xiaobright/dsh-anchored-standard
Two-phase DeepSeek Harness preset: Minimal-aligned bootstrap, then full Standard tools (Project2 98/99)
安装Install
dsh plugin --profile web add github:xiaobright/dsh-anchored-standard把 xiaobright/dsh-anchored-standard 加入你的 DSH 配置（web profile）即可启用。
READMEREADME
dsh-anchored-standard
Experimental DeepSeek Harness agent presets — a base mode, two live-anchor variants, and one seeded prefab mode — that anchor a session's model trajectory on the Minimal condition (real Minimal tool schema, no auto-injected context), then promote to a small resident catalog once the session is durable, unlocking heavier Standard tools on demand.
This is a community project. It is not an official DeepSeek preset and is not affiliated with or endorsed by DeepSeek.
Feel free to submit feedback on the plugin in the form of Issues or PRs. For ideas for new plugins or useful findings, please submit them under the repository.
Project status (2026-08-17)
Following the price increases on both the DeepSeek official API and the opencode go subscription, active development of this project has effectively stopped: the evaluation loops these presets depend on (Project2-class runs and multi-trial roll/probe experiments) are no longer affordable. The repository stays available as-is and receives maintenance only (bug fixes and harness-compatibility updates when feasible). The mechanism findings, the dose-response data, and the tooling (context-gate, the prefab pipeline, the probe suite) remain valid and are largely model-agnostic. A personal note from the maintainer: FAREWELL.md (Chinese). Contributors and collaborators are listed in ACKNOWLEDGEMENTS.md.
Community projects that users report perform better in some scenarios:
- dsh-routing-suite — a runtime injector plus task-aware thinking-mode routing presets (the router-standard family).
- J-Space Cognition Suite — a model-agnostic inference-time cognitive control layer packaged as a Skill.
Modes at a glance
| Mode | Directory | First model request | Anchor mechanism | Promotion signal | Cost |
|---|---|---|---|---|---|
| Anchored Standard | preset/ |
2 tools (the Minimal pair) | Minimal tool schema | first durable tool/call or assistant/message (promoteOn: either) |
none |
| Zero-Anchored Standard | zero-anchored-standard/ |
0 tools | one fixed anchor turn | the anchor reply (assistant/message) |
+1 model call |
| Whoami Standard | whoami-standard/ |
0 tools | one "你是谁" self-introduction turn | the self-introduction reply (assistant/message) |
+1 model call |
| Prefab Anchored Standard | prefab/ |
seeded rolled history | bundled successful trajectory | already promoted in the seed | no model call to instantiate |
| Eternal Minimal | eternal-minimal/ |
2 tools, forever | the visible catalog never grows; heavier tools run via the dshx bash gateway |
none (no phases) | none |
| Wire Think-Execute Standard | wire-think-standard/ |
tools present, tool_choice: none on the wire |
sibling provider route per think step | per-turn: the steer itself | +1 model call/turn, prefix-cache churn |
| Combo Anchored | combo-anchored/ |
0 tools, on every user turn | think/execute split + depth gate + deliberation drip as three independent rows | per-mechanism | +1 model call/turn |
Every mode directory is self-contained and installs alone under whatever id you copy it to (see Install). The prefab hydrates the blank session in place when its preset is selected; no per-workspace import is required.
Terminology
- trajectory — the style of the model's first reasoning chain. The Minimal condition produces "We need…" first lines; the Standard condition produces "Let me…" ("standard-like") first lines.
- anchor — the first-request conditions that select the trajectory. Issue #11 isolated three levers: the tool schema, the output budget, and the injected reminders.
- bootstrap phase — request #1 of a session: the bootstrap tool pair, no auto-injected context, optional output cap.
- promotion — the durable session event that ends the bootstrap phase.
Base mode: first
tool/callorassistant/message, whichever comes first. Variants: the anchor reply. - durable — recorded in the session event log. Phase state is derived from durable events, so resume and reload preserve it.
- resident catalog — the promoted tool set: the bootstrap pair plus the discovery tools plus every tool the model explicitly unlocked.
