Use InferRSI from openJiuwen JiuwenSwarm / WorkSwarm v0.2.7 to research an inference bottleneck, prepare a comparison or inspect existing evidence. This is the openJiuwen adapter slice of RFC #152. DSH, three-host acceptance and measured usability remain separate work.
From the GitHub Releases
page, download inferrsi-workswarm-plugin-v<version>.zip and SHA256SUMS for
the same version. Verify the archive's SHA-256 against that file. Select the
WorkSwarm archive; the Codex ZIP has a different native manifest. Release builds
stamp the native manifest and bundled runtime with the same release version.
For a local build from the InferRSI source checkout:
python tools/build_workswarm_plugin.py
The local deliverable is dist/workswarm-plugin.zip. Its inspectable package
directory is dist/workswarm-plugin/inferrsi/. Both local and release ZIPs have
exactly one native package root:
inferrsi/
manifest.json
tools/runtime_info_tool.py
skills/inferrsi/SKILL.md
skills/inferrsi/scripts/inferrsi.py
runtime/manifest.json
runtime/payload.zip
runtime/launcher.py
LICENSE
README.md
The native manifest declares package_type: plugin, id: inferrsi, one
skill with mode: all and the read-only inferrsi_runtime_info tool. Codex
uses a different installation manifest; use the
archive built for your host. Both builders copy the same canonical runtime,
knowledge, resource notices and shared launcher. WorkSwarm keeps that runtime
tree inside runtime/payload.zip, so its importer never copies deep resource
paths. The entry bootstrap verifies every payload file, its content identity
and the shared launcher before using a private per-user runtime cache. The
source template alone is incomplete.
Choose a fresh --output <directory> for another build; existing outputs and
archives are never overwritten. Building does not register the plugin in a host.
In the WorkSwarm v0.2.7 web/desktop UI:
- Open Extensions (扩展) → My extensions (我的扩展) → Plugins (插件).
- Choose Create (创建) → Upload file (上传文件创建) and select the WorkSwarm ZIP.
- Confirm import and installation, then check that InferRSI is installed.
- In a conversation in your target project, choose + → Extensions → Plugins and select InferRSI.
The v0.2.7 dialog selects a path on the backend service's filesystem. For a
remote backend, transfer the ZIP there first; a browser-local file path does not
identify a remote file. Native import and installation are separate backend
operations (plugin_packages.import_local, then plugin_packages.install).
Import alone does not make a package active. Conversation selection uses
plugin_names; there is no package-wide enable switch in this contract.
The legacy TUI /plugin install <path> uses skill import and is not the native
package installation route. Public Hub publication is separate.
The bundled core needs Python 3.10+; WorkSwarm v0.2.7 itself needs Python 3.11–3.13. Select an interpreter on the execution host and keep using it. Core knowledge operations need no InferRSI pip install. PyYAML is optional for filesystem skill operations; image validators and target engines retain their own dependencies. The host's model connection is still required for chat.
Open your target repository and ask:
Use InferRSI to research this engine's memory bottleneck. Cite relevant knowledge and prepare a comparison without starting an experiment.
Or provide an existing run and ask InferRSI to explain its verdict, missing
evidence and next action. After skill_tool loads the entry skill, call
inferrsi_runtime_info({}). Its model-facing JSON includes success and the
installed absolute launcher path. The host hides skill_directory from the
rendered skill result; neither that metadata nor the basename-only directory
appendix is a bootstrap path. The locator derives the path from its installed
module, accepts no arguments, writes no files and executes no command. If it
fails, reinstall/select the built native plugin rather than guessing a profile
path. The adapter keeps the target as cwd and delegates to existing research,
Collect/Read/Resource, plan/run and status/report operations. For manual diagnosis:
python -X utf8 "<PLUGIN_ROOT>/skills/inferrsi/scripts/inferrsi.py" --plugin-info
python -X utf8 "<PLUGIN_ROOT>/skills/inferrsi/scripts/inferrsi.py" --help
--plugin-info reports runtime/plugin versions and the build's content identity.
The WorkSwarm bootstrap verifies the payload and cached files before every
invocation. The bundle manifest records per-file SHA-256 values. A missing or
modified runtime fails with a repair message. Knowledge
reads retain their qualified identity, exact revision and attribution; stale
revisions are rejected rather than silently replaced.
plan previews a specification and starts no experiment. Freeze validators and
accepted gates before candidate search. Installation grants no new experiment
authority. Use the existing local executor on a prepared target; the adapter
does not install engines, download weights or dispatch SSH jobs.
For swarm work, designate one execution owner for a comparison. Other agents must not duplicate its runs, edit measured sources/frozen validators or contend for allocated devices. Carry workspace, interpreter, selected revisions and existing process/run identities across a handoff. A RunSpec snapshot is not a distributed device lease.
