@mikode13/harness 2.1.0 → 2.3.0

package/dist/tools/infrastructure/boundedWorkspace.js

@@ -0,0 +1,160 @@
+import { createReadStream } from 'node:fs';
+import { lstat, open } from 'node:fs/promises';
+import { isAbsolute, join, matchesGlob, relative, resolve, sep } from 'node:path';
+import { createInterface } from 'node:readline';
+// A NUL byte this early means a binary file, as git and ripgrep both judge it.
+const binaryProbeBytes = 8000;
+// What a search holds while it sorts. A minified file can put a whole bundle on one line, and
+// a careless pattern can match every line of a repository.
+const maxStoredMatches = 50_000;
+const maxStoredLineLength = 1_000;
+// Far beyond what ripgrep needs on a large repository, so only a runaway search reaches it,
+// such as a pattern that backtracks without end in git's Perl engine.
+const defaultTimeoutMs = 30_000;
+function compareText(a, b) {
+    // Plain code-unit order: `localeCompare` would sort differently on every machine.
+    return a < b ? -1 : a > b ? 1 : 0;
+}
+/**
+ * Everything both implementations share: the boundary and the shape of every answer. A
+ * subclass only says which files are visible and where a pattern matches, across the whole
+ * root; the scope a model asks for is applied here, to those results, and never handed to the
+ * program underneath. ripgrep stops honouring `.gitignore` for a path it is given explicitly,
+ * so passing the model's scope down is how a secret would leak. See "`.gitignore` is the
+ * boundary of what an agent reads" in decisions.md.
+ *
+ * Every operation also stops after `timeoutMs`. A timeout is a failure the model reads and
+ * narrows its query from, not a cancellation: only the run's own signal cancels the run.
+ */
+export class BoundedWorkspace {
+    root;
+    timeoutMs;
+    constructor({ root, timeoutMs = defaultTimeoutMs }) {
+        this.root = root;
+        this.timeoutMs = timeoutMs;
+    }
+    listFiles(query, signal) {
+        return this.withTimeout(signal, async (bounded) => {
+            const inScope = [...new Set(await this.listCandidates(bounded))].filter(path => this.inScope(path, query));
+            const files = (await this.regularFiles(inScope)).sort(compareText);
+            return {
+                files: files.slice(0, query.limit),
+                total: files.length,
+                truncated: files.length > query.limit,
+            };
+        });
+    }
+    searchText(query, signal) {
+        return this.withTimeout(signal, async (bounded) => {
+            const listed = new Set(await this.listCandidates(bounded));
+            // The scope applies as the output arrives, so narrowing a search is what makes it
+            // lighter: matches outside it are never stored, however many the program prints.
+            const kept = [];
+            await this.findMatches(query.pattern, query.ignoreCase, bounded, match => {
+                if (!listed.has(match.path) || !this.inScope(match.path, query)) {
+                    return;
+                }
+                if (kept.length === maxStoredMatches) {
+                    throw new Error(`More than ${String(maxStoredMatches)} lines match; narrow the pattern, path or glob`);
+                }
+                kept.push({ ...match, text: match.text.slice(0, maxStoredLineLength) });
+            });
+            // Only the files that matched are checked on disk, not the whole repository.
+            const regular = new Set(await this.regularFiles([...new Set(kept.map(match => match.path))]));
+            const matches = kept
+                .filter(match => regular.has(match.path))
+                .sort((a, b) => compareText(a.path, b.path) || a.line - b.line);
+            return {
+                matches: matches.slice(0, query.limit),
+                total: matches.length,
+                truncated: matches.length > query.limit,
+            };
+        });
+    }
+    readFile(query, signal) {
+        return this.withTimeout(signal, async (bounded) => {
+            const path = this.toRelative(query.path);
+            const visible = path !== undefined &&
+                (await this.listCandidates(bounded)).includes(path) &&
+                (await this.regularFiles([path])).length === 1;
+            // One answer for a missing file and an ignored one, or the model would learn the secret exists.
+            if (!visible) {
+                throw new Error(`No such file: ${query.path}`);
+            }
+            const absolute = join(this.root, path);
+            if (await isBinary(absolute)) {
+                throw new Error(`Not a text file: ${query.path}`);
+            }
+            // Streamed, so a large file costs its length in lines, not in memory.
+            const reader = createInterface({
+                input: createReadStream(absolute, { encoding: 'utf8', signal: bounded }),
+                crlfDelay: Infinity,
+            });
+            const last = query.fromLine + query.lineCount - 1;
+            const lines = [];
+            let totalLines = 0;
+            for await (const line of reader) {
+                totalLines++;
+                if (totalLines >= query.fromLine && totalLines <= last) {
+                    lines.push(line);
+                }
+            }
+            return { lines, totalLines, truncated: totalLines > last };
+        });
+    }
+    /**
+     * Runs an operation under the run's signal and this workspace's time limit together. The
+     * run's cancellation escapes unchanged; the time limit becomes an error the model can act on.
+     */
+    async withTimeout(signal, operation) {
+        signal.throwIfAborted();
+        const timeout = AbortSignal.timeout(this.timeoutMs);
+        try {
+            return await operation(AbortSignal.any([signal, timeout]));
+        }
+        catch (error) {
+            signal.throwIfAborted();
+            if (timeout.aborted) {
+                throw new Error(`The operation took longer than ${String(this.timeoutMs / 1000)} s; narrow the pattern, path or glob`, { cause: error });
+            }
+            throw error;
+        }
+    }
+    /** Regular files only: a symlink could point out of the root, and a deleted file is gone. */
+    async regularFiles(paths) {
+        const regular = await Promise.all(paths.map(path => lstat(join(this.root, path)).then(stats => stats.isFile(), () => false)));
+        return paths.filter((_, index) => regular[index]);
+    }
+    inScope(path, { path: scope, glob }) {
+        if (scope !== undefined) {
+            const folder = this.toRelative(scope);
+            if (folder === undefined) {
+                return false;
+            }
+            if (folder !== '' && path !== folder && !path.startsWith(`${folder}/`)) {
+                return false;
+            }
+        }
+        return glob === undefined || matchesGlob(path, glob);
+    }
+    /** A path relative to the root with `/`, '' for the root itself, or undefined when it leaves it. */
+    toRelative(path) {
+        const relativePath = relative(this.root, resolve(this.root, path));
+        if (relativePath === '..' || relativePath.startsWith(`..${sep}`) || isAbsolute(relativePath)) {
+            return undefined;
+        }
+        return relativePath.split(sep).join('/');
+    }
+}
+async function isBinary(path) {
+    const file = await open(path);
+    try {
+        const probe = Buffer.alloc(binaryProbeBytes);
+        const { bytesRead } = await file.read(probe, 0, probe.length, 0);
+        return probe.subarray(0, bytesRead).includes(0);
+    }
+    finally {
+        await file.close();
+    }
+}
+//# sourceMappingURL=boundedWorkspace.js.map