- discovery tools —
dev_tool_search,skill_search,skill_load: the on-demand unlock surface for heavier Standard tools. - materialized copy — the committed copy of a
shared/plugin inside a mode directory, generated bynpm run sync.
How it works
The base mode in one request lifecycle (the variants change only the first turn — see their sections):
user's first message
│
▼
┌ request #1 ─ bootstrap phase ──────────────────────────────┐
│ tools : bash + str_replace_editor (Minimal's real pair) │
│ context : no AGENTS.md digest, no skill-catalog reminder │
│ budget : adapter default (`bootstrapMaxTokens` optional) │
└─────────────────────────────────────────────────────────────┘
│ first durable tool/call OR assistant/message
▼ PROMOTION — derived from durable events, resume-safe
┌ request #2+ ─ resident phase ──────────────────────────────┐
│ tools : bootstrap pair + discovery tools + unlocked │
│ context : standard injections restored │
│ budget : adapter default (a cap is stripped on promote) │
└─────────────────────────────────────────────────────────────┘
Three first-request levers decide the trajectory (issue #11):
- Tool schema — the decisive variable at the adapter-default maxTokens (256000). The real Minimal pair anchored 5/5; every standard-family schema fell standard-like 11/11.
- Output budget — a 1024 first-request cap also anchored the trajectory
(26/32), independent of the tool descriptions. The base mode leaves this
lever unset (
bootstrapMaxTokensis opt-in). - Injected reminders — the AGENTS.md/CLAUDE.md digest and the
available-skills reminder. With the skill catalog present, the anchor did
not reproduce at all (0/9). The base mode now suppresses EVERY automatic
injection during bootstrap at the harness's two unified injection paths
(the
context-gateplugin), not just the two measured ones.
Why
DeepSeek V4 Pro conditions strongly on the API-visible tool catalog. In the Project2 evaluation, Standard and PTC produced scores of 91 and 92, while the official Minimal preset produced 99 and 96. Permanently staying on Minimal, however, gives up the Standard preset's broader tool set.
Anchored Standard separates initial trajectory selection from later tool use:
- Keep the complete Minimal system prompt.
- Expose the Minimal preset's REAL tool schemas — persistent
bash+str_replace_editor, byte-identical to the official Minimal composition — on the first model request (lever 1 above). - Suppress EVERY auto-injected context on that first request — at the
harness's unified injection paths, not per source name (the
context-gaterow, mounted FIRST; lever 3). While the session is unpromoted the assembly's dynamic runtime-context contributions are blanked (the wholeSystemPrompt.context()family: sandbox/approval policy snapshots and any third-party context provider), and the pre-step waterfall keeps only the CLAIMED message batch plus a small kind allowlist (a user-initiated skill gesture survives; skill catalog, AGENTS.md digest, time/tmux context, hooks, and unknown third-party injections are stripped by default). After promotion the gate opens and the loop's own snapshot projection diffs exactly ONE fresh runtime-context message into the next request — minimal first round, injections on the second round. Acompaction/endboundary re-closes the gate the same way. - After the session records its first durable promotion signal — a
tool/callor the firstassistant/message, whichever comes first — promote to the RESIDENT catalog: the bootstrap pair plus the discovery tools plus whatever the model has explicitly unlocked viadev_tool_search. Dumping the full Standard catalog at promotion pulled the trajectory back to standard-like behavior (the post-promotion regression), so heavier tools —web_search,subagent,workflow, … — stay onedev_tool_searchcall away. Request #1 always sees the bootstrap catalog; request #2 always sees the resident catalog, so a text-only first reply can no longer trap the session in bootstrap. (promoteOnin thetool-bootstraprow selects the trigger:eitherdefault,tool-call, orassistant-message.) - Derive the phase from durable session events so resume and reload preserve it.
The bootstrap catalog is the same on every platform: the Minimal pair
(bash/str_replace_editor). The preset's shell is the persistent PTY bash
(the sandboxed Standard bash row is disabled — both register the bash name
into the same layer, and the tools registry rejects duplicates; Windows never
had the sandboxed bash anyway). pwsh remains available in the promoted
catalog on Windows.
