@archal/cli 0.9.1 → 0.9.5

package/harnesses/openclaw/agent.mjs

@@ -1,229 +0,0 @@
-/**
- * OpenClaw Harness Agent — bridges OpenClaw to Archal twin infrastructure.
- *
- * Native OpenClaw CLI execution only:
- *
- * 1. **Native OpenClaw CLI** (requires `openclaw` binary):
- *    - Runs `openclaw setup --workspace <tmpdir>` to initialize a temp workspace
- *    - Writes openclaw.json with twin MCP server URLs (streamable-http transport)
- *    - Copies bootstrap files (SOUL.md, AGENTS.md, TOOLS.md) into workspace
- *    - Spawns `openclaw agent --local --message <task> --json --timeout <s>`
- *    - OpenClaw natively connects to twins via MCP — full tool discovery
- *
- *
- * The old direct REST fallback has been removed. Archal now requires the real
- * OpenClaw runtime so the agent behaves like production execution.
- */
-
-import { execSync, spawn } from 'node:child_process';
-import { existsSync, writeFileSync, mkdirSync, readFileSync, rmSync } from 'node:fs';
-import { join, dirname } from 'node:path';
-import { tmpdir } from 'node:os';
-import { randomUUID } from 'node:crypto';
-import { collectTwinUrls } from '../_lib/rest-client.mjs';
-import { writeMetrics } from '../_lib/metrics.mjs';
-
-const TASK = (process.env['ARCHAL_ENGINE_TASK'] || '').trim();
-const MODEL = process.env['ARCHAL_ENGINE_MODEL'] || 'openclaw:main';
-if (!TASK) {
-  console.error('[openclaw] ARCHAL_ENGINE_TASK not set or empty');
-  process.exit(1);
-}
-
-// ── Detect OpenClaw installation ─────────────────────────────────────
-
-function isOpenClawInstalled() {
-  try {
-    execSync('openclaw --version', { stdio: 'pipe', timeout: 5000 });
-    return true;
-  } catch {
-    return false;
-  }
-}
-
-// ── Mode 1: Native OpenClaw with MCP twin connections ────────────────
-//
-// Validated against OpenClaw docs (docs.openclaw.ai):
-// - `openclaw setup --workspace <dir>` initializes a workspace at custom path
-// - `openclaw agent --local --message <text> --json --timeout <s>` runs locally
-// - MCP config in openclaw.json under mcpServers key
-// - Streamable HTTP transport uses { url: "..." } format
-// - No --workspace or --agent flags on `agent` subcommand
-// - Workspace override is via openclaw.json `agent.workspace` or setup flag
-
-async function runWithOpenClawCli() {
-  const twinUrls = collectTwinUrls();
-  const twinNames = Object.keys(twinUrls);
-  const harnessDir = dirname(new URL(import.meta.url).pathname);
-
-  if (twinNames.length === 0) {
-    console.error('[openclaw] No twin URLs found. Check ARCHAL_TWIN_NAMES and ARCHAL_<TWIN>_URL env vars.');
-    process.exit(1);
-  }
-
-  // Create a temp workspace directory
-  const workspaceDir = join(tmpdir(), `archal-openclaw-${randomUUID().slice(0, 8)}`);
-  mkdirSync(workspaceDir, { recursive: true });
-
-  // Build MCP server config for twin endpoints (streamable-http transport).
-  // OpenClaw reads mcpServers from openclaw.json — for HTTP transport,
-  // each entry needs just a `url` field pointing at the MCP endpoint.
-  const mcpServers = {};
-  for (const [twinName, baseUrl] of Object.entries(twinUrls)) {
-    const trimmed = baseUrl.trim().replace(/\/+$/, '');
-    const mcpUrl = trimmed.endsWith('/mcp') ? trimmed : `${trimmed}/mcp`;
-    mcpServers[`archal-${twinName}`] = { url: mcpUrl };
-  }
-
-  // Write openclaw.json config — this is the canonical config location
-  // that OpenClaw reads on startup. We set agent.workspace to this dir
-  // and configure mcpServers with twin endpoints.
-  const openclawConfig = {
-    agent: {
-      workspace: workspaceDir,
-    },
-    mcpServers,
-  };
-  // OpenClaw looks for openclaw.json in ~/.openclaw/ by default,
-  // but with --local mode it also checks the current working directory.
-  // We write both locations to be safe.
-  const dotOpenclawDir = join(workspaceDir, '.openclaw');
-  mkdirSync(dotOpenclawDir, { recursive: true });
-  writeFileSync(
-    join(dotOpenclawDir, 'openclaw.json'),
-    JSON.stringify(openclawConfig, null, 2),
-  );
-  // Also write a .mcp.json in workspace root (project-level MCP config)
-  writeFileSync(
-    join(workspaceDir, '.mcp.json'),
-    JSON.stringify({ mcpServers }, null, 2),
-  );
-
-  // Copy bootstrap files from harness into workspace
-  for (const file of ['SOUL.md', 'AGENTS.md', 'TOOLS.md', 'IDENTITY.md']) {
-    const src = join(harnessDir, file);
-    if (existsSync(src)) {
-      writeFileSync(join(workspaceDir, file), readFileSync(src, 'utf-8'));
-    }
-  }
-
-  // Build environment for the OpenClaw process
