@sweny-ai/core 0.1.132 → 0.1.134

package/dist/cli/comment-output.d.ts

@@ -0,0 +1,29 @@
+/**
+ * PR billboard: the markdown a CI runner posts as one sticky comment on a
+ * pull request. Pure and runner-agnostic; the CLI writes it with
+ * `--comment-file`, and any runner (the GitHub Action, GitLab, a script) posts it.
+ *
+ * METADATA ONLY. Node names, statuses, counts, and durations. Never prompts,
+ * tool inputs, node outputs, errors, or model prose.
+ */
+import type { ExecutionTrace, NodeResult, Workflow } from "../types.js";
+import { type RunSummary } from "./run-output.js";
+export declare const RUN_COMMENT_FOOTER = "Run with [SWEny](https://sweny.ai): `npx @sweny-ai/core new`";
+export interface RunCommentOptions {
+    trace?: ExecutionTrace;
+    /** Per-node wall-clock duration in ms, keyed by node id. Missing entries render as "-". */
+    durationsMs?: Record<string, number>;
+}
+/** Hidden marker the poster searches for to update the comment in place. */
+export declare function runCommentMarker(workflowId: string): string;
+export declare function formatRunComment(workflow: Workflow, results: Map<string, NodeResult>, summary: RunSummary, opts?: RunCommentOptions): string;
+/**
+ * Minimal comment for a run that crashed before producing node results: no DAG
+ * and no error text (thrown messages can embed model prose). Replaces any stale
+ * success comment for the same workflow.
+ */
+export declare function formatCrashComment(workflow: Workflow, summary: RunSummary): string;
+/** Write the comment markdown to `file`. Never throws; returns true only when written. */
+export declare function writeRunComment(file: string, workflow: Workflow, results: Map<string, NodeResult>, summary: RunSummary, opts?: RunCommentOptions & {
+    crashed?: boolean;
+}): boolean;

package/dist/cli/comment-output.js

@@ -0,0 +1,82 @@
+/**
+ * PR billboard: the markdown a CI runner posts as one sticky comment on a
+ * pull request. Pure and runner-agnostic; the CLI writes it with
+ * `--comment-file`, and any runner (the GitHub Action, GitLab, a script) posts it.
+ *
+ * METADATA ONLY. Node names, statuses, counts, and durations. Never prompts,
+ * tool inputs, node outputs, errors, or model prose.
+ */
+import fs from "node:fs";
+import path from "node:path";
+import { toMermaidBlock } from "../mermaid.js";
+import { STEP_SUMMARY_CLASS_DEFS, formatReceipt, formatReceiptDuration, nodeStates, } from "./run-output.js";
+export const RUN_COMMENT_FOOTER = "Run with [SWEny](https://sweny.ai): `npx @sweny-ai/core new`";
+/** Hidden marker the poster searches for to update the comment in place. */
+export function runCommentMarker(workflowId) {
+    const safe = workflowId.replace(/[^A-Za-z0-9._-]+/g, "_").replace(/-{2,}/g, "-") || "workflow";
+    return `<!-- sweny-run-comment:${safe} -->`;
+}
+/** Table-cell-safe text: no pipes, newlines, or backticks. */
+function cell(text) {
+    return text.replace(/\s+/g, " ").replace(/\|/g, "\\|").replace(/`/g, "'").trim();
+}
+const STATUS_LABEL = { success: "✅ success", failed: "❌ failed", skipped: "⏭️ skipped" };
+export function formatRunComment(workflow, results, summary, opts = {}) {
+    const state = nodeStates(results);
+    const rows = [...results].map(([id, r]) => {
+        const name = cell(workflow.nodes[id]?.name ?? id);
+        const ms = opts.durationsMs?.[id];
+        const dur = typeof ms === "number" && Number.isFinite(ms) ? formatReceiptDuration(ms) : "-";
+        return `| ${name} | ${STATUS_LABEL[r.status]} | ${dur} |`;
+    });
+    return [
+        runCommentMarker(workflow.id),
+        `## ${summary.ok ? "✅" : "❌"} ${cell(workflow.name)}`,
+        "",
+        `\`${formatReceipt(summary)}\``,
+        "",
+        toMermaidBlock(workflow, { state, trace: opts.trace, classDefs: STEP_SUMMARY_CLASS_DEFS }),
+        "",
+        "<details>",
+        "<summary>Nodes</summary>",
+        "",
+        "| Node | Status | Duration |",
+        "| --- | --- | --- |",
+        ...rows,
+        "",
+        "</details>",
+        "",
+        `<sub>${RUN_COMMENT_FOOTER}</sub>`,
+        "",
+    ].join("\n");
+}
+/**
+ * Minimal comment for a run that crashed before producing node results: no DAG
+ * and no error text (thrown messages can embed model prose). Replaces any stale
+ * success comment for the same workflow.
+ */
+export function formatCrashComment(workflow, summary) {
+    return [
+        runCommentMarker(workflow.id),
+        `## ❌ ${cell(workflow.name)}`,
+        "",
+        `\`${formatReceipt(summary)} · crashed\``,
+        "",
+        "The run stopped before finishing. See the job log for details.",
+        "",
+        `<sub>${RUN_COMMENT_FOOTER}</sub>`,
+        "",
+    ].join("\n");
+}
+/** Write the comment markdown to `file`. Never throws; returns true only when written. */
+export function writeRunComment(file, workflow, results, summary, opts = {}) {
+    try {
+        fs.mkdirSync(path.dirname(file), { recursive: true });
+        fs.writeFileSync(file, opts.crashed ? formatCrashComment(workflow, summary) : formatRunComment(workflow, results, summary, opts));
+        return true;
+    }
+    catch (err) {
+        process.stderr.write(`  ⚠ could not write comment file: ${err instanceof Error ? err.message : err}\n`);
+        return false;
+    }
+}

