@gethmy/mcp 2.24.0 → 2.26.0

package/src/playbook-metric-warnings.ts

@@ -0,0 +1,56 @@
+import {
+  declaredGateMetricsFromAgents,
+  referencedGateMetrics,
+  type WorkspaceAgent,
+} from "@harmony/shared";
+
+/**
+ * Non-blocking warnings for a playbook write whose `custom` gates name metrics
+ * no reporting agent in the workspace declares (card #922). The write always
+ * succeeds — a metric name can legitimately precede its declaration — but the
+ * author learns about the mismatch NOW, not after the first bound card burns a
+ * stage run on it. Inert until at least one agent has reported its declared
+ * names (an older daemon, or none running, must never raise a false alarm).
+ */
+export function playbookMetricWarnings(
+  agents: ReadonlyArray<Pick<WorkspaceAgent, "declared_gate_metrics">>,
+  steps: unknown,
+): string[] {
+  if (!Array.isArray(steps)) return [];
+  const declared = declaredGateMetricsFromAgents(agents);
+  if (!declared.known) return [];
+
+  const warnings: string[] = [];
+  const seen = new Set<string>();
+  for (const ref of referencedGateMetrics({ steps, steps_version: 2 })) {
+    if (declared.names.has(ref.metric) || seen.has(ref.metric)) continue;
+    seen.add(ref.metric);
+    warnings.push(
+      `Gate metric "${ref.metric}" (stage "${ref.stageName}") is not declared by any agent in this workspace — a stage run gating on it will hold until a daemon declares it under agent.playbooks.metrics.${ref.metric}.`,
+    );
+  }
+  return warnings;
+}
+
+/**
+ * Fetch the workspace's agents and compute the warnings, swallowing every
+ * error: the playbook write already succeeded, and a warning lookup must never
+ * turn that success into a tool failure.
+ */
+export async function collectPlaybookMetricWarnings(
+  client: {
+    listWorkspaceAgents(
+      workspaceId: string,
+    ): Promise<{ agents: WorkspaceAgent[] }>;
+  },
+  workspaceId: string | undefined,
+  steps: unknown,
+): Promise<string[]> {
+  if (!workspaceId || !Array.isArray(steps)) return [];
+  try {
+    const { agents } = await client.listWorkspaceAgents(workspaceId);
+    return playbookMetricWarnings(agents, steps);
+  } catch {
+    return [];
+  }
+}

package/src/prompt-builder.ts

@@ -402,8 +402,6 @@ export interface GeneratePromptOptions {
   contextOptions?: Partial<PromptContextOptions>;
   customConstraints?: string;
   memories?: MemoryData[];
-  /** Pre-assembled context string from context assembly engine */
-  assembledContext?: string;
   /** Assembly ID for manifest tracking */
   assemblyId?: string;
 }
@@ -414,15 +412,8 @@ export interface GeneratePromptOptions {
 export function generatePrompt(
   options: GeneratePromptOptions,
 ): GeneratedPrompt {
-  const {
-    card,
-    column,
-    variant,
-    customConstraints,
-    memories,
-    assembledContext,
-    assemblyId,
-  } = options;
+  const { card, column, variant, customConstraints, memories, assemblyId } =
+    options;
 
   // Merge context options with defaults
   const contextOpts: PromptContextOptions = {
@@ -537,11 +528,7 @@ export function generatePrompt(
   });
 
   // Relevant memories from knowledge graph
-  if (assembledContext) {
-    // Use pre-assembled context from context assembly engine
-    sections.push(`\n${assembledContext}`);
-  } else if (memories && memories.length > 0) {
-    // Fallback: legacy memory format
+  if (memories && memories.length > 0) {
     sections.push(`\n## Relevant Memories`);
     sections.push(
       `*${memories.length} memories recalled from knowledge graph:*`,
@@ -556,12 +543,7 @@ export function generatePrompt(
   }
 
