@gotgenes/pi-permission-system 25.2.1 → 25.2.2

package/CHANGELOG.md

@@ -5,6 +5,21 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [25.2.2](https://github.com/gotgenes/pi-packages/compare/pi-permission-system-v25.2.1...pi-permission-system-v25.2.2) (2026-08-15)
+
+
+### Bug Fixes
+
+* **pi-permission-system:** gate commands hosted in bash redirect targets ([1015bb8](https://github.com/gotgenes/pi-packages/commit/1015bb879234d1d96595b9f995ca88037588e16f)), closes [#741](https://github.com/gotgenes/pi-packages/issues/741)
+* **pi-permission-system:** gate commands hosted in interpolating heredoc bodies ([48978d2](https://github.com/gotgenes/pi-packages/commit/48978d2016c67aa90f291db55d98908ba930f0d1)), closes [#741](https://github.com/gotgenes/pi-packages/issues/741)
+* **pi-permission-system:** project path operands of heredoc-hosted nested commands ([8e2fbee](https://github.com/gotgenes/pi-packages/commit/8e2fbee404677d02c7cc565cf23d1e778994d5f0)), closes [#741](https://github.com/gotgenes/pi-packages/issues/741)
+* **pi-permission-system:** project path operands of redirect-hosted nested commands ([12164f3](https://github.com/gotgenes/pi-packages/commit/12164f3b61b6a4b50281d84b6d741a3736516991)), closes [#741](https://github.com/gotgenes/pi-packages/issues/741)
+
+
+### Documentation
+
+* **pi-permission-system:** document hosted nested-command evaluation ([f24b338](https://github.com/gotgenes/pi-packages/commit/f24b338a42855f96038fdf2efdfe440cef8501c8)), closes [#741](https://github.com/gotgenes/pi-packages/issues/741)
+
 ## [25.2.1](https://github.com/gotgenes/pi-packages/compare/pi-permission-system-v25.2.0...pi-permission-system-v25.2.1) (2026-08-15)
 
 

package/docs/configuration.md

@@ -328,6 +328,11 @@ Quotes are respected (an operator inside `'…'` or `"…"` does not split the c
 Commands nested inside command substitution (`$(…)`, backticks), process substitution (`<(…)`/`>(…)`), and subshells (`( … )`) are evaluated against the bash patterns too, in addition to their enclosing command — since those inner commands really execute.
 So `echo $(rm -rf foo)` evaluates both `echo $(rm -rf foo)` and the inner `rm -rf foo`; if `rm *` is denied, the whole invocation is denied.
 The deny reason and the approval prompt note the nested origin (e.g. `inside command substitution`).
+
+This holds wherever the substitution appears, not only in argument position.
+A substitution in a **redirect target** (`echo hi > $(rm *.txt)`, `cat < <(rm c)`, ``echo hi 2> `rm d` ``) and one in an **interpolating heredoc body** (`cat <<EOF` with `$(rm e)` in the body) are evaluated the same way.
+A quoted heredoc delimiter (`<<'EOF'` or `<<"EOF"`) does not interpolate, so its body is literal text and nothing in it is evaluated as a command.
+The enclosing command is still matched without its redirect, so a rule like `npm install` keeps matching `npm install > out.txt`.
 Control-flow bodies (`if`/`while`/`for`/`case`) and `{ … }` brace groups are not descended into; their contents are matched as part of the enclosing statement's text.
 
 A leading environment-variable assignment prefix is stripped before matching, so the rule gates the underlying command rather than the prefix.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gotgenes/pi-permission-system",
-  "version": "25.2.1",
+  "version": "25.2.2",
   "description": "Permission enforcement extension for the Pi coding agent.",
   "type": "module",
   "exports": {

package/src/access-intent/bash/command-enumeration.ts

@@ -1,3 +1,7 @@
+import {
+  EXECUTION_HOST_TYPES,
+  forEachNestedExecution,
+} from "#src/access-intent/bash/nested-execution";
 import type { TSNode } from "#src/access-intent/bash/parser";
 import type { BashCommandContext } from "#src/types";
 
