frontier-council 0.1.2__tar.gz → 0.2.0__tar.gz

{frontier_council-0.1.2 → frontier_council-0.2.0}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: frontier-council
-Version: 0.1.2
-Summary: Multi-model deliberation for important decisions. 5 frontier LLMs debate, then a judge synthesizes consensus.
+Version: 0.2.0
+Summary: Multi-model deliberation for important decisions. 4 frontier LLMs debate with rotating challenger, then Claude judges.
 Project-URL: Homepage, https://github.com/terry-li-hm/frontier-council
 Project-URL: Repository, https://github.com/terry-li-hm/frontier-council
 Project-URL: Issues, https://github.com/terry-li-hm/frontier-council/issues
@@ -19,22 +19,25 @@ Classifier: Programming Language :: Python :: 3.12
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.11
 Requires-Dist: httpx>=0.25.0
+Requires-Dist: pydantic>=2.0
+Requires-Dist: pyyaml>=6.0
 Description-Content-Type: text/markdown
 
 # Frontier Council
 
-Multi-model deliberation for important decisions. 5 frontier LLMs debate a question, then a judge synthesizes consensus.
+Multi-model deliberation for important decisions. 4 frontier LLMs debate a question, then Claude judges and synthesizes.
 
-Inspired by [Andrej Karpathy's LLM Council](https://github.com/karpathy/llm-council), with added blind phase (anti-anchoring), explicit engagement requirements, devil's advocate role, and social calibration mode.
+Inspired by [Andrej Karpathy's LLM Council](https://github.com/karpathy/llm-council), with added blind phase (anti-anchoring), explicit engagement requirements, rotating challenger role, and social calibration mode.
 
 ## Models
 
-- Claude (claude-opus-4.5)
+**Council (deliberators):**
 - GPT (gpt-5.2-pro)
 - Gemini (gemini-3-pro-preview)
 - Grok (grok-4)
 - Kimi (kimi-k2.5)
-- Judge: Claude Opus 4.5
+
+**Judge:** Claude Opus 4.5 (synthesizes + adds own perspective)
 
 ## Installation
 
@@ -99,7 +102,9 @@ All sessions are auto-saved to `~/.frontier-council/sessions/` for later review.
 | `--share` | Upload transcript to secret GitHub Gist |
 | `--social` | Enable social calibration mode (auto-detected for interview/networking) |
 | `--persona TEXT` | Context about the person asking |
-| `--advocate N` | Which speaker (1-5) should be devil's advocate (default: random) |
+| `--challenger MODEL` | Which model starts as challenger (gpt/gemini/grok/kimi). Rotates each round. |
+| `--domain DOMAIN` | Regulatory domain context (banking, healthcare, eu, fintech, bio) |
+| `--followup` | Enable interactive drill-down after judge synthesis |
 | `--quiet` | Suppress progress output |
 | `--sessions` | List recent saved sessions |
 | `--no-save` | Don't auto-save transcript to ~/.frontier-council/sessions/ |
@@ -114,9 +119,15 @@ All sessions are auto-saved to `~/.frontier-council/sessions/` for later review.
 **Deliberation Protocol:**
 1. All models see everyone's blind claims, then deliberate
 2. Each model MUST explicitly AGREE, DISAGREE, or BUILD ON previous speakers by name
-3. After each round, the system checks for consensus (4/5 agreement triggers early exit)
+3. After each round, the system checks for consensus (3/4 non-challengers agreeing triggers early exit)
 4. Judge synthesizes the full deliberation
 
+**Rotating Challenger:**
+- One model each round is assigned the "challenger" role
+- The challenger MUST argue the contrarian position and identify weaknesses in emerging consensus
+- Role rotates each round (GPT R1 → Gemini R2 → Grok R3 → Kimi R4...) to ensure sustained disagreement
+- Challenger is excluded from consensus detection (forced disagreement shouldn't block early exit)
+
 **Anonymous Deliberation:**
 - Models see each other as "Speaker 1", "Speaker 2", etc. during deliberation
 - Prevents models from playing favorites based on vendor reputation

{frontier_council-0.1.2 → frontier_council-0.2.0}/README.md

@@ -1,17 +1,18 @@
 # Frontier Council
 
-Multi-model deliberation for important decisions. 5 frontier LLMs debate a question, then a judge synthesizes consensus.
+Multi-model deliberation for important decisions. 4 frontier LLMs debate a question, then Claude judges and synthesizes.
 
