* fix(process)!: retain observed cleanup intervention
Process capture derived `settlement.cleanup_outcome` from the settlement
state alone in three places, so a quiesced capture reported
`not_required` even after final owned-process cleanup SIGKILLed a live
token-owned descendant. `cleanupOwnedProcessGroup` already knew it had
signaled, but the capture lifecycle discarded that fact.
Carry the observed signal out of ownership cleanup (also on failure
paths), accumulate it across cleanup retries as true/false/unknown, and
derive settlement and the intervention residual unknown once in
`processCaptureCleanupFacts` for full, partial and harness results.
BREAKING CHANGE: a quiesced settlement may now report
`cleanup_outcome: "cleaned"`. Saved captures stay valid; readers that
accepted only `not_required` for quiesced settlements must accept
`cleaned`.
Fixes#8032
Squashed from #8043 and rebased onto main (adapted to the resource-scope
release API and the inspection-failure cleanup path).
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(process): keep a sent signal when cleanup throws an inspection failure
Final ownership verification can rethrow a typed inspection failure (for
example cancellation) after SIGKILL succeeded. The capture owner then
recorded the attempt as unknown and dropped the intervention fact.
cleanupOwnedProcessGroup now marks a failure it throws after signaling,
without changing the failure's identity, and the capture owner keeps
the signal for such failures.
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: wunianze666-netizen <wunianze666@gmail.com>
Co-authored-by: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The testing guide requires file-owned createMcpSchemaValidation instances
to be cleared in afterAll, so shared workers do not retain schema strings
and compiled validators after a file finishes.
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After tools/list, the MCP SDK client builds an output validator for every
listed tool before its first tools/call. Against REA's 255-tool catalog
that compiled for about 15 seconds of CPU (5 seconds for
inspect_analysis_view alone), whichever tool a test called. The shared
test validation now snapshots each schema when asked and compiles it when
it first validates.
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* refactor(cli): run one-shot commands through one shared owner
scripts/rea.mjs and the test CLI session worker each implemented the
one-shot command sequence: argument normalization, output admission,
configuration errors, version documents, the Amp scope refusal, and
result delivery. The worker copy had drifted: it never rendered
formatted version reports or the Amp refusal.
Move the sequence into src/cliOneShot.ts. Both callers supply a host
for their streams, exit and resource ownership. The launcher keeps
closed-pipe status, worker shutdown and process exit; the session
worker captures output per command.
The cliOutput bundle entry is superseded and removed. The dispatcher
test now replaces only the command runtime, so output admission runs
production code. A new parity test compares the file session with a
fresh process.
* test(cli): bundle the one-shot entry in the runtime bundle fixture
* perf(tests): run plain CLI commands on the file's shared process
`cli.run` in `cliTest` files started a fresh compiled CLI for every
command, about 0.45s and 260 MB each before any work. The file-scoped
session already existed, but callers had to opt in through
`runFileCli` or `sharedCliRun`, so 1,300 call sites still spawned.
Make the fixture choose: `cli.run` uses the shared process when it
reports the command exactly as a fresh process would, and starts one
otherwise (stdin, explicit deadlines, MCP and MCP doctor routing,
NODE_OPTIONS, or enabled logging). `freshProcess: true` keeps
explicit process-level checks. The session worker now presents the
launcher's entrypoint, so registration identity matches a fresh
process.
`runFileCli` and `sharedCliRun` are superseded and removed; their
call sites use `cli.run`, and paired fresh/shared checks use
`freshProcess`.
* style(tests): format the managed member summary call site
* test(cli): give each shared-session command an empty stdin
The shared CLI process reads session frames from its own stdin. A
command that reads stdin without supplied input, such as
`compare-wasm-interfaces - --json`, read that pipe and hung, where a
fresh process receives EOF. The worker now presents an ended, empty
stdin to each command and restores its own afterwards.
The parity test compared the routed path with itself after cli.run
began choosing the session. It now compares a fresh process with the
session directly, including the `mcp add` refusal and a command
reading absent stdin.
* test(cli): keep default command logging and launcher checks on fresh processes
environment-serve and environment-build log at info when REA_LOG_LEVEL is
unset, through a logger that writes to stderr directly. The shared CLI
process captures only process.stderr writes, so routed runs lost those
lines. Declare the default-logging commands beside the command names,
have infoLogger read that set, and start a fresh process for them.
The log-level precedence checks compared the shared process with itself
after routing; their fresh side now passes freshProcess: true.