package/dist/cli/e2e.js

@@ -10,7 +10,7 @@ import * as p from "@clack/prompts";
 import chalk from "chalk";
 import { parse as parseYaml, stringify as stringifyYaml } from "yaml";
 import { execute } from "../executor.js";
-import { ClaudeClient } from "../claude.js";
+import { createHarness } from "../harness/index.js";
 import { createSkillMap } from "../skills/index.js";
 import { configuredSkills } from "../skills/custom-loader.js";
 import { consoleLogger } from "../types.js";
@@ -906,7 +906,7 @@ export async function runE2eRun(options) {
         }
         // Build skills + Claude client
         const skills = createSkillMap(configuredSkills(process.env, cwd));
-        const claude = new ClaudeClient({
+        const claude = createHarness("claude-code", {
             maxTurns: 80,
             cwd,
             logger: consoleLogger,

package/dist/cli/main.d.ts

@@ -5,6 +5,7 @@ export declare function workflowRunAction(file: string | undefined, options: Rec
     json?: boolean;
     stream?: boolean;
     mermaid?: boolean;
+    commentFile?: string;
     verbose?: boolean;
     timeout?: string;
     maxSteps?: string;

package/dist/cli/main.js

@@ -9,7 +9,7 @@ import chalk from "chalk";
 import { execute } from "../executor.js";
 import { triageWorkflow, implementWorkflow, seedContentWorkflow } from "../workflows/index.js";
 import { consoleLogger } from "../types.js";
-import { ClaudeClient } from "../claude.js";
+import { createHarness } from "../harness/index.js";
 import { builtinSkills, createSkillMap, validateWorkflowSkills } from "../skills/index.js";
 import { formatMissingSkillLines, skillEnvWarnings } from "./skill-env.js";
 import { configuredSkills, configuredSkillsWithDiagnostics } from "../skills/custom-loader.js";
@@ -32,6 +32,7 @@ import { registerRunsCommand } from "./runs.js";
 import { runE2eRun, runWithWallClockBudget, DEFAULT_WORKFLOW_TIMEOUT_MS } from "./e2e.js";
 import { createVerboseToolObserver } from "./verbose-observer.js";
 import { createRunLogger, renderReceiptLine, summarizeRun, writeStepSummary, WORKFLOW_RUN_DESCRIPTION, WORKFLOW_RUN_OPTIONS, } from "./run-output.js";
+import { writeRunComment } from "./comment-output.js";
 import { registerTriageCommand, registerImplementCommand, parseCliInputs, validateInputs, validateWarnings, parsePositiveInt, } from "./config.js";
 import { c, formatBanner, getStepDetails, formatStepLine, formatDagResultHuman, formatDagResultMarkdown, formatResultJson, formatValidationErrors, formatCrashError, formatCheckResults, } from "./output.js";
 import { checkProviderConnectivity, detectClaudeCodeLogin, discoverWorkflowSkillIds, validateCheckInputs, } from "./check.js";
@@ -202,7 +203,7 @@ triageCmd.action(async (options) => {
     const skills = createSkillMap(triageSkillDiscovery.skills);
     const mcpAutoConfig = buildMcpAutoConfig(config);
     const mcpServers = buildAutoMcpServers(mcpAutoConfig);
-    const claude = new ClaudeClient({
+    const claude = createHarness("claude-code", {
         maxTurns: config.maxInvestigateTurns || 50,
