@whittlelabs/sifter 0.39.0 → 0.40.0

package/README.md

@@ -98,6 +98,7 @@ The Sifter subscribes to your job pool and waits for review jobs. Leave it runni
 A review checks out the pull request and runs Claude Code on it. The pull request is someone else's code, so the review is held to reading:
 
 - It is offered three tools, Read, Grep and Glob, and nothing else: no shell, no edits, no network tools, no subagents.
+- When the review draws a guided tour, it may also write files into one empty folder of its own (`out/`, beside the checkout), and nowhere else. The Sifter puts each drawing written there back into the review and deletes the folder with the checkout. A review that loads your Claude Code user settings writes nothing, because your allow rules could widen where it writes.
 - It loads no Claude Code settings file, neither the repository's nor yours. Your allow rules, permission mode, hooks and plugins do not apply to reviews, and no MCP server starts.
 
 Your login still works: Claude Code keeps it in your keychain or credentials file, or reads `ANTHROPIC_API_KEY`. If instead your login or provider is configured in your Claude Code settings (an `apiKeyHelper`, or an `env` block naming Bedrock, Vertex or a custom base URL), reviews will fail as signed out. Tell the Sifter to load your user settings in its `sifter.yaml` (`~/.sifter/sifter.yaml`, or `~/.sifter/<profile>/sifter.yaml` for a named profile):

package/bin.js

@@ -3214,6 +3214,13 @@ var require_prompt_execution = __commonJS({
           }
         }
       }
+      if (spec.answerFiles !== void 0) {
+        const f = spec.answerFiles;
+        const isName = (v) => typeof v === "string" && v.length > 0;
+        if (typeof f !== "object" || f === null || Array.isArray(f) || !isName(f.at) || !isName(f.reference) || !isName(f.into) || !Array.isArray(f.extensions) || f.extensions.length === 0 || !f.extensions.every((e) => typeof e === "string" && /^\.[a-z0-9]+$/.test(e)) || typeof f.maxBytes !== "number" || !Number.isInteger(f.maxBytes) || f.maxBytes <= 0) {
+          throw new Error('prompt-execution: spec.answerFiles must be { at, reference, into, extensions: [".ext", \u2026], maxBytes }');
+        }
+      }
       return spec;
     }
     function renderPromptExecution(inputs) {
@@ -3454,6 +3461,8 @@ var require_prompt_execution2 = __commonJS({
         spec.progressChannel = options.progressChannel;
       if (options.outputGates && options.outputGates.length > 0)
         spec.gates = options.outputGates;
+      if (options.answerFiles)
+        spec.answerFiles = options.answerFiles;
       if (options.rendering === "producer" && !options.prompt) {
         throw new Error("rendering='producer' requires `prompt`");
       }
@@ -9648,6 +9657,7 @@ var require_canonicalise = __commonJS({
           promptTemplatePath: promptOptions.promptTemplate,
           outputSchema: toJsonSchema(promptOptions.outputSchema),
           ...promptOptions.outputGates ? { outputGates: promptOptions.outputGates } : {},
+          ...promptOptions.answerFiles ? { answerFiles: promptOptions.answerFiles } : {},
           outputLabel,
           rendering: promptOptions.rendering,
           ...promptOptions.maxAttempts !== void 0 ? { maxAttempts: promptOptions.maxAttempts } : {},
@@ -10826,19 +10836,25 @@ var require_claude_code = __commonJS({
     var DEFAULT_TIMEOUT = 6e5;
     var REVIEW_TOOLS = ["Read", "Grep", "Glob"];
     var THINKING_DISPLAY_FLOOR = "2.1.94";
-    function reviewRunArgs(settingSources) {
+    function reviewRunArgs(settingSources, writesAnswerFiles = false) {
       return [
         "--setting-sources",
         settingSources,
         "--strict-mcp-config",
         "--tools",
-        REVIEW_TOOLS.join(","),
+        (writesAnswerFiles ? [...REVIEW_TOOLS, "Write"] : REVIEW_TOOLS).join(","),
         "--permission-mode",
         "default",
         "--settings",
         JSON.stringify({ disableAllHooks: true })
       ];
     }
+    function outputDirRule(cwd, outputDir) {
+      const relative = path_1.default.relative(cwd, outputDir);
+      if (relative !== path_1.default.basename(outputDir) || relative.startsWith("."))
