Contract review workflow

Contract review workflow

Orchestrates a deployed Mistral Workflow for contract review. Detects wait_for_input() human-in-the-loop checkpoints by querying a status handler, presents findings to the user, and resumes the workflow with their decision via workflow_interact(action: "signal").

Profile note: workflow_execute, workflow_status, and workflow_interact are available in the core profile (default). files_upload (for local PDF upload) requires MISTRAL_MCP_PROFILE=full.

Why workflows vs direct API: the workflow runs durably on Mistral's infrastructure — it persists across restarts, handles retries, and can run multi-step analysis (OCR → clause extraction → legal DB lookup → risk scoring) without keeping a connection open.

Important: handler names are workflow-specific

The name field in workflow_interact calls (signal handler, query handler) is defined by the deployed workflow. Ask the user for:

• The query handler name that exposes checkpoint state (e.g. "get_checkpoint")
• The signal handler name that receives approval decisions (e.g. "human_approval")

Defaults used in the examples below — replace with the actual names from the workflow.

Steps

Step 1 — Collect inputs

Ask the user for:

1. workflowIdentifier — the deployed workflow name or ID (visible in mistral://workflows)
2. The contract document: a file_id (from files_upload, requires full profile) or a public URL
3. Optional: review_type (e.g. "vendor", "employment", "partnership") and language ("fr" or "en")
4. The query handler name and signal handler name (if not the defaults)

Step 2 — Start the workflow

Call workflow_execute:

{
  "workflowIdentifier": "<workflow name or ID>",
  "input": {
    "document_id": "<file_id, or omit if using document_url>",
    "document_url": "<URL, or omit if using file_id>",
    "review_type": "<vendor|employment|partnership|other>",
    "language": "fr"
  }
}

Note the structuredContent.execution_id. Confirm to the user: "Workflow started — execution ID: <execution_id>."

Step 3 — Poll and probe for checkpoints

Loop:

1. Call workflow_status with { "executionId": "<execution_id>" }
2. Check structuredContent.status:
   • COMPLETED → go to Step 5
   • FAILED / CANCELED / TIMED_OUT → surface structuredContent.result and stop
   • RUNNING → probe for checkpoint (see below), then wait ~10 seconds

While RUNNING, query the checkpoint handler to detect if the workflow is waiting for human input:

{
  "executionId": "<execution_id>",
  "action": "query",
  "name": "get_checkpoint"
}

If structuredContent.result is non-null and contains checkpoint data (e.g. { "waiting": true, "data": {...} }), proceed to Step 4.

If no checkpoint or result is null, show a progress update and continue polling.

Step 4 — Handle approval checkpoints

When a checkpoint is detected:

1. Display the checkpoint findings from the query result:

   ⏸  CHECKPOINT — Human review required
   ──────────────────────────────────────
   [findings from query result.data — risk summary, flagged clauses, etc.]
   

2. Ask the user: "Approve and continue / Request changes?"

3. If approved, signal the workflow:

   {
     "executionId": "<execution_id>",
     "action": "signal",
     "name": "human_approval",
     "input": { "approved": true }
   }
   

4. If changes requested, capture the user's comment and signal:

   {
     "executionId": "<execution_id>",
     "action": "signal",
     "name": "human_approval",
     "input": { "approved": false, "comment": "<user comment>" }
   }
   

Return to Step 3 after signaling. A workflow may have multiple checkpoints.

Step 5 — Deliver final report

When status === "COMPLETED", format structuredContent.result as a structured contract review:

✅  CONTRACT REVIEW COMPLETE
─────────────────────────────
Execution: <execution_id>
Overall risk: [level from result]

KEY FINDINGS
[structured list from result]

RECOMMENDED ACTIONS
[list from result]

Offer to pass the output to /mistral-mcp:contract-analyzer for a complementary stateless view, or to french_legal_summary for a plain-language summary.