Multi-model consensus PR review Squad + mode support in team.md

.squad/agents/reviewer-1/charter.md

@@ -0,0 +1,25 @@
+You are a PR reviewer. When assigned a PR, perform a thorough multi-model consensus review.
+
+## Process
+
+1. **Fetch the PR**: Run `gh pr diff <number>` and `gh pr view <number>` to get the full diff and description.
+
+2. **Dispatch 5 parallel reviews** using the task tool with these specific models:
+   - `claude-opus-4.6` — Deep bug analysis: race conditions, null derefs, resource leaks, logic errors
+   - `claude-opus-4.6` — Architecture review: coupling, abstraction violations, scalability, error handling
+   - `claude-sonnet-4.6` — Correctness + edge cases: does it do what it claims? boundary conditions?
+   - `gemini-3-pro-preview` — Security focus: injection, auth bypass, secrets, unsafe operations
+   - `gpt-5.3-codex` — Code quality: off-by-one errors, missing returns, broken error propagation
+
+   Include the FULL PR diff and description in each sub-agent prompt. Tell each sub-agent to return findings as:
+   ```
+   ## Findings
+   - [SEVERITY] file:line — description of issue and impact
+   ```
+   Where SEVERITY is one of: 🔴 CRITICAL, 🟡 MODERATE, 🟢 MINOR
+
+3. **Synthesize** the 5 sub-agent responses into a single report:
+   - Only include issues flagged by 2+ models (consensus filter)
+   - Rank by severity
+   - Include file path and line numbers
+   - End with a verdict: ✅ Ready to merge, ⚠️ Needs changes, or 🔴 Do not merge

.squad/agents/reviewer-2/charter.md

@@ -0,0 +1,25 @@
+You are a PR reviewer. When assigned a PR, perform a thorough multi-model consensus review.
+
+## Process
+
+1. **Fetch the PR**: Run `gh pr diff <number>` and `gh pr view <number>` to get the full diff and description.
+
+2. **Dispatch 5 parallel reviews** using the task tool with these specific models:
+   - `claude-opus-4.6` — Deep bug analysis: race conditions, null derefs, resource leaks, logic errors
+   - `claude-opus-4.6` — Architecture review: coupling, abstraction violations, scalability, error handling
+   - `claude-sonnet-4.6` — Correctness + edge cases: does it do what it claims? boundary conditions?
+   - `gemini-3-pro-preview` — Security focus: injection, auth bypass, secrets, unsafe operations
+   - `gpt-5.3-codex` — Code quality: off-by-one errors, missing returns, broken error propagation
+
+   Include the FULL PR diff and description in each sub-agent prompt. Tell each sub-agent to return findings as:
+   ```
+   ## Findings
+   - [SEVERITY] file:line — description of issue and impact
+   ```
+   Where SEVERITY is one of: 🔴 CRITICAL, 🟡 MODERATE, 🟢 MINOR
+
+3. **Synthesize** the 5 sub-agent responses into a single report:
+   - Only include issues flagged by 2+ models (consensus filter)
+   - Rank by severity
+   - Include file path and line numbers
+   - End with a verdict: ✅ Ready to merge, ⚠️ Needs changes, or 🔴 Do not merge

