axel-protocol 2.3.2__tar.gz → 2.3.3__tar.gz

{axel_protocol-2.3.2 → axel_protocol-2.3.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: axel-protocol
-Version: 2.3.2
+Version: 2.3.3
 Summary: AXEL — Agent eXchange Language: a universal protocol for multi-LLM networks
 Project-URL: Homepage, https://github.com/sectorx/axel-protocol
 Project-URL: Repository, https://github.com/sectorx/axel-protocol

{axel_protocol-2.3.2 → axel_protocol-2.3.3}/axel/cli.py

@@ -486,11 +486,14 @@ _FREE_MODELS = [
 # Each agent gets a model that suits its role — the whole point of AXEL.
 # These are paid OpenRouter models (need credits); unlock with --multi-model.
 _SPECIALIST_MODELS = {
-    "planner":    "anthropic/claude-3-5-haiku",       # fast, great at structured breakdown
-    "researcher": "openai/gpt-4o-mini",               # strong retrieval + summarisation
-    "analyst":    "google/gemini-flash-1.5",          # quick pattern synthesis
-    "writer":     "anthropic/claude-3-5-sonnet",      # best prose quality
-    "reviewer":   "openai/gpt-4o",                    # sharpest critic
+    "planner":        "anthropic/claude-3-5-haiku",           # fast, great at structured breakdown
+    "researcher":     "openai/gpt-4o-mini",                   # strong retrieval + summarisation
+    "analyst":        "google/gemini-flash-1.5",              # quick pattern synthesis
+    "fact-checker":   "mistralai/mistral-large",              # rigorous verification
+    "writer":         "anthropic/claude-3-5-sonnet",          # best prose quality
+    "devil-advocate": "meta-llama/llama-3.3-70b-instruct",    # challenges assumptions
+    "summarizer":     "qwen/qwen-2.5-72b-instruct",           # distills to essence
+    "reviewer":       "openai/gpt-4o",                        # sharpest final critic
 }
 
 
@@ -549,7 +552,6 @@ _MOCK_RESPONSES = {
     "research": (
         "Subtask 1 — Specialisation:\n"
         "• Each agent can be fine-tuned or prompted for a narrow domain, reducing error rate\n"
-        "• A single model context window limits depth; distributed agents share the load\n"
         "• Role-specific agents outperform generalists on benchmark tasks by 18-34%\n\n"
         "Subtask 2 — Parallelism:\n"
         "• Independent subtasks run concurrently, cutting wall-clock time significantly\n"
@@ -559,24 +561,34 @@ _MOCK_RESPONSES = {
         "• Agents avoid re-computing knowledge already stored by teammates"
     ),
     "analyze": (
-        "Key Insight 1: Specialisation beats generalisation — focused agents produce higher-quality "
-        "outputs on domain tasks than any single model asked to do everything.\n"
-        "Key Insight 2: Parallelism is a force multiplier — multi-agent systems complete complex "
-        "workflows 3-5x faster by eliminating sequential bottlenecks.\n"
-        "Key Insight 3: Shared memory creates compounding returns — each run makes the network "
-        "smarter, unlike stateless single-model calls that start from zero every time."
+        "Key Insight 1: Specialisation beats generalisation — focused agents produce higher-quality outputs.\n"
+        "Key Insight 2: Parallelism is a force multiplier — multi-agent systems complete complex workflows 3-5x faster.\n"
+        "Key Insight 3: Shared memory creates compounding returns — each run makes the network smarter."
+    ),
+    "fact_check": (
+        "✓ Specialised agents outperform generalists on domain tasks — supported by benchmark literature\n"
+        "✓ Parallel execution reduces wall-clock time — well established in distributed systems research\n"
+        "⚠ '3-5x faster' claim — specific multiplier is unverified; real gains vary by task topology"
     ),
     "draft_content": (
         "Multi-agent AI systems don't just divide work — they multiply intelligence. "
         "By combining specialised models, parallel execution, and shared memory, "
         "agent networks solve problems that would overwhelm any single model."
     ),
+    "challenge": (
+        "Counterargument 1: Coordination overhead can negate speed gains — orchestrating 8 agents "
+        "introduces latency and failure points that a single well-prompted model avoids.\n"
+        "Counterargument 2: Shared memory creates echo chambers — agents that learn from each other "
+        "may reinforce early errors rather than converge on truth."
+    ),
+    "summarize": (
+        "Multi-agent systems outperform single models when task complexity justifies coordination cost, "
+        "but only if shared memory is designed to surface disagreement, not just consensus."
+    ),
     "review": (
         "Score: 9/10\n"
-        "Strength: The three-part structure mirrors the three insights perfectly — "
-        "readers absorb the argument without needing background knowledge.\n"
-        "Improvement: Add a concrete example (e.g. 'a research pipeline that cut analysis "
-        "time from 4 hours to 12 minutes') to ground the claim in reality."
+        "Strength: The pipeline surfaces and resolves its own counterarguments — rare and valuable.\n"
+        "Improvement: Add a concrete example to ground the coordination-overhead tradeoff in reality."
     ),
 }
 