+        return null;
+      return `Edit(./${relative}/**)`;
+    }
     var RETIRED_READ_TOOLS = ["LS"];
     var HOSTED_ENV_PASSTHROUGH = [
       "PATH",
@@ -10927,7 +10943,11 @@ var require_claude_code = __commonJS({
           materializeOptions: {
             ...tmpBase ? { tmpDir: tmpBase } : {},
             overlayAllowedHosts: this.config.overlayAllowedHosts ?? []
-          }
+          },
+          // Writes are confined by one allow rule (`outputDirRule`), which an
+          // operator's own settings could widen: a run that loads them writes
+          // nothing, and its answer inlines (ADR 0133).
+          writesAnswerFiles: this.config.claudeSettings !== "user"
         };
       }
       async execute(dispatch, signal, ctx) {
@@ -10979,7 +10999,8 @@ var require_claude_code = __commonJS({
             // machine's, and the runner pauses on it. With a revealed key it is one
             // customer's, and only their job fails.
             modelEnv === null,
-            effort
+            effort,
+            ctx?.outputDir
           );
           if (workspace) {
             response.outputs.push(...(0, workspace_telemetry_1.workspaceOutputs)(workspace, runCwd, readPaths, toolInputs));
@@ -11065,7 +11086,7 @@ var require_claude_code = __commonJS({
       async listModels() {
         return CLAUDE_CODE_MODELS;
       }
-      async runClaude(prompt, spec, cwd, childEnv, signal, modelOverride, reporter, resumeSessionId, transcriptName, local = false, effortOverride) {
+      async runClaude(prompt, spec, cwd, childEnv, signal, modelOverride, reporter, resumeSessionId, transcriptName, local = false, effortOverride, outputDir) {
         const outputLabel = spec.outputLabel;
         const model = modelOverride ?? this.config.model;
         const effort = effortOverride ?? this.config.reasoningEffort;
@@ -11074,7 +11095,14 @@ var require_claude_code = __commonJS({
           throw new Error("claude-code: `claudeSettings: user` is for an installed Sifter whose login lives in Claude Code settings; a hosted run authenticates with the key it is handed and loads no settings");
         }
         const settingSources = local && this.config.claudeSettings === "user" ? "user" : "";
-        const args = ["--print", "--output-format", "stream-json", "--verbose", ...reviewRunArgs(settingSources)];
+        const writeRule = outputDir ? outputDirRule(cwd, outputDir) : null;
+        const args = [
+          "--print",
+          "--output-format",
+          "stream-json",
+          "--verbose",
+          ...reviewRunArgs(settingSources, writeRule !== null)
+        ];
         if (resumeSessionId)
           args.push("--resume", resumeSessionId);
         if (reporter)
@@ -11092,8 +11120,9 @@ var require_claude_code = __commonJS({
         }
         if (this.config.maxTurns !== void 0)
           args.push("--max-turns", String(this.config.maxTurns));
-        if (allowedTools && allowedTools.length > 0) {
-          args.push("--allowedTools", allowedTools.join(","));
+        const allowRules = [...allowedTools ?? [], ...writeRule ? [writeRule] : []];
+        if (allowRules.length > 0) {
+          args.push("--allowedTools", allowRules.join(","));
         }
         const tap = reporter ? new progress_1.StreamJsonTap({
           onReasoning: (t) => reporter.reasoningBlock(t),
@@ -11530,8 +11559,10 @@ var require_codex = __commonJS({
     var workspace_telemetry_1 = require_workspace_telemetry();
     var apply_1 = require_apply();
     var CODEX_PROFILE = "sift-review";
-    function reviewProfile(installDir, platform = process.platform) {
+    function reviewProfile(installDir, platform = process.platform, outputDir) {
       const rules = ['":root"="deny"', '":minimal"="read"', '":workspace_roots"="read"', `${JSON.stringify(installDir)}="read"`];
+      if (outputDir)
+        rules.push(`${JSON.stringify(outputDir)}="write"`);
       if (platform === "linux") {
         rules.push(`${JSON.stringify(path_1.default.join(codexHome(), "tmp", "arg0"))}="read"`);
       } else {
@@ -11582,14 +11613,14 @@ var require_codex = __commonJS({
       }
       return codex;
