agentic-env-presets
Manifest validStandard and PTC agent presets whose system prompt describes a plain agentic coding environment instead of a branded harness. Install: dsh plugin --profile web add dsh-agentic-env-presets
Agentic environment presets
Two agent presets — exact clones of the shipped Standard and PTC
compositions — whose model-facing text describes a plain agentic coding
environment instead of naming the harness. Tools, plan mode, compaction,
delegation, subagents, prompt caching and every config value are copied
verbatim from the shipped declarations. The only composition change is one
extra row, env-identity.
"Model-facing" is wider than the system prompt: it also covers the runtime context snapshot, tool schema descriptions, and the skill catalog message injected every step. Those four kinds of text belong to four different subsystems, which is why the plugin has five layers instead of one — see How the trick works.
Install
Requires a DSH whose prompt API matches this bundle — 0.1.7 through 0.2.x
(see Compatibility). Nothing else is needed: no build step, no
native dependency, no post-install script.
From the plugin page (GUI)
The Web GUI's plugin manager installs the same three things through one text box (Settings → Plugins → install). Paste any one of these:
| Paste this | Installs | Resolves as |
|---|---|---|
dsh-agentic-env-presets | latest from npm | registry |
dsh-agentic-env-presets@1.2.1 | that exact npm version | registry |
https://github.com/Jing494/agentic-env-presets | GitHub default branch | git |
github:Jing494/agentic-env-presets | GitHub default branch | git |
/absolute/path/to/agentic-env-presets | a local checkout | path |
The box asks for the package name, a GitHub repository URL, or a local directory path — exactly those three shapes. Things that look right but are not accepted, each confirmed against the installer's own parser:
| Instead of | Reason it fails |
|---|---|
https://www.npmjs.com/package/dsh-agentic-env-presets | the npm web page is not a repository URL. The installer reads it as a git host and the clone fails. Paste the bare package name instead. |
npm install dsh-agentic-env-presets / dsh plugin add … | a whole command line is not a spec; paste only the part after add. |
Jing494/agentic-env-presets | not a valid package name, and not a URL — use the full https://github.com/… form. |
./agentic-env-presets | a local path must be absolute. |
Install scripts are never needed by this bundle, so the plugin page will not ask you to approve any build step. If a package it does install wants to run one, the page stops and asks first — that prompt is the gate, not an error.
From npm (recommended)
dsh plugin --profile web add dsh-agentic-env-presets
Pinned to an exact version, if you would rather control when it moves:
dsh plugin --profile web add dsh-agentic-env-presets@1.2.1
From GitHub
Tracks the repository's default branch — you get whatever is committed, including changes that have not been released:
dsh plugin --profile web add https://github.com/Jing494/agentic-env-presets
# or the shorthand pnpm accepts:
dsh plugin --profile web add github:Jing494/agentic-env-presets
From a local checkout
For editing the presets or the rewriter itself. link: keeps the Host pointed at
this working tree, so a change is visible without reinstalling:
dsh plugin --profile web add link:/path/to/agentic-env-presets
After installing
Both commands register the bundle as a profile layer automatically. dsh then
writes it into $DSH_HOME/profiles/<profile>/package.json under
dsh.profile.bundles; check with:
dsh --profile web --dump-config | grep -E 'preset-(standard|ptc)-env'
Restart the profile, then pick 标准模式(中性环境) or PTC 模式(中性环境)
when creating a new session — the picker lists them next to the shipped
presets. Existing sessions keep the preset they started with.
Verify the install
dsh --profile web --dump-config | grep -c env-identity # expect 2
If a bundle cannot load, dsh prints skipping profile bundle "<name>": <reason>
and boots without it — the profile is never taken down by a bad bundle. Re-running
add after fixing the cause reinstalls it.
Uninstall
dsh plugin --profile web remove dsh-agentic-env-presets
This drops the dependency and its dsh.profile.bundles entry in one step, so no
stale layer is left behind. Sessions already running on one of these presets keep
it until they restart.
