@forwardimpact/libeval 0.1.20 → 0.1.22

package/bin/fit-trace.js

@@ -20,6 +20,9 @@ import {
   runReasoningCommand,
   runTimelineCommand,
   runStatsCommand,
+  runInitCommand,
+  runTurnCommand,
+  runFilterCommand,
 } from "../src/commands/trace.js";
 
 const { version: VERSION } = JSON.parse(
@@ -99,6 +102,10 @@ const definition = {
           type: "string",
           description: "Surrounding turns per hit (default: 0)",
         },
+        full: {
+          type: "boolean",
+          description: "Full content block in match descriptions",
+        },
       },
     },
     {
@@ -135,11 +142,45 @@ const definition = {
       args: "<file>",
       description: "Token usage and cost breakdown",
     },
+    {
+      name: "init",
+      args: "<file>",
+      description: "Full system/init event",
+    },
+    {
+      name: "turn",
+      args: "<file> <index>",
+      description: "Single turn by index",
+    },
+    {
+      name: "filter",
+      args: "<file>",
+      description: "Filter turns by structural properties",
+      options: {
+        role: {
+          type: "string",
+          description: "Turn role (system, user, assistant, tool_result)",
+        },
+        tool: {
+          type: "string",
+          description: "Tool name (matches assistant turns)",
+        },
+        error: {
+          type: "boolean",
+          description:
+            "Error tool_result turns only (flag-only; for non-errors use the API)",
+        },
+      },
+    },
   ],
   globalOptions: {
     help: { type: "boolean", short: "h", description: "Show this help" },
     version: { type: "boolean", description: "Show version" },
     json: { type: "boolean", description: "Output help as JSON" },
+    signatures: {
+      type: "boolean",
+      description: "Include thinking.signature blobs in output",
+    },
   },
   examples: [
     "fit-trace runs --lookback 7d",
@@ -149,6 +190,11 @@ const definition = {
     "fit-trace search structured.json 'error|fail' --context 1",
     "fit-trace tool structured.json Bash",
     "fit-trace batch structured.json 0 20",
+    "fit-trace init structured.json",
+    "fit-trace turn structured.json 3",
+    "fit-trace filter structured.json --role system",
+    "fit-trace filter structured.json --tool Bash --role assistant",
+    "fit-trace search structured.json 'error' --full",
   ],
 };
 
@@ -170,6 +216,9 @@ const COMMANDS = {
   reasoning: runReasoningCommand,
   timeline: runTimelineCommand,
   stats: runStatsCommand,
+  init: runInitCommand,
+  turn: runTurnCommand,
+  filter: runFilterCommand,
 };
 
 async function main() {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@forwardimpact/libeval",
-  "version": "0.1.20",
+  "version": "0.1.22",
   "description": "Process Claude Code stream-json output into structured traces",
   "license": "Apache-2.0",
   "author": "D. Olsson <hi@senzilla.io>",
@@ -25,15 +25,18 @@
     "node": ">=18.0.0"
   },
   "scripts": {
-    "test": "bun run node --test test/*.test.js"
+    "test": "bun test test/*.test.js"
   },
   "dependencies": {
-    "@anthropic-ai/claude-agent-sdk": "^0.2.112",
+    "@anthropic-ai/claude-agent-sdk": "0.2.112",
     "@forwardimpact/libcli": "^0.1.0",
     "@forwardimpact/libconfig": "^0.1.0",
     "@forwardimpact/libtelemetry": "^0.1.22",
     "zod": "^4.3.6"
   },
+  "devDependencies": {
+    "@forwardimpact/libharness": "^0.1.14"
+  },
   "publishConfig": {
     "access": "public"
   }

package/src/agent-runner.js

@@ -28,6 +28,7 @@ function applyDefaults(deps) {
     systemPrompt: deps.systemPrompt ?? null,
     disallowedTools: deps.disallowedTools ?? [],
     mcpServers: deps.mcpServers ?? null,
+    taskAmend: deps.taskAmend ?? null,
   };
 }
 
