dsh-deliberation-presets

中文说明

Keep DeepSeek V4 Pro deliberating deeply while its tools keep working.

Three experimental DeepSeek Harness (DSH) agent presets. The design premise comes from hands-on sessions with V4 Pro: its strongest reasoning — the "We …" chain — is highly conditional on the request surface, and the three modes below restore it at different moments of a conversation without ever taking the tools away:

Mode	Directory	One line
Wire Think-Execute Standard	wire-think-standard/	every turn opens with a think step that keeps the tool catalog visible while the wire forbids invocation (tool_choice: "none"), then executes on the official route
Eternal Minimal	eternal-minimal/	the visible catalog stays the Minimal pair forever; heavier Standard tools run FOR REAL behind the dshx bash gateway
Combo Anchored	combo-anchored/	the combination package — think/execute split + deliberation gate + deliberation drip as three independent rows

Not an official DeepSeek product. Sibling of dsh-anchored-standard (the first-request anchor modes); this repository holds the whole-conversation deliberation mechanisms.

Background

Observed on DeepSeek V4 Pro in our own sessions (the sibling repository documents the evaluation methodology):

1. The "We …" reasoning chain survives when the request carries no callable tool catalog, and collapses into a thin act-now loop the moment one appears — pre-action deliberation shrinks to a fraction of its no-tools depth.
2. Keeping the tool DEFINITIONS visible while the wire forbids invocation (tool_choice: "none") preserves the full planning context during deliberation — the model still sees everything it will be allowed to use — and leaves tools one parameter away from working again.
3. A directive push-back while tools stay live is the intervention shape that both prompts deeper reasoning and keeps tool calls working.

The three modes operationalize these: wire-think-standard builds its think steps on (2), eternal-minimal keeps the surface that (1) rewards for the whole session while routing real tool execution behind a gateway, and combo-anchored combines a zero-tool think opening, a (3)-style depth gate for the first action, and a maintenance drip for the long middle of tool loops.

Modes at a glance

Mode	Directory	First model request	Anchor mechanism	Promotion signal	Cost
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, tools-prefix cache switch
Eternal Minimal	eternal-minimal/	2 tools, forever	the visible catalog never grows; heavier tools run via the dshx bash gateway	none (no phases)	none
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).

Terminology

• trajectory — the style of the model's first reasoning chain; the Minimal-family conditions produce "We need…" first lines, the Standard condition "Let me…".
• think step / execute step — step 0 of a user turn (deliberation, tools withheld or wire-forbidden) vs every later step of the same turn.
• resident catalog — the narrow promoted tool set: the shells + str_replace_editor + the discovery tools (dev_tool_search, skill_search, skill_load) plus everything the model explicitly unlocked.
• sibling route — a second DeepSeek provider id pointing at the same backend/key, registered by a preset-local adapter that owns its wire serialization.
• steer — agent.steer(...) at the agent/turn-stopping checkpoint; forces one more step instead of letting a text-only think reply close the turn.
• materialized copy — the committed copy of a shared/ plugin inside a mode directory, generated by npm run sync.

Wire Think-Execute Standard (experimental)

The wire-level mode: every user turn opens with one think step on condition (2) — tool definitions PRESENT in the request, tool_choice: "none" on the wire — 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:

1. Sibling route: toolchoice-adapter.mjs (row 1) registers a zero-dependency DeepSeek chat-completions adapter under its OWN provider id (deepseek-wire-think) that puts tool_choice: "none" on the wire whenever tool definitions are present. The official DeepSeekAdapter cannot 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_content replayed only on tool-call turns, tool results as role: "tool" with an (no output) fallback) and the same usage/finish translation. Connection facts resolve row config > llm-deepseek settings section > env, exactly like the official row, so the same DEEPSEEK_API_KEY serves both routes.
2. Per-step routing: wire-think.mjs keeps the think step's assembled catalog UNTOUCHED (that is the condition being reproduced) and swaps ONLY the provider in the agent/request waterfall — 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.
3. Steer + resident: agent/turn-stopping steers exactly once per turn (resume-safe from durable steering/message events), 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).