Deliverables
| File | What it is |
|---|---|
presets-standard.patch.yml | Standard clone, preset id standard-env, Loader row preset-standard-env |
presets-ptc.patch.yml | PTC clone (run_code-only executor, workflow / plugin-market tools disabled), preset id ptc-env |
index.js | The env-identity rewriter plugin (five layers, config-driven phrase table and three switches) |
verify.mjs | Reproducible 20-case verification against the real shipped plugins, plus mutation-test instructions |
probe-restrict.mjs | Proves a tool shadow does not bypass a subagent's tools.restrict() |
probe-robust.mjs | Proves missing services/fields degrade to a no-op instead of throwing (with a control case) |
gen-presets.mjs | Regenerates both patches from the shipped files (fidelity check/upgrade tool) |
package.json | Bundle manifest; dsh.bundle.patch lists both patch files |
icon.svg, locale/en.json, locale/zh.json | Official-style icon and localized bundle title/description |
docs/DETAIL.md, docs/DETAIL.zh.md | The detail document each language's meta.detail points at |
docs/PUBLISHING.md, docs/PUBLISHING.zh.md | Repository requirements for publishing and being indexed as a DSH plugin |
LICENSE | MIT |
README.zh.md | Chinese README |
移植说明.md | Chinese design/pitfall/porting notes |
PROMPT.md | The original task brief |
How the trick works
includeHarnessIdentity is a deployment-level switch, and an agent preset
cannot mount the prompt registry itself — so no preset can turn it off. What a
preset can do is contribute sections in its own scope, and the registry's
scoped layers are keyed by section name: ScopedLayers.merge() lets a scoped
section shadow a global one of the same name, farthest-ancestor-first, and
dsh-agent-presets chains the agent scope to its preset's standing scope with
bindScopeParent(agentKey, presetKey). That is the whole trick:
global layer harness:identity, tool guidance, harness:source,
app:web-surface, sandbox:policy, every tool …
▲ shadowed by
preset layer harness:identity (new text), harness:source = "",
app:web-surface = "", sandbox:policy = delegating wrapper,
each tool = same name, scrubbed description
▲ inherited by
agent scope (per-session additions)
The empty shadows still exist as assembly entries, but renderPrompt drops
empty sections and empty contexts disappear from the joined snapshot. The
global owner stays mounted for every other preset.
Brand text lives in five places, not one
This was 1.1.1's core blind spot: it covered only the two prompt sections, so a real session still introduced itself with the brand name.
| Surface | Owner | Seam |
|---|---|---|
| System-prompt sections | dsh-system-prompt | same-name section() |
| Runtime-context entries | each owner; re-resolved after assembly | delegating wrapper in the agent's layer; in-place text for this preset's own |
| Tool schema descriptions | dsh-tools | same-name tools.register(), forwarding the executables |
| Loop-injected user messages (skill catalog) | dsh-tool-skill and friends | agent/pre-step, and it must be {prepend:true} |
| Everything left over | anyone | system-prompt/assemble waterfall |
Two of these are counter-intuitive enough to spell out:
- Runtime context is re-resolved after the waterfall.
dsh-agent-looprunsassemble()→ waterfall →refreshContext()→ render, andrefreshContext()re-resolves every context provider. A waterfall-only scrub therefore looks perfect in a unit test and leaks in the real request. This package shadows each context with a delegating wrapper instead, so the re-resolution lands on our function. - The skill catalog is appended after
next(). So theagent/pre-steplistener must be outermost ({prepend:true}); a default (inner) registration has already finished scanning an empty decision by the time the catalog — brand text and all — is appended.verify.mjspins this with a dedicated case.