     }
-    function codexReviewArgs(installDir, platform = process.platform) {
+    function codexReviewArgs(installDir, platform = process.platform, outputDir) {
       return [
         "--json",
         "--skip-git-repo-check",
         "--ignore-user-config",
         "--ignore-rules",
         "-c",
-        reviewProfile(installDir, platform),
+        reviewProfile(installDir, platform, outputDir),
         "-c",
         `default_permissions="${CODEX_PROFILE}"`,
         "-c",
@@ -11632,7 +11663,7 @@ var require_codex = __commonJS({
         };
       }
       async prepareEnvironment(dispatch) {
-        return (0, agent_common_1.agentEnvironment)(dispatch, this.config);
+        return { ...(0, agent_common_1.agentEnvironment)(dispatch, this.config), writesAnswerFiles: true };
       }
       async execute(dispatch, signal, ctx) {
         (0, agent_common_1.refuseHosted)("codex", Boolean(ctx?.credentials));
@@ -11650,9 +11681,10 @@ var require_codex = __commonJS({
             runCwd = path_1.default.join(scratch, "run");
             await fs_1.promises.mkdir(runCwd);
           }
+          const outputDir = ctx?.outputDir ? await fs_1.promises.realpath(ctx.outputDir) : void 0;
           const { response, commands } = await this.runCodex({
             codex,
-            args: ["exec", ...codexReviewArgs(codex.installDir), "-C", runCwd],
+            args: ["exec", ...codexReviewArgs(codex.installDir, process.platform, outputDir), "-C", runCwd],
             prompt,
             spec,
             cwd: runCwd,
@@ -12016,14 +12048,22 @@ var require_copilot = __commonJS({
     var apply_1 = require_apply();
     var REVIEW_TOOLS = ["view", "grep", "glob"];
     var DENIED_TOOLS = ["shell", "write"];
-    function copilotReviewArgs() {
+    function copilotReviewArgs(outputDir) {
+      const permissions = outputDir ? [
+        `--available-tools=${[...REVIEW_TOOLS, "create"].join(",")}`,
+        ...REVIEW_TOOLS.map((t) => `--allow-tool=${t}`),
+        `--allow-tool=write(${path_1.default.join(outputDir, "*")})`,
+        ...DENIED_TOOLS.filter((t) => t !== "write").map((t) => `--deny-tool=${t}`)
+      ] : [
+        "--allow-all-tools",
+        `--available-tools=${REVIEW_TOOLS.join(",")}`,
+        ...DENIED_TOOLS.map((t) => `--deny-tool=${t}`)
+      ];
       return [
         "--output-format",
         "json",
         "--no-color",
-        "--allow-all-tools",
-        `--available-tools=${REVIEW_TOOLS.join(",")}`,
-        ...DENIED_TOOLS.map((t) => `--deny-tool=${t}`),
+        ...permissions,
         "--disable-builtin-mcps",
         "--disallow-temp-dir",
         "--no-ask-user",
@@ -12065,7 +12105,7 @@ var require_copilot = __commonJS({
         };
       }
       async prepareEnvironment(dispatch) {
-        return (0, agent_common_1.agentEnvironment)(dispatch, this.config);
+        return { ...(0, agent_common_1.agentEnvironment)(dispatch, this.config), writesAnswerFiles: true };
       }
       async execute(dispatch, signal, ctx) {
         (0, agent_common_1.refuseHosted)("copilot", Boolean(ctx?.credentials));
@@ -12088,8 +12128,9 @@ var require_copilot = __commonJS({
             await fs_1.promises.mkdir(runCwd);
           }
           const sessionId = (0, crypto_1.randomUUID)();
+          const outputDir = ctx?.outputDir ? await fs_1.promises.realpath(ctx.outputDir) : void 0;
           const { response, reads, toolInputs } = await this.runCopilot({
-            args: [...copilotReviewArgs(), `--session-id=${sessionId}`, "-C", runCwd],
+            args: [...copilotReviewArgs(outputDir), `--session-id=${sessionId}`, "-C", runCwd],
             prompt: withSchema(prompt, spec),
             spec,
             cwd: runCwd,
@@ -27192,6 +27233,96 @@ var require_workspace = __commonJS({
   }
 });
 
+// ../../packages/shuttle/dist/core/answer-files.js
+var require_answer_files = __commonJS({
+  "../../packages/shuttle/dist/core/answer-files.js"(exports2) {
+    "use strict";
+    var __importDefault = exports2 && exports2.__importDefault || function(mod) {
+      return mod && mod.__esModule ? mod : { "default": mod };
+    };