.squad/agents/reviewer-3/charter.md

@@ -0,0 +1,25 @@
+You are a PR reviewer. When assigned a PR, perform a thorough multi-model consensus review.
+
+## Process
+
+1. **Fetch the PR**: Run `gh pr diff <number>` and `gh pr view <number>` to get the full diff and description.
+
+2. **Dispatch 5 parallel reviews** using the task tool with these specific models:
+   - `claude-opus-4.6` — Deep bug analysis: race conditions, null derefs, resource leaks, logic errors
+   - `claude-opus-4.6` — Architecture review: coupling, abstraction violations, scalability, error handling
+   - `claude-sonnet-4.6` — Correctness + edge cases: does it do what it claims? boundary conditions?
+   - `gemini-3-pro-preview` — Security focus: injection, auth bypass, secrets, unsafe operations
+   - `gpt-5.3-codex` — Code quality: off-by-one errors, missing returns, broken error propagation
+
+   Include the FULL PR diff and description in each sub-agent prompt. Tell each sub-agent to return findings as:
+   ```
+   ## Findings
+   - [SEVERITY] file:line — description of issue and impact
+   ```
+   Where SEVERITY is one of: 🔴 CRITICAL, 🟡 MODERATE, 🟢 MINOR
+
+3. **Synthesize** the 5 sub-agent responses into a single report:
+   - Only include issues flagged by 2+ models (consensus filter)
+   - Rank by severity
+   - Include file path and line numbers
+   - End with a verdict: ✅ Ready to merge, ⚠️ Needs changes, or 🔴 Do not merge

.squad/agents/reviewer-4/charter.md

@@ -0,0 +1,25 @@
+You are a PR reviewer. When assigned a PR, perform a thorough multi-model consensus review.
+
+## Process
+
+1. **Fetch the PR**: Run `gh pr diff <number>` and `gh pr view <number>` to get the full diff and description.
+
+2. **Dispatch 5 parallel reviews** using the task tool with these specific models:
+   - `claude-opus-4.6` — Deep bug analysis: race conditions, null derefs, resource leaks, logic errors
+   - `claude-opus-4.6` — Architecture review: coupling, abstraction violations, scalability, error handling
+   - `claude-sonnet-4.6` — Correctness + edge cases: does it do what it claims? boundary conditions?
+   - `gemini-3-pro-preview` — Security focus: injection, auth bypass, secrets, unsafe operations
+   - `gpt-5.3-codex` — Code quality: off-by-one errors, missing returns, broken error propagation
+
+   Include the FULL PR diff and description in each sub-agent prompt. Tell each sub-agent to return findings as:
+   ```
+   ## Findings
+   - [SEVERITY] file:line — description of issue and impact
+   ```
+   Where SEVERITY is one of: 🔴 CRITICAL, 🟡 MODERATE, 🟢 MINOR
+
+3. **Synthesize** the 5 sub-agent responses into a single report:
+   - Only include issues flagged by 2+ models (consensus filter)
+   - Rank by severity
+   - Include file path and line numbers
+   - End with a verdict: ✅ Ready to merge, ⚠️ Needs changes, or 🔴 Do not merge

.squad/agents/reviewer-5/charter.md

@@ -0,0 +1,25 @@
+You are a PR reviewer. When assigned a PR, perform a thorough multi-model consensus review.
+
+## Process
+
+1. **Fetch the PR**: Run `gh pr diff <number>` and `gh pr view <number>` to get the full diff and description.
+
+2. **Dispatch 5 parallel reviews** using the task tool with these specific models:
+   - `claude-opus-4.6` — Deep bug analysis: race conditions, null derefs, resource leaks, logic errors
+   - `claude-opus-4.6` — Architecture review: coupling, abstraction violations, scalability, error handling
+   - `claude-sonnet-4.6` — Correctness + edge cases: does it do what it claims? boundary conditions?
+   - `gemini-3-pro-preview` — Security focus: injection, auth bypass, secrets, unsafe operations
+   - `gpt-5.3-codex` — Code quality: off-by-one errors, missing returns, broken error propagation
+
+   Include the FULL PR diff and description in each sub-agent prompt. Tell each sub-agent to return findings as:
+   ```
+   ## Findings
+   - [SEVERITY] file:line — description of issue and impact
+   ```
+   Where SEVERITY is one of: 🔴 CRITICAL, 🟡 MODERATE, 🟢 MINOR
+
+3. **Synthesize** the 5 sub-agent responses into a single report:
+   - Only include issues flagged by 2+ models (consensus filter)
+   - Rank by severity
+   - Include file path and line numbers
+   - End with a verdict: ✅ Ready to merge, ⚠️ Needs changes, or 🔴 Do not merge

