@odla-ai/harness 0.10.1 → 0.10.3

package/dist/code-runtime-cli.cjs

@@ -3,7 +3,7 @@
 
 // src/code-runtime-cli.ts
 var import_node_os4 = require("os");
-var import_promises12 = require("fs/promises");
+var import_promises13 = require("fs/promises");
 
 // src/code-runtime-client-validation.ts
 var import_code = require("@odla-ai/camel/code");
@@ -2071,10 +2071,12 @@ Prefer these over listing the tree \u2014 a full listing of a real repository is
 of thousands of tokens and you will carry it for the rest of the session.
 
 Then odla_search for a literal string, odla_read for a bounded range, and
-odla_apply_git_diff to change something. A patch must start with
-"diff --git a/<path> b/<path>", include matching "---" and "+++" headers and
-numbered "@@" hunks with at least one line of surrounding context, and must never
-use "*** Begin Patch" wrappers.
+odla_edit_file to change an existing file: give it the exact current text
+(copied verbatim from odla_read) and the text that replaces it; it must match
+once. Use odla_apply_git_diff only to create or delete whole files. A patch must
+start with "diff --git a/<path> b/<path>", include matching "---" and "+++"
+headers and numbered "@@" hunks with at least one line of surrounding context,
+and must never use "*** Begin Patch" wrappers.
 
 The workspace, model, and tool effects are controlled by the host broker.
 Never claim a build or test passed unless odla_run_recipe returned that result.`;
@@ -2134,6 +2136,21 @@ function codeSkill(opts) {
     },
     handler: (input, ctx) => call("sandbox.apply_patch", input, ctx.signal)
   };
+  const editFile = {
+    name: "odla_edit_file",
+    description: "Change an existing file by replacing one exact stretch of its current text with new text. oldText must match the file exactly once (copy it verbatim from odla_read, including indentation and blank lines; widen it if it would match more than once). Prefer this over odla_apply_git_diff for every change to an existing file; use odla_apply_git_diff only to create or delete whole files.",
+    inputSchema: {
+      type: "object",
+      required: ["path", "oldText", "newText"],
+      properties: {
+        path: { type: "string", minLength: 1, maxLength: 1024 },
+        oldText: { type: "string", minLength: 1, maxLength: 262144 },
+        newText: { type: "string", maxLength: 262144 }
+      },
+      additionalProperties: false
+    },
+    handler: (input, ctx) => call("sandbox.edit", input, ctx.signal)
+  };
   const runRecipe = {
     name: "odla_run_recipe",
     description: `Run one app-registered build or test recipe through CaMeL policy.${opts.recipeIds?.length ? ` Available recipes: ${opts.recipeIds.join(", ")}.` : ""}`,
@@ -2220,8 +2237,9 @@ function codeSkill(opts) {
       true
     )
   ];
-  const effects = opts.readOnly ? [] : [applyPatch, runRecipe];
-  const tools = opts.surface === "v3" ? [...orientation, searchFiles, read2, ...effects] : opts.surface === "v2" ? [listFiles, searchFiles, read2, ...effects] : [read2, ...effects];
+  const runsRecipes = opts.recipes ?? !opts.readOnly;
+  const effects = [...opts.readOnly ? [] : [applyPatch], ...runsRecipes ? [runRecipe] : []];