-Inspired by [Andrej Karpathy's LLM Council](https://github.com/karpathy/llm-council), with added blind phase (anti-anchoring), explicit engagement requirements, devil's advocate role, and social calibration mode.
+Inspired by [Andrej Karpathy's LLM Council](https://github.com/karpathy/llm-council), with added blind phase (anti-anchoring), explicit engagement requirements, rotating challenger role, and social calibration mode.
 
 ## Models
 
-- Claude (claude-opus-4.5)
+**Council (deliberators):**
 - GPT (gpt-5.2-pro)
 - Gemini (gemini-3-pro-preview)
 - Grok (grok-4)
 - Kimi (kimi-k2.5)
-- Judge: Claude Opus 4.5
+
+**Judge:** Claude Opus 4.5 (synthesizes + adds own perspective)
 
 ## Installation
 
@@ -76,7 +77,9 @@ All sessions are auto-saved to `~/.frontier-council/sessions/` for later review.
 | `--share` | Upload transcript to secret GitHub Gist |
 | `--social` | Enable social calibration mode (auto-detected for interview/networking) |
 | `--persona TEXT` | Context about the person asking |
-| `--advocate N` | Which speaker (1-5) should be devil's advocate (default: random) |
+| `--challenger MODEL` | Which model starts as challenger (gpt/gemini/grok/kimi). Rotates each round. |
+| `--domain DOMAIN` | Regulatory domain context (banking, healthcare, eu, fintech, bio) |
+| `--followup` | Enable interactive drill-down after judge synthesis |
 | `--quiet` | Suppress progress output |
 | `--sessions` | List recent saved sessions |
 | `--no-save` | Don't auto-save transcript to ~/.frontier-council/sessions/ |
@@ -91,9 +94,15 @@ All sessions are auto-saved to `~/.frontier-council/sessions/` for later review.
 **Deliberation Protocol:**
 1. All models see everyone's blind claims, then deliberate
 2. Each model MUST explicitly AGREE, DISAGREE, or BUILD ON previous speakers by name
-3. After each round, the system checks for consensus (4/5 agreement triggers early exit)
+3. After each round, the system checks for consensus (3/4 non-challengers agreeing triggers early exit)
 4. Judge synthesizes the full deliberation
 
+**Rotating Challenger:**
+- One model each round is assigned the "challenger" role
+- The challenger MUST argue the contrarian position and identify weaknesses in emerging consensus
+- Role rotates each round (GPT R1 → Gemini R2 → Grok R3 → Kimi R4...) to ensure sustained disagreement
+- Challenger is excluded from consensus detection (forced disagreement shouldn't block early exit)
+
 **Anonymous Deliberation:**
 - Models see each other as "Speaker 1", "Speaker 2", etc. during deliberation
 - Prevents models from playing favorites based on vendor reputation

frontier_council-0.2.0/docs/brainstorms/2026-02-05-rotating-challenger-brainstorm.md

