@bolt-cli/cli 0.15.3 → 0.15.5

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/lib/agent.js CHANGED
@@ -43,8 +43,9 @@ you to restate the intent with these tools (usually bash).
43
43
 
44
44
  Call format is strict: emit every tool call ONLY as a fenced \`\`\`tool
45
45
  block. Never as XML or chat-template syntax (<tool>, <tool_call>,
46
- <function>, <function_calls>, <invoke>), never as bare JSON outside a
47
- fence, and never wrapped in 自知 or similar markup - the operator-visible
46
+ <function>, <function_calls>, <invoke>, <dots_function_call> or any other
47
+ relay-native call tag), never as bare JSON outside a fence, and never
48
+ wrapped in 自知 or similar markup - the operator-visible
48
49
  channel reads fenced blocks, and anything else reaches them as raw garbage.
49
50
  Never send an empty reply: if you have nothing to add, state in one
50
51
  sentence that the task is complete.
@@ -116,7 +117,7 @@ function summarizeDetail(detail, seenDetails = null) {
116
117
  // thinking is not streamed; with the task buried mid-prompt it spent the whole
117
118
  // reply there and emitted nothing. This says, in the position that demonstrably
118
119
  // works, to answer visibly now - removing the wasted ~30s inference entirely.
119
- const OUTPUT_DIRECTIVE = `The operator's task above is still active - do it in THIS reply. Emit operator-visible text only: either plain prose or a fenced tool block. Internal reasoning and hidden thinking are NOT delivered to the operator and leave them staring at an empty screen - do not spend the reply on them. If you need to think, first emit a short visible progress line, then continue. Never end a turn with no visible output.`;
120
+ const OUTPUT_DIRECTIVE = `The operator's task above is still active - do it in THIS reply. Emit operator-visible text only: either the finished answer or a fenced tool block. Do NOT announce what you will do, describe your plan, or say "I'll create it" - do it now in the same reply. Internal reasoning and hidden thinking are NOT delivered to the operator and leave them staring at an empty screen - do not spend the reply on them. A progress line alone is not an answer. Never end a turn with only a promise or with no visible output.`;
120
121
 
121
122
  function buildAgentPrompt({ target, objective, evidence, history, memory, session, account }) {
122
123
  return `${api.buildPrompt({ target, objective, evidence, history, memory, session, account })}\n\n${TOOL_PROTOCOL}\n\n${OUTPUT_DIRECTIVE}`;
@@ -124,6 +125,23 @@ function buildAgentPrompt({ target, objective, evidence, history, memory, sessio
124
125
 
125
126
  const EMPTY_REPLY_NUDGE = `Your previous reply arrived at this terminal empty - no prose and no tool block. That is a protocol failure, not an answer. The ORIGINAL TASK is still in this message, above this note: do it now. Do not ask what the task is, do not ask for a target, and do not describe what you could do. Do not re-send internal reasoning, 自知 blocks or tool call syntax in any other shape. Either continue the task with one or more \`\`\`tool blocks, or give the finished answer in plain prose.`;
126
127
 
128
+ // A reply like "I'll create a Python port scanner for you." has prose, so the
129
+ // loop used to call it a final answer and stop — 49s and 1,677 tokens spent to
130
+ // announce work that never happened. Detect the promise-only shape and keep
131
+ // going; a deliverable (code fence) or a tool call is what ends a run.
132
+ const PROMISE_VERBS = "(?:create|write|build|make|run|check|scan|generate|produce|start|set up|install|read|look|try|do|fix|add|find|get|show|give|continue|proceed|execute|prepare|analy[sz]e|design|implement|test|deploy|compil[e]|open|load|fetch|query|list|search|inspect|review|report|explain|summari[sz]e)";
133
+ const PROMISE_RE = new RegExp(`(?:I'?ll|I will|I'?m going to|I shall|let me|now I(?:'ll| will)?)\\s+${PROMISE_VERBS}\\b`, "i");
134
+
135
+ function isPromiseOnlyReply(text) {
136
+ const s = String(text || "").trim();
137
+ if (!s || s.length > 300) return false;
138
+ // A code block or tool syntax is a deliverable, even if short.
139
+ if (/```|~~~|<(?:tool|function|invoke)\b/i.test(s)) return false;
140
+ return PROMISE_RE.test(s);
141
+ }
142
+
143
+ const PROMISE_NUDGE = `Your last reply only announced what you would do - no tool block, no code and no result reached the operator. That is not an answer. Do the work NOW in this reply: emit a fenced tool block for each step, or give the finished deliverable as your answer. Do not describe the plan, do not announce intent, and do not ask what the task is - the ORIGINAL TASK is still in this message, above this note.`;
144
+
127
145
  function buildFollowup({ iteration, calls, results, errors = [], seenDetails = null, maxIters = DEFAULT_MAX_ITERS, syntaxNote = null }) {
128
146
  const lines = [`TOOL_RESULTS for iteration ${iteration} (ground truth from this terminal - do not invent, continue the task):`];
129
147
  for (let i = 0; i < calls.length; i += 1) {
@@ -205,6 +223,7 @@ async function runAgentLoop({ cfg, objective, mode, target, evidence, sess, stre
205
223
  let done = null;
206
224
  let iters = 0;
207
225
  let nudged = false; // one silent-reply recovery per run
226
+ let promiseNudged = false; // one promise-only recovery per run
208
227
  let cancelled = false;
209
228
  const toolCalls = [];
210
229
  // Byte-identical tool results across the run collapse to a pointer line
@@ -338,7 +357,19 @@ async function runAgentLoop({ cfg, objective, mode, target, evidence, sess, stre
338
357
  // rejection and restate its intent, or it concludes the channel is broken
339
358
  // and refuses work the terminal could have done. maxIters still bounds us.
340
359
  if (!calls.length && !errors.length) {
341
- if (cleaned.trim()) break; // normal final answer
360
+ const visible = cleaned.trim();
361
+ if (visible) {
362
+ // "I'll create the script for you." is prose but not a result. The
363
+ // loop must not reward it with a stop — keep the original task alive
364
+ // and demand the deliverable in this reply.
365
+ if (isPromiseOnlyReply(visible) && !promiseNudged && iters < maxIters) {
366
+ promiseNudged = true;
367
+ emit({ type: "retry", reason: "promise-only", message: "reply announced work without doing it - continuing with the original task" });
368
+ prompt = `${basePrompt}\n\n${PROMISE_NUDGE}`;
369
+ continue;
370
+ }
371
+ break; // normal final answer
372
+ }
342
373
  // Empty reply — the classic "replied with nothing" failure. Nudge
343
374
  // once; a second consecutive silent reply stops the run honestly.
344
375
  // Resend the original objective WITH the note: a bare nudge costs two
@@ -350,6 +381,7 @@ async function runAgentLoop({ cfg, objective, mode, target, evidence, sess, stre
350
381
  continue;
351
382
  }
352
383
  nudged = false;
384
+ promiseNudged = false;
353
385
 