package/dist/tools/infrastructure/boundedWorkspace.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"boundedWorkspace.js","sourceRoot":"","sources":["../../../src/tools/infrastructure/boundedWorkspace.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,gBAAgB,EAAE,MAAM,SAAS,CAAC;AAC3C,OAAO,EAAE,KAAK,EAAE,IAAI,EAAE,MAAM,kBAAkB,CAAC;AAC/C,OAAO,EAAE,UAAU,EAAE,IAAI,EAAE,WAAW,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,EAAE,MAAM,WAAW,CAAC;AAClF,OAAO,EAAE,eAAe,EAAE,MAAM,eAAe,CAAC;AAGhD,+EAA+E;AAC/E,MAAM,gBAAgB,GAAG,IAAI,CAAC;AAE9B,8FAA8F;AAC9F,2DAA2D;AAC3D,MAAM,gBAAgB,GAAG,MAAM,CAAC;AAChC,MAAM,mBAAmB,GAAG,KAAK,CAAC;AAElC,4FAA4F;AAC5F,sEAAsE;AACtE,MAAM,gBAAgB,GAAG,MAAM,CAAC;AAEhC,SAAS,WAAW,CAAC,CAAS,EAAE,CAAS;IACxC,kFAAkF;IAClF,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AACnC,CAAC;AAED;;;;;;;;;;GAUG;AACH,MAAM,OAAgB,gBAAgB;IAClB,IAAI,CAAS;IACf,SAAS,CAAS;IAEnC,YAAY,EAAE,IAAI,EAAE,SAAS,GAAG,gBAAgB,EAAwC;QACvF,IAAI,CAAC,IAAI,GAAG,IAAI,CAAC;QACjB,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC;IAC5B,CAAC;IAgBD,SAAS,CACR,KAAsD,EACtD,MAAmB;QAEnB,OAAO,IAAI,CAAC,WAAW,CAAC,MAAM,EAAE,KAAK,EAAC,OAAO,EAAC,EAAE;YAC/C,MAAM,OAAO,GAAG,CAAC,GAAG,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,cAAc,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,EAAE,CAC9E,IAAI,CAAC,OAAO,CAAC,IAAI,EAAE,KAAK,CAAC,CACzB,CAAC;YACF,MAAM,KAAK,GAAG,CAAC,MAAM,IAAI,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC;YAEnE,OAAO;gBACN,KAAK,EAAE,KAAK,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,KAAK,CAAC;gBAClC,KAAK,EAAE,KAAK,CAAC,MAAM;gBACnB,SAAS,EAAE,KAAK,CAAC,MAAM,GAAG,KAAK,CAAC,KAAK;aACrC,CAAC;QACH,CAAC,CAAC,CAAC;IACJ,CAAC;IAED,UAAU,CACT,KAA4F,EAC5F,MAAmB;QAEnB,OAAO,IAAI,CAAC,WAAW,CAAC,MAAM,EAAE,KAAK,EAAC,OAAO,EAAC,EAAE;YAC/C,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,cAAc,CAAC,OAAO,CAAC,CAAC,CAAC;YAC3D,kFAAkF;YAClF,iFAAiF;YACjF,MAAM,IAAI,GAAgB,EAAE,CAAC;YAC7B,MAAM,IAAI,CAAC,WAAW,CAAC,KAAK,CAAC,OAAO,EAAE,KAAK,CAAC,UAAU,EAAE,OAAO,EAAE,KAAK,CAAC,EAAE;gBACxE,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,IAAI,EAAE,KAAK,CAAC,EAAE,CAAC;oBACjE,OAAO;gBACR,CAAC;gBACD,IAAI,IAAI,CAAC,MAAM,KAAK,gBAAgB,EAAE,CAAC;oBACtC,MAAM,IAAI,KAAK,CACd,aAAa,MAAM,CAAC,gBAAgB,CAAC,gDAAgD,CACrF,CAAC;gBACH,CAAC;gBACD,IAAI,CAAC,IAAI,CAAC,EAAE,GAAG,KAAK,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,mBAAmB,CAAC,EAAE,CAAC,CAAC;YACzE,CAAC,CAAC,CAAC;YAEH,6EAA6E;YAC7E,MAAM,OAAO,GAAG,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,YAAY,CAAC,CAAC,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAC9F,MAAM,OAAO,GAAG,IAAI;iBAClB,MAAM,CAAC,KAAK,CAAC,EAAE,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;iBACxC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,WAAW,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;YAEjE,OAAO;gBACN,OAAO,EAAE,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,KAAK,CAAC;gBACtC,KAAK,EAAE,OAAO,CAAC,MAAM;gBACrB,SAAS,EAAE,OAAO,CAAC,MAAM,GAAG,KAAK,CAAC,KAAK;aACvC,CAAC;QACH,CAAC,CAAC,CAAC;IACJ,CAAC;IAED,QAAQ,CACP,KAA4D,EAC5D,MAAmB;QAEnB,OAAO,IAAI,CAAC,WAAW,CAAC,MAAM,EAAE,KAAK,EAAC,OAAO,EAAC,EAAE;YAC/C,MAAM,IAAI,GAAG,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;YACzC,MAAM,OAAO,GACZ,IAAI,KAAK,SAAS;gBAClB,CAAC,MAAM,IAAI,CAAC,cAAc,CAAC,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC;gBACnD,CAAC,MAAM,IAAI,CAAC,YAAY,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,MAAM,KAAK,CAAC,CAAC;YAChD,gGAAgG;YAChG,IAAI,CAAC,OAAO,EAAE,CAAC;gBACd,MAAM,IAAI,KAAK,CAAC,iBAAiB,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC;YAChD,CAAC;YAED,MAAM,QAAQ,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;YACvC,IAAI,MAAM,QAAQ,CAAC,QAAQ,CAAC,EAAE,CAAC;gBAC9B,MAAM,IAAI,KAAK,CAAC,oBAAoB,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC;YACnD,CAAC;YAED,sEAAsE;YACtE,MAAM,MAAM,GAAG,eAAe,CAAC;gBAC9B,KAAK,EAAE,gBAAgB,CAAC,QAAQ,EAAE,EAAE,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC;gBACxE,SAAS,EAAE,QAAQ;aACnB,CAAC,CAAC;YACH,MAAM,IAAI,GAAG,KAAK,CAAC,QAAQ,GAAG,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC;YAClD,MAAM,KAAK,GAAa,EAAE,CAAC;YAC3B,IAAI,UAAU,GAAG,CAAC,CAAC;YACnB,IAAI,KAAK,EAAE,MAAM,IAAI,IAAI,MAAM,EAAE,CAAC;gBACjC,UAAU,EAAE,CAAC;gBACb,IAAI,UAAU,IAAI,KAAK,CAAC,QAAQ,IAAI,UAAU,IAAI,IAAI,EAAE,CAAC;oBACxD,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;gBAClB,CAAC;YACF,CAAC;YAED,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE,SAAS,EAAE,UAAU,GAAG,IAAI,EAAE,CAAC;QAC5D,CAAC,CAAC,CAAC;IACJ,CAAC;IAED;;;OAGG;IACK,KAAK,CAAC,WAAW,CACxB,MAAmB,EACnB,SAAoD;QAEpD,MAAM,CAAC,cAAc,EAAE,CAAC;QACxB,MAAM,OAAO,GAAG,WAAW,CAAC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QACpD,IAAI,CAAC;YACJ,OAAO,MAAM,SAAS,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC;QAC5D,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,CAAC,cAAc,EAAE,CAAC;YACxB,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC;gBACrB,MAAM,IAAI,KAAK,CACd,kCAAkC,MAAM,CAAC,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,sCAAsC,EACrG,EAAE,KAAK,EAAE,KAAK,EAAE,CAChB,CAAC;YACH,CAAC;YACD,MAAM,KAAK,CAAC;QACb,CAAC;IACF,CAAC;IAED,6FAA6F;IACrF,KAAK,CAAC,YAAY,CAAC,KAAe;QACzC,MAAM,OAAO,GAAG,MAAM,OAAO,CAAC,GAAG,CAChC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CAChB,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC,IAAI,CAChC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,MAAM,EAAE,EACvB,GAAG,EAAE,CAAC,KAAK,CACX,CACD,CACD,CAAC;QAEF,OAAO,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,EAAE,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC;IACnD,CAAC;IAEO,OAAO,CAAC,IAAY,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAoC;QACpF,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;YACzB,MAAM,MAAM,GAAG,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;YACtC,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;gBAC1B,OAAO,KAAK,CAAC;YACd,CAAC;YACD,IAAI,MAAM,KAAK,EAAE,IAAI,IAAI,KAAK,MAAM,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,MAAM,GAAG,CAAC,EAAE,CAAC;gBACxE,OAAO,KAAK,CAAC;YACd,CAAC;QACF,CAAC;QAED,OAAO,IAAI,KAAK,SAAS,IAAI,WAAW,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;IACtD,CAAC;IAED,oGAAoG;IAC5F,UAAU,CAAC,IAAY;QAC9B,MAAM,YAAY,GAAG,QAAQ,CAAC,IAAI,CAAC,IAAI,EAAE,OAAO,CAAC,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;QACnE,IAAI,YAAY,KAAK,IAAI,IAAI,YAAY,CAAC,UAAU,CAAC,KAAK,GAAG,EAAE,CAAC,IAAI,UAAU,CAAC,YAAY,CAAC,EAAE,CAAC;YAC9F,OAAO,SAAS,CAAC;QAClB,CAAC;QAED,OAAO,YAAY,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IAC1C,CAAC;CACD;AAED,KAAK,UAAU,QAAQ,CAAC,IAAY;IACnC,MAAM,IAAI,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,CAAC;IAC9B,IAAI,CAAC;QACJ,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK,CAAC,gBAAgB,CAAC,CAAC;QAC7C,MAAM,EAAE,SAAS,EAAE,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,KAAK,EAAE,CAAC,EAAE,KAAK,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;QACjE,OAAO,KAAK,CAAC,QAAQ,CAAC,CAAC,EAAE,SAAS,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC;IACjD,CAAC;YAAS,CAAC;QACV,MAAM,IAAI,CAAC,KAAK,EAAE,CAAC;IACpB,CAAC;AACF,CAAC","sourcesContent":["import { createReadStream } from 'node:fs';\nimport { lstat, open } from 'node:fs/promises';\nimport { isAbsolute, join, matchesGlob, relative, resolve, sep } from 'node:path';\nimport { createInterface } from 'node:readline';\nimport type { TextMatch, Workspace } from '../domain/workspace.ts';\n\n// A NUL byte this early means a binary file, as git and ripgrep both judge it.\nconst binaryProbeBytes = 8000;\n\n// What a search holds while it sorts. A minified file can put a whole bundle on one line, and\n// a careless pattern can match every line of a repository.\nconst maxStoredMatches = 50_000;\nconst maxStoredLineLength = 1_000;\n\n// Far beyond what ripgrep needs on a large repository, so only a runaway search reaches it,\n// such as a pattern that backtracks without end in git's Perl engine.