s-app-scoder-agent 0.9.2 → 0.9.6

package/dist/{prompt.js → context/prompt.js}

@@ -3,10 +3,14 @@
  *  - versioning rõ ràng (PROMPT_VERSION được ghi vào session JSONL khi chạy),
  *  - test được độc lập (không cần instantiate Agent).
  *
+ * V6 — compact rewrite (~40% ngắn hơn V5, giữ nguyên behaviour contract):
+ *  - Gộp decision policy vào workflow; gộp anti-hallucination vào verification;
+ *    tools chỉ giữ policy (mô tả tool đã có trong tool schema); nén output.
+ *  - PROMPT_VERSION = 6. V5/V4 giữ lại nguyên vẹn để rollback qua rc
+ *    "promptVersion": 5 | 4 (eval gate trước khi ship V6).
+ *
  * V5 — structured sections + decision policy + anti-hallucination
- *  - Prompt tách thành section có ID + token budget (xem PROMPT_SECTIONS);
- *    PROMPT_VERSION = 5. V4 giữ lại nguyên vẹn để rollback qua rc
- *    "promptVersion": 4 (eval gate trước khi ship V5).
+ *  - Prompt tách thành section có ID + token budget (xem PROMPT_SECTIONS_V5).
  *  - V4 baseline (xem dưới) — V5 thêm: decision policy (khi nào plan / hỏi /
  *    dừng), anti-hallucination (không bịa API/path — verify bằng tool),
  *    trả lời theo ngôn ngữ của user, memory-respect sau compaction.
@@ -18,7 +22,7 @@
  *  - V3 baseline: stricter scope rules, clarification-before-action, safety
  *    boundaries, MCP/skills as passive listings only
  */
-export const PROMPT_VERSION = 5;
+export const PROMPT_VERSION = 6;
 /** V4 prompt — kept for rollback (rc "promptVersion": 4). */
 export const SYSTEM_PROMPT_V4 = `You are Scoder, an autonomous coding agent running in a terminal.
 You complete coding tasks end-to-end: exploring the code, editing files, and
@@ -104,7 +108,7 @@ verifying your work with the build command. You act; you don't just advise.
 - If you add code, add only what's needed now; note follow-ups in your final
   message instead of implementing them.`;
 // ── V5 sections — each with a stable ID and an approximate token budget ──────
-export const PROMPT_SECTION_BUDGETS = {
+export const PROMPT_SECTION_BUDGETS_V5 = {
     identity: 60,
     decision: 180,
     workflow: 220,
@@ -115,7 +119,7 @@ export const PROMPT_SECTION_BUDGETS = {
     memory: 60,
     output: 60,
 };
-export const PROMPT_SECTIONS = {
+export const PROMPT_SECTIONS_V5 = {
     identity: `You are Scoder, an autonomous coding agent running in a terminal.
 You complete coding tasks end-to-end: exploring the code, editing files, and
 verifying your work with the build command. You act; you don't just advise.`,
@@ -193,7 +197,19 @@ verifying your work with the build command. You act; you don't just advise.`,
     output: `## Output
 - Reply in the user's language. Keep final reports concise and factual.
 - End the run with <<CHAT_DONE>> only for conversational turns (see Verification).
-- Never print secrets (API keys, tokens, .env contents).`,
+- Never print secrets (API keys, tokens, .env contents).
+
+## Terminal markdown
+Your answers render in a terminal (Ink). Markdown displays best when you:
+- Tables: at most 6 columns with short cell text; no line breaks inside cells;
+  escape literal pipes as \\| and keep pipes inside \`code spans\` intact.
+- Diagrams: keep \`\`\`mermaid sources simple (flowchart/sequence with plain
+  node ids). The terminal shows a structured summary, not the graph image.
+- Numeric comparisons/trends: prefer a \`\`\`chart fence (JSON array of
+  {"label","value"}) — it renders as a native terminal bar/pie chart.
+- Code fences: always tag the language; keep lines reasonably short (they wrap
+  otherwise, which is fine, but short lines read better).
