dsh-model-sync

dsh-model-sync

Mirror each configured llm-pi-ai provider's advertised model list into the Settings → Models section, so the model dropdown shows the provider's full model list without hand-entry. Discovery goes through the llm service's discoverModels('llm-pi-ai', …) — GET {baseURL}/models for OpenAI-compatible routes, the installed catalog for catalog routes.

Manual sync

Every activation of the plugin runs one sync pass: the first boot, and any manual re-enable. To refresh the model lists on demand, open the Plugins page (sidebar → Plugins), expand the dsh-model-sync card, and flick the model-sync row's switch off, then on again — the re-enable restarts the plugin, so the pass runs ~1.5 s later and re-reads every provider's advertised list. The settings section is written only when a sync changed at least one provider's models.

Each activation pass also logs an outcome summary per provider: [dsh-model-sync] activation summary: … lists the model ids the pass added (or pruned, with prune: true), or no models added or removed when a provider changed nothing. Interval passes print no summary.

The same outcome is persisted in the model-sync settings namespace (lastSync) and rendered right on the bundle's Plugins page: open the dsh-model-sync card and a Last sync panel sits between the description and the Components list, showing the run time and one line per provider — the added/deleted id lists, or no models added or removed. The panel follows the settings document, so it refreshes by itself ~1.5 s after the re-enable that triggered the pass.

Sync rules

• Adds — discovered ids that are not configured yet are appended with an owner: 'model-sync' marker (plus name / contextWindow / maxTokens / input metadata when the listing reports them). Hand-entered entries carry no owner; existing entries are never rewritten — ids, order, and any hand-entered capacities survive every sync.
• Prune (prune: true) — entries marked owner: 'model-sync' that are no longer advertised are removed. Hand-entered entries are never pruned.
• Rejected — when an owned id that is still advertised disappears between two syncs, it is treated as a deliberate deletion, remembered in a plugin state file (below), and not re-added. Manually re-adding a rejected id clears the rejection. Detection diffs consecutive syncs within one activation, so a manual re-enable never triggers it — the re-enable resets the in-memory baseline, and a single activation sync re-adds a deleted id. Deletion detection stays live only with a positive intervalMinutes.

Provenance for pruning rides on the per-entry owner marker (single source — no duplicated model list). The small per-provider rejected id lists are the sync's own bookkeeping, so they live in the plugin's state file — a JSON document at <dshHome>/storages/model-sync/state.json (configurable via stateFile) — never in the settings document, which is user configuration and holds only the lastSync outcome the Plugins page renders. Entries configured before this version have no owner and count as hand-entered; clearing the provider's models list once and letting the plugin rebuild it gives every entry full tracking.

Configuration

The bundle inserts a model-sync row:

- id: model-sync
  name: dsh-model-sync
  config:
    intervalMinutes: 0   # 0 = once per activation (boot or re-enable);
                         # N = also repeat every N minutes
    #     NB: a positive interval also enables deletion detection (rejected)
    prune: false         # true = remove auto-added ids no longer advertised
    # providers: [amd-server]   # optional: restrict which providers to sync
    # stateFile:                # optional: rejection-baseline state file
    #                           # (default: <dshHome>/storages/model-sync/state.json)

Every activation — first boot or a manual re-enable from the Plugins page — runs one sync ~1.5 s after the llm/settings services are ready; a positive intervalMinutes repeats it on that cadence.

Notes

• A provider whose endpoint is down is skipped with a warning; a discovery error never fails the row.
• Every pass logs its trigger — [dsh-model-sync] activation sync: … for boot/re-enable, … interval sync: … for the repeating timer — so a manual toggle is visible in the host log. Activation passes add an outcome summary line, [dsh-model-sync] activation summary: ….
• The rejection baseline moves out of the settings document in 0.7.0: a pre-0.7.0 providers key found in the model-sync namespace is migrated into the state file and stripped from the settings document on the next sync.
• npm test (node:test fake timers) needs Node ≥ 20.4, and ≥ 21.3 to silence the --disable-warning flag; the plugin itself only requires engines's >=18.