@@ -0,0 +1,59 @@
+# Brainstorm: Rotating Challenger for Sustained Disagreement
+
+**Date:** 2026-02-05
+**Status:** Ready for planning
+
+## What We're Building
+
+Modify the frontier-council deliberation architecture so that the challenger role rotates each round instead of only firing in Round 1. This ensures someone is always structurally incentivized to push back, preventing the premature convergence observed in current transcripts.
+
+Additionally, strengthen the challenger prompt to produce sharper disagreement.
+
+## Why This Approach
+
+### Problem Observed
+
+From transcript analysis:
+- Best insights came from genuine pushback (Kimi's "stop weaving, start haunting", Gemini's "dismantle the OpenCode delegation")
+- These contrarian takes got softened by Round 2
+- Models converge quickly because LLMs are trained to agree
+- Current challenger/advocate roles only fire in Round 1, exactly when they're least needed
+
+### Why Rotating Challenger
+
+| Alternative | Why Not |
+|-------------|---------|
+| Position Locking | Requires state tracking, feels artificial, models may defend positions they don't believe |
+| Adversarial Pairing | Major redesign, doesn't work for all question types |
+| Just fix the judge | Symptom not cause — the deliberation itself converges too fast |
+
+Rotating challenger is:
+- Minimal code change (move the `if round_num == 0` check)
+- Immediate impact on deliberation dynamics
+- Easy to measure (compare transcripts before/after)
+
+## Key Decisions
+
+1. **Rotation pattern:** Sequential through council order (Claude → GPT → Gemini → Grok → Kimi → repeat)
+2. **Prompt strengthening:** Add explicit requirements to challenger prompt:
+   - Must name one specific thing that would make the emerging consensus WRONG
+   - Must identify the weakest assumption being made
+   - Cannot use phrases like "building on" or "adding nuance"
+3. **Merge advocate and challenger:** Remove the redundant devil's advocate role. One challenger role is enough.
+4. **All rounds:** Challenger fires every round, not just Round 1
+
+## Open Questions
+
+1. Should challenger be excluded from consensus detection? (If the challenger is forced to disagree, they shouldn't count toward "4/5 agree")
+2. Should we track which model was challenger in the output metadata?
+3. Does strengthening the prompt risk making disagreement feel forced/artificial?
+
+## Success Criteria
+
+- Transcripts show sustained disagreement through Round 2+
+- Contrarian perspectives survive to judge synthesis
+- Judge explicitly notes unresolved tensions (may need separate prompt tweak)
+
+## Next Steps
+
+Run `/workflows:plan` to create implementation plan.

