dsh-model-pipeline
Verifieddsh-model-pipeline · v0.10.4 · MIT · Web UI
Context-budgeted, quality-gated DSH collaboration with role fallbacks, tool checks, and bounded refusal recovery.
Install
dsh plugin add dsh-model-pipeline Confirm the layer applied with dsh --profile default --dump-config — see the install guide.
Source
Published to npm without a public repository. Inspect the package contents before installing.
Tags
Creators
Readme
Model collaboration (DSH plugin)
This plugin gives the current DSH agent a role-based model collaboration loop. Every role can use a different provider/model; roles may share a model when a smaller deployment is preferred.
task + conversation + constraints
-> classifier: detect explicit, indirect, partial, evasive, or stalled refusal
-> interrogator: ask the main model for missing definitions, context, and parameters
-> planner: give clues, a narrow assignment, boundaries, and acceptance criteria
-> finisher: perform the specific last-mile work the main model did not do (tools allowed)
-> reviewer: check every claim, parameter, edge case, and acceptance criterion
-> (optional correction rounds)
finisher revises -> reviewer checks again
-> current main agent runs pending checks and delivers the original task
If the planner initially refuses to reason, the interrogator's concrete questions are sent back in a coaching-only retry that asks for clues, decomposition, definitions, and checks instead of a finished answer. The finisher and reviewer run as DSH subagents when that service is available, so they can use the tools authorized for the parent task. Direct model calls remain as a compatibility fallback.
Tools
collaborate_run accepts task, optional context, constraints, and mainDraft. It returns the
questions, main-model briefing, focused contribution, per-detail verdicts (pass, fail, or
unverified), pending tool checks, revision trail, and instructions for the main agent.
pipeline_run remains as a compatibility alias with the same collaboration behavior.
The reviewer scores correctness, completeness, evidence, and actionability from 0 to 100. Every
dimension must meet qualityThreshold; ready: true alone cannot pass the gate. Failed role calls
advance through roleFallbacks, so one unavailable or uncooperative model does not terminate the
task.
If full review is refused, the plugin decomposes it into at most maxReviewChecks narrow audits over
acceptance criteria, constraints, facts, and expected checks. Each audit tries the reviewer followed
by reviewerFallbackRoles. An audit that all strong roles refuse becomes unverified with a concrete
parent-tool check; it does not terminate the collaboration.
The default role selection deliberately uses the registered localqwen model only for interrogator and finisher.
main, planner, and reviewer remain strong-model roles. If you add fallbacks for those roles,
they should be models of comparable quality; local output is never an authority or final answer.
Thinking models
A model that spends its whole budget in reasoning_content returns no visible text. Both call paths handle that:
the direct HTTP path already fell back to reasoning_content, and the DSH path now does the same for a stream that
carries only reasoning deltas. The reasoning text is a last resort, not a preferred answer, so keep outputRatio
large enough that the model can finish thinking and still answer.
Automatic handoff
With auto: true, the plugin watches each completed main-model response. Deterministic rules catch
obvious failures and authorization boundaries; with semanticRefusalDetection: true, a dedicated
classifier also detects indirect, partial, evasive, and stalled refusals. Its structured result records
the refusal type, confidence, and whether collaboration can recover it. Classification failures fall
back to the deterministic result. Policy, safety, and authorization refusals follow the same recovery path; with
recoverSafetyRefusals: false they are recorded but not
delegated to a less constrained model. When DSH exposes collaborate_run, the plugin substitutes a native tool call: DSH runs the
collaboration, returns the work packet to the same main-agent loop, and the main model can continue
using its normal tools. A direct main-model synthesis is retained as a compatibility fallback. If
collaboration or synthesis fails, the original response is preserved. Normal apologies and
corrections do not trigger automatic handoff. A continued inability can trigger another collaboration
round with the prior packet in context, up to maxAutoHandoffs; this bounds recovery without making
the first refusal terminal.
Configuration
Version 0.9.0 includes a graphical configuration page. Open DSH Settings, choose Plugins, open the
model-pipeline row, and select Configure. The page edits every role's registered DSH model/output
limit, role context budgets, collaboration rounds, timeouts, and the
quality gate. Every role uses the same provider-grouped model catalog as DSH's /model picker; endpoint,
credentials, protocol, and discovery remain owned by DSH. Changes are read dynamically by subsequent
calls; cordis.patch.yml remains available for declarative deployments.
- insert:
- id: model-pipeline
name: dsh-model-pipeline
config:
auto: true
semanticRefusalDetection: true
classifierConfidence: 75
collaborationRounds: 2
qualityThreshold: 80
maxAutoHandoffs: 2
maxReviewChecks: 10
reviewerFallbackRoles: [main, planner]
requestTimeoutMs: 120000
# Budgets derive from each role model's context window; these are ceilings.
budgetMode: model # model | fixed
charsPerToken: 3 # 1..6, chars per token when converting a window
contextRatio: 0.75 # share of the window a role may use
windowRatio: 0.9 # prompt + reply must fit inside this share
outputRatio: 0.15 # share of the window a single reply may generate
# Two fuses only; every role budget comes from that model's own window.
contextChars: 150000 # global prompt-character ceiling
maxTokens: 12000 # global reply-token ceiling
roles: # provider + model only; maxTokens is derived
classifier: { provider: opencode, model: deepseek-v4.1-flash }
main: { provider: opencode, model: deepseek-v4.1-flash }
interrogator: { provider: localqwen, model: /model/HuggingFace/Vincentt/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF-Q4_K_P/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf }
planner: { provider: opencode, model: deepseek-v4.1-flash }
finisher: { provider: localqwen, model: /model/HuggingFace/Vincentt/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF-Q4_K_P/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-Q4_K_P.gguf }
reviewer: { provider: opencode, model: deepseek-v4.1-flash }
# Optional per-role overrides, still bounded by that model's window:
# roleContextChars: { interrogator: 60000 }
roleFallbacks: {}
Legacy roles.local / roles.drafter configure both local roles, while roles.fixer configures the
reviewer. fixRounds maps to collaborationRounds.
Context budgets
contextChars is the global ceiling. roleContextChars applies a smaller ceiling per role. The
interrogator receives a relevance-selected excerpt, while the finisher receives only the focused
assignment, compact facts and criteria, a small source excerpt, and bounded revision feedback. The
planner and reviewer retain larger budgets because they are the strong-model roles responsible for
semantic compression and verification. Long context is selected from the beginning, relevant
segments, and recent tail rather than being cut at one arbitrary point.
Each role may also set maxTokens for its output, preventing a small local context window from being
consumed by an unnecessarily long completion.
State and status
GET /model-pipeline.json returns the latest collaboration report. The state file keeps the full
latest work packet across restarts. The JSON-lines log stores run metadata only, without prompts or
model output. The plugin does not directly spawn shell commands. Tool-capable DSH subagents may use
only tools exposed and authorized by the host for the parent task, so normal host approval and
sandbox rules still apply.
Quality evaluation
npm run evaluate compares the configured main model alone with the full collaboration chain on
benchmark/cases.example.json. A reviewer scores both final answers and reports the average score
delta and win counts. Use -- --config config.json --output report.json to supply model settings and
save the report. Evaluation makes real model calls and is never run automatically.
License
MIT