\nconst defaultTimeoutMs = 30_000;\n\nfunction compareText(a: string, b: string): number {\n\t// Plain code-unit order: `localeCompare` would sort differently on every machine.\n\treturn a < b ? -1 : a > b ? 1 : 0;\n}\n\n/**\n * Everything both implementations share: the boundary and the shape of every answer. A\n * subclass only says which files are visible and where a pattern matches, across the whole\n * root; the scope a model asks for is applied here, to those results, and never handed to the\n * program underneath. ripgrep stops honouring `.gitignore` for a path it is given explicitly,\n * so passing the model's scope down is how a secret would leak. See \"`.gitignore` is the\n * boundary of what an agent reads\" in decisions.md.\n *\n * Every operation also stops after `timeoutMs`. A timeout is a failure the model reads and\n * narrows its query from, not a cancellation: only the run's own signal cancels the run.\n */\nexport abstract class BoundedWorkspace implements Workspace {\n\tprotected readonly root: string;\n\tprivate readonly timeoutMs: number;\n\n\tconstructor({ root, timeoutMs = defaultTimeoutMs }: { root: string; timeoutMs?: number }) {\n\t\tthis.root = root;\n\t\tthis.timeoutMs = timeoutMs;\n\t}\n\n\t/** Every file the program does not ignore, relative to the root, in any order. */\n\tprotected abstract listCandidates(signal: AbortSignal): Promise<string[]>;\n\n\t/**\n\t * Hands `keep` every line matching `pattern` in a file the program does not ignore, in any\n\t * order, as the program prints it. Nothing is held here: `keep` decides what to store.\n\t */\n\tprotected abstract findMatches(\n\t\tpattern: string,\n\t\tignoreCase: boolean,\n\t\tsignal: AbortSignal,\n\t\tkeep: (match: TextMatch) => void,\n\t): Promise<void>;\n\n\tlistFiles(\n\t\tquery: { path?: string; glob?: string; limit: number },\n\t\tsignal: AbortSignal,\n\t): Promise<{ files: string[]; total: number; truncated: boolean }> {\n\t\treturn this.withTimeout(signal, async bounded => {\n\t\t\tconst inScope = [...new Set(await this.listCandidates(bounded))].filter(path =>\n\t\t\t\tthis.inScope(path, query),\n\t\t\t);\n\t\t\tconst files = (await this.regularFiles(inScope)).sort(compareText);\n\n\t\t\treturn {\n\t\t\t\tfiles: files.slice(0, query.limit),\n\t\t\t\ttotal: files.length,\n\t\t\t\ttruncated: files.length > query.limit,\n\t\t\t};\n\t\t});\n\t}\n\n\tsearchText(\n\t\tquery: { pattern: string; ignoreCase: boolean; path?: string; glob?: string; limit: number },\n\t\tsignal: AbortSignal,\n\t): Promise<{ matches: TextMatch[]; total: number; truncated: boolean }> {\n\t\treturn this.withTimeout(signal, async bounded => {\n\t\t\tconst listed = new Set(await this.listCandidates(bounded));\n\t\t\t// The scope applies as the output arrives, so narrowing a search is what makes it\n\t\t\t// lighter: matches outside it are never stored, however many the program prints.\n\t\t\tconst kept: TextMatch[] = [];\n\t\t\tawait this.findMatches(query.pattern, query.ignoreCase, bounded, match => {\n\t\t\t\tif (!listed.has(match.path) || !this.inScope(match.path, query)) {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tif (kept.length === maxStoredMatches) {\n\t\t\t\t\tthrow new Error(\n\t\t\t\t\t\t`More than ${String(maxStoredMatches)} lines match; narrow the pattern, path or glob`,\n\t\t\t\t\t);\n\t\t\t\t}\n\t\t\t\tkept.push({ ...match, text: match.text.slice(0, maxStoredLineLength) });\n\t\t\t});\n\n\t\t\t// Only the files that matched are checked on disk, not the whole repository.\n\t\t\tconst regular = new Set(await this.regularFiles([...new Set(kept.map(match => match.path))]));\n\t\t\tconst matches = kept\n\t\t\t\t.filter(match => regular.has(match.path))\n\t\t\t\t.sort((a, b) => compareText(a.path, b.path) || a.line - b.line);\n\n\t\t\treturn {\n\t\t\t\tmatches: matches.slice(0, query.limit),\n\t\t\t\ttotal: matches.length,\n\t\t\t\ttruncated: matches.length > query.limit,\n\t\t\t};\n\t\t});\n\t}\n\n\treadFile(\n\t\tquery: { path: string; fromLine: number; lineCount: number },\n\t\tsignal: AbortSignal,\n\t): Promise<{ lines: string[]; totalLines: number; truncated: boolean }> {\n\t\treturn this.withTimeout(signal, async bounded => {\n\t\t\tconst path = this.toRelative(query.path);\n\t\t\tconst visible =\n\t\t\t\tpath !== undefined &&\n\t\t\t\t(await this.listCandidates(bounded)).includes(path) &&\n\t\t\t\t(await this.regularFiles([path])).length === 1;\n\t\t\t// One answer for a missing file and an ignored one, or the model would learn the secret exists.\n\t\t\tif (!visible) {\n\t\t\t\tthrow new Error(`No such file: ${query.path}`);\n\t\t\t}\n\n\t\t\tconst absolute = join(this.root, path);\n\t\t\tif (await isBinary(absolute)) {\n\t\t\t\tthrow new Error(`Not a text file: ${query.path}`);\n\t\t\t}\n\n\t\t\t// Streamed, so a large file costs its length in lines, not in memory.\n\t\t\tconst reader = createInterface({\n\t\t\t\tinput: createReadStream(absolute, { encoding: 'utf8', signal: bounded }),\n\t\t\t\tcrlfDelay: Infinity,\n\t\t\t});\n\t\t\tconst last = query.fromLine + query.lineCount - 1;\n\t\t\tconst lines: string[] = [];\n\t\t\tlet totalLines = 0;\n\t\t\tfor await (const line of reader) {\n\t\t\t\ttotalLines++;\n\t\t\t\tif (totalLines >= query.fromLine && totalLines <= last) {\n\t\t\t\t\tlines.push(line);\n\t\t\t\t}\n\t\t\t}\n\n\t\t\treturn { lines, totalLines, truncated: totalLines > last };\n\t\t});\n\t}\n\n\t/**\n\t * Runs an operation under the run's signal and this workspace's time limit together. The\n\t * run's cancellation escapes unchanged; the time limit becomes an error the model can act on.\n\t */\n\tprivate async withTimeout<Result>(\n\t\tsignal: AbortSignal,\n\t\toperation: (bounded: AbortSignal) => Promise<Result>,\n\t): Promise<Result> {\n\t\tsignal.throwIfAborted();\n\t\tconst timeout = AbortSignal.timeout(this.timeoutMs);\n\t\ttry {\n\t\t\treturn await operation(AbortSignal.any([signal, timeout]));\n\t\t} catch (error) {\n\t\t\tsignal.throwIfAborted();\n\t\t\tif (timeout.aborted) {\n\t\t\t\tthrow new Error(\n\t\t\t\t\t`The operation took longer than ${String(this.timeoutMs / 1000)} s; narrow the pattern, path or glob`,\n\t\t\t\t\t{ cause: error },\n\t\t\t\t);\n\t\t\t}\n\t\t\tthrow error;\n\t\t}\n\t}\n\n\t/** Regular files only: a symlink could point out of the root, and a deleted file is gone. */\n\tprivate async regularFiles(paths: string[]): Promise<string[]> {\n\t\tconst regular = await Promise.all(\n\t\t\tpaths.map(path =>\n\t\t\t\tlstat(join(this.root, path)).then(\n\t\t\t\t\tstats => stats.isFile(),\n\t\t\t\t\t() => false,\n\t\t\t\t),\n\t\t\t),\n\t\t);\n\n\t\treturn paths.filter((_, index) => regular[index]);\n\t}\n\n\tprivate inScope(path: string, { path: scope, glob }: { path?: string; glob?: string }): boolean {\n\t\tif (scope !== undefined) {\n\t\t\tconst folder = this.toRelative(scope);\n\t\t\tif (folder === undefined) {\n\t\t\t\treturn false;\n\t\t\t}\n\t\t\tif (folder !== '' && path !== folder && !path.startsWith(`${folder}/`)) {\n\t\t\t\treturn false;\n\t\t\t}\n\t\t}\n\n\t\treturn glob === undefined || matchesGlob(path, glob);\n\t}\n\n\t/** A path relative to the root with `/`, '' for the root itself, or undefined when it leaves it. */\n\tprivate toRelative(path: string): string | undefined {\n\t\tconst relativePath = relative(this.root, resolve(this.root, path));\n\t\tif (relativePath === '..' || relativePath.startsWith(`..${sep}`) || isAbsolute(relativePath)) {\n\t\t\treturn undefined;\n\t\t}\n\n\t\treturn relativePath.split(sep).join('/');\n\t}\n}\n\nasync function isBinary(path: string): Promise<boolean> {\n\tconst file = await open(path);\n\ttry {\n\t\tconst probe = Buffer.alloc(binaryProbeBytes);\n\t\tconst { bytesRead } = await file.read(probe, 0, probe.length, 0);\n\t\treturn probe.subarray(0, bytesRead).includes(0);\n\t} finally {\n\t\tawait file.close();\n\t}\n}\n"]}