frontier_council-0.2.0/docs/plans/2026-02-05-feat-rotating-challenger-plan.md

@@ -0,0 +1,280 @@
+---
+title: "feat: Rotating Challenger for Sustained Disagreement"
+type: feat
+date: 2026-02-05
+brainstorm: docs/brainstorms/2026-02-05-rotating-challenger-brainstorm.md
+---
+
+# feat: Rotating Challenger for Sustained Disagreement
+
+## Overview
+
+Modify frontier-council so the challenger role rotates each round instead of only firing in Round 1. This ensures someone is always structurally incentivized to push back, preventing the premature convergence observed in transcripts.
+
+## Problem Statement
+
+From transcript analysis:
+- Best insights came from genuine pushback (Kimi's "stop weaving, start haunting", Gemini's "dismantle delegation")
+- These contrarian takes got softened by Round 2
+- Current challenger/advocate roles only apply in `round_num == 0`
+- Devil's advocate and challenger prompts are **nearly identical** — redundant complexity
+
+## Proposed Solution
+
+1. **Merge advocate and challenger** into single "challenger" role
+2. **Rotate challenger each round** — Claude R1 → GPT R2 → Gemini R3 → Grok R4 → Kimi R5 → wrap
+3. **Strengthen challenger prompt** with explicit anti-convergence requirements
+4. **Exclude challenger from consensus detection** so forced disagreement doesn't block early exit
+
+## Technical Approach
+
+### Files to Modify
+
+| File | Changes |
+|------|---------|
+| `council.py` | Remove `devils_advocate_addition`, modify role application logic, update consensus detection |
+| `cli.py` | Deprecate `--advocate` with warning, update `--challenger` semantics |
+
+### Implementation Steps
+
+#### Step 1: Merge Advocate and Challenger Prompts
+
+**File:** `council.py` lines 783-837
+
+Remove `devils_advocate_addition` (lines 783-794). Keep only `challenger_addition` and strengthen it:
+
+```python
+# council.py ~line 826
+challenger_addition = """
+
+SPECIAL ROLE: You are the CHALLENGER for this round. Your job is to argue the CONTRARIAN position.
+
+REQUIREMENTS:
+1. You MUST explicitly DISAGREE with at least one major point from the other speakers
+2. Identify the weakest assumption in the emerging consensus and attack it
+3. Name ONE specific thing that would make the consensus WRONG
+4. You CANNOT use phrases like "building on", "adding nuance", or "I largely agree"
+5. If everyone is converging too fast, that's a red flag — find the hidden complexity
+
+Even if you ultimately agree with the direction, you MUST articulate the strongest possible counter-argument.
+If you can't find real disagreement, explain why the consensus might be groupthink."""
+```
+
+#### Step 2: Rotate Challenger Each Round
+
+**File:** `council.py` lines 879-883
+
+Current code:
+```python
+if idx == advocate_idx and round_num == 0:
+    system_prompt += devils_advocate_addition
+
+if idx == challenger_idx and round_num == 0:
+    system_prompt += challenger_addition
+```
+
+Replace with:
+```python
+# Calculate rotating challenger for this round
+if challenger_idx is not None:
+    # Explicit --challenger sets starting point, then rotates
+    current_challenger = (challenger_idx + round_num) % len(council_config)
+else:
+    # Default: start with Claude (index 0), rotate through council
+    current_challenger = round_num % len(council_config)
+
+if idx == current_challenger:
+    system_prompt += challenger_addition
+```
+
+#### Step 3: Update Consensus Detection
+
+**File:** `council.py` lines 545-564
+
+Modify `detect_consensus` to accept and exclude challenger:
+
+```python
+def detect_consensus(
+    conversation: list[tuple[str, str]],
+    council_config: list,
+    current_challenger_idx: int | None = None
+) -> tuple[bool, str]:
+    """Detect if council has converged. Returns (converged, reason)."""
+    council_size = len(council_config)
+
+    if len(conversation) < council_size:
+        return False, "insufficient responses"
+
+    recent = conversation[-council_size:]
+
+    # Exclude challenger from consensus count
+    if current_challenger_idx is not None:
+        challenger_name = council_config[current_challenger_idx][0]
+        recent = [(name, text) for name, text in recent if name != challenger_name]
+
+    effective_size = len(recent)
+    threshold = effective_size - 1  # Need all-but-one non-challengers to agree
+
+    consensus_count = sum(1 for _, text in recent if "CONSENSUS:" in text.upper())
+    if consensus_count >= threshold:
+        return True, "explicit consensus signals"
+
+    agreement_phrases = ["i agree with", "i concur", "we all agree", "consensus emerging"]
+    agreement_count = sum(
+        1 for _, text in recent
+        if any(phrase in text.lower() for phrase in agreement_phrases)
+    )
+    if agreement_count >= threshold:
+        return True, "agreement language detected"
+
+    return False, "no consensus"
+```
+
+Update the call site (~line 943):
+```python
+current_challenger = (challenger_idx + round_num) % len(council_config) if challenger_idx is not None else round_num % len(council_config)
+converged, reason = detect_consensus(conversation, council_config, current_challenger)
+```
+
