@dharma-ai-labs/agent-fabric-provider-adapters
Capability-scoped local adapters for Codex, Claude Code, Agy, and Hermes Agent. Each adapter reports evidence discovery, task execution, continuation, Skill installation, activation, rollback, and usage support independently. Evidence support never implies that another capability is available. Agy 1.1.13 remains limited to partial read-only execution and evidence. Agy 1.1.15 adds content-bound Skill activation and transactional rollback through the Dharma CLI, but consequential task effects remain unavailable because Agy still lacks path- and command-scoped effect acknowledgements.
Hermes 0.20.4 uses its supported redacted JSONL session export, safe-mode one-shot task interface, project Skill trust command, and local Skill listing. The adapter permits read-only tasks only after Hermes reports a configured inference provider. Skill discovery and activation do not imply that model execution is configured.
The experimental retained Codex app-server session bridge is qualified on Linux
only. Native Windows probes accepted a permission-profile declaration without
enforcing it when the sandbox was not configured; the configured sandbox did not
pass the permitted-read control. Other hosts therefore fail before provider
dispatch with codex_session_sandbox_unqualified. Do not widen filesystem access
or disable the sandbox to work around this gate. This restriction does not change
ordinary enrollment, relay startup, or the separate provider task interfaces.
Named local Codex work retains the exact turn/start request and bounded,
exact-thread/turn notifications in the encrypted local vault. Version 2 captures
bind that request's text, workspace and restricted permission profile by hash;
version 1 notification-only captures remain readable. Work prompts containing
recognizable credentials are rejected before intent persistence or dispatch.
Its success or failure receipt references hashes and explicit coverage; raw
requests and notifications are never added to CLI output.
Early events are attributed only after the provider identifies the turn.
Missing scope, excluded credentials, interrupted turns and capture limits remain
visible limitations. Existing raw-evidence retention applies to these captures.
The approved evidence outbox includes the request under the existing disclosure
policy. This is request-and-notification evidence, not retained session context,
verified skill adoption or an attestation of the execution model.
acceptedLearningObservation remains false: local
completion and capture do not establish server intake, an automatic policy
counter, a replay evaluation, canary approval or a signed learning release.
Licensed under MIT.
Task Knowledge Transport
The task-local knowledge server uses MCP SDK server 2.3.0 over stdio, with a fresh server factory for each transport and an explicit 1,024-element raw tool-input budget. The existing document, page and response limits remain in force. The normal valid client continues after an oversized request is rejected; separate stdio clients retain their own receipt-pinned catalog scope.
This adapter does not expose an HTTP transport or bearer-authentication endpoint. These tests do not qualify an HTTP deployment, shared HTTP transport, audience validation or HTTP body limits. Introducing such a deployment requires separate transport and authorization qualification, not reuse of stdio evidence.