dsh-codex-reasoning-router

简体中文 | English | AI / LLM context

DeepSeek Harness (dsh) preset and plugin: GPT-5.6 Luna executes with the full Standard toolset; tool-less GPT-5.6 Sol supplies compact reasoning advice.

Luna owns action. Sol owns advice.

This DeepSeek Harness plugin keeps openai-codex / gpt-5.6-luna as the root execution agent and uses openai-codex / gpt-5.6-sol only for short, independent reasoning-advisor calls.

Sol is not a second coding agent. Sol never touches the workspace, never receives tools, never creates subagents, and never answers the user directly. Luna owns every tool, file change, test, skill, MCP call, subagent, permission decision, and final response.

Requirements and preset installation

Requirements:

• a current DeepSeek Harness installation;
• dsh-codex >= 0.2.3, installed and authenticated;
• access to openai-codex/gpt-5.6-luna and openai-codex/gpt-5.6-sol;
• Node.js 22 and pnpm for source development (verified with Node 22.23 and pnpm 11.7).

Install and authenticate the current dsh-codex first. Add this package to the profile as a dependency only; do not add it to dsh.profile.bundles. Then copy the shipped preset directory into DSH home:

pnpm dsh plugin --profile web add dsh-codex
pnpm --dir /path/to/.dsh/profiles/web add link:/absolute/path/to/dsh-codex-reasoning-router
cp -R /absolute/path/to/dsh-codex-reasoning-router/preset/luna-sol-reasoning-router /path/to/.dsh/.agent-presets/

Install directly from GitHub instead of a local checkout:

pnpm --dir /path/to/.dsh/profiles/web add github:chenmzh/dsh-codex-reasoning-router
cp -R /path/to/.dsh/profiles/web/node_modules/dsh-codex-reasoning-router/preset/luna-sol-reasoning-router /path/to/.dsh/.agent-presets/

For a published package, replace the link dependency with dsh-codex-reasoning-router and copy the same directory from its installed package. Restart DSH, then explicitly select Luna + Sol Reasoning Router for a new session. The existing default preset is not changed.

The preset is a complete copy of the official Standard composition, preserving its normal tools, Skills, MCP, compaction, and subagent surface. It adds only the Router row. The plugin also checks the effective durable session preset via the public resolveSessionPreset API; accidental global loading does not attach it to other presets.

DSH presets do not own the host model route. Select openai-codex / gpt-5.6-luna before using this preset. If a saved route differs, the plugin stops that session with a diagnostic and never silently switches the main model.

Configuration

- id: reasoning-router
  name: dsh-codex-reasoning-router
  inject: [openAICodex]
  config:
    requiredPresetId: luna-sol-reasoning-router
    lunaProvider: openai-codex
    lunaModel: gpt-5.6-luna
    solProvider: openai-codex
    solModel: gpt-5.6-sol
    initialSolReasoning: medium
    escalatedSolReasoning: high
    solAdviceMaxTokens: 2000
    solTimeoutMs: 30000
    initialConsultEnabled: true
    failOpen: true

Both effort fields accept only medium or high and default to medium -> high. The narrow domain type and runtime guard reject xhigh and max. Startup checks the provider catalog and exact model metadata; a missing configured model is an error, not a fallback.

Confirm current IDs with the DSH model picker (/model) or the public LLM registry used by a diagnostic plugin:

await ctx.llm.listModels('openai-codex')
await ctx.llm.resolveModelInfo('openai-codex', 'gpt-5.6-luna')
await ctx.llm.resolveModelInfo('openai-codex', 'gpt-5.6-sol')

At implementation time, the installed pi-ai catalog contains gpt-5.6-luna and gpt-5.6-sol.

