@retaw106/mmc-oc 2.5.0 → 2.7.1

package/src/shared/prompt-surface-a1-golden.md

@@ -1,6 +1,6 @@
 # Magic Context agent-facing surface (exported from source)
 
-Generated by `packages/plugin/scripts/export-agent-surface.ts` at 2026-10-01.
+Generated by `packages/plugin/scripts/export-agent-surface.ts` at 2026-11-10.
 Token counts are Claude BPE estimates on the raw text.
 
 ## 1. System-prompt guidance section
@@ -201,7 +201,7 @@ Finer recovery:
 }
 ```
 
-### mmc_note — description ~319 tokens, params ~203 tokens (total ~522)
+### mmc_note — description ~319 tokens, params ~196 tokens (total ~515)
 
 **Description:**
 
@@ -228,7 +228,7 @@ First line is the title (under 80 chars), followed by detail. Operations:
 ```json
 {
   "action": {
-    "description": "write | read | update | dismiss. Defaults to write when content is given, else read.",
+    "description": "write | read | update | dismiss.",
     "type": "string",
     "enum": [
       "write",
@@ -259,8 +259,7 @@ First line is the title (under 80 chars), followed by detail. Operations:
     "type": "number"
   },
   "note_ids": {
-    "description": "Note ids: one for update, 1–50 for dismiss or read (read returns full bodies). Ignored by write.",
-    "minItems": 1,
+    "description": "Note ids: one for update, 1–50 for dismiss, or 1–50 to read those notes in full. Write and a plain read take an empty list.",
     "maxItems": 50,
     "type": "array",
     "items": {
@@ -272,7 +271,7 @@ First line is the title (under 80 chars), followed by detail. Operations:
 }
 ```
 
-### mmc_memory — description ~279 tokens, params ~172 tokens (total ~451)
+### mmc_memory — description ~279 tokens, params ~198 tokens (total ~477)
 
 **Description:**
 
@@ -306,11 +305,11 @@ Examples: category="CONFIG_VALUES", content="OpenCode source is at ~/Work/OSS/op
     ]
   },
   "content": {
-    "description": "The memory text — one standalone fact (write, update, merge).",
+    "description": "The memory text — one standalone fact (write, update, merge). Archive and get take an empty string.",
     "type": "string"
   },
   "category": {
-    "description": "Kind of fact (required for write; on update/merge optional, omitted keeps the current category).",
+    "description": "Kind of fact: the new fact's kind for write; on update/merge, the memory's current category to keep it. Archive and get ignore it.",
     "type": "string",
     "enum": [
       "PROJECT_RULES",
@@ -321,7 +320,7 @@ Examples: category="CONFIG_VALUES", content="OpenCode source is at ~/Work/OSS/op
     ]
   },
   "ids": {
-    "description": "Memory ids from <project-memory>: one for update, one or more for archive, two or more for merge, 1–20 for get.",
+    "description": "Memory ids from <project-memory>: one for update, one or more for archive, two or more for merge, 1–20 for get. Write takes an empty list.",
     "type": "array",
     "items": {
       "type": "number"

package/src/shared/search-text.test.ts

@@ -1,5 +1,40 @@
 import { expect, test } from "bun:test";
-import { cleanSearchText, keywordRelevance, matchedSearchSnippet } from "./search-text";
+import {
+    cleanSearchText,
+    hintKeywordRelevance,
+    keywordRelevance,
+    matchedSearchSnippet,
+} from "./search-text";
+
+test("hint scoring preserves exact English and other space-separated keyword scores", () => {
+    const cases: Array<[string, string, number]> = [
+        ["explain historian cache wiring", "cache icons", 0.2],
+        ["historian cache wiring", "historian cache wiring details", 0.8000000000000002],
+        ["please explain cache wiring to me", "cache wiring", 0.5333333333333333],
+        ["CACHE cache cache_timeout", "cache_timeout", 0.4],
+        ["how are you", "how are you", 0],
+        ["historian cache", "<task>historian cache</task>", 0],
+        ["résumé über prüfen", "über prüfen", 0.5333333333333333],
+        ["память кеш", "кеш", 0.4],
+        ["ذاكرة تخزين", "تخزين", 0.4],
+    ];
+    for (const [query, content, expected] of cases) {
+        expect(keywordRelevance(query, content)).toBe(expected);
+        expect(hintKeywordRelevance(query, content)).toBe(expected);
+    }
+});
+
+test("only hint scoring splits unspaced CJK runs into overlapping character bigrams", () => {
+    expect(keywordRelevance("MC_Maze 残差怎样汇总？", "MC_Maze 残差按种子汇总。")).toBe(0.4);
+    expect(hintKeywordRelevance("残差", "残留")).toBe(0);
+    expect(hintKeywordRelevance("差", "残差")).toBe(0.8);
+    expect(hintKeywordRelevance("𠮷野家", "𠮷野川")).toBe(0.8);
+    expect(hintKeywordRelevance("かながき", "かながきです")).toBe(0.8);