What the plugin does (index.js)
It is a plain Cordis function plugin (name, inject, apply) with no
runtime dependencies, mounted as one row inside each preset. Five layers:
- Replace the identity. Register
harness:identityin the preset scope withconfig.identity. Same name ⇒ the global one is replaced, not duplicated. - Silence harness-maintenance context. When
silenceDevContext(defaulttrue), registerharness:sourceandapp:web-surface— both the section and the runtime-context halves, four names total — with empty text. This is what removes the checkout path and GUI URL, without unmounting their owner. - Shadow the runtime contexts. The wrapper is installed in the agent's
own layer, not the preset's, and its
textresolves whatever provider currently wins the name before scrubbing. Doing it in the preset layer would occupy the name:dsh-basemounts a globaltool-subagent, every preset re-declares it, and the preset's own row would then fail activation withalready registered in this scope— one failed row blocks every new session. The agent's layer competes with nobody and still wins, becauseassemble()/refreshContext()merge along the same scope chain. Providers declared in the preset's own layer are rewritten in place instead, for the same reason: re-registering them there cannot win. - Shadow tool descriptions. Every branded tool gets all three
documentation surfaces rewritten: the top-level
description, every nesteddescription/titleinsideparameters, and the same fields insideoutput.schema.enum,const,default,$refand property NAMES are data or identifiers and stay byte-identical, as doexecute,renderandwhen. PTC mode inlines all three surfaces into itstools:sdkprompt section, so both executors are covered by the one fix. Inherited tools are re-registered under their own name; tools declared in the preset's own layer are rewritten in place, becausetools.view()without a scope reports the global view — empty in a deployment that disables its tool rows globally and re-declares them per preset, which left every such description unscrubbed. The reservedrun_codename is skipped. A shadow does not bypass a subagent'stools.restrict(): a preset layer is an inherited layer and still has to passadmits(). - Rewrite a
complete: trueprompt. Such a section is restored byassemble()after the waterfall, from the pre-waterfall snapshot, as the entire prompt — none of the four layers above can reach it. In an ancestor layer a same-name re-registration (keepingcomplete) works; in this preset's own layer that re-registration must throw, so the section object is rewritten in place (after checking that it is neither frozen nor shared between layers, and restored on unload). The shippedminimalpreset setscomplete: true, so this is a real path, not a hypothetical one. - Scrub what is left. A scope-filtered
system-prompt/assemblewaterfall listener as a last resort, plus an outermostagent/pre-steplistener that rewrites the injected skill catalog and skill bodies (skill-catalog,skill-invocation, …) while excluding the caller's own messages by id.
Layer 6's phrase table is literal split/join and order matters (longest
first): the space-suffixed DeepSeek Harness pair runs before the bare one, so
"the DeepSeek Harness implementation checkout" becomes "the implementation
checkout" instead of "the the …".
A bare DSH left in prose is caught by a narrow regex with three
assertions, (?<![\w$.\-])DSH(?![\w\-]). That reverses 1.1.1's "never
regex" rule, because a pure phrase table always lags whatever a plugin author
writes next. A substitution at the head of a sentence is capitalised
(DSH rules apply becomes The … rules apply), while CJK context gets no
English article (DSH 的规则 never becomes The … 的规则). The assertions keep
these byte-identical:
$DSH_HOME DSH_ENV_PREFIX __DSH_BOOT__ engines.dsh
@deepseek-ai/dsh-tool-bash DSH-1234 myDSHThing
The listener returns copies of the section/context entries. Assemblies are shared with the registry until the waterfall returns, and mutating them in place would leak into other agents' later requests.
Deliberately not done: no global patch to framework prompt strings; no
complete: true section (that would swallow tool guidance, plan mode and the
SDK docs and break prompt caching); no renaming of tools (that is the call
surface — a rename breaks tool dispatch outright); no touching $DSH_*
identifiers.
