* 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
* refactor(application): own the finalization precedence in runOwned
runOwned runs owned work, closes its owner and applies the documented
precedence once: a recorded failure stays primary, cancellation before
commit keeps the completed value as a partial observation, cancellation
after commit keeps the value, and unconfirmed cleanup reports
cleanup_incomplete while keeping the completed value. finishAnalysisCommand
now projects its Finalized outcome; its 31 callers are unchanged.
* fix(application): preserve finalization outcomes
* fix(application): classify finalization by the primary error and keep cancellation through recovered publication
A producer composite whose primary is a cancellation finalized as failed because the kind read only the top-level error class; it now follows primaryAnalysisError. A cancellation that arrived while a recovered value's publication failed was dropped because the signal was read before that last await; it is now attached to the publication failure as a secondary.
* fix(cli): keep the initial failure snapshot whenever close itself succeeded
* refactor(parsers): report format failures through one ParseFailure algebra
Add ParseFailure (malformed, truncated, unsupported, limit, with format,
offset and field) and its one carrier, ParseFailureError, in the domain.
BinaryReader now names truncation and the field it read, and its native
format and resource failures are ParseFailureError subclasses.
Providers projected a format failure with their own instanceof ladder,
and some dropped the selected path from the invalid_format message.
parseFailureAnalysisError now owns that projection for the native,
Go, devicetree, V8, Windows Shortcut and Electron fuse providers.
The V8 code cache, Windows Shortcut and Electron fuse format classes
are removed in favour of the algebra.
The architecture guard counts raw Buffer and DataView reads, plain
Error throws in parser folders and plain JSON.parse of parser input
against baselines that can only shrink.
* fix(parsers): stop reading every RangeError as a format failure
Mach-O, Swift and Objective-C metadata parsers threw RangeError for a
truncated or malformed record, so their catch sites treated any
RangeError as a partial decode: a call stack overflow or a host string
limit became an "unresolved" metadata reason, and the IL2CPP Mach-O
path reported it as a malformed container.
Those parsers and the CIL method body reader now throw ParseFailureError.
hostRangeErrorKind is the one interpreter of a host RangeError (stack
overflow, host capacity, raw out-of-bounds read), and
parseFailureFromCause admits only a parse failure or a raw
out-of-bounds read as evidence about the input. Fixup stream catch-alls
rethrow host limits. The architecture guard counts the remaining direct
instanceof RangeError checks.
* fix(parsers): preserve Apple metadata failure ownership
* test(architecture): shrink parser failure baseline
* fix: preserve uncertainty for host range errors
Migrate inspect-go-binary, inspect-v8-code-cache, inspect-devicetree and
inspect-native-exports onto the contract command description and delete
their hand-written options and input spreads.
The description gains evidenceDelivery. It adds the shared --detail and
--evidence_output options, and the run context carries the selected
delivery. A contract detail field selects that delivery and is not passed
to the analysis, as the hand-written commands did. Go's cross-field
budget rules stay with the service's contract parse, so standalone
--max-string-literals still returns invalid_request.
The parity guard accepts the declared CLI-only delivery options as the
only options without a contract field.
* docs: refuse literal tool counts in living prose
docs:check now refuses a count of ten or more tools in README, AGENTS,
CONTRIBUTING, docs/ (except dated releases, specs, verification and ADR
records), the authored skill and the public website. The catalog owns that
number and every hand-written copy goes stale when a tool is added. The three
current statements (240, 172 and 254 tools, all measured on earlier builds)
now refer to the catalog of the measured build instead.
* fix(docs): exclude generated pages from prose checks
* fix(docs): rerun the tool-count check when any scanned prose changes
The docs:check:built cache inputs and the CI docs lane did not cover AGENTS.md, CONTRIBUTING.md, every .agents/skills tree or website/public, so an edit there could reuse a cached pass or skip the docs job. The scanner now owns one isLivingProsePath predicate; the CI plan selects the docs lane with it, and a test binds the turbo inputs to the scanned roots.
* refactor(cli): derive contract-backed command options from the tool contract
Each contract-backed CLI command restated its contract's input fields as
hand-written incur options and mapped them back to snake_case input with
conditional spreads. Nothing checked that the two schemas agreed.
Add registerContractCommand: a per-command projection names only what the
CLI spells differently (positionals, caller-relative path fields, option
renames, CLI help wording, omitted optional fields with a reason). Option
schemas, decimal argv tokens and the contract's cross-field rules are
derived from the contract input schema, and parsed options map back to the
contract input mechanically.
Migrate inspect-sqlite-database and the thirteen flat binary-diagnostics
inspectors. Their --schema output is unchanged apart from property order,
and help lists options in contract order. Delete the SQLite superRefine that
mirrored the contract and the per-command spreads.
Add a parity test: every projected command exposes each contract field or
declares its omission, and each field admits exactly what the contract
admits for empty, NUL, relative-path, bound and numeric-spelling probes.
* refactor(cli): refuse unprojected contract input forms at registration
Address review findings on the contract projection:
- Bind inspect-sqlite-database to its contract by name instead of catalog
position.
