producer. Only declaration arrays and comments stay out of discovery.
2026-11-11 19:39:06 +08:00
Cursor Agent f2d55aba7b fix(limits): ignore allowlists and require tracked limits to be settings
Discovery no longer treats comments or registration-list entries as producers. The three limits moving on their own PRs must join ANALYSIS_SETTINGS; the exception registry cannot classify them.
2026-11-11 19:39:06 +08:00
Cursor Agent 710d2af41c fix(limits): classify endpoint candidates and ignore registry declarations
The numeric-limit check now treats REA_WEB_BUNDLE_ENDPOINT_LITERAL_CANDIDATES as a count limit and does not count registry text as a producer.
2026-11-11 19:39:06 +08:00
morluto 7bcefead69 test: strengthen and prune audited test coverage (#8573)
* test: strengthen and prune audited test coverage

* test: require delivered DeepSeek skill instructions

* test: keep schema fixture types private

* test: verify Gemini receives skill instructions

* test: bind Qwen skill evidence to its tool result

* test: report NativeAOT verifier checks precisely

* fix: verify Ghidra cleanup paths reliably

* fix: verify Command Code skill activation

* test: tighten cleanup and fixture checks

* test: verify packet capture evidence is recorded

* fix: report HTTP3 capture provenance accurately

* test: preserve resource coverage through MCP

* test: prune repeated assertions and retain cancellation identity

* fix: validate client verifier fixture paths

* test: skip Copilot verifier cases on Windows

* test: keep verifier setup target relative
2026-11-02 02:13:32 +08:00
gensokyo-dev e18594b247 refactor(composition): unify capability bindings and cleanup ownership (#8441)
Bind typed service creation, observed availability and retryable cleanup in one composition owner. Reuse selected CLI/MCP factories and canonical contract groups without copying provider operation inventories.

Closes #8429. Refs #740.

Validated paired real application workflows, actual local Codex CLI Terra/high use, complete Evidence and Unknown checkpoint parity, compiler-backed import boundaries, SDK schemas and final real mixed routing. Automated review findings were fixed at their producing owners. Existing exact-base CodeQL alerts are documented on the PR under the owner-authorized unrelated-check exception; no global alerts or security rules were changed.

Assisted-by: Suika (gensokyo-skills:suika-scatter-gather-planning)
2026-11-01 13:30:21 +08:00
morluto 60149282fc test: strengthen integration oracles and verification (#8295)
* test: strengthen integration oracles and verification

* test: bound lifecycle gates and strengthen workflow checks

* test: verify complete paged and extracted results

* test: preserve bytecode and cancellation evidence

* test: strengthen retained evidence and fixture cleanup

* test: require cross-projection bridge identities

* test: retain fixture owners through cleanup failures

* test: verify graph examples and project selectors

* test: tighten provider and lifecycle regressions

* test: consolidate hex coverage and bound cutter cleanup

* test: bound subprocess output waits and strengthen graph checks

* test: check mcp recovery and preserve cleanup owners

* test: verify structured mcp argument failures

* test: preserve artifact cleanup and gdb limit evidence

* test: preserve evidence identities and bound provider gates

* test: retain browser and comparison cleanup owners

* test: bound startup and strengthen fixture oracles

* test: bound lifecycle gates and verify recovered evidence

* test: verify public outcomes and lifecycle ownership

* test: assert precise evidence and release lifecycle gates

* test: verify selected facts and bound cleanup waits

* test: check retained evidence and decoded records

* test: verify final reports and retain cleanup ownership

* test: bound reviewed startup gates and retain cleanup

* test: verify PE errors and preserve MCP cleanup

* test: preserve MCP fixture cleanup ownership

* test: release provider gates before late startup
2026-11-01 00:15:28 +08:00
morlutoandCursor Agent 77b09314ff test(ci): require provider checks and reduce test startup (#8145)
* test(ci): require provider coverage and reuse test sessions

* fix(ci): prepare the runtime before Python producer tests

The fast Python job ran the suite on a fresh checkout, so analysis-efficiency
checks failed without compiled process modules. Install locked dependencies
and prepare the runtime before the suite.

* fix(ci): keep provider test reports inside the checkout

The provider-report reader accepted an arbitrary path from argv. Resolve the
path and reject anything outside the checkout before reading it. The packet
and signal workflows write those reports under .cache.

* fix(ci): require a test module in each named Python bridge family

Discovery still runs every bridge test. The mitmproxy, native, and pwndbg
families must each keep a test module, so deleting one cannot pass through
the others.

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
2026-10-30 19:34:19 +08:00
morlutoandCursor Agent 249b90c5e3 refactor(javascript): break the semantic analysis import cycle (#8063)
* refactor(javascript): break the semantic analysis import cycle

Split shared semantic facts out of the analysis hub so projectors of
effects, evaluation, and identity no longer import one another in a cycle.

* refactor(javascript): share the import specifier reader

Read static imports in one module used by the boundary checker and both
cycle guards, so those rules cannot drift apart.

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
2026-10-30 06:07:49 +08:00
morluto 365fb03e73 chore: archive analysis efficiency experiments 2026-10-27 14:10:07 +08:00
morluto c8269ca497 fix(ci): own mock isolation and normalize Git patch directories
Convert native directory separators only at Git patch argument production. Exercise fresh, applied and invalid dependency patches in private copies, preserving protectNTFS, and select the regression for Windows CI.

Replace the spelling-based mock inventory with one binding-aware Oxlint rule. Canonical test projects own isolation for both Vitest and lint; external temporary CLI probes avoid concurrent test discovery.

