Govern Prime Agent tool calls with bounded policy, principal-authority checks, redacted evidence, and local receipts. Use before Prime Agent writes, uses networks, deploys, publishes, changes trust, or performs payment-related actions.
Resources
12Install
npx skillscat add rhein1/agoragentic-integrations/agoragentic-prime-agent Install via the SkillsCat registry.
SKILL.md
Agoragentic for Prime Agent
- Start with read-only inspection and no-spend proof.
- Classify every proposed tool call as read, write, network, spend, deploy, publish, trust, or unknown.
- For spend, deploy, publish, or trust actions, require a short-lived grant bound to the exact principal, agent, session, tool call, capability, and input hash.
- Ask interactively for ordinary write/network actions when policy requires review.
- Fail closed when review is required but no UI is available.
- Require a host-trusted verifier to validate authority integrity; policy allowlists, UI confirmation, and grant fields alone are not authority.
- Consume each accepted authority ID and action hash once; retries require a new principal-approved action.
- Never let the agent approve its own authority request, expand its own budget, fund its own wallet, or convert a local receipt into settlement proof.
- Record hashes and bounded redacted evidence, not raw prompts, credentials, wallet material, or unrestricted tool output.
- Reconcile ambiguous paid outcomes before retrying.
A Prime Agent extension is an application policy layer. Prime Agent's worker and kernel processes are not security sandboxes. Payment-bearing and production work still requires a restricted runtime plus enforced network, filesystem, process, and payment chokepoints.
Qualification boundary
- Bind compatibility claims to Prime Agent
v0.7.2, commit83a0f9f9566219551fcb6ffaf7f519a815749a58, and release SHA-256bc5471f2a626d727b88a45eb745fff93b10c554a3c4fc5912f25d8c64b987f5e. - Verify the raw artifact before extraction; materialize the committed dependency lock in isolation with lifecycle scripts disabled; select the exact Node/platform closure tuple from the schema-closed integrity profile; bind the extracted first-party tree, installed dependency tree, profile, and exact source-extension manifest before spawn.
- Run with
PI_OFFLINE=1,PRIME_AGENT_TELEMETRY=0,--offline,--no-session,--no-builtin-tools,--no-extensions,--no-skills,--no-prompt-templates,--no-themes, and--no-context-files, loading only the exact local extension with-e. - Treat the provider-free released-host matrix only as
runtime_compatibilityevidence. The captured high-severityextract-zipadvisory blocks promotion, so keep the candidate level empty and the effective level atsource_adapter; do not make a public compatibility claim. The matrix does not prove real policy interception, restricted exact runtime, hosting, production activation, or adoption. - Newer-release observation never grants automatic update or promotion authority.
- Preserve the acyclic manifest → profile → receipt → qualification evidence → Marketplace record chain; verify all runtime-request refs/digests and never feed a downstream hash back into the source manifest.
- Keep credentials, provider calls, spend, wallet, settlement, deployment, publication, outreach, public compatibility claims, trust, and ranking mutations false, and keep the source-only package centrally held.