-  const env = { ...process.env };
-  // Use OPENCLAW_PROFILE to isolate this run's config from user's default
-  const profileName = `archal-${randomUUID().slice(0, 6)}`;
-  env['OPENCLAW_PROFILE'] = profileName;
-  // Pass gateway token if available
-  if (process.env['ARCHAL_TOKEN'] && !env['OPENCLAW_GATEWAY_TOKEN']) {
-    env['OPENCLAW_GATEWAY_TOKEN'] = process.env['ARCHAL_TOKEN'];
-  }
-
-  const timeoutSeconds = parseInt(process.env['ARCHAL_ENGINE_TIMEOUT'] || '240', 10);
-  const runStart = Date.now();
-
-  return new Promise((resolve, reject) => {
-    // OpenClaw agent CLI: --local runs embedded, --message is the task,
-    // --json gives machine-readable output, --timeout sets deadline
-    const args = [
-      'agent',
-      '--local',
-      '--message', TASK,
-      '--json',
-      '--timeout', String(timeoutSeconds),
-    ];
-
-    console.error(`[openclaw] Spawning: openclaw ${args.slice(0, 3).join(' ')} ... --timeout ${timeoutSeconds}`);
-    console.error(`[openclaw] Workspace: ${workspaceDir}`);
-    console.error(`[openclaw] Twins: ${twinNames.join(', ')} (MCP streamable-http)`);
-    console.error(`[openclaw] Profile: ${profileName}`);
-
-    const child = spawn('openclaw', args, {
-      env,
-      cwd: workspaceDir, // Run from workspace so .mcp.json is discovered
-      stdio: ['pipe', 'pipe', 'pipe'],
-      timeout: (timeoutSeconds + 30) * 1000, // Buffer above agent timeout
-    });
-
-    let stdout = '';
-    let stderr = '';
-
-    child.stdout.on('data', (data) => {
-      stdout += data.toString();
-    });
-
-    child.stderr.on('data', (data) => {
-      const text = data.toString();
-      stderr += text;
-      process.stderr.write(text);
-    });
-
-    child.on('close', (code) => {
-      const totalTimeMs = Date.now() - runStart;
-
-      // Parse structured JSON output from OpenClaw
-      let parsedOutput = null;
-      try {
-        // OpenClaw --json may output multiple JSON objects; take the last one
-        const jsonLines = stdout.trim().split('\n').filter((l) => l.startsWith('{'));
-        if (jsonLines.length > 0) {
-          parsedOutput = JSON.parse(jsonLines[jsonLines.length - 1]);
-        }
-      } catch {
-        // Non-JSON output — extract what we can
-      }
-
-      // Extract metrics from OpenClaw's structured output
-      const metrics = {
-        inputTokens: parsedOutput?.usage?.input_tokens ?? parsedOutput?.usage?.inputTokens ?? 0,
-        outputTokens: parsedOutput?.usage?.output_tokens ?? parsedOutput?.usage?.outputTokens ?? 0,
-        llmCallCount: parsedOutput?.turns ?? parsedOutput?.steps ?? 0,
-        toolCallCount: parsedOutput?.tool_calls ?? parsedOutput?.toolCalls ?? 0,
-        toolErrorCount: parsedOutput?.tool_errors ?? parsedOutput?.toolErrors ?? 0,
-        totalTimeMs,
-        exitReason: code === 0 ? 'completed' : (code === null ? 'timeout' : 'error'),
-        provider: 'openclaw',
-        model: MODEL,
-      };
-
-      writeMetrics(metrics);
-
-      // Write output for the orchestrator
-      if (stdout) {
-        process.stdout.write(stdout);
-      }
-
-      if (code !== 0) {
-        console.error(`[openclaw] Process exited with code ${code}`);
-        if (stderr.includes('unknown option') || stderr.includes('Unknown flag')) {
-          console.error('[openclaw] Hint: OpenClaw CLI version may be incompatible. Try updating: npm install -g openclaw@latest');
-        }
-      }
-
-      // Cleanup temp workspace (best-effort)
-      try { rmSync(workspaceDir, { recursive: true, force: true }); } catch { /* ignore */ }
-
-      resolve(code ?? 1);
-    });
-
-    child.on('error', (err) => {
-      console.error(`[openclaw] Failed to spawn: ${err.message}`);
-      try { rmSync(workspaceDir, { recursive: true, force: true }); } catch { /* ignore */ }
-      reject(err);
-    });
-  });
-}
-
-// ── Main ─────────────────────────────────────────────────────────────
-
-const useOpenClawCli = isOpenClawInstalled();
-if (!useOpenClawCli) {
-  console.error('[openclaw] OpenClaw CLI not found. Install OpenClaw to run this harness.');
-  console.error('[openclaw] Use sandbox mode (`archal run ... --sandbox`) or install openclaw locally.');
-  process.exit(1);
-}
-
-console.error('[openclaw] Mode: native OpenClaw CLI');
-console.error(`[openclaw] Model: ${MODEL}`);
-console.error(`[openclaw] Task: ${TASK.slice(0, 200)}${TASK.length > 200 ? '...' : ''}`);
-
-const exitCode = await runWithOpenClawCli();
-process.exit(exitCode);