These fixtures' descendants exit on their own after 20 or 30 seconds. The
shorter explicit deadline is what proves cleanup stopped them first.
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Registries created directly with createTestProcessRegistry did not get the
processTest default, so a hung child there had no deadline once its explicit
value was removed; Vitest's test timeout does not settle the awaited run, so
shutdown was never reached. Derive the default in the registry from the
running test instead, using the real clock.
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
239 literal timeoutMs values on cli.run and processes.run only restated
a guess at host speed. On hosted runners with four workers, cliOutput's
3-second value failed while the command still produced correct output.
Deadlines that are the behavior under test, derived from an operation
budget, or that bound a regression's wait remain.
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A processTest child run without its own timeoutMs now gets the test's
remaining deadline less a reporting reserve (at least 5 seconds, or a
tenth of the deadline). A hung child still fails with its captured output,
without a fixed per-call value that assumes a host's speed.
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Antigravity keeps its MCP configuration and skills under
~/.gemini/config, and Gemini CLI was detected by the bare ~/.gemini
directory. Setting up Antigravity therefore made broad setup select
Gemini CLI and write a registration nobody asked for.
Gemini CLI's applicability now follows its own evidence: settings.json,
its trusted-folder records (honoring GEMINI_CLI_TRUSTED_FOLDERS_PATH) and
its .gemini/tmp session state, through the existing discoveryPaths
owner that detection, status and uninstall share. Selecting
--client gemini_cli explicitly is unchanged.
Fixes#7567
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Linux Hopper setup verifies the package signature, then compares the
launcher it contains with the release feed's main_exe_hash. When only
that comparison failed, as with the Fedora 6.6.1 RPM, setup reported the
generic integrity failure and advised retrying or updating REA, which
cannot change a stale vendor feed.
The publisher now reports that case as launcher_digest_mismatch. Setup
keeps the public integrity_mismatch code and the refusal, and its
recovery says the download is authentic, the feed describes another
launcher build, and the mismatch needs a corrected feed. The two digests
stay in the detail. Every other integrity failure keeps its recovery.
Fixes#8138
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Doctor and configuration failures for a missing or rejected
REA_JADX_MCP_JAR asked for a "caller-supplied" jadx-headless-mcp 0.7.1
JAR without saying where it comes from. The upstream project has since
renamed its packages, so users concluded the JAR could not be obtained.
The shared JADX JAR remediation now names the audited release URL and
SHA-256 that REA already pins. The two capability-availability texts
point to rea doctor --provider jadx for them instead of repeating the
release data outside its owner.
Fixes#7768
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`rea doctor --target ""` passed an empty string to target readiness,
which checked the working directory instead of reporting that no app
was named. Use the shared local path schema, so an empty target fails
validation before any readiness check, as every other target does.
Fixes#7561
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On Windows `where rea` lists the npm shims rea, rea.cmd and rea.ps1 that
one global install writes side by side, and doctor counted each listed
path as an installation, so a single install reported "multiple".
The system host now pairs each PATH launcher with the installation it
resolves to: its directory after symlink resolution, compared
case-insensitively on Windows. Doctor counts distinct installations and
still lists every launcher path. Two POSIX launchers that resolve to one
package also count once; an unresolvable launcher keeps its own path.
Fixes#8122
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Report identity as unobserved when the selected tool families omit
binary_session, preserving the current health decision and exit status.
Cover scoped production sessions and retain validation of successful
and schema-invalid probe responses.
Fixes#7438
[gensokyo-dev: rebased onto main; report "identity not observed" for every
request-flow reply that was not received (skipped scope, transport failure or
rejected output schema), not only skipped scopes; corrected the replayed
inventory-mode expectation, whose identity is observed and matches.]
Fixes#7438
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The owned transport is the doctor's only cleanup owner, and Client.close
only delegates to it. Close it once and report its typed failure instead
of aggregating a duplicate SDK close and clearing that failure after a
retry. Pass the filtered caller environment as the launcher's base only,
and name the doctor operation in transport cleanup errors.
Cover a descendant in its own session, which only the token sweep can
reach, and a late cleanup denial that keeps the startup diagnostic and
reports the retained run. Bound the fixture descendant's lifetime so a
failing run cannot leak it. Document the doctor's process-family
ownership and output bound.
