@brutalist/mcp 1.8.1 → 1.9.0

package/dist/debate/debate-orchestrator.js

@@ -0,0 +1,699 @@
+/**
+ * DebateOrchestrator — debate orchestration extracted from brutalist-server.ts.
+ *
+ * This module encapsulates the entire debate subsystem:
+ *   - handleDebateToolExecution(): cache-aware entry point for debate tool calls
+ *   - executeCLIDebate(): core debate engine with 3-tier escalation
+ *
+ * Dependencies are injected via constructor, making brutalist-server.ts a pure
+ * composition root that wires and delegates.
+ *
+ * Extracted from brutalist-server.ts lines 665-1348.
+ */
+import { existsSync } from 'fs';
+import { join as pathJoin, resolve as pathResolve } from 'path';
+import { mediateTranscript } from '../utils/transcript-mediator.js';
+import { parseCursor, PAGINATION_DEFAULTS } from '../utils/pagination.js';
+import { safeMetric as sharedSafeMetric, } from '../metrics/index.js';
+import { detectRefusal } from './refusal-detection.js';
+import { constitutionalAnchor } from './constitutional.js';
+import { synthesizeDebate } from './synthesis.js';
+/**
+ * Rank of each debate tier for computing the MAX tier reached across all
+ * turns of a debate. Used to derive the `tier` label on the debate
+ * duration histogram (higher rank wins).
+ */
+const TIER_RANK = {
+    standard: 0,
+    escalated: 1,
+    decomposed: 2,
+};
+/**
+ * DebateOrchestrator encapsulates all debate orchestration logic.
+ *
+ * It accepts dependencies via constructor injection so that brutalist-server.ts
+ * remains a thin composition root.
+ */
+export class DebateOrchestrator {
+    /** Mutable so test harnesses can replace cliOrchestrator on BrutalistServer. */
+    _cliOrchestrator;
+    responseCache;
+    formatter;
+    config;
+    onStreamingEvent;
+    onProgressUpdate;
+    metrics;
+    log;
+    get cliOrchestrator() {
+        return this._cliOrchestrator;
+    }
+    set cliOrchestrator(value) {
+        this._cliOrchestrator = value;
+    }
+    constructor(deps) {
+        this._cliOrchestrator = deps.cliOrchestrator;
+        this.responseCache = deps.responseCache;
+        this.formatter = deps.formatter;
+        this.config = deps.config;
+        this.onStreamingEvent = deps.onStreamingEvent;
+        this.onProgressUpdate = deps.onProgressUpdate;
+        this.metrics = deps.metrics;
+        this.log = deps.log;
+    }
+    /**
+     * Isolate metric writes from business control flow.
+     *
+     * Delegates to the shared `safeMetric` helper in
+     * `src/metrics/safe-metric.ts`. The private method is retained so
+     * existing call sites inside DebateOrchestrator
+     * (`this.safeMetric(op, fn)`) keep working without a touch, and so
+     * any debate-specific metric-error instrumentation can be layered in
+     * one place in the future.
+     *
+     * Parity note: `CLIAgentOrchestrator` uses the same shared helper
+     * directly (no private method) to prevent metric throws from
+     * propagating into the outer spawn try/catch. See Cycle 3 rework
+     * Task CLI-B' in phases/instrument_cli_spawn/phase.md.
+     */
+    safeMetric(op, fn) {
+        sharedSafeMetric(this.log, op, fn);
+    }
+    /**
+     * Handle debate tool execution with constitutional position anchoring.
+     * Uses 2 randomly selected agents (or user-specified) with explicit PRO/CON positions.
+     *
+     * This is the entry point called from the roast_cli_debate tool registration.
+     *
+     * Instrumentation (intent #1): every exit path records the debate
+     * orchestration duration histogram exactly once. The `tier` label is the
+     * MAX tier reached across all turns of the underlying `executeCLIDebate`
+     * call; cache-hit paths short-circuit before any CLI agent runs, so their
+     * tier is always `'standard'`. The outer try/finally placement ensures
+     * error paths, refusal paths, and cache-hit paths all emit exactly one
+     * observation — `executeCLIDebate` itself has NO timer block to avoid
+     * double-observation.