Results
The anchored family was validated on Project2 with three V4 Pro scores of
98, 99, and 99. Provenance note (issue #60): those three runs predate the
current composition — they used the Minimal system prompt with a first-request
pwsh + read surface and promoted to the full 25-tool Standard catalog;
the exact Minimal pair (persistent bash + str_replace_editor) with the
small resident catalog was introduced afterward. The bundled generic prefab
removes Project2-specific warm-up facts and was not re-benchmarked before the
API price change, so those scores must not be attributed to the generic
template.
Independent replications: the trajectory anchoring reproduces strongly, but the ability gap is unresolved at small n — see #65 (anchoring 9/9 separated by preset; anchored−standard +3.3, 95% CI [−2.6, +9.3]) and #51 (multi-env, Ability 85–90, 98/99 not reproduced). Treat the scores above as our original observations, not a settled effect size.
Research write-ups live in the companion exploration repository DeepseekCotexplorations (data and methodology; this repository keeps the code):
- Tool-surface dose-response + Project2 replication — methodology, per-run scope, tool-surface experiments, and limitations.
- Anchor-mass quantification + single-request probe methodology — the prefab template quality model and the post-price-hike low-cost evaluation loop.
Development-process records (what was done, why, and the pitfall lists) are
kept in this repository as HANDOFF.md and
HANDOFF-2.md.
Configuration reference
All knobs are rows in each mode's agent.cordis.yml. Unknown keys fail at
preset mount.
context-gate (mounted FIRST in preset/, zero-anchored-standard/, and
whoami-standard/ — waterfall registration order makes the gate the outermost
transform; the plugin lives in shared/context-gate.mjs and is reusable by
any other composition that wants unified injection control alone):
| Key | Default | Meaning |
|---|---|---|
promoteOn |
either |
Promotion trigger: either, tool-call, or assistant-message (the variants use assistant-message). |
includeSubagents |
false |
Gate subagents too (true in the base mode and whoami; keep in sync with the bootstrap row). |
enabled |
true |
false disables both interception paths (A/B testing without touching the row set). |
allowKinds |
[skill-invocation] |
source.kind values allowed beyond the claimed batch; [] keeps ONLY the claimed batch. |
Injection control division of labor: session-phase suppression (everything
keyed on a promotion boundary) belongs to context-gate. Two documented
exceptions keep their own enumerated suppressedContextSources strip because
the gate's phase machine does not map onto their scope: the think-step strip
in think-phase/wire-think (per-step, not per-session-phase) and the
permanent every-request strip in eternal-minimal (no promotion boundary;
frozen to the configuration its recorded measurements were taken under).
tool-bootstrap (in preset/agent.cordis.yml; mount right after
context-gate):
| Key | Default | Meaning |
|---|---|---|
bootstrapTools |
[bash, str_replace_editor] |
Tools visible on request #1. |
promoteOn |
either |
Promotion trigger: either, tool-call, or assistant-message. |
bootstrapMaxTokens |
unset | Optional output cap for request #1; stripped after promotion. |
includeSubagents |
false |
Subagents take the bootstrap phase too (true in the base mode). |
compactionTools |
[] |
Extra tools available between a compaction boundary and re-promotion. |
zero-tool-bootstrap (in zero-anchored-standard/ and whoami-standard/):
compactionTools has the same semantics (promotion is always the first
assistant/message), plus includeSubagents, whether subagents also take the
anchor phase (set true in whoami-standard, false in
zero-anchored-standard). Context suppression is NOT here — both variants
mount the context-gate row (above) with promoteOn: assistant-message;
the bootstrap's former suppressedContextSources key now fails at mount.
anchor-turn (in both variants): text — the synthetic first user message
(default "This round is a test. Tools are not open yet; all tools will open
next round." in zero-anchored, "你是谁" in whoami); includeSubagents —
whether subagents also take the anchor turn.