   // "One Thing" synthesis — highest-leverage next action
-  const oneThingLine = synthesizeOneThing(
-    card,
-    subtasks,
-    links,
-    assembledContext,
-  );
+  const oneThingLine = synthesizeOneThing(card, subtasks, links);
   if (oneThingLine) {
     sections.push(`\n## Recommended Next Step\n${oneThingLine}`);
   }
@@ -590,9 +572,7 @@ Keep \`currentTask\` specific (e.g., "Refactoring auth middleware" not "Working
 
   const prompt = sections.join("\n");
 
-  const memoryCount = assembledContext
-    ? (assembledContext.match(/^### /gm) || []).length
-    : memories?.length || 0;
+  const memoryCount = memories?.length ?? 0;
 
   return {
     prompt,
@@ -616,69 +596,7 @@ Keep \`currentTask\` specific (e.g., "Refactoring auth middleware" not "Working
 }
 
 /**
- * Extract session insights from assembled context string.
- * Parses session summaries, blockers, and progress data.
- */
-function extractSessionInsights(assembledContext: string): {
-  lastSessionStatus: "completed" | "paused" | null;
-  lastSessionTask: string | null;
-  lastSessionProgress: number | null;
-  blockers: string[];
-  procedureNextStep: string | null;
-} {
-  const result = {
-    lastSessionStatus: null as "completed" | "paused" | null,
-    lastSessionTask: null as string | null,
-    lastSessionProgress: null as number | null,
-    blockers: [] as string[],
-    procedureNextStep: null as string | null,
-  };
-
-  // Find the most recent session summary with status
-  const sessionMatches = assembledContext.match(
-    /### Session:.*?\n([\s\S]*?)(?=\n###|\n## |\n---|\n\*Assembly|$)/g,
-  );
-  if (sessionMatches && sessionMatches.length > 0) {
-    const latest = sessionMatches[0];
-    if (/Completed work on/i.test(latest)) {
-      result.lastSessionStatus = "completed";
-    } else if (/Paused work on|status:\s*paused/i.test(latest)) {
-      result.lastSessionStatus = "paused";
-    }
-    const taskMatch = latest.match(/Final task:\s*(.+)/);
-    if (taskMatch) result.lastSessionTask = taskMatch[1].trim();
-    const progressMatch = latest.match(/Progress:\s*(\d+)%/);
-    if (progressMatch)
-      result.lastSessionProgress = parseInt(progressMatch[1], 10);
-  }
-
-  // Extract blockers from context
-  const blockerMatches = assembledContext.match(
-    /(?:blocker|blocked by|blocking):\s*(.+)/gi,
-  );
-  if (blockerMatches) {
-    result.blockers = blockerMatches.map((m) =>
-      m.replace(/(?:blocker|blocked by|blocking):\s*/i, "").trim(),
-    );
-  }
-
-  // Extract next procedure step (first uncompleted step)
-  const stepMatches = assembledContext.match(
-    /^\d+\.\s+(?!.*\*\*\[key step\]\*\*.*✓)(.+?)(?:\s*\*\*\[key step\]\*\*)?$/gm,
-  );
-  if (stepMatches && stepMatches.length > 0) {
-    result.procedureNextStep = stepMatches[0]
-      .replace(/^\d+\.\s+/, "")
-      .replace(/\s*\*\*\[key step\]\*\*.*$/, "")
-      .trim();
-  }
-
-  return result;
-}
-
-/**
- * Synthesize the single highest-leverage next action from card state
- * and assembled context (session history, blockers, procedures).
+ * Synthesize the single highest-leverage next action from card state.
  * Inspired by ArtemXTech's "One Thing" pattern in /recall.
  */
 function synthesizeOneThing(
@@ -689,7 +607,6 @@ function synthesizeOneThing(
     display_type: string;
     direction: "outgoing" | "incoming";
   }>,
-  assembledContext?: string,
 ): string | null {
   // Priority 1: Card is already done
   if (card.done) return null;
@@ -703,25 +620,7 @@ function synthesizeOneThing(
     return `Unblock first: resolve #${blocker.target_card.short_id} "${blocker.target_card.title}" which is blocking this card.`;
   }
 