+     */
+    async handleDebateToolExecution(args, extra) {
+        const handleToolLog = this.log.forOperation('handle_tool');
+        const t0 = Date.now();
+        // Histogram labels — DEBATE_DURATION_LABELS = ['outcome', 'tier'] as const.
+        // outcome is derived from the debate result's behavior (refused vs. success)
+        // or forced to 'error' in the catch branch.
+        let outcome = 'success';
+        let tier = 'standard';
+        try {
+            // Build pagination params
+            const paginationParams = {
+                offset: args.offset || 0,
+                limit: args.limit || PAGINATION_DEFAULTS.DEFAULT_LIMIT_TOKENS
+            };
+            if (args.cursor) {
+                const cursorParams = parseCursor(args.cursor);
+                Object.assign(paginationParams, cursorParams);
+            }
+            const explicitPaginationRequested = args.offset !== undefined ||
+                args.limit !== undefined ||
+                args.cursor !== undefined ||
+                args.context_id !== undefined;
+            // Extract session ID early — needed for cache session isolation
+            const sessionId = extra?.sessionId ||
+                extra?._meta?.sessionId ||
+                extra?.headers?.['mcp-session-id'] ||
+                'anonymous';
+            // Validate resume flag requires context_id
+            if (args.resume && !args.context_id) {
+                throw new Error(`The 'resume' flag requires a 'context_id' from a previous debate. ` +
+                    `Run an initial debate first, then use the returned context_id with resume: true.`);
+            }
+            // Check cache if context_id provided
+            let conversationHistory;
+            if (args.context_id && !args.force_refresh) {
+                const cachedResponse = await this.responseCache.getByContextId(args.context_id, sessionId);
+                if (cachedResponse) {
+                    handleToolLog.info(`🎯 Debate cache HIT for context_id: ${args.context_id}`);
+                    if (args.resume === true) {
+                        // CONVERSATION CONTINUATION: Continue the debate
+                        if (!args.topic || args.topic.trim() === '') {
+                            throw new Error(`Debate continuation (resume: true) requires a new prompt/question. ` +
+                                `Provide your follow-up in the topic field.`);
+                        }
+                        // Security: avoid logging user-provided topic text at info level.
+                        // Emit length only; if a developer needs the preview, run at debug.
+                        handleToolLog.info('Debate continuation - new prompt received', {
+                            topicLength: args.topic.length,
+                        });
+                        conversationHistory = cachedResponse.conversationHistory || [];
+                        // Fall through to execute new debate round with history
+                    }
+                    else {
+                        // PAGINATION: Return cached debate result — no agent ran,
+                        // outcome='success' and tier='standard' (their initial values).
+                        handleToolLog.info(`📖 Debate pagination request - returning cached response`);
+                        const cachedResult = {
+                            success: true,
+                            responses: [{
+                                    agent: 'cached',
+                                    success: true,
+                                    output: cachedResponse.content,
+                                    executionTime: 0
+                                }]
+                        };
+                        return this.formatter.formatToolResponse(cachedResult, args.verbose, paginationParams, args.context_id, explicitPaginationRequested);
+                    }
+                }
+                else {
+                    handleToolLog.warn(`❌ Debate cache MISS for context_id: ${args.context_id}`);
+                    throw new Error(`Context ID "${args.context_id}" not found in cache. ` +
+                        `It may have expired (2 hour TTL) or belong to a different session. ` +
+                        `Remove context_id parameter to run a new debate.`);
+                }
+            }
+            // Generate cache key for this debate
+            const cacheKey = this.responseCache.generateCacheKey({
+                tool: 'roast_cli_debate',
+                topic: args.topic,
+                proPosition: args.proPosition,
+                conPosition: args.conPosition,
+                agents: args.agents,
+                rounds: args.rounds,
+                context: args.context
+            });
+            // Check cache for identical request (if not resuming)
+            if (!args.force_refresh && !args.resume) {
+                const cachedContent = await this.responseCache.get(cacheKey);
+                if (cachedContent) {
+                    const existingContextId = this.responseCache.findContextIdForKey(cacheKey);
+                    const contextId = existingContextId
+                        ? this.responseCache.createAlias(existingContextId, cacheKey)
+                        : this.responseCache.generateContextId(cacheKey);
+                    handleToolLog.info(`🎯 Debate cache hit for new request, using context_id: ${contextId}`);
+                    const cachedResult = {
+                        success: true,
+                        responses: [{
+                                agent: 'cached',
+                                success: true,
+                                output: cachedContent,
+                                executionTime: 0
+                            }]
+                    };
+                    // Cache hit: outcome='success', tier='standard' (no agent ran).