@@ -67,9 +68,12 @@ export class AgentRunner {
   async run(task) {
     const abortController = new AbortController();
     this.currentAbortController = abortController;
+    const effectiveTask = this.taskAmend
+      ? `${task}\n\n${this.taskAmend}`
+      : task;
     try {
       const iterator = this.query({
-        prompt: task,
+        prompt: effectiveTask,
         options: {
           cwd: this.cwd,
           allowedTools: this.allowedTools,

package/src/commands/facilitate.js

@@ -30,8 +30,7 @@ function parseFacilitateOptions(values) {
     throw new Error("--task-file or --task-text is required");
 
   const taskAmend = values["task-amend"] ?? undefined;
-  let taskContent = taskFile ? readFileSync(taskFile, "utf8") : taskText;
-  if (taskAmend) taskContent += `\n\n${taskAmend}`;
+  const taskContent = taskFile ? readFileSync(taskFile, "utf8") : taskText;
 
   const profilesRaw = values["agent-profiles"];
   if (!profilesRaw) throw new Error("--agent-profiles is required");
@@ -42,6 +41,7 @@ function parseFacilitateOptions(values) {
 
   return {
     taskContent,
+    taskAmend,
     agentConfigs,
     facilitatorCwd: resolve(values["facilitator-cwd"] ?? "."),
     model: values.model ?? "opus",
@@ -82,6 +82,7 @@ export async function runFacilitateCommand(values, _args) {
     model: opts.model,
     maxTurns: opts.maxTurns,
     facilitatorProfile: opts.facilitatorProfile,
+    taskAmend: opts.taskAmend,
   });
 
   const result = await facilitator.run(opts.taskContent);

package/src/commands/run.js

@@ -21,11 +21,11 @@ function parseRunOptions(values) {
 
   const maxTurnsRaw = values["max-turns"] ?? "50";
   const taskAmend = values["task-amend"] ?? undefined;
-  let taskContent = taskFile ? readFileSync(taskFile, "utf8") : taskText;
-  if (taskAmend) taskContent += `\n\n${taskAmend}`;
+  const taskContent = taskFile ? readFileSync(taskFile, "utf8") : taskText;
 
   return {
     taskContent,
+    taskAmend,
     cwd: resolve(values.cwd ?? "."),
     model: values.model ?? "opus",
     maxTurns: maxTurnsRaw === "0" ? 0 : parseInt(maxTurnsRaw, 10),
@@ -49,6 +49,7 @@ function parseRunOptions(values) {
 export async function runRunCommand(values, _args) {
   const {
     taskContent,
+    taskAmend,
     cwd,
     model,
     maxTurns,
@@ -94,6 +95,7 @@ export async function runRunCommand(values, _args) {
     onLine,
     settingSources: ["project"],
     systemPrompt,
+    taskAmend,
   });
 
   const result = await runner.run(taskContent);

package/src/commands/supervise.js

@@ -20,11 +20,11 @@ function parseSuperviseOptions(values) {
   const supervisorAllowedToolsRaw = values["supervisor-allowed-tools"];
 
   const taskAmend = values["task-amend"] ?? undefined;
-  let taskContent = taskFile ? readFileSync(taskFile, "utf8") : taskText;
-  if (taskAmend) taskContent += `\n\n${taskAmend}`;
+  const taskContent = taskFile ? readFileSync(taskFile, "utf8") : taskText;
 
   return {
     taskContent,
+    taskAmend,
     supervisorCwd: resolve(values["supervisor-cwd"] ?? "."),
     agentCwd: resolve(
       values["agent-cwd"] ?? mkdtempSync(join(tmpdir(), "fit-eval-agent-")),
@@ -83,6 +83,7 @@ export async function runSuperviseCommand(values, _args) {
     supervisorAllowedTools: opts.supervisorAllowedTools,
     supervisorProfile: opts.supervisorProfile,
     agentProfile: opts.agentProfile,
+    taskAmend: opts.taskAmend,
   });
 
   const result = await supervisor.run(opts.taskContent);

package/src/commands/trace.js

@@ -3,6 +3,7 @@ import { join } from "node:path";
 import { createTraceCollector } from "@forwardimpact/libeval";
 import { createTraceQuery } from "../trace-query.js";
 import { createTraceGitHub } from "../trace-github.js";
+import { stripSignatures } from "../signature-filter.js";
 
 // --- GitHub commands ---
 
@@ -20,7 +21,7 @@ export async function runRunsCommand(values, args, ctx) {
   const pattern = args[0] ?? "agent";
   const lookback = values.lookback ?? "7d";
   const runs = await gh.listRuns({ pattern, lookback });
-  writeJSON(runs);
+  writeJSON(runs, values);
 }
 
 /**
@@ -51,14 +52,14 @@ export async function runDownloadCommand(values, args, ctx) {
     result.files.push("structured.json");
   }
 
-  writeJSON(result);
+  writeJSON(result, values);
 }
 
 // --- Query commands ---
 
 /** @param {object} values @param {string[]} args - [file] */
 export async function runOverviewCommand(values, args) {
-  writeJSON(loadTrace(args[0]).overview());
+  writeJSON(loadTrace(args[0]).overview(), values);
 }
 