eternal-minimal (in eternal-minimal/; the row must stay FIRST):
| Key | Default | Meaning |
|---|---|---|
guide |
true |
Append the short dshx capability guide to the system prompt; false keeps the persona byte-pure. |
gateway |
true |
Intercept dshx shell commands and execute the real tools; false leaves the bare Minimal pair. |
gatewayCommand |
dshx |
The interception word. |
maxGatewayChars |
12000 |
Cap on one gateway result payload. |
suppressedContextSources |
[agent-instructions, skill-catalog] |
Stripped on every request (no promotion boundary; the enum intentionally stays — see the division-of-labor note above). |
cot-drip (in combo-anchored/):
| Key | Default | Meaning |
|---|---|---|
every |
4 |
Attach one deliberation beat after every Nth tool result; 0 disables the drip. |
maxPerTurn |
1 |
Beats per turn. |
text |
built-in beat | The reminder text (one "We …" sentence restating the remaining goal). |
includeSubagents |
false |
Whether subagent calls are dripped too. |
toolchoice-adapter (in wire-think-standard/; the row must stay the first LOCAL row):
| Key | Default | Meaning |
|---|---|---|
provider |
deepseek-wire-think |
The sibling route id the adapter owns; registering an id twice throws DUPLICATE_ADAPTER (caught, degraded). |
toolChoice |
none |
The wire tool_choice sent whenever tool definitions are present. |
baseURL / apiKeyEnv |
settings/env | Row config first, then the llm-deepseek settings section, then DEEPSEEK_BASE_URL / DEEPSEEK_API_KEY. |
logprobs |
false |
Opt-in research hook: request token logprobs and log a per-request mean summary (no StreamChunk surface exists). |
wire-think (in wire-think-standard/): same mode / suppressedContextSources /
includeSubagents / steerText semantics as think-phase, plus
provider (must match the toolchoice-adapter row's id) and defaultProvider
(the route execute steps restore onto, default deepseek-official).
instruction-hint (all modes): promoteOn matching the mode's promotion
semantics (either in the base mode, assistant-message in the variants) —
the one-shot "instruction files exist, read them before acting" hint waits
for promotion.
Repository layout
preset/ Anchored Standard — the base mode
zero-anchored-standard/ variant: fixed zero-tool anchor turn
whoami-standard/ variant: "你是谁" anchor turn, subagents inherit
eternal-minimal/ variant: Minimal pair forever + dshx bash gateway
wire-think-standard/ variant: wire-level condition (tools + tool_choice=none)
combo-anchored/ combination package: think split + gate + drip rows
shared/ single source of truth for plugins used by 2+ modes
scripts/sync-modes.mjs materializes shared/ plugins into every mode dir
test/ zero-dependency test suite (npm test)
verify/ one-shot headless verification runner
prefab/ Prefab Anchored Standard + bundled session template
prefab/ ships a generic template by default and a Project2-specific template
as an explicit opt-in. Both contain real model reasoning; read the mode's
installation notes before use.
Invariants, enforced by npm run check:
- Every mode directory is self-contained: installable by copying it alone;
agent.cordis.ymlrows may reference only./local.mjsfiles, never../. - Plugins shared by several modes live once in
shared/; the copies in mode directories are generated. Editshared/, runnpm run sync, commit both — never edit a materialized copy. - The
context-gaterow stays the FIRST row ofpreset/agent.cordis.yml(the gate must register before every injecting plugin), withtool-bootstrapright after it. The same first-row rule holds for thecontext-gaterow inzero-anchored-standard/andwhoami-standard/.
This repository deliberately ships no AGENTS.md/CLAUDE.md: the presets' whole mechanism is a clean request #1, stripping exactly those instruction-file digests from it (issue #6: 0/9 anchored with the injection present). Shipping one would only feed later rounds and contradict the mechanism being documented. Everything an assistant needs is in this README.
Compatibility
Developed and tested against:
- DeepSeek Harness
0.1.0-rc.5 - repository commit
47f9438 - Node.js 24 on Windows
The persistent shell resolves shellPath adaptively: it keeps the
terminal-bash plugin default /bin/bash on hosts where that absolute path
exists, and falls back to bash (PATH lookup) otherwise — e.g. NixOS, where
bash lives under the Nix store. Hosts that ship /bin/bash keep the previous
behavior exactly; the fallback only activates where the default would make
every bash call fail with "PTY shell exited during startup".