         cwd: process.cwd(),
         logger: consoleLogger,
@@ -512,7 +513,7 @@ implementCmd.action(async (issueId, options) => {
     const skills = createSkillMap(implementSkillDiscovery.skills);
     const mcpAutoConfig = buildMcpAutoConfig(config);
     const mcpServers = buildAutoMcpServers(mcpAutoConfig);
-    const claude = new ClaudeClient({
+    const claude = createHarness("claude-code", {
         maxTurns: config.maxImplementTurns || 40,
         cwd: process.cwd(),
         logger: consoleLogger,
@@ -756,7 +757,7 @@ export async function workflowRunAction(file, options) {
     });
     // Raw [info]/[debug] lines are verbose-only; warnings are rendered (#383).
     const runLogger = createRunLogger({ verbose: Boolean(options.verbose), tty: isTTY });
-    const claude = new ClaudeClient({
+    const claude = createHarness("claude-code", {
         maxTurns: config.maxInvestigateTurns || 50,
         cwd: process.cwd(),
         logger: runLogger,
@@ -948,6 +949,14 @@ export async function workflowRunAction(file, options) {
         catch {
             // silent
         }
+        // PR billboard markdown (metadata only). Written before any early exit so
+        // --json runs and failed runs still get a comment.
+        if (options.commentFile) {
+            writeRunComment(options.commentFile, workflow, results, summarizeRun(results, wfDurationMs), {
+                trace,
+                durationsMs: Object.fromEntries(nodeTimer.durations),
+            });
+        }
         if (isJson) {
             process.stdout.write(JSON.stringify(Object.fromEntries(results), null, 2) + "\n");
             process.exit(wfHasFailed ? 1 : 0);
@@ -986,6 +995,12 @@ export async function workflowRunAction(file, options) {
         runLogger.flush();
         recordHistory(nodeTimer.lastResults, undefined, true);
         console.error(`  ${renderReceiptLine(summarizeRun(new Map(), Date.now() - runStart, true), isTTY)}\n`);
+        // A crash must not leave a stale success comment behind.
+        if (options.commentFile) {
+            writeRunComment(options.commentFile, workflow, new Map(), summarizeRun(new Map(), Date.now() - runStart, true), {
+                crashed: true,
+            });
+        }
         // Finalize the cloud run as failed (covers thrown errors, incl.
         // RouteEvaluationError). Without this a crashed workflow run stays
         // "running" in cloud forever.
@@ -1086,7 +1101,7 @@ workflowCmd
         console.warn("\x1B[33m  ⚠  `sweny workflow create` is deprecated. Use `sweny new` instead.\x1B[0m\n");
     }
     const skills = configuredSkills();
-    const claude = new ClaudeClient({
+    const claude = createHarness("claude-code", {
         maxTurns: 3,
         cwd: process.cwd(),
         logger: consoleLogger,
@@ -1153,7 +1168,7 @@ workflowCmd
         return;
     }
     const skills = configuredSkills();
-    const claude = new ClaudeClient({
+    const claude = createHarness("claude-code", {
         maxTurns: 3,
         cwd: process.cwd(),
         logger: consoleLogger,