Keep execution attached to the host's process tool. Cancellation must reach that process and its cleanup path. After a timeout, disconnect or reconnect, inspect process state and existing evidence before another invocation; never silently retry. Existing-run inspection does not resume execution. Preserve PASS / FAIL / INCONCLUSIVE and distinguish synthetic integration checks from model-quality or service-performance evidence.
Deselect InferRSI in the conversation selector to stop selecting it. Start a new conversation if its prior instructions are already in context; deselection does not erase messages or stop an existing process.
For updates, finish or inspect owned work, deselect and uninstall the old package, then import/install the fresh ZIP and select it in a new conversation. Duplicate package IDs are rejected; do not bypass the host registry. Use the native uninstall action to remove the package. Keep experiment evidence outside the package directory. Uninstall does not cancel an owned run or delete evidence in target directories. The adapter adds no service, hook or scheduler.
The verified runtime cache is under %LOCALAPPDATA%/InferRSI/workswarm-runtime/
on Windows, or ~/.cache/InferRSI/workswarm-runtime/ elsewhere. It is independent
of the host profile and target repository. Content-addressed entries are
published atomically and verified before use, including concurrent first use.
The shared launcher runs in the same foreground process, preserving cancellation,
cwd and exit codes. The cache persists after commands and uninstall so paths
returned by skills reference and skills resource remain usable. After all
owned invocations finish, that cache directory can be removed to reclaim space;
the next invocation rebuilds its required entry. Modified cache entries are
rejected rather than executed. Remove the affected entry and retry from a
verified package to repair it. No host configuration or Windows policy is changed.
The native contract is based on the released v0.2.7 host, rather than an assumed universal openJiuwen SDK/plugin format:
Run the host's validator in its prepared environment against the built package:
python -X utf8 "<HOST_SOURCE>/jiuwenswarm/resources/agent/workspace/skills/plugin-creator/scripts/validate_plugin.py" "<BUILT_PACKAGE>"
--no-hot-load checks only structure/quality. A skipped hot-load layer does not
establish actual loading. The host validator's L0/L1/L2 labels refer to plugin
validation, independently of InferRSI's campaign validation ladder.
For the native backend lifecycle, use the host's prepared interpreter from a scratch directory outside either checkout:
python -X utf8 "<INFERRSI_SOURCE>/tests/check_workswarm_installed.py" --host-source "<HOST_SOURCE>" --archive "<BUILT_ZIP>"
This check creates its own temporary profile, imports and installs the archive,
checks selected/empty conversation metadata and the installed launcher, then
uninstalls and reinstalls it. It verifies installed runtime hashes, exact reads,
stale-revision rejection, and preservation of an unrelated registry entry and
existing evidence. A real DeepAgent loads the native package and dispatches
skill_tool followed by inferrsi_runtime_info through AbilityManager.execute.
The launcher is parsed only from the resulting model-facing ToolMessage and
then invoked from an external target workspace. Structured path metadata is
discarded. Extension unload is also checked. The probe runs in a disposable
subprocess so its SDK log handles close before the parent removes the temporary
profile. The replay invokes no model and does not change the desktop user's profile.
Windows v0.2.7's importer stages and copies the entire package under its data
root. The earlier unpacked layout reproduced WinError 206 / missing-file
errors in the desktop's ordinary profile. The compact outer package avoids
those deep import paths; the bootstrap uses extended-length paths only for
its own verified Windows cache. Native lifecycle checks use an ordinary data
root without a \\?\ prefix, rather than masking the importer failure by
changing the host environment.
Validation on 2026-11-03, Windows / Python 3.12:
| Check | Observed scope |
|---|---|
| Official v0.2.7 validator | Quality, static checks and real hot-load passed with zero errors/warnings; one skill and one read-only locator bound |
| Native backend lifecycle | Ordinary isolated data-root import/install/uninstall replay; reinstall, selection metadata and unrelated registry/evidence preservation checks |
| Model-facing bootstrap | Native skill discovery and locator ToolMessages resolve the installed launcher; isolated CLI invocation, unknown-argument rejection and extension unload checked without calling a model |
| Extracted runtime | Both host distributions exercise canonical bytes/notices, isolated exact knowledge reads, stale-revision rejection, optional dependency guidance, preview without execution and synthetic CPU run/status/report for all three verdicts |
| Running desktop backend | Compact ZIP imported and installed through the running desktop's native RPCs in its unchanged default profile; installed launcher and resource reads checked |
| Desktop click workflow, LLM dispatch, cancellation/reconnect | UI error text inspected; click automation unavailable, so a complete UI installation journey and model/cancellation tests remain unverified |
Plain unbuilt source reports runtime version 0+unknown; --plugin-info
reports that development state explicitly. A versioned build can use
--release-version <version> only when the already-built runtime agrees;
this stamps generated host metadata without modifying canonical source bytes.
Actual end-to-end desktop/LLM journeys, cancellation/reconnect, other host releases and Linux/macOS installation require their own verification. Packaging or hot-load success alone does not complete the RFC or demonstrate a usability improvement.