negotium 0.2.18 → 0.2.19

package/dist/runtime/src/platform/mcp-config.ts

@@ -474,7 +474,7 @@ const MCP_CATALOG: Record<string, RuntimeMcpCatalogEntry> = {
       }
       const args = [`--user-id=${userId}`];
       if (topicId) args.push(`--current-topic-id=${topicId}`);
-      // Prefer the REST topic id so save_wiki_entry writes filenames and SQLite
+      // Prefer the REST topic id so wiki_write writes filenames and SQLite
       // briefs keyed the same way the topic wiki API reads them.
       const resolvedWikiTopicId =
         wikiTopicId ?? topicId ?? (session !== "dm" ? session : undefined);

package/dist/runtime/src/prompts/agents/wiki-archiver.md

@@ -1,32 +1,38 @@
 ---
 name: wiki-archiver
 type: programmatic
-description: Agent that extracts key content from session logs and updates wiki/summaries/, wiki/articles/, wiki/topic/, wiki/skills/, wiki/article-index.md, wiki/topic-index.md, and wiki/skill-index.md
+description: Agent that extracts key content from session logs and updates wiki/summaries/, wiki/articles/, wiki/topic/, wiki/skills/, wiki/summary-index.md, wiki/article-index.md, wiki/topic-index.md, and wiki/skill-index.md
 model: deepseek-pro
 tools:
   - Read
-  - Write
   - Glob
-  - mcp__wiki__save_wiki_entry
-  - mcp__wiki__save_topic_brief
+  - mcp__wiki__wiki_write
+  - mcp__wiki__wiki_read
   - mcp__wiki__skill_save
   - mcp__wiki__skill_query
   - mcp__wiki__index_upsert
 ---
 
 You are a wiki archiver agent. Extract key information from session logs and save it into the
-wiki knowledge base. The wiki has **two index files** that must be kept in sync with the underlying
-content directories:
+wiki knowledge base. Every document has exactly one catalog, and **`wiki_write` writes the document
+and its catalog row in the same call** — so you cannot leave a document unindexed:
 
 ```
 wiki/
-  summaries/<date>-<topic>.md   <- session summaries (write-once)
-  articles/<slug>.md            <- curated concept pages (mergeable)
-  topic/<topic>.md              <- accumulated persona brief (one file per title)
-  article-index.md              <- catalog: articles + summaries
-  topic-index.md                <- catalog: topic briefs only
+  summaries/<date>-<topic>.md   <- session summaries (write-once)   -> summary-index.md
+  articles/<slug>.md            <- curated concept pages (mergeable) -> article-index.md
+  topic/<topic>.md              <- accumulated persona brief         -> topic-index.md
+  .wiki-search-index.sqlite     <- derived body-search cache (never edit or archive)
 ```
 
+`wiki_write` also feeds the derived cache, so a document is searchable by its body the moment it is
+written. The cache is disposable: if a call reports that it was not updated, the catalogs still
+work and `wiki_reindex` refills it.
+
+**You have no file-writing tool.** `wiki_write` is the only way to create or update a wiki
+document; `description` is mandatory because the catalog is what makes retrieval work. Use `Read`
+and `Glob` to inspect existing files before merging.
+
 ## Output language
 
 Write all human-readable prose (summaries, the persona brief, article bodies, skill descriptions, your final report) in the `output_language:` given in the prompt; default to **English** if absent. `output_language` tracks the user's mother tongue (e.g. `English`, `Korean`, `한국어`, `Japanese`).
@@ -40,7 +46,7 @@ Keep structural tokens in English regardless: frontmatter keys, `type:` values,
    - Read the first chunk with `Read(archive_path, offset: 1, limit: 2000)`.
    - If the result ends with a truncation notice like `lines X-Y of N`, call the next chunk with `offset = Y + 1`. Stop when there is no notice or `Y == N`.
    - For each chunk, **only accumulate** key items (decisions / facts / tools / files / patterns …) into a short in-memory bullet buffer.
-   - **Do not call wiki write / save / index_upsert per chunk.** After reading every chunk, run steps 3 onward exactly once against the accumulated buffer. Saving per chunk splits one session across multiple summary files and piles up duplicate index_upsert calls.
+   - **Do not call `wiki_write` per chunk.** After reading every chunk, run steps 3 onward exactly once against the accumulated buffer. Writing per chunk splits one session across multiple summary files.
    - If the buffer grows large (e.g. 3000+ bullets, ≥ 5 chunks), compress/drop trivial earlier items and keep only the essential decisions / facts / patterns. Preserving cross-references through the last chunk takes priority.
 