package/dist/cli/new.js

@@ -14,7 +14,7 @@ import chalk from "chalk";
 import { parse as parseYaml, stringify as stringifyYaml } from "yaml";
 import { WORKFLOW_TEMPLATES } from "./templates.js";
 import { buildWorkflow, refineWorkflow } from "../workflow-builder.js";
-import { ClaudeClient } from "../claude.js";
+import { createHarness } from "../harness/index.js";
 import { consoleLogger } from "../types.js";
 import { configuredSkills } from "../skills/custom-loader.js";
 import { builtinSkills } from "../skills/index.js";
@@ -541,7 +541,7 @@ export async function runNew(options) {
         const { installMarketplaceWorkflow } = await import("./marketplace.js");
         const allSkills = configuredSkills(process.env, cwd);
         const hasAgent = !!process.env.ANTHROPIC_API_KEY;
-        const claude = hasAgent ? new ClaudeClient({ maxTurns: 3, cwd, logger: consoleLogger }) : null;
+        const claude = hasAgent ? createHarness("claude-code", { maxTurns: 3, cwd, logger: consoleLogger }) : null;
         if (!options.skipIntro)
             p.intro(`Installing ${options.marketplaceId} from marketplace`);
         // Pre-compute source-control from git-remote (same method as the wizard path).
@@ -833,7 +833,7 @@ async function runCustomWorkflowBuilder(skills) {
     });
     if (p.isCancel(description))
         return null;
-    const claude = new ClaudeClient({
+    const claude = createHarness("claude-code", {
         maxTurns: 3,
         cwd: process.cwd(),
         logger: consoleLogger,

package/dist/cli/run-history.d.ts

@@ -48,6 +48,11 @@ export interface RunRecord {
         to: string;
     }>;
     totals: RunTotals;
+    /** Harness that ran the nodes (id + version). Absent for runs with no harness-tagged result. */
+    harness?: {
+        id: string;
+        version: string;
+    };
 }
 /** sha256 (hex) of the workflow as canonical JSON: keys sorted, undefined dropped. */
 export declare function hashWorkflow(workflow: Workflow): string;

package/dist/cli/run-history.js

@@ -91,6 +91,7 @@ export function buildRunRecord(i) {
             retries: retries.get(id) ?? 0,
         };
     });
+    const harness = [...i.results.values()].find((r) => r.harness)?.harness;
     return {
         schema_version: RUN_HISTORY_SCHEMA_VERSION,
         run_id: i.runId,
@@ -109,6 +110,7 @@ export function buildRunRecord(i) {
             tokens: summary.tokens ?? null,
             cost_usd: summary.costUsd ?? null,
         },
+        ...(harness ? { harness: { id: harness.id, version: harness.version } } : {}),
     };
 }
 // ── Write, prune, read ──────────────────────────────────────────