-  // Extract session insights from assembled context
-  const session = assembledContext
-    ? extractSessionInsights(assembledContext)
-    : null;
-
-  // Priority 3: Blockers detected in session context
-  if (session?.blockers && session.blockers.length > 0) {
-    return `Resolve blocker: ${session.blockers[0]}`;
-  }
-
-  // Priority 4: Previous session was paused — resume where it left off
-  if (session?.lastSessionStatus === "paused" && session.lastSessionTask) {
-    const progress = session.lastSessionProgress
-      ? ` (was ${session.lastSessionProgress}% complete)`
-      : "";
-    return `Resume previous session${progress}: "${session.lastSessionTask}".`;
-  }
-
-  // Priority 5: Has subtasks — find the first incomplete one
+  // Priority 3: Has subtasks — find the first incomplete one
   if (subtasks.length > 0) {
     const completed = subtasks.filter((s) => s.completed).length;
     if (completed === subtasks.length) {
@@ -733,17 +632,7 @@ function synthesizeOneThing(
     }
   }
 
-  // Priority 6: Procedure has a next step
-  if (session?.procedureNextStep) {
-    return `Follow procedure: ${session.procedureNextStep}`;
-  }
-
-  // Priority 7: Previous session completed — build on it
-  if (session?.lastSessionStatus === "completed" && session.lastSessionTask) {
-    return `Previous session completed ("${session.lastSessionTask}"). Review results and continue with remaining work.`;
-  }
-
-  // Priority 8: High/urgent priority with due date
+  // Priority 4: High/urgent priority with due date
   if (
     card.due_date &&
     (card.priority === "urgent" || card.priority === "high")
@@ -751,7 +640,7 @@ function synthesizeOneThing(
     return `High-priority task with deadline ${card.due_date}. Start implementation immediately.`;
   }
 
-  // Priority 9: Has description — start working
+  // Priority 5: Has description — start working
   if (card.description) {
     return "Analyze the description, identify the approach, and begin implementation.";
   }

package/src/read-consumer.ts

@@ -0,0 +1,16 @@
+// Read-consumer labels for `record_entity_reads` (migration
+// 20260828120000_knowledge_entity_reads.sql).
+//
+// Duplicated in `supabase/functions/_shared/read-consumer.ts` rather than
+// shared via `packages/harmony-shared`: this package has no dependency on
+// `@harmony/shared` (so it could not import the type from there), and
+// `packages/mobile` DOES depend on `@harmony/shared` (so adding it there
+// would drag mobile's typecheck CI into a change mobile never uses). Both
+// copies must match the SQL `CHECK` constraint exactly — six values, no more.
+export type ReadConsumer =
+  | "agent-prompt"
+  | "assistant"
+  | "mcp-tool"
+  | "analyze"
+  | "search"
+  | "browse";

package/src/server.ts

@@ -69,6 +69,7 @@ import {
 } from "./memory-session.js";
 import { lintTags, normalizeTags } from "./memory-tags.js";
 import { onboardNewUser } from "./onboard.js";
+import { collectPlaybookMetricWarnings } from "./playbook-metric-warnings.js";
 import { stripSkillPreamble } from "./skills.js";
 