+    Object.defineProperty(exports2, "__esModule", { value: true });
+    exports2.OUTPUT_DIR = void 0;
+    exports2.makeOutputDir = makeOutputDir;
+    exports2.collectAnswerFiles = collectAnswerFiles;
+    exports2.spliceAnswerFiles = spliceAnswerFiles;
+    var node_fs_1 = require("node:fs");
+    var node_path_1 = __importDefault(require("node:path"));
+    var output_gates_1 = require_output_gates();
+    exports2.OUTPUT_DIR = "out";
+    var FINDING_GATE = "answer-files";
+    async function makeOutputDir(runRoot) {
+      const dir = node_path_1.default.join(runRoot, exports2.OUTPUT_DIR);
+      await node_fs_1.promises.mkdir(dir);
+      return dir;
+    }
+    async function collectAnswerFiles(dir, declaration) {
+      const files = /* @__PURE__ */ new Map();
+      let entries;
+      try {
+        entries = await node_fs_1.promises.readdir(dir, { withFileTypes: true });
+      } catch {
+        return files;
+      }
+      for (const entry of entries) {
+        if (!entry.isFile())
+          continue;
+        if (!declaration.extensions.includes(node_path_1.default.extname(entry.name).toLowerCase()))
+          continue;
+        const full = node_path_1.default.join(dir, entry.name);
+        const { size } = await node_fs_1.promises.stat(full);
+        if (size > declaration.maxBytes) {
+          files.set(entry.name, { problem: `is ${size} bytes, over the ${declaration.maxBytes}-byte limit` });
+          continue;
+        }
+        files.set(entry.name, { text: await node_fs_1.promises.readFile(full, "utf8") });
+      }
+      return files;
+    }
+    function spliceAnswerFiles(response, outputLabel, declaration, files) {
+      if (!response.parsed)
+        return { response, findings: [] };
+      const answer = structuredClone(response.parsed);
+      const findings = [];
+      const selected = (0, output_gates_1.resolvePath)(answer, declaration.at);
+      for (const target of Array.isArray(selected) ? selected : [selected]) {
+        if (typeof target !== "object" || target === null || Array.isArray(target))
+          continue;
+        const holder = target;
+        const name = holder[declaration.reference];
+        if (name === void 0)
+          continue;
+        const refused = (why) => findings.push({
+          gate: FINDING_GATE,
+          message: `${declaration.at}.${declaration.reference} ${JSON.stringify(name)} ${why}. Put the content inline as \`${declaration.into}\` and remove \`${declaration.reference}\`.`
+        });
+        if (typeof name !== "string" || name !== node_path_1.default.basename(name) || name.startsWith(".")) {
+          refused(`is not a plain file name in ${exports2.OUTPUT_DIR}/`);
+          continue;
+        }
+        const file = files.get(name);
+        if (!file) {
+          refused(`names no ${declaration.extensions.join(" or ")} file written to ${exports2.OUTPUT_DIR}/`);
+          continue;
+        }
+        if ("problem" in file) {
+          refused(file.problem);
+          continue;
+        }
+        holder[declaration.into] = file.text;
+        delete holder[declaration.reference];
+      }
+      return {
+        response: {
+          ...response,
+          parsed: answer,
+          outputs: response.outputs.map((o) => o.label === outputLabel ? { ...o, content: answer } : o)
+        },
+        findings
+      };
+    }
+  }
+});
+
 // ../../packages/shuttle/dist/core/run-dispatch.js
 var require_run_dispatch = __commonJS({
   "../../packages/shuttle/dist/core/run-dispatch.js"(exports2) {
@@ -27205,6 +27336,7 @@ var require_run_dispatch = __commonJS({
     var output_gates_1 = require_output_gates();
     var logger_1 = require_logger();
     var workspace_1 = require_workspace();
+    var answer_files_1 = require_answer_files();
     exports2.GATE_REPAIR_BUDGET = 2;
     async function runDispatch(dispatch, executor, options) {
       const { credentials, signal } = options;
@@ -27225,46 +27357,78 @@ var require_run_dispatch = __commonJS({
           workspace = null;
         }
       }
-      const ctx = {
+      const declaration = answerFilesDeclaration(dispatch);
+      let outputDir;