package/dist/tools/infrastructure/createWorkspace.d.ts

@@ -0,0 +1,14 @@
+import type { ILogger } from '#src/shared/domain/logger';
+import type { Workspace } from '../domain/workspace.ts';
+/**
+ * Picks how to read `root`: the ripgrep that ships with the package, or git when this platform
+ * has no ripgrep binary, which is logged because it is a degraded path. Plain `grep` is never a
+ * fallback: it knows nothing of `.gitignore`. See "`.gitignore` is the boundary of what an agent
+ * reads" in decisions.md.
+ *
+ * @throws {UnrecoverableError} when neither ripgrep nor a git repository is available.
+ */
+export declare function createWorkspace({ root, logger, }: {
+    root: string;
+    logger: ILogger;
+}): Promise<Workspace>;

package/dist/tools/infrastructure/createWorkspace.js

@@ -0,0 +1,54 @@
+import { UnrecoverableError } from '#src/agent/domain/errors';
+import { classifyHostFailure, treatErrors } from '#src/agent/domain/providerFailure';
+import { GitWorkspace } from "./gitWorkspace.js";
+import { RipgrepWorkspace } from "./ripgrepWorkspace.js";
+import { runProcess } from "./runProcess.js";
+// Long enough for a slow disk; short enough that a hung git does not hold up building an agent.
+const gitCheckTimeoutMs = 5_000;
+async function loadRipgrep() {
+    try {
+        // Dynamic: the package throws on import when the platform has no binary.
+        const { rgPath } = await import('@vscode/ripgrep');
+        return rgPath;
+    }
+    catch {
+        return undefined;
+    }
+}
+async function isGitWorkTree(root) {
+    try {
+        const { stdout, exitCode } = await runProcess('git', ['rev-parse', '--is-inside-work-tree'], {
+            cwd: root,
+            signal: AbortSignal.timeout(gitCheckTimeoutMs),
+        });
+        return exitCode === 0 && stdout.trim() === 'true';
+    }
+    catch {
+        // No git at all, or one that never answered.
+        return false;
+    }
+}
+/**
+ * Picks how to read `root`: the ripgrep that ships with the package, or git when this platform
+ * has no ripgrep binary, which is logged because it is a degraded path. Plain `grep` is never a
+ * fallback: it knows nothing of `.gitignore`. See "`.gitignore` is the boundary of what an agent
+ * reads" in decisions.md.
+ *
+ * @throws {UnrecoverableError} when neither ripgrep nor a git repository is available.
+ */
+export async function createWorkspace({ root, logger, }) {
+    const ripgrepPath = await loadRipgrep();
+    if (ripgrepPath !== undefined) {
+        return new RipgrepWorkspace({ root, ripgrepPath });
+    }
+    if (await isGitWorkTree(root)) {
+        treatErrors(() => {
+            logger.warn('ripgrep is not available on this platform; reading the repository through git');
+        }, classifyHostFailure, 'Workspace logger failed while choosing an implementation');
+        return new GitWorkspace({ root });
+    }
+    throw new UnrecoverableError('No program can read the workspace', {
+        cause: `ripgrep is not available on this platform, and ${root} is not a git repository.`,
+    });
+}
+//# sourceMappingURL=createWorkspace.js.map

package/dist/tools/infrastructure/createWorkspace.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"createWorkspace.js","sourceRoot":"","sources":["../../../src/tools/infrastructure/createWorkspace.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,kBAAkB,EAAE,MAAM,0BAA0B,CAAC;AAC9D,OAAO,EAAE,mBAAmB,EAAE,WAAW,EAAE,MAAM,mCAAmC,CAAC;AAGrF,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AACjD,OAAO,EAAE,gBAAgB,EAAE,MAAM,uBAAuB,CAAC;AACzD,OAAO,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAE7C,gGAAgG;AAChG,MAAM,iBAAiB,GAAG,KAAK,CAAC;AAEhC,KAAK,UAAU,WAAW;IACzB,IAAI,CAAC;QACJ,yEAAyE;QACzE,MAAM,EAAE,MAAM,EAAE,GAAG,MAAM,MAAM,CAAC,iBAAiB,CAAC,CAAC;QACnD,OAAO,MAAM,CAAC;IACf,CAAC;IAAC,MAAM,CAAC;QACR,OAAO,SAAS,CAAC;IAClB,CAAC;AACF,CAAC;AAED,KAAK,UAAU,aAAa,CAAC,IAAY;IACxC,IAAI,CAAC;QACJ,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,GAAG,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,WAAW,EAAE,uBAAuB,CAAC,EAAE;YAC5F,GAAG,EAAE,IAAI;YACT,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,iBAAiB,CAAC;SAC9C,CAAC,CAAC;QACH,OAAO,QAAQ,KAAK,CAAC,IAAI,MAAM,CAAC,IAAI,EAAE,KAAK,MAAM,CAAC;IACnD,CAAC;IAAC,MAAM,CAAC;QACR,6CAA6C;QAC7C,OAAO,KAAK,CAAC;IACd,CAAC;AACF,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,eAAe,CAAC,EACrC,IAAI,EACJ,MAAM,GAIN;IACA,MAAM,WAAW,GAAG,MAAM,WAAW,EAAE,CAAC;IACxC,IAAI,WAAW,KAAK,SAAS,EAAE,CAAC;QAC/B,OAAO,IAAI,gBAAgB,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,CAAC,CAAC;IACpD,CAAC;IAED,IAAI,MAAM,aAAa,CAAC,IAAI,CAAC,EAAE,CAAC;QAC/B,WAAW,CACV,GAAG,EAAE;YACJ,MAAM,CAAC,IAAI,CACV,+EAA+E,CAC/E,CAAC;QACH,CAAC,EACD,mBAAmB,EACnB,0DAA0D,CAC1D,CAAC;QACF,OAAO,IAAI,YAAY,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC;IACnC,CAAC;IAED,MAAM,IAAI,kBAAkB,CAAC,mCAAmC,EAAE;QACjE,KAAK,EAAE,kDAAkD,IAAI,2BAA2B;KACxF,CAAC,CAAC;AACJ,CAAC","sourcesContent":["import { UnrecoverableError } from '#src/agent/domain/errors';\nimport { classifyHostFailure, treatErrors } from '#src/agent/domain/providerFailure';\nimport type { ILogger } from '#src/shared/domain/logger';\nimport type { Workspace } from '../domain/workspace.ts';\nimport { GitWorkspace } from './gitWorkspace.ts';\nimport { RipgrepWorkspace } from './ripgrepWorkspace.ts';\nimport { runProcess } from './runProcess.ts';\n\n// Long enough for a slow disk; short enough that a hung git does not hold up building an agent.\nconst gitCheckTimeoutMs = 5_000;\n\nasync function loadRipgrep(): Promise<string | undefined> {\n\ttry {\n\t\t// Dynamic: the package throws on import when the platform has no binary.\n\t\tconst { rgPath } = await import('@vscode/ripgrep');\n\t\treturn rgPath;\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nasync function isGitWorkTree(root: string): Promise<boolean> {\n\ttry {\n\t\tconst { stdout, exitCode } = await runProcess('git', ['rev-parse', '--is-inside-work-tree'], {\n\t\t\tcwd: root,\n\t\t\tsignal: AbortSignal.timeout(gitCheckTimeoutMs),\n\t\t});\n\t\treturn exitCode === 0 && stdout.trim() === 'true';\n\t} catch {\n\t\t// No git at all, or one that never answered.\n\t\treturn false;\n\t}\n}\n\n/**\n * Picks how to read `root`: the ripgrep that ships with the package, or git when this platform\n * has no ripgrep binary, which is logged because it is a degraded path. Plain `grep` is never a\n * fallback: it knows nothing of `.gitignore`. See \"`.gitignore` is the boundary of what an agent\n * reads\" in decisions.md.\n *\n * @throws {UnrecoverableError} when neither ripgrep nor a git repository is available.\n */\nexport async function createWorkspace({\n\troot,\n\tlogger,\n}: {\n\troot: string;\n\tlogger: ILogger;\n}): Promise<Workspace> {\n\tconst ripgrepPath = await loadRipgrep();\n\tif (ripgrepPath !== undefined) {\n\t\treturn new RipgrepWorkspace({ root, ripgrepPath });\n\t}\n\n\tif (await isGitWorkTree(root)) {\n\t\ttreatErrors(\n\t\t\t() => {\n\t\t\t\tlogger.warn(\n\t\t\t\t\t'ripgrep is not available on this platform; reading the repository through git',\n\t\t\t\t);\n\t\t\t},\n\t\t\tclassifyHostFailure,\n\t\t\t'Workspace logger failed while choosing an implementation',\n\t\t);\n\t\treturn new GitWorkspace({ root });\n\t}\n\n\tthrow new UnrecoverableError('No program can read the workspace', {\n\t\tcause: `ripgrep is not available on this platform, and ${root} is not a git repository.`,\n\t});\n}\n"]}

package/dist/tools/infrastructure/defineTool.d.ts

@@ -0,0 +1,15 @@
+import { z } from 'zod';
+import type { Tool } from '../domain/tool.ts';
+/**
+ * Builds a `Tool` from one Zod schema, which gives the model its JSON Schema, validates what the
+ * model sends, and types the input `execute` receives. Invalid input is rejected with a message
+ * naming what is wrong, which the agent hands back to the model as an error result.
+ *
+ * @throws {InvalidAgentConfigError} when the schema falls outside what strict mode supports.
+ */
+export declare function defineTool<Schema extends z.ZodObject>({ name, description, input, execute, }: {
+    name: string;
+    description: string;
+    input: Schema;
+    execute: (input: z.infer<Schema>, signal: AbortSignal) => Promise<string>;
+}): Tool;