 2. **Extract** key information (decisions, facts, patterns, tools — skip greetings, debug noise, repeated questions).
@@ -55,22 +61,28 @@ Keep structural tokens in English regardless: frontmatter keys, `type:` values,
    - Use `canonical_topic` for every summary, brief, and topic-index write below. This keeps a
      differently named room's summary and persona update in one canonical namespace.
    - If `sent_files:` is in the prompt, include those entries under `## Files Sent`
-   - If the session yielded **no extractable substance** (pure debug, ≤2 short exchanges, all greeting), save only a single-line immutable summary via `save_wiki_entry`, then STOP. Do not modify the accumulated persona brief, articles, or indexes.
+   - If the session yielded **no extractable substance** (pure debug, ≤2 short exchanges, all greeting), save only a single-line immutable summary via `wiki_write(kind="summary", ...)`, then STOP. Do not touch the persona brief or articles.
 
 > **Ordering principle:** pipeline is **archive → summary → brief**. The summary is *this* session's raw distillation; the persona brief is the slow-moving user model that folds each summary in. Save the summary first and update the brief last, so the brief sees everything the session produced.
 
-3. **Save the immutable session summary** via `mcp__wiki__save_wiki_entry(canonical_topic, content)`.
-   The MCP handles file naming + dedup → returns the saved path (e.g. `wiki/summaries/2026-05-08-dev.md`).
-   It also records `latest_summary_md` + `summary_date` in SQLite (it does **not** touch the brief —
-   that is done in step 5 via `save_topic_brief`). Use the **summary format** below.
+3. **Save the immutable session summary** via
+   `mcp__wiki__wiki_write(kind="summary", topic=canonical_topic, content, description)`.
+   The MCP handles file naming + dedup, writes the `summary-index.md` row from your `description`,
+   and records `latest_summary_md` + `summary_date` in SQLite (it does **not** touch the brief —
+   that is step 5). Use the **summary format** below.
 4. **Update articles** — for each genuinely reusable concept/decision/tool/pattern:
-   - Glob existing articles: `Glob(wiki/articles/*.md)`
-   - If a matching article exists (by slug or topic): Read it, then Write merged content.
+   - Glob existing articles: `Glob(wiki/articles/*.md)`, and `Read(wiki/article-index.md)` once to
+     see the existing `## ...` sections.
+   - If a matching article exists (by slug or topic): `Read` it, merge, then write the merged body.
      - **Preserve frontmatter `date:` (first-seen) and `status:`.** Only refresh `updated:`.
      - **Preserve manually written body sections.** Only append/update what the session adds.
-   - If new: Write `wiki/articles/<slug>.md` using the **article format** below.
+   - Write with
+     `wiki_write(kind="article", slug=<kebab-slug>, section=<H2 without "## ">, content, description)`.
+     `section` is required: pick the closest existing header, or a short new domain title in
+     `output_language` (e.g. `Business / Career`, `Physical AI / Robotics`).
    - Skip session-specific noise. If nothing qualifies, no articles change — that's fine.
-5. **Update the accumulated persona brief last** via `mcp__wiki__save_topic_brief(canonical_topic, content)`.
+5. **Update the accumulated persona brief last** via
+   `mcp__wiki__wiki_write(kind="topic", topic=canonical_topic, content, description)`.
    This is the culmination of the run — the brief is not a worklog, it is the wiki's evolving
    **persona/user-model** for this topic: who the user is, how they want to be served, and where
    things stand. It is injected verbatim at the next session's start, so write it as durable memory,
@@ -81,31 +93,23 @@ Keep structural tokens in English regardless: frontmatter keys, `type:` values,
      Decisions into the **persona layer** (accumulate — the user-model is slow-moving). Refresh only
      the volatile layers (`## Recent Work`, `## Current State`) from this session. Preserve still-valid
      prior persona traits; remove a trait only when the new session explicitly supersedes it.
-   - Keep one canonical file per topic memory key — `save_topic_brief` handles the path + SQLite mirror.
-     Never add a UUID or room id. Write a fresh compact brief using the **brief format** below.
-6. **Update the dual indexes via `mcp__wiki__index_upsert` — one call per entry.**
-   The MCP handles in-place updates, section insertion, and the `created` vs `updated` date split. Do **not** Read/Write the index files manually.
-
-   **For each new or updated article** (from step 4):
-   - First, scan `wiki/article-index.md` once with `Read` to see existing `## ...` headers, then pick the closest matching section. If no section fits, choose a short domain title in `output_language` (e.g. `Business / Career`, `Physical AI / Robotics`) — the MCP will create the new H2 above `## Source Summaries`.
-   - Call: `index_upsert(slug=<article-slug>, description=<one-line>, kind="article", section=<chosen-header-without-"## ">)`
-   - The MCP preserves the original `created` date on update — do not pass it.
-
-   **For the new session summary** (from step 3):
-   - Call: `index_upsert(slug=<summary-slug>, description=<one-line>, kind="summary")`
-   - Goes under `## Source Summaries` automatically.
+   - Keep one canonical file per topic memory key — `wiki_write` handles the path, the
+     `topic-index.md` row, and the SQLite mirror. Never add a UUID or room id. Write a fresh compact
+     brief using the **brief format** below, and let `description` be a one-line state of recent work.
+6. **There is no separate indexing step.** Steps 3–5 already wrote every catalog row, so the
+   catalogs cannot drift from the documents. Two rules remain:
 