Validated actual local Git and Oxlint execution, resolved-policy parity, CI selection and report tests. Native Windows verification remains with CI. Refs: #7050
2026-10-25 09:49:26 +08:00
morluto ab720b4216 test(evaluation): preserve partial coverage in the Notes oracle 2026-10-25 09:07:57 +08:00
morluto be698d4761 test: verify assertion contracts and the production evaluation gate 2026-10-25 08:27:16 +08:00
morluto 42c16ba631 fix(tests): resolve lint plugins in boundary regression 2026-10-25 08:18:09 +08:00
morluto 76afdd35ae chore(lint): guard validation ownership and asynchronous assertions 2026-10-25 07:57:07 +08:00
morluto 92f32981da refactor(mcp): remove the duplicate result delivery class
Construct the canonical ToolResultDelivery through the native continuation policy factory. Move type and lifecycle spy consumers to the core owner while preserving explicit response budgets and the deferred runtime composition.

Validation: runtime build and 335 focused tests passed; 41 Linux-only firmware cases were skipped on macOS. Tests use authored local fixtures, not real external providers.
2026-10-25 02:28:09 +08:00
gensokyo-dev 9895f1870c test(evaluation): find failure gate within cleanup callbacks (#6988) 2026-10-24 10:50:50 +00:00
morluto d9660eb032 fix(verifiers): share package command and workspace ownership
Join canonical owned command cleanup before releasing package verifier workspaces. Replace raw package command wrappers, the updater supervisor, WASM cleanup composite, and boolean-only retention with explicit capability ownership.

Preserve selected environments, npm arguments, no-deadline calls, clean exit status, and distinct aggregate/per-stream output budgets. Exercise actual timeout descendants, cancellation, retained cleanup retry, and workspace admission.
2026-10-24 16:58:18 +08:00
morluto 3ad465a666 refactor(evaluation): move agent grading into development ownership
Keep the maintained manual evaluator, fixture rubrics and semantic tests together under tests/evaluation/model-evals instead of shipping them in dist. Load the same typed owners from the manual runner and update the architecture and client guide. Grading/report contracts and executable logic are unchanged.
2026-10-24 11:50:11 +08:00
morluto d3700f709b fix(evaluation): align scenario prompts and outcome gates 2026-10-24 02:36:00 +08:00
86655eac66 fix(hopper): keep automated Linux setup off non-x86_64 hosts (#3422)
* fix(hopper): keep automated Linux setup off non-x86_64 hosts

Every pinned Linux Hopper package (deb, rpm, Arch) is x86_64-only, but
hopperSetupSupported was derived from os-release alone. Setup therefore
planned and confirmed an amd64 package install on arm64 Ubuntu 24.04 and
the other tested distributions.

Classify installer support with the host architecture where the
distribution is parsed, so setup planning, preflight, installation and
package verification share one decision. Setup now takes its Linux
distribution from the doctor host, as it already does for the macOS
version, and the supported-host remediation text has a single owner.

Fixes #3419

* fix(hopper): install Linux Hopper from the setup host's distribution facts

systemLinuxHopperInstallHost read os-release with process.arch on its own,
while setup planning used the doctor host's injected architecture. Both
agree in production, but a doctor host composed with another architecture
could plan one answer and install with another.

The install host now takes its distribution reader from the caller, and
systemSetupHost passes the doctor host's, so planning, preflight and
installation share one host-fact owner. The unused default host on
installLinuxHopper, the other ambient reader, is removed.

* fix(verify-package): require Hopper readiness only where Linux setup installs it

packageSetupSupport required doctor Hopper readiness on every Linux host
whose host check passed, including installer-untested distributions and,
after the architecture gate, arm64 hosts where setup never provides
Hopper. On Linux, readiness now also requires automated installer
applicability; macOS is unchanged.

This narrows the readiness assertion from #2973 on hosts outside the
installer matrix, such as Kali and ubuntu-24.04-arm.

---------

Co-authored-by: DavidFB-creator <220631175+DavidFB-creator@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
2026-10-23 02:39:05 +08:00
morluto 79dea89f1f fix(evaluation): own transcript result parsing and error certainty 2026-10-22 17:00:42 +08:00
70cd4cbfb8 fix(hopper)!: verify current Linux releases with the publisher key
Replace fixed Linux Hopper release pins with current vendor feed metadata,
complete detached-PGP package verification under the stable publisher primary
key, and authenticated package evidence for offline admission. Share parsed
publisher facts across approved setup, doctor and the pre-execution guard.
Preserve existing ownership, cancellation, private-display and cleanup rules.

Fixes #3577

Normalize empty XDG settings, preserve backups during approved cache repair,
and retain publisher cleanup through cancellation and shutdown. Exercise cold
offline verification without reusing process-local proof state. Serialize
approved cache publication across processes, restore old evidence on failure,
and share the existing lease owner from process without changing its identities.

Validation: 438 focused tests (six platform-related skips), full static/documentation checks, real Ubuntu
24.04 official Hopper 6.6.0 setup/CLI/MCP, offline ordinary-user admission,
launcher mutation rejection, and cold/warm verification measurements. Official
RPM metadata mismatch is tracked separately in #3578 and fails closed.

BREAKING CHANGE: Older Linux Hopper builds admitted by the fixed allowlist
now require retained publisher-signed package evidence or an approved update
to the current public demo. Update explicit HOPPER_LAUNCHER_PATH values when
migrating; the live feed exposes only its current release.

Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
2026-10-15 12:15:55 +08:00