package/dist/tools/infrastructure/defineTool.js

@@ -0,0 +1,87 @@
+import { z } from 'zod';
+import { InvalidAgentConfigError } from '#src/agent/domain/errors';
+// Both providers take tools in strict mode, which supports only part of JSON Schema. Claude
+// rejects these bounds; a tool that needs one checks it in its own code instead.
+const unsupportedKeywords = [
+    'minimum',
+    'maximum',
+    'exclusiveMinimum',
+    'exclusiveMaximum',
+    'multipleOf',
+    'minLength',
+    'maxLength',
+];
+// Keywords whose value is one schema, or a list or map of them. Anything else, such as the
+// keys of `properties`, is a name or a value, never a schema to check.
+const schemaKeywords = ['items', 'not', 'additionalProperties'];
+const schemaListKeywords = ['anyOf', 'oneOf', 'allOf', 'prefixItems'];
+const schemaMapKeywords = ['properties', '$defs', 'definitions'];
+function isSchema(value) {
+    return typeof value === 'object' && value !== null && !Array.isArray(value);
+}
+/**
+ * Walks the whole schema, nested objects included, and visits only schemas: a property named
+ * `maxLength` is a name, not the keyword. A schema the providers reject would fail every
+ * request with a 400, so it fails here, when the tool is defined.
+ */
+function assertStrict(schema, toolName) {
+    if (!isSchema(schema)) {
+        return;
+    }
+    const keyword = unsupportedKeywords.find(candidate => candidate in schema);
+    if (keyword) {
+        throw new InvalidAgentConfigError(`Tool "${toolName}" uses "${keyword}", which strict mode does not support; check it in the tool's code`);
+    }
+    if (schema.type === 'object') {
+        // A record or a loose object takes keys strict mode cannot know in advance.
+        if (schema.additionalProperties !== false) {
+            throw new InvalidAgentConfigError(`Tool "${toolName}" has an object open to any key; strict mode needs every key declared`);
+        }
+        const required = (schema.required ?? []);
+        // OpenAI requires every property: an optional one is a union with `null`.
+        const optional = Object.keys(schema.properties ?? {}).filter(property => !required.includes(property));
+        if (optional.length > 0) {
+            throw new InvalidAgentConfigError(`Tool "${toolName}" has optional fields (${optional.join(', ')}); make them .nullable() instead`);
+        }
+    }
+    for (const key of schemaKeywords)
+        assertStrict(schema[key], toolName);
+    for (const key of schemaListKeywords) {
+        const list = schema[key];
+        if (Array.isArray(list))
+            for (const item of list)
+                assertStrict(item, toolName);
+    }
+    for (const key of schemaMapKeywords) {
+        const map = schema[key];
+        if (isSchema(map))
+            for (const item of Object.values(map))
+                assertStrict(item, toolName);
+    }
+}
+/**
+ * Builds a `Tool` from one Zod schema, which gives the model its JSON Schema, validates what the
+ * model sends, and types the input `execute` receives. Invalid input is rejected with a message
+ * naming what is wrong, which the agent hands back to the model as an error result.
+ *
+ * @throws {InvalidAgentConfigError} when the schema falls outside what strict mode supports.
+ */
+export function defineTool({ name, description, input, execute, }) {
+    const inputSchema = { ...z.toJSONSchema(input) };
+    // The dialect marker is noise to both APIs.
+    delete inputSchema.$schema;
+    assertStrict(inputSchema, name);
+    return {
+        name,
+        description,
+        inputSchema: inputSchema,
+        execute: (raw, signal) => {
+            const parsed = input.safeParse(raw);
+            if (!parsed.success) {
+                return Promise.reject(new Error(z.prettifyError(parsed.error)));
+            }
+            return execute(parsed.data, signal);
+        },
+    };
+}
+//# sourceMappingURL=defineTool.js.map

package/dist/tools/infrastructure/defineTool.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"defineTool.js","sourceRoot":"","sources":["../../../src/tools/infrastructure/defineTool.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EAAE,uBAAuB,EAAE,MAAM,0BAA0B,CAAC;AAInE,4FAA4F;AAC5F,iFAAiF;AACjF,MAAM,mBAAmB,GAAG;IAC3B,SAAS;IACT,SAAS;IACT,kBAAkB;IAClB,kBAAkB;IAClB,YAAY;IACZ,WAAW;IACX,WAAW;CACX,CAAC;AAEF,2FAA2F;AAC3F,uEAAuE;AACvE,MAAM,cAAc,GAAG,CAAC,OAAO,EAAE,KAAK,EAAE,sBAAsB,CAAC,CAAC;AAChE,MAAM,kBAAkB,GAAG,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,aAAa,CAAC,CAAC;AACtE,MAAM,iBAAiB,GAAG,CAAC,YAAY,EAAE,OAAO,EAAE,aAAa,CAAC,CAAC;AAEjE,SAAS,QAAQ,CAAC,KAAc;IAC/B,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC7E,CAAC;AAED;;;;GAIG;AACH,SAAS,YAAY,CAAC,MAAe,EAAE,QAAgB;IACtD,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;QACvB,OAAO;IACR,CAAC;IAED,MAAM,OAAO,GAAG,mBAAmB,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,CAAC,SAAS,IAAI,MAAM,CAAC,CAAC;IAC3E,IAAI,OAAO,EAAE,CAAC;QACb,MAAM,IAAI,uBAAuB,CAChC,SAAS,QAAQ,WAAW,OAAO,oEAAoE,CACvG,CAAC;IACH,CAAC;IAED,IAAI,MAAM,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;QAC9B,4EAA4E;QAC5E,IAAI,MAAM,CAAC,oBAAoB,KAAK,KAAK,EAAE,CAAC;YAC3C,MAAM,IAAI,uBAAuB,CAChC,SAAS,QAAQ,uEAAuE,CACxF,CAAC;QACH,CAAC;QAED,MAAM,QAAQ,GAAG,CAAC,MAAM,CAAC,QAAQ,IAAI,EAAE,CAAa,CAAC;QACrD,0EAA0E;QAC1E,MAAM,QAAQ,GAAG,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,UAAU,IAAI,EAAE,CAAC,CAAC,MAAM,CAC3D,QAAQ,CAAC,EAAE,CAAC,CAAC,QAAQ,CAAC,QAAQ,CAAC,QAAQ,CAAC,CACxC,CAAC;QACF,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACzB,MAAM,IAAI,uBAAuB,CAChC,SAAS,QAAQ,0BAA0B,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,kCAAkC,CAChG,CAAC;QACH,CAAC;IACF,CAAC;IAED,KAAK,MAAM,GAAG,IAAI,cAAc;QAAE,YAAY,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE,QAAQ,CAAC,CAAC;IACtE,KAAK,MAAM,GAAG,IAAI,kBAAkB,EAAE,CAAC;QACtC,MAAM,IAAI,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC;QACzB,IAAI,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC;YAAE,KAAK,MAAM,IAAI,IAAI,IAAI;gBAAE,YAAY,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC;IAChF,CAAC;IACD,KAAK,MAAM,GAAG,IAAI,iBAAiB,EAAE,CAAC;QACrC,MAAM,GAAG,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC;QACxB,IAAI,QAAQ,CAAC,GAAG,CAAC;YAAE,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,MAAM,CAAC,GAAG,CAAC;gBAAE,YAAY,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC;IACxF,CAAC;AACF,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,UAAU,CAA6B,EACtD,IAAI,EACJ,WAAW,EACX,KAAK,EACL,OAAO,GAMP;IACA,MAAM,WAAW,GAA4B,EAAE,GAAG,CAAC,CAAC,YAAY,CAAC,KAAK,CAAC,EAAE,CAAC;IAC1E,4CAA4C;IAC5C,OAAO,WAAW,CAAC,OAAO,CAAC;IAC3B,YAAY,CAAC,WAAW,EAAE,IAAI,CAAC,CAAC;IAEhC,OAAO;QACN,IAAI;QACJ,WAAW;QACX,WAAW,EAAE,WAAyB;QACtC,OAAO,EAAE,CAAC,GAAG,EAAE,MAAM,EAAE,EAAE;YACxB,MAAM,MAAM,GAAG,KAAK,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC;YACpC,IAAI,CAAC,MAAM,CAAC,OAAO,EAAE,CAAC;gBACrB,OAAO,OAAO,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,CAAC,CAAC,aAAa,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;YACjE,CAAC;YAED,OAAO,OAAO,CAAC,MAAM,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;QACrC,CAAC;KACD,CAAC;AACH,CAAC","sourcesContent":["import { z } from 'zod';\nimport { InvalidAgentConfigError } from '#src/agent/domain/errors';\nimport type { JSONSchema } from '#src/llm/domain/tool';\nimport type { Tool } from '../domain/tool.ts';\n\n// Both providers take tools in strict mode, which supports only part of JSON Schema. Claude\n// rejects these bounds; a tool that needs one checks it in its own code instead.\nconst unsupportedKeywords = [\n\t'minimum',\n\t'maximum',\n\t'exclusiveMinimum',\n\t'exclusiveMaximum',\n\t'multipleOf',\n\t'minLength',\n\t'maxLength',\n];\n\n// Keywords whose value is one schema, or a list or map of them. Anything else, such as the\n// keys of `properties`, is a name or a value, never a schema to check.\nconst schemaKeywords = ['items', 'not', 'additionalProperties'];\nconst schemaListKeywords = ['anyOf', 'oneOf', 'allOf', 'prefixItems'];\nconst schemaMapKeywords = ['properties', '$defs', 'definitions'];\n\nfunction isSchema(value: unknown): value is Record<string, unknown> {\n\treturn typeof value === 'object' && value !== null && !Array.isArray(value);\n}\n\n/**\n * Walks the whole schema, nested objects included, and visits only schemas: a property named\n * `maxLength` is a name, not the keyword. A schema the providers reject would fail every\n * request with a 400, so it fails here, when the tool is defined.\n */\nfunction assertStrict(schema: unknown, toolName: string): void {\n\tif (!isSchema(schema)) {\n\t\treturn;\n\t}\n\n\tconst keyword = unsupportedKeywords.find(candidate => candidate in schema);\n\tif (keyword) {\n\t\tthrow new InvalidAgentConfigError(\n\t\t\t`Tool \"${toolName}\" uses \"${keyword}\", which strict mode does not support; check it in the tool's code`,\n\t\t);\n\t}\n\n\tif (schema.type === 'object') {\n\t\t// A record or a loose object takes keys strict mode cannot know in advance.\n\t\tif (schema.additionalProperties !== false) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`Tool \"${toolName}\" has an object open to any key; strict mode needs every key declared`,\n\t\t\t);\n\t\t}\n\n\t\tconst required = (schema.required ?? []) as string[];\n\t\t// OpenAI requires every property: an optional one is a union with `null`.\n\t\tconst optional = Object.keys(schema.properties ?? {}).filter(\n\t\t\tproperty => !required.includes(property),\n\t\t);\n\t\tif (optional.length > 0) {\n\t\t\tthrow new InvalidAgentConfigError(\n\t\t\t\t`Tool \"${toolName}\" has optional fields (${optional.join(', ')}); make them .nullable() instead`,\n\t\t\t);\n\t\t}\n\t}\n\n\tfor (const key of schemaKeywords) assertStrict(schema[key], toolName);\n\tfor (const key of schemaListKeywords) {\n\t\tconst list = schema[key];\n\t\tif (Array.isArray(list)) for (const item of list) assertStrict(item, toolName);\n\t}\n\tfor (const key of schemaMapKeywords) {\n\t\tconst map = schema[key];\n\t\tif (isSchema(map)) for (const item of Object.values(map)) assertStrict(item, toolName);\n\t}\n}\n\n/**\n * Builds a `Tool` from one Zod schema, which gives the model its JSON Schema, validates what the\n * model sends, and types the input `execute` receives. Invalid input is rejected with a message\n * naming what is wrong, which the agent hands back to the model as an error result.\n *\n * @throws {InvalidAgentConfigError} when the schema falls outside what strict mode supports.\n */\nexport function defineTool<Schema extends z.ZodObject>({\n\tname,\n\tdescription,\n\tinput,\n\texecute,\n}: {\n\tname: string;\n\tdescription: string;\n\tinput: Schema;\n\texecute: (input: z.infer<Schema>, signal: AbortSignal) => Promise<string>;\n}): Tool {\n\tconst inputSchema: Record<string, unknown> = { ...z.toJSONSchema(input) };\n\t// The dialect marker is noise to both APIs.\n\tdelete inputSchema.$schema;\n\tassertStrict(inputSchema, name);\n\n\treturn {\n\t\tname,\n\t\tdescription,\n\t\tinputSchema: inputSchema as JSONSchema,\n\t\texecute: (raw, signal) => {\n\t\t\tconst parsed = input.safeParse(raw);\n\t\t\tif (!parsed.success) {\n\t\t\t\treturn Promise.reject(new Error(z.prettifyError(parsed.error)));\n\t\t\t}\n\n\t\t\treturn execute(parsed.data, signal);\n\t\t},\n\t};\n}\n"]}

