@holoscript/holoscript-agent 2.0.0 → 2.0.2

package/dist/runner.js

@@ -37,7 +37,18 @@ function brainClassOf(brain) {
   return "unknown";
 }
 function buildCaelRecord(input) {
-  const { identity, brain, task, messages, finalText, usage, costUsd, spentUsd, prevChain, runtimeVersion } = input;
+  const {
+    identity,
+    brain,
+    task,
+    messages,
+    finalText,
+    usage,
+    costUsd,
+    spentUsd,
+    prevChain,
+    runtimeVersion
+  } = input;
   const l0 = sha(brain.systemPrompt);
   const l1 = sha(`${task.id}|${task.title}|${task.description ?? ""}`);
   const l2 = sha(JSON.stringify(messages));
@@ -53,15 +64,16 @@ function buildCaelRecord(input) {
     prev_hash: prevChain,
     fnv1a_chain,
     version_vector_fingerprint: `agent@${runtimeVersion}|brain@${brainClassOf(brain)}|provider@${identity.llmProvider}|model@${identity.llmModel}`,
-    brain_class: brainClassOf(brain)
+    brain_class: brainClassOf(brain),
+    trust_epoch: "post-w107"
   };
 }
 
 // src/tools.ts
 import { readFile, writeFile, readdir, mkdir, stat } from "fs/promises";
-import { resolve, dirname } from "path";
+import { resolve, dirname, delimiter, isAbsolute, sep } from "path";
 import { spawn } from "child_process";
-var ALLOWED_READ_ROOTS = [
+var FLEET_READ_ROOTS = [
   "/root/msc-paper-22",
   // Paper 22 mechanization inputs (scp'd by deploy)
   "/root/holoscript-mesh",
@@ -69,15 +81,24 @@ var ALLOWED_READ_ROOTS = [
   "/root/agent-output"
   // Read back what we wrote
 ];
-var ALLOWED_WRITE_ROOTS = [
+var FLEET_WRITE_ROOTS = [
   "/root/agent-output"
   // Single write sink — keeps deliverables in one place
 ];
-var BASH_WHITELIST = [
-  "lake build",
-  "lake env",
-  "lake clean",
-  "lean ",
+function parseRootsEnv(raw, fallback) {
+  if (!raw) return fallback;
+  const roots = raw.split(delimiter).map((r) => r.trim()).filter((r) => r.length > 0 && isAbsolute(r));
+  return roots.length > 0 ? roots : fallback;
+}
+var ALLOWED_READ_ROOTS = parseRootsEnv(
+  process.env.HOLOSCRIPT_AGENT_READ_ROOTS,
+  FLEET_READ_ROOTS
+);
+var ALLOWED_WRITE_ROOTS = parseRootsEnv(
+  process.env.HOLOSCRIPT_AGENT_WRITE_ROOTS,
+  FLEET_WRITE_ROOTS
+);
+var BASH_READ_ONLY_PREFIXES = [
   "ls ",
   "ls\n",
   "ls$",
@@ -92,16 +113,36 @@ var BASH_WHITELIST = [
   "git log",
   "git diff",
   "git show",
+  "pwd",
+  "echo ",
+  "lake env"
+];
+var BASH_PRODUCTIVE_PREFIXES = [
+  "lake build",
+  "lake clean",
+  "lean ",
   "pnpm --filter",
   "pnpm vitest",
   "vitest run",
-  "pwd",
-  "echo "
+  // Robotics / edge-node (Jetson) productive commands — without these, every
+  // ros2/colcon/tegrastats task fails the W.107 artifact gate and is abandoned
+  // as no-artifact. (jetson-orin-01 lane.)