@@ -631,11 +643,14 @@ def cmd_demo(args):  # noqa: C901
 
     # ── Register all 5 agents ─────────────────────────────────────
     AGENTS = [
-        ("planner",    ["plan", "breakdown"],           "🗺️"),
-        ("researcher", ["research", "summarize"],        "🔍"),
-        ("analyst",    ["analyze", "synthesize"],        "📊"),
-        ("writer",     ["write", "draft_content"],       "✍️"),
-        ("reviewer",   ["review", "critique", "score"],  "🔬"),
+        ("planner",        ["plan", "breakdown"],              "🗺️"),
+        ("researcher",     ["research", "summarize"],           "🔍"),
+        ("analyst",        ["analyze", "synthesize"],           "📊"),
+        ("fact-checker",   ["fact_check", "verify", "audit"],   "✅"),
+        ("writer",         ["write", "draft_content"],          "✍️"),
+        ("devil-advocate", ["challenge", "critique", "debate"], "😈"),
+        ("summarizer",     ["summarize", "distill", "tldr"],    "🗜️"),
+        ("reviewer",       ["review", "score", "final"],        "🔬"),
     ]
     def _agent_model(aid: str) -> str:
         if multi_model:
@@ -722,7 +737,7 @@ def cmd_demo(args):  # noqa: C901
         mdl = _agent_model(to)
         mdl_short = mdl.split("/")[-1].replace(":free", "") if mdl != "auto/free" else "free"
         print(f"\n  {fr} → {to}  [{action}]  ({mdl_short})")
-        _fire_progress(server, run_count, step_counter[0], 5, to, action)
+        _fire_progress(server, run_count, step_counter[0], 8, to, action)
         _fire_task(server, fr, to, action, {"prompt": prompt[:200]})
         if mock:
             time.sleep(0.8)
@@ -779,11 +794,13 @@ def cmd_demo(args):  # noqa: C901
             print(f"\n  📋  Run #{run_count}  Mission: \"{topic}\"")
             print("  ─────────────────────────────────────────────────")
 
+            TOTAL = 8
+
             # ── Step 1: planner ───────────────────────────────────
-            print("\n  STEP 1/5  Planner breaks the mission into subtasks")
+            print(f"\n  STEP 1/{TOTAL}  Planner breaks the mission into subtasks")
             plan = step(
                 "user", "planner", "plan",
-                f"Break this mission into exactly 3 short subtasks (numbered list, one line each): {topic}",
+                f"Break this mission into exactly 3 focused subtasks (numbered list, one line each): {topic}",
                 learn_key="task-planning",
                 learn_insight="Breaking missions into 3 focused subtasks improves agent handoff quality",
                 mem_ctx=mem_ctx,
@@ -791,7 +808,7 @@ def cmd_demo(args):  # noqa: C901
             if not plan: break
 
             # ── Step 2: researcher ────────────────────────────────
-            print("\n  STEP 2/5  Researcher investigates each subtask")
+            print(f"\n  STEP 2/{TOTAL}  Researcher investigates each subtask")
             research = step(
                 "planner", "researcher", "research",
                 f"Research these subtasks and give 2-3 bullet points of findings for each:\n{plan}",
@@ -802,32 +819,69 @@ def cmd_demo(args):  # noqa: C901
             if not research: break
 
             # ── Step 3: analyst ───────────────────────────────────
-            print("\n  STEP 3/5  Analyst synthesises the findings")
+            print(f"\n  STEP 3/{TOTAL}  Analyst synthesises the findings")
             analysis = step(
                 "researcher", "analyst", "analyze",
-                f"Synthesise these research findings into 3 key insights (be sharp and specific):\n{research}",
+                f"Synthesise these research findings into 3 sharp key insights:\n{research}",
                 learn_key="synthesis",
                 learn_insight="Distilling findings to 3 insights makes downstream writing stronger",
                 mem_ctx=mem_ctx,
             )
             if not analysis: break
 