-   **For this session's topic brief** (from step 5):
-   - Call: `index_upsert(slug=<canonical_topic>, description=<one-line summary of recent work>, kind="topic")`
-   - Pass the bare canonical key (no `topic/` prefix); the MCP wikilinks it as `[[topic/<canonical_topic>]]`.
+   - **`description` carries the retrieval weight.** Write it as the one line you would want to see
+     when searching six months from now: what changed and why it matters, not "session summary".
+   - **`index_upsert` is for corrections only.** It can refine the description or section of an entry
+     that already exists, and it refuses to create a row for a document that was never written.
+     **Never delete entries** — pruning is a `wiki lint` concern.
 
-   **Never delete entries** — `index_upsert` is insert-or-update only; pruning is a `wiki lint` concern.
-
-## Section rules (recap, used when calling `index_upsert(kind="article", section=...)`)
+## Section rules (recap, used for `wiki_write(kind="article", section=...)`)
 
 1. Scan existing `## ...` headers in `article-index.md` first. Pick the closest match.
-2. If none fits, pass a short title in `output_language` for the article's domain — the MCP inserts a new H2 above `## Source Summaries`.
-3. `## Source Summaries` is always the last section; don't try to push articles below it.
+2. If none fits, pass a short title in `output_language` for the article's domain — the MCP appends
+   the new H2. Do not rely on a particular ordering of sections.
 
 ## wiki/summaries/ entry format
 
@@ -226,9 +230,21 @@ _Last updated: {YYYY-MM-DD}_
 
 ## {Domain section 2}
 - ...
+```
+
+## summary-index.md structure (skeleton)
+
+```
+---
+{frontmatter — preserve as-is}
+---
+
+# Wiki Index — Summaries
+
+_Last updated: {YYYY-MM-DD}_
 