+                    return this.formatter.formatToolResponse(cachedResult, args.verbose, paginationParams, contextId, explicitPaginationRequested);
+                }
+            }
+            // Build context with conversation history if resuming
+            let debateContext = args.context || '';
+            if (conversationHistory && conversationHistory.length > 0) {
+                const previousDebate = conversationHistory.map(msg => {
+                    const role = msg.role === 'user' ? 'User Question' : 'Debate Response';
+                    return `${role}:\n${msg.content}`;
+                }).join('\n\n---\n\n');
+                debateContext = `## Previous Debate Context\n\n${previousDebate}\n\n---\n\n## New Follow-up Question\n\nThe user wants to continue this debate with a new question or direction.\n\n${debateContext}`;
+                handleToolLog.info(`💬 Injected ${conversationHistory.length} previous messages into debate context`);
+            }
+            // Extract streaming context from extra
+            const progressToken = extra?._meta?.progressToken;
+            // Execute the debate
+            const numRounds = Math.min(args.rounds || 3, 3);
+            const result = await this.executeCLIDebate({
+                topic: args.topic,
+                proPosition: args.proPosition,
+                conPosition: args.conPosition,
+                agents: args.agents,
+                rounds: numRounds,
+                context: debateContext,
+                workingDirectory: args.workingDirectory,
+                models: args.models,
+                onStreamingEvent: this.onStreamingEvent,
+                progressToken,
+                onProgress: progressToken && sessionId ?
+                    (progress, total, message) => this.onProgressUpdate(progressToken, progress, total, message, sessionId) : undefined,
+                sessionId,
+                mcp_servers: args.mcp_servers,
+            });
+            // Derive outcome and tier from the debate result for the histogram
+            // observation that fires in the finally block below. The counter for
+            // per-turn escalation tier already fired inside executeCLIDebate; this
+            // block only extracts the histogram labels — no metric emissions here.
+            const turns = result.debateBehavior?.turns ?? [];
+            if (turns.length > 0) {
+                // Tier = MAX tier reached across all turns (higher rank wins).
+                tier = turns.reduce((max, t) => TIER_RANK[t.tier] > TIER_RANK[max] ? t.tier : max, 'standard');
+                // Outcome = 'refused' when every turn's engaged=false AND at least
+                // one turn refused. Otherwise 'success'. The catch branch below
+                // overrides to 'error'.
+                const allDisengaged = turns.every(t => !t.engaged);
+                const anyRefused = turns.some(t => t.refused);
+                if (allDisengaged && anyRefused) {
+                    outcome = 'refused';
+                }
+            }
+            // Cache the result
+            let contextId;
+            if (result.success && result.responses.length > 0) {
+                const fullContent = this.formatter.extractFullContent(result);
+                if (fullContent) {
+                    const now = Date.now();
+                    const updatedConversation = [
+                        ...(conversationHistory || []),
+                        { role: 'user', content: args.topic, timestamp: now },
+                        { role: 'assistant', content: fullContent, timestamp: now }
+                    ];
+                    if (args.resume && args.context_id && conversationHistory) {
+                        // Update existing cache entry
+                        contextId = args.context_id;
+                        await this.responseCache.updateByContextId(contextId, fullContent, updatedConversation, sessionId);
+                        this.log.forOperation('cache').info(`✅ Updated debate conversation ${contextId} (now ${updatedConversation.length} messages)`);
+                    }
+                    else {
+                        // New debate - create new context_id
+                        const { contextId: newId } = await this.responseCache.set({ tool: 'roast_cli_debate', topic: args.topic }, fullContent, cacheKey, sessionId, undefined, updatedConversation);
+                        contextId = newId;
+                        this.log.forOperation('cache').info(`✅ Cached new debate with context ID: ${contextId}`);
+                    }
+                }
+            }
+            return this.formatter.formatToolResponse(result, args.verbose, paginationParams, contextId, explicitPaginationRequested);
+        }
+        catch (error) {
+            outcome = 'error';
+            return this.formatter.formatErrorResponse(error);
+        }
+        finally {
+            // Record the debate duration exactly once per invocation. This is the
+            // SINGLE histogram observation point for debate orchestration — do
+            // NOT add another observe() call inside executeCLIDebate or any
+            // inner path. The typed label record below references
+            // DEBATE_DURATION_LABELS so a future label-set change triggers a
+            // compile error at this call site.