@@ -49,32 +53,19 @@ const COMMAND_ENUM_DESCEND = new Set([
 ]);
 
 /**
- * Named node types skipped during command enumeration: redirect targets,
- * comments, and heredoc bodies — none is a command to evaluate.
+ * Named node types abandoned during command enumeration: they are neither
+ * commands nor able to host one, so nothing in their subtree ever runs.
+ *
+ * A redirect and a heredoc body are deliberately NOT listed here. Neither is a
+ * command, but each can host a substitution that really executes, so both are
+ * {@link EXECUTION_HOST_TYPES} members instead — conflating the two questions
+ * ("is this a command?" and "can this host one?") is the bypass #741 fixed.
+ *
  * Anonymous tokens (chain operators `&&`/`;`/`|`, substitution and subshell
  * delimiters `$(`/`)`/`` ` ``/`(`) are filtered by the `isNamed` guard, not
  * listed here.
  */
-const COMMAND_ENUM_SKIP = new Set([
-  "file_redirect",
-  "heredoc_redirect",
-  "herestring_redirect",
-  "comment",
-  "heredoc_body",
-  "heredoc_end",
-]);
-
-/**
- * Nested execution contexts whose interior commands really execute and must be
- * evaluated too: command substitution (`$(…)`, backticks) and process
- * substitution (`<(…)`/`>(…)`).
- * Subshells (`( … )`) are handled separately because they are also emitted
- * whole.
- */
-const NESTED_EXECUTION_CONTEXTS = new Map<string, BashCommandContext>([
-  ["command_substitution", "command_substitution"],
-  ["process_substitution", "process_substitution"],
-]);
+const COMMAND_ENUM_SKIP = new Set(["comment", "heredoc_end"]);
 
 /**
  * Enumerate the command units of a bash program, in source order.
@@ -119,7 +110,14 @@ function collectCommandsInto(
     );
     // A command's text already contains any substitution; descend its subtree
     // to ALSO emit the inner commands of command/process substitutions.
-    collectSubstitutionCommands(node, out);
+    collectHostedCommands(node, out);
+    return;
+  }
+
+  if (EXECUTION_HOST_TYPES.has(node.type)) {
+    // Not a command itself, but its subtree can host one that really runs
+    // (`> $(rm x)`, `< <(rm c)`). Emit only what it hosts (#741).
+    collectHostedCommands(node, out);
     return;
   }
 
@@ -297,20 +295,15 @@ function descendCommandChildren(
 }
 
 /**
- * Search a command's subtree for command/process substitutions and enumerate
- * the commands inside them, tagged with the substitution's execution context.
- * A substitution can nest under `command_name` (when the whole command is
- * `$(…)`) or under an argument, so the entire subtree is searched.
+ * Enumerate the commands of every nested execution context in a subtree, each
+ * tagged with the context it was found in.
+ *
+ * The traversal itself lives in `nested-execution.ts` so the bash path surface
+ * shares one definition of what counts as a nested execution (#741); this
+ * function supplies the command-surface interpretation of each one found.
  */
-function collectSubstitutionCommands(node: TSNode, out: BashCommand[]): void {
-  for (let i = 0; i < node.childCount; i++) {
-    const child = node.child(i);
-    if (!child) continue;
-    const nestedContext = NESTED_EXECUTION_CONTEXTS.get(child.type);
-    if (nestedContext) {
-      descendCommandChildren(child, nestedContext, out);
-    } else {
-      collectSubstitutionCommands(child, out);
-    }
-  }
+function collectHostedCommands(node: TSNode, out: BashCommand[]): void {
+  forEachNestedExecution(node, (contextNode, context) => {
+    descendCommandChildren(contextNode, context, out);
+  });
 }