+  "ros2 launch",
+  "ros2 topic pub",
+  "ros2 service call",
+  "colcon build",
+  "tegrastats"
 ];
+var BASH_WHITELIST = [...BASH_READ_ONLY_PREFIXES, ...BASH_PRODUCTIVE_PREFIXES];
+function isProductiveBashCommand(cmd) {
+  const trimmed = String(cmd ?? "").trim();
+  if (!trimmed) return false;
+  return BASH_PRODUCTIVE_PREFIXES.some((prefix) => trimmed.startsWith(prefix.trim()));
+}
 var MESH_TOOLS = [
   {
     name: "read_file",
-    description: "Read a file from the agent sandbox. Allowed roots: /root/msc-paper-22, /root/holoscript-mesh, /root/agent-output. Returns the file content as text. Use this to inspect inputs scp'd to the instance (e.g. MSC/Invariants.lean).",
+    description: `Read a file from the agent sandbox. Allowed roots: ${ALLOWED_READ_ROOTS.join(", ")}. Returns the file content as text. Use this to inspect task inputs and the read-only repo view.`,
     input_schema: {
       type: "object",
       properties: {
@@ -123,11 +164,11 @@ var MESH_TOOLS = [
   },
   {
     name: "write_file",
-    description: "Write a file to /root/agent-output/. This is the deliverable sink \u2014 anything you want to emit as task output (a Lean proof, a markdown report, a JSON dataset) goes here. Creates parent directories. Will refuse paths outside the write root.",
+    description: `Write a file to the deliverable sink (write roots: ${ALLOWED_WRITE_ROOTS.join(", ")}). Anything you want to emit as task output (a Lean proof, a markdown report, a JSON dataset, a .holo scene) goes here. Creates parent directories. Will refuse paths outside the write root(s).`,
     input_schema: {
       type: "object",
       properties: {
-        path: { type: "string", description: "Absolute path under /root/agent-output/" },
+        path: { type: "string", description: `Absolute path under a write root: ${ALLOWED_WRITE_ROOTS.join(", ")}` },
         content: { type: "string", description: "File content to write (UTF-8)" }
       },
       required: ["path", "content"]
@@ -135,7 +176,7 @@ var MESH_TOOLS = [
   },
   {
     name: "bash",
-    description: "Run a shell command. Whitelisted prefixes only: lake build, lean, ls, cat, grep, find, wc, head, tail, git status/log/diff/show, pnpm --filter, vitest run, pwd, echo. Hard 60s wall timeout, 1MB stdout cap. Use for lake build / lean kernel-checks, git inspection, repo greps. Refuses rm, curl, ssh, sudo, eval.",
+    description: "Run a shell command. Whitelisted prefixes only: lake build, lean, ls, cat, grep, find, wc, head, tail, git status/log/diff/show, pnpm --filter, vitest run, pwd, echo, ros2 launch/topic/service, colcon build, tegrastats. Hard 60s wall timeout, 1MB stdout cap. Use for builds, tests, hardware probes. Refuses rm, curl, ssh, sudo, eval.",
     input_schema: {
       type: "object",
       properties: {
@@ -144,22 +185,52 @@ var MESH_TOOLS = [
       },
       required: ["cmd"]
     }
+  },
+  {
+    name: "emit_hardware_receipt",
+    description: "Emit a portable hardware receipt (PortableHardwareReceiptMetadata v1) capturing device identity, runtime, and measured performance. Writes a JSON receipt to the agent output dir. Use after running tegrastats or colcon build to record hardware evidence for the CAEL audit chain. Accepts either pre-parsed measurements or raw tegrastats output (the tool parses it automatically).",
+    input_schema: {
+      type: "object",
+      properties: {
+        device_kind: {
+          type: "string",
+          description: 'Device identifier, e.g. "jetson-orin-nano-super", "raspberry-pi-5"'
+        },
+        accelerator: {
+          description: 'Accelerator string, e.g. "NVIDIA CUDA 8.7", or null for CPU-only'
+        },
+        runtime_name: { type: "string", description: 'Inference runtime, e.g. "Ollama", "llama.cpp"' },
+        runtime_version: { type: "string", description: 'Runtime version, e.g. "0.30.8"' },
+        host_os: { type: "string", description: 'OS + firmware, e.g. "JetPack 6.2.1 / Ubuntu 22.04"' },
+        composition_id: { type: "string", description: 'Brain composition reference, e.g. "jetson-orin-brain"' },
+        measurements: {
+          type: "array",
+          description: "Pre-parsed measurements. Each item: {metric: string, value: number, unit: string}",
+          items: { type: "object" }
+        },
+        tegrastats_output: {