package/harnesses/openclaw/archal-harness.json

@@ -1,28 +0,0 @@
-{
-  "version": 1,
-  "name": "openclaw",
-  "description": "OpenClaw agent harness. Runs the real OpenClaw CLI against Archal twins; sandbox mode is the recommended path for production-fidelity evaluations.",
-  "defaultModel": "openclaw:main",
-  "promptFiles": [
-    "SOUL.md",
-    "AGENTS.md",
-    "TOOLS.md"
-  ],
-  "local": {
-    "command": "node",
-    "args": ["agent.mjs"]
-  },
-  "maxSteps": 80,
-  "supportedProviders": ["openclaw"],
-  "requiredEnvVars": [
-    "ARCHAL_ENGINE_TASK",
-    "ARCHAL_ENGINE_MODEL"
-  ],
-  "configDefaults": {
-    "maxSteps": 80,
-    "systemPrompt": true,
-    "errorHandling": true,
-    "retryOnTransient": true,
-    "maxConsecutiveErrors": 5
-  }
-}

package/harnesses/react/agent.mjs

@@ -1,233 +0,0 @@
-/**
- * ReAct Agent — the "good" bundled harness.
- *
- * Full ReAct (Reason + Act) loop with:
- * - Multi-provider support (Gemini, OpenAI, Anthropic)
- * - Structured system prompt encouraging step-by-step reasoning
- * - Error recovery with retries on transient failures
- * - Context-aware done detection
- * - Configurable step limit (default 80, cap 200 via ARCHAL_MAX_STEPS)
- * - Token usage and timing instrumentation
- *
- * Env vars (set by archal orchestrator):
- *   ARCHAL_ENGINE_TASK   — the scenario task to complete
- *   ARCHAL_ENGINE_MODEL  — model identifier (e.g. gemini-2.0-flash)
- *   ARCHAL_<TWIN>_URL    — twin REST base URL (per twin)
- *   ARCHAL_ENGINE_API_KEY / GEMINI_API_KEY / OPENAI_API_KEY / ANTHROPIC_API_KEY
- */
-import {
-  createHarnessContext,
-  runAgentLoop,
-  appendUserInstruction,
-} from '../_lib/harness-runner.mjs';
-import {
-  appendToolResults,
-} from '../_lib/providers.mjs';
-import { parseEnvInt } from '../_lib/env-utils.mjs';
-import { classifyTask, selectStepTools } from './tool-selection.mjs';
-
-const MAX_STEPS = parseEnvInt('ARCHAL_MAX_STEPS', 80, { min: 1, max: 200 });
-const MAX_CONSECUTIVE_ERRORS = parseEnvInt('ARCHAL_MAX_CONSECUTIVE_ERRORS', 8, { min: 1, max: 20 });
-const MAX_INITIAL_NO_TOOL_RECOVERIES = parseEnvInt('ARCHAL_MAX_INITIAL_NO_TOOL_RECOVERIES', 2, { min: 1, max: 5 });
-
-const SYSTEM_PROMPT = `You are a capable AI agent performing a task using tools. Think step by step.
-
-APPROACH:
-1. Before each action, briefly reason about what you need to do next and why.
-2. Use the available tools to gather information and take actions.
-3. If a tool call fails, analyze the error and try an alternative approach.
-4. When you have completed the task, stop calling tools.
-
-GUIDELINES:
-- Read the task carefully and identify all requirements before starting.
-- Work methodically — complete one sub-task before moving to the next.
-- Pay attention to tool output — it contains the information you need.
-- If you're unsure about something, gather more information first.
-- Do NOT repeat the same failed tool call — try a different approach.
-- Do not create new entities unless the task explicitly asks for creation.
-- Do not create or edit repository files as a substitute for issue, ticket, label, or message updates.
-- If the task spans multiple systems, do not stop after the first system mutation. Complete the required follow-up in every mentioned system.
-- When done, provide a brief summary of what you accomplished.`;
-
-const MUTATING_TOOL_NAME = /(?:^|_)(create|update|add|post|reply|delete|close|merge|approve|archive|send)(?:_|$)/i;
-const REPO_CONTENT_MUTATION_TOOL = /(?:^|_)(create_or_update_file|delete_file|create_branch|create_commit)(?:_|$)/i;
-const CREATE_ISSUE_TOOL = /(?:^|_)create_issue(?:_|$)/i;
