Mounts and the Consumption Skill

RememberStack is designed for coding agents that already know how to navigate files well. A deployment can therefore give an agent four read-only filesystem views, plus a generated SKILL.md that explains how to use those views and the query API together.

The four read-only views

The raw view is deliberately outside the normal navigation tree. This keeps Markdown-first browsing cheap while preserving access to a recording, image, or other original when conversion has lost something important.

The self-host mount publisher returns the exact paths as a typed PublishedMounts value. A mount set is bound to one deployment and is always marked read-only; the consumption-skill renderer refuses paths from a different deployment.

Filesystem first, query when files cannot do the job

When mounts are available, prefer them for navigation, reading, and grep. Use the API, CLI, or MCP tools for operations without a filesystem equivalent:

• semantic search;
• graph neighborhoods and paths;
• valid_at or believed_at time-travel;
• fact-to-evidence hydration;
• decision transcripts;
• change feeds and deltas.

An environment that cannot mount the deployment still has the API/CLI/MCP route. The generated skill states that it is unmounted and directs the agent to those surfaces instead of advertising paths that do not exist.

The generated consumption skill

ConsumptionSkillSurface reads the live deployment name, language, special-purpose scopes, current Plane K page count, and latest active recipe versions from PostgreSQL. It combines them with the published mount locations and renders a versioned, standard SKILL.md.

The skill teaches one default motion:

1. Orient on Plane K. Read the knowledge checkout or use an orientation recipe such as pages_about.
2. Verify on the spine. Use fact-grain relations or observations for anything that answers “is this true now?”
3. Audit on evidence. Hydrate the fact to claims, source spans, documents, and raw media when the stakes demand it.

Plane K is not implemented yet, so current deployments honestly render its state as known-empty. A cold agent is instructed to attempt orientation, then fall back to P3 when mounted or search when unmounted—never to invent the missing synthesis.

The renderer can return the artifact without writing it, or publish it atomically so a harness never reads a partially-written skill:

rendered = skill_surface.render(mounts=published_mounts)
skill_path = skill_surface.publish(
    directory=harness_skill_directory,
    rendered=rendered,
)

What the skill prevents

The fixed curriculum is the correctness boundary for a cold agent:

• claims are source testimony; relations and observations are current, adjudicated facts;
• claims_as_of means assertion history, never current truth; the generated skill also says explicitly when that recipe is not enabled;
• superseded testimony is excluded by default;
• support: withdrawn means a fact still stands but needs a caveat and audit;
• contradiction co-members must be reported together instead of silently choosing one;
• compiled pages carry their own freshness and must be verified when load-bearing;
• a media locator, world-validity time, and system-belief time are three different coordinates.

The S58 cold-agent check

The retrieval eval suite now includes S58. It gives a model only the rendered skill and a task—no design documents or repository context—and asks for a structured plan. The check passes only if the agent orients on K, falls back honestly when K is empty, routes current truth to facts, keeps fact and evidence grains separate, treats withdrawn support cautiously, reports contradiction co-members, prefers mounts for readable work, and hydrates to sources for an audit.

The canary is seeded per deployment and each run is recorded in eval_runs against the skill revision. A wrong plan, including using claim search as the answer to a current-truth question, fails the suite.