+          type: "string",
+          description: "Raw tegrastats output line(s) \u2014 tool auto-parses GPU%, RAM, temp, power"
+        }
+      },
+      required: ["device_kind", "runtime_name", "runtime_version", "host_os"]
+    }
   }
 ];
 function isUnderRoot(absPath, root) {
   const resolved = resolve(absPath);
   const rootResolved = resolve(root);
-  return resolved === rootResolved || resolved.startsWith(rootResolved + "/");
+  return resolved === rootResolved || resolved.startsWith(rootResolved + sep);
 }
 function checkReadAllowed(path) {
-  if (!path.startsWith("/")) return `path must be absolute, got "${path}"`;
+  if (!isAbsolute(path)) return `path must be absolute, got "${path}"`;
   for (const root of ALLOWED_READ_ROOTS) {
     if (isUnderRoot(path, root)) return null;
   }
   return `read denied \u2014 path "${path}" not under allowed roots: ${ALLOWED_READ_ROOTS.join(", ")}`;
 }
 function checkWriteAllowed(path) {
-  if (!path.startsWith("/")) return `path must be absolute, got "${path}"`;
+  if (!isAbsolute(path)) return `path must be absolute, got "${path}"`;
   for (const root of ALLOWED_WRITE_ROOTS) {
     if (isUnderRoot(path, root)) return null;
   }
@@ -214,12 +285,113 @@ async function runTool(use) {
       return result.code === 0 ? okResult(use.id, result.stdout) : errResult(use.id, `exit=${result.code}
 ${result.stderr || result.stdout}`);
     }
+    if (use.name === "emit_hardware_receipt") {
+      const deviceKind = String(use.input.device_kind ?? "unknown-device");
+      const accelerator = use.input.accelerator === null || use.input.accelerator === "null" ? null : String(use.input.accelerator ?? "").trim() || null;
+      const runtimeName = String(use.input.runtime_name ?? "Ollama");
+      const runtimeVersion = String(use.input.runtime_version ?? "unknown");
+      const hostOs = String(use.input.host_os ?? "unknown");
+      const compositionId = String(use.input.composition_id ?? "unknown");
+      let measurements = [];
+      if (Array.isArray(use.input.measurements)) {
+        for (const m of use.input.measurements) {
+          const metric = String(m.metric ?? "");
+          const value = Number(m.value ?? 0);
+          const unit = String(m.unit ?? "");
+          if (metric && Number.isFinite(value)) {
+            measurements.push({ metric, value, unit, method: "measured" });
+          }
+        }
+      }
+      if (typeof use.input.tegrastats_output === "string" && use.input.tegrastats_output.length > 0) {
+        measurements = [...measurements, ...parseTegrastats(use.input.tegrastats_output)];
+      }
+      if (measurements.length === 0) {
+        measurements.push({ metric: "agent-tick", value: 1, unit: "count", method: "presence" });
+      }
+      const capturedAt = (/* @__PURE__ */ new Date()).toISOString();
+      const receipt = {
+        schemaVersion: "holoscript.hardware-receipt-metadata.v1",
+        target: {
+          id: `${deviceKind}-${Date.now()}`,
+          kind: deviceKind,
+          architecture: /jetson|orin|nano|agx|xavier/i.test(deviceKind) ? "arm64" : "unknown",
+          artifactKind: "measurement-trace"
+        },
+        device: {
+          vendor: /jetson|orin|nvidia/i.test(deviceKind) ? "nvidia" : "unknown",
+          model: deviceKind,
+          accelerator
+        },
+        runtime: { name: runtimeName, version: runtimeVersion, hostOS: hostOs },
+        compilerVersion: "holoscript-agent-1.0.0",
+        constraints: [],
+        measuredResults: measurements,
+        replayInputs: [
+          { kind: "composition-ref", uri: `compositions/${compositionId}`, sha256: "unknown" }
+        ],
+        provenance: {
+          capturedAt,
+          sourceCompositionHash: compositionId
+        },
+        owner: {
+          agent: process.env.HOLOSCRIPT_AGENT_HANDLE ?? "unknown",
+          ...process.env.HOLOMESH_TEAM_ID ? { team: process.env.HOLOMESH_TEAM_ID } : {}
+        }
+      };
+      const ts = capturedAt.replace(/[:.]/g, "-");
+      const outPath = resolve(ALLOWED_WRITE_ROOTS[0], `hardware-receipt-${ts}.json`);
+      const denied = checkWriteAllowed(outPath);
+      if (denied) return errResult(use.id, `Cannot write receipt: ${denied}`);
+      await mkdir(dirname(outPath), { recursive: true });
+      await writeFile(outPath, JSON.stringify(receipt, null, 2), "utf8");