- Accept only field forms argv can carry (string, number, boolean, enum,
and arrays of strings or enums) instead of rejecting a short list, so
arrays of objects, tuples and other forms fail at registration.
- Refuse root input keywords other than the projected object constraints
and presentation annotations, so a contract rule such as `not` cannot
disappear from --schema.
* refactor(cli): describe contract commands independently of the CLI framework
Split the contract projection into a framework-neutral description and a
thin Incur adapter, so a later REA-owned CLI layer (#10583) can render the
same description.
- contractCliDescription.ts derives each positional and option from the
contract and projection: key and short flag, help, argv admission
schema, boolean-flag/repeatable/decimal/caller-path facts, object
constraints renamed to CLI keys, the optional argv-time cross-field
check, and the mapping back to contract input.
- contractCommand.ts only renders that description as Incur args and
options and runs the handler.
Cross-field rules now refuse at argv parsing only when a projection asks
(inspect-sqlite-database does, keeping its behavior); otherwise the
service's contract parse refuses them as before. Exact-optional fields are
projected as optional because argv cannot carry an explicit undefined and
Incur's help reads flag and choice types only through optional.
The parity guard reads the descriptions, and also checks that boolean
fields parse as bare flags and that no projected field stays
exact-optional.
* test(mcp): check advertised schemas against a portable dialect in every profile
* fix(mcp): validate advertised dialect by selected profile
* fix(mcp): refuse zero-padded repeat bounds and record host-specific path patterns
* refactor(errors): require a reason for every analysis cancellation
Add cancelledBy, the one constructor for AnalysisCancelledError. Its
reason argument is required: an aborted signal, an earlier cancellation,
a reason carried from an inner failure, or an OwnerCancellation when REA
closed the work's owner itself.
Closed adb, apktool, firmware, sqlite, gdb, web-bundle analyzer and
binary-session record owners now report OwnerCancellation.ownerClosed
instead of an absent reason. Firmware's predecessor wait returns its
cancellation as a Result, removing an unreachable reason-less fallback.
An architecture guard rejects new direct constructions, with a counted
per-file baseline that can only shrink.
* refactor(errors): share one aborted-signal predicate with cancelledBy
* refactor(process): return a departing caller's reason from a shared startup wait
* refactor(providers): build decompiler-provider cancellations with a required reason (#10587)
Migrate every AnalysisCancelledError construction under src/ghidra,
src/ida, src/cutter, src/jeb and src/kuna to cancelledBy and lower the
guard baseline from 671 to 572. Closed Kuna target commands report
OwnerCancellation.ownerClosed instead of an absent reason, and Kuna and
JEB close abort their lifetimes with that typed reason.
* fix(errors): read the reason of any signal-shaped cancellation source
* test(ci): cover the counted-baseline comparison behind the architecture guard
* refactor(parsers): report format failures through one ParseFailure algebra
Add ParseFailure (malformed, truncated, unsupported, limit, with format,
offset and field) and its one carrier, ParseFailureError, in the domain.
BinaryReader now names truncation and the field it read, and its native
format and resource failures are ParseFailureError subclasses.
Providers projected a format failure with their own instanceof ladder,
and some dropped the selected path from the invalid_format message.
parseFailureAnalysisError now owns that projection for the native,
Go, devicetree, V8, Windows Shortcut and Electron fuse providers.
The V8 code cache, Windows Shortcut and Electron fuse format classes
are removed in favour of the algebra.
The architecture guard counts raw Buffer and DataView reads, plain
Error throws in parser folders and plain JSON.parse of parser input
against baselines that can only shrink.
* fix(parsers): retain malformed range offsets
Every rea process imported and registered all 52 command registrars
(about 1,900 modules) before running any command. A one-shot command
now loads only the registrar that defines its top-level command.
- src/cli/cliCommandRegistrations.ts holds every registrar in one table,
grouped by tool family, with its module loader and its call. The
complete createCli registers the table with static imports, and the
types check that each entry matches its module.
- src/cli/cliShell.ts owns what does not depend on which commands are
registered: identity, globals, argument policy, completions and the
registration builtins. The full and lazy paths share it.
- scripts/record-cli-commands.mjs records, after tsc, which registrar
defines each top-level command, from the same compiled table. The
bundle keeps the record as an external dist module.
- cli/cliLoader.ts selects the recorded registrar when a known top-level
command runs. It loads the complete CLI for help, --schema, --llms,
completions, builtins, unknown names ("did you mean"), or a missing
record.
- Process-owned CLI resources register their closers when their module
loads (cli/cliResources.ts), so the shell no longer imports
JavaScript-analysis and guidance owners just to close them.
Measured on one pinned core (median of 5 warm runs, base main vs this
change):
- guidance: 1.95 s / 253 MB → 0.42 s / 101 MB
- list-cutter-sessions: 1.97 s → 0.49 s
- inspect-go-binary: 1.95 s → 0.95 s
- setup --dry-run: 2.40 s → 1.88 s
- analyze: 2.39 s → 1.78 s
- --help: 1.99 s → 2.04 s (complete CLI, as before)
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
- 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>
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>
The testing guides stated lane and shard counts in prose. The counts drifted
when the workflow changed: "all eight fast built shards" survived the move to
twelve, "Eight Linux real-system lanes" now covers twelve workflow calls, and
the SQLite guide named a Node 24.15.0 lane that the matrix no longer runs.