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

1. Eternal pair: system-prompt/assemble keeps only the shells + str_replace_editor on every request (think steps, post-compaction, subagents — everything), and auto-injected context is stripped everywhere (there is no promotion boundary to key suppression on).
2. Gateway: a tools/pre-execute listener intercepts bash commands starting with dshx, dispatches them through ctx.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.
3. Guide: a short dshx capability guide is appended to the system prompt (guide: false for 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.

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.

Configuration reference

All knobs are rows in each mode's agent.cordis.yml. Unknown keys fail at preset mount.

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.

wire-think (in wire-think-standard/):

Key	Default	Meaning
mode	every-turn	every-turn opens every user turn with a wire-forbidden think step; first-turn limits it to the session's first user turn.
provider	deepseek-wire-think	Must match the toolchoice-adapter row's id.
defaultProvider	deepseek-official	The route execute steps restore onto.
suppressedContextSources	[agent-instructions, skill-catalog]	source.kind values stripped during think steps; [] disables.
includeSubagents	false	Whether subagents also think first.
steerText	built-in notice	The steering message that opens the execute phase.

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 (there is no promotion boundary).

think-phase (in combo-anchored/), deliberation-gate (in combo-anchored/), cot-drip (in combo-anchored/): see the mode directory headers — mode / suppressedContextSources / includeSubagents / steerText for the split; minChars (default 400, 0 disables) / maxGatesPerTurn (default 1) / gateText / includeSubagents for the gate; every (default 4, 0 disables) / maxPerTurn (default 1) / text / includeSubagents for the drip.

Repository layout

wire-think-standard/    wire-level think/execute split
eternal-minimal/        Minimal pair forever + dshx bash gateway
combo-anchored/         combination package: think split + gate + drip
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

Invariants, enforced by npm run check:

• Every mode directory is self-contained: installable by copying it alone; agent.cordis.yml rows may reference only ./local.mjs files, never ../.
• Plugins shared by several modes live once in shared/; the copies in mode directories are generated. Edit shared/, run npm run sync, commit both — never edit a materialized copy.

Compatibility

Developed and tested against:

• DeepSeek Harness 0.1.0-rc.5
• repository commit 47f9438
• Node.js 24 on Windows

DeepSeek Harness is currently a developer preview and explicitly permits breaking changes. Every mode is a full snapshot of the Standard composition, so review upstream changes before using them with a newer release.

The wire mode additionally assumes the DeepSeek backend accepts the OpenAI-style tool_choice parameter (the official adapter simply never sends it). If the backend rejects it, the think step fails through the normal request-error path — degrade to another mode or mode: first-turn if that happens.

Install

Clone this repository, then copy a mode directory into the user preset root. Every mode installs alone; the order fields (11/9/12) keep the three from colliding with the dsh-anchored-standard family if you install both.

Linux/macOS:

dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
for m in wire-think-standard eternal-minimal combo-anchored; do
  test ! -e "$dsh_home/.agent-presets/$m" && cp -R "$m" "$dsh_home/.agent-presets/$m"
done

PowerShell (one mode at a time):

$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\wire-think-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 '.\wire-think-standard' -Destination $target

Fully restart DeepSeek Harness, create a blank session, and select the mode by name. Do not switch an active session from a different preset. The wire mode needs DEEPSEEK_API_KEY visible to the process (same key as the official route).

Verify

Export the session JSONL and inspect request/header events:

• Wire mode: think-step headers show provider: "deepseek-wire-think" with the tool catalog present; execute-step headers show the official provider and the resident catalog; turn/end reasons show the steered continuation.
• Eternal Minimal: every header's tools array is exactly ["bash", "str_replace_editor"]; tool/call events show dshx … commands whose results carry the real tool output.
• Combo: think steps show zero tools, gated calls show one deny result with the directive text, and drip beats appear as short plugin-sourced user messages after tool results.

Run the local zero-dependency tests with:

npm test

For a headless one-shot check against a real harness checkout, see verify/run-verify.mjs --help.

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. See the official CONTRIBUTING.md.

License

MIT. The mode compositions derive from the DeepSeek Harness Standard preset, and shared/toolchoice-adapter.mjs adapts a subset of the official llm-deepseek adapter pipeline; the original DeepSeek copyright and MIT notice are retained in NOTICE.