Two ways to make installation painless: make the installer simpler, or make the
install disappear. This is the second one.
standalone/moyun-armor/SKILL.md is the persona, discipline and output
anti-censorship rules as plain text. It registers no tools, edits no config,
opens no network connection, writes nothing outside the target skill directory,
and needs no restart.
Three ways to use it:
1. copy into <project>/.dsh/skills/moyun-armor/ (stays inside the workspace,
which means an agent can do it without asking for anything)
2. copy into ~/.agents/skills/moyun-armor/
3. paste the file contents straight into the conversation
It sits outside skills/ so the 63 built-in security skills keep their count.
README has the capability comparison table.
The command-line path is dead once someone refuses to execute a third-party
script. This adds the other path: one click in the DSH plugin market.
- market/index.json: format 1 catalog index, source.type = npm. The market
fetches the package from npm itself, so no git and no script are involved.
- publish-npm.ps1: publish helper with login check, version-conflict warning
and post-publish verification.
- package.json gains repository / homepage / bugs / publishConfig for npm.
- README: "install with one click" section with the three-step flow and the
index-source fallback URL.
Why this is cleaner: the market writes the profile, the user clicks Install, and
that click IS the authorization. It is not bypassing anyone's judgement, it just
moves the action to a button they already trust.
Prerequisite: publish to npm first (the name moyun-ultimate is unclaimed).
Someone reviewed the plugin and objected to three things: it writes outside the
workspace, it asks you to run a third-party install.ps1, and after installing it
you get full-disk access with approvals auto-granted plus a silent update check.
All three are accurate. This release does not argue with that; it makes each one
optional and documents it.
- SECURITY.md lists every path written, every network call, every permission
change, and for each one says whether it can be turned off and how.
- syncSkills: false - do not copy the 63 skills into \/skills
- cordis.patch.prompt-only.yml - the smallest footprint: six prompt sections
only, no skill sync, no network, no approval answerer, no permission override.
- switch-mode.ps1 gains a third mode: full / complement / prompt-only
- install.ps1 -Mode prompt-only installs in that shape in one step
In prompt-only mode the plugin touches nothing except its own directories and the
two lines it adds to package.json.
The old captures were taken before the repository had its current numbers, so
anyone comparing image to page would notice a mismatch. Retaken from the live
site with headless Edge:
01-gitea-home.png home page, shows 3.5k / 3.2k / 2.4k
03-gitea-files.png file tree at main
04-gitea-skills.png skills directory (63 entries)
09-badge-zoom.png status badge, 3x upscale, reads
"墨昀爱搬砖 · 六层就绪 · 权限由用户层接管"
Also: repository description corrected from five-layer to six-layer and the
anti-censor layer added; dropped unused 05-signup.png.
Completes the 0.6.0 release that was pushed with code only:
- README: how to learn about a new version (three ways) and how to upgrade
(one command, three variants, how to roll back, how to disable the check)
- package.json: version 0.6.0, files += update.ps1 (13 entries)
- cordis.patch.yml: checkUpdate: true toggle
- CHANGELOG: also backfills 0.5.0 and 0.5.1 which earlier edits missed
The goal is users who dare to install it. Three additions:
1. doctor.ps1 -Fix
Runs the 12 checks, then repairs what is safe to repair: re-copy the runtime
copy, add exports["./client"], add the dependencies entry, sync missing
skills. It deliberately does NOT touch the profile's own cordis.patch.yml
(that could break other plugins) - it only prints a hint.
2. install.ps1 -Rollback
Restores package.json from the newest package.json.bak-* and removes the
plugin directories. One command back to the pre-install state.
3. Friendly install ending
Prints the three next steps (quit/reopen, doctor, moyun_ultimate_profile)
plus the two escape commands, instead of a single terse line.
doctor also detects and reports whether the install is in FULL or COMPLEMENT
mode by reading the bundle patch, so complement mode is no longer misreported.
Instead of removing other people's plugins (which breaks their environment and
cannot be rolled back), this detects and yields.
1. install.ps1 scans every installed bundle before copying, reads the config ids
each one actually rewrites, and intersects with ours:
scanned plugins : 19
hard conflicts : none
~ dsh-purge shares: system-prompt (safe)
2. Three modes. auto (default) picks complement when a hard conflict exists,
full otherwise.
full permission layer owned by this plugin
complement permission layer yields, prompt layers + 63 skills still apply
auto decided by the scan
3. switch-mode.ps1 flips between them at any time; the full patch is kept as
cordis.patch.yml.full so switching back is one command.
