@adhdev/mesh-shared 1.0.29 → 1.0.30-rc.2

package/src/mesh-tool-names.ts

@@ -70,6 +70,7 @@ export const CANONICAL_MESH_TOOL_NAMES = [
     'mesh_magi_kind_panel_list',
     'mesh_node_slots_set',
     'mesh_node_slots_list',
+    'mesh_node_slots_propose',
 ] as const;
 
 export type CanonicalMeshToolName = typeof CANONICAL_MESH_TOOL_NAMES[number];

package/src/slot-proposal.ts

@@ -0,0 +1,327 @@
+/**
+ * CLI auto-detect → capability-slot / MAGI-panel PROPOSAL generator.
+ *
+ * Detection of installed CLI providers already exists per node (the status
+ * snapshot's `availableProviders`), and applying a slot profile already exists
+ * (`mesh_node_slots_set`, dry-run by default). What was missing is the bit in
+ * between: turning "these CLIs are installed on this node" into a concrete
+ * `NodeCapabilitySlot[]` draft the operator can review. This module is that
+ * bridge, and nothing more — it is a PURE proposal generator. It never writes;
+ * the caller feeds its output into the existing dry-run/approve tools.
+ *
+ * ─── Why the mapping is a static table ───────────────────────────────────────
+ *
+ * Provider manifests carry NO difficulty or capability-grade information. The
+ * fields that look like they might (`modelOptions`, `thinkingLevelOptions`)
+ * describe what a provider ACCEPTS, not what it is GOOD AT — a provider listing
+ * `opus` says nothing about whether opus should get the hard tasks. So there is
+ * no honest way to derive difficulty from the manifest today.
+ *
+ * The table below is therefore seeded from the operator's real, in-use slot
+ * configuration rather than from a guess. That makes it a starting point with
+ * actual provenance, and it is deliberately isolated in ONE constant so the
+ * planned usage-data-driven replacement has a single, obvious swap point.
+ *
+ * Kept in the dependency-free mesh-shared leaf because both daemon-core (which
+ * has the detection data) and mcp-server (which owns the propose/apply tools)
+ * need it, and it is pure data + pure functions on plain objects.
+ */
+import {
+    normalizeNodeCapabilitySlot,
+    type MeshTaskDifficulty,
+    type NodeCapabilitySlot,
+} from './brain-routing'
+import type { MagiSlot } from './magi'
+
+/**
+ * One provider's seeded slot recipe. A provider may map to MORE THAN ONE slot
+ * (claude-cli does: a wide sonnet slot plus a narrow opus slot for the hard
+ * work), which is why the table's values are arrays.
+ */
+export interface CliSlotRecipe {
+    /** Optional model to pin on the slot. Best-effort at launch. */
+    model?: string
+    /** Optional thinking level, in the provider's own vocabulary. */
+    thinkingLevel?: string
+    /** Difficulty range this slot handles. Empty = general-purpose. */
+    difficulty?: MeshTaskDifficulty[]
+    /** Per-slot concurrency cap. */
+    maxParallel?: number
+    /**
+     * Set when this recipe is an unvalidated guess rather than a transcription
+     * of a slot the operator actually runs. Surfaced on the proposal so the
+     * reviewer knows which lines to scrutinize.
+     */
+    provisional?: boolean
+    /** Short human-readable justification, echoed into the proposal. */
+    rationale?: string
+}
+
+/**
+ * ★ THE MAPPING TABLE — the single swap point.
+ *
+ * Seeded (2026-08-03) from the operator's live slot configuration, on the
+ * reasoning that a transcription of what is actually in production beats an
+ * invented heuristic. Replace wholesale once usage data (success rate, cost,
+ * turn latency per provider×difficulty) can drive it.
+ *
+ * Every entry except `hermes-cli` reflects a real configured slot. `hermes-cli`
+ * is a conservative GUESS — see its `provisional` flag.
+ */
+export const CLI_SLOT_RECIPES: Readonly<Record<string, readonly CliSlotRecipe[]>> = Object.freeze<Record<string, readonly CliSlotRecipe[]>>({
+    'claude-cli': [
+        {
+            model: 'sonnet',
+            thinkingLevel: 'high',
+            difficulty: ['medium', 'easy'],
+            maxParallel: 5,
+            rationale: 'Primary workhorse — widest parallelism for routine work.',
+        },
+        {
+            model: 'opus',
+            thinkingLevel: 'high',
+            difficulty: ['difficult'],
+            maxParallel: 1,
+            rationale: 'Reserved for hard tasks; capped at 1 to bound cost.',
+        },
+    ],
+    'kimi': [
+        {
+            model: 'kimi-code/k3',
+            difficulty: ['medium', 'difficult'],
+            maxParallel: 2,
+            rationale: 'Independent second opinion on mid/hard work.',
+        },
+    ],
+    'codex-cli': [
+        {
+            difficulty: ['medium', 'difficult', 'freeform'],
+            maxParallel: 2,