Request lifecycle

WAIT_FIRST_USER
  -> SOL_INITIAL_MEDIUM
  -> LUNA_EXECUTING
       -> new blocker: SOL_MEDIUM -> LUNA_EXECUTING
       -> same blocker, medium evaluated: SOL_HIGH -> LUNA_EXECUTING
       -> same blocker again: ESCALATION_EXHAUSTED (no model call)

agent/pre-step is an awaited public waterfall. On a root session's first direct user message, the listener awaits Sol medium, creates a plugin-sourced user context containing only the Advisory Packet, and then returns kind: enter. Only after that does the agent loop log the step, assemble the prompt/tools, and make Luna's first request. Failure appends a warning event and returns the original messages when failOpen is enabled.

The later sol_consult tool has a fixed schema and is registered once in the root agent scope. Its second call for the same fingerprint requires medium_advice_evaluation; this prevents transport failure or an untried suggestion from being treated as grounds for high escalation.

Why Sol cannot act

Every Sol call is a hand-built ctx.llm.stream request with:

• the configured Sol provider/model;
• only the static advisor prompt and one compact evidence message;
• no tools property;
• no session continuation identity;
• no filesystem, shell, MCP, skill, web, or subagent interface.

The response is consumed directly with BlockAssembler. It never enters the DSH agent loop or tool dispatcher. A returned tool-call block is a SolProtocolError; it is never executed. Only visible text becomes an Advisory Packet for Luna.

Recursion and durable state

Internal calls run under an AsyncLocalStorage marker whose purpose is sol-advisory. The llm/stream hook checks this marker and delegates immediately. It does not infer internal calls from the model name. DSH rc.6 exposes only compaction | session-title in GenerateOptions.purpose, so a custom purpose is not forged into that public field.

Consultation results are append-only session events:

• reasoning-router/initial-consult
• reasoning-router/consult-medium
• reasoning-router/consult-high
• reasoning-router/consult-failed
• reasoning-router/escalation-exhausted

The issue-state fold reads successful medium/high events from the durable log. A process-local map is not the source of truth. Failed network/provider calls log consult-failed but do not set mediumUsed or highUsed. Resume reconstructs the state; compaction may replace model-visible surface nodes but does not erase these log-only events. Only compact Advisory Packets—not private reasoning—are retained.

Fingerprints hash normalized goal, problem, and stable file/error/test anchors. Attempts, the question wording, timestamps, and random values are excluded.

Compatibility

The plugin uses the existing openai-codex adapter and credential lifecycle. It does not read OAuth files or tokens and does not call private ChatGPT endpoints. Sol one-shots intentionally omit sessionId; normal Luna turns remain owned by dsh-codex and retain their standard WebSocket context reuse and native/basic compaction behavior.

The Router is mounted only by the opt-in preset and adds one scoped system section and one scoped tool. It does not replace system prompt sections, contexts, the normal tool catalog, skills, MCP, subagent orchestration, compaction, permissions, or the agent loop. Subagents do not receive sol_consult from this plugin.

Observability and verification

The event names above are visible in the session log, and concise secret-free info/warning messages report consultations. To exercise escalation:

1. Start a new Luna root session and send one user message; confirm initial-consult precedes the first Luna assistant chunk.
2. Ask Luna to call sol_consult for a concrete blocker; confirm consult-medium.
3. Call again with the same goal/problem/evidence anchors and a non-empty medium_advice_evaluation; confirm consult-high.
4. Call a third time; confirm escalation-exhausted and no provider request.
5. Resume the session and repeat step 4; exhaustion must remain restored.

Development checks:

pnpm install --offline
pnpm run typecheck
pnpm run test
pnpm run build

Public DSH APIs used

• ctx.llm.listModels, resolveModelInfo, and stream
• awaited agent/pre-step and agent/request waterfalls
• agent/created and agent/disposed
• ctx.agents.roots() and agent-scoped agent.ctx
• resolveSessionPreset from @deepseek-ai/dsh-agent-presets
• agent.ctx.systemPrompt.section
• agent.ctx.tools.register with defineTool
• Session.append, Session.events, and the extensible KNOWN_SESSION_EVENT_TYPES
• createUserMessage, BlockAssembler, and ReasoningEffortId

No DSH private source, private runtime object, credential file, or undocumented backend endpoint is used.