+#### Step 4: Deprecate --advocate Flag
+
+**File:** `cli.py` lines 100-105 and 212
+
+Add deprecation warning:
+```python
+# cli.py ~line 212
+if args.advocate:
+    print("Warning: --advocate is deprecated. Use --challenger instead.", file=sys.stderr)
+    # Map speaker number (1-5) to model name for backward compat
+    model_names = [n for n, _, _ in COUNCIL]
+    mapped_model = model_names[args.advocate - 1]
+    print(f"  Mapping --advocate {args.advocate} to --challenger {mapped_model.lower()}", file=sys.stderr)
+    if not args.challenger:
+        args.challenger = mapped_model.lower()
+```
+
+#### Step 5: Update Transcript Output
+
+Show challenger indicator in round headers:
+
+```python
+# council.py ~line 908 (in the speaker output section)
+challenger_indicator = " (challenger)" if idx == current_challenger else ""
+output_parts.append(f"### {name}{challenger_indicator}\n{response}")
+```
+
+### Function Signature Changes
+
+**`run_council`** (lines 706-723):
+- Remove `advocate_idx` parameter
+- Keep `challenger_idx` (now means "starting challenger")
+
+```python
+def run_council(
+    question: str,
+    council_config: list[tuple[str, str, tuple[str, str] | None]],
+    api_key: str,
+    google_api_key: str | None = None,
+    moonshot_api_key: str | None = None,
+    rounds: int = 1,
+    verbose: bool = True,
+    anonymous: bool = True,
+    blind: bool = True,
+    context: str | None = None,
+    social_mode: bool = False,
+    persona: str | None = None,
+    # advocate_idx removed
+    domain: str | None = None,
+    challenger_idx: int | None = None,  # Now means "starting challenger"
+    format: str = "prose",
+) -> tuple[str, list[str]]:
+```
+
+## Acceptance Criteria
+
+### Functional Requirements
+- [x] Challenger role rotates each round (R1: model 0, R2: model 1, etc.)
+- [x] `--challenger X` sets starting point, then rotates
+- [x] `--advocate` shows deprecation warning and maps to `--challenger`
+- [x] Challenger excluded from consensus detection
+- [x] Transcript shows which model is challenger each round
+- [x] Claude removed from council (judge-only)
+- [x] Judge has own voice with "Judge's Own Take" section
+
+### Non-Functional Requirements
+- [x] No breaking changes to existing scripts (deprecation, not removal)
+- [x] Tests pass for new rotation logic
+- [x] README updated with new behavior
+- [x] Tests updated for 4-model council
+
+## Success Metrics
+
+Compare transcripts before/after:
+- Sustained disagreement through Round 2+
+- Contrarian perspectives survive to judge synthesis
+- Judge notes unresolved tensions (may need separate prompt tweak)
+
+## Testing Plan
+
+### Unit Tests
+
+Add to `tests/test_utils.py`:
+
+```python
+class TestRotatingChallenger:
+    def test_challenger_rotates_default(self):
+        """Challenger rotates through council order by default."""
+        # R0: index 0, R1: index 1, R2: index 2...
+        assert get_challenger_for_round(None, 0, 5) == 0
+        assert get_challenger_for_round(None, 1, 5) == 1
+        assert get_challenger_for_round(None, 4, 5) == 4
+        assert get_challenger_for_round(None, 5, 5) == 0  # wraps
+
+    def test_challenger_rotates_from_explicit(self):
+        """Explicit --challenger sets starting point."""
+        # --challenger gemini (index 2): R0=2, R1=3, R2=4, R3=0...
+        assert get_challenger_for_round(2, 0, 5) == 2
+        assert get_challenger_for_round(2, 1, 5) == 3
+        assert get_challenger_for_round(2, 3, 5) == 0  # wraps
+
+class TestConsensusWithChallenger:
+    def test_consensus_excludes_challenger(self):
+        """Challenger's agreement doesn't count toward consensus."""
+        conversation = [
+            ("Claude", "CONSENSUS: I agree"),
+            ("GPT", "CONSENSUS: agreed"),
+            ("Gemini", "CONSENSUS: yes"),  # challenger
+            ("Grok", "CONSENSUS: agreed"),
+            ("Kimi", "different view"),
+        ]
+        council_config = [("Claude",), ("GPT",), ("Gemini",), ("Grok",), ("Kimi",)]
+        # Gemini (index 2) is challenger, excluded
+        # 3 of 4 non-challengers agree = consensus
+        converged, _ = detect_consensus(conversation, council_config, 2)
+        assert converged
+```
+
+### Integration Test
+
+```bash
+# Run with 3 rounds, verify rotation in transcript
+frontier-council "test question" --rounds 3 --output /tmp/test.md
+grep -E "### .+ \(challenger\)" /tmp/test.md
+# Should show 3 different models as challenger
+```
+
+## Risks and Mitigations
+
+| Risk | Mitigation |
+|------|------------|
+| Forced disagreement feels artificial | Prompt says "even if you ultimately agree" — models can agree after challenging |
+| Breaking scripts using `--advocate` | Deprecation warning + automatic mapping, not hard removal |
+| Consensus detection edge cases | Thorough unit tests for threshold math |
+
+## References
+
+- Brainstorm: `docs/brainstorms/2026-02-05-rotating-challenger-brainstorm.md`
+- Current challenger impl: `council.py:826-837`
+- Current advocate impl: `council.py:783-794` (to be removed)
+- Consensus detection: `council.py:545-564`
+- Deliberation loop: `council.py:839-947`