+      return okResult(
+        use.id,
+        `Hardware receipt written to ${outPath} \u2014 ${measurements.length} measurements, accelerator=${accelerator ?? "none"}`
+      );
+    }
     return errResult(use.id, `unknown tool: ${use.name}`);
   } catch (err) {
     return errResult(use.id, err instanceof Error ? err.message : String(err));
   }
 }
+function parseTegrastats(raw) {
+  const results = [];
+  const m = (pattern, metric, unit, transform) => {
+    const match = raw.match(pattern);
+    if (match?.[1]) {
+      const value = transform ? transform(match[1]) : Number(match[1]);
+      if (Number.isFinite(value)) results.push({ metric, value, unit, method: "tegrastats" });
+    }
+  };
+  const ram = raw.match(/RAM\s+(\d+)\/(\d+)MB/);
+  if (ram) {
+    const used = Number(ram[1]);
+    const total = Number(ram[2]);
+    results.push({ metric: "ram-used", value: used, unit: "MB", method: "tegrastats" });
+    results.push({ metric: "ram-total", value: total, unit: "MB", method: "tegrastats" });
+    if (total > 0)
+      results.push({ metric: "ram-pct", value: Math.round(used / total * 100), unit: "%", method: "tegrastats" });
+  }
+  m(/GR3D_FREQ\s+(\d+)%/, "gpu-util", "%");
+  m(/EMC_FREQ\s+(\d+)%/, "emc-freq-pct", "%");
+  m(/tj@([\d.]+)C/, "temp-tj", "C", parseFloat);
+  m(/cpu@([\d.]+)C/, "temp-cpu", "C", parseFloat);
+  m(/gpu@([\d.]+)C/, "temp-gpu", "C", parseFloat);
+  m(/VDD_SOC\s+(\d+)mW/, "power-soc", "mW");
+  m(/VDD_CPU_CV\s+(\d+)mW/, "power-cpu-cv", "mW");
+  m(/VDD_IN\s+(\d+)mW/, "power-total", "mW");
+  m(/CPU\s+\[(\d+)%/, "cpu-util-core0", "%");
+  return results;
+}
 function runBash(cmd, cwd) {
+  if (process.env.VITEST === "true" || process.env.NODE_ENV === "test") {
+    return Promise.resolve({
+      code: 0,
+      stdout: `[mock-bash under vitest] cmd="${cmd}" cwd="${cwd}"`,
+      stderr: ""
+    });
+  }
   return new Promise((resolveProm) => {
     const child = spawn("bash", ["-c", cmd], { cwd, env: process.env });
     let stdout = "";
@@ -288,6 +460,35 @@ var AgentRunner = class {
     const { identity, brain, mesh, costGuard, provider, logger } = this.opts;
     const log = logger ?? (() => void 0);
     await this.heartbeatWithAutoRejoin();
+    if (this.opts.messageHandler) {
+      try {
+        const receipts = await this.opts.messageHandler.processMessages();
+        if (receipts.length > 0) {
+          log({
+            ev: "messages-processed",
+            count: receipts.length,
+            statuses: receipts.map((r) => r.status)
+          });
+          if (brain.capabilityTags.length === 0 || brain.capabilityTags.every((t) => t.startsWith("delegated"))) {
+            return {
+              action: "messages-processed",
+              spentUsd: costGuard.getState().spentUsd,
+              remainingUsd: costGuard.getRemainingUsd(),
+              receipts: receipts.map((r) => ({
+                status: r.status,
+                action: r.action,
+                reason: r.reason
+              }))
+            };
+          }
+        }
+      } catch (err) {
+        log({
+          ev: "message-handler-error",
+          message: err instanceof Error ? err.message : String(err)
+        });
+      }
+    }
     if (costGuard.isOverBudget()) {
       const state = costGuard.getState();
       log({ ev: "over-budget", spentUsd: state.spentUsd, budget: identity.budgetUsdPerDay });
@@ -321,6 +522,8 @@ var AgentRunner = class {
     const MAX_TOOL_ITERS = 30;
     let lastResponse;
     const toolsCalled = /* @__PURE__ */ new Set();
+    let productiveCallCount = 0;
+    let lastCommitHash;
     while (true) {
       iters++;
       if (iters > MAX_TOOL_ITERS) {
@@ -328,12 +531,16 @@ var AgentRunner = class {
         finalText = finalText || `[tool-loop hit ${MAX_TOOL_ITERS}-iter cap before final text]`;
         break;
       }
+      const activeTools = brain.requires.includes("local-llm") ? MESH_TOOLS.filter((t) => t.name === "write_file") : MESH_TOOLS;
       const resp = await provider.complete(
         {
           messages,
-          maxTokens: 4096,
+          // 8192 for local thinking models (qwen3:4b uses ~3800 tokens on thinking
+          // before the tool-call JSON; 4096 cuts off mid-generation). Frontier
+          // models ignore this ceiling and stop naturally earlier.
