macOS alpha, 0.32.0-alpha.1

Kupo routing

Settings reference: behavior, defaults and safeguards for this page of desktop Settings.

Published unchanged from docs/settings/kupo-routing.md in Moguri's source repository at commit b686ddc (0.32.0-alpha.1).

Technical reference for Settings / Kupo routing. The app shows a short summary; this page keeps the full detail.

Add model scores and prices

Previously "Review model evaluations and reference prices". Contracts: docs/architecture/automatic-kupo.md, docs/architecture/routing.md, docs/architecture/model-evidence.md.

  • Balanced can use declared capabilities. Savings and Performance need a current exact-model source review, including frontier or economy classification and API-equivalent reference prices. Unknown is never treated as frontier, cheap, free or independent.
  • Scores are source-reviewed normalized observations, not Moguri quality guarantees. Comparable Cascade models need the same evaluation-set SHA-256. Critique also needs a reviewed independent lineage pair and a matching user-defined cost class. Evidence expires within 30 days of observation. No benchmark or metadata request runs here.
  • The review shows the source SHA-256, evaluation set and cost class.
  • Reference prices are reference only: not billed cost or a rate-policy change. For local models with basis local-no-api-fee, having no hosted API fee does not erase token usage, configured rates or unknown accounting.
  • Consent previously read: "I reviewed this exact source, classification, scores, price and limitations".
  • Adding a profile does not select a mode, call a model or grant permissions. The complete policy must be saved explicitly after review (button previously "Add to unsaved Kupo policy").

Independent model pair

Previously titled "Actual-user source claim". Contracts: docs/architecture/compound-workflows.md, docs/architecture/model-evidence.md.

  • This prepares a local policy draft only. No source URL is fetched, no provider metadata is inspected, no policy is saved and no model is called. Settings are saved separately after reviewing the complete policy.
  • The panel previously showed the full COMPOUND_EVIDENCE_WARNING: "Lineage eligibility is an actual-user-reviewed source claim, not proof about unobservable training or model quality. Provider declarations, aliases and matching output contracts are not that review. Passing project tests covers only those commands and exact bytes; critics do not grant permissions or accept source changes."
  • Old field labels: "Exact draft model", "Exact critic model", "Source-backed draft lineage identifier", "Source-backed critic lineage identifier", "Reviewed source kind" (reviewed training report, reviewed model card, reviewed independent audit), "HTTPS source URL", "Exact reviewed source text", "Explicit review expiry, UTC ISO timestamp", "Evidence limitations".
  • Consent previously read: "I reviewed this exact source and model pair. I am making a limited eligibility claim, not proving independent training or code quality."
  • After adding, the exact model/source claim is added only to the local policy draft and matching Critique routes; it implies no model call or source acceptance.

Workflow policy options

The options were renamed: "Explicit Single policy" is now "Single (approved solvers)" and "Reviewed worktree Cascade / Critique" is now "Cascade / Critique in a worktree". Values (auto, single, compound) are unchanged.

Automatic Kupo

Kupo uses your configured models automatically. No eligible model means no inference. One eligible model always runs directly in Single (without a classifier). With multiple models, Kupo chooses Single, gated Cascade or an eligible independent Critique with at most one revision.

Configured model pool

Selecting Kupo uses all eligible configured models by default, including models configured later. Turn off network routing to restrict it to execution-host loopback. An optional nonempty approvedModelIds list in the advanced policy narrows the pool. Disable Kupo to allow none. Configuration never starts inference or changes autonomy.

Model capabilities (previously "Declared capability inventory")

Categories describe supported inputs, reasoning controls and context, not measured coding quality or latency. Unknown tool support stays unknown. No paid benchmark or extra provider inspection runs during discovery. When no model is eligible, inspect the provider, privacy and environment diagnostics shown under Saved-policy availability.

Automatic context compaction

The switch was previously labeled "Allow bounded automatic text summaries at safe model boundaries".

Until you change it here, it follows Context compaction, which is on by default since protocol 110; the page says "Follows Context compaction until you change it here." Changing it saves an explicit Kupo policy that permits (or forbids) metered compaction calls and source-bound summaries, not permissions or human answers. Complete recent turns and original history remain retained and readable. Images are not compacted or silently dropped. The advanced policy sets the exact compactor, trigger fraction, recent turns, output bound and per-turn call limit.

Automatic mode warning

Only a selected original response reaches ordinary permissions. New modes require comparable reviewed Cascade evidence (models evaluated on the same evaluation set) or an independent same-cost-class Critique pair. Missing System One/Max transports remain blocked. Sandbox needs fresh exact native-byte qualification; there is no Local fallback.

Compound routing policy

Declare exact Cascade/Critique routes, approved concrete models, hosted-data consent and complete role bounds. The classifier chooses only an eligible declared route. Configuration never selects a model, spends budget, stages code or runs tests.

Each Cascade route needs id, pattern: cascade, draftId and escalationId. Each Critique route needs id, pattern: critique, draftId, criticId and lineageReviewId. Use canonical model IDs (listed under "Available model IDs", previously "Available exact model IDs"), not names or provider family strings. Up to four routes and eight approved role models.

Compound warning

Public Compound requires exact selected-Env source and stages and a separate actual-user staging/test review. Local and pinned SSH use their original roots; native-private (Sandbox) work remains physically unqualified until fresh exact-runtime evidence passes. Finite budgets need the selected-boundary reserved finishing (wind-down) policy. Max needs an explicit compatible Decision provider and cannot bypass reviews or hard policy. No fallback.

Single policy

One tool-less classifier chooses one explicitly approved solver. Both calls are metered and use the existing agent, permissions and execution env. Configuration never selects a model or starts inference.

Approved solvers

Approve up to eight models. Availability, env, context capacity and output bounds are rechecked before each call; this list is not a guarantee of eligibility. With no configured models, connect Copilot (or another provider) or configure a local endpoint first. An unavailable solver shows "unavailable for Kupo": inspect environment, privacy, context and output bounds.

Hosted or network routing data (Single)

The switch was previously labeled "Allow approved task history to hosted or network classifiers and solvers". Leave this off for explicitly configured loopback endpoints. A provider's local label does not make a network URL local. Loopback is on the execution host, not proof that a configured proxy never forwards data. Model redirects are blocked. Remote also needs its separately approved SSH host and scoped Copilot grant.

Classifier confidence threshold

This is model-supplied advice, not a calibrated probability or permission. Invalid or low-confidence plans pause without another model or fallback.

Single warning

Finite Single needs an explicitly configured qualified finishing (wind-down) policy on its selected execution boundary. Images, documents and exact window/effort controls require every actual classifier/solver capability; unsupported context is preserved, not dropped. This explicit policy remains Single; automatic Cascade/Critique and separately reviewed Compound have their own contracts. Native execution still needs fresh physical qualification. Hard policy is never lifted.

Saved-policy availability

After saving, explicitly select kupo in the composer. Changing this policy invalidates pending routed permissions; calls already in flight may still incur usage.

Classifier prompt (previously "Read the fixed classifier contract")

The editable system prompt still governs the solver. Classifier output is validated data, never solver instructions or a user answer.

Back to the user guide