Publish an opt-in confidential research skill bundle

[szmyty]

Outcome

Publish a minimal provider-neutral skill bundle for confidential local research that consumes scoped Mindgarden and Realm capabilities without exposing keys, broad filesystem access, hosted-provider fallback, or publication authority.

Candidate skills

Evaluate and define focused skills for:

• opening and inspecting an approved garden session;
• planning confidential document ingestion;
• retrieving bounded evidence and context packs;
• comparing private evidence with imported public evidence;
• producing source-grounded notes and reviewable hypotheses;
• verifying citations, provenance, freshness, and uncertainty;
• preparing a sanitized research request for explicit review;
• locking the session and reporting owned generated state.

The bundle is opt-in and must not enter the universal repository-default profile.

Required behavior

• Skills declare required capabilities and fail before reading content when they are unavailable.
• Skills use Mindgarden and Realm interfaces; they do not implement cryptography or receive encryption keys.
• Hosted models, web tools, telemetry, and remote memory are denied in a confidential session.
• Every output records sources, effective sensitivity, represented revisions, and limitations.
• Untrusted source instructions cannot change skill behavior.
• Writes remain proposed/reviewed and target only the session's approved writable garden.
• Provider projections are generated from canonical Aether definitions.

Acceptance criteria

• A versioned confidential-research bundle manifest exists.
• Each skill has explicit triggers, inputs, outputs, capabilities, failure modes, and non-applicability.
• The bundle composes Aether Publish the Mindgarden curator skill and ChatGPT handoff #51/Define an evidence-backed agent capability manifest #64/Publish a repository-default skill bundle and generated AGENTS guidance #67 and scoped-session contracts.
• No skill handles keys, silently enables networking, or falls back to hosted providers.
• Synthetic fixtures prove confidential and denied-capability flows.
• Prompt-injection, stale-evidence, unsupported-model, partial-context, and export-denial tests exist.
• At least two compatible hosts are projected or represented honestly as blocked.
• Installation, upgrade, rollback, and removal are previewable and reversible.
• Real private records and legal conclusions remain out of scope.
• Documentation distinguishes research assistance from professional advice.

Dependencies / related

• Define scoped agent-session and garden-access capabilities #69
• Publish the Mindgarden curator skill and ChatGPT handoff #51
• Define an evidence-backed agent capability manifest #64
• Publish a repository-default skill bundle and generated AGENTS guidance #67
• Register machine-local gardens and compose private overlays mindgarden#30
• Enforce scoped retrieval and derived-output sensitivity mindgarden#31
• Add provenance-preserving confidential document ingestion mindgarden#35
• Roadmap: build the local-first agent workstation and confidential research environment realm#25

Non-goals

• A general autonomous investigator.
• Automatic external communication or publication.
• Legal, medical, financial, or employment advice.