/aryaamin
Log in
OpenClasp
docs/SHIELD.md 41 lines · 2.4 KB

OpenClasp Shield

Shield is an independent AI decision-support agent that works beside a connected agent. It is for consequential interactions where persuasion, unsupported claims, policy conflicts, unnecessary data requests, changed payment details, or ambiguous authority could cause a bad decision.

Agent flow

  1. Call openclasp_shield_open_case with the goal, counterparty type, proposed action, and any bounded facts, evidence, or policy.
  2. Call openclasp_shield_consult when the agent needs help. Provide only the minimum current-turn context needed for the decision. The default analysisDepth: "fast" performs one bounded model generation for inline decisions. Use analysisDepth: "deep" only for an explicit investigation.
  3. Use Shield's conversational reply and structured disposition to gather evidence, modify the plan, seek approval, proceed with safeguards, or stop.
  4. Call openclasp_shield_close_case with the action taken and observed result.

Owners can review cases, run deep consultations, and add authenticated guidance in the Shield dashboard. Agent access tokens cannot add owner guidance.

What is stored

OpenClasp stores the case structure, Shield's structured assessments, model/prompt metadata, input digests, authenticated owner guidance, and reported outcomes. The consultation message and transient situation context are sent to the configured model provider for generation and are not stored by OpenClasp. Do not submit secrets or full transcripts.

Set ANTHROPIC_API_KEY to enable model-backed investigation. OPENCLASP_SHIELD_MODEL can override the default Anthropic model. OPENCLASP_SHIELD_TIMEOUT_MS can set a 5–55 second generation deadline; the default is 50 seconds to remain below the hosted function limit. If generation is unavailable, Shield returns an explicit low-confidence fallback and does not pretend that an AI investigation occurred. Each structured consultation records whether the fast or deep strategy ran, its duration, and available token usage.

τ³ benchmark

The reproducible benchmark adapter is in benchmarks/tau3. It runs matched baseline, generic-second-review, and Shield conditions, withholds hidden evaluator data from Shield, records model/token metadata and outcomes, and includes a comparison script. Start with the documented airline cancellation smoke test, then run the three-condition sample before making performance claims.