Describe each set by its selector and point at the workflow that owns the
exact versions, so later workflow changes cannot make the text wrong.
Fixes#7616Fixes#7617
Co-Authored-By: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Across 29 successful runs, the slowest of eight shards ran 24% above the
mean shard (797 s against 646 s), with no shard index consistently slow:
duration-balanced shards still vary with hosted runner speed. Resampling
the observed per-shard speed factors predicts the slowest of twelve
shards at about 9.2 minutes instead of 13.2, for roughly 5% more runner
time.
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.
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>
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>
* 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
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>
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.
Hosted CI jobs own a four-core runner but the suite ran on two workers.
Other hosts keep the two-worker bound because they are shared.
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
* feat(setup): support Cline as a client
* test(setup): cover Cline doctor and uninstall ownership
* docs(setup): document Cline support
* test(setup): verify Cline against an installed client
* test(setup): include Cline in the exhaustive client detection list
* docs(setup): note that a hand-written rea entry blocks Cline setup
* fix(setup): isolate Cline verification and await MCP discovery
* fix(verification): own and isolate Cline discovery
* fix(setup): preserve native Cline transport settings
---------
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
Cursor Marketplace installs are Git checkouts, so the plugin manifest
points at a synced copy of the authored skill and starts
npx -y rea-agents@<release> mcp. Release Please and the release
checkpoint keep that pin aligned with the package version.
The manifest, skill, and icon checks follow the local plugin bundle
contract from #7619, adapted for a checkout that has no dist/.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: GamerDubz <105694977+GamerDubz@users.noreply.github.com>
* fix(call-path): keep computed-call targets out of complete callee coverage
Ghidra reported complete callee coverage when a computed call had one
ReferenceManager target, so trace_call_path, build_call_path and the
function comparisons turned the static candidate into a definite negative.
The bridge now keeps those targets as observed callees and marks callee
coverage partial, naming the computed call sites. IDA callee reads cannot
classify their call sites, so their callee coverage is partial too. The
Ghidra call-coverage profile revision refuses older snapshots.
A source-owned x86-64 bridge probe in the ghidra-bridge-sources job checks
the slot, register and direct-call controls against real Ghidra.
Fixes#8277
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ci): match the observed call coverage probe marker
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
* docs: organize canonical documentation layout (URGENTE)
Group all existing authored guides by domain and ownership, expose complete navigation, and preserve their public routes. Update physical path consumers and release configuration while retaining historical records. Document domain-based specs and ADR slugs without global numbering.
Co-Authored-By: Codex <noreply@openai.com>
* fix(docs): honor home layout in all-page route verification
* style: format reviewed source changes
* fix(docs): preserve routes and navigation cache inputs
* fix: repair the static-check failures on the default branch (#2895)
* fix: repair the static-check failures on the default branch
The default branch fails its own required `Static checks` job three ways, so
every pull request reports that job as failed regardless of its own change.
- `npm run format:check` reports `src/domain/functionComparison.test.ts` and
`src/ida/IdaOperationRunner.ts`.
- `npm run lint` reports a duplicate `import { join } from "node:path"` in
`tests/boundary/mcp/ghidraInventoryMcp.test.ts` (lines 4 and 17).
- `npm run typecheck` reports `format: "devin"` missing from the client format
union in `src/application/SupportedClients.ts`, and an unchecked index in
`tests/boundary/cli/clientProfileLifecycle.test.ts`, whose pair array is
inferred as `string[][]`.
Apply the pinned formatter, drop the duplicate import, list the `"devin"` format
that its own definition already uses, and mark the pair array `as const`.
Validation: `npm run check:fast` and `npm run format:check` pass.
* fix(setup): keep Devin on the JSON configuration dialect
---------
Co-authored-by: 吴佳浩 <wujiahao@supremind.com>
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
* fix(docs): update links after guide relocation
* docs: correct verified documentation route counts
* fix(docs): update browser capture guide link after move (#2820)
Refresh gettext protected-link comments alongside the authored href.
Co-authored-by: Codex <noreply@openai.com>
* fix(docs): include http3 lab in stable routes
* fix(docs): preserve historical release evidence and links
* fix(ghidra): restore Windows verifier identity import
* docs(wakaru): preserve pinned upstream policy link
* docs(index): retain recently added guide links
* docs(mcp): preserve parent retention clarification
* fix(cli): preserve standalone import guidance
* fix(test): preserve translated guide link drift coverage
---------
Co-authored-by: Matheus Breguêz <matbrgz@gmail.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: morluto <17176810+morluto@users.noreply.github.com>
Co-authored-by: 吴佳浩Alben <35291372+XiaoBinGan@users.noreply.github.com>
Co-authored-by: 吴佳浩 <wujiahao@supremind.com>
morluto