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The owned launcher merges its host base back under the explicit overrides, so
filtering only the override object let an ambient HOPPER_TARGET_* variable
through and the diagnostic child could open the target the doctor promises to
skip. Pass the filtered environment as the host base too, and opt into the
token sweep so a detached descendant that outlives the launcher is still
reached by cleanup.
Also keep the stderr observer attached until bounded cleanup and draining have
settled, report the typed cleanup resources instead of a generic close_binary
message, and drop a first-attempt cleanup failure once the owned retry proves
release.
runProductionMcpDoctor wrapped the SDK StdioClientTransport and only joined its
close calls. That transport clears its direct-child reference, ends stdin and
signals only that child, so a server whose launcher exited after forking an
inherited-stdio descendant left the descendant running while the doctor
reported nothing but the transport deadline.
Run the doctor on OwnedMcpStdioTransport, which spawns the child in a dedicated
POSIX process group or Windows job and verifies group and token-owned
descendant cleanup before close returns. Feed stderr through a sink on
acquisition so diagnostics survive an early child exit and still honor the
64 KiB retained prefix and full byte count.
Covered by a fixture that forks a descendant, exits, and is timed out: the
regression asserts the descendant is gone after the doctor returns.
Replace the simulated-host sweep test with one that parses PF_KTHREAD
from a controlled /proc/<pid>/stat record through the system ownership
host, with root-matching credentials and an ESRCH environ read. The
user-process control with identical facts stays fail-closed. Name the
task flag constants and correct the comments: the reported kernel
threads were live, not exiting.
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An unrelated kernel thread that appeared during a provider run has no
readable /proc/<pid>/environ, so the ownership sweep recorded it as a live
candidate with an unreadable token and failed cleanup with
cleanup_incomplete. Kernel threads exit quickly, so the read usually
reports ESRCH and is not attributable to environment withholding, which
meant a cleanly finished provider run discarded its result.
Kernel threads are forked by the kernel, never by a provider, so they can
never carry that run's token. Read PF_KTHREAD from the same /proc/<pid>/stat
identity observation and treat a kernel thread with no readable token as
unverified evidence, like an unrelated process whose environment the OS
withholds. The candidate still has to be otherwise unrelated and holds no
signal authority.
A terminal call on one branch does not end the whole function: other
branches can still reach the following code, which stays in the body and
pseudocode. The function-level limitation now says Ghidra excluded that
call's fall-through and that code reachable only through it is missing
if the callee returns. It is listed before the call-specific facts so a
bounded composed summary that admits limitations in order keeps it.
The body-coverage tests now check the counted records and their reasons
instead of freezing the explanatory sentence.
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ghidra's Non-Returning Functions - Discovered analyzer can mark a
returning callee, such as an imported operator new thunk, as
non-returning. Ghidra then clears the fallthrough at each call site and
refits the callers, so a constructor decompiles to a few plausible
lines. REA reported the terminal call with the FunctionManager flag but
did not say the result omits the code after it or how to compare.
That analyzer bookmarks each entry it marks with the Analysis comment
"Non-Returning Function Found" (Ghidra 11.0 through 12.1.4 source); the
name-list Known analyzer writes "Non-Returning Function Identified".
When a terminal call's callee, or the function a thunk forwards to,
carries the Discovered bookmark, the call's limitation now says the
flag was inferred, and the function adds one address-free limitation
that its body and pseudocode end at that call and omit the following
code if the callee returns, with the REA_GHIDRA_DISABLED_ANALYZERS
comparison. A flag without the bookmark keeps an unknown origin.
The function-boundary observation contract moves to v5 so snapshots
from earlier bridges are not replayed without the qualification.
Refs #10365
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
decompile_all set its top-level "procedure_info is unavailable" caveat
from a per-call flag that only a newly attempted function could raise.
A resume that skipped every verified file, or selected nothing, rewrote
the manifest and decompile-index.json with an empty limitations list
while every retained record still reported an unavailable body.
The export owner now answers from the records it retains: the summary
counts every non-external record whose body is unavailable, including
records an earlier run wrote, and repeats the reason when all share one.
The per-call flag is removed. A fresh export whose procedure_info works
but leaves some bodies unavailable now reports that caveat too; a later
attempt that records a known body clears it.
Fixes#8351
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The handler regression added in #10390 listed tools under the full input
profile and matched every advertised schema before calling each tool. It
took about 14 s locally and passed the 30 s deadline on loaded CI runners.
advertisedPatternsVMode.test.ts already proves that no input or output
profile advertises the surrogate escape, so this case now checks only the
runtime refusal through the default-profile handlers, as the
relative-path case does.