Config
- id: env-identity
name: 'dsh-agentic-env-presets'
config:
identity: You are an AI coding agent working inside an agentic coding environment.
silenceDevContext: true # optional, default true
scrubToolDescriptions: true # optional, default true; false = prompt+context only
scrubLoopMessages: true # optional, default true; false = leave loops alone
bareToken: true # optional, default true; false = phrase table only
replacements: # optional, defaults to DEFAULT_REPLACEMENTS (whole table)
- ['DSH(DeepSeek Harness)', 'the agentic coding environment']
- ['DeepSeek Harness ', '']
- ['DeepSeek Harness', 'the agentic coding environment']
- ['DSH Marketplace', 'plugin marketplace']
- ['DSH file policy', 'file policy']
- ['DSH file sandbox', 'file sandbox']
- ['extend DSH itself', 'extend the environment itself']
The three switches exist for diagnosis and porting, not daily use: when one
layer is suspected of over-reaching, switching layers off one at a time
localizes it. bareToken: false restores the old literal-phrase-only behavior.
Compatibility
The manifest declares no engines.dsh on purpose. A narrow range such as
^0.2.2-rc.1 covers only 0.2.2–0.2.5, so the bundle would be rejected
outright the moment the engine moves past that window — exactly the failure
mode where upgrading the engine forces upgrading the plugin first. Instead:
peerDependenciesdeclares@deepseek-ai/dsh-system-prompt": ">=0.1.0-0", whichsemvermatches withincludePrerelease, so prerelease engines count.- The sections the plugin touches (
harness:identity,harness:source,app:web-surface) and the prompt API it calls (getSectionOrder,getContextOrder,system-prompt/assemble) exist from0.1.7through0.2.x, which is what the range is anchored on.
If a future engine renames those sections, the bundle is skipped at startup with a diagnostic naming the missing piece, and the profile still boots — a bad bundle never takes the harness down with it.
Verification
verify.mjs mounts the real dsh-system-prompt,
dsh-harness-dev-context, dsh-persona, dsh-tools,
dsh-tool-executor-native, dsh-agent and dsh-scope packages, plus the two
Host services dsh-harness-dev-context injects (webApp,
sessionProjections). It opens a preset scope, chains an agent scope beneath
it, then follows the loop's real order — assemble, refresh, render, resolve
tools, run the pre-step chain — and asserts with exact substring checks. Nine
cases:
| Case | Assertion |
|---|---|
| 1 default config | identity is the new text; maintenance sections/contexts are empty; no brand text on any of the four surfaces; no doubled "the the" |
2 silenceDevContext: false | maintenance paragraphs stay (that is what disabling silence means), but the scrub still removes every brand phrase |
| 3 baseline preset (no env row) | brand text is still present — proving nothing leaked into the global/deployment layer |
4 after refreshContext() | the loop re-resolves contexts after the waterfall; the result must still be clean, and the file-policy text must not have been emptied |
| 5 tool schemas | top-level descriptions and nested parameters documentation scrubbed; identifiers byte-identical; enum/default/property names (data) preserved; the baseline tool description still original |
| 6 skill catalog | the catalog message is actually rewritten (not merely skipped), and source.entries is unpolluted |
| 7 identifier safety | $DSH_HOME, __DSH_BOOT__, engines.dsh, DSH-1234 and friends survive verbatim |
| 8 mount-order independence | the same plugins in two different mount orders must both be clean — the shadow scan was once aborted by its own empty shadows and leaked in only one order |
5b PTC tools:sdk render | Renders the shadowed definition through the real renderToolsSdkPy(): prose must be clean and the enum data (Literal["DSH", …]) must survive |
9 complete: true in either layer | In an ancestor layer a same-name shadow works; in this preset's own layer (a sibling dsh-persona row) re-registration must throw, so the section object is rewritten in place. Neither placement may leak, and a fixed placement must not warn |
| 10 Readable wording | Sentence-initial substitution is capitalised, mid-sentence stays lowercase, CJK context gets no English article, and identifiers are byte-identical |
It runs from any directory (it locates the dsh installation itself and imports
by absolute path, because ESM does not honor NODE_PATH and this filesystem
disallows symlinks):
node verify.mjs # or: node verify.mjs <dsh-install>
echo $? # 0 = all pass, 1 = failure
Each case prints the exact text that surface would send to the model, so the result can be eyeballed as well as asserted.