+          maxTokens: 8192,
           temperature: 0.4,
-          tools: MESH_TOOLS
+          tools: activeTools
         },
         identity.llmModel
       );
@@ -344,13 +551,39 @@ var AgentRunner = class {
         totalTokens: aggUsage.totalTokens + resp.usage.totalTokens
       };
       if (resp.finishReason === "tool_use" && resp.toolUses && resp.toolUses.length > 0) {
-        log({ ev: "tool-call", taskId: target.id, iter: iters, tools: resp.toolUses.map((t) => t.name) });
-        for (const u of resp.toolUses) toolsCalled.add(u.name);
+        log({
+          ev: "tool-call",
+          taskId: target.id,
+          iter: iters,
+          tools: resp.toolUses.map((t) => t.name)
+        });
+        for (const u of resp.toolUses) {
+          toolsCalled.add(u.name);
+          if (u.name === "write_file") {
+            const content = String(u.input?.content ?? "");
+            if (content.length > 0) productiveCallCount++;
+          } else if (u.name === "bash") {
+            const cmd = String(u.input?.cmd ?? "");
+            if (isProductiveBashCommand(cmd)) productiveCallCount++;
+          } else if (u.name === "emit_hardware_receipt") {
+            productiveCallCount++;
+          }
+        }
         messages.push({
           role: "assistant",
           content: resp.assistantBlocks ?? []
         });
         const toolResults = await Promise.all(resp.toolUses.map((u) => runTool(u)));
+        for (let ti = 0; ti < resp.toolUses.length; ti++) {
+          const tu = resp.toolUses[ti];
+          if (tu.name === "bash") {
+            const tr = toolResults[ti];
+            if (tr && !tr.is_error) {
+              const shaMatch = tr.content.match(/\b([0-9a-f]{7,40})\b/);
+              if (shaMatch) lastCommitHash = shaMatch[1];
+            }
+          }
+        }
         messages.push({
           role: "user",
           content: toolResults
@@ -361,24 +594,75 @@ var AgentRunner = class {
       break;
     }
     const durationMs = Date.now() - start;
-    const SIDE_EFFECTING_TOOLS = /* @__PURE__ */ new Set(["write_file", "bash"]);
-    const sideEffectingCalled = [...toolsCalled].some((t) => SIDE_EFFECTING_TOOLS.has(t));
-    if (!sideEffectingCalled) {
+    if (productiveCallCount === 0) {
       log({
         ev: "no-artifact",
         taskId: target.id,
         tool_iters: iters,
         toolsCalled: [...toolsCalled],
-        message: "task execution called no side-effecting tool (write_file/bash) \u2014 refusing to mark executed. Likely a pure-text or read-only-inspection response. Task remains open for a grounded attempt."
+        productiveCallCount,
+        message: "task execution did not produce a real artifact \u2014 refusing to mark executed. Required: write_file with non-empty content OR bash with a productive prefix (lake build / pnpm --filter / vitest run / lean / pnpm vitest). Pure-text, read-only inspection, and trivial-bash-bypass (`echo`, `cat`, etc.) do not satisfy the gate."
       });
       return {
         action: "no-artifact",
         taskId: target.id,
         spentUsd: costGuard.getState().spentUsd,
         remainingUsd: costGuard.getRemainingUsd(),
-        message: `no side-effecting tool called (toolsCalled=[${[...toolsCalled].join(",")}], iters=${iters})`
+        message: `no productive tool call observed (toolsCalled=[${[...toolsCalled].join(",")}], productiveCallCount=${productiveCallCount}, iters=${iters})`
       };
     }
+    let reflectVerdict;
+    if (brain.reflect) {
+      try {
+        const reflectResp = await provider.complete(
+          {
+            messages: [
+              {
+                role: "system",
+                content: "You are a strict reviewer. Evaluate the work against the criteria; do not rewrite it."
+              },
+              {
+                role: "user",
+                content: `Reflect on the artifact produced for this task. Evaluate it for: ${brain.reflect.criteria}.