+            const durationSec = (Date.now() - t0) / 1000;
+            const durationLabels = {
+                outcome,
+                tier,
+            };
+            this.safeMetric('observe:debate_duration', () => this.metrics.debateOrchestrationDurationSeconds.observe(durationLabels, durationSec));
+        }
+    }
+    /**
+     * Execute CLI debate with constitutional position anchoring.
+     * 2 agents, explicit PRO/CON positions, context compression between rounds.
+     *
+     * This is the core debate engine. It manages:
+     *   - Agent selection and position assignment
+     *   - Round execution with 3-tier refusal escalation
+     *   - Transcript mediation between rounds
+     *   - Behavioral metadata and asymmetry detection
+     *   - Synthesis generation
+     */
+    async executeCLIDebate(args) {
+        const { topic, proPosition, conPosition, rounds, context, workingDirectory, models, onStreamingEvent, progressToken, onProgress, sessionId } = args;
+        const debateLog = this.log.forOperation('execute_debate');
+        const escalateLog = this.log.forOperation('escalate');
+        // Security (Cycle 3 F32): the debug-level emission previously leaked
+        // user-provided topic/proPosition/conPosition text into logs whenever
+        // BRUTALIST_LOG_LEVEL=debug was set — identical disclosure channel to
+        // the info-level site already redacted at :263. Emit length-only
+        // fields matching that pattern; a developer needing the raw text
+        // should inspect the transcript passed to executeCLIDebate directly.
+        debateLog.debug("Executing CLI debate", {
+            topicLength: topic.length,
+            proPositionLength: proPosition.length,
+            conPositionLength: conPosition.length,
+            rounds,
+        });
+        try {
+            // Get available CLIs
+            const cliContext = await this.cliOrchestrator.detectCLIContext();
+            const availableCLIs = cliContext.availableCLIs;
+            if (availableCLIs.length < 2) {
+                throw new Error(`Need at least 2 CLI agents for debate. Available: ${availableCLIs.join(', ')}`);
+            }
+            // Select 2 agents: use specified or random selection
+            let selectedAgents;
+            if (args.agents && args.agents.length === 2) {
+                // Validate specified agents are available
+                const unavailable = args.agents.filter(a => !availableCLIs.includes(a));
+                if (unavailable.length > 0) {
+                    throw new Error(`Specified agents not available: ${unavailable.join(', ')}. Available: ${availableCLIs.join(', ')}`);
+                }
+                selectedAgents = args.agents;
+            }
+            else {
+                // Random selection of 2 agents
+                const shuffled = [...availableCLIs].sort(() => Math.random() - 0.5);
+                selectedAgents = shuffled.slice(0, 2);
+            }
+            // Randomly assign PRO/CON positions
+            const shuffledAgents = [...selectedAgents].sort(() => Math.random() - 0.5);
+            const proAgent = shuffledAgents[0];
+            const conAgent = shuffledAgents[1];
+            debateLog.info(`🎭 Debate: ${proAgent.toUpperCase()} (PRO) vs ${conAgent.toUpperCase()} (CON)`);
+            const debateResponses = [];