package/dist/tools/infrastructure/gitWorkspace.d.ts

@@ -0,0 +1,13 @@
+import type { TextMatch } from '../domain/workspace.ts';
+import { BoundedWorkspace } from './boundedWorkspace.ts';
+/**
+ * The fallback `Workspace`, on the git the repository already has. git applies `.gitignore`
+ * by itself, even to an explicit path, but only inside a repository: outside one, every
+ * operation fails.
+ */
+export declare class GitWorkspace extends BoundedWorkspace {
+    protected listCandidates(signal: AbortSignal): Promise<string[]>;
+    protected findMatches(pattern: string, ignoreCase: boolean, signal: AbortSignal, keep: (match: TextMatch) => void): Promise<void>;
+    private listFiles0;
+    private git;
+}

package/dist/tools/infrastructure/gitWorkspace.js

@@ -0,0 +1,61 @@
+import { BoundedWorkspace } from "./boundedWorkspace.js";
+import { firstLine, runProcess } from "./runProcess.js";
+/**
+ * The fallback `Workspace`, on the git the repository already has. git applies `.gitignore`
+ * by itself, even to an explicit path, but only inside a repository: outside one, every
+ * operation fails.
+ */
+export class GitWorkspace extends BoundedWorkspace {
+    async listCandidates(signal) {
+        // Tracked files and new ones not yet added; then the tracked ones `.gitignore` matches
+        // anyway, such as a `.env` committed before it was ignored. ripgrep hides those, and the
+        // fallback must not show what the preferred implementation hides.
+        const [listed, ignored] = await Promise.all([
+            this.listFiles0(['--cached', '--others', '--exclude-standard'], signal),
+            this.listFiles0(['--cached', '--ignored', '--exclude-standard'], signal),
+        ]);
+        const hidden = new Set(ignored);
+        return listed.filter(path => !hidden.has(path));
+    }
+    async findMatches(pattern, ignoreCase, signal, keep) {
+        const { stderr, exitCode } = await this.git([
+            'grep',
+            // Untracked files too, or the agent could not search a file it has just created.
+            '--untracked',
+            // Binary files left out, as ripgrep does.
+            '-I',
+            '--line-number',
+            // path NUL line NUL text, so a colon in a path cannot split it.
+            '-z',
+            // Perl syntax, so `\d` and `\b` mean what they mean to ripgrep.
+            '-P',
+            ...(ignoreCase ? ['--ignore-case'] : []),
+            // `-e`, so a pattern starting with a dash is never read as an option.
+            '-e',
+            pattern,
+        ], signal, line => {
+            const [path, number, ...text] = line.split('\0');
+            if (path && number) {
+                keep({ path, line: Number(number), text: text.join('\0').replace(/\r$/, '') });
+            }
+        });
+        // 1 means no match; anything above is a failure, such as a pattern that does not compile.
+        if (exitCode > 1) {
+            throw new Error(`git could not search for the pattern: ${firstLine(stderr)}`);
+        }
+    }
+    async listFiles0(options, signal) {
+        const { stdout, stderr, exitCode } = await this.git(['ls-files', '-z', ...options], signal);
+        if (exitCode !== 0) {
+            throw new Error(`git could not list the files: ${firstLine(stderr)}`);
+        }
+        return stdout.split('\0').filter(path => path !== '');
+    }
+    git(args, signal, onLine) {
+        // A user's settings must not change the output this parses: colour codes would end up in
+        // the text, `grep.column` adds a field, and `grep.fullName` makes paths relative to the
+        // repository instead of the root.
+        return runProcess('git', ['-c', 'color.ui=never', '-c', 'grep.column=false', '-c', 'grep.fullName=false', ...args], { cwd: this.root, signal, onLine });
+    }
+}
+//# sourceMappingURL=gitWorkspace.js.map

