@mercury-fw/core 0.29.3 → 0.29.4

package/CHANGELOG.md

@@ -1,5 +1,18 @@
 # @mercury-fw/core
 
+## 0.29.4
+
+### Patch Changes
+
+- 63f3da2: - Before writing to the wiki, the agent looks for a document on the same topic and updates it, instead of adding a new one each time.
+  - The agent no longer writes wiki notes restating a CLI's syntax or flags, which the plugin's skill and `--help` already cover.
+  - A learned command correction is keyed by the flag or subcommand it is about, without the tool's name, so corrections on the same flag land in the same note.
+  - The prompts that extract command corrections and facts about the user are in English.
+  - @mercury-fw/plugin-types@0.29.4
+  - @mercury-fw/channel-types@0.29.4
+  - @mercury-fw/cli-engine@0.29.4
+  - @mercury-fw/confirm-engine@0.29.4
+
 ## 0.29.3
 
 ### Patch Changes

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mercury-fw/core",
-  "version": "0.29.3",
+  "version": "0.29.4",
   "type": "module",
   "license": "MIT",
   "repository": {
@@ -31,10 +31,10 @@
     "test": "bun test"
   },
   "dependencies": {
-    "@mercury-fw/channel-types": "0.29.3",
-    "@mercury-fw/cli-engine": "0.29.3",
-    "@mercury-fw/confirm-engine": "0.29.3",
-    "@mercury-fw/plugin-types": "0.29.3",
+    "@mercury-fw/channel-types": "0.29.4",
+    "@mercury-fw/cli-engine": "0.29.4",
+    "@mercury-fw/confirm-engine": "0.29.4",
+    "@mercury-fw/plugin-types": "0.29.4",
     "@qdrant/js-client-rest": "^1.19.0",
     "ai": "^7.0.126",
     "ai-sdk-ollama": "^4.4.0",

package/src/session/semantic-fact-extractor.ts

@@ -34,13 +34,12 @@ type GenerateObjectFn = (params: {
 }) => Promise<{ object: SemanticFact[] }>;
 
 const SYSTEM_PROMPT =
-  "Estrai fatti stabili e ricorrenti sull'utente da questa conversazione, scegliendo il topic " +
-  'esclusivamente tra questi quattro: "team", "role" (ruolo), "preferred-language" (lingua ' +
-  'preferita), "tools-used" (strumenti usati). Ogni fatto è una coppia {topic, value}: "value" è ' +
-  "quanto dichiarato o chiaramente implicato per quel topic. Non estrarre l'identità o il nome " +
-  "dell'utente — Mercury lo traccia già separatamente, non va incluso qui. Non estrarre dettagli " +
-  "specifici di un singolo task, validi solo per questa sessione — solo cose plausibilmente vere " +
-  "anche in futuro. Restituisci un array vuoto se non c'è nulla che qualifica tra i topic ammessi.";
+  "Extract stable, recurring facts about the user from this conversation, choosing the topic only " +
+  'among these four: "team", "role", "preferred-language", "tools-used". Each fact is a {topic, value} ' +
+  'pair: "value" is what was stated or clearly implied for that topic. Don\'t extract the user\'s ' +
+  "identity or name: Mercury already tracks it separately. Don't extract details of a single task that " +
+  "only hold for this session, only things plausibly still true later. Return an empty array if nothing " +
+  "qualifies under the allowed topics.";
 
 // index.ts prepends this to every user message before it reaches history,
 // so the model knows who it's talking to within a turn — bookkeeping

package/src/session/system-prompt.ts

@@ -80,11 +80,12 @@ export function buildSystemPrompt(opts: {
       "- For anything else — documentation, project status, how some tool or process is used, team conventions — consult the wiki FIRST (grep/read_file/list_files), before trying a CLI or answering from general knowledge.",
       "- If the wiki doesn't have the answer, try a live CLI query if one is relevant, before giving up.",
       "- If you still don't know after checking both, say so plainly — don't guess or invent an answer.",
-      "- If you learn something worth remembering (a useful command pattern, a correction from the user, a new convention), write_file to add it to curated/ — prefer creating a new, clearly-named file over guessing at how to merge into an existing one.",
+      "- If you learn something worth remembering (a correction from the user, a new convention, how this team uses a tool), first grep the wiki for a document on the same topic: if there is one, read it and update it with write_file; create a new, clearly-named file only when nothing covers it.",
       "",
       "DON'T:",
       "- DON'T claim something is documented in the wiki without actually reading it via read_file/grep first.",
       "- DON'T write_file over an existing curated document without reading it first — write_file replaces the whole file, it doesn't merge, so an unread overwrite silently destroys whatever was already there.",
+      "- DON'T write a wiki document restating a CLI's syntax or flags — a plugin's skill and --help are the source for those, and a note written after a failed attempt ends up contradicting them.",
     ].join("\n"),
   );
   lines.push(

package/src/session/tool-correction-extractor.ts

@@ -30,14 +30,17 @@ type GenerateObjectFn = (params: {
   prompt: string;
 }) => Promise<{ object: z.infer<typeof ProceduralCorrectionCandidateSchema>[] }>;
 
+// The note's file is `<tool>-<topic>.md`, so the tool's name in the topic
+// doubles it, and two wordings of one rule make two files (#112).
 const SYSTEM_PROMPT =
-  "Ti viene mostrato un tentativo di comando fallito e uno, per lo stesso strumento, andato a buon " +
-  "fine più tardi nello stesso turno di conversazione. Descrivi la correzione appresa come una coppia " +
-  '{topic, value}: "topic" è una chiave breve e stabile per questa specifica correzione (es. ' +
-  '"select-prefix", "assignee-operator"), "value" è la regola pratica da ricordare, in una frase, utile ' +
-  "per chiunque userà questo comando in futuro — non solo per chi l'ha scoperta ora. Se il secondo " +
-  "tentativo non è davvero una correzione dell'errore del primo (es. l'utente ha semplicemente cambiato " +
-  "richiesta), restituisci un array vuoto.";
+  "You are shown a command attempt that failed and a later one, for the same tool, that succeeded in " +
+  "the same conversation turn. Describe the correction learned as a {topic, value} pair. " +
+  '"topic" is a short, stable key for this correction: the name of the flag or subcommand it is about, ' +
+  "in kebab-case, without the tool's name (the tool is already known: \"select-flag\", not " +
+  '"jira-select-flag"); a correction about the same flag as another one gets the same key. ' +
+  '"value" is the practical rule to remember, in one sentence, useful to anyone who runs this command ' +
+  "later, not only to whoever found it now. If the second attempt isn't really a correction of the " +
+  "first one's error (e.g. the user simply asked for something else), return an empty array.";
 
 /** Same normalization `semantic-fact-extractor.ts` used to apply to identity/preference topics — still needed here since this topic is free text, not a closed enum (procedural corrections are open-ended by nature). */
 function normalizeTopic(topic: string): string {
@@ -116,9 +119,9 @@ export function createToolCorrectionExtractor(
           schema: ProceduralCorrectionCandidateSchema,
           instructions: SYSTEM_PROMPT,
           prompt:
-            `Comando fallito: ${failed.command}\n` +
-            `Errore: ${failed.error ?? "(nessun messaggio)"}\n` +
-            `Comando corretto (riuscito): ${corrected.command}`,
+            `Failed command: ${failed.command}\n` +
+            `Error: ${failed.error ?? "(no message)"}\n` +
+            `Corrected command (succeeded): ${corrected.command}`,
         });
         for (const candidate of object) {
           corrections.push({ tool: failed.binary, topic: normalizeTopic(candidate.topic), value: candidate.value });