agent-passport-system-openclaw-plugin

OpenClaw plugin: Agent Passport System trust verification provider. Targets Agent Trust Verification Provider Pattern v0.1.

Known issue (verified 2026-09-08): the published 0.1.1 npm artifact and the 0.2.1 ClawHub artifact do not register successfully on OpenClaw 2026.9.3 and do not provide the gating described below. Registration fails on the first hook, before the enforcement hooks and the gateway methods are installed. 0.3.0 is the first release verified end-to-end against OpenClaw 2026.9.3.

To check an installation, run openclaw plugins inspect aps --runtime. On an affected version it reports a registration error. openclaw plugins list --json may still report the plugin as loaded, so use runtime inspection for this issue.

The plugin gates plugin installs against the APS public trust registry on the Gateway plugins.install path, gates high-risk tool calls behind explicit approval, and exposes APS primitives (grade lookup, delegation verification, message signing) via OpenClaw gateway RPC. It runs entirely in the OpenClaw plugin lifecycle and adds no requirement on OpenClaw core.

Delegation verification is implemented through agent-passport-system 6.0.1. The plugin contains no independent verifier. Delegation verification uses the authority-aware chain verifier with trust anchors the operator configures, so an integrity result is never returned as an authorization decision (SDK 6.0.0, advisory GHSA-r2fw-x6mg-f6h8).

Install

Version 0.3.0 is the corrected release verified end-to-end against OpenClaw 2026.9.3. Until 0.3.0 is published, these commands install an affected version and should not be relied on for enforcement.

clawhub install agent-passport-system-openclaw-plugin
# or
npm install agent-passport-system-openclaw-plugin

Configuration

Config is read from, in order:

1. $OPENCLAW_APS_CONFIG_PATH (env var). If set and the file is missing, loading fails rather than falling back.
2. ~/.openclaw/aps.config.json
3. Built-in defaults (permissive-with-warnings)

Configuration is not read from OpenClaw's plugin config. plugins.entries.aps.config.* is rejected by the manifest schema rather than accepted and ignored, because the plugin does not read api.pluginConfig.

A file that exists but cannot be read is a configuration error, not an absent file: an unreadable policy never silently becomes the permissive default. Malformed values, unknown enum values and out-of-range numbers are configuration errors for the same reason.

Some previously documented keys have been removed. endpoints.jwks still loads and warns, because the plugin no longer contacts that service. The removed security controls (policy.skillAuthor.minGrade, policy.toolCalls.enforceScope, everything under policy.inboundMessages, and highRiskBehavior: "warn") now fail to load: they never did anything, and accepting them silently would leave you believing a control was in force. Any other unrecognized key is also a configuration error.

Custom passportPath and auditLogPath must be absolute. Node does not expand ~, so a tilde path is rejected instead of resolving to a literal ./~/... directory. The built-in defaults shown below resolve through your home directory.

Schema (targets spec section 8):

{
  "provider": "aps",
  "endpoints": {
    "verifier": "https://gateway.aeoess.com/api/v1/public/trust"
  },
  "credentials": { "passportPath": "/absolute/path/to/aps-credentials.json" },
  "signing": {
    "enabled": false,
    "allowedCallers": [],
    "requireApproval": false,
    "auditLogPath": "/absolute/path/to/aps-signing-audit.log"
  },
  "policy": {
    "skillAuthor": { "warnBelow": 1, "blockBelow": null },
    "toolCalls": { "highRiskTools": ["bash", "exec", "fetch"], "highRiskBehavior": "approval" }
  }
}

Where the install gate applies

The APS before_install gate is verified on Gateway plugins.install. CLI-driven installs do not load the APS runtime hook and are not gated. Other install flows, including skill installation, have not been exercised and are not claimed.

What was tested is the Gateway plugin-install path: with a scoped package name, the plugin resolves the npm scope as the author, looks it up against the configured verifier, and a grade below blockBelow blocks the install, with the block reason returned in the Gateway response. Other install flows, including skill installation, have not been exercised and nothing is claimed about them.

Hook coverage (Pattern v0.1)

The Behavior column describes intended behaviour. On OpenClaw 2026.9.3, the published 0.1.0 and 0.1.1 npm artifacts and the 0.2.1 ClawHub artifact do not register any of these handlers.

Gateway RPC methods

The plugin defines the following namespaced RPC handlers. All three answer over the Gateway boundary on OpenClaw 2026.9.3, and each has an integration regression.

• aps.checkGrade with params: { agentId } → TrustProfile | null from the public APS gateway. That is the successful-response contract: a profile for a known author, null for an author the registry does not know. It is not the only outcome. If the verifier is unreachable or answers with a profile the plugin cannot trust, the call fails with a Gateway error (aps_verifier_unavailable or aps_verifier_malformed) rather than returning null, because reporting a transport failure as "author not known" is the conflation this release removed.
• aps.verifyDelegation with params: { chain } → result of APS SDK verifyAuthorityDelegationChain(). chain is the delegation chain as an array, root first, not a single token. Gateway methods take one options object from the host; 0.2.0 read positional arguments and both RPCs were unusable. Trust anchors come from policy.delegation.trustedIssuers; with none configured nothing verifies, which is the default. Revocation resolves to unknown because this plugin carries no revocation feed, so a cryptographically sound chain returns state: "indeterminate" rather than a valid: true it cannot establish.
• aps.signMessage({ message }) → { signature, domain, digest }. Ed25519 signature over a domain-separated input, using the local passport's private key. Off unless the operator turns it on. See Signing.