On the 0.1.0-rc.5 source checkout, bootstrapMaxTokens reaches the actual
first request (the first request/header records the cap, adapterDefaults
stays empty), because llm.prepareCall only materializes a default maxTokens
when the proposed config has none. One prebuilt profile package observed in
issue #11 (CLI launcher reporting 0.1.0-rc.6) overwrote the proposed cap
with adapterDefaults.maxTokens; there the cap is a no-op. The default
composition therefore relies on the Minimal tool schema alone (which anchors
at the adapter default with no cap) and leaves bootstrapMaxTokens as an
opt-in for standard-schema bootstraps.
DeepSeek Harness is currently a developer preview and explicitly permits breaking changes. This preset is a full snapshot of the Standard composition, so review upstream changes before using it with a newer release.
Install
For the prefab mode, the recommended path is AI-assisted one-command setup.
Give your coding agent this repository and ask it to follow the
installation-agent contract. When it reports
INSTALL READY, start DSH, select Prefab Anchored Standard, create a new
session in the target workspace, and send the real task prompt. This installs
the generic template; the Project2 benchmark template requires an explicit
--template project2 selection and installs under a separate preset id.
Clone this repository, then copy the entire preset directory into the user
preset root under the id anchored-standard. Every mode directory in this
repository is self-contained: the zero-anchored-standard/,
whoami-standard/, prefab/, eternal-minimal/, wire-think-standard/, and
combo-anchored/ variants install the same way, alone or together, with no
other directory required (see their sections below). prefab/ automatically
hydrates newly selected sessions; follow prefab/README.md.
PowerShell:
$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\anchored-standard'
if (Test-Path -LiteralPath $target) { throw "Preset already exists: $target" }
New-Item -ItemType Directory -Force -Path (Split-Path -Parent $target) | Out-Null
Copy-Item -Recurse -LiteralPath '.\preset' -Destination $target
Linux/macOS:
dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
test ! -e "$dsh_home/.agent-presets/anchored-standard"
cp -R preset "$dsh_home/.agent-presets/anchored-standard"
Fully restart DeepSeek Harness, create a blank session, and select Anchored Standard (experimental). Do not switch an active session from a different preset.
Verify
Export the session JSONL and inspect request/header events. Reproduction
checklist (issue #11 asks for the first two explicitly, because both are the
variables that decide the anchor):
- First-request
config.maxTokensvalue: withbootstrapMaxTokensunset (the default), the first header records the adapter default (e.g. 256000 withadapterDefaults.maxTokens: true); with a cap configured it records the cap (e.g. 1024 with no maxTokens adapterDefault). - First-request tool schema source: the first header's
toolsarray must be exactly["bash", "str_replace_editor"]— the official Minimal preset's real schemas, not Standard'spwsh/read. - the first request's messages should contain no AGENTS.md/CLAUDE.md digest and no available-skills reminder — only the user message and the minimal persona system prompt;
- after the first tool call or the first assistant reply, the next changed
header should contain the promoted resident catalog: the bootstrap pair plus
dev_tool_search/skill_search/skill_loadplus any tools the model already unlocked; - subsequent requests should keep that resident set (it grows only through
explicit
dev_tool_searchunlocks) and restore the standard context injections.
Run the local zero-dependency tests with:
npm test
Important behavior
- With the default
promoteOn: either, the session promotes after its first durabletool/callOR its firstassistant/message, whichever comes first — request #1 sees the bootstrap catalog and every later request sees the resident catalog. A text-only first reply therefore still promotes at request #2; setpromoteOn: tool-callto restore the original behavior, where a first response that makes no tool call never promotes. - A failed tool execution still promotes the session because the durable
tool/callalready exists. - The first request's output budget is NOT capped by default: the Minimal tool
schema anchors at the adapter-default maxTokens, so
bootstrapMaxTokensis opt-in. When set, the first request is capped and the cap is explicitly stripped after promotion (the next request's seed proposal carries the previous header's maxTokens forward). - The promoted catalog is the RESIDENT set — the bootstrap pair plus the
discovery tools plus everything the model unlocked via
dev_tool_search— not the full Standard dump. The Standard sandboxedbashrow stays disabled in favor of the persistent shell (same tool name, same layer; see Why). When unlocked, theread/write/edittools keep the sandboxed filesystem whilestr_replace_editoruses the preset's local fs. - A missing bootstrap tool degrades to the full catalog with a one-time
warning instead of failing requests, so a composition drift cannot brick a
session; invalid
promoteOnvalues fail at preset mount instead. - Promotion decisions are memoized per session for the process lifetime; the durable event scan runs once per session per process.