+  const tools = opts.surface === "v3" ? [...orientation, searchFiles, read2, ...opts.readOnly ? [] : [editFile], ...effects] : opts.surface === "v2" ? [listFiles, searchFiles, read2, ...effects] : [read2, ...effects];
   return { name: "code", tools };
 }
 
@@ -2278,6 +2296,7 @@ async function runCodeAgentAttempt(options) {
       model: "brokered",
       surface,
       ...options.readOnly === void 0 ? {} : { readOnly: options.readOnly },
+      ...options.recipes === void 0 ? {} : { recipes: options.recipes },
       ...options.recipeIds ? { recipeIds: options.recipeIds } : {},
       ...options.extraSkills ? { extraSkills: options.extraSkills } : {},
       ...options.maxSteps === void 0 ? {} : { maxSteps: options.maxSteps },
@@ -2511,6 +2530,10 @@ function codeToolRequestPresentation(request) {
   if (request.tool === "sandbox.apply_patch") {
     return { kind: "patch", ...patchStats(input.patch) };
   }
+  if (request.tool === "sandbox.edit") {
+    const lines = (value) => typeof value === "string" ? value.slice(0, 262144).split("\n").length : 0;
+    return { kind: "patch", additions: lines(input.newText), deletions: lines(input.oldText) };
+  }
   const recipeId = text(input.recipeId, 120);
   return recipeId ? { kind: "recipe", recipeId } : void 0;
 }
@@ -2638,6 +2661,11 @@ var PATCH = descriptor("sandbox.apply_patch", "reversible_mutation", {
   authority: "authority",
   patch: "payload"
 });
+var EDIT = descriptor("sandbox.edit", "reversible_mutation", {
+  workspace: "destination",
+  authority: "authority",
+  patch: "payload"
+});
 var RECIPE = descriptor("sandbox.run_recipe", "code_execution", {
   workspace: "destination",
   authority: "authority",
@@ -2705,6 +2733,14 @@ function createCodePolicyGate(options) {
         patch: { role: "payload", value: patch2 }
       }, []);
     },
+    edit: async (input) => {
+      const base = await environment(input, options, "sandbox.edit");
+      const patch2 = unsafe(base, input.patch, "patch");
+      return authorize(input, options, base, EDIT, {
+        ...base.fixedArgs,
+        patch: { role: "payload", value: patch2 }
+      }, []);
+    },
     recipe: async (input) => {
       const base = await environment(input, options, "sandbox.run_recipe");
       const conversions = await conversionRegistry([
@@ -2849,13 +2885,134 @@ function response(request, ok, content, details) {
   return { requestId: request.requestId, ok, content, details: { ...details, failureReason } };
 }
 
+// src/code-tool-edit.ts
+var import_promises10 = require("fs/promises");
+
+// src/code-tool-graph.ts
+var import_promises9 = require("fs/promises");
+var import_node_path10 = require("path");
+var import_graph = require("@odla-ai/graph");
+var import_code4 = require("@odla-ai/graph/code");
+var cache = /* @__PURE__ */ new Map();
+function workspaceGraphs(workspaceDir, paths2) {
+  const existing = cache.get(workspaceDir);
+  if (existing) return existing;
+  const read2 = (path) => (0, import_promises9.readFile)((0, import_node_path10.join)(workspaceDir, path), "utf8");
+  const built = (async () => ({
+    // No knownTables: a staged workspace may not carry migrations, and a filter
+    // that silently drops every table is worse than an unfiltered one. Callers
+    // with ground truth should build the graph themselves.