{frontier_council-0.1.2 → frontier_council-0.2.0}/frontier_council/__init__.py

@@ -1,6 +1,6 @@
 """Frontier Council - Multi-model deliberation for important decisions."""
 
-__version__ = "0.1.2"
+__version__ = "0.1.3"
 
 from .council import (
     run_council,

{frontier_council-0.1.2 → frontier_council-0.2.0}/frontier_council/cli.py

@@ -29,6 +29,8 @@ from .council import (
     COUNCIL,
     detect_social_context,
     run_council,
+    DOMAIN_CONTEXTS,
+    run_followup_discussion,
 )
 
 
@@ -42,9 +44,10 @@ Examples:
   frontier-council "What questions should I ask?" --social
   frontier-council "Career decision" --persona "builder who hates process work"
   frontier-council "Architecture choice" --rounds 3 --output transcript.md
+  frontier-council "Decision" --domain banking --followup --output counsel.md
         """,
     )
-    parser.add_argument("question", help="The question for the council to deliberate")
+    parser.add_argument("question", nargs="?", help="The question for the council to deliberate")
     parser.add_argument(
         "--rounds",
         type=int,
@@ -74,6 +77,12 @@ Examples:
         "--context", "-c",
         help="Context hint for the judge (e.g., 'architecture decision', 'ethics question')",
     )
+    parser.add_argument(
+        "--format", "-f",
+        choices=["json", "yaml", "prose"],
+        default="prose",
+        help="Output format: json (machine-parseable), yaml (structured), prose (default)",
+    )
     parser.add_argument(
         "--share",
         action="store_true",
@@ -92,7 +101,20 @@ Examples:
         "--advocate",
         type=int,
         choices=[1, 2, 3, 4, 5],
-        help="Which speaker (1-5) should be devil's advocate (default: random)",
+        help="DEPRECATED: Use --challenger instead. Maps to --challenger by model name.",
+    )
+    parser.add_argument(
+        "--domain",
+        help="Regulatory domain context (banking, healthcare, eu, fintech, bio)",
+    )
+    parser.add_argument(
+        "--challenger",
+        help="Which model should argue contrarian (claude, gpt, gemini, grok, kimi). Default: claude",
+    )
+    parser.add_argument(
+        "--followup",
+        action="store_true",
+        help="Enable followup mode to drill into specific points after judge synthesis",
     )
     parser.add_argument(
         "--no-save",
@@ -121,12 +143,44 @@ Examples:
                 print(f"\n  ... and {len(sessions) - 20} more")
         sys.exit(0)
 
+    # Require question for normal operation
+    if not args.question:
+        parser.error("the following arguments are required: question")
+
     # Auto-detect social context if not explicitly set
     social_mode = args.social or detect_social_context(args.question)
     if social_mode and not args.social and not args.quiet:
         print("(Auto-detected social context - enabling social calibration mode)")
         print()
 