-## Source Summaries
-- [[2026-05-08-topic]] — desc (2026-05-08)
+## Session Summaries
+- [[summaries/2026-05-08-topic]] — desc (2026-05-08)
 - ...
 ```
 
@@ -302,7 +318,6 @@ description: "keyword1, keyword2, likely trigger phrases the user would type —
 Summarize in `output_language` (report in pipeline order — summary → articles → brief):
 - 📝 summary: saved `wiki/summaries/<filename>`
 - 📄 articles: created/updated N pages (slugs)
-- 🗂 brief: `wiki/topic/<topic>.md` persona merge (`save_topic_brief`)
-- 📇 article-index: N `index_upsert` calls (article + summary)
-- 📇 topic-index: 1 `index_upsert` call (this topic)
+- 🗂 brief: `wiki/topic/<topic>.md` persona merge
+- 📇 catalogs: summary-index 1 row, article-index N rows, topic-index 1 row (all written by `wiki_write`)
 - 🛠 skill: created/updated `wiki/skills/<name>/skill.md` (or 'none')

package/dist/runtime/src/version.ts

@@ -1 +1 @@
-export const NEGOTIUM_VERSION = "0.2.18";
+export const NEGOTIUM_VERSION = "0.2.19";

package/dist/types/packages/core/src/agents/archiver.d.ts

@@ -73,7 +73,7 @@ export interface ArchiverRuntime {
 export declare function createArchiverRuntime(host: ArchiverHost): ArchiverRuntime;
 /**
  * Locate the summary the archiver turn just wrote. The wiki MCP's
- * `save_wiki_entry` writes `summaries/<date>-<title>[~N].md`.
+ * `wiki_write(kind="summary")` writes `summaries/<date>-<title>[~N].md`.
  */
 export declare function findSummaryFile(storage: ArchiverStorageHost, topicTitle: string, date: string, sinceMs: number, topicId?: string): string | null;
 export declare function runArchiverTurn(params: RunArchiverTurnParams): boolean;

package/dist/types/packages/core/src/mcp/wiki-search-index.d.ts

@@ -0,0 +1,59 @@
+export type IndexedWikiKind = "article" | "summary";
+export interface WikiSearchIndexSource {
+    kind: IndexedWikiKind;
+    key: string;
+    path: string;
+    mtimeMs: number;
+    size: number;
+}
+export interface PreparedWikiSearchDocument {
+    title: string;
+    text: string;
+    tokens: string[];
+    date?: string;
+    family?: string;
+}
+export interface WikiSearchIndexDocument extends PreparedWikiSearchDocument {
+    id: string;
+    kind: IndexedWikiKind;
+    key: string;
+    path: string;
+}
+export interface WikiSearchIndexMetadata {
+    id: string;
+    kind: IndexedWikiKind;
+    key: string;
+    title: string;
+    date?: string;
+}
+export interface WikiSummaryTemporalQuery {
+    year?: string;
+    month?: string;
+    direction: "latest" | "oldest";
+    limit: number;
+}
+export interface WikiSearchIndexSyncResult {
+    added: number;
+    updated: number;
+    removed: number;
+}
+export declare class WikiSearchIndex {
+    #private;
+    readonly path: string;
+    constructor(path: string);
+    close(): void;
+    /**
+     * Record one document that was just written.
+     *
+     * The write path already holds the document text, so the index is maintained
+     * incrementally and never needs to scan the content directories. `sync` stays
+     * for the explicit backfill of documents that predate the write tool.
+     */
+    upsert(source: WikiSearchIndexSource, document: PreparedWikiSearchDocument): void;
+    sync(sources: WikiSearchIndexSource[], prepare: (source: WikiSearchIndexSource, text: string) => PreparedWikiSearchDocument): WikiSearchIndexSyncResult;
+    metadata(kinds: IndexedWikiKind[]): WikiSearchIndexMetadata[];
+    summaryMetadataByDate(query: WikiSummaryTemporalQuery): WikiSearchIndexMetadata[];
+    matchingDocumentIds(kinds: IndexedWikiKind[], terms: string[]): Set<string>;
+    documents(ids: Set<string>): WikiSearchIndexDocument[];
+}
+export declare function resetCorruptWikiSearchIndex(path: string): void;

package/dist/types/packages/core/src/version.d.ts

@@ -1 +1 @@
-export declare const NEGOTIUM_VERSION = "0.2.18";
+export declare const NEGOTIUM_VERSION = "0.2.19";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "negotium",
-  "version": "0.2.18",
+  "version": "0.2.19",
   "type": "module",
   "description": "Install the Negotium multi-agent runtime and CLI with one package",
   "license": "Apache-2.0",