-
-function isMutatingToolName(toolName) {
-  return MUTATING_TOOL_NAME.test(toolName);
-}
-
-const ctx = await createHarnessContext('react');
-const TASK_FLAGS = classifyTask(ctx.task);
-const TASK_LOWER = ctx.task.toLowerCase();
-const TASK_ALLOWS_REPO_CONTENT_MUTATION = /\b(file|files|code|commit|branch|pull request|pull requests|pr|readme|source|implementation|repository)\b/i.test(TASK_LOWER);
-
-const knownTwinNames = new Set(Object.keys(ctx.twinUrls));
-const updatedTwins = new Set();
-let mutatedTwinsThisStep = new Set();
-let pendingFollowupTwins = null;
-let repoContentGuardRecoveries = 0;
-
-await runAgentLoop(ctx, {
-  systemPrompt: SYSTEM_PROMPT,
-  maxSteps: MAX_STEPS,
-  useRetry: true,
-  retryCount: 4,
-  useTrace: true,
-  maxConsecutiveErrors: MAX_CONSECUTIVE_ERRORS,
-  maxInitialNoToolRecoveries: MAX_INITIAL_NO_TOOL_RECOVERIES,
-
-  initMessages(provider, messages) {
-    if (TASK_FLAGS.isExistingIssueTriage) {
-      return appendUserInstruction(
-        provider,
-        messages,
-        'This task is issue triage on the existing repository issues. Update those issues in place. ' +
-          'Do not use comments, files, or duplicate issues as a substitute for labels. ' +
-          'If the task asks you to prioritize bug reports, every bug issue must also receive an appropriate priority label. ' +
-          'Use the repository priority labels exactly as named: priority:high, priority:medium, or priority:low.',
-      );
-    }
-    return messages;
-  },
-
-  selectTools(_ctx, _state) {
-    return selectStepTools(ctx.allTools, TASK_FLAGS, ctx.toolToTwin, pendingFollowupTwins);
-  },
-
-  onBeforeToolExecution(_ctx, state, stepResult) {
-    const { toolCalls, thinking, text, iterDurationMs, step } = stepResult;
-
-    // Block repo content mutations when the task doesn't warrant them
-    const proposedRepoContentMutation = toolCalls.some((tc) => REPO_CONTENT_MUTATION_TOOL.test(tc.name));
-    if (proposedRepoContentMutation && (!TASK_ALLOWS_REPO_CONTENT_MUTATION || TASK_FLAGS.isExistingIssueTriage) && repoContentGuardRecoveries < 2) {
-      repoContentGuardRecoveries++;
-      if (state.agentTrace) {
-        state.agentTrace.addStep({
-          step,
-          thinking,
-          text,
-          toolCalls: toolCalls.map((tc) => ({ name: tc.name, arguments: tc.arguments })),
-          durationMs: iterDurationMs,
-        });
-      }
-      state.messages = appendToolResults(
-        ctx.provider,
-        state.messages,
-        toolCalls,
-        toolCalls.map(() =>
-          'Blocked by harness: this task must update the existing issue or message state directly, not repository files or commits.',
-        ),
-      );
-      state.messages = appendUserInstruction(
-        ctx.provider,
-        state.messages,
-        'This task is about updating existing issues/messages, not repository content. ' +
-          'Do not create or edit files or commits as a substitute for labels, issue state changes, or replies. ' +
-          'Use the issue or messaging mutation tools directly.',
-      );
-      ctx.log.info('repo_content_mutation_blocked', {
-        step,
-        attemptedTools: toolCalls.map((tc) => tc.name),
-      });
-      return 'continue';
-    }
-
-    // Block issue creation during triage tasks
-    if (TASK_FLAGS.isExistingIssueTriage && toolCalls.some((tc) => CREATE_ISSUE_TOOL.test(tc.name)) && repoContentGuardRecoveries < 2) {
-      repoContentGuardRecoveries++;
-      if (state.agentTrace) {
-        state.agentTrace.addStep({
-          step,
-          thinking,
-          text,