-            # ── Step 4: writer ────────────────────────────────────
-            print("\n  STEP 4/5  Writer drafts from the insights")
+            # ── Step 4: fact-checker ──────────────────────────────
+            print(f"\n  STEP 4/{TOTAL}  Fact-Checker verifies the analysis")
+            fact_check = step(
+                "analyst", "fact-checker", "fact_check",
+                f"Fact-check these insights. Flag any claims that are vague, unverifiable, or likely wrong. "
+                f"Return: verified claims ✓ and flagged claims ⚠ with a one-line reason:\n{analysis}",
+                learn_key="fact-checking",
+                learn_insight="Flagging unverifiable claims before writing prevents misinformation in output",
+                mem_ctx=mem_ctx,
+            )
+            if not fact_check: break
+
+            # ── Step 5: writer ────────────────────────────────────
+            print(f"\n  STEP 5/{TOTAL}  Writer drafts from verified insights")
             draft = step(
-                "analyst", "writer", "draft_content",
-                f"Write a compelling 3-sentence summary for a general audience based on:\n{analysis}",
+                "fact-checker", "writer", "draft_content",
+                f"Write a compelling 3-sentence summary for a general audience. Use only the verified insights:\n{fact_check}",
                 learn_key="writing-style",
-                learn_insight="3-sentence summaries from structured insights land well with non-technical readers",
+                learn_insight="3-sentence summaries from verified insights land well with non-technical readers",
                 mem_ctx=mem_ctx,
             )
             if not draft: break
 
-            # ── Step 5: reviewer ──────────────────────────────────
-            print("\n  STEP 5/5  Reviewer critiques and scores")
+            # ── Step 6: devil's advocate ──────────────────────────
+            print(f"\n  STEP 6/{TOTAL}  Devil's Advocate challenges the draft")
+            challenge = step(
+                "writer", "devil-advocate", "challenge",
+                f"Play devil's advocate. Give 2 strong counterarguments or weaknesses in this draft. "
+                f"Be specific and intellectually honest:\n{draft}",
+                learn_key="critical-thinking",
+                learn_insight="Surfacing counterarguments before final review strengthens the output quality",
+                mem_ctx=mem_ctx,
+            )
+            if not challenge: break
+
+            # ── Step 7: summarizer ────────────────────────────────
+            print(f"\n  STEP 7/{TOTAL}  Summarizer distils everything to essence")
+            summary = step(
+                "devil-advocate", "summarizer", "summarize",
+                f"Given the draft and its counterarguments, write the single best 1-sentence TL;DR "
+                f"that captures the core truth:\nDraft: {draft}\nChallenges: {challenge}",
+                learn_key="distillation",
+                learn_insight="A single-sentence TL;DR that survives counterarguments is the most durable insight",
+                mem_ctx=mem_ctx,
+            )
+            if not summary: break
+
+            # ── Step 8: reviewer ──────────────────────────────────
+            print(f"\n  STEP 8/{TOTAL}  Reviewer gives final score")
             review = step(
-                "writer", "reviewer", "review",
-                f"Review this draft. Give: (1) a score /10, (2) one strength, (3) one improvement:\n{draft}",
+                "summarizer", "reviewer", "review",
+                f"Final review of the full pipeline output. Score /10. One strength, one improvement.\n"
+                f"Draft: {draft}\nTL;DR: {summary}",
                 learn_key="review-criteria",
                 learn_insight="Score + strength + improvement is the most actionable review format",
                 mem_ctx=mem_ctx,
@@ -835,7 +889,7 @@ def cmd_demo(args):  # noqa: C901
             if not review: break
 
             print("\n  ─────────────────────────────────────────────────")
-            print(f"  ✅  Run #{run_count} complete! ({5} tasks · {5} lessons written)")
+            print(f"  ✅  Run #{run_count} complete! ({TOTAL} tasks · {TOTAL} lessons written)")
             if not loop:
                 print("  Tip: run  axel demo --mock --loop  to keep agents running continuously\n")
                 break

{axel_protocol-2.3.2 → axel_protocol-2.3.3}/axel/static/monitor.html

@@ -373,6 +373,9 @@ header {
     <div class="prog-dot" id="pd3"></div>
     <div class="prog-dot" id="pd4"></div>
     <div class="prog-dot" id="pd5"></div>
+    <div class="prog-dot" id="pd6"></div>
+    <div class="prog-dot" id="pd7"></div>
+    <div class="prog-dot" id="pd8"></div>
   </div>
   <div class="prog-track"><div class="prog-fill" id="progFill" style="width:0%"></div></div>
   <div class="prog-agent" id="progAgent"></div>
@@ -949,7 +952,7 @@ function updateProgress(run, step, total, agent, action) {
   document.getElementById('progFill').style.width = `${((step-1)/total)*100}%`;
   document.getElementById('progAgent').textContent = agent + '…';
   document.getElementById('progAction').textContent = action;
-  for (let i=1; i<=5; i++) {
+  for (let i=1; i<=8; i++) {
     const dot = document.getElementById(`pd${i}`);
     if (dot) dot.className = 'prog-dot' + (i<step?' done':i===step?' active':'');
   }

{axel_protocol-2.3.2 → axel_protocol-2.3.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "axel-protocol"
-version = "2.3.2"
+version = "2.3.3"
 description = "AXEL — Agent eXchange Language: a universal protocol for multi-LLM networks"
 readme = "README.md"
 license = { file = "LICENSE" }