+            const transcript = [];
+            const turnMetadata = [];
+            let compressedContext = '';
+            const totalTurns = rounds * 2; // 2 agents per round
+            let completedTurns = 0;
+            // Frontier 1: Detect self-referential working directory (Codex reading its own control prompts)
+            const resolvedWorkDir = args.target || workingDirectory || this.config.workingDirectory || process.cwd();
+            const absWorkDir = pathResolve(resolvedWorkDir);
+            const isSelfReferential = existsSync(pathJoin(absWorkDir, 'src', 'brutalist-server.ts'))
+                || existsSync(pathJoin(absWorkDir, 'dist', 'brutalist-server.js'));
+            if (isSelfReferential) {
+                debateLog.info(`🔒 Debate working directory is brutalist repo — Codex will be sandboxed`);
+            }
+            // Execute rounds
+            for (let round = 1; round <= rounds; round++) {
+                debateLog.info(`📢 Round ${round}/${rounds}`);
+                // Both agents argue in each round
+                for (const [agent, position, thesis] of [
+                    [proAgent, 'PRO', proPosition],
+                    [conAgent, 'CON', conPosition]
+                ]) {
+                    let prompt;
+                    debateLog.info(`  ⚔️ ${agent.toUpperCase()} (${position}) arguing...`);
+                    // Build prompt-generation function so we can rebuild on escalation
+                    const mcpBlock = args.mcp_servers?.length
+                        ? `\nEXTERNAL TOOL ACCESS: You have MCP tools available (${args.mcp_servers.join(', ')}). Use them to gather evidence supporting your position. You MUST NOT modify the codebase.\n`
+                        : '';
+                    const buildPrompt = (tier) => {
+                        if (round === 1) {
+                            return `${constitutionalAnchor(agent, position, thesis, tier)}
+${mcpBlock}
+TOPIC: ${topic}
+${context ? `CONTEXT: ${context}` : ''}
+
+Round 1: Opening analysis.
+
+Present your ${position} analysis. Structure your response:
+
+<thesis_statement>
+Your core analytical position
+</thesis_statement>
+
+<key_arguments>
+Three strongest arguments grounding your position in evidence and reasoning
+</key_arguments>
+
+<preemptive_rebuttal>
+Address the strongest counterargument and show why it does not defeat your position
+</preemptive_rebuttal>
+
+<conclusion>
+Reinforce why your analysis holds
+</conclusion>`;
+                        }
+                        else {
+                            const rawOpponent = transcript
+                                .filter(t => t.agent !== agent && t.round === round - 1)
+                                .map(t => t.content)
+                                .join('\n\n');
+                            const { sanitized: opponentTranscript, patternsDetected: opponentPatterns } = mediateTranscript(rawOpponent, 'sanitize', 4000);
+                            if (opponentPatterns.length > 0) {
+                                debateLog.info(`🛡️ Mediated ${opponentPatterns.length} patterns from opponent transcript for ${agent}`, { opponentPatterns });
+                            }
+                            return `${constitutionalAnchor(agent, position, thesis, tier)}
+${mcpBlock}
+TOPIC: ${topic}
+
+Round ${round}: Engage with your counterpart's analysis.