package/src/access-intent/bash/nested-execution.ts

@@ -0,0 +1,76 @@
+import type { TSNode } from "#src/access-intent/bash/parser";
+import type { BashCommandContext } from "#src/types";
+
+/**
+ * AST node types whose interior commands really execute when the shell runs the
+ * program: command substitution (`$(…)`, backticks) and process substitution
+ * (`<(…)`/`>(…)`).
+ *
+ * Subshells (`( … )`) are deliberately absent — a subshell is also a command
+ * unit in its own right, so the command enumerator emits it whole and descends
+ * it separately rather than treating it as a pure nesting wrapper.
+ *
+ * This map is the single vocabulary shared by the bash command surface and the
+ * bash path surface, so the two cannot disagree about what counts as a nested
+ * execution (#741).
+ */
+export const NESTED_EXECUTION_CONTEXTS: ReadonlyMap<
+  string,
+  BashCommandContext
+> = new Map([
+  ["command_substitution", "command_substitution"],
+  ["process_substitution", "process_substitution"],
+] satisfies [string, BashCommandContext][]);
+
+/**
+ * AST node types that are neither commands nor argument values themselves, but
+ * whose subtree can host a nested execution context that really runs.
+ *
+ * A redirect destination is the motivating case: tree-sitter-bash parses
+ * `echo hi > $(rm x)` with the `file_redirect` as a *sibling* of the `command`,
+ * so a consumer that abandons the redirect never sees the substitution inside
+ * it — the bypass #741 fixed.
+ *
+ * An interpolating heredoc body is the second case: `cat <<EOF` with `$(rm e)`
+ * in the body really runs `rm e`. Quoting needs no special handling here —
+ * tree-sitter-bash emits a `command_substitution` node under `heredoc_body`
+ * only for a bare `<<EOF`, never for `<<'EOF'` or `<<"EOF"`, so the parser
+ * already encodes the interpolation rule.
+ *
+ * Membership means "do not read this subtree's own text, but do descend it for
+ * executions"; each consumer keeps its own handling of the destination tokens.
+ */
+export const EXECUTION_HOST_TYPES: ReadonlySet<string> = new Set([
+  "file_redirect",
+  "heredoc_redirect",
+  "herestring_redirect",
+  "heredoc_body",
+]);
+
+/**
+ * Visit every nested execution context in `node`'s subtree, in source order.
+ *
+ * The walk does not descend *past* a context it finds: `visit` receives the
+ * context node itself and decides how to treat its interior (the command
+ * enumerator enumerates commands there; the path collector collects operand
+ * tokens), which keeps recursion policy with the consumer that understands it.
+ *
+ * A substitution can nest under `command_name` (when the whole command is
+ * `$(…)`), under an argument, inside a redirect destination, or inside an
+ * interpolating heredoc body, so the entire subtree is searched.
+ */
+export function forEachNestedExecution(
+  node: TSNode,
+  visit: (contextNode: TSNode, context: BashCommandContext) => void,
+): void {
+  for (let i = 0; i < node.childCount; i++) {
+    const child = node.child(i);
+    if (!child) continue;
+    const context = NESTED_EXECUTION_CONTEXTS.get(child.type);
+    if (context) {
+      visit(child, context);
+    } else {
+      forEachNestedExecution(child, visit);
+    }
+  }
+}

package/src/access-intent/bash/node-text.ts

@@ -2,8 +2,13 @@ import type { TSNode } from "#src/access-intent/bash/parser";
 import { resolvePlainVariableExpansion } from "#src/access-intent/bash/shell-variable-expansion";
 
 /**
- * Node types whose subtrees must never be descended into for
- * path extraction — their text content is not a command argument.
+ * Node types whose text content is never a command argument, so no path
+ * candidate is ever read from it.
+ *
+ * This governs the subtree's *text*, not whether it is visited at all: an
+ * interpolating `heredoc_body` is also an execution host, so it is still
+ * descended for the commands it runs while its prose stays out of the path
+ * surface (#741). See `EXECUTION_HOST_TYPES` in `nested-execution.ts`.
  */
 export const SKIP_SUBTREE_TYPES = new Set([
   "heredoc_body",