- While a session is unpromoted, the
context-gateplugin closes BOTH unified injection paths: the assembly's runtime-context contributions are blanked (the wholeSystemPrompt.context()family, without enumerating sources), and the pre-step waterfall keeps only the claimed batch plus theallowKindsentries. At promotion the loop's snapshot projection diffs in exactly ONE fresh runtime-context message; a gate failure degrades to keeping every message rather than eating context. - The tool catalog changes at promotion and again whenever
dev_tool_searchunlocks a new tool; request-prefix cache continuity breaks at those points. - The preset has the same trust level as shell access. Review its files before installation.
- The plugin performs no network requests and adds no telemetry.
Zero-Anchored Standard (experimental)
An extra test mode that does not change the Anchored Standard logic above. It uses the same Minimal-aligned system prompt, but instead of exposing two tools on the first request it injects one fixed zero-tool anchor turn:
- When the user sends their first message, the
anchor-turnplugin prepends a fixed user message — "This round is a test. Tools are not open yet; all tools will open next round." — ahead of it. - The first real model request carries ZERO tools, so the session's first reasoning chain follows the zero-injection "we" trajectory.
- Once that anchor response is durable, the resident catalog is exposed and the real message proceeds with it.
Anchoring on the first message — not on session creation — keeps the blank-session preset switcher usable. Subagents always see the resident catalog.
Measured behavior (opencode-go, DeepSeek V4 Pro, reasoningEffort=max): the
anchor request is stable "we"-style with zero let me; the following
tool-bearing requests return to the "The user wants…/Let me" style. This mode
is a comparison point for whether the zero-tool first turn is worth the extra
model call — not a claim that tool rounds stay "we"-style.
Install as a separate preset id:
dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
test ! -e "$dsh_home/.agent-presets/zero-anchored-standard"
cp -R zero-anchored-standard "$dsh_home/.agent-presets/zero-anchored-standard"
Restart DeepSeek Harness, create a blank session, select Zero-Anchored Standard (experimental), then send your first message.
Whoami Standard (experimental)
A usability-oriented variant of the zero-tool anchor idea: the first turn is a natural self-introduction prompt instead of a fixed test message, and the user's real first message is deferred to the next turn. Whatever the user types first, the session warms up exactly one round and everything is ready when the real message is processed:
- When the user sends their first message, the
anchor-turnplugin prepends a fixed user message — "你是谁" (who are you) — ahead of it in thenext-turninbox queue. - dsh claims exactly ONE
next-turnmessage per turn, so the first model request sees only the anchor on an EMPTY tool surface and replies with a self-introduction; that reply is the promotion signal. - The real message is claimed by the NEXT turn, with the promoted resident
catalog (shells,
str_replace_editor, the discovery tools) already unlocked — heavier Standard tools are onedev_tool_searchaway.
The anchor text is configurable via the anchor-turn row's text option
(default "你是谁"). Anchoring on the first message — not session creation —
keeps the blank-session preset switcher usable.
Full-powered subagents
Whoami Standard ships with includeSubagents: true on both the
zero-tool-bootstrap and anchor-turn rows, so subagents spawned from a
session inherit the same anchor flow as top-level sessions:
- A newly spawned subagent's first model request sees only the "你是谁" anchor on an empty tool catalog.
- The subagent's self-introduction reply is the promotion signal.
- The delegated prompt runs on the next turn with the promoted resident
catalog (shells,
str_replace_editor, and the discovery tools).
Set includeSubagents: false on both rows to restore the plain behavior,
where subagents start with the resident catalog immediately. Each subagent
costs one extra model call for its anchor turn — a delegation-heavy session
pays it per subagent.
zero-anchored-standard keeps subagents plain by default; enabling the same
flow there means setting includeSubagents: true on its zero-tool-bootstrap
and anchor-turn rows (its anchor text stays the fixed test notice).
