Fix Pi engine report_incomplete emission failing silently in sandboxed agent container (EROFS)

actions/setup/js/pi_provider.cjs

@@ -28,8 +28,8 @@
 "use strict";
 
 const { fetchAWFReflect, AWF_API_PROXY_REFLECT_URL, AWF_REFLECT_OUTPUT_PATH, AWF_REFLECT_TIMEOUT_MS, AWF_MODELS_URL_TIMEOUT_MS } = require("./awf_reflect.cjs");
+const { emitInfrastructureIncomplete } = require("./safeoutputs_cli.cjs");
 const fs = require("fs");
-const path = require("path");
 const { getErrorMessage } = require("./error_helpers.cjs");
 
 // Default logger: prefixed with "[gh-aw/pi-provider]" for easy grepping.
@@ -142,23 +142,16 @@ function formatResponseHeaderNames(headers) {
 }
 
 /**
- * Build a structured report_incomplete payload for infrastructure failures.
- *
- * @param {string} details
- * @returns {string}
- */
-function buildInfrastructureIncompletePayload(details) {
-  return JSON.stringify({
-    type: "report_incomplete",
-    reason: "infrastructure_error",
-    details,
-  });
-}
-
-/**
- * Append a report_incomplete safe output when provider infrastructure fails
+ * Emit a report_incomplete safe output when provider infrastructure fails
  * before any safe outputs have been recorded.
  *
+ * This Pi extension runs inside the AWF agent sandbox, where the directory
+ * backing GH_AW_SAFE_OUTPUTS is mounted read-only (writes fail with EROFS).
+ * Emission is therefore delegated to the `safeoutputs` CLI (see
+ * safeoutputs_cli.cjs), which forwards the call to the MCP gateway process
+ * that owns write access to the real outputs file — the same channel used
+ * by every other safe-output tool call the agent makes.
+ *
  * @param {string} details
  * @param {(msg: string) => void} logger
  * @returns {void}
@@ -176,14 +169,11 @@ function emitInfrastructureIncompleteIfNoSafeOutputs(details, logger) {
       logger(`report_incomplete skipped: safe outputs already recorded at ${safeOutputsPath}`);
       return;
     }
-
-    fs.mkdirSync(path.dirname(safeOutputsPath), { recursive: true });
-    fs.appendFileSync(safeOutputsPath, buildInfrastructureIncompletePayload(details) + "\n", { encoding: "utf8" });
-    logger(`report_incomplete emitted: ${safeOutputsPath}`);
   } catch (error) {
-    const message = getErrorMessage(error);
-    logger(`report_incomplete emission failed: ${message}`);
+    logger(`report_incomplete pre-check failed, proceeding with emission: ${getErrorMessage(error)}`);
   }
+
+  emitInfrastructureIncomplete(details, { safeOutputsPath, logger });
 }
 
 /**
@@ -411,6 +401,5 @@ _piExports.resolveGatewayUrl = resolveGatewayUrl;
 _piExports.registerConfiguredProviders = registerConfiguredProviders;
 _piExports.resolveProviderRequestTarget = resolveProviderRequestTarget;
 _piExports.formatResponseHeaderNames = formatResponseHeaderNames;
-_piExports.buildInfrastructureIncompletePayload = buildInfrastructureIncompletePayload;
 _piExports.emitInfrastructureIncompleteIfNoSafeOutputs = emitInfrastructureIncompleteIfNoSafeOutputs;
 _piExports.logReflectFailure = logReflectFailure;

actions/setup/js/pi_provider.test.cjs

@@ -116,10 +116,14 @@ describe("pi_provider.cjs", () => {
     expect(stderrOutput.some(line => line.includes("provider_response provider=copilot model=claude-sonnet-4 status=503 method=POST url=http://api-proxy:10002/v1/chat/completions response_headers=content-type,x-request-id"))).toBe(true);
   });
 
-  it("logs assistant inference errors with the last request target", async () => {
+  // Triggers the message_end infrastructure-error handler with a given stand-in
+  // GH_AW_SAFEOUTPUTS_CLI override ('true' simulates a successful CLI call, 'false'
+  // simulates a failed one) and returns the handlers/stderr output for assertions.
+  async function triggerConnectionError(cliOverride) {
     process.env.GH_AW_PI_MODEL = "copilot/claude-sonnet-4";
     const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), "pi-provider-"));
     process.env.GH_AW_SAFE_OUTPUTS = path.join(tempDir, "outputs.jsonl");
+    process.env.GH_AW_SAFEOUTPUTS_CLI = cliOverride;
 
     const handlers = {};
     const pi = {
@@ -150,14 +154,30 @@ describe("pi_provider.cjs", () => {
         errorMessage: "Connection error.",
       },
     });
+  }
+
+  it("logs assistant inference errors with the last request target", async () => {
+    // 'true' is a stand-in safeoutputs CLI binary that always exits 0, simulating a
+    // successful emission through the CLI channel instead of a direct fs append.
+    await triggerConnectionError("true");
 
     expect(
       stderrOutput.some(line =>
         line.includes('provider_error provider=aw-gateway model=claude-sonnet-4 api=openai-completions status=no-response method=POST url=http://api-proxy:10002/v1/chat/completions response_headers=none error="Connection error."')
       )
     ).toBe(true);
-    expect(fs.readFileSync(process.env.GH_AW_SAFE_OUTPUTS, "utf8")).toContain('"type":"report_incomplete"');
-    expect(fs.readFileSync(process.env.GH_AW_SAFE_OUTPUTS, "utf8")).toContain("Pi provider request failed before safe outputs were emitted");
+    // Emission goes through the safeoutputs CLI channel (not a direct fs append), so it
+    // survives the read-only sandbox mount that backs GH_AW_SAFE_OUTPUTS in production.
+    expect(stderrOutput.some(line => line.includes("report_incomplete emitted via safeoutputs CLI"))).toBe(true);
+  });
+
+  it("logs a failure when the safeoutputs CLI channel is unavailable (e.g. read-only sandbox)", async () => {
+    // 'false' is a stand-in safeoutputs CLI binary that always exits 1, simulating a
+    // failed CLI invocation without ever touching the filesystem directly.
+    await triggerConnectionError("false");
+
+    expect(stderrOutput.some(line => line.includes("report_incomplete emission failed"))).toBe(true);
+    expect(fs.existsSync(process.env.GH_AW_SAFE_OUTPUTS)).toBe(false);
   });
 
   it("skips synthetic report_incomplete emission when safe outputs already exist", () => {
@@ -166,9 +186,11 @@ describe("pi_provider.cjs", () => {
     process.env.GH_AW_SAFE_OUTPUTS = safeOutputsPath;
     fs.writeFileSync(safeOutputsPath, '{"type":"add_comment","body":"done"}\n');
 
-    module.emitInfrastructureIncompleteIfNoSafeOutputs("temporary outage", () => {});
+    const logs = [];
+    module.emitInfrastructureIncompleteIfNoSafeOutputs("temporary outage", message => logs.push(message));
 
     expect(fs.readFileSync(safeOutputsPath, "utf8")).toBe('{"type":"add_comment","body":"done"}\n');
+    expect(logs.some(m => m.includes("skipped: safe outputs already recorded"))).toBe(true);
   });
 
   it("calls /reflect on the management port (10000) when AWF_REFLECT_ENABLED is set", async () => {