+    expect(hintKeywordRelevance("カタカナ", "カタカナです")).toBe(0.8);
+    expect(hintKeywordRelevance("프로젝트캐시", "프로젝트의캐시")).toBe(0.8);
+    expect(hintKeywordRelevance("cache残差", "cache残差按种子汇总")).toBe(0.8);
+    expect(hintKeywordRelevance("差", "<task>残差</task> unrelated")).toBe(0);
+});
 
 test("notification trees never contribute relevance or display text", () => {
     expect(

package/src/shared/search-text.ts

@@ -35,6 +35,54 @@ export function keywordRelevance(query: string, content: string): number {
     return (0.8 * needles.filter((term) => words.has(term)).length) / needles.length;
 }
 
+const CJK_RUN = /^[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}\p{Script=Hangul}]+$/u;
+const CJK_RUN_SPLIT = /([\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}\p{Script=Hangul}]+)/u;
+
+function hintTerms(text: string): string[] {
+    return [
+        ...new Set(
+            terms(text)
+                .flatMap((term) => term.split(CJK_RUN_SPLIT))
+                .filter((term) => term.length > 0 && !QUERY_STOP_WORDS.has(term)),
+        ),
+    ];
+}
+
+function characterBigrams(run: string): string[] {
+    const characters = [...run];
+    return characters.length === 1
+        ? characters
+        : characters.slice(1).map((character, index) => characters[index] + character);
+}
+
+/** Match unspaced clauses for hints without changing retrieval or snippet scoring. */
+export function hintKeywordRelevance(query: string, content: string): number {
+    const needles = hintTerms(query);
+    if (needles.length === 0) return 0;
+    const words = new Set(hintTerms(cleanSearchText(content)));
+    const runs = [...words].filter((word) => CJK_RUN.test(word));
+    const bigrams = new Set(runs.flatMap(characterBigrams));
+    let matched = 0;
+    for (const needle of needles) {
+        if (!CJK_RUN.test(needle)) {
+            matched += Number(words.has(needle));
+            continue;
+        }
+        const grams = characterBigrams(needle);
+        if ([...needle].length === 1) {
+            matched += Number(runs.some((run) => run.includes(needle)));
+            continue;
+        }
+        // Count each clause as one term so filler bigrams across word boundaries
+        // do not dilute it; require half coverage so shared function words alone do not pass.
+        matched += Math.min(
+            1,
+            grams.filter((gram) => bigrams.has(gram)).length / Math.ceil(grams.length / 2),
+        );
+    }
+    return (0.8 * matched) / needles.length;
+}
+
 /** Pick the densest matching passage before compression discards matched words. */
 export function matchedSearchSnippet(content: string, query: string, cap = 220): string {
     const text = cleanSearchText(content);

package/src/shared/sqlite-transaction-mode-guard.test.ts

@@ -27,6 +27,8 @@ const READ_ONLY_DEFERRED_TRANSACTIONS: Record<string, string> = {
         "pins one read snapshot so the session's rowid coverage proof and the filtered message search see the same rows; never writes (read-only connections use it too)",
     "packages/plugin/src/hooks/magic-context/lkg-persist.ts#loadPersistedLkgSlot":
         "reads an LKG slot row and its prefix slices as one consistent snapshot even while another connection saves the slot; never writes (a failed verification clear opens its own transaction)",
+    "packages/plugin/src/hooks/magic-context/compaction-marker-manager.ts#reconcileForkOrphanedCompactionMarkers":
+        "reads publication intent and finalized ownership as one context.db snapshot while host-order probes use the already-open host connection; never writes context.db",
 };
 
 const REPOSITORY_ROOT = resolve(import.meta.dir, "../../../..");

package/src/shared/tag-transcript.ts

@@ -75,6 +75,8 @@ import type { Transcript, TranscriptMessage, TranscriptPart } from "./transcript
 export { TEXT_TAG_IDENTITY_MARKER };
 
 export interface TagTranscriptOptions {
+    /** Independently validated text sources; does not authorize tool accounting reuse. */
+    reuseTextMessageIds?: ReadonlySet<string>;
     /**
      * When true, skip injecting `§N§` prefix into visible text. Tags
      * still get assigned in the DB so historian/drops can reference
@@ -305,6 +307,12 @@ interface ToolAggregate {
     maxByteSize: number;
     /** Token count paired with maxByteSize (the same output occurrence). */
     maxTokenCount: number;
+    /**
+     * Whether maxTokenCount is a real count. Legacy rows can store bytes with a
+     * NULL token_count; seeding 0 from that must not look like a stored count,
+     * or unchanged results would never be recounted.