The trade-off of the mode itself is one extra model call per session: the anchor turn is always taken, even when the first message is urgent.
The directory is self-contained; install it alone or alongside any other mode.
Install as a separate preset id:
dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
test ! -e "$dsh_home/.agent-presets/whoami-standard"
cp -R whoami-standard "$dsh_home/.agent-presets/whoami-standard"
Restart DeepSeek Harness, create a blank session, select Whoami Standard (experimental), then send your first message — the self-introduction round runs first, and your message is answered with the full tooling on the next turn.
Eternal Minimal (experimental)
The "make the model believe it never left Minimal" mode: the model-visible
catalog stays EXACTLY the Minimal pair (bash + str_replace_editor) for
the WHOLE session — no anchor round, no promotion, no discovery tools, no
catalog growth — while the full Standard toolset stays registered and
executes FOR REAL behind the dshx bash gateway:
dshx list # list every gateway tool
dshx web_search '{"query": "..."}' # execute the real web_search
dshx read_image '{"path": "..."}' # execute the real read_image
- Eternal pair:
system-prompt/assemblekeeps only the shells +str_replace_editoron every request (think steps, post-compaction, subagents — everything), and auto-injected context is stripped everywhere (there is no promotion boundary to key suppression on). - Gateway: a
tools/pre-executelistener intercepts bash commands starting withdshx, dispatches them throughctx.tools.execute()(the full registry pipeline — policy, guards, execution, rendering), and returns the rendered output as the command result. The deny channel is the only sanctioned pre-dispatch way to substitute a result, so gateway payloads arrive flagged as errors — every payload states plainly that the tool executed and its output follows, so the model reads it as output. The real tool really ran: the user sees genuine effects (files, searches, subagents) exactly as if it had been called by name. - Guide: a short
dshxcapability guide is appended to the system prompt (guide: falsefor a byte-pure Minimal persona) so the model knows the gateway exists without a third visible tool.
The gateway refuses to dispatch the shells/str_replace_editor themselves
("invoke them directly"), which also makes recursion impossible. Unknown
tools, malformed JSON, and tool failures all come back as readable payloads.
Set gateway: false for a bare two-tool session with no interception.
Install as a separate preset id:
dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
test ! -e "$dsh_home/.agent-presets/eternal-minimal"
cp -R eternal-minimal "$dsh_home/.agent-presets/eternal-minimal"
Restart DeepSeek Harness, create a blank session, select Eternal Minimal
(experimental), then work as usual — the model composes shell commands,
and dshx … lines run the heavier Standard tools for real.
Wire Think-Execute Standard (experimental)
The wire-level mode: every user turn opens with one think step that keeps
tool definitions PRESENT in the request while the wire forbids invocation
(tool_choice: "none"), then a steering notice opens the execute phase on
the official provider with the resident catalog.
tool_choice is outside the harness GenerateOptions vocabulary (the
official deepseek adapter documents the mapping as an MVP cut), so reaching
this condition takes the sanctioned wire seam:
- Sibling route:
toolchoice-adapter.mjs(row 1) registers a zero-dependency DeepSeek chat-completions adapter under its OWN provider id (deepseek-wire-think) that putstool_choice: "none"on the wire whenever tool definitions are present. The officialDeepSeekAdaptercannot be wrapped (its wire body is built inside a private generator), so this file vendors a minimal, protocol-faithful subset of the official serialize/SSE/translate pipeline — the same assistant-message nuances (content: ""never null,reasoning_contentreplayed only on tool-call turns, tool results asrole: "tool"with an(no output)fallback) and the same usage/finish translation. Connection facts resolve row config >llm-deepseeksettings section > env, exactly like the official row, so the sameDEEPSEEK_API_KEYserves both routes. - Per-step routing:
wire-think.mjskeeps the think step's assembled catalog UNTOUCHED (that is the condition being reproduced) and swaps ONLY the provider in theagent/requestwaterfall — the frozen loop-built request and the log-reconstructability invariant are preserved. Execute steps (and every subagent) are routed back to the captured original provider even when the folded session header seeds them with the think route. - Steer + resident:
agent/turn-stoppingsteers exactly once per turn (resume-safe from durablesteering/messageevents), and execute steps see the promoted RESIDENT set.