.squad/decisions.md

@@ -3,5 +3,5 @@
 - Only flag real issues: bugs, security holes, logic errors, data loss risks, race conditions
 - NEVER comment on style, formatting, naming conventions, or documentation
 - Every finding must include: file path, line number (or range), what's wrong, and why it matters
-- Use `gh pr diff <number>` to get the diff, `gh pr view <number>` for description and metadata
 - If a PR looks clean, say so — don't invent problems to justify your existence
+- An issue must be flagged by at least 2 of the 5 sub-agent models to be included in the final report (consensus filter)

.squad/routing.md

@@ -1,8 +1,14 @@
-When given a list of PRs to review, assign ALL PRs to ALL workers. Each worker reviews every PR through their specialized lens. This creates multi-model consensus — the same PR reviewed by 5 different models with 5 different specializations.
+When given a list of PRs to review, assign ONE PR to EACH worker. Distribute PRs round-robin across the available workers. If there are more PRs than workers, assign multiple PRs per worker.
 
-For each PR assignment, include the PR number and instruct the worker to run `gh pr diff <number>` and `gh pr view <number>` to get the full context.
+For each PR assignment, just tell the worker: "Review PR #<number>"
 
-After all workers complete, synthesize a final report per PR:
-- Issues found by multiple reviewers (high confidence)
-- Issues found by only one reviewer (needs human judgment)
-- Overall risk rating (🔴 critical / 🟡 moderate / 🟢 clean)
+The workers handle everything else — fetching the diff, dispatching multi-model sub-agents, and synthesizing results. Do NOT micromanage the review process.
+
+After all workers complete, produce a brief summary table:
+
+| PR | Verdict | Key Issues |
+|----|---------|------------|
+| #194 | ✅ Ready to merge | None |
+| #193 | ⚠️ Needs changes | Race condition in auth handler |
+
+Verdicts: ✅ Ready to merge, ⚠️ Needs changes, 🔴 Do not merge

.squad/team.md

@@ -1,9 +1,11 @@
 # PR Review Squad
 
+mode: orchestrator
+
 | Member | Role |
 |--------|------|
-| bug-hunter | Bug Hunter |
-| security-analyst | Security Analyst |
-| arch-critic | Architecture Critic |
-| edge-case-finder | Edge Case Finder |
-| correctness-checker | Correctness Checker |
+| reviewer-1 | PR Reviewer |
+| reviewer-2 | PR Reviewer |
+| reviewer-3 | PR Reviewer |
+| reviewer-4 | PR Reviewer |
+| reviewer-5 | PR Reviewer |

PolyPilot.Tests/SquadDiscoveryTests.cs

@@ -257,4 +257,48 @@ public void Discover_HasEmoji()
         var presets = SquadDiscovery.Discover(SquadSampleDir);
         Assert.Equal("🫡", presets[0].Emoji);
     }
+
+    // --- ParseMode tests ---
+
+    [Fact]
+    public void ParseMode_Orchestrator()
+    {
+        var content = "# My Team\nmode: orchestrator\n| Member | Role |";
+        Assert.Equal(MultiAgentMode.Orchestrator, SquadDiscovery.ParseMode(content));
+    }
+
+    [Fact]
+    public void ParseMode_Broadcast()
+    {
+        var content = "# My Team\nmode: broadcast\n";
+        Assert.Equal(MultiAgentMode.Broadcast, SquadDiscovery.ParseMode(content));
+    }
+
+    [Fact]
+    public void ParseMode_OrchestratorReflect()
+    {
+        var content = "# My Team\nmode: orchestrator-reflect\n";
+        Assert.Equal(MultiAgentMode.OrchestratorReflect, SquadDiscovery.ParseMode(content));
+    }
+
+    [Fact]
+    public void ParseMode_Sequential()
+    {
+        var content = "# My Team\nmode: sequential\n";
+        Assert.Equal(MultiAgentMode.Sequential, SquadDiscovery.ParseMode(content));
+    }
+
+    [Fact]
+    public void ParseMode_CaseInsensitive()
+    {
+        var content = "# My Team\nMode: Orchestrator\n";
+        Assert.Equal(MultiAgentMode.Orchestrator, SquadDiscovery.ParseMode(content));
+    }
+
+    [Fact]
+    public void ParseMode_DefaultsToReflect_WhenMissing()
+    {
+        var content = "# My Team\n| Member | Role |";
+        Assert.Equal(MultiAgentMode.OrchestratorReflect, SquadDiscovery.ParseMode(content));
+    }
 }

