@basein/runner 0.3.20 → 0.3.21

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -360,13 +360,13 @@ same command, the same file, the same kind of arguments); any other call runs as
360
360
  written and is asked about as usual. A *directly executed* step never reaches the
361
361
  permission system at all — it is dispatched to a `bir-proxy` and called on a
362
362
  connection that is already open, or, for Claude Code's own `Read`, `Glob`,
363
- `Grep`, `Edit`, `Write` and `Bash`, run by the recorder itself in the project
364
- folder — and `run_scenario`, the call that starts the
363
+ `Grep`, `Edit`, `Write` and `Bash`, run by the recorder itself from the project
364
+ folder, on any file the agent could reach — and `run_scenario`, the call that starts the
365
365
  run, is answered `allow` while the recorder has announced the plan. In plan mode
366
366
  nothing is announced and nothing is approved. Set `BIR_REPLAY_ALLOW_SERVERS` to
367
367
  the servers you are content to have called unattended (it also covers a steered
368
- plan's recovery), `BIR_BUILTIN_TOOLS` and `BIR_BASH_ALLOW` to the built-ins and
369
- commands, and read [the guide's §5.1](docs/calculatedReplayGuide.md)
368
+ plan's recovery), `BIR_BUILTIN_TOOLS`, `BIR_PATH_ALLOW` and `BIR_BASH_ALLOW` to
369
+ the built-ins, folders and commands, and read [the guide's §5.1](docs/calculatedReplayGuide.md)
370
370
  before enabling it anywhere that matters.
371
371
 
372
372
  What it buys: on a scenario whose every step is a wrapped MCP tool or one of
@@ -489,8 +489,9 @@ what to do when they register, so there is one switch and no second copy of it:
489
489
  |---|---|
490
490
  | `BIR_REPLAY` | `0` turns replay off and any other value on, over the project's stored switch (`bir replay on` / `off`; a first plain `bir install` stores off). Nothing below matters while it is off. |
491
491
  | `BIR_REPLAY_ALLOW_SERVERS` | Server keys bir may call unattended: **direct** execution, and a steered plan's recovery. Unset means the list `bir replay allow` stored for the project, else every wrapped server; narrowing it is the recommendation. |
492
- | `BIR_BUILTIN_TOOLS` | Built-in tools bir may execute itself during a replay — in the session's own project directory, and confined to it. Default all six: `Read,Glob,Grep,Edit,Write,Bash` (up to 0.3.7 only the read-only `Read,Glob,Grep`). A runner-side step never reaches Claude Code's permission prompts, so leave off the list any tool you want the model to call itself. `none` turns it off, and built-in steps are steered in the session. |
492
+ | `BIR_BUILTIN_TOOLS` | Built-in tools bir may execute itself during a replay — from the session's own project directory; which folders the file tools may touch is `BIR_PATH_ALLOW`. Default all six: `Read,Glob,Grep,Edit,Write,Bash` (up to 0.3.7 only the read-only `Read,Glob,Grep`). A runner-side step never reaches Claude Code's permission prompts, so leave off the list any tool you want the model to call itself. `none` turns it off, and built-in steps are steered in the session. |
493
493
  | `BIR_BASH_ALLOW` | Command prefixes a runner-side `Bash` step may run. Unset means any command (`*`); a list such as `git,npm test` narrows it; `none` allows no command, and `Bash` steps are then steered in the session. A plan's Bash steps share a working directory, as Claude Code's do (variables and functions do not persist, also as in Claude Code). |
494
+ | `BIR_PATH_ALLOW` | Folders a runner-side `Read`, `Glob`, `Grep`, `Edit` or `Write` may touch. Unset means any folder (`*`), as the agent's own tools: a plan that reads the agent's memory file under `~/.claude` runs (from 0.3.21; up to 0.3.20 only the project directory). `project` keeps them inside the session's project directory; add folders after a comma, `~` for your home folder: `project,~/.claude`. `none` allows no folder, and file steps are then steered in the session. |
494
495