+
+--- artifact / final response ---
+${finalText.slice(0, 4e3)}
+--- end ---
+
+Give a one-line reason, then end with exactly "VERDICT: PASS" or "VERDICT: FAIL".`
+              }
+            ],
+            maxTokens: 512,
+            temperature: 0.1
+          },
+          identity.llmModel
+        );
+        aggUsage = {
+          promptTokens: aggUsage.promptTokens + reflectResp.usage.promptTokens,
+          completionTokens: aggUsage.completionTokens + reflectResp.usage.completionTokens,
+          totalTokens: aggUsage.totalTokens + reflectResp.usage.totalTokens
+        };
+        const verdictMatch = /VERDICT:\s*(PASS|FAIL)/i.exec(reflectResp.content);
+        const pass = verdictMatch ? verdictMatch[1].toUpperCase() === "PASS" : true;
+        reflectVerdict = {
+          pass,
+          reason: reflectResp.content.replace(/VERDICT:\s*(PASS|FAIL)/i, "").trim().slice(0, 300)
+        };
+        log({
+          ev: "reflect",
+          taskId: target.id,
+          pass,
+          escalateOnFail: brain.reflect.escalateOnFail,
+          reason: reflectVerdict.reason.slice(0, 120)
+        });
+      } catch (err) {
+        log({
+          ev: "reflect-error",
+          taskId: target.id,
+          message: err instanceof Error ? err.message : String(err)
+        });
+      }
+    }
     const cost = costGuard.recordUsage(identity.llmModel, aggUsage);
     log({
       ev: "executed",
@@ -388,7 +672,11 @@ var AgentRunner = class {
       tokens: aggUsage.totalTokens,
       tool_iters: iters
     });
-    const response = { ...lastResponse ?? { content: finalText, usage: aggUsage }, content: finalText, usage: aggUsage };
+    const response = {
+      ...lastResponse ?? { content: finalText, usage: aggUsage },
+      content: finalText,
+      usage: aggUsage
+    };
     const execResult = {
       taskId: target.id,
       responseText: response.content,
@@ -422,10 +710,32 @@ var AgentRunner = class {
       });
       const posted = await mesh.postAuditRecords(identity.handle, [caelRecord]);
       this.prevCaelChain = caelRecord.fnv1a_chain;
-      log({ ev: "cael-posted", taskId: target.id, appended: posted.appended, rejected: posted.rejected });
+      log({
+        ev: "cael-posted",
+        taskId: target.id,
+        appended: posted.appended,
+        rejected: posted.rejected
+      });
     } catch (err) {
       log({ ev: "cael-post-error", message: err instanceof Error ? err.message : String(err) });
     }
+    if (reflectVerdict && !reflectVerdict.pass && brain.reflect?.escalateOnFail) {
+      try {
+        await mesh.sendMessageOnTask(
+          target.id,
+          `[${identity.handle}] reflect gate FAILED \u2014 escalating to the fleet instead of marking done. Reason: ${reflectVerdict.reason}`
+        );
+      } catch {
+      }
+      log({ ev: "reflect-escalate", taskId: target.id, reason: reflectVerdict.reason.slice(0, 120) });
+      return {
+        action: "reflect-escalate",
+        taskId: target.id,
+        spentUsd: costGuard.getState().spentUsd,
+        remainingUsd: costGuard.getRemainingUsd(),
+        message: `reflect self-evaluation failed; escalated to fleet (reason: ${reflectVerdict.reason.slice(0, 120)})`
+      };
+    }
     if (this.opts.onTaskExecuted) {
       await this.opts.onTaskExecuted(execResult, target);
     } else {
@@ -436,6 +746,16 @@ var AgentRunner = class {
 ${response.content}`
       );
     }
+    try {
+      await mesh.markDone(target.id, finalText.slice(0, 500), lastCommitHash);
+      log({ ev: "mark-done", taskId: target.id, commitHash: lastCommitHash });
+    } catch (err) {
+      log({
+        ev: "mark-done-error",
+        taskId: target.id,
+        message: err instanceof Error ? err.message : String(err)
+      });
+    }
     return {
       action: "executed",
       taskId: target.id,
@@ -528,7 +848,7 @@ function buildTaskPrompt(task) {
     "Description:",
     task.description ?? "(no description)",
     "",
-    "Produce the deliverable described in the task. Apply your brain composition rules \u2014 anti-patterns, decision loop, and scope tier all bind. Return the response as plain text suitable for posting to /room as a message on this task."
+    "Produce the deliverable: call write_file (or bash with a build command) to create all required output files FIRST. Apply your brain composition rules \u2014 anti-patterns, decision loop, and scope tier all bind. After calling the tool(s), return a short plain-text summary of what you did for posting to /room."
   ].join("\n");
 }
 function sleep(ms) {