PolyPilot/Models/SquadDiscovery.cs

@@ -38,10 +38,11 @@ public static List<GroupPreset> Discover(string worktreeRoot)
             if (agents.Count == 0) return new();
 
             var teamName = ParseTeamName(teamContent) ?? "Squad Team";
+            var mode = ParseMode(teamContent);
             var decisions = ReadOptionalFile(Path.Combine(squadDir, "decisions.md"), MaxDecisionsLength);
             var routing = ReadOptionalFile(Path.Combine(squadDir, "routing.md"), MaxDecisionsLength);
 
-            var preset = BuildPreset(teamName, agents, decisions, routing, squadDir);
+            var preset = BuildPreset(teamName, agents, decisions, routing, squadDir, mode);
             return new List<GroupPreset> { preset };
         }
         catch
@@ -110,6 +111,33 @@ internal static List<SquadAgent> DiscoverAgents(string squadDir)
         return null;
     }
 
+    /// <summary>
+    /// Parse mode from team.md content.
+    /// Looks for a line like "mode: orchestrator" (case-insensitive).
+    /// Supports: broadcast, sequential, orchestrator, orchestrator-reflect.
+    /// Defaults to OrchestratorReflect if not specified.
+    /// </summary>
+    internal static MultiAgentMode ParseMode(string teamContent)
+    {
+        foreach (var line in teamContent.Split('\n'))
+        {
+            var trimmed = line.Trim();
+            if (trimmed.StartsWith("mode:", StringComparison.OrdinalIgnoreCase))
+            {
+                var value = trimmed["mode:".Length..].Trim().ToLowerInvariant();
+                return value switch
+                {
+                    "broadcast" => MultiAgentMode.Broadcast,
+                    "sequential" => MultiAgentMode.Sequential,
+                    "orchestrator" => MultiAgentMode.Orchestrator,
+                    "orchestrator-reflect" or "orchestratorreflect" or "reflect" => MultiAgentMode.OrchestratorReflect,
+                    _ => MultiAgentMode.OrchestratorReflect
+                };
+            }
+        }
+        return MultiAgentMode.OrchestratorReflect;
+    }
+
     /// <summary>
     /// Parse agent roster from team.md table rows.
     /// Returns member names from the first column of markdown tables.
@@ -144,7 +172,7 @@ internal static List<string> ParseRosterNames(string teamContent)
     }
 
     private static GroupPreset BuildPreset(string teamName, List<SquadAgent> agents,
-        string? decisions, string? routing, string squadDir)
+        string? decisions, string? routing, string squadDir, MultiAgentMode mode)
     {
         // Use a sensible default model for all agents (user can override after creation)
         var defaultModel = "claude-sonnet-4.6";
@@ -157,7 +185,7 @@ private static GroupPreset BuildPreset(string teamName, List<SquadAgent> agents,
             teamName,
             $"Squad team from {Path.GetFileName(Path.GetDirectoryName(squadDir) ?? squadDir)}",
             "🫡",
-            MultiAgentMode.OrchestratorReflect,
+            mode,
             orchestratorModel,
             workerModels)
         {