-          toolCalls: toolCalls.map((tc) => ({ name: tc.name, arguments: tc.arguments })),
-          durationMs: iterDurationMs,
-        });
-      }
-      state.messages = appendToolResults(
-        ctx.provider,
-        state.messages,
-        toolCalls,
-        toolCalls.map(() =>
-          'Blocked by harness: this task is to triage the existing issues in the repository, not create duplicate issues.',
-        ),
-      );
-      state.messages = appendUserInstruction(
-        ctx.provider,
-        state.messages,
-        'This task is to triage the existing issues that are already in the repository. ' +
-          'Do not create duplicate issues. Inspect the current issues and use the issue update tools to apply category labels and priority labels directly to those existing issues.',
-      );
-      ctx.log.info('issue_creation_blocked_for_triage', {
-        step,
-        attemptedTools: toolCalls.map((tc) => tc.name),
-      });
-      return 'continue';
-    }
-    // NOTE: Do NOT reset repoContentGuardRecoveries here. The counter must
-    // persist across the entire run so alternating clean/blocked steps cannot
-    // bypass the 2-attempt safety limit indefinitely.
-  },
-
-  onToolSuccess(tc) {
-    if (isMutatingToolName(tc.name)) {
-      const twinName = ctx.toolToTwin[tc.name]?.twinName;
-      if (twinName) {
-        updatedTwins.add(twinName);
-        mutatedTwinsThisStep.add(twinName);
-      }
-    }
-  },
-
-  onAfterToolExecution(_ctx, state, stepResult) {
-    const { step } = stepResult;
-    // Capture and reset per-step tracking (populated by onToolSuccess)
-    const stepMutations = mutatedTwinsThisStep;
-    mutatedTwinsThisStep = new Set();
-
-    // Clear pending followup if a pending twin was mutated
-    if (pendingFollowupTwins && pendingFollowupTwins.size > 0) {
-      const completedFollowups = [...stepMutations].filter((twin) => pendingFollowupTwins.has(twin));
-      if (completedFollowups.length > 0) {
-        pendingFollowupTwins = null;
-      }
-    }
-
-    // Trigger cross-system followup when the task spans multiple services
-    if (TASK_FLAGS.requiresCrossSystemFollowup && !pendingFollowupTwins && knownTwinNames.size > 1 && stepMutations.size > 0) {
-      const untouchedTwins = [...knownTwinNames].filter((twinName) => !updatedTwins.has(twinName));
-      if (untouchedTwins.length > 0) {
-        pendingFollowupTwins = new Set(untouchedTwins);
-        state.messages = appendUserInstruction(
-          ctx.provider,
-          state.messages,
-          `You have updated ${[...updatedTwins].join(', ')} but not ${untouchedTwins.join(', ')}. ` +
-            'Continue and finish the remaining required actions in the untouched system before you conclude.',
-        );
-        ctx.log.info('cross_system_followup_required', {
-          step,
-          updatedTwins: [...updatedTwins],
-          remainingTwins: untouchedTwins,
-        });
-      }
-    }
-  },
-
-  onNoToolCalls(_ctx, state, stepResult) {
-    if (pendingFollowupTwins && pendingFollowupTwins.size > 0) {
-      const remainingTwins = [...pendingFollowupTwins].join(', ');
-      state.messages = appendUserInstruction(
-        ctx.provider,
-        state.messages,
-        `You have not finished the required follow-up in ${remainingTwins}. ` +
-          'Continue using the remaining system tools until those actions are complete before you conclude.',
-      );
-      ctx.log.info('cross_system_followup_reprompt', {
-        step: stepResult.step,
-        remainingTwins,
-      });
-      return 'continue';
-    }
-  },
-});