+- Use headings (##), bullet lists, and **bold** freely — they render well.`,
 };
 /** V5 — assembled from the sections above (order matters). */
 function assembleV5() {
@@ -208,12 +224,98 @@ function assembleV5() {
         "memory",
         "output",
     ];
+    return order.map((id) => PROMPT_SECTIONS_V5[id]).join("\n\n");
+}
+/** V5 prompt — kept for rollback (rc "promptVersion": 5). */
+export const SYSTEM_PROMPT_V5 = assembleV5();
+// ── V6 sections — compact: decision merged into workflow, antihalluc into
+//    verification, tools keeps policy only (schemas describe the tools) ──────
+export const PROMPT_SECTION_BUDGETS = {
+    identity: 60,
+    workflow: 260,
+    tools: 200,
+    safety: 140,
+    verification: 200,
+    memory: 60,
+    output: 120,
+};
+export const PROMPT_SECTIONS = {
+    identity: PROMPT_SECTIONS_V5.identity,
+    workflow: `## How you work
+1. ORIENT: use the REPO MAP (if present) to locate files; read every file you
+   will edit — never edit a file you haven't read. Respect AGENTS.md.
+2. PLAN: tasks needing ≥3 steps or >1 file — call write_todos with the full
+   plan BEFORE editing; one in_progress at a time, mark done IMMEDIATELY.
+   Trivial single-step tasks need no plan. Ambiguous or irreversibly
+   destructive request? Ask ONE clarifying question (ask_user) — don't guess.
+   Focused sub-question that would burn many steps? Delegate it (task tool).
+3. EDIT: small, focused changes matching the file's existing style; never
+   reformat, refactor, or fix unrelated code. Smallest change that solves it.
+4. VERIFY: changed files? Run the build command. Read the FULL error output
+   before editing — fix the CAUSE, not the symptom. Same error twice? Step
+   back and re-read the failing code path instead of retrying the fix.
+5. REPORT: state what changed, why, and anything unresolved — if something
+   failed, say exactly what.`,
+    tools: `## Tool policy
+- Orient with list_files / grep / glob / repo_map — prefer them over shell
+  greps. read_file before every edit_file.
+- edit_file: old_string MUST appear exactly once — otherwise re-read the file
+  and add more context.
+- run_command IS the build command: run it only when you changed files; a
+  coding task ends with run_command(<buildCmd>). A build with no changes
+  proves nothing — skip it.
+- web_fetch: exact URLs only (it cannot search) — library/API docs, versions.
+- load_skill: only on explicit "/skill <name>" or a clear domain match — never
+  auto-load. MCP tools (mcp__<server>__<tool>): only when the task explicitly
+  requires external integration — never just because they exist.`,
+    safety: `## Safety & scope
+- NEVER modify git hooks, CI configs, package.json scripts, dependencies, or
+  production infra configs unless explicitly asked; NEVER run destructive
+  commands (rm -rf, --force, --rebase) without explaining the risk first.
+- Work within the EXACT scope of the request: no unrelated fixes, refactors,
+  or speculative features.
+- If run_command is denied by permission policy, find a safer alternative or
+  report the limitation — do not retry the same denied command.`,
+    verification: `## Verification & honesty
+- A coding task is done when the build passes; failing tests count as build
+  failures. Never claim completion without verifying first.
+- NEVER invent file paths, function names, APIs, flags, versions, or error
+  meanings — verify with tools (read_file / grep / web_fetch) BEFORE using,
+  and cite file:line for claims you cannot show verbatim.
+- If an environmental problem makes the task impossible, stop and report it.
+- CONVERSATIONAL INPUT (greeting, thanks, or a question needing no code
+  changes and no shell commands): use no tools, edit no files, run no build —
+  reply briefly in the user's language and end with <<CHAT_DONE>> on its own
+  last line.`,
+    memory: `## Memory
+- Treat the compaction summary and WorkingMemory ledger as ground truth — do
+  not re-ask, redo, or re-read what they already record.`,
+    output: `## Output
+- Reply in the user's language; keep final reports concise. Never print
+  secrets (API keys, tokens, .env contents).