+    graph: await (0, import_code4.buildCodeGraph)({ paths: paths2, read: read2, data: { ignore: (path) => path.includes(".generated.") } })
+  }))();
+  cache.set(workspaceDir, built);
+  return built;
+}
+function forgetWorkspaceGraphs(workspaceDir) {
+  cache.delete(workspaceDir);
+}
+var shortId = (id) => id.slice(id.indexOf(":") + 1);
+function renderOverview(graphs, prefix) {
+  const rows = (0, import_graph.rollup)(graphs.graph, import_code4.FILE, prefix === void 0 ? {} : { prefix });
+  if (rows.length === 0) return prefix ? `No source under "${prefix}".` : "No source files.";
+  const lines = rows.slice(0, 60).map((row) => `${row.prefix}  (${row.count})  e.g. ${row.examples[0] ?? ""}`);
+  const total = (0, import_graph.nodesOfKind)(graphs.graph, import_code4.FILE).length;
+  return [`${total} source files. Directories, largest first \u2014 read one with sandbox.list --prefix.`, ...lines].join("\n");
+}
+function renderWhereIs(graphs, symbol) {
+  const sites = (0, import_graph.neighbors)(graphs.graph, (0, import_graph.nodeId)(import_code4.SYMBOL, symbol), { direction: "in", kinds: ["exports"] }).map((id) => ({
+    path: shortId(id),
+    pkg: (0, import_graph.neighbors)(graphs.graph, id, { direction: "in", kinds: ["contains"] })[0],
+    dependents: (0, import_graph.incident)(graphs.graph, id, { direction: "in", kinds: [import_code4.IMPORTS] }).length
+  })).sort((left, right) => right.dependents - left.dependents || left.path.localeCompare(right.path));
+  if (sites.length === 0) return `No exported symbol named "${symbol}". Try sandbox.search for a textual match.`;
+  return sites.slice(0, 20).map((site) => `${site.path}${site.pkg ? `  [${shortId(site.pkg)}]` : ""}  ${site.dependents} dependents`).join("\n");
+}
+function renderWhoImports(graphs, path) {
+  const id = (0, import_graph.nodeId)(import_code4.FILE, path);
+  const importers = (0, import_graph.neighbors)(graphs.graph, id, { direction: "in", kinds: [import_code4.IMPORTS] });
+  if (importers.length === 0) {
+    return graphs.graph.nodes.has(id) ? `Nothing imports ${path}. It is a leaf.` : `${path} is not a source file in this workspace.`;
+  }
+  return importers.slice(0, 40).map(shortId).sort().join("\n");
+}
+function renderWhoTouches(graphs, query) {
+  const needle = query.toLowerCase();
+  const hits = [...graphs.graph.nodes.values()].filter((node) => (node.kind === "table" || node.kind === "namespace") && node.name.toLowerCase().includes(needle)).slice(0, 10);
+  if (hits.length === 0) return `No table or namespace matching "${query}".`;
+  return hits.map((hit) => {
+    const side = (kind) => (0, import_graph.neighbors)(graphs.graph, hit.id, { direction: "in", kinds: [kind] }).map(shortId).sort().slice(0, 8);
+    return [
+      `${hit.name}  (${hit.kind})`,
+      `  writes: ${side(import_code4.WRITES).join(", ") || "(none)"}`,
+      `  reads:  ${side(import_code4.READS).join(", ") || "(none)"}`
+    ].join("\n");
+  }).join("\n\n");
+}
+
+// src/code-tool-edit.ts
+var MAX_TEXT_BYTES = 256 * 1024;
+function editAsDiff(path, oldText, newText) {
+  const removed = oldText.split("\n").map((line) => `-${line}`);
+  const added = newText.split("\n").map((line) => `+${line}`);
+  return `diff --git a/${path} b/${path}
+--- a/${path}
++++ b/${path}
+@@ -1,${removed.length} +1,${added.length} @@
+${[...removed, ...added].join("\n")}
+`;
+}
+function occurrences(haystack, needle) {
+  let count = 0;
+  for (let at = haystack.indexOf(needle); at !== -1; at = haystack.indexOf(needle, at + needle.length)) count += 1;
+  return count;
+}
+async function editCodeFile(workspaceDir, path, oldText, newText) {
+  if (!oldText) throw new TypeError("oldText must not be empty; to create a file use sandbox.apply_patch with a new-file diff");
+  if (Buffer.byteLength(oldText) > MAX_TEXT_BYTES || Buffer.byteLength(newText) > MAX_TEXT_BYTES) {
