feat(prompts): work-first envelope with spawn/followup tool calls at top and bottom

2026-10-05 05:43:26 +00:00
parent dad33dcaa3
commit 74d093d259
3 changed files with 115 additions and 26 deletions
@@ -528,28 +528,10 @@ def main():
        "step_n": step_n,
    })

-    # Determine thread url / context for the runtime envelope
-    target_clean = str(target)
-    thread_url = f"https://box.muse-dev.online/thread/{target_clean}" if re.fullmatch(r"[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}", target_clean.lower()) else f"https://box.muse-dev.online/sidechat/{target_clean}"
-
-    # Build active tool calling envelope for https://box.muse-dev.online/
-    tool_envelope = (
-        f"\n\n--- NetVM Runtime & Tools ({agent} on {target_clean}) ---\n"
-        f"Thread Console: {thread_url}\n"
-        f"Available Tools via [TOOL <op> <json_args>] or curl to https://box.muse-dev.online/exec:\n"
-        f"  • Status: [TOOL health.check {{}}]\n"
-        f"  • Service: [TOOL service.status {{\"unit\": \"board.service\"}}]\n"
-        f"  • Follow-up: [TOOL followup.create {{\"in_m\": 10, \"prompt\": \"check progress\"}}]\n"
-        f"  • Swarm: [TOOL swarm.spawn {{\"count\": 2, \"task\": \"parallel work\"}}]\n"
-        f"Direct Curl (Agent Bearer Auth):\n"
-        f"  curl -sk -X POST https://box.muse-dev.online/exec -H \"Authorization: Bearer <token>\" "
-        f"-H \"Content-Type: application/json\" -d '{{\"op\":\"<op>\",\"args\":{{...}}}}'\n\n"
-        f"When finished, end your response with:\n"
-        f"[RESULT {job_id}] <OK|FAILED>: <summary of output>"
-    )
-
-    if "[RESULT" not in rendered:
-        rendered = rendered.rstrip() + tool_envelope
+    # Work-first envelope: executable swarm.spawn/followup.create at TOP and BOTTOM
+    # (see bin/prompt_envelope.py). Always applied, even if the template has its own [RESULT.
+    import prompt_envelope
+    rendered = prompt_envelope.wrap(job_name, job_id, agent, target, rendered)