  | `BIR_DERIVE_API_KEY` | **Optional.** Working out what a new request is about — which fleet, which file, which date — is done for you by the service on its own key, as long as you are signed in. Set this (an Anthropic key) only to keep that reading on your machine, on your key. Signed out *and* unset, a scenario whose values change between requests is declined rather than replayed on stale ones. |
495
496
  | `ANTHROPIC_API_KEY` | Not used by the background recorder, which has Claude Code's environment and would bill your own key — unless `BIR_USE_ANTHROPIC_API_KEY=1`. A `bir-hooks` run by hand in a terminal uses that terminal's key. |
496
497
  | `BIR_DERIVE_MODEL` | Derivation model, when you set a key of your own (default `claude-haiku-4-5-20251001`). |
@@ -62,9 +62,9 @@
62
62
  * BIR_REPLAY_ALLOW_SERVERS comma-separated server keys eligible for *direct*
63
63
  * execution. Unset means every wrapped server
64
64
  * BIR_BUILTIN_TOOLS comma-separated built-ins the runner may execute
65
- * itself during a replay, in the session's own
66
- * project directory and confined to it
67
- * (replay/builtins.ts). Default all six:
65
+ * itself during a replay, from the session's own
66
+ * project directory (replay/builtins.ts; which
67
+ * folders: BIR_PATH_ALLOW). Default all six:
68
68
  * Read,Glob,Grep,Edit,Write,Bash. A runner-side
69
69
  * step never reaches the host's permission
70
70
  * prompts; Read,Glob,Grep is the read-only set
@@ -73,6 +73,13 @@
73
73
  * BIR_BASH_ALLOW command prefixes a runner-side Bash step may run
74
74
  * (`*` for all). Unset means `*`; `none` means no
75
75
  * command, and Bash steps are steered
76
+ * BIR_PATH_ALLOW folders a runner-side Read, Glob, Grep, Edit or
77
+ * Write may touch (`*` for any). Unset means `*`:
78
+ * any path, as the agent's own tools (up to 0.3.20
79
+ * the project directory alone). `project` is the
80
+ * session's project directory, `~` the home
81
+ * folder: `project,~/.claude`. `none` means no
82
+ * folder, and the file steps are steered
76
83
  * BIR_DERIVE_API_KEY OPTIONAL. Derivation — reading what this turn acts
77
84
  * on — is done by the service for a signed-in
78
85
  * runner. Set this (an Anthropic key) only to keep
@@ -62,9 +62,9 @@
62
62
  * BIR_REPLAY_ALLOW_SERVERS comma-separated server keys eligible for *direct*
63
63
  * execution. Unset means every wrapped server
64
64
  * BIR_BUILTIN_TOOLS comma-separated built-ins the runner may execute
65
- * itself during a replay, in the session's own
66
- * project directory and confined to it
67
- * (replay/builtins.ts). Default all six:
65
+ * itself during a replay, from the session's own
66
+ * project directory (replay/builtins.ts; which
67
+ * folders: BIR_PATH_ALLOW). Default all six:
68
68
  * Read,Glob,Grep,Edit,Write,Bash. A runner-side
69
69
  * step never reaches the host's permission
70
70
  * prompts; Read,Glob,Grep is the read-only set
@@ -73,6 +73,13 @@
73
73
  * BIR_BASH_ALLOW command prefixes a runner-side Bash step may run
74
74
  * (`*` for all). Unset means `*`; `none` means no
75
75
  * command, and Bash steps are steered
76
+ * BIR_PATH_ALLOW folders a runner-side Read, Glob, Grep, Edit or
77
+ * Write may touch (`*` for any). Unset means `*`:
78
+ * any path, as the agent's own tools (up to 0.3.20
79
+ * the project directory alone). `project` is the
80
+ * session's project directory, `~` the home
81
+ * folder: `project,~/.claude`. `none` means no
82
+ * folder, and the file steps are steered
76
83
  * BIR_DERIVE_API_KEY OPTIONAL. Derivation — reading what this turn acts
77
84
  * on — is done by the service for a signed-in
78
85
  * runner. Set this (an Anthropic key) only to keep
@@ -413,7 +413,7 @@ export function localFindings(turn) {
413
413
  const elsewhere = names.filter((n) => !EXECUTABLE_BUILTINS.has(n));
414
414
  const fixes = [];
415
415
  if (offList.length > 0) {
416
- fixes.push(`The runner runs ${offList.join(", ")} itself when ${offList.length === 1 ? "it is" : "they are"} on \`BIR_BUILTIN_TOOLS\` (unset means all six, from runner 0.3.8) and, for Bash, \`BIR_BASH_ALLOW\` is not \`none\`: fix those where the recorder starts, then \`bir up --restart\`.`);