+    # Validate and resolve domain
+    domain_context = None
+    if args.domain:
+        if args.domain.lower() not in DOMAIN_CONTEXTS:
+            print(f"Error: Unknown domain '{args.domain}'. Valid domains: {', '.join(DOMAIN_CONTEXTS.keys())}", file=sys.stderr)
+            sys.exit(1)
+        domain_context = args.domain.lower()
+
+    # Resolve challenger model
+    challenger_idx = None
+    if args.challenger:
+        challenger_lower = args.challenger.lower()
+        model_name_map = {n.lower(): i for i, (n, _, _) in enumerate(COUNCIL)}
+        if challenger_lower not in model_name_map:
+            print(f"Error: Unknown model '{args.challenger}'. Valid models: {', '.join(n for n, _, _ in COUNCIL)}", file=sys.stderr)
+            sys.exit(1)
+        challenger_idx = model_name_map[challenger_lower]
+    elif args.domain:
+        # Default challenger: GPT (index 0) when domain is set
+        # Reasoning: Grok is naturally contrarian anyway, so assigning GPT as challenger
+        # gives you two sources of pushback
+        challenger_idx = 0
+
+    if not args.quiet and challenger_idx is not None:
+        challenger_name = COUNCIL[challenger_idx][0]
+        print(f"(Contrainian challenger: {challenger_name})")
+        print()
+
     # Get API keys
     api_key = os.environ.get("OPENROUTER_API_KEY")
     if not api_key:
@@ -155,14 +209,28 @@ Examples:
         print()
 
     try:
-        advocate_idx = (args.advocate - 1) if args.advocate else random.randint(0, len(COUNCIL) - 1)
+        # Handle deprecated --advocate flag
+        if args.advocate:
+            print("Warning: --advocate is deprecated. Use --challenger instead.", file=sys.stderr)
+            model_names = [n for n, _, _ in COUNCIL]
+            mapped_model = model_names[args.advocate - 1].lower()
+            print(f"  Mapping --advocate {args.advocate} to --challenger {mapped_model}", file=sys.stderr)
+            if not args.challenger:
+                args.challenger = mapped_model
+            # Re-resolve challenger_idx after mapping
+            challenger_lower = args.challenger.lower()
+            model_name_map = {n.lower(): i for i, (n, _, _) in enumerate(COUNCIL)}
+            challenger_idx = model_name_map.get(challenger_lower, 0)
 
         if not args.quiet and args.persona:
             print(f"(Persona context: {args.persona})")
             print()
+
+        # Show starting challenger (now rotates each round)
         if not args.quiet:
-            advocate_name = COUNCIL[advocate_idx][0]
-            print(f"(Devil's advocate: {advocate_name})")
+            starting_challenger_idx = challenger_idx if challenger_idx is not None else 0
+            starting_challenger_name = COUNCIL[starting_challenger_idx][0]
+            print(f"(Starting challenger: {starting_challenger_name}, rotates each round)")
             print()
 
         transcript, failed_models = run_council(
@@ -178,9 +246,32 @@ Examples:
             context=args.context,
             social_mode=social_mode,
             persona=args.persona,
-            advocate_idx=advocate_idx,
+            domain=domain_context,
+            challenger_idx=challenger_idx,
+            format=args.format,
         )
 
+        # Followup mode
+        followup_transcript = ""
+        if args.followup and not args.quiet:
+            print("\n" + "=" * 60)
+            print("Enter topic to explore further (or 'done'): ", end="", flush=True)
+            topic = input().strip()
+            
+            if topic and topic.lower() != "done":
+                domain_ctxt = DOMAIN_CONTEXTS.get(domain_context, "") if domain_context else ""
+                followup_transcript = run_followup_discussion(
+                    question=args.question,
+                    topic=topic,
+                    council_config=COUNCIL,
+                    api_key=api_key,
+                    domain_context=domain_ctxt,
+                    social_mode=social_mode,
+                    persona=args.persona,
+                    verbose=not args.quiet,
+                )
+                transcript += "\n\n" + followup_transcript
+
         # Print failure summary
         if failed_models and not args.quiet:
             print()