feat: Task-aware budget tracking for MCP Tasks (SEP-1686)

[matt-dean-git]

Summary

MCP's Tasks primitive (SEP-1686) enables long-running async operations. The 2026 MCP roadmap confirms retry semantics and expiry policies are coming next.

SatGate currently enforces budgets per-request. When Tasks land with retries, we need to track costs at the task level, not just the request level — otherwise retries silently drain budgets.

Problem

1. Agent calls tool → task starts → costs N credits → task fails transiently
2. Agent retries same task → costs another N credits
3. Budget enforcement has no concept of "same logical operation"
4. After K retries, agent burns K×N credits for something that should cost N

Proposed Solution

Phase 1: Task-ID Correlation

• Parse task_id from MCP task responses
• Group spend records by task_id in the audit log
• Expose task-level cost in the dashboard (not just per-request)

Phase 2: Retry Cost Policy

• New policy field: retryCostMode: full | first-only | reduced
  • full — charge full cost per retry (default, backward-compatible)
  • first-only — only charge the first attempt, retries are free
  • reduced — charge a configurable fraction on retries
• Budget hold: reserve credits on first attempt, release on task completion/failure

Phase 3: Task Expiry Alignment

• When MCP finalizes expiry policies, align budget holds with task TTLs
• Auto-release held budget when task expires
• Configurable grace period for late completions

Context

• MCP 2026 roadmap: Enterprise Readiness is a top-4 priority
• Tasks retry semantics are in active development
• SatGate should be ready when the spec lands, not scrambling after

Labels

enhancement, mcp, budget-enforcement