package/harnesses/react/archal-harness.json

@@ -1,22 +0,0 @@
-{
-  "version": 1,
-  "name": "react",
-  "description": "Full ReAct (Reason + Act) harness. Step-by-step reasoning prompt, error recovery with retries, consecutive-error bailout. Recommended for production evaluations.",
-  "local": {
-    "command": "node",
-    "args": ["agent.mjs"]
-  },
-  "maxSteps": 50,
-  "supportedProviders": ["openai", "anthropic", "gemini"],
-  "requiredEnvVars": [
-    "ARCHAL_ENGINE_TASK",
-    "ARCHAL_ENGINE_MODEL"
-  ],
-  "configDefaults": {
-    "maxSteps": 50,
-    "systemPrompt": true,
-    "errorHandling": true,
-    "retryOnTransient": true,
-    "maxConsecutiveErrors": 5
-  }
-}

package/harnesses/react/package.json

@@ -1,12 +0,0 @@
-{
-  "name": "@archal/harness-react",
-  "version": "0.0.0",
-  "private": true,
-  "type": "module",
-  "scripts": {
-    "start": "node agent.mjs"
-  },
-  "dependencies": {
-    "@modelcontextprotocol/sdk": "^1.27.1"
-  }
-}

package/harnesses/react/tool-selection.mjs

@@ -1,66 +0,0 @@
-const ISSUE_TRIAGE_TOOL = /(?:^|_)(list_issues|get_issue|update_issue)(?:_|$)/i;
-const SLACK_CHANNEL_POST_TOOL = /(?:^|_)slack_post_message(?:_|$)/i;
-
-/**
- * Patterns that identify distinct service domains in task text.
- * Used to detect whether a task genuinely spans multiple systems.
- */
-const SERVICE_DOMAIN_PATTERNS = [
-  { name: 'github', pattern: /\b(github|pull request|pr\s*#\d|merge|branch|commit|repository|repo)\b/i },
-  { name: 'slack', pattern: /\b(slack|#\w[\w-]*|channel|thread|post\s+(?:a\s+)?(?:message|summary|update))\b/i },
-  { name: 'linear', pattern: /\b(linear|[A-Z]{2,5}-\d+)\b/ },
-  { name: 'jira', pattern: /\b(jira|sprint|epic|story|CHG-\d+)\b/i },
-  { name: 'stripe', pattern: /\b(stripe|payment|charge|refund|invoice|subscription)\b/i },
-  { name: 'supabase', pattern: /\b(supabase|database|table|row|query|migration)\b/i },
-];
-
-function countMentionedServiceDomains(taskText) {
-  const matched = new Set();
-  for (const { name, pattern } of SERVICE_DOMAIN_PATTERNS) {
-    if (pattern.test(taskText)) matched.add(name);
-  }
-  return matched.size;
-}
-
-export function classifyTask(task) {
-  const taskLower = task.toLowerCase();
-  return {
-    taskLower,
-    isExistingIssueTriage: /\ball open issues?\b/.test(taskLower)
-      || (/\bissues?\b/.test(taskLower)
-        && /\b(triage|prioriti[sz]e|categor(?:ize|ization)|classif(?:y|ication))\b/.test(taskLower)),
-    requiresThreadReply: /\bthread\b/.test(taskLower)
-      && /\b(reply|replies|respond|post back)\b/.test(taskLower),
-    requiresCrossSystemFollowup: countMentionedServiceDomains(task) >= 2,
-  };
-}
-
-export function getToolsForTwins(tools, twinNames, toolToTwin) {
-  if (!twinNames || twinNames.size === 0) return tools;
-  return tools.filter((tool) => twinNames.has(toolToTwin[tool.name]?.twinName));
-}
-
-function canPerformIssueTriage(tools) {
-  return tools.some((tool) => ISSUE_TRIAGE_TOOL.test(tool.name));
-}
-
-export function filterToolsForTask(tools, taskFlags, { enforceIssueTriageAllowlist = true } = {}) {
-  let filtered = tools;
-  if (taskFlags.isExistingIssueTriage && enforceIssueTriageAllowlist) {
-    filtered = filtered.filter((tool) => ISSUE_TRIAGE_TOOL.test(tool.name));
-  }
-  if (taskFlags.requiresThreadReply) {
-    filtered = filtered.filter((tool) => !SLACK_CHANNEL_POST_TOOL.test(tool.name));
-  }
-  return filtered;
-}
-
-export function selectStepTools(tools, taskFlags, toolToTwin, pendingFollowupTwins) {
-  const twinScopedTools = getToolsForTwins(tools, pendingFollowupTwins, toolToTwin);
-  return filterToolsForTask(twinScopedTools, taskFlags, {
-    // Follow-up routing is the harder constraint. If the scoped twin cannot
-    // satisfy the generic issue-triage allowlist, keep its reply/mutation tools
-    // available so the agent can finish the required cross-system work.
-    enforceIssueTriageAllowlist: !taskFlags.isExistingIssueTriage || canPerformIssueTriage(twinScopedTools),
-  });
-}