Always run the mutation tests
A suite that always passes is not a suite. Run at least these two; both must FAIL:
# drop prepend: the skill-catalog case (6) must fail
sed -i "s/onPreStep, { prepend: true }/onPreStep/" index.js && node verify.mjs
# disable the context shadow: the refresh case (4) must fail
sed -i "s/^ rescanContexts();/ \/\/ rescanContexts();/" index.js && node verify.mjs
git checkout index.js # restore
Both were run against 1.2.0: they hit cases 6 and 4 respectively, with every other case still passing.
Four more, added in 1.2.1 for the defects that broke a real deployment:
# scan every scoped layer again: case 20 (sibling-preset collision) must fail
# (replace the ownLayerOf(...) block in shadowContexts with a loop over
# prompt.layers.scoped.values())
# ask tools.view() without a scope: case 21 (preset-layer tools) must fail
sed -i "s/tools.view(state.scope)/tools.view()/" index.js && node verify.mjs
# close over the provider instead of resolving it: case 23 must fail
# (replace `const current = resolveCurrent();` with `const current = provider;`)
# register the agent-layer shadow on the preset context: case 22 must fail
# (replace `agent.ctx.effect(` with `ctx.effect(` in shadowContextsForAgent)
Two more probes cover the "do not break the harness" side; exit code 0 means pass:
node probe-restrict.mjs # a shadow does not resurrect a tools.restrict()-denied tool
node probe-robust.mjs # no tools service / no layers / empty config / a throwing provider
The last case in probe-robust.mjs is a control experiment: when a context
provider throws on its own, it throws with the plugin absent too. The assertion
is "identical with and without the preset", not "never throws".
Limitations / what this cannot change
- This changes the model-facing text, not the product. The Web GUI, the
CLI help (
dsh --help), the plugin manager, the marketplace copy, and the onboarding text still name the harness. A preset cannot reach any of them, and should not try. - It is per-agent, by construction. Other presets, other sessions, and the deployment layer keep the original prompt, so an agent that maintains the harness can still describe it accurately. If the brand text ever disappears from a baseline preset, the shadowing leaked — that is exactly what case 3 checks.
- Empty-shadowed sections still exist. They render to nothing, but
SystemPrompt.assemble()lists them; a tool inspecting the raw assembly sees two empty entries. Intentional and harmless. - It is harness-shaped. It relies on
ScopedLayersshadowing by section name,system-prompt/assemblebeing a scope-filtered waterfall,refreshContext()reading the registry, theagent/pre-stepwaterfall order, and the preset→agent scope chain. On a platform without those seams, see the porting section in移植说明.md. Tool parameter descriptions are not scrubbed: fixed in 1.2.0. Nesteddescription/titlefields are now rewritten recursively, whileenum/const/default/$refand property names are left alone — those are data or identifiers, and changing them would break argument construction.Skill bodies are not scrubbed: fixed in 1.2.0.skill-invocationis appended by the sameagent/pre-stephook as the catalog, so the body is scrubbed too (seeverify.mjscase 6).A: fixed in 1.2.0. An ancestor-layer section is shadowed by name; a same-layer one has its registry entry rewritten in place (case 9). Only a frozen or cross-layer-shared object falls back to a warning instead of a fix, and that is logged loudly rather than silently.complete: trueprompt takeover- Skill NAMES (e.g.
dsh-plugin-store) and the on-diskSKILL.mdfiles are untouched. The name is the call surface — renaming it breaksskill(). The file itself lives outside the prompt layer; this package scrubs it only as it enters the model context. - The phrase table is deployment-specific. Shipped wording is the default;
a deployment that changed its prompts should edit
config.replacements. A bareDSHin prose is caught by the guarded regex; add new brand terms to the table. Re-runverify.mjs— and one mutation test — after any framework upgrade.