 /** @param {object} values @param {string[]} args - [file] */
@@ -70,48 +71,53 @@ export async function runCountCommand(values, args) {
 export async function runBatchCommand(values, args) {
   writeJSON(
     loadTrace(args[0]).batch(parseInt(args[1], 10), parseInt(args[2], 10)),
+    values,
   );
 }
 
 /** @param {object} values @param {string[]} args - [file, N?] */
 export async function runHeadCommand(values, args) {
   const n = args[1] ? parseInt(args[1], 10) : 10;
-  writeJSON(loadTrace(args[0]).head(n));
+  writeJSON(loadTrace(args[0]).head(n), values);
 }
 
 /** @param {object} values @param {string[]} args - [file, N?] */
 export async function runTailCommand(values, args) {
   const n = args[1] ? parseInt(args[1], 10) : 10;
-  writeJSON(loadTrace(args[0]).tail(n));
+  writeJSON(loadTrace(args[0]).tail(n), values);
 }
 
 /** @param {object} values @param {string[]} args - [file, pattern] */
 export async function runSearchCommand(values, args) {
   const limit = values.limit ? parseInt(values.limit, 10) : 50;
   const context = values.context ? parseInt(values.context, 10) : 0;
-  writeJSON(loadTrace(args[0]).search(args[1], { limit, context }));
+  const full = values.full ?? false;
+  writeJSON(
+    loadTrace(args[0]).search(args[1], { limit, context, full }),
+    values,
+  );
 }
 
 /** @param {object} values @param {string[]} args - [file] */
 export async function runToolsCommand(values, args) {
-  writeJSON(loadTrace(args[0]).toolFrequency());
+  writeJSON(loadTrace(args[0]).toolFrequency(), values);
 }
 
 /** @param {object} values @param {string[]} args - [file, name] */
 export async function runToolCommand(values, args) {
-  writeJSON(loadTrace(args[0]).tool(args[1]));
+  writeJSON(loadTrace(args[0]).tool(args[1]), values);
 }
 
 /** @param {object} values @param {string[]} args - [file] */
 export async function runErrorsCommand(values, args) {
-  writeJSON(loadTrace(args[0]).errors());
+  writeJSON(loadTrace(args[0]).errors(), values);
 }
 
 /** @param {object} values @param {string[]} args - [file] */
 export async function runReasoningCommand(values, args) {
   const from = values.from ? parseInt(values.from, 10) : undefined;
   const to = values.to ? parseInt(values.to, 10) : undefined;
-  writeJSON(loadTrace(args[0]).reasoning({ from, to }));
+  writeJSON(loadTrace(args[0]).reasoning({ from, to }), values);
 }
 
 /** @param {object} values @param {string[]} args - [file] */
@@ -122,7 +128,26 @@ export async function runTimelineCommand(values, args) {
 
 /** @param {object} values @param {string[]} args - [file] */
 export async function runStatsCommand(values, args) {
-  writeJSON(loadTrace(args[0]).stats());
+  writeJSON(loadTrace(args[0]).stats(), values);
+}
+
+/** @param {object} values @param {string[]} args - [file] */
+export async function runInitCommand(values, args) {
+  writeJSON(loadTrace(args[0]).init(), values);
+}
+
+/** @param {object} values @param {string[]} args - [file, index] */
+export async function runTurnCommand(values, args) {
+  writeJSON(loadTrace(args[0]).turn(parseInt(args[1], 10)), values);
+}
+
+/** @param {object} values @param {string[]} args - [file] */
+export async function runFilterCommand(values, args) {
+  const opts = {};
+  if (values.role) opts.role = values.role;
+  if (values.tool) opts.toolName = values.tool;
+  if (values.error) opts.isError = true;
+  writeJSON(loadTrace(args[0]).filter(opts), values);
 }
 
 // --- Shared helpers ---
@@ -151,7 +176,14 @@ function loadTrace(file) {
   return createTraceQuery(collector.toJSON());
 }
 
-/** @param {object} data */
-function writeJSON(data) {
-  process.stdout.write(JSON.stringify(data, null, 2) + "\n");
+/**
+ * Write JSON output to stdout. By default strips `thinking.signature`
+ * base64 blobs from the payload so they don't dominate terminal output;
+ * pass `--signatures` (surfaced as `values.signatures`) to keep them.
+ * @param {*} data
+ * @param {object} [values]
+ */
+function writeJSON(data, values = {}) {
+  const output = values.signatures ? data : stripSignatures(data);
+  process.stdout.write(JSON.stringify(output, null, 2) + "\n");
 }