+
+YOUR COUNTERPART'S PREVIOUS ANALYSIS:
+${opponentTranscript || 'No previous analysis recorded'}
+
+${compressedContext ? `ANALYSIS CONTEXT SO FAR:\n${compressedContext}\n` : ''}
+
+<counterpart_gaps>
+Identify the specific weaknesses in their reasoning and evidence
+</counterpart_gaps>
+
+<deepening_analysis>
+Advance new evidence and reasoning that strengthens your position
+</deepening_analysis>
+
+<reinforcement>
+Show why your position holds against their strongest points
+</reinforcement>`;
+                        }
+                    };
+                    try {
+                        const turnRequestId = `debate-${sessionId || 'anon'}-${round}-${agent}-${Date.now()}`;
+                        // Emit agent_start streaming event
+                        if (onStreamingEvent) {
+                            onStreamingEvent({
+                                type: 'agent_start',
+                                agent,
+                                content: `Round ${round}/${rounds}: ${agent.toUpperCase()} (${position}) arguing...`,
+                                timestamp: Date.now(),
+                                sessionId,
+                            });
+                        }
+                        // Working directory: debateMode suppresses Codex shell exploration via prompt,
+                        // so no need to redirect — Codex still needs a git repo to function
+                        const agentWorkDir = workingDirectory || this.config.workingDirectory;
+                        const cliOptions = {
+                            workingDirectory: agentWorkDir,
+                            timeout: (this.config.defaultTimeout || 60000) * 2,
+                            models,
+                            onStreamingEvent,
+                            progressToken,
+                            onProgress,
+                            sessionId,
+                            requestId: turnRequestId,
+                            debateMode: true, // Frontier 1: suppress Codex shell exploration
+                            mcpServers: args.mcp_servers, // MCP servers for evidence-backed debate
+                        };
+                        // Three-tier escalation: standard -> escalated -> decomposed
+                        prompt = buildPrompt('standard');
+                        let wasRefused = false;
+                        let wasEscalated = false;
+                        let engagedAfterEscalation = false;
+                        let finalTier = 'standard';
+                        let response = await this.cliOrchestrator.executeSingleCLI(agent, prompt, prompt, cliOptions);
+                        // Tier 2: Detect refusal -> retry with analytical framing
+                        if (response.success && response.output && detectRefusal(response.output)) {
+                            wasRefused = true;
+                            wasEscalated = true;
+                            finalTier = 'escalated';
+                            escalateLog.warn(`🛡️ ${agent.toUpperCase()} (${position}) refused — escalating to analytical framing (tier 2)`);
+                            const escalatedPrompt = buildPrompt('escalated');
+                            const retryResponse = await this.cliOrchestrator.executeSingleCLI(agent, escalatedPrompt, escalatedPrompt, { ...cliOptions, requestId: `${turnRequestId}-escalated` });
+                            if (retryResponse.success && retryResponse.output && !detectRefusal(retryResponse.output)) {
+                                escalateLog.info(`✅ ${agent.toUpperCase()} (${position}) engaged after tier 2 escalation`);
+                                engagedAfterEscalation = true;
+                                response = retryResponse;
+                            }
+                            else {
+                                // Tier 3: Decomposed — scholarly steelman framing
+                                finalTier = 'decomposed';
+                                escalateLog.warn(`🛡️ ${agent.toUpperCase()} (${position}) refused tier 2 — escalating to decomposed framing (tier 3)`);
+                                const decomposedPrompt = buildPrompt('decomposed');
+                                const decomposedResponse = await this.cliOrchestrator.executeSingleCLI(agent, decomposedPrompt, decomposedPrompt, { ...cliOptions, requestId: `${turnRequestId}-decomposed` });
+                                if (decomposedResponse.success && decomposedResponse.output && !detectRefusal(decomposedResponse.output)) {
+                                    escalateLog.info(`✅ ${agent.toUpperCase()} (${position}) engaged after tier 3 decomposition`);
+                                    engagedAfterEscalation = true;
+                                    response = decomposedResponse;
+                                }
+                                else {
+                                    escalateLog.warn(`⚠️ ${agent.toUpperCase()} (${position}) refused all 3 tiers — using best response`);
+                                    // Use decomposed response if available (likely less meta-commentary)
+                                    if (decomposedResponse.success && decomposedResponse.output) {
+                                        response = decomposedResponse;
+                                    }
+                                }
+                            }
+                        }
+                        // Always add response (success or failure) for visibility
+                        debateResponses.push(response);
+                        completedTurns++;
+                        // Emit agent_complete streaming event
+                        if (onStreamingEvent) {
+                            onStreamingEvent({
+                                type: 'agent_complete',
+                                agent,
+                                content: `Round ${round}/${rounds}: ${agent.toUpperCase()} (${position}) ${response.success ? 'finished' : 'failed'}`,
+                                timestamp: Date.now(),
+                                sessionId,
+                            });
+                        }
+                        // Emit progress update
+                        if (onProgress) {
+                            onProgress(completedTurns, totalTurns, `Debate: ${completedTurns}/${totalTurns} turns complete`);
+                        }
+                        // Frontier 3: Track behavioral metadata
+                        const finalRefused = response.success && response.output ? detectRefusal(response.output) : false;
+                        turnMetadata.push({
+                            agent: agent,
+                            position: position,
+                            round,
+                            engaged: response.success && !!response.output && !finalRefused,
+                            refused: wasRefused,
+                            escalated: wasEscalated,
+                            engagedAfterEscalation,
+                            responseLength: response.output?.length || 0,
+                            executionTime: response.executionTime,
+                            tier: engagedAfterEscalation ? finalTier : (wasEscalated ? finalTier : 'standard'),
+                        });
+                        // Escalation-tier counter: fires exactly ONCE per turn, labeled
+                        // with this turn's FINAL tier (standard/escalated/decomposed).