package/dist/cli/run-output.d.ts

@@ -6,6 +6,7 @@
  * model prose: only counts, durations, token totals, and cost.
  */
 import { type ExecutionTrace, type Logger, type NodeResult, type Workflow } from "../types.js";
+import { type NodeStatus } from "../mermaid.js";
 export interface RunSummary {
     ok: boolean;
     nodesOk: number;
@@ -31,6 +32,12 @@ export declare function formatCost(usd: number): string;
 export declare function formatReceipt(s: RunSummary): string;
 /** Receipt for a terminal: colored only when `color` is true. */
 export declare function renderReceiptLine(s: RunSummary, color: boolean): string;
+/** Brand blue (blue-600 / blue-700), not indigo. Applied to successful nodes. */
+export declare const STEP_SUMMARY_CLASS_DEFS: {
+    readonly success: "fill:#2563eb,stroke:#1d4ed8,color:#fff,stroke-width:2px";
+};
+/** Per-node Mermaid status from run results. */
+export declare function nodeStates(results: Map<string, NodeResult>): Record<string, NodeStatus>;
 /** Markdown for `$GITHUB_STEP_SUMMARY`: receipt + status-colored Mermaid DAG. */
 export declare function formatStepSummary(workflow: Workflow, results: Map<string, NodeResult>, summary: RunSummary, trace?: ExecutionTrace): string;
 /**

package/dist/cli/run-output.js

@@ -94,15 +94,20 @@ export function renderReceiptLine(s, color) {
 }
 // ── Step summary ────────────────────────────────────────────────
 /** Brand blue (blue-600 / blue-700), not indigo. Applied to successful nodes. */
-const STEP_SUMMARY_CLASS_DEFS = {
+export const STEP_SUMMARY_CLASS_DEFS = {
     success: "fill:#2563eb,stroke:#1d4ed8,color:#fff,stroke-width:2px",
 };
-/** Markdown for `$GITHUB_STEP_SUMMARY`: receipt + status-colored Mermaid DAG. */
-export function formatStepSummary(workflow, results, summary, trace) {
+/** Per-node Mermaid status from run results. */
+export function nodeStates(results) {
     const state = {};
     for (const [id, r] of results) {
         state[id] = r.status === "success" ? "success" : r.status === "failed" ? "failed" : "skipped";
     }
+    return state;
+}
+/** Markdown for `$GITHUB_STEP_SUMMARY`: receipt + status-colored Mermaid DAG. */
+export function formatStepSummary(workflow, results, summary, trace) {
+    const state = nodeStates(results);
     return [
         `## ${summary.ok ? "✅" : "❌"} ${workflow.name}`,
         "",
@@ -185,5 +190,9 @@ export const WORKFLOW_RUN_OPTIONS = [
         "Show raw log lines and each tool call's input and output inline (human-readable, truncated). Use --stream for full untruncated NDJSON.",
     ],
     ["--mermaid", "Output a Mermaid diagram with execution state after run"],
+    [
+        "--comment-file <path>",
+        "Write a PR-comment markdown (run receipt, status-colored DAG, per-node table; metadata only) to <path> so any CI can post it",
+    ],
     ["--input <json>", "JSON string of input data to pass to the workflow"],
 ];

package/dist/eval/index.js

@@ -8,7 +8,10 @@ import { evaluateFunctionRule } from "./function.js";
 import { evaluateJudge } from "./judge.js";
 /** Reasoning cap, matching the spec's ~500-character ceiling. */
 const MAX_REASONING_LENGTH = 500;
-/** Default judge model when neither evaluator nor node nor workflow specify one. */
+/**
+ * Judge model when neither evaluator, node, workflow nor the harness names one.
+ * Harnesses declare their own via `defaultJudgeModel` (Claude Code: this value).
+ */
 const DEFAULT_JUDGE_MODEL = "claude-haiku-4-5";
 function capReasoning(reasoning) {
     if (!reasoning)
@@ -17,8 +20,8 @@ function capReasoning(reasoning) {
         return reasoning;
     return reasoning.slice(0, MAX_REASONING_LENGTH - 1) + "…";
 }
-function resolveJudgeModel(evaluator, node, workflow) {
-    return evaluator.model ?? node?.judge_model ?? workflow?.judge_model ?? DEFAULT_JUDGE_MODEL;
+function resolveJudgeModel(evaluator, node, workflow, harnessDefault) {
+    return evaluator.model ?? node?.judge_model ?? workflow?.judge_model ?? harnessDefault ?? DEFAULT_JUDGE_MODEL;
 }
 /**
  * Run every declared evaluator against the node's result and return one
@@ -56,7 +59,7 @@ export async function evaluateAll(evaluators, result, opts = {}) {
             if (!opts.claude) {
                 throw new Error(`evaluator '${e.name}' is kind: judge but no Claude client was provided to evaluateAll`);
             }
-            const model = resolveJudgeModel(e, opts.node, opts.workflow);
+            const model = resolveJudgeModel(e, opts.node, opts.workflow, opts.claude.defaultJudgeModel);
             const verdict = await evaluateJudge(e, result, opts.claude, { model });
             out.push({
                 name: e.name,