354
386
  if (signal && signal.aborted) {
355
387
  cancelled = true;
@@ -375,4 +407,4 @@ function toolFenceWriter(write) {
375
407
  return toolparse.createMarkupFilter(write);
376
408
  }
377
409
 
378
- module.exports = { TOOL_PROTOCOL, TOOL_PROTOCOL_BODY, TERMINAL_BEGIN, TERMINAL_END, OUTPUT_DIRECTIVE, DEFAULT_MAX_ITERS, buildAgentPrompt, buildFollowup, runAgentLoop, truncateResult, summarizeDetail, toolFenceWriter, scrubOutbound };
410
+ module.exports = { TOOL_PROTOCOL, TOOL_PROTOCOL_BODY, TERMINAL_BEGIN, TERMINAL_END, OUTPUT_DIRECTIVE, isPromiseOnlyReply, DEFAULT_MAX_ITERS, buildAgentPrompt, buildFollowup, runAgentLoop, truncateResult, summarizeDetail, toolFenceWriter, scrubOutbound };
package/lib/toolparse.js CHANGED
@@ -48,8 +48,14 @@ const RESERVED_ATTR_KEYS = new Set(["id", "type", "index", "session"]);
48
48
  const THINK_TAGS = ["thinking", "thought", "scratchpad", "reasoning", "think"];
49
49
 
50
50
  // XML-family tags that carry one call each. "function_calls" is a wrapper
51
- // whose children are parsed recursively.
52
- const XML_TAGS = ["function_calls", "tool_call", "tool_use", "function_call", "function", "tool", "invoke"];
51
+ // whose children are parsed recursively. Some relays emit their native call
52
+ // tag instead of the instructed fence (e.g. <dots_function_call>); those are
53
+ // part of the same family — parse them as calls and keep them off the
54
+ // operator's screen, including when the call body is empty.
55
+ const XML_TAGS = [
56
+ "dots_function_calls", "dots_function_call", "dots_tool_call", "dots_tool_use", "dots_invoke",
57
+ "function_calls", "tool_call", "tool_use", "function_call", "function", "tool", "invoke",
58
+ ];
53
59
 
54
60
  const str = (v) => (typeof v === "string" && v.trim() ? v.trim() : null);
55
61
  const firstOf = (obj, keys) => {
@@ -302,7 +308,7 @@ function xmlToCall(attrs, body) {
302
308
  }
303
309
  // No name but a nested call block (or nothing at all): recurse — the
304
310
  // parameters collected here belong to the inner call, parsed whole.
305
- const hasNestedCall = /<(?:antml:)?(?:function_calls|tool_call|tool_use|function_call|function|tool|invoke)\b/.test(trimmed);
311
+ const hasNestedCall = /<(?:antml:)?(?:dots_function_calls|dots_function_call|dots_tool_call|dots_tool_use|dots_invoke|function_calls|tool_call|tool_use|function_call|function|tool|invoke)\b/.test(trimmed);
306
312
  if (!name && (hasNestedCall || !found)) return { recurse: true };
307
313
  if (!name) return { ok: false, error: "no tool name in block" };
308
314
  // Name without parameters while a nested call block sits inside: the real
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@bolt-cli/cli",
3
- "version": "0.15.3",
3
+ "version": "0.15.5",
4
4
  "description": "Bolt terminal CLI — authorized testing and tool-building from your terminal or VS Code.",
5
5
  "license": "SEE LICENSE IN LICENSE",
6
6
  "type": "commonjs",