package/dist/tools/infrastructure/gitWorkspace.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"gitWorkspace.js","sourceRoot":"","sources":["../../../src/tools/infrastructure/gitWorkspace.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,gBAAgB,EAAE,MAAM,uBAAuB,CAAC;AACzD,OAAO,EAAE,SAAS,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAExD;;;;GAIG;AACH,MAAM,OAAO,YAAa,SAAQ,gBAAgB;IACvC,KAAK,CAAC,cAAc,CAAC,MAAmB;QACjD,uFAAuF;QACvF,yFAAyF;QACzF,kEAAkE;QAClE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;YAC3C,IAAI,CAAC,UAAU,CAAC,CAAC,UAAU,EAAE,UAAU,EAAE,oBAAoB,CAAC,EAAE,MAAM,CAAC;YACvE,IAAI,CAAC,UAAU,CAAC,CAAC,UAAU,EAAE,WAAW,EAAE,oBAAoB,CAAC,EAAE,MAAM,CAAC;SACxE,CAAC,CAAC;QACH,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,CAAC;QAEhC,OAAO,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC;IACjD,CAAC;IAES,KAAK,CAAC,WAAW,CAC1B,OAAe,EACf,UAAmB,EACnB,MAAmB,EACnB,IAAgC;QAEhC,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,GAAG,MAAM,IAAI,CAAC,GAAG,CAC1C;YACC,MAAM;YACN,iFAAiF;YACjF,aAAa;YACb,0CAA0C;YAC1C,IAAI;YACJ,eAAe;YACf,gEAAgE;YAChE,IAAI;YACJ,gEAAgE;YAChE,IAAI;YACJ,GAAG,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YACxC,sEAAsE;YACtE,IAAI;YACJ,OAAO;SACP,EACD,MAAM,EACN,IAAI,CAAC,EAAE;YACN,MAAM,CAAC,IAAI,EAAE,MAAM,EAAE,GAAG,IAAI,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;YACjD,IAAI,IAAI,IAAI,MAAM,EAAE,CAAC;gBACpB,IAAI,CAAC,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,CAAC,MAAM,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,EAAE,CAAC,CAAC;YAChF,CAAC;QACF,CAAC,CACD,CAAC;QACF,0FAA0F;QAC1F,IAAI,QAAQ,GAAG,CAAC,EAAE,CAAC;YAClB,MAAM,IAAI,KAAK,CAAC,yCAAyC,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QAC/E,CAAC;IACF,CAAC;IAEO,KAAK,CAAC,UAAU,CAAC,OAAiB,EAAE,MAAmB;QAC9D,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,GAAG,MAAM,IAAI,CAAC,GAAG,CAAC,CAAC,UAAU,EAAE,IAAI,EAAE,GAAG,OAAO,CAAC,EAAE,MAAM,CAAC,CAAC;QAC5F,IAAI,QAAQ,KAAK,CAAC,EAAE,CAAC;YACpB,MAAM,IAAI,KAAK,CAAC,iCAAiC,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QACvE,CAAC;QAED,OAAO,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,KAAK,EAAE,CAAC,CAAC;IACvD,CAAC;IAEO,GAAG,CAAC,IAAc,EAAE,MAAmB,EAAE,MAA+B;QAC/E,yFAAyF;QACzF,wFAAwF;QACxF,kCAAkC;QAClC,OAAO,UAAU,CAChB,KAAK,EACL,CAAC,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,qBAAqB,EAAE,GAAG,IAAI,CAAC,EACzF,EAAE,GAAG,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,CAClC,CAAC;IACH,CAAC;CACD","sourcesContent":["import type { TextMatch } from '../domain/workspace.ts';\nimport { BoundedWorkspace } from './boundedWorkspace.ts';\nimport { firstLine, runProcess } from './runProcess.ts';\n\n/**\n * The fallback `Workspace`, on the git the repository already has. git applies `.gitignore`\n * by itself, even to an explicit path, but only inside a repository: outside one, every\n * operation fails.\n */\nexport class GitWorkspace extends BoundedWorkspace {\n\tprotected async listCandidates(signal: AbortSignal): Promise<string[]> {\n\t\t// Tracked files and new ones not yet added; then the tracked ones `.gitignore` matches\n\t\t// anyway, such as a `.env` committed before it was ignored. ripgrep hides those, and the\n\t\t// fallback must not show what the preferred implementation hides.\n\t\tconst [listed, ignored] = await Promise.all([\n\t\t\tthis.listFiles0(['--cached', '--others', '--exclude-standard'], signal),\n\t\t\tthis.listFiles0(['--cached', '--ignored', '--exclude-standard'], signal),\n\t\t]);\n\t\tconst hidden = new Set(ignored);\n\n\t\treturn listed.filter(path => !hidden.has(path));\n\t}\n\n\tprotected async findMatches(\n\t\tpattern: string,\n\t\tignoreCase: boolean,\n\t\tsignal: AbortSignal,\n\t\tkeep: (match: TextMatch) => void,\n\t): Promise<void> {\n\t\tconst { stderr, exitCode } = await this.git(\n\t\t\t[\n\t\t\t\t'grep',\n\t\t\t\t// Untracked files too, or the agent could not search a file it has just created.\n\t\t\t\t'--untracked',\n\t\t\t\t// Binary files left out, as ripgrep does.\n\t\t\t\t'-I',\n\t\t\t\t'--line-number',\n\t\t\t\t// path NUL line NUL text, so a colon in a path cannot split it.\n\t\t\t\t'-z',\n\t\t\t\t// Perl syntax, so `\\d` and `\\b` mean what they mean to ripgrep.\n\t\t\t\t'-P',\n\t\t\t\t...(ignoreCase ? ['--ignore-case'] : []),\n\t\t\t\t// `-e`, so a pattern starting with a dash is never read as an option.\n\t\t\t\t'-e',\n\t\t\t\tpattern,\n\t\t\t],\n\t\t\tsignal,\n\t\t\tline => {\n\t\t\t\tconst [path, number, ...text] = line.split('\\0');\n\t\t\t\tif (path && number) {\n\t\t\t\t\tkeep({ path, line: Number(number), text: text.join('\\0').replace(/\\r$/, '') });\n\t\t\t\t}\n\t\t\t},\n\t\t);\n\t\t// 1 means no match; anything above is a failure, such as a pattern that does not compile.\n\t\tif (exitCode > 1) {\n\t\t\tthrow new Error(`git could not search for the pattern: ${firstLine(stderr)}`);\n\t\t}\n\t}\n\n\tprivate async listFiles0(options: string[], signal: AbortSignal): Promise<string[]> {\n\t\tconst { stdout, stderr, exitCode } = await this.git(['ls-files', '-z', ...options], signal);\n\t\tif (exitCode !== 0) {\n\t\t\tthrow new Error(`git could not list the files: ${firstLine(stderr)}`);\n\t\t}\n\n\t\treturn stdout.split('\\0').filter(path => path !== '');\n\t}\n\n\tprivate git(args: string[], signal: AbortSignal, onLine?: (line: string) => void) {\n\t\t// A user's settings must not change the output this parses: colour codes would end up in\n\t\t// the text, `grep.column` adds a field, and `grep.fullName` makes paths relative to the\n\t\t// repository instead of the root.\n\t\treturn runProcess(\n\t\t\t'git',\n\t\t\t['-c', 'color.ui=never', '-c', 'grep.column=false', '-c', 'grep.fullName=false', ...args],\n\t\t\t{ cwd: this.root, signal, onLine },\n\t\t);\n\t}\n}\n"]}

package/dist/tools/infrastructure/ripgrepWorkspace.d.ts

@@ -0,0 +1,20 @@
+import type { TextMatch } from '../domain/workspace.ts';
+import { BoundedWorkspace } from './boundedWorkspace.ts';
+/**
+ * The preferred `Workspace`, on the ripgrep binary that ships with the package. It always runs
+ * from the root with no path: ripgrep ignores nothing that it is given explicitly.
+ *
+ * It takes the binary's path instead of importing `@vscode/ripgrep`, which throws on import when
+ * the platform has no binary; `createWorkspace` loads it where that failure can be handled.
+ */
+export declare class RipgrepWorkspace extends BoundedWorkspace {
+    private readonly ripgrepPath;
+    constructor({ ripgrepPath, ...options }: {
+        root: string;
+        ripgrepPath: string;
+        timeoutMs?: number;
+    });
+    protected listCandidates(signal: AbortSignal): Promise<string[]>;
+    protected findMatches(pattern: string, ignoreCase: boolean, signal: AbortSignal, keep: (match: TextMatch) => void): Promise<void>;
+    private ripgrep;
+}

package/dist/tools/infrastructure/ripgrepWorkspace.js

@@ -0,0 +1,85 @@
+import { BoundedWorkspace } from "./boundedWorkspace.js";
+import { firstLine, runProcess } from "./runProcess.js";
+// Every call sees what git would. Each flag is ours; nothing a model writes ever reaches the
+// command line except the pattern, and that only after `-e`.
+const sameViewAsGit = [
+    // A user's ripgrep configuration could switch the ignore rules off.
+    '--no-config',
+    // ripgrep's own `.ignore` and `.rgignore` outrank `.gitignore`, so a `!.env` line in either
+    // would show a secret git hides. Only git's ignore sources count.
+    '--no-ignore-dot',
+    // Hidden files such as `.github/` are part of the repository; ripgrep skips them by default.
+    '--hidden',
+    // `.gitignore` applies even outside a git repository.
+    '--no-require-git',
+    // `--hidden` would otherwise search git's own folder.
+    '--glob',
+    '!.git',
+    // Windows would print `\`; every path in a `Workspace` is written with `/`.
+    '--path-separator',
+    '/',
+];
+function decode({ text, bytes }) {
+    return text ?? Buffer.from(bytes ?? '', 'base64').toString('utf8');
+}
+/**
+ * The preferred `Workspace`, on the ripgrep binary that ships with the package. It always runs
+ * from the root with no path: ripgrep ignores nothing that it is given explicitly.
+ *
+ * It takes the binary's path instead of importing `@vscode/ripgrep`, which throws on import when
+ * the platform has no binary; `createWorkspace` loads it where that failure can be handled.
+ */
+export class RipgrepWorkspace extends BoundedWorkspace {
+    ripgrepPath;
+    constructor({ ripgrepPath, ...options }) {
+        super(options);
+        this.ripgrepPath = ripgrepPath;
+    }
+    async listCandidates(signal) {
+        const { stdout, stderr, exitCode } = await this.ripgrep(['--files', '--null'], signal);
+        // 1 means nothing to list, which an empty repository is. 2 with a listing means some
+        // folder could not be read, which git skips with a warning too; without one, it failed.
+        if (exitCode > 1 && stdout === '') {
+            throw new Error(`ripgrep could not list the files: ${firstLine(stderr)}`);
+        }
+        return stdout
+            .split('\0')
+            .filter(path => path !== '')
+            .map(path => path.replace(/^\.\//, ''));
+    }
+    async findMatches(pattern, ignoreCase, signal, keep) {
+        // ripgrep ends every search it ran with a summary. A failure before searching, such as a
+        // pattern that does not compile, prints none. A count, not a flag: TypeScript cannot see
+        // the callback change it.
+        let summaries = 0;
+        const { stderr, exitCode } = await this.ripgrep(['--json', ...(ignoreCase ? ['--ignore-case'] : []), '-e', pattern], signal, line => {
+            if (line === '') {
+                return;
+            }
+            const event = JSON.parse(line);
+            if (event.type === 'summary') {
+                summaries++;
+            }
+            if (event.type === 'match') {
+                keep({
+                    path: decode(event.data.path).replace(/^\.\//, ''),
+                    line: event.data.line_number,
+                    text: decode(event.data.lines).replace(/\r?\n$/, ''),
+                });
+            }
+        });
+        // 1 means no match. 2 after a search means some file could not be read, and the rest
+        // stands, as it does on git; 2 with no search is the failure.
+        if (exitCode > 1 && summaries === 0) {
+            throw new Error(`ripgrep could not search for the pattern: ${firstLine(stderr)}`);
+        }
+    }
+    ripgrep(args, signal, onLine) {
+        return runProcess(this.ripgrepPath, [...sameViewAsGit, ...args], {
+            cwd: this.root,
+            signal,
+            onLine,
+        });
+    }
+}
+//# sourceMappingURL=ripgrepWorkspace.js.map