    # Format as JOB DM
    dm_message = f"[JOB {job_id}] {rendered}"
@@ -0,0 +1,91 @@
+#!/usr/bin/env python3
+"""prompt_envelope.py — wraps every dispatched job prompt so agents DO work.
+
+Why: agents treated job DMs as informational and replied with prose; swarms and
+timers never got created. The envelope puts an executable subagent-spawn tool
+call at the TOP (before the task) and repeats it at the BOTTOM (after the task),
+with a work-first mandate and explicit tool/curl surface in between.
+
+Profiles tune the spawn recipe per job family; add new ones to PROFILES.
+"""
+import json
+import re
+
+BOX_EXEC = "https://box.muse-dev.online/exec"
+UUID_RE = re.compile(r"[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}")
+
+# profile -> (spawn count, followup minutes, spawn task hint)
+PROFILES = {
+    "pulse":  (2, 30, "execute your scope's standing work: set timers, verify them, report per-slot verdicts"),
+    "health": (2, 60, "run the prescribed health gates read-only and report per-gate verdicts"),
+    "swarm":  (2, 20, "sweep swarms and sidechats read-only and report stale or unarchived items"),
+    "work":   (2, 15, "split the task above into two independent halves and complete each"),
+}
+
+
+def pick_profile(job_name):
+    n = job_name or ""
+    if n.startswith("autonomy-pulse"):
+        return "pulse"
+    if "health" in n:
+        return "health"
+    if "swarm" in n:
+        return "swarm"
+    return "work"
+
+
+def _tool(op, args):
+    # no ']' inside the JSON: the harvester's [TOOL ...] regex stops at the first one
+    return "[TOOL %s %s]" % (op, json.dumps(args, separators=(", ", ": ")))
+
+
+def spawn_call(job_id, job_name, profile):
+    count, _, hint = PROFILES[profile]
+    # no brackets in the task text: the harvester's [TOOL ...] regex stops at the first ']'
+    task = "Subagent for job %s (%s): %s." % (job_id, job_name, hint)
+    return _tool("swarm.spawn", {"count": count, "task": task[:900], "label": (job_name or "job")[:60]})
+
+
+def followup_call(job_id, job_name, profile, agent):
+    _, mins, _ = PROFILES[profile]
+    return _tool("followup.create", {
+        "agent": agent, "in_m": mins,
+        "prompt": "Continue job %s (%s): check swarm results with swarm.list then act or close." % (job_id, job_name)})
+
+
+def thread_url(target):
+    t = str(target)
+    return "https://box.muse-dev.online/%s/%s" % ("thread" if UUID_RE.fullmatch(t.lower()) else "sidechat", t)
+
+
+def wrap(job_name, job_id, agent, target, rendered):
+    """Return the full DM body (without the [JOB id] prefix)."""
+    profile = pick_profile(job_name)
+    spawn = spawn_call(job_id, job_name, profile)
+    follow = followup_call(job_id, job_name, profile, agent)
+    top = (
+        "WORK ORDER - ACTION REQUIRED, NOT INFORMATIONAL.\n"
+        "Your FIRST output must be tool calls, not prose. Start now:\n"
+        "  1) %s\n"
+        "  2) %s\n"
+        "Each tool call returns a result message; keep going until the task is done.\n"
+        "---- TASK ----\n" % (spawn, follow)
+    )
+    has_result = "[RESULT" in rendered
+    bottom = (
+        "\n---- END TASK ----\n"
+        "REMINDER (importance: critical): work is only real if a tool call ran. "
+        "If you have not yet spawned subagents or set a follow-up timer, do it now:\n"
+        "  %s\n"
+        "  %s\n"
+        "Other tools: [TOOL health.check {}] [TOOL swarm.list {}] [TOOL swarm.status {\"swarm_id\": \"<id>\"}] "
+        "[TOOL swarm.results {\"swarm_id\": \"<id>\"}] [TOOL quality.check {}]\n"
+        "Direct: curl -sk -X POST %s -H \"Authorization: Bearer <token>\" "
+        "-H \"Content-Type: application/json\" -d '{\"op\":\"<op>\",\"args\":{...}}'\n"
+        "Thread: %s\n" % (spawn, follow, BOX_EXEC, thread_url(target))
+    )
+    if not has_result:
+        bottom += "Close with: [RESULT %s] <OK|FAILED>: <summary of tool outputs>\n" % job_id
+    else:
+        bottom += "Close using the [RESULT ...] line specified in the task, summarising tool outputs.\n"
+    return top + rendered.strip() + "\n" + bottom
@@ -163,7 +163,16 @@ def parse_tool_calls(text):
        except Exception:
            pass

-    return calls
+    # De-duplicate identical calls within one message (envelope repeats the spawn
+    # call at top and bottom; an agent echoing both must not run it twice).
+    seen, unique = set(), []
+    for op, args in calls:
+        key = (op, json.dumps(args, sort_keys=True, default=str))
+        if key in seen:
+            continue
+        seen.add(key)
+        unique.append((op, args))
+    return unique


 def execute_agent_tool(agent, op, args, timeout=30):
@@ -1065,12 +1074,19 @@ def maybe_nudge_untagged_sidechat(agent, thread_id, thread_name, mid, text, dry_
    # 1st turn: Strict rejection and tool invocation prompt
    if n_count == 0:
        thread_url = f"https://box.muse-dev.online/thread/{thread_id}"
+        try:
+            import prompt_envelope
+            _spawn = prompt_envelope.spawn_call(
+                matching_job_id, matching_job_id, prompt_envelope.pick_profile(matching_job_id))
+        except Exception:
+            _spawn = '[TOOL swarm.spawn {"count": 2, "task": "continue the job work"}]'
        nudge_msg = (
+            f"{_spawn}\n"
            f"[STRICT ENFORCEMENT: Conversational commentary is rejected. Work requires active execution.]\n"
            f"Thread Console: {thread_url}\n"
-            f"You must emit an executable tool call: [TOOL <op> <args>] or curl command against https://box.muse-dev.online/exec\n"
-            f"Example: [TOOL service.status {{\"unit\": \"board.service\"}}] or [TOOL followup.create {{\"in_m\": 10, \"prompt\": \"continue\"}}]\n"
-            f"When all operations are finished, close strictly with [RESULT {matching_job_id}] <outcome>."
+            f"Emit executable tool calls now: [TOOL <op> <args>] or curl against https://box.muse-dev.online/exec\n"
+            f"When all operations are finished, close strictly with [RESULT {matching_job_id}] <outcome>.\n"
+            f"{_spawn}"
        )
        try:
            import muse_hybrid