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rea mcp --http printed "Choose another --port or stop the process using
it" for every listener failure except a missing host allowlist, so a host
lookup failure, an unassigned address or a permission denial got port
advice. StreamableHttpListenError already classifies the native failure
and owns a recovery for each kind.
Startup now always prints that owned recovery behind one shared lead-in,
and the fixed message is removed. The native cause text stays out of
stderr as before.
Fixes#7346
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Planning for win32 refuses PTY capture before root admission runs, so the
cancellation cases now plan for Linux on any host. The ENOTDIR admission
case skips on Windows, which reports ENOENT for a path below a file and
admits it as a root the scenario may create.
Squashed from #9306 onto current main.
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Ghidra profile fixture simulates a Linux host, so posix.isAbsolute
refused the C:\ install directory a Windows test run supplied. The
fixture now owns a POSIX-shaped install directory with shared ownership
and cleanup, used by the DOL, MIPS, PowerPC, RISC-V and language
override suites.
Squashed from #9284 onto current main.
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The snapshot identity test writes its file and admits it at once with the
default ChangeTime mode. On Windows that ChangeTime is inside the 2 s drift
window, so admission keeps re-reading the descriptor until it settles. The
returned baseline is then a later observation, not the one the test
captured, and the descriptor stat runs more than once.
Admit the fixture as written. That is the documented mode for a file that
was just written, and it reads the descriptor once on every host. Settling
keeps its own coverage in ChangeTimeBaseline.test.ts.
The shared reader lifecycle case waits for a member's writable side to
finish before it reads anything. That only happens while the member fits
below the readable high-water mark. Node's default is 16 KiB on Windows and
64 KiB elsewhere, and the ESP flash fixture's first member is a 32 KiB
bootloader region, so both cases hung until the 30 s timeout on Windows.
Pick the first nonempty member that fits the default buffer and assert
that it fits the opened stream before waiting.
* refactor(cli): run one-shot commands through one shared owner
scripts/rea.mjs and the test CLI session worker each implemented the
one-shot command sequence: argument normalization, output admission,
configuration errors, version documents, the Amp scope refusal, and
result delivery. The worker copy had drifted: it never rendered
formatted version reports or the Amp refusal.
Move the sequence into src/cliOneShot.ts. Both callers supply a host
for their streams, exit and resource ownership. The launcher keeps
closed-pipe status, worker shutdown and process exit; the session
worker captures output per command.
The cliOutput bundle entry is superseded and removed. The dispatcher
test now replaces only the command runtime, so output admission runs
production code. A new parity test compares the file session with a
fresh process.
* test(cli): bundle the one-shot entry in the runtime bundle fixture
The test bounded each hdiutil and CLI step with its own 10-30 s process
deadline. On hosted macOS runners the test takes 32-45 s (about three
times a local run), and the busy-image CLI call, which verifies the whole
image before its attach is refused, exceeded 30 s and was killed in two
recent main runs. Remove those per-step deadlines; the project deadline
still bounds a hang.
* test(mcp): open stdio clients through the shared local-tools fixture
Twelve files built their own StdioClientTransport and Client for
`rea mcp`. Those clients never listed tools, so the SDK did not
validate structured results against advertised output schemas, and each
file repeated transport setup, stderr draining and close ordering.
Use connectLocalToolsMcp, which lists tools, validates replies,
silences logging, leaves deep providers unconfigured and owns cleanup.
Per-test clients register test-scoped cleanup; module-level clients
collect cleanup for afterAll.
* test(mcp): keep failed file-scoped MCP cleanup owned
The three module-level clients removed every cleanup callback before
running them, so a failed cleanup lost its only handle. Add
fileLocalToolsMcp, which connects before the file's tests and removes
each cleanup only after it succeeds, and use it in those files.
* perf(mcp): build advertised tool schemas once per contract
The MCP SDK reprojects every registered tool schema on each tools/list.
REA memoized that projection, but per wrapper, and each server built new
wrappers for every contract, so each new server reconverted the whole
catalog: about 280 ms per tools/list for 255 tools with the full input
profile. Each wrapper also registered a new schema in zod's global
registry, so heap grew with every server.
Contracts are immutable module values, so build each advertised input
and output wrapper once per contract. tools/list on later servers drops
to about 26 ms (full profile) or 48 ms (compact).