+            rationale: 'Broad range including freeform; no model pin.',
+        },
+    ],
+    'antigravity-cli': [
+        {
+            model: 'Gemini 3.1 Pro (High)',
+            difficulty: ['easy'],
+            maxParallel: 2,
+            rationale: 'Cheap capacity for easy tasks.',
+        },
+    ],
+    'cursor-cli': [
+        {
+            model: 'auto',
+            difficulty: ['easy'],
+            maxParallel: 1,
+            rationale: 'Easy tasks only; auto model selection.',
+        },
+    ],
+    'hermes-cli': [
+        {
+            difficulty: ['medium'],
+            maxParallel: 2,
+            provisional: true,
+            // NOTE: ESTIMATE, NOT OBSERVED. hermes-cli is absent from the live
+            // slot configuration this table was seeded from, so `medium` is a
+            // conservative placement rather than a transcription. Revisit once
+            // it has real usage data.
+            rationale: 'ESTIMATE — no live slot to transcribe; conservative mid placement. Adjust after real use.',
+        },
+    ],
+})
+
+/**
+ * Fallback for a detected CLI provider absent from {@link CLI_SLOT_RECIPES}.
+ * Deliberately timid: one general-purpose slot at the lowest parallelism, so an
+ * unrecognized provider can be used but never silently soaks up the queue.
+ */
+export const UNKNOWN_CLI_SLOT_RECIPE: Readonly<CliSlotRecipe> = Object.freeze<CliSlotRecipe>({
+    difficulty: ['medium'],
+    maxParallel: 1,
+    provisional: true,
+    rationale: 'Unrecognized provider — conservative default (medium, maxParallel 1). Review before relying on it.',
+})
+
+/** A detected, installed CLI provider on one node — the generator's input. */
+export interface DetectedCliProvider {
+    /** Provider type id, e.g. 'claude-cli'. */
+    type: string
+    /** Human-readable name, for display in the proposal. */
+    displayName?: string
+    /** Detected version, when known. Display only — never affects the mapping. */
+    version?: string
+}
+
+/** One proposed slot plus why it was proposed. */
+export interface ProposedSlotEntry {
+    slot: NodeCapabilitySlot
+    /** True when the provider had no table entry and took the conservative fallback. */
+    unknownProvider: boolean
+    /** True when the recipe behind this slot is flagged as an unvalidated estimate. */
+    provisional: boolean
+    rationale?: string
+}
+
+/** The full slot proposal for one node, including what a write would DESTROY. */
+export interface SlotProposal {
+    /** The draft slot list — a WHOLESALE replacement for the node's policy.slots. */
+    proposedSlots: NodeCapabilitySlot[]
+    /** Per-slot provenance, index-aligned with `proposedSlots`. */
+    entries: ProposedSlotEntry[]
+    /** Provider types detected but not present in the mapping table. */
+    unknownProviders: string[]
+    /** Provider types whose proposal rests on an unvalidated estimate. */
+    provisionalProviders: string[]
+    /**
+     * Slots currently configured on the node that the proposal does NOT
+     * reproduce — i.e. what applying this proposal would DELETE. Slot writes are
+     * wholesale replacements, so an operator-hand-tuned slot absent from the
+     * detection-derived draft is silently destroyed unless it is named here.
+     */
+    droppedSlots: NodeCapabilitySlot[]
+    /** Provider types that appear in `droppedSlots` but in no proposed slot at all. */
+    droppedProviders: string[]
+    /** True when applying the proposal would remove at least one existing slot. */
+    destructive: boolean
+}
+
+/** Stable key identifying a slot's identity for current-vs-proposed diffing. */
+function slotKey(slot: NodeCapabilitySlot): string {
+    return [
+        slot.provider,
+        slot.model ?? '',
+        slot.thinkingLevel ?? '',
+        [...(slot.difficulty ?? [])].sort().join('|'),
+        [...(slot.capability ?? [])].sort().join('|'),
+        slot.maxParallel ?? '',
+    ].join('')
+}
+
+/** Dedupe detected providers by type, preserving first-seen order. */
+function dedupeDetected(detected: readonly DetectedCliProvider[]): DetectedCliProvider[] {
+    const seen = new Set<string>()
+    const out: DetectedCliProvider[] = []
+    for (const d of detected) {
+        const type = typeof d?.type === 'string' ? d.type.trim() : ''
+        if (!type || seen.has(type)) continue
+        seen.add(type)
+        out.push({ ...d, type })
+    }
+    return out
+}
+
+/**
+ * Build a capability-slot proposal from a node's detected CLI providers.
+ *
+ * Pure and total: zero detections yields an empty proposal (never throws), which
+ * the caller should treat as "nothing to propose" rather than "replace the
+ * node's slots with nothing".
+ *
+ * `currentSlots` is optional but strongly recommended — it is the only way the