package/dist/tools/infrastructure/ripgrepWorkspace.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ripgrepWorkspace.js","sourceRoot":"","sources":["../../../src/tools/infrastructure/ripgrepWorkspace.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,gBAAgB,EAAE,MAAM,uBAAuB,CAAC;AACzD,OAAO,EAAE,SAAS,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAExD,6FAA6F;AAC7F,6DAA6D;AAC7D,MAAM,aAAa,GAAG;IACrB,oEAAoE;IACpE,aAAa;IACb,4FAA4F;IAC5F,kEAAkE;IAClE,iBAAiB;IACjB,6FAA6F;IAC7F,UAAU;IACV,sDAAsD;IACtD,kBAAkB;IAClB,sDAAsD;IACtD,QAAQ;IACR,OAAO;IACP,4EAA4E;IAC5E,kBAAkB;IAClB,GAAG;CACH,CAAC;AAaF,SAAS,MAAM,CAAC,EAAE,IAAI,EAAE,KAAK,EAAe;IAC3C,OAAO,IAAI,IAAI,MAAM,CAAC,IAAI,CAAC,KAAK,IAAI,EAAE,EAAE,QAAQ,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC;AACpE,CAAC;AAED;;;;;;GAMG;AACH,MAAM,OAAO,gBAAiB,SAAQ,gBAAgB;IACpC,WAAW,CAAS;IAErC,YAAY,EACX,WAAW,EACX,GAAG,OAAO,EAKV;QACA,KAAK,CAAC,OAAO,CAAC,CAAC;QACf,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;IAChC,CAAC;IAES,KAAK,CAAC,cAAc,CAAC,MAAmB;QACjD,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,CAAC,SAAS,EAAE,QAAQ,CAAC,EAAE,MAAM,CAAC,CAAC;QACvF,qFAAqF;QACrF,wFAAwF;QACxF,IAAI,QAAQ,GAAG,CAAC,IAAI,MAAM,KAAK,EAAE,EAAE,CAAC;YACnC,MAAM,IAAI,KAAK,CAAC,qCAAqC,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QAC3E,CAAC;QAED,OAAO,MAAM;aACX,KAAK,CAAC,IAAI,CAAC;aACX,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,KAAK,EAAE,CAAC;aAC3B,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,CAAC;IAC1C,CAAC;IAES,KAAK,CAAC,WAAW,CAC1B,OAAe,EACf,UAAmB,EACnB,MAAmB,EACnB,IAAgC;QAEhC,yFAAyF;QACzF,yFAAyF;QACzF,0BAA0B;QAC1B,IAAI,SAAS,GAAG,CAAC,CAAC;QAClB,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,GAAG,MAAM,IAAI,CAAC,OAAO,CAC9C,CAAC,QAAQ,EAAE,GAAG,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,EACnE,MAAM,EACN,IAAI,CAAC,EAAE;YACN,IAAI,IAAI,KAAK,EAAE,EAAE,CAAC;gBACjB,OAAO;YACR,CAAC;YACD,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAiB,CAAC;YAC/C,IAAI,KAAK,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;gBAC9B,SAAS,EAAE,CAAC;YACb,CAAC;YACD,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;gBAC5B,IAAI,CAAC;oBACJ,IAAI,EAAE,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC;oBAClD,IAAI,EAAE,KAAK,CAAC,IAAI,CAAC,WAAW;oBAC5B,IAAI,EAAE,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,OAAO,CAAC,QAAQ,EAAE,EAAE,CAAC;iBACpD,CAAC,CAAC;YACJ,CAAC;QACF,CAAC,CACD,CAAC;QAEF,qFAAqF;QACrF,8DAA8D;QAC9D,IAAI,QAAQ,GAAG,CAAC,IAAI,SAAS,KAAK,CAAC,EAAE,CAAC;YACrC,MAAM,IAAI,KAAK,CAAC,6CAA6C,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QACnF,CAAC;IACF,CAAC;IAEO,OAAO,CAAC,IAAc,EAAE,MAAmB,EAAE,MAA+B;QACnF,OAAO,UAAU,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,GAAG,aAAa,EAAE,GAAG,IAAI,CAAC,EAAE;YAChE,GAAG,EAAE,IAAI,CAAC,IAAI;YACd,MAAM;YACN,MAAM;SACN,CAAC,CAAC;IACJ,CAAC;CACD","sourcesContent":["import type { TextMatch } from '../domain/workspace.ts';\nimport { BoundedWorkspace } from './boundedWorkspace.ts';\nimport { firstLine, runProcess } from './runProcess.ts';\n\n// Every call sees what git would. Each flag is ours; nothing a model writes ever reaches the\n// command line except the pattern, and that only after `-e`.\nconst sameViewAsGit = [\n\t// A user's ripgrep configuration could switch the ignore rules off.\n\t'--no-config',\n\t// ripgrep's own `.ignore` and `.rgignore` outrank `.gitignore`, so a `!.env` line in either\n\t// would show a secret git hides. Only git's ignore sources count.\n\t'--no-ignore-dot',\n\t// Hidden files such as `.github/` are part of the repository; ripgrep skips them by default.\n\t'--hidden',\n\t// `.gitignore` applies even outside a git repository.\n\t'--no-require-git',\n\t// `--hidden` would otherwise search git's own folder.\n\t'--glob',\n\t'!.git',\n\t// Windows would print `\\`; every path in a `Workspace` is written with `/`.\n\t'--path-separator',\n\t'/',\n];\n\ninterface RipgrepText {\n\ttext?: string;\n\t// ripgrep sends what is not valid UTF-8 as base64 instead.\n\tbytes?: string;\n}\n\ninterface RipgrepEvent {\n\ttype: string;\n\tdata: { path: RipgrepText; line_number: number; lines: RipgrepText };\n}\n\nfunction decode({ text, bytes }: RipgrepText): string {\n\treturn text ?? Buffer.from(bytes ?? '', 'base64').toString('utf8');\n}\n\n/**\n * The preferred `Workspace`, on the ripgrep binary that ships with the package. It always runs\n * from the root with no path: ripgrep ignores nothing that it is given explicitly.\n *\n * It takes the binary's path instead of importing `@vscode/ripgrep`, which throws on import when\n * the platform has no binary; `createWorkspace` loads it where that failure can be handled.\n */\nexport class RipgrepWorkspace extends BoundedWorkspace {\n\tprivate readonly ripgrepPath: string;\n\n\tconstructor({\n\t\tripgrepPath,\n\t\t...options\n\t}: {\n\t\troot: string;\n\t\tripgrepPath: string;\n\t\ttimeoutMs?: number;\n\t}) {\n\t\tsuper(options);\n\t\tthis.ripgrepPath = ripgrepPath;\n\t}\n\n\tprotected async listCandidates(signal: AbortSignal): Promise<string[]> {\n\t\tconst { stdout, stderr, exitCode } = await this.ripgrep(['--files', '--null'], signal);\n\t\t// 1 means nothing to list, which an empty repository is. 2 with a listing means some\n\t\t// folder could not be read, which git skips with a warning too; without one, it failed.\n\t\tif (exitCode > 1 && stdout === '') {\n\t\t\tthrow new Error(`ripgrep could not list the files: ${firstLine(stderr)}`);\n\t\t}\n\n\t\treturn stdout\n\t\t\t.split('\\0')\n\t\t\t.filter(path => path !== '')\n\t\t\t.map(path => path.replace(/^\\.\\//, ''));\n\t}\n\n\tprotected async findMatches(\n\t\tpattern: string,\n\t\tignoreCase: boolean,\n\t\tsignal: AbortSignal,\n\t\tkeep: (match: TextMatch) => void,\n\t): Promise<void> {\n\t\t// ripgrep ends every search it ran with a summary. A failure before searching, such as a\n\t\t// pattern that does not compile, prints none. A count, not a flag: TypeScript cannot see\n\t\t// the callback change it.\n\t\tlet summaries = 0;\n\t\tconst { stderr, exitCode } = await this.ripgrep(\n\t\t\t['--json', ...(ignoreCase ? ['--ignore-case'] : []), '-e', pattern],\n\t\t\tsignal,\n\t\t\tline => {\n\t\t\t\tif (line === '') {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tconst event = JSON.parse(line) as RipgrepEvent;\n\t\t\t\tif (event.type === 'summary') {\n\t\t\t\t\tsummaries++;\n\t\t\t\t}\n\t\t\t\tif (event.type === 'match') {\n\t\t\t\t\tkeep({\n\t\t\t\t\t\tpath: decode(event.data.path).replace(/^\\.\\//, ''),\n\t\t\t\t\t\tline: event.data.line_number,\n\t\t\t\t\t\ttext: decode(event.data.lines).replace(/\\r?\\n$/, ''),\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t},\n\t\t);\n\n\t\t// 1 means no match. 2 after a search means some file could not be read, and the rest\n\t\t// stands, as it does on git; 2 with no search is the failure.\n\t\tif (exitCode > 1 && summaries === 0) {\n\t\t\tthrow new Error(`ripgrep could not search for the pattern: ${firstLine(stderr)}`);\n\t\t}\n\t}\n\n\tprivate ripgrep(args: string[], signal: AbortSignal, onLine?: (line: string) => void) {\n\t\treturn runProcess(this.ripgrepPath, [...sameViewAsGit, ...args], {\n\t\t\tcwd: this.root,\n\t\t\tsignal,\n\t\t\tonLine,\n\t\t});\n\t}\n}\n"]}

package/dist/tools/infrastructure/runProcess.d.ts

@@ -0,0 +1,23 @@
+export interface ProcessResult {
+    /** Empty when the output went to `onLine` instead. */
+    stdout: string;
+    stderr: string;
+    exitCode: number;
+}
+/**
+ * Runs a program with its arguments as an array, never through a shell, so nothing a model
+ * writes can become a command. Standard input is closed: ripgrep given no path searches its
+ * standard input whenever that is a pipe. A cancelled signal kills the process and rejects
+ * with the signal's `AbortError`.
+ *
+ * With `onLine`, each line of the output is handed over as it arrives and nothing is kept, so
+ * the caller decides what to hold; only a single line is then bound by the size limit. An
+ * error thrown by `onLine` stops the program and rejects with that error.
+ */
+export declare function runProcess(command: string, args: string[], { cwd, signal, onLine }: {
+    cwd: string;
+    signal: AbortSignal;
+    onLine?: (line: string) => void;
+}): Promise<ProcessResult>;
+/** The first line of a program's error output, which is the part a model can act on. */
+export declare function firstLine(text: string): string;