* test(mcp): load only the contract family under test
* docs(ci): move dated shard measurements into a verification record
The testing guide and testing-lanes guide mixed the October 2026 shard
budget and stale-history measurements into current rules. Move them to
docs/verification/ci-shard-scheduling-2026-10-20.md and link it from
the rules they justify.
* docs(ci): drop dated shard measurements instead of keeping a record
After REA stopped a Rizin debug session (frame or output limit, deadline,
cancellation, protocol failure) or the debugger exited, later commands
returned a generic invalid_request and status gave no reason. GDB
collapsed the same states the same way.
Both managers now keep a published session addressable until
close_*_session releases it, record the first reason it stopped, report
it as closed_reason in session status, and refuse later commands with
capability_unavailable naming that reason and the close-then-restart
recovery. An ID the server never returned, or that close released, still
fails as invalid_request with its own recovery. Retained stopped
sessions keep their slot, which the capacity recovery now says. A GDB
command deadline now records its reason like cancellation already did.
Fixes#8024
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rizin exits 0 when its shell refuses a statement, so execute_rizin_command
reported completion_status complete with exit code 0 and an empty stdout
that read as a genuine empty result.
Rizin 0.6 through 0.9.1 log exactly four refusal lines from
librz/core/cmd/cmd.c: an unknown command, a wrong argument count, an
invalid statement (RZ_CMD_STATUS_INVALID, which also stops the remaining
statements) and a parse error. The Rizin command profile now classifies a
completed process's stderr against those lines; the shared one-shot
service reports completion_status "rejected" and a limitation that names
the refusal and says stdout holds only earlier statements' output. Other
ERROR lines come from commands that ran and stay diagnostics.
Fixes#10364
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`cliIntegerArgument` piped a regex-guarded string into a number but never
published Zod's input view, so `--schema` advertised a bare integer for 15
fields while the parser accepted only decimal argv text. `environment-serve
--port` had the same gap through the transforming MCP_PORT owner.
Remove `cliIntegerArgument` and route every call site through
`cliDecimalField`, binding each field to its domain or contract owner where
one exists so bounds, defaults and optionality come from that owner. Export
the input-view publisher as `publishArgvInput` and apply it to `--port`,
which keeps the MCP_PORT grammar shared with `rea mcp --http`.
The decimal catalog test now requires every numeric field's advertisement
to admit its decimal argv token, covers repeated options and
string-advertised token grammars, and checks `--schema` output directly.
The string branch's missing numeric range (#9348) is unchanged.
Fixes#9344
Co-Authored-By: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Passing runs of the macOS artifact lane took 10.2-14.9 minutes (median
13.6) across recent runs, and 4 of 12 runs that reached their steps were
cancelled at the 15-minute limit without a failing test. Allow 25
minutes so runner speed alone does not fail the lane.
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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>
Four wide-input regressions used a short per-test deadline as their only
check against a slow but terminating implementation. Assert the same
budgets explicitly so the project deadline only detects hangs. The
composed-body test keeps its exact read count instead.
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Explicit deadlines at or below the project deadline only shortened it:
a literal 30-second value also disabled the coverage scaling, which timed
out wide-capacity regressions on main. Workloads above the built
deadline use the long-running tag. Deadlines derived from an operation
budget remain.
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Source projects allow 30 seconds and built projects 120 seconds, both
doubled under coverage, and hooks share their project's deadline instead
of Vitest's 10-second default. A long-running tag covers bounded
workloads that need more.
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(javascript): give the wide fingerprint regression a capacity deadline
Each case runs about 22 s of CPU-bound projection on a typical hosted
runner. Its explicit 30 s deadline failed on slower runners at both two
and four Vitest workers, and it also overrode the suite's 60 s
coverage-mode allowance.
* test(ci): double test deadlines on hosted runners
Hosted CI jobs run four Vitest workers on a four-core runner. CPU-bound
capacity cases then share every core with other workers and their
subprocesses; three such cases exceeded the 30 s deadline at four
workers (16.8-22.7 s at two). The deadline only needs to bound a hang,
so hosted CI doubles it: 60 s, or 120 s with coverage. Local runs keep
30 s and 60 s.
* test(ci): check worker and deadline policy for local and hosted runners
The policy test loaded vitest.config.ts in a child that inherited the
parent's GITHUB_ACTIONS, so on CI it observed the hosted deadlines and
failed. Select the runner explicitly and check both: two workers and the
base deadline locally, four workers and twice the deadline on hosted CI.