+      if (declaration && workspace && environment?.writesAnswerFiles) {
+        try {
+          outputDir = await (0, answer_files_1.makeOutputDir)(workspace.cwd);
+        } catch (err) {
+          (0, logger_1.getLogger)().warn("could not make the output directory; the answer will inline its files", {
+            jobId: dispatch.job.id,
+            error: err instanceof Error ? err.message : String(err)
+          });
+        }
+      }
+      const repairCtx = {
         ...options.reportProgress ? { reportProgress: options.reportProgress } : {},
         workspace,
         credentials
       };
+      const ctx = { ...repairCtx, ...outputDir ? { outputDir } : {} };
       let response;
+      let files = /* @__PURE__ */ new Map();
       try {
         response = await executor.execute(dispatch, signal, ctx);
+        if (outputDir && declaration)
+          files = await (0, answer_files_1.collectAnswerFiles)(outputDir, declaration);
       } catch (error) {
         return { kind: "executor-failed", error };
       } finally {
         if (workspace)
           await workspace.cleanup();
       }
-      const gated = await runOutputGates(dispatch, response, executor, signal, ctx);
+      const gated = await runOutputGates(dispatch, response, executor, signal, repairCtx, files);
       if (!gated.ok)
         return { kind: "gate-failed", reason: gated.reason };
       return { kind: "completed", response: gated.response };
     }
-    async function runOutputGates(dispatch, response, executor, signal, ctx) {
+    function answerFilesDeclaration(dispatch) {
+      try {
+        return (0, loom_1.readPromptExecutionSpec)(dispatch.inputs).answerFiles ?? null;
+      } catch {
+        return null;
+      }
+    }
+    async function runOutputGates(dispatch, response, executor, signal, ctx, files = /* @__PURE__ */ new Map()) {
       let spec;
       try {
         spec = (0, loom_1.readPromptExecutionSpec)(dispatch.inputs);
       } catch {
         return { ok: true, response };
       }
-      const gates = spec.gates;
-      if (!gates || gates.length === 0)
+      const gates = spec.gates ?? [];
+      const declaration = spec.answerFiles;
+      if (gates.length === 0 && !declaration)
         return { ok: true, response };
-      let current = response;
-      let result = (0, output_gates_1.evaluateOutputGates)(gates, current.parsed ?? {}, dispatch.inputs);
+      const check = (candidate) => {
+        const spliced = declaration ? (0, answer_files_1.spliceAnswerFiles)(candidate, spec.outputLabel, declaration, files) : { response: candidate, findings: [] };
+        const gated = (0, output_gates_1.evaluateOutputGates)(gates, spliced.response.parsed ?? {}, dispatch.inputs);
+        const findings = [...spliced.findings, ...gated.findings];
+        return { response: spliced.response, result: { ok: findings.length === 0, findings } };
+      };
+      let checked = check(response);
+      let current = checked.response;
+      let result = checked.result;
       let attempts = 0;
       while (!result.ok && attempts < exports2.GATE_REPAIR_BUDGET && current.continuation !== void 0 && typeof executor.continue === "function") {
         try {
           const repaired = await executor.continue(current.continuation, (0, output_gates_1.renderRepairPrompt)(result.findings), signal, ctx);
-          current = mergeRepairedResponse(current, repaired);
+          checked = check(mergeRepairedResponse(current, repaired));
         } catch {
           break;
         }
-        result = (0, output_gates_1.evaluateOutputGates)(gates, current.parsed ?? {}, dispatch.inputs);
+        current = checked.response;
+        result = checked.result;
         attempts += 1;
       }
       if (result.ok)
@@ -29469,7 +29633,7 @@ var import_path = require("path");
 var import_promises = require("fs/promises");
 var import_yaml = __toESM(require_dist3());
 var import_shuttle = __toESM(require_dist5());
-var buildVersion = true ? "0.39.0" : pkg.version;
+var buildVersion = true ? "0.40.0" : pkg.version;
 var AGENT_EXECUTORS = ["claude-code", "codex", "copilot"];
 var sifterBrand = {
   product: {