+                        // Retries within a single turn are NOT counted separately —
+                        // they are represented by the final tier value on the pushed
+                        // turnMetadata record. The typed label record references
+                        // ESCALATION_TIER_LABELS so a future label-set change
+                        // triggers a compile error at this call site. The call is
+                        // wrapped in safeMetric so a metric throw cannot corrupt the
+                        // surrounding turn try/catch (would otherwise double-push
+                        // metadata and double-count completedTurns).
+                        const successTierLabels = {
+                            tier: turnMetadata[turnMetadata.length - 1].tier,
+                        };
+                        this.safeMetric('inc:escalation_tier', () => this.metrics.debateEscalationTierTotal.inc(successTierLabels, 1));
+                        if (response.success && response.output) {
+                            transcript.push({
+                                agent,
+                                position,
+                                round,
+                                content: response.output
+                            });
+                        }
+                        else {
+                            // Security (Cycle 3 F33 Pattern A): response.error can carry
+                            // CLI-subprocess stderr tail, which in turn may echo model-
+                            // generated or prompt-echoed text. Emit a presence-only flag
+                            // at warn level instead of the raw string; operators with
+                            // debug file-logging can still correlate via agent/position/
+                            // round, and the transcript is the canonical source of truth
+                            // for the actual failure text.
+                            debateLog.warn(`⚠️ ${agent.toUpperCase()} (${position}) failed`, {
+                                agent,
+                                position,
+                                error: response.error ? '<redacted>' : undefined,
+                                hasOutput: Boolean(response.output),
+                            });
+                        }
+                    }
+                    catch (error) {
+                        // Security (Cycle 3 F33): the StructuredLogger emitError path
+                        // serializes the raw Error verbatim (message, stack, name)
+                        // into NDJSON. Passing the original `error` leaks any
+                        // CLI-subprocess stderr tail or prompt-echoed text embedded
+                        // in error.message. Pass a sanitized Error-shaped shim that
+                        // preserves `name` for diagnostic triage while redacting the
+                        // payload. `.stack` is omitted from the shim (undefined) so
+                        // the file-side fileData record carries only name+message.
+                        const errorName = error instanceof Error ? error.name : 'Error';
+                        const errorShim = { name: errorName, message: '<redacted>' };
+                        debateLog.error(`❌ ${agent.toUpperCase()} (${position}) threw error`, errorShim);
+                        completedTurns++;
+                        // Security (Cycle 4 F7/F17): the same raw caught error.message
+                        // that Cycle 3 redacted at the logger sink was still flowing
+                        // through two adjacent sinks — the streaming event content
+                        // (remote subscribers) and the debateResponses push (flows
+                        // back out as `responses` at the return site, and downstream
+                        // into synthesis.ts and response-formatter.ts). Emit a static
+                        // classifier that retains the agent identity for operator
+                        // triage but carries no subprocess/prompt-derived payload.
+                        const redactedTurnError = `${agent.toUpperCase()} execution failed. See internal logs for details.`;
+                        if (onStreamingEvent) {
+                            onStreamingEvent({
+                                type: 'agent_error',
+                                agent,
+                                content: `Round ${round}/${rounds}: ${agent.toUpperCase()} (${position}) error: ${redactedTurnError}`,
+                                timestamp: Date.now(),
+                                sessionId,
+                            });
+                        }
+                        turnMetadata.push({
+                            agent: agent,
+                            position: position,
+                            round,
+                            engaged: false,
+                            refused: false,
+                            escalated: false,
+                            engagedAfterEscalation: false,
+                            responseLength: 0,
+                            executionTime: 0,
+                            tier: 'standard',
+                        });
+                        // Error-path turn: still counts exactly ONCE per turn. Tier is
+                        // 'standard' because the turn never reached the refusal-retry
+                        // branches — it threw before any escalation decision. The
+                        // typed label record references ESCALATION_TIER_LABELS so a
+                        // future label-set change triggers a compile error at this
+                        // call site. Wrapped in safeMetric so a metric throw cannot
+                        // re-enter the catch path and double-count the turn.