+     */
+    tokenCountKnown: boolean;
     /** Tool name from the first occurrence we see one on. */
     toolName: string | null;
     /** Input byte size from the invocation occurrence (for storage projection). */
@@ -417,7 +425,10 @@ export function tagTranscript(
                     targets,
                     skipPrefixInjection,
                     entryFingerprint: options.entryFingerprintByMessageId?.get(messageId) ?? null,
-                    reuseIdentity: reuseIdentity || contentDerivedTextIds !== undefined,
+                    reuseIdentity:
+                        reuseIdentity ||
+                        options.reuseTextMessageIds?.has(messageId) === true ||
+                        contentDerivedTextIds !== undefined,
                     timing,
                     textIdentitySourceCache: options.textIdentitySourceCache,
                     textTokenCache: options.textTokenCache,
@@ -508,6 +519,13 @@ export function tagTranscript(
                             timing,
                             options.toolTokenCache,
                             tokenCacheKey,
+                            {
+                                byteSize: existing.maxByteSize,
+                                tokenCount: existing.tokenCountKnown
+                                    ? existing.maxTokenCount
+                                    : null,
+                                inputTokenCount: existing.inputTokenCount,
+                            },
                         );
                         text = accounting.text;
                         if (part.kind === "tool_result") {
@@ -596,6 +614,7 @@ export function tagTranscript(
                         // genuinely larger folded result can still bump accounting.
                         maxByteSize: reusableAccounting.byteSize,
                         maxTokenCount: reusableAccounting.tokenCount ?? 0,
+                        tokenCountKnown: reusableAccounting.tokenCount !== null,
                         toolName: null,
                         inputByteSize: reusableAccounting.inputByteSize,
                         inputTokenCount: reusableAccounting.inputTokenCount,
@@ -623,6 +642,7 @@ export function tagTranscript(
                         timing,
                         options.toolTokenCache,
                         tokenCacheKey,
+                        tagger.getToolTagAccounting(sessionId, callId, messageId),
                     );
                     text = accounting.text;
                     const outputByteSize = part.kind === "tool_result" ? accounting.byteSize : 0;
@@ -657,6 +677,11 @@ export function tagTranscript(
                         occurrences: [{ message, part, kind: part.kind }],
                         maxByteSize: persistedAccounting?.byteSize ?? outputByteSize,
                         maxTokenCount: persistedAccounting?.tokenCount ?? outputTokenCount,
+                        // A fresh allocation just wrote outputTokenCount (0 for an
+                        // invocation); an existing row keeps its NULL as unknown.
+                        tokenCountKnown: persistedAccounting
+                            ? persistedAccounting.tokenCount !== null
+                            : true,
                         toolName: accounting.toolName,
                         inputByteSize:
                             persistedAccounting?.inputByteSize ??
@@ -745,7 +770,9 @@ function syncToolAggregateAccounting(
 ): void {
     tagger.setToolTagAccounting(sessionId, aggregate.tagId, {
         byteSize: aggregate.maxByteSize,
-        tokenCount: aggregate.maxTokenCount,
+        // Unknown stays unknown so the next occurrence/pass recounts instead of
+        // caching a fabricated 0 for a legacy row.
+        tokenCount: aggregate.tokenCountKnown ? aggregate.maxTokenCount : null,
         inputByteSize: aggregate.inputByteSize,
         inputTokenCount: aggregate.inputTokenCount,
     });
@@ -770,6 +797,7 @@ function applyGrownToolResultAccounting(args: GrownToolResultAccountingArgs): vo
     }
     args.aggregate.maxByteSize = args.byteSize;
     args.aggregate.maxTokenCount = tokenCount;
+    args.aggregate.tokenCountKnown = true;
     updateTagByteSize(args.db, args.sessionId, args.aggregate.tagId, args.byteSize);
     updateTagTokenCount(args.db, args.sessionId, args.aggregate.tagId, tokenCount);
     syncToolAggregateAccounting(args.tagger, args.sessionId, args.aggregate);
@@ -780,12 +808,22 @@ function readAggregateToolAccounting(
     timing: TagTranscriptTiming | undefined,
     tokenCache?: Map<string, { text: string; tokenCount: number }>,
     tokenCacheKey?: string,
+    persisted?: { byteSize: number; tokenCount: number | null; inputTokenCount: number | null },
 ): AggregateToolAccounting {
     const text = part.getText() ?? "";
     const byteSize = getToolPartByteSize(part, text);
-    const metadata = part.getToolMetadata();
+    // Persisted tool-result accounting tracks the largest byte size seen. Skip
+    // output tokenization when that size has not grown and a count is stored;
+    // known input counts also need no tokenization. Growth and missing counts still take
+    // the original counting path, even when the identity was not reusable.