 // --- Signed-upload handshake (artifacts & card attachments) ---
@@ -1206,7 +1207,7 @@ export const TOOLS = {
   },
   harmony_classify_card: {
     description:
-      "Classify a card with the LLM classifier: sets `intent` (plan/think/implement/review), `complexity_score` (0-10), `model_tier` (simple/advanced/research), stamps `classified_at`, and applies the type label (feature/bug/idea). Call right after creating a card to classify it in-flow. Idempotent; never touches the user-owned `model_override`.",
+      "DEPRECATED — run sizing now happens at daemon pickup and is run-scoped, so nothing reads `model_tier`, `intent` or `complexity_score` any more; this tool still writes them, but only the type label has an effect. Prefer letting card creation apply the type label. Sets `intent` (plan/think/implement/review), `complexity_score` (0-10), `model_tier` (simple/advanced/research), stamps `classified_at`, and applies the type label (feature/bug/idea). Idempotent; never touches the user-owned `model_override`.",
     inputSchema: {
       type: "object",
       properties: {
@@ -4680,6 +4681,12 @@ async function handleToolCall(
                 : undefined,
             include_superseded: includeSuperseded,
             include_episodes: includeEpisodes,
+            // No consumer here on purpose (fix round 1, task 4): this fetches
+            // an over-fetched candidate pool (up to `fetchLimit`, itself
+            // padded above `requestedLimit`/`topK`) for Park to rescore, not
+            // what the tool actually returns. Recording it would overstate
+            // access_count by up to fetchLimit/finalCount per call. What we
+            // actually deliver is recorded below, after trimming.
           },
         );
         entities = (searchResult.entities ?? []) as any[];
@@ -4741,19 +4748,19 @@ async function handleToolCall(
         trimmed = fitToBudget(trimmed, budgetTokens);
       }
 
-      // Touch access counters for the entities we returned (#273, task 4).
-      // Previously this wrote `metadata._last_recall` via updateMemoryEntity,
-      // which left `access_count` at 0 forever and never updated
-      // `last_accessed_at` — the column the Park recency term actually reads.
-      // Route through batch-touch, which calls the atomic
-      // `batch_touch_knowledge_entities` RPC (access_count += 1, last_accessed_at
-      // = now) in a single round-trip. Non-blocking; best-effort.
+      // Read accounting (task 4, fix round 1): record only the entities
+      // actually delivered to the caller — after trimming and the budget
+      // pass, not the larger over-fetched candidate pool the search above
+      // deliberately left unrecorded. Routed through batch-touch, which
+      // dispatches to the same recordEntityReads/record_entity_reads path
+      // the search chokepoint uses when a consumer is given. Non-blocking;
+      // best-effort.
       if (trimmed.length > 0) {
         const touchIds = trimmed
           .map(({ entity }: { entity: any }) => entity?.id)
           .filter((id: unknown): id is string => typeof id === "string");
         if (touchIds.length > 0) {
-          client.batchTouchMemoryEntities(touchIds).catch(() => {});
+          client.batchTouchMemoryEntities(touchIds, "mcp-tool").catch(() => {});
         }
       }
 
@@ -5370,7 +5377,18 @@ async function handleToolCall(
         autoBind: args.autoBind,
         catalogId: args.catalogId as string | undefined,
       });
-      return { success: true, playbook: result.playbook };
+      // #922: surface custom-gate metrics no agent declares — non-blocking,
+      // the playbook is already created.
+      const warnings = await collectPlaybookMetricWarnings(
+        client,
+        workspaceId,
+        args.steps,
+      );
+      return {
+        success: true,
+        playbook: result.playbook,
+        ...(warnings.length > 0 ? { warnings } : {}),
+      };
     }
 
     case "harmony_update_playbook": {
@@ -5386,7 +5404,19 @@ async function handleToolCall(
         triggerType: args.triggerType as string | undefined,
         ...("autoBind" in args ? { autoBind: args.autoBind } : {}),
       });
-      return { success: true, playbook: result.playbook };
+      // #922: same non-blocking metric check as create. The playbook's own
+      // workspace scopes the lookup — the active context may point elsewhere.
+      const warnings = await collectPlaybookMetricWarnings(
+        client,
+        (result.playbook as { workspace_id?: string } | undefined)
+          ?.workspace_id,
+        args.steps,
+      );
+      return {
+        success: true,
+        playbook: result.playbook,
+        ...(warnings.length > 0 ? { warnings } : {}),
+      };
     }
 
     // Deprecated (#612) — see harmony_run_playbook above.