package/dist/executor.d.ts

@@ -13,13 +13,22 @@
  * This replaces ~8k lines of engine + recipe step code.
  */
 import type { Workflow, Skill, Tool, Claude, Observer, Logger, ExecutionResult } from "./types.js";
+import type { AgentHarness } from "./harness/types.js";
 export interface ExecuteOptions {
     /** Registered skills (id → Skill) */
     skills: Map<string, Skill>;
     /** Config values — env vars + explicit overrides */
     config?: Record<string, string>;
-    /** Claude client */
-    claude: Claude;
+    /**
+     * The agent harness that runs nodes (`createHarness("claude-code")`).
+     * Required unless the deprecated `claude` option is given.
+     */
+    harness?: AgentHarness;
+    /**
+     * Claude client.
+     * @deprecated Pass `harness`. Still accepted (see `claudeCompat()`); when both are set, `harness` wins.
+     */
+    claude?: Claude;
     /** Event observer for streaming/logging */
     observer?: Observer;
     /** Logger */

package/dist/executor.js

@@ -22,6 +22,7 @@ import { buildToolAliases } from "./skills/index.js";
 import { validateWorkflow } from "./schema.js";
 import { resolveAgentAccess } from "./agent-env.js";
 import { fenceUntrusted } from "./untrusted.js";
+import { asClaude } from "./harness/compat.js";
 /**
  * Default workflow-level step budget. Counts total node executions across the
  * whole run (loop iterations included). Generous enough for legitimate deep
@@ -47,7 +48,12 @@ function throwIfAborted(signal) {
  * - `trace`: full ordered execution trace including loops and routing decisions
  */
 export async function execute(workflow, input, options) {
-    const { claude, observer, signal, timeoutMs } = options;
+    const { observer, signal, timeoutMs } = options;
+    // `harness` is the seam; a legacy `claude` object is used as-is (what `asClaude(claudeCompat(claude))` yields).
+    const claude = options.harness ? asClaude(options.harness) : options.claude;
+    if (!claude) {
+        throw new Error("execute() needs options.harness (or the deprecated options.claude)");
+    }
     // If the caller already aborted before we started, fail fast.
     throwIfAborted(signal);
     const dryRun = isDryRunInput(input);

package/dist/harness/capabilities.d.ts

@@ -0,0 +1,7 @@
+/**
+ * Declared capabilities per adapter. No runtime imports, so the executor can use
+ * them without pulling in an agent SDK.
+ */
+import type { HarnessCapabilities } from "./types.js";
+/** Claude Code enforces every sweny opinion natively, so `policyGate` never degrades or refuses it. */
+export declare const CLAUDE_CODE_CAPABILITIES: HarnessCapabilities;

package/dist/harness/capabilities.js

@@ -0,0 +1,17 @@
+/**
+ * Declared capabilities per adapter. No runtime imports, so the executor can use
+ * them without pulling in an agent SDK.
+ */
+/** Claude Code enforces every sweny opinion natively, so `policyGate` never degrades or refuses it. */
+export const CLAUDE_CODE_CAPABILITIES = {
+    structuredOutput: "native",
+    toolTrace: "full",
+    builtinDeny: "by-name",
+    mcp: { inject: true, exclusive: "native" },
+    sandbox: { fs: true, network: true },
+    readOnly: "native",
+    turnLimit: "native",
+    usage: { tokens: true, costUsd: true, live: false },
+    cancel: "signal",
+    resume: false,
+};