dsh-mnemon-strategy-default-three-tier

One ordinary Cordis Strategy plugin. It consumes public Source facts and proposes a three-tier View; it imports no Source implementation and grants no authority.

compose is pure. createTurn owns the executing turn's retrieval policy: one Documents query (4 results / 6,000 content characters), two different Recall queries sharing 6 results / 4,800 content characters, duplicate replay, relevance admission, and one Related traversal from admitted evidence. Named tools and generic View routes use the same state. Independent child turns never share quotas, even when they inherit the same immutable View.

Source schemas and provenance are the public read protocol. This plugin formats the familiar compact model response; Core retains the evidence/authority envelope for auditing without serializing it again into every model tool result.

The default plugin also owns the existing three-tier instructions and reminders. They are pinned in ViewSpec.guidance, separately from Source projections. The Host renders them for the selected turn; a different Strategy does not inherit this plugin's instructions. DSH-assisted write/archival workflows remain Host adapters over Source operations; Core does not schedule autonomous maintenance.

Additive plugins

The selected complete Strategy still produces one View. Its optional, exclusive slots are public through dsh-mnemon-strategy-default-three-tier/extension-sdk:

| Slot | Value | Optional plugin | | --- | --- | --- | | selection | Ordered Source keys and an optional writable subset | dsh-mnemon-strategy-scoped | | projection | A ceiling within the Host's projection-character budget | dsh-mnemon-strategy-light-context | | capture | In-turn recording guidance, explicit recording Action ids and optional Source keys | dsh-mnemon-strategy-auto-capture |

Enable these ordinary Cordis Entries alongside this Strategy; no strategyId change is needed. They can coexist because each owns a different slot. Core rejects duplicate slot owners rather than letting installation order win. Slot names and value schemas belong to this Strategy, not to Core. External authors use defineThreeTierExtension and installMemory(ctx, { strategyExtensions: [...] }).

All Sources share the existing turn's read quotas and one projection budget. Replay and Related admission include the exact Source instance identity. Writes still require an offered Action and Host authorization, including implicit capacity maintenance. Explicit operator management is a separate channel.

With no extensions the original three-tier composition is retained. Disabling one extension only removes its contribution for new turns; running turns keep their pinned View. If disabling explicit selection leaves multiple Sources in the same role, the original ambiguity check applies rather than silently choosing a Source. Projection limits are not token limits or delta injection; capture guidance is not a background worker or a guarantee of model behavior.

Its own directory is a standalone project: install its declared dependencies, then run pnpm verify. When developing against an unreleased Mnemon, install a packed dsh-mnemon artifact as the peer instead of linking its source tree.

The Starter or a user's Profile explicitly mounts its Entry. This package has no auto-activation patch; installing it does not override a configured Strategy.

Installation and verification

The default Starter already mounts and selects this Strategy. In an independent Profile, install the package, mount its Entry and explicitly select default-three-tier; another installed Strategy does not win by import order.

Installing an npm package is not the same as activating a contribution. From a source checkout, run pnpm verify with declared public peers installed; use a packed peer for unreleased SDK work.

Plugin development and independent fixtures · 中文指南