+    throw new TypeError(`oldText and newText must each stay under ${MAX_TEXT_BYTES} bytes; split the change`);
+  }
+  if (oldText.includes("\0") || newText.includes("\0")) throw new TypeError("edit text contains NUL bytes; use plain text");
+  const target = resolveCodePath(workspaceDir, path);
+  const current = await (0, import_promises10.readFile)(target, "utf8");
+  const found = occurrences(current, oldText);
+  if (found === 0) {
+    throw new TypeError(`oldText was not found in "${path}"; sandbox.read the current lines and copy them exactly, including indentation and blank lines`);
+  }
+  if (found > 1) {
+    throw new TypeError(`oldText occurs ${found} times in "${path}"; include more of the surrounding lines so it matches exactly once`);
+  }
+  const at = current.indexOf(oldText);
+  await (0, import_promises10.writeFile)(target, `${current.slice(0, at)}${newText}${current.slice(at + oldText.length)}`, "utf8");
+  return { deletions: oldText.split("\n").length, additions: newText.split("\n").length };
+}
+async function edit(context, request, options, policy, registry) {
+  exactKeys(request.input, ["path", "oldText", "newText"]);
+  const path = stringField(request.input, "path");
+  const oldText = stringField(request.input, "oldText");
+  const newText = typeof request.input.newText === "string" ? request.input.newText : stringField(request.input, "newText");
+  const paths2 = await registry.files(context.workspaceDir);
+  if (!paths2.includes(path)) {
+    throw new TypeError(`no such file in the staged workspace: "${path}". Use sandbox.overview, sandbox.where_is or sandbox.search to find the correct path.`);
+  }
+  if (options.readOnlyPrefixes?.some((prefix) => path === prefix || path.startsWith(`${prefix}/`))) {
+    throw new TypeError("edit targets a read-only reference source");
+  }
+  const patch2 = editAsDiff(path, oldText, newText);
+  const allowed = await policy.edit(policyContext(context, request, options, { patch: patch2 }));
+  if (!allowed) return response(request, false, "tool denied by CaMeL policy");
+  const changed = await editCodeFile(context.workspaceDir, path, oldText, newText);
+  registry.invalidate(context.workspaceDir);
+  forgetWorkspaceGraphs(context.workspaceDir);
+  return response(request, true, `Edited ${path}: -${changed.deletions} +${changed.additions} line(s).`, { paths: [path], ...changed });
+}
+
 // src/code-tool-reads.ts
-var import_promises11 = require("fs/promises");
+var import_promises12 = require("fs/promises");
 
 // src/code-tool-discovery.ts
 var import_node_child_process5 = require("child_process");
-var import_promises9 = require("fs/promises");
-var import_node_path10 = require("path");
+var import_promises11 = require("fs/promises");
+var import_node_path11 = require("path");
 var DEFAULT_MAX_FILES = 2e4;
 var DEFAULT_MAX_RESULTS = 100;
 var DEFAULT_MAX_FILE_BYTES = 512 * 1024;
@@ -2880,13 +3037,13 @@ function createWorkspaceFileRegistry(limit = DEFAULT_MAX_FILES, enumerate = regi
 async function registeredFiles(root, limit = DEFAULT_MAX_FILES) {
   const paths2 = [];
   const walk = async (directory) => {
-    for (const entry of await (0, import_promises9.readdir)(directory, { withFileTypes: true })) {
+    for (const entry of await (0, import_promises11.readdir)(directory, { withFileTypes: true })) {
       if (SKIP_WORKSPACE_DIRS.has(entry.name)) continue;
       if (entry.isSymbolicLink()) throw new TypeError("workspace contains a symbolic link");
-      const target = (0, import_node_path10.resolve)(directory, entry.name);
+      const target = (0, import_node_path11.resolve)(directory, entry.name);
       if (entry.isDirectory()) await walk(target);
       else if (entry.isFile()) {
-        const path = (0, import_node_path10.relative)(root, target).split("\\").join("/");