Degradation ladder: if the sibling route is not registered (row removed, or
a second preset already mounted the same id — DUPLICATE_ADAPTER is caught
and warned), think steps fall back to the zero-tool condition, so a
composition mistake can never brick a session. mode: first-turn limits the
routing (and its costs) to the session's first user turn.
Costs to know before adopting: the think/execute alternation switches the
tools block of the request prefix twice per turn, so DeepSeek prefix-cache
reuse breaks from the first changed token each switch (provider id itself is
invisible to the backend cache; the tools block is what diverges). Each swap
appends a request/header change event. Set logprobs: true on the adapter
row for the opt-in research hook — the adapter requests token logprobs and
logs a per-request mean summary (the harness StreamChunk vocabulary has no
surface for logprob data, so logging is all a plugin can do today; that log
stream is exactly what offline trajectory analysis would consume).
Install as a separate preset id:
dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
test ! -e "$dsh_home/.agent-presets/wire-think-standard"
cp -R wire-think-standard "$dsh_home/.agent-presets/wire-think-standard"
Combo Anchored (experimental) — the combination package
The everything-is-a-plugin showcase: THREE orthogonal anchoring mechanisms
composed as independent rows, each with its own knobs, each removable or
retunable by editing one line of agent.cordis.yml. They attack the
pre-tool deliberation collapse at different moments of a turn:
| Row | Mechanism | Owns |
|---|---|---|
think-phase |
zero-tool think step + steering notice | the turn OPENING |
deliberation-gate |
depth gate denies the first tool call of a shallow turn | the FIRST ACTION |
cot-drip |
one "We …" beat after every Nth tool result (tools/post-execute additionalContexts — never blocking, never erroring) |
the LONG MIDDLE |
With mode: every-turn the think step opens every turn, the gate catches
the paths that skip it (steering continuations, resumed sessions,
straight-to-tools follow-ups), and the drip sustains deliberation across
long tool loops. Defaults are deliberately gentle (minChars: 400,
every: 4, one beat per turn); tune per workload. Swapping the
think-phase row for wire-think + toolchoice-adapter upgrades the
opening to the wire-level condition (see above) at the cost of the sibling
route and its prefix-cache churn.
Explored and rejected for this package: pure Code Mode presentation
(presentAs('code') collapses the catalog into one run_code tool) — a
single-tool surface measurably underperforms the two-tool condition in the
sibling project's evaluations; and text-only fake tools or ghost tool-call
histories — both proved unreliable anchors in practice.
Official ecosystem guidance
DeepSeek currently asks community plugin authors to publish plugins in their own
GitHub projects and add the dsh-plugin
repository topic for discovery. The official repository does not currently
accept external pull requests and does not mandate a community repository
template. See the official
CONTRIBUTING.md.
License
MIT. preset/agent.cordis.yml is derived from the DeepSeek Harness Standard
preset; the original DeepSeek copyright and MIT notice are retained in
NOTICE.
相关推荐Related
- deepseek-ai/deepseek-harness DeepSeek Harness: Everything is a Plugin.
- deepseek-ai/deepseek-harness DeepSeek Harness (dsh) — 由 DeepSeek AI 开发的开源 agent harness（智能体框架）。一切皆插件：核心 + 插件 + 技能 = 可热插拔的 AI 编程工作台。DeepSeek Harness: …
-
ruvnet/ruflo
🌊 The original agent meta-harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build c…
-
amruthpillai/reactive-resume
A one-of-a-kind resume builder that keeps your privacy in mind. Completely secure, customizable, portable, open-source a…
-
volcengine/OpenViking
Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.
-
Molunerfinn/PicGo
:rocket: The Ultimate Image Uploader for Efficient Creators. Supports Obsidian, Typora, VS Code etc. and 60+ image hosti…
-
titanwings/colleague-skill
将冰冷的离别化为温暖的 Skill，欢迎加入数字生命1.0！Transforming cold farewells into warm skills? It's giving rebirth era. Welcome to Digital …
-
nocobase/nocobase
NocoBase is an open-source AI + no-code platform for building business systems fast. Instead of generating everything fr…