New file cordis.patch.complement.yml drops the five permission-related config
overrides entirely, so coexistence is structural rather than hopeful.
Deliberately NOT doing: uninstalling other plugins. Reasons are in the README.
The old heuristic matched names like purge|ungate|jailbreak and reported
dsh-purge as a sibling. That was wrong: dsh-purge patches host source, it does
not touch approval/permission config ids at all.
Now the plugin reads every installed bundle's cordis.patch.yml, extracts the
config ids it actually rewrites, and intersects with its own set:
hardConflicts - other plugin rewrites approval/permission/sandbox-policy/...
-> last-applied wins, one silently stops working -> pick one
softOverlaps - other plugin only rewrites system-prompt (identity prefix)
-> negligible difference -> safe to coexist
Measured on this machine (22 bundles): scanned 19, hardConflicts [],
softOverlaps [{dsh-purge, system-prompt}]. So dsh-purge and this plugin do NOT
conflict.
Eleven comments from the video, each marked resolved / partially resolved /
cannot be resolved, with a pointer to the file that handles it.
Honest status:
- 9 resolved (install, verify, purge conflict, "second fork", missing binaries,
model gets dumber, not-working)
- 1 partially resolved (safety refusal: host permission fixed, model layer
fixed, server-side moderation improved by provider switch + layer 6 but not
removable)
- 1 cannot be resolved (the one about bypassing limits on minors - out of scope)
When the projection payload (armor) had not arrived yet, permOn evaluated to
false and the badge said the permission layer was closed, even though
config.permission is true. Only trust armor.permissionOn when it is an actual
boolean; otherwise assume on.
Real permission state should be read from moyun_ultimate_profile:
permission.effective = effective | overridden-by-user-layer | disabled-by-config
Server-side output moderation was the one thing the prompt layers could not
touch. This release does the two things that actually work:
1. prompts/50-anticensor.md (order 220, 1663 B): how to tell apart model-layer
refusal from output truncation, and five re-expression strategies that raise
pass rates (code-block form / segmentation / terminology / counterfactual
framing / language mix), plus the correct retry posture (change form, do not
resend verbatim).
2. moyun_ultimate_profile now reports a provider block with censorRisk:
high (ModelScope / Qwen / Zhipu), medium (DeepSeek official), low (overseas),
none (local). Changing provider is usually more effective than rewording, and
now the user can see which class they are on.
Local measurement: modelscope-token-plan + DeepSeek-V4.1-Flash-1017 -> high
(Aliyun content safety, over-blocks technical content).
Six layers total 23,017 B.
Answers the top complaint "it does not work and I cannot tell why" by turning
hidden state into explicit output.
- moyun_ultimate_profile now reports permission.builtin / .effective / .verdict
and environment.profile / .userLayerOverrides / .siblingPlugins / .winner,
so the user can see whether the permission layer actually took effect
- new tool moyun_skills_audit: which of the 63 skills run out of the box and
which are missing their external binaries (nuclei/fscan/IDA/jadx/...)
- badge now colors itself by real state: green (this package owns permission),
amber (user layer overrode it / sibling plugin present / disabled), red (refusal)
- skill audit speed: 2933 ms -> 234 ms by indexing PATH instead of spawning
"where" per command (0 ms once cached)
Answers the issues raised in the comments:
1. "keeps triggering safety refusal" -> README now separates three layers:
host permission (this package fixes it), model willingness (prompt layer),
provider-side moderation (nothing can fix it). TROUBLESHOOTING gives a
decision tree with commands.
2. "how do I verify it worked" -> doctor.ps1 checks 11 items and prints
[OK]/[WARN]/[FAIL]. Verified locally: OK=11 WARN=2 FAIL=0.
3. "does it conflict with purge" -> documented: both target the same config
ids, last-applied wins. doctor warns when a user-layer override or a
conflicting plugin is detected.
4. "incomplete / missing stuff" -> documented that the 63 skills need their
external binaries (nuclei/fscan/IDA/jadx...), which the package does not ship.
5. "stronger armor makes the model dumber" -> confirmed in README as a real
trade-off with per-layer guidance on what to turn off.
Five ordered prompt sections (persona core / refusal kernel / execution
intensity / anti-dilution mirror / skill routing), 63 built-in skills synced
idempotently into DSH_HOME/skills, two tools (moyun_frame, moyun_ultimate_profile),
a live refusal-score badge, and five independent layer toggles.
Ships no host source patches: the permission layer stays with host-patch
plugins, this package owns persona, strategy and skills.