+        const path = (0, import_node_path11.relative)(root, target).split("\\").join("/");
         try {
           validateRelativePath(path);
         } catch {
@@ -2897,7 +3054,7 @@ async function registeredFiles(root, limit = DEFAULT_MAX_FILES) {
       }
     }
   };
-  await walk((0, import_node_path10.resolve)(root));
+  await walk((0, import_node_path11.resolve)(root));
   return paths2.sort();
 }
 function listWorkspace(paths2, options = {}) {
@@ -3011,7 +3168,7 @@ async function fallbackSearch(root, scoped, options) {
     if (matches.length >= options.maxResults) break;
     let source;
     try {
-      source = await (0, import_promises9.readFile)((0, import_node_path10.resolve)(root, path));
+      source = await (0, import_promises11.readFile)((0, import_node_path11.resolve)(root, path));
     } catch {
       continue;
     }
@@ -3028,67 +3185,6 @@ async function fallbackSearch(root, scoped, options) {
   return matches;
 }
 
-// src/code-tool-graph.ts
-var import_promises10 = require("fs/promises");
-var import_node_path11 = require("path");
-var import_graph = require("@odla-ai/graph");
-var import_code4 = require("@odla-ai/graph/code");
-var cache = /* @__PURE__ */ new Map();
-function workspaceGraphs(workspaceDir, paths2) {
-  const existing = cache.get(workspaceDir);
-  if (existing) return existing;
-  const read2 = (path) => (0, import_promises10.readFile)((0, import_node_path11.join)(workspaceDir, path), "utf8");
-  const built = (async () => ({
-    // No knownTables: a staged workspace may not carry migrations, and a filter
-    // that silently drops every table is worse than an unfiltered one. Callers
-    // with ground truth should build the graph themselves.
-    graph: await (0, import_code4.buildCodeGraph)({ paths: paths2, read: read2, data: { ignore: (path) => path.includes(".generated.") } })
-  }))();
-  cache.set(workspaceDir, built);
-  return built;
-}
-function forgetWorkspaceGraphs(workspaceDir) {
-  cache.delete(workspaceDir);
-}
-var shortId = (id) => id.slice(id.indexOf(":") + 1);
-function renderOverview(graphs, prefix) {
-  const rows = (0, import_graph.rollup)(graphs.graph, import_code4.FILE, prefix === void 0 ? {} : { prefix });
-  if (rows.length === 0) return prefix ? `No source under "${prefix}".` : "No source files.";
-  const lines = rows.slice(0, 60).map((row) => `${row.prefix}  (${row.count})  e.g. ${row.examples[0] ?? ""}`);
-  const total = (0, import_graph.nodesOfKind)(graphs.graph, import_code4.FILE).length;
-  return [`${total} source files. Directories, largest first \u2014 read one with sandbox.list --prefix.`, ...lines].join("\n");
-}
-function renderWhereIs(graphs, symbol) {
-  const sites = (0, import_graph.neighbors)(graphs.graph, (0, import_graph.nodeId)(import_code4.SYMBOL, symbol), { direction: "in", kinds: ["exports"] }).map((id) => ({
-    path: shortId(id),
-    pkg: (0, import_graph.neighbors)(graphs.graph, id, { direction: "in", kinds: ["contains"] })[0],
-    dependents: (0, import_graph.incident)(graphs.graph, id, { direction: "in", kinds: [import_code4.IMPORTS] }).length
-  })).sort((left, right) => right.dependents - left.dependents || left.path.localeCompare(right.path));
-  if (sites.length === 0) return `No exported symbol named "${symbol}". Try sandbox.search for a textual match.`;
-  return sites.slice(0, 20).map((site) => `${site.path}${site.pkg ? `  [${shortId(site.pkg)}]` : ""}  ${site.dependents} dependents`).join("\n");
-}
-function renderWhoImports(graphs, path) {
-  const id = (0, import_graph.nodeId)(import_code4.FILE, path);
-  const importers = (0, import_graph.neighbors)(graphs.graph, id, { direction: "in", kinds: [import_code4.IMPORTS] });
-  if (importers.length === 0) {
-    return graphs.graph.nodes.has(id) ? `Nothing imports ${path}. It is a leaf.` : `${path} is not a source file in this workspace.`;
