Turn one decision into a multi-model debate.

01 / ARCHITECTURE

Keep intent, control, and execution in separate layers

The Claude skill is a thin operating procedure: it resolves the question, source paths, and output paths. Knodr MCP is the control plane that validates and starts the graph. The .knodrflow owns all orchestration, including concurrency, write ordering, failure policy, and the three-iteration cap.

Four discovery agents inspect the same basis from different angles: Codex for code and implementation constraints, Claude for documents and decision history, Gemini for alternatives and prior art, and Grok for contradictions and risk. Claude compresses their raw briefs into the dossier before the debate begins.

02 / INPUTS & OUTPUTS

Six explicit inputs, three durable artifacts

start_run.inputs binds values by Manual Input node ID. The IDs below are part of the example’s public contract.

What comes out

03 / INSIDE THE LOOP

Rework is a graph edge, not a promise in a prompt

A control.loop node enforces n=3. Its payload is JSON—{ "i", "dossier", "workdir", "final" }—so the complete iteration context travels through the back edge. The body never depends on a Manual Input firing again.

1. 01
   Increment and unpack context

   A script increments i; json.text nodes extract the dossier, working directory, and iteration number for the body.

2. 02
   Author competing concepts

   Claude and Grok run in parallel. From iteration two onward, both must address the previous verdict’s directives explicitly.

3. 03
   Append before critique

   The concepts are written to the dossier. The resulting file.write.path output activates all four critics, enforcing disk write order.

4. 04
   Critique through independent lenses

   Two DeepSeek roles cover completeness and delivery cost; MiniMax tests complexity and trade-offs; Gemini applies security and domain risk.

5. 05
   Arbitrate with a safe parser

   Codex emits strict JSON. A parser accepts only pass or rework; missing or malformed output is forced to rework.

6. 06
   Break or feed back

   control.gate routes pass to loop.break and rework to loop.feedback. After a pass or the third iteration, Claude writes the final document.

Every agent uses a non-fatal error policy. One unavailable vendor leaves an empty section instead of destroying the run; the dossier and final summary make the gap visible.

04 / SET UP THE EXAMPLE

Put the flow under the MCP project root

1. Download the flow and save it as C:\work\knodr-examples\flows\decision-panel.knodrflow.
2. Download the Claude skill and save it as ~/.claude/skills/decision-panel/SKILL.md.
3. Register knodr mcp --project C:\work\knodr-examples in Claude’s MCP configuration, then restart Claude.
4. Authenticate the local claude and codex CLIs.
5. Add Knodr secrets named gemini, grok, deepseek, and minimax, or provide the matching KNODR_CREDENTIAL_* environment variables.

{
  "mcpServers": {
    "knodr": {
      "command": "C:\\Users\\you\\AppData\\Local\\Knodr\\current\\knodr.exe",
      "args": [
        "mcp",
        "--project",
        "C:\\work\\knodr-examples"
      ]
    }
  }
}

05 / THE CLAUDE REQUEST

Give the skill a decision, not a topic

Install a Claude skill named decision-panel. Its procedure should resolve material paths, choose a safe common working directory, allocate the three output files, call Knodr MCP, poll to a terminal state, and present the final file plus the dossier path.

---
name: decision-panel
description: Run a bounded multi-model architecture decision panel through Knodr.
---

1. Require the Knodr MCP tools and the decision-panel flow.
2. Turn the argument into one decidable question.
3. Resolve every material to an absolute path.
4. Use their deepest safe common ancestor as `workdir`.
5. Allocate dossier, raw-discovery, and final Markdown paths.
6. Call `validate_flow`, then `start_run` with all six inputs.
7. Poll `get_run_status` at intervals of at least 60 seconds.
8. Never call `get_run_snapshot` for this flow.
9. Read `finalPath`; report the decision, unresolved risks,
   final path, dossier path, and runId.
10. On failure, read the partial dossier and call `get_run_logs`.

This text is not sent directly to the MCP server. Claude interprets it under the skill, resolves the file contract, and turns it into typed MCP tool arguments.

06 / MCP PROCESSING TRACE

See the prompt become a running graph

For the request above, Claude first checks the downloaded graph. Knodr loads the JSON, resolves installed node definitions, and returns every structural diagnostic.

{
  "path": "flows/decision-panel.knodrflow"
}

{
  "ok": true,
  "nodes": 63,
  "connections": 116,
  "diagnostics": []
}

Claude then allocates the run files and binds the six Manual Input nodes by ID.

{
  "path": "flows/decision-panel.knodrflow",
  "inputs": {
    "question": "Should Atlas keep one modular monolith or split billing into a service before the EU launch, given a six-person team, a four-month deadline, GDPR, and a zero-downtime requirement?",
    "materials": "- doc: C:\\work\\atlas\\docs\\architecture.md — current boundaries and known coupling\n- doc: C:\\work\\atlas\\docs\\eu-launch.md — launch constraints and GDPR requirements\n- repo: C:\\work\\atlas\\src — implementation; inspect billing, identity, and data access",
    "workdir": "C:\\work\\atlas",
    "dossierPath": "C:\\work\\atlas\\.panel-tmp\\dossier-20260723-1430.md",
    "rawPath": "C:\\work\\atlas\\.panel-tmp\\discovery-20260723-1430.md",
    "finalPath": "C:\\work\\atlas\\.panel-tmp\\final-20260723-1430.md"
  },
  "timeoutSeconds": 12000
}

{
  "runId": "run_01KYA1EXAMPLE8YJ7Q2",
  "state": "running"
}

start_run validates again, binds the inputs to a fresh in-memory graph, applies the timeout, starts the eager scheduler, and returns immediately. The returned run ID means “accepted and running,” not “finished.”

{
  "runId": "run_01KYA1EXAMPLE8YJ7Q2"
}

{
  "runId": "run_01KYA1EXAMPLE8YJ7Q2",
  "state": "running",
  "flowName": "decision-panel",
  "startedAt": "2026-07-23T14:30:00Z",
  "elapsedMs": 512438,
  "error": null,
  "cancelReason": null,
  "nodes": [
    { "nodeId": "d-codex", "status": "completed" },
    { "nodeId": "archA", "status": "running" },
    { "nodeId": "archB", "status": "running" }
  ]
}

07 / READ THE RESULT

Return the decision and its evidence

When status becomes success, Claude reads finalPath with its filesystem tool and searches the dossier for ### Arbiter verdict. It reports whether the panel passed at iteration 1, 2, or 3—or reached the cap without convergence—then summarizes the recommendation and unresolved risk.