+ * returned proposal can report what a wholesale write would destroy.
+ */
+export function buildSlotProposal(
+    detected: readonly DetectedCliProvider[],
+    currentSlots: readonly NodeCapabilitySlot[] = [],
+): SlotProposal {
+    const providers = dedupeDetected(detected ?? [])
+    const proposedSlots: NodeCapabilitySlot[] = []
+    const entries: ProposedSlotEntry[] = []
+    const unknownProviders: string[] = []
+    const provisionalProviders: string[] = []
+
+    for (const provider of providers) {
+        const known = CLI_SLOT_RECIPES[provider.type]
+        const recipes: readonly CliSlotRecipe[] = known ?? [UNKNOWN_CLI_SLOT_RECIPE]
+        const isUnknown = !known
+        if (isUnknown) unknownProviders.push(provider.type)
+
+        let providerProvisional = false
+        for (const recipe of recipes) {
+            // Normalize through the SAME normalizer the daemon applies on write, so
+            // a proposal can never preview a shape the write would reshape.
+            const slot = normalizeNodeCapabilitySlot({
+                provider: provider.type,
+                model: recipe.model,
+                thinkingLevel: recipe.thinkingLevel,
+                difficulty: recipe.difficulty,
+                maxParallel: recipe.maxParallel,
+            })
+            if (!slot) continue
+            const provisional = recipe.provisional === true
+            if (provisional) providerProvisional = true
+            proposedSlots.push(slot)
+            entries.push({
+                slot,
+                unknownProvider: isUnknown,
+                provisional,
+                ...(recipe.rationale ? { rationale: recipe.rationale } : {}),
+            })
+        }
+        if (providerProvisional) provisionalProviders.push(provider.type)
+    }
+
+    // What a wholesale write would destroy: every currently-configured slot with
+    // no exact counterpart in the draft.
+    const proposedKeys = new Set(proposedSlots.map(slotKey))
+    const droppedSlots = currentSlots.filter(slot => !proposedKeys.has(slotKey(slot)))
+    const proposedProviders = new Set(proposedSlots.map(s => s.provider))
+    const droppedProviders = [...new Set(
+        droppedSlots.map(s => s.provider).filter(p => !proposedProviders.has(p)),
+    )]
+
+    return {
+        proposedSlots,
+        entries,
+        unknownProviders,
+        provisionalProviders,
+        droppedSlots,
+        droppedProviders,
+        destructive: droppedSlots.length > 0,
+    }
+}
+
+/**
+ * Build a MAGI panel proposal from the same detections.
+ *
+ * ─── Deliberately narrow ──────────────────────────────────────────────────────
+ *
+ * MAGI's value is provider INDEPENDENCE: replicas from different providers
+ * (ideally different machines) answering the same question, so agreement means
+ * something. Detection tells us which providers exist — that is exactly enough
+ * to propose one panel of distinct providers, and no more.
+ *
+ * What detection does NOT tell us is which provider suits which review KIND
+ * (rca vs design vs claim_audit). Nothing in any manifest grades a provider for
+ * root-cause analysis over design review, and inventing a per-kind assignment
+ * would fabricate a rationale that does not exist. So this proposes ONE panel of
+ * the detected providers and leaves the kind binding to the operator; the caller
+ * decides which `task_kind` to bind it to via the existing dry-run tool.
+ *
+ * Ordering follows {@link CLI_SLOT_RECIPES} insertion order (recipe-known
+ * providers first, in table order), so the panel leads with the providers whose
+ * suitability is actually attested.
+ */
+export function buildMagiPanelProposal(
+    detected: readonly DetectedCliProvider[],
+    opts: { nodeId?: string; maxSlots?: number } = {},
+): MagiSlot[] {
+    const providers = dedupeDetected(detected ?? [])
+    const tableOrder = Object.keys(CLI_SLOT_RECIPES)
+    const rank = (type: string): number => {
+        const i = tableOrder.indexOf(type)
+        return i === -1 ? Number.MAX_SAFE_INTEGER : i
+    }
+    const ordered = [...providers].sort((a, b) => rank(a.type) - rank(b.type))
+    const limit = Number.isFinite(opts.maxSlots) && (opts.maxSlots as number) > 0
+        ? Math.floor(opts.maxSlots as number)
+        : ordered.length
+
+    return ordered.slice(0, limit).map((p): MagiSlot => ({
+        ...(opts.nodeId ? { nodeId: opts.nodeId } : {}),
+        provider: p.type,
+        // A model is intentionally NOT pinned: the panel's job is cross-provider
+        // independence, and pinning models here would silently couple the panel
+        // to this table's cost assumptions rather than to review quality.
+    }))
+}