Upgrading the clones
Both patches are verbatim clones. After a DSH upgrade, regenerate and diff:
node gen-presets.mjs <dsh-install>/node_modules/@deepseek-ai/dsh-web-app/presets
diff -u <shipped>/standard.patch.yml presets-standard.patch.yml
The diff must be exactly the header, the row id, the metadata block, and the
single env-identity row.
Changelog
2026-11-14 — 1.2.1: stop claiming context names other presets own.
Creating a session failed outright:
agent-preset/invalid: prompt context "tool:subagent:depth" is already registered in this scope
The context scan walked every scoped layer, so a name owned by a different
preset looked like an unshadowed global. This plugin then claimed it in its own
(later, farther) layer, and the preset's own tool-subagent row — mounted after
env-identity — could not register it in that same layer. One failed row fails
the whole preset, so every new session was refused.
Two more defects had the same root: asking a scoped registry the wrong question.
- The same name can legitimately exist globally and in the preset. Shadowing the global copy from the preset layer occupied the name and broke the preset's own row. Inherited contexts are now shadowed in the agent's layer; own-layer providers are rewritten in place. (Verified: without this plugin, a global and a preset copy of one name coexist fine — so this was our bug, not the presets'.)
tools.view()without a scope reports the global view, which is empty in a deployment that disables its tool rows globally and re-declares them per preset. Every preset-layer tool kept its brand text.
And an agent-layer wrapper must resolve the current winner per call. Closing over the provider froze the text: a replaced provider never reached the model.
verify.mjs grew to 24 cases, each with a mutation test that must fail.
2026-11-14 — 1.2.0: the scrub now reaches what the model actually reads.
A session running the standard-env preset still introduced itself as
"running inside DSH (DeepSeek Harness)". Pulling that request out of the
session log showed the leak came from three surfaces 1.1.1 never touched:
| Surface | Observed |
|---|---|
| Runtime context | Current DSH file policy: … the DSH file sandbox … |
| Tool schema descriptions | 11 occurrences, e.g. Search the DSH Marketplace catalogue of DeepSeek Harness plugins |
| Skill-catalog user message | 3 occurrences, Find and install DeepSeek Harness (DSH) plugins |
And the runtime context it thought it covered was only covered at the instant
assemble() runs: dsh-agent-loop calls refreshContext() after the
waterfall, re-resolving every provider and putting the original sentence back.
The old verify.mjs asserted on assemble()'s output, which is why all three
of its cases passed.
Fixes and hardening:
- Delegating context shadows, so
refreshContext()resolves a scrubbed function instead of the original provider. - Tool shadowing (forwarding
execute/output/when), which also fixes PTC mode inlining the same descriptions intotools:sdk. - Skill-catalog scrubbing, registered
{prepend:true}so it runs outermost — the catalog is appended afternext(), so a default inner registration would silently do nothing. - Bare
DSHnow falls to a guarded narrow regex; identifiers such as$DSH_HOMEstay byte-identical. verify.mjsgrew from 3 cases to 7, each pinning one of the paths above, with mutation tests proving they can actually fail.
2026-11-11 — re-cloned against DSH 0.2.2-rc.1 (fixes a broken row).
Both files had drifted from the shipped declarations. The visible symptom was
tool-subagent-list-agents (@deepseek-ai/dsh-tool-subagent-control/list-agents): never started
reported by @deepseek-ai/dsh-agent-preset-registry while mounting
standard-env / ptc-env. The row imported
@deepseek-ai/dsh-tool-subagent-control/list-agents, but that package's
exports map has no ./list-agents subpath — list_agents is registered
by its main entry (apply() calls registerListAgents(ctx)). The import could
therefore only fail with ERR_PACKAGE_PATH_NOT_EXPORTED, the Loader logged the
import error and left entry.fiber undefined, and the audit rendered it as
"never started". The shipped standard/ptc presets never declare such a row.
