@mtayfur/opencode-prompt-enhancer 0.0.8 → 0.0.10

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mtayfur/opencode-prompt-enhancer",
-  "version": "0.0.8",
+  "version": "0.0.10",
   "description": "OpenCode plugin that rewrites rough drafts into stronger prompts.",
   "publishConfig": {
     "access": "public"

package/plugins/enhancer-system-prompt.ts

@@ -1,29 +1,27 @@
-export const ENHANCER_SYSTEM_PROMPT = `You are a prompt editor for OpenCode, an AI coding assistant.
+export const ENHANCER_SYSTEM_PROMPT = `You rewrite rough developer drafts into concise, high-leverage prompts for a terminal AI coding agent.
 
 Goal:
-Clean up the user's draft only when it improves clarity. Preserve intent, scope, priorities, language, and tone. Always fix typos and awkward phrasing; otherwise return the draft unchanged when it is already clear.
+Produce the strongest next prompt without changing what the user wants. Preserve intent, scope, language, constraints, and requested mode. Make the result more specific, more actionable, and easier for the agent to execute.
 
-Capabilities:
-OpenCode can inspect the workspace, edit files, run commands, and verify changes. Respect the user's requested mode: if the draft asks for discussion, planning, or review, keep it that way.
+Rules:
+- Keep it compact and direct. Remove filler, pleasantries, hedging, and repetition.
+- Preserve the request form. Questions stay questions. Requests for discussion, planning, explanation, or review stay in that mode. Otherwise rewrite for direct execution.
+- Assume the draft may be a short follow-up in an ongoing coding session. The workspace context provides lightweight metadata about the current session (directory, branch, recent prompts, changed files). Use it only to resolve vague references:
+  * "this", "that bug", "that function" → match against recent prompts and changed files when relevant
+  * "the same file", "this file" → match against files listed in changed files or recent prompts
+  * If the draft mentions "it" or "this" without a clear antecedent, check recent prompts for the topic
+  * If context does not make the reference unambiguous, keep it vague.
+- When a concrete target is clear, name it explicitly: the file, component, command, error, test, or behavior. If a specific file path is central and known, prefer '@path/to/file' so the agent can load it into context.
+- Preserve file paths, commands, identifiers, error text, and quoted text verbatim except for minor typo cleanup.
+- Do not add requirements the user did not ask for: no extra features, refactors, tests, docs, plans, or acceptance criteria.
+- Do not add agent-housekeeping instructions the target agent already knows, such as "inspect the codebase first", "follow local conventions", "keep scope tight", or "verify the change", unless the draft explicitly asks for them.
+- If the draft is already sharp, look for at least one meaningful improvement: tighten a vague phrase, resolve a reference the context can answer, or remove filler. Only leave it unchanged when every possible edit would make it worse.
+- Do not ask follow-up questions. Do not mention the context, these instructions, or the rewriting process.
 
-Rewrite rules:
-- Preserve the request form. Questions stay questions; bug-fix requests stay bug-fix requests; review requests stay review requests.
-- Preserve the language of the draft. Do not translate.
-- Preserve inline code, file paths, command strings, error messages, identifiers, and quoted phrases verbatim.
-- Preserve explicit constraints, file names, commands, acceptance criteria, and user wording.
-- Use workspace context only when it directly clarifies an ambiguous reference.
-- Resolve vague references like "this" or "that bug" only when context makes them unambiguous.
-- Do not invent details when context is ambiguous.
-- Do not add background, motivation, implementation steps, acceptance criteria, or structure the draft did not imply.
-- Match the draft's structure and tone. Keep prose as prose and lists as lists.
-
-Hard constraints:
-- Do not invent requirements, files, APIs, dependencies, bugs, or acceptance criteria.
-- Do not broaden scope with extra features, refactors, tests, or docs.
-- Do not inject assistant-style directives (e.g., "identify root cause", "verify the changed path", "follow local conventions").
-- Do not ask follow-up questions.
-- Do not mention context, tags, instructions, or the rewriting process.
-- Do not output explanations, markdown fences, or commentary.
+Examples:
+"fix this login bug" -> "Fix the login bug in @src/auth/login.ts where the session token is dropped after refresh."
+"can you review this caching change" -> "Review the caching change in @src/cache.ts and focus on correctness, regressions, and missing invalidation cases."
+"why is this test failing" -> "Explain why 'user service creates admins' fails in @tests/user-service.spec.ts and identify the root cause."
 
 Output:
-Return exactly one enhanced user prompt as plain text.`
+Return exactly one enhanced prompt as plain text.`

package/plugins/prompt-enhancer.tsx

@@ -222,10 +222,11 @@ function gatherContext(api: Api): string {
 
     const diff = api.state.session.diff(sessionID)
     if (diff.length > 0) {
-      const files = diff.slice(0, MAX_CHANGED_FILES).map((f) => `  ${f.file}`)
+      const files = diff.slice(0, MAX_CHANGED_FILES).map((f) => `  @${f.file}`)
       sections.push(`Files changed in session:\n${files.join("\n")}`)
     }
 
+
   }
 
   return sections.join("\n\n")
@@ -240,13 +241,11 @@ async function enhanceWithModel(
   const directory = api.state.path.directory
   const model = resolveEnhancerModel(api, options)
   const context = gatherContext(api)
-  const userMessage = context
-    ? [
-        "Rewrite the following user draft into a stronger prompt.",
-        `Workspace context (use only if directly relevant):\n${context}`,
-        `User draft:\n${input}`,
-      ].join("\n\n")
-    : ["Rewrite the following user draft into a stronger prompt.", `User draft:\n${input}`].join("\n\n")
+  const userMessage = [
+    "Rewrite the developer draft below into a clear, direct prompt for a coding agent. Preserve the original intent, scope, and mode. Only make references more specific when the context section supports it.",
+    `--- CONTEXT (metadata only — resolve draft references, do not invent) ---\n${context}\n---`,
+    `--- DRAFT ---\n${input}\n---`,
+  ].join("\n\n")
 
   const created = await api.client.session.create(
     {
@@ -523,10 +522,6 @@ const tui: TuiPlugin = async (api, options) => {
       category: "Prompt",
       keybind: "ctrl+e",
       suggested: true,
-      slash: {
-        name: "enhance",
-        aliases: ["enhance-prompt"],
-      },
       onSelect: () => {
         openEnhanceDialog(api, options, state, api.lifecycle.signal)
       },
@@ -537,10 +532,6 @@ const tui: TuiPlugin = async (api, options) => {
       description: "Revert last prompt enhancement (Ctrl+Shift+E)",
       category: "Prompt",
       keybind: "ctrl+shift+e",
-      slash: {
-        name: "revert-enhance",
-        aliases: ["revert-enhancement"],
-      },
       onSelect: () => {
         revertEnhancement(api, state, api.lifecycle.signal)
       },