+    const metadata = part.getToolMetadata({
+        skipInputTokenCount: persisted?.inputTokenCount != null,
+    });
     let tokenCount = 0;
-    if (part.kind === "tool_result") {
+    if (
+        part.kind === "tool_result" &&
+        (!persisted || persisted.tokenCount === null || byteSize > persisted.byteSize)
+    ) {
         const tokenStart = timing ? performance.now() : 0;
         tokenCount = getCachedToolPartTokenCount(part, text, tokenCache, tokenCacheKey);
         if (timing) timing.tokenCounting += performance.now() - tokenStart;
@@ -1414,6 +1452,11 @@ function buildAggregateTarget(
         },
         // Open invocations still belong to the current turn; only complete arcs reclaim.
         prepareDrop: () => !occurrences.some((occ) => occ.part.canRemove?.() === "defer"),
+        prepareReplayDrop: () => {
+            if (complete() && !requiresToolArcSkeleton) {
+                for (const occ of occurrences) occ.part.canRemove?.();
+            }
+        },
         canDrop: () => complete() && !occurrences.some((occ) => occ.part.hasUserAnswer?.()),
         requiresToolArcSkeleton,
         // Non-mutating read of the invocation input (the tool_use occurrence

package/src/shared/transcript.ts

@@ -132,7 +132,7 @@ export interface TranscriptPart {
      *
      * For non-tool parts both byte fields are undefined/0.
      */
-    getToolMetadata(): {
+    getToolMetadata(options?: { skipInputTokenCount?: boolean }): {
         toolName: string | undefined;
         inputByteSize: number;
         inputTokenCount: number;

package/src/shared/write-transaction-attribution.test.ts

@@ -14,11 +14,7 @@ const fencedSources: Array<{ path: string; sites: string[]; reporterCalls?: stri
     },
     {
         path: "packages/plugin/src/features/magic-context/message-index.ts",
-        sites: [
-            "message_index_incremental",
-            "message_index_reconcile",
-            "message_index_orphan_sweep",
-        ],
+        sites: ["message_index_incremental", "message_index_reconcile"],
     },
     {
         path: "packages/plugin/src/features/magic-context/storage-meta-persisted.ts",
@@ -101,7 +97,6 @@ const newlyCoveredSites = [
     "opencode_soft_refresh_cache",
     "message_index_incremental",
     "message_index_reconcile",
-    "message_index_orphan_sweep",
     "message_index_clear",
     "message_fts_rowid_backfill",
     "storage_meta_wrapup_expiry_cleanup",

package/tui.js

@@ -0,0 +1,4 @@
+// mmc fork: root TUI entry mirror (T-003 B-tier), shape mirrors upstream's
+// tui.js: the host's "./tui" export resolves here, re-exporting the bundled
+// dist entry.
+export { default } from "./dist/tui.js";

package/dist/features/magic-context/message-index-sweep.d.ts

@@ -1,11 +0,0 @@
-import type { Database } from "../../shared/sqlite";
-import { type MessageHistoryOrphanSweepOptions, type MessageHistoryOrphanSweepResult } from "./message-index";
-/**
- * Scope wrapper for the background sweep. It lives beside `message-index` rather
- * than inside it: importing the transform-pass scope helpers into
- * `message-index.ts` itself changes the bundled module graph in a way that hits
- * a Bun 1.4.2 packaging bug (a top-level `__promiseAll` init call is emitted
- * without its definition, which crashes node hosts at import time).
- */
-export declare function sweepOrphanedOpenCodeMessageIndexes(db: Database, openReadableOpenCodeDb: () => Database | null, options?: MessageHistoryOrphanSweepOptions): MessageHistoryOrphanSweepResult;
-//# sourceMappingURL=message-index-sweep.d.ts.map

package/dist/features/magic-context/message-index-sweep.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"message-index-sweep.d.ts","sourceRoot":"","sources":["../../../src/features/magic-context/message-index-sweep.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,qBAAqB,CAAC;AAEpD,OAAO,EACH,KAAK,gCAAgC,EACrC,KAAK,+BAA+B,EAEvC,MAAM,iBAAiB,CAAC;AAEzB;;;;;;GAMG;AACH,wBAAgB,mCAAmC,CAC/C,EAAE,EAAE,QAAQ,EACZ,sBAAsB,EAAE,MAAM,QAAQ,GAAG,IAAI,EAC7C,OAAO,GAAE,gCAAqC,GAC/C,+BAA+B,CAMjC"}