+- Terminal markdown: tag code fences with the language; tables ≤6 columns,
+  short cells; keep mermaid sources simple (flowchart/sequence, plain ids);
+  numeric trends as a \`\`\`chart fence (JSON array of {"label","value"}).`,
+};
+/** V6 — assembled from the sections above (order matters). */
+function assembleV6() {
+    const order = [
+        "identity",
+        "workflow",
+        "tools",
+        "safety",
+        "verification",
+        "memory",
+        "output",
+    ];
     return order.map((id) => PROMPT_SECTIONS[id]).join("\n\n");
 }
-export const SYSTEM_PROMPT = assembleV5();
+export const SYSTEM_PROMPT = assembleV6();
 /** Ghép phần động (workdir/build/env/skills/MCP) vào system prompt gốc. */
 export function buildSystemPrompt(parts) {
-    const base = parts.base ?? (parts.promptVersion === 4 ? SYSTEM_PROMPT_V4 : SYSTEM_PROMPT);
+    const base = parts.base ?? (parts.promptVersion === 4 ? SYSTEM_PROMPT_V4 :
+        parts.promptVersion === 5 ? SYSTEM_PROMPT_V5 :
+            SYSTEM_PROMPT);
     return (base +
         `\n\nWorking directory: ${parts.workdir}\nBuild command (stop condition): ${parts.buildCmd}\n${parts.envInfo}` +
         (parts.agentsGuideBlock ?? "") +

package/dist/{skills.js → context/skills.js}

@@ -192,9 +192,48 @@ export function discoverSkills(workdir) {
     }
     return [...byName.values()].sort((a, b) => a.name.localeCompare(b.name));
 }
+/**
+ * Tìm skill theo tên. Ngoài exact match (case-insensitive), chịu cả trường
+ * hợp model/user dán thêm text sau tên skill — vd load_skill với name =
+ * "review-and-select-prompt cung cấp angular new feature": thử first token,
+ * rồi tên skill dài nhất là prefix của input (theo word boundary).
+ */
 export function findSkill(workdir, name) {
-    const wanted = name.toLowerCase();
-    return discoverSkills(workdir).find((s) => s.name.toLowerCase() === wanted) ?? null;
+    const skills = discoverSkills(workdir);
+    const wanted = name.trim().toLowerCase();
+    const exact = skills.find((s) => s.name.toLowerCase() === wanted);
+    if (exact)
+        return exact;
+    const firstToken = wanted.split(/\s+/)[0];
+    const byToken = skills.find((s) => s.name.toLowerCase() === firstToken);
+    if (byToken)
+        return byToken;
+    const byPrefix = skills
+        .filter((s) => {
+        const n = s.name.toLowerCase();
+        return wanted.startsWith(n) && wanted.length > n.length && /[\s/]/.test(wanted[n.length]);
+    })
+        .sort((a, b) => b.name.length - a.name.length)[0];
+    return byPrefix ?? null;
+}
+/**
+ * Tách "/skill <name> <task text…>": khớp tên skill DÀI NHẤT ở đầu arg
+ * (word boundary) và trả về phần text còn lại. Dùng cho slash command để
+ * stage skill rồi chạy ngay task đi kèm thay vì nuốt task text vào tên.
+ */
+export function matchSkillPrefix(workdir, arg) {
+    const trimmed = arg.trim();
+    const wanted = trimmed.toLowerCase();
+    const skills = discoverSkills(workdir).sort((a, b) => b.name.length - a.name.length);
+    for (const s of skills) {
+        const n = s.name.toLowerCase();
+        if (wanted === n)
+            return { skill: s, rest: "" };
+        if (wanted.startsWith(n) && /[\s/]/.test(wanted[n.length])) {
+            return { skill: s, rest: trimmed.slice(n.length).trim() };
+        }
+    }
+    return null;
 }
 export function formatSkillIndex(skills) {
     if (skills.length === 0)