-  }
-  return importers.slice(0, 40).map(shortId).sort().join("\n");
-}
-function renderWhoTouches(graphs, query) {
-  const needle = query.toLowerCase();
-  const hits = [...graphs.graph.nodes.values()].filter((node) => (node.kind === "table" || node.kind === "namespace") && node.name.toLowerCase().includes(needle)).slice(0, 10);
-  if (hits.length === 0) return `No table or namespace matching "${query}".`;
-  return hits.map((hit) => {
-    const side = (kind) => (0, import_graph.neighbors)(graphs.graph, hit.id, { direction: "in", kinds: [kind] }).map(shortId).sort().slice(0, 8);
-    return [
-      `${hit.name}  (${hit.kind})`,
-      `  writes: ${side(import_code4.WRITES).join(", ") || "(none)"}`,
-      `  reads:  ${side(import_code4.READS).join(", ") || "(none)"}`
-    ].join("\n");
-  }).join("\n\n");
-}
-
 // src/code-tool-reads.ts
 var GRAPH_TOOLS = /* @__PURE__ */ new Set([
   "sandbox.overview",
@@ -3111,11 +3207,11 @@ async function read(context, request, options, policy, registry) {
   const allowed = await policy.read(policyContext(context, request, options, { paths: paths2, path, startLine, endLine }));
   if (!allowed) return response(request, false, "tool denied by CaMeL policy");
   const target = resolveCodePath(context.workspaceDir, path);
-  const info = await (0, import_promises11.stat)(target);
+  const info = await (0, import_promises12.stat)(target);
   if (!info.isFile() || info.size > Math.max(options.maxReadBytes ?? 128 * 1024, 2 * 1024 * 1024)) {
     throw new TypeError("file is not a bounded regular source file");
   }
-  const source = await (0, import_promises11.readFile)(target);
+  const source = await (0, import_promises12.readFile)(target);
   if (source.includes(0)) throw new TypeError("binary files are not readable through this tool");
   const lines = source.toString("utf8").split("\n");
   const content = lines.slice(startLine - 1, endLine).join("\n");
@@ -3230,6 +3326,7 @@ async function route(context, request, options, recipes, policy, registry) {
     if (request.tool === "sandbox.search") return await search(context, request, options, policy, registry);
     if (GRAPH_TOOLS.has(request.tool)) return await graphQuery(context, request, options, policy, registry);
     if (request.tool === "sandbox.apply_patch") return await patch(context, request, options, policy, registry);
+    if (request.tool === "sandbox.edit") return await edit(context, request, options, policy, registry);
     return await recipe(context, request, options, recipes, policy);
   } catch (reason) {
     return response(request, false, toolFailureMessage(reason));
@@ -3916,6 +4013,9 @@ var TheseusRuntimeEngine = class {
       prompt: metadata.prompt,
       signal: active.abort.signal,
       readOnly: metadata.readOnly,
+      // A review-role session is read-only and still runs the proof; a planner
+      // marked read-only by its payload runs nothing.
+      recipes: metadata.role === "review" || !metadata.readOnly,
       recipeIds: this.options.recipes.map((recipe2) => recipe2.id),
       ...extraSkills.length ? { extraSkills } : {}
     });
@@ -4025,7 +4125,7 @@ function parse(argv) {
   };
 }
 async function readPolicy(path) {
-  const value = JSON.parse(await (0, import_promises12.readFile)(path, "utf8"));
+  const value = JSON.parse(await (0, import_promises13.readFile)(path, "utf8"));
   if (!value || Object.keys(value).some((key) => !["recipes", "recipeAuthorization"].includes(key)) || !Array.isArray(value.recipes) || !value.recipes.length || value.recipeAuthorization !== void 0 && value.recipeAuthorization !== "registered_recipe" && value.recipeAuthorization !== "exact_approval") throw new TypeError("invalid Code build policy file");
   const recipes = value.recipes;
   for (const recipe2 of recipes) assertCodeBuildRecipe(recipe2);