+                        const errorTierLabels = {
+                            tier: turnMetadata[turnMetadata.length - 1].tier,
+                        };
+                        this.safeMetric('inc:escalation_tier', () => this.metrics.debateEscalationTierTotal.inc(errorTierLabels, 1));
+                        debateResponses.push({
+                            agent,
+                            success: false,
+                            output: '',
+                            error: redactedTurnError,
+                            executionTime: 0
+                        });
+                    }
+                }
+                // Compress context for next round with mediation (if not final round)
+                if (round < rounds) {
+                    const roundTranscript = transcript
+                        .filter(t => t.round === round)
+                        .map(t => {
+                        const { sanitized } = mediateTranscript(t.content, 'sanitize', 1500);
+                        return `${t.agent.toUpperCase()} (${t.position}): ${sanitized}`;
+                    })
+                        .join('\n\n---\n\n');
+                    compressedContext = `Round ${round} Summary:\n${roundTranscript}`;
+                }
+            }
+            // Compute position-dependent asymmetry summary
+            const proTurns = turnMetadata.filter(t => t.position === 'PRO');
+            const conTurns = turnMetadata.filter(t => t.position === 'CON');
+            const proRefusalRate = proTurns.length > 0
+                ? proTurns.filter(t => t.refused).length / proTurns.length : 0;
+            const conRefusalRate = conTurns.length > 0
+                ? conTurns.filter(t => t.refused).length / conTurns.length : 0;
+            const debateAgents = [...new Set(turnMetadata.map(t => t.agent))];
+            const agentAsymmetries = debateAgents.map(a => {
+                const aPro = turnMetadata.filter(t => t.agent === a && t.position === 'PRO');
+                const aCon = turnMetadata.filter(t => t.agent === a && t.position === 'CON');
+                const proEngaged = aPro.some(t => t.engaged);
+                const conEngaged = aCon.some(t => t.engaged);
+                return { agent: a, proEngaged, conEngaged, asymmetric: proEngaged !== conEngaged };
+            });
+            const asymmetryDetected = Math.abs(proRefusalRate - conRefusalRate) > 0.3
+                || agentAsymmetries.some(a => a.asymmetric);
+            const behaviorSummary = {
+                topic, proPosition, conPosition,
+                turns: turnMetadata,
+                asymmetry: {
+                    detected: asymmetryDetected,
+                    description: asymmetryDetected
+                        ? `Position-dependent asymmetry: PRO refusal ${(proRefusalRate * 100).toFixed(0)}%, CON refusal ${(conRefusalRate * 100).toFixed(0)}%`
+                        : 'No significant position-dependent asymmetry detected',
+                    proRefusalRate,
+                    conRefusalRate,
+                    agentAsymmetries,
+                }
+            };
+            if (asymmetryDetected) {
+                debateLog.warn(`🎭 Alignment asymmetry detected: ${behaviorSummary.asymmetry.description}`);
+            }
+            // Build synthesis with behavioral data
+            const synthesis = synthesizeDebate(debateResponses, topic, rounds, new Map([[proAgent, `PRO: ${proPosition}`], [conAgent, `CON: ${conPosition}`]]), behaviorSummary);
+            return {
+                success: debateResponses.some(r => r.success),
+                responses: debateResponses,
+                synthesis,
+                debateBehavior: behaviorSummary,
+                analysisType: 'cli_debate',
+                topic
+            };
+        }
+        catch (error) {
+            debateLog.error("CLI debate execution failed", error);
+            throw error;
+        }
+    }
+}
+//# sourceMappingURL=debate-orchestrator.js.map