package/harnesses/zero-shot/agent.mjs

@@ -1,31 +0,0 @@
-/**
- * Zero-Shot Agent — the "medium" bundled harness.
- *
- * Sends the full task with all tools in one shot, minimal guidance.
- * - Multi-provider support (Gemini, OpenAI, Anthropic)
- * - Minimal system prompt — no reasoning encouragement
- * - Basic error handling (log and continue, no retry)
- * - Max 40 steps
- *
- * Env vars (set by archal orchestrator):
- *   ARCHAL_ENGINE_TASK   — the scenario task to complete
- *   ARCHAL_ENGINE_MODEL  — model identifier
- *   ARCHAL_<TWIN>_URL    — twin REST base URL (per twin)
- *   ARCHAL_ENGINE_API_KEY / GEMINI_API_KEY / OPENAI_API_KEY / ANTHROPIC_API_KEY
- */
-import { createHarnessContext, runAgentLoop } from '../_lib/harness-runner.mjs';
-import { parseEnvInt } from '../_lib/env-utils.mjs';
-
-const MAX_STEPS = 40;
-const MAX_INITIAL_NO_TOOL_RECOVERIES = parseEnvInt('ARCHAL_MAX_INITIAL_NO_TOOL_RECOVERIES', 2, { min: 1, max: 5 });
-
-const SYSTEM_PROMPT = 'Complete the task. Use the tools provided.';
-
-const ctx = await createHarnessContext('zero-shot');
-
-await runAgentLoop(ctx, {
-  systemPrompt: SYSTEM_PROMPT,
-  maxSteps: MAX_STEPS,
-  useTrace: true,
-  maxInitialNoToolRecoveries: MAX_INITIAL_NO_TOOL_RECOVERIES,
-});

package/harnesses/zero-shot/archal-harness.json

@@ -1,21 +0,0 @@
-{
-  "version": 1,
-  "name": "zero-shot",
-  "description": "Medium-quality harness. Minimal system prompt, basic error handling (log and continue), no retry. Good for testing model raw capability without agent scaffolding.",
-  "local": {
-    "command": "node",
-    "args": ["agent.mjs"]
-  },
-  "maxSteps": 40,
-  "supportedProviders": ["openai", "anthropic", "gemini"],
-  "requiredEnvVars": [
-    "ARCHAL_ENGINE_TASK",
-    "ARCHAL_ENGINE_MODEL"
-  ],
-  "configDefaults": {
-    "maxSteps": 40,
-    "systemPrompt": true,
-    "errorHandling": true,
-    "retryOnTransient": false
-  }
-}

package/harnesses/zero-shot/package.json

@@ -1,12 +0,0 @@
-{
-  "name": "@archal/harness-zero-shot",
-  "version": "0.0.0",
-  "private": true,
-  "type": "module",
-  "scripts": {
-    "start": "node agent.mjs"
-  },
-  "dependencies": {
-    "@modelcontextprotocol/sdk": "^1.27.1"
-  }
-}