Fixed by regenerating both files verbatim from the 0.2.2-rc.1 declarations, so the row is gone. The same regeneration also restored three drift items that had silently changed behavior:
| Row | Before | After (matches shipped) |
|---|---|---|
tool-fs | no config (applyPatch defaulted off, but the comment was gone) | applyPatch: false |
tool-subagent | modelSelectionSettings: true (not a config key) | modelSelection: true + modelDiscovery: true |
tool-executor-ptc (ptc only) | @deepseek-ai/dsh-agent-tool-presentation (mode: ptc) | @deepseek-ai/dsh-tool-executor-ptc |
2026-11-13 — publishable manifest. Renamed to the unscoped
dsh-agentic-env-presets, dropped private, added license: MIT,
repository, icon, keywords, main, and the locale meta.detail docs;
wide peerDependencies replaces any engines.dsh pin (see Compatibility).
2026-11-14 (same day, after the compatibility audit) — two more real gaps closed; 9 cases.
Auditing upgrade cost and cross-version compatibility surfaced two more surfaces:
| Gap | Symptom |
|---|---|
Documentation inside output.schema | dsh-tools never shows it to the model, but PTC's renderSdk() renders it into tools:sdk, where nested description becomes a # comment |
| Skill bodies | skill-invocation is appended by the same agent/pre-step hook as the catalog; only the catalog was being scrubbed |
Also measured: 0.1.7-rc.2 and 0.2.2-rc.3 both work in either direction, and
gen-presets.mjs absorbs every shipped-preset change automatically (3 of them
between rc.1 and rc.3). The suite grew to 9 cases, with case 5b asserting on the
output of the real renderToolsSdkPy() while keeping enum data intact.
License
MIT © 2026 Jing494
Comments
Loading…
Similar plugins
by KannaKuron
Agent preset for DeepSeek Harness that layers the self-referential Cordis toolset (dynamic plugins, runtime inspection) and preset-authoring skills onto the PTC Code-Mode preset; materialized files ca
★ 9
↓ 1.4k/wk
MIT
JavaScript
Nov 12, 2026
dsh plugin --profile web add dsh-ptc-cordis-presetby lishize20040802-rgb
Deep PTC agent preset for DeepSeek Harness: built-in PTC tooling plus visible progress, confident tool use, and early clarification.
★ 52
↓ 3.3k/wk
MIT
JavaScript
Nov 17, 2026
dsh plugin --profile web add dsh-deep-ptc-modeby KouzakiUmi
DeepSeek Harness 插件:按 Agent preset 限制每个 Agent 可见的工具,保留全局 provider 挂载;基于 agent/created 与 ctx.tools.restrict()。
★ 0
MIT
JavaScript
Oct 19, 2026
dsh plugin --profile web add dsh-preset-tool-guardby YukimuraSol
DeepSeek Harness plugin: renders AGENTS.md as an independent system-prompt section, re-read on every prompt assembly instead of injected as user messages.
★ 0
MIT
JavaScript
Oct 28, 2026
dsh plugin --profile web add dsh-agents-sectionby QlzqQlzq
Installs two selectable Agent Presets: a shell-restricted general-purpose agent and a repository-focused coding agent.
★ 1
MIT
JavaScript
Aug 15, 2026
dsh plugin --profile web add dsh-dual-agent-presetsby lilinfangrelax
Connect DeepSeek Harness (dsh) to the coding agents already installed on your machine over ACP: local detection, official ACP Registry, native chat routing and subagent delegation.
★ 0
MIT
JavaScript
Nov 13, 2026
dsh plugin --profile web add dsh-acp-agents