package/src/access-intent/bash/token-collection.ts

@@ -1,4 +1,9 @@
 import { basename } from "node:path";
+import {
+  EXECUTION_HOST_TYPES,
+  forEachNestedExecution,
+  NESTED_EXECUTION_CONTEXTS,
+} from "#src/access-intent/bash/nested-execution";
 import {
   ARG_NODE_TYPES,
   resolveNodeText,
@@ -12,7 +17,12 @@ import type { TSNode } from "#src/access-intent/bash/parser";
  * Recursively visit the AST and collect resolved text of nodes that
  * represent command arguments or redirect destinations.
  *
- * Skips `heredoc_body`, `heredoc_end`, and `comment` subtrees entirely.
+ * Reads no text from `heredoc_body`, `heredoc_end`, or `comment` subtrees, but
+ * still descends an execution host for the commands it hosts — an interpolating
+ * heredoc body runs its substitution even though its prose is never an operand
+ * (#741). That is why the {@link EXECUTION_HOST_TYPES} branch sits above the
+ * {@link SKIP_SUBTREE_TYPES} check: `heredoc_body` is in both sets, and the
+ * host reading is the one that must win.
  *
  * For commands in `PATTERN_FIRST_COMMANDS`, uses position-based
  * argument skipping to avoid collecting inline patterns/scripts
@@ -20,9 +30,12 @@ import type { TSNode } from "#src/access-intent/bash/parser";
  * arguments generically.
  */
 export function collectPathCandidateTokens(node: TSNode): string[] {
-  if (SKIP_SUBTREE_TYPES.has(node.type)) return [];
   if (node.type === "command") return collectCommandTokens(node);
   if (node.type === "file_redirect") return collectRedirectTokens(node);
+  if (EXECUTION_HOST_TYPES.has(node.type)) {
+    return collectHostedExecutionTokens(node);
+  }
+  if (SKIP_SUBTREE_TYPES.has(node.type)) return [];
 
   const tokens: string[] = [];
   for (let i = 0; i < node.childCount; i++) {
@@ -50,6 +63,15 @@ export function collectCommandTokens(node: TSNode): string[] {
 
 /**
  * Collect redirect-destination tokens from a `file_redirect` node.
+ *
+ * The destination itself is an argument value (`> out.txt`), but it can also
+ * host a command that really runs (`> $(cat /etc/shadow)`, `< <(cmd)`), whose
+ * own operands are path candidates too — so each child is both read for its
+ * text and searched for nested executions (#741).
+ *
+ * Both passes are needed: a substitution can be the destination outright, or be
+ * concatenated into it (`> ${DIR}/$(cmd)`), and a `concatenation` is itself an
+ * argument node.
  */
 export function collectRedirectTokens(node: TSNode): string[] {
   const tokens: string[] = [];
@@ -59,10 +81,33 @@ export function collectRedirectTokens(node: TSNode): string[] {
     if (ARG_NODE_TYPES.has(child.type)) {
       tokens.push(resolveNodeText(child));
     }
+    tokens.push(...collectHostedExecutionTokens(child));
   }
   return tokens;
 }
 
+/**
+ * Collect the path-candidate tokens of every command nested inside `node`'s
+ * execution contexts, reading none of the host subtree's own text.
+ *
+ * This is what lets a heredoc body contribute its substitution's operands while
+ * its prose stays out of the path surface entirely.
+ *
+ * `node` may be a context outright (`> $(cmd)`) or merely contain one
+ * (`> ${DIR}/$(cmd)`); `forEachNestedExecution` searches strictly within a
+ * subtree, so the first case is checked here.
+ */
+function collectHostedExecutionTokens(node: TSNode): string[] {
+  if (NESTED_EXECUTION_CONTEXTS.has(node.type)) {
+    return collectPathCandidateTokens(node);
+  }
+  const tokens: string[] = [];
+  forEachNestedExecution(node, (contextNode) => {
+    tokens.push(...collectPathCandidateTokens(contextNode));
+  });
+  return tokens;
+}
+
 /**
  * Extract the command name from a `command` node.
  * Returns the basename (e.g. `/usr/bin/sed` → `sed`), or undefined