Other plugins can call these by their namespaced names.

Signing

The aps.signMessage handler is designed to sign with the private key in the passport file at credentials.passportPath. A gateway method registered by a plugin is not private to that plugin: OpenClaw dispatches it for authenticated gateway clients, and other plugins can reach it through the in-process runtime. The ClawHub review of 0.2.0 identified that the original handler would have made that signing capability reachable too broadly. On OpenClaw 2026.9.3 the signing RPC is registered. With signing disabled it refuses through the Gateway response channel and does not open the passport file. Caller allowlisting inside the handler has unit-level regressions.

What the reviewers told installers still holds. The points below describe intended behaviour:

• Signing is off by default. With signing.enabled false, aps.signMessage refuses every request and the passport file is never opened.
• Turning it on is a decision about every plugin on the host, not just this one. Turn it on only if you trust each installed plugin that could call aps.signMessage, and prefer a limited-purpose passport identity over your main one.
• signing.allowedCallers names who may sign. Entries are OpenClaw plugin ids, or the literal gateway-client for an authenticated gateway client that the host did not identify as a plugin. An empty list, the default, allows nobody.
• signing.requireApproval defaults to false, and the allowlist is the gate: signing is off unless signing.enabled is true, and an empty allowedCallers refuses everything. Setting requireApproval to true refuses every request, because OpenClaw 2026.9.3, the version this release is built and proved against, gives a gateway RPC handler no channel to ask a person for a decision: the mechanism the tool-call gate uses is a return value of the before_tool_call hook, and the SDK helper that reaches the host approval method is limited to plugin HTTP routes. It exists so an operator can hard-stop signing without editing the allowlist, and it will become a real approval once the host offers one. Citations are in the header of src/signing.ts.

Every signature is minted over APS-OPENCLAW-PLUGIN-SIGN-MESSAGE-V1\0 plus the message, following the domain-separation convention the SDK uses for authority delegations. A signature produced here therefore does not verify as a passport, attestation or delegation signature over the same bytes, and cannot be replayed into one of those contexts.

The plugin attempts to append one JSON line per signing request and refusal to signing.auditLogPath, by default aps-signing-audit.log in your .openclaw directory. Each line records the timestamp, the outcome, the caller, the domain prefix and the sha256 digest of the message. The message body is never written. A write failure does not abort the request: it is logged and signing continues, and that log is itself best-effort, so the audit log is not a guaranteed record.

One limit worth knowing: the caller name comes from what the host supplies. A plugin dispatching through the trusted in-process runtime is named exactly, because OpenClaw stamps the plugin id itself and never takes it from request parameters. A plugin that instead forwards a request from one of its own HTTP routes arrives carrying the original client and is indistinguishable from gateway-client. Allowlisting gateway-client is therefore broader than allowlisting a plugin id.

Conformance status

0.3.0 does not currently claim Pattern v0.1 conformance. The previous checklist was tied to an earlier implementation and configuration surface and has not been re-evaluated against the final 0.3.0 behaviour. The registration failure that previously blocked criterion 1 is fixed and proved against OpenClaw 2026.9.3.

v0.1 scope and known limitations

• High-risk-tool gate is the only before_tool_call enforcement. Full delegation-scope verification requires the agent to be running with an APS passport context, and is not implemented. Caching to keep the typical-case latency under 100ms is planned alongside it (spec section 9 #5).
• inbound_claim and before_dispatch deferred. The agent runtime context for inter-agent messaging is still being formalized; these hooks are planned once the surface is stable.
• Author identifier extraction is best-effort. OpenClaw's hook event types expose no author on event.skill or event.plugin, so binding to the host types removed the branches that read one. The only identifier available is the npm scope of a scoped packageName; everything else, including local archives, is treated as missing-author (warning, not block).
• No retry layer. If the gateway is slow (>500ms) the install proceeds with a warning. Caching of grade lookups is also planned, not implemented.

Examples

The examples below describe the configuration semantics.

Default (permissive-with-warnings)

No config file needed. The defaults block nothing, surface warnings for unknown or low-grade authors, and require user approval for bash/exec/fetch.

Strict mode

{
  "provider": "aps",
  "policy": {
    "skillAuthor": { "warnBelow": 2, "blockBelow": 1 },
    "toolCalls": { "highRiskTools": ["bash", "exec", "fetch", "shell"], "highRiskBehavior": "block" }
  }
}

Place at ~/.openclaw/aps.config.json or set $OPENCLAW_APS_CONFIG_PATH to its location.

Development

npm install
npm run build      # tsc -> dist/
npm test           # vitest
npm run lint       # tsc --noEmit --strict --noUnusedLocals --noUnusedParameters

To link locally into an OpenClaw checkout:

cd /path/to/openclaw-plugin-aps && npm link
cd /path/to/openclaw && npm link agent-passport-system-openclaw-plugin

License

Apache-2.0. Copyright 2026 Tymofii Pidlisnyi.