416
+ fixes.push(`The runner runs ${offList.join(", ")} itself when ${offList.length === 1 ? "it is" : "they are"} on \`BIR_BUILTIN_TOOLS\` (unset means all six, from runner 0.3.8) and, for Bash, \`BIR_BASH_ALLOW\` is not \`none\` (for the file tools, \`BIR_PATH_ALLOW\`): fix those where the recorder starts, then \`bir up --restart\`.`);
417
417
  }
418
418
  if (elsewhere.length > 0) {
419
419
  fixes.push(`Put the data ${elsewhere.join(", ")} ${elsewhere.length === 1 ? "reaches" : "reach"} behind an MCP server, wrap it (\`bir install --server <name>\`), and record the prompt again so its steps are MCP calls.`);
@@ -26,7 +26,7 @@ import { DISCOVERY_CHECK_MS, HEARTBEAT_MS, pidAlive, readDiscovery, removeDiscov
26
26
  import { DAEMON_CWD_ENV, DAEMON_ENV, DAEMON_LOG_ENV, DERIVE_KEY_ENV, DaemonLog, USE_API_KEY_ENV, clearPortProblem, compareVersions, probeHealth, projectEndpoint, rediscover, validPort, writePortProblem, } from "./daemon.js";
27
27
  import { configDir, normalizePath, supersededLogPath } from "./paths.js";
28
28
  import { RecorderKeeper } from "./recorder-keeper.js";
29
- import { parseBashAllow, parseBuiltinTools } from "../replay/builtins.js";
29
+ import { FILE_BUILTINS, parseBashAllow, parseBuiltinTools, parsePathAllow } from "../replay/builtins.js";
30
30
  import { parseAllowList } from "../replay/coverage.js";
31
31
  import { sweepDeliveries } from "../replay/bash-channel.js";
32
32
  import { exemptLoopbackFromProxy, isWrapped, projectRecord, projectToken, readSidecar, } from "../config/generate.js";
@@ -197,15 +197,18 @@ function buildReplayOptions(keeper, policy, key) {
197
197
  : undefined;
198
198
  // Which built-ins the runner may execute itself (replay/builtins.ts):
199
199
  // all six by default, `BIR_BUILTIN_TOOLS` to narrow them, `BIR_BASH_ALLOW`
200
- // to narrow the commands a `Bash` step may run. Environment-only knobs,
201
- // like `BIR_REPLAY_RECORDED_OUTPUTS`.
200
+ // to narrow the commands a `Bash` step may run, `BIR_PATH_ALLOW` the
201
+ // folders a file step may touch (unset: any, as the agent's own tools).
202
+ // Environment-only knobs, like `BIR_REPLAY_RECORDED_OUTPUTS`.
202
203
  const builtinTools = parseBuiltinTools(process.env.BIR_BUILTIN_TOOLS);
203
204
  const bashAllow = parseBashAllow(process.env.BIR_BASH_ALLOW);
205
+ const pathAllow = parsePathAllow(process.env.BIR_PATH_ALLOW);
204
206
  const opts = {
205
207
  enabled,
206
208
  allowServers,
207
209
  builtinTools,
208
210
  bashAllow,
211
+ pathAllow,
209
212
  budgets: {
210
213
  // What a prompt may wait for its match (K5). Past it the prompt goes on
211
214
  // without one; a service slower than this must not cost every prompt.
@@ -263,6 +266,15 @@ function buildReplayOptions(keeper, policy, key) {
263
266
  // all — loud for the same reason as the rest of this line.
264
267
  builtins: builtinTools.size > 0 ? [...builtinTools].join(",") : "(none)",
265
268
  bashAllow: builtinTools.has("Bash") ? (bashAllow.length > 0 ? bashAllow.join(",") : "(none — Bash steps are steered)") : undefined,
269
+ // Where the file tools may reach: anywhere, as the agent's own, unless
270
+ // BIR_PATH_ALLOW names folders (up to 0.3.20 the project alone).
271
+ pathAllow: [...builtinTools].some((name) => FILE_BUILTINS.has(name))
272
+ ? pathAllow.includes("*")
273
+ ? "(any folder)"
274
+ : pathAllow.length > 0
275
+ ? pathAllow.join(",")
276
+ : "(none — file steps are steered)"
277
+ : undefined,
266
278
  // Who reads what this turn acts on (segmented.md R-PARAM-5). Named at
267
279
  // startup because the third state is a replay that silently never runs.
268
280
  derive: key.source ??
@@ -2212,8 +2212,8 @@ export class ControlServer {
2212
2212
  }
2213
2213
  state.sessionId = session.sessionId;
2214
2214
  // The project directory an allow-listed built-in step executes in, and
2215
- // the jail it may not leave (replay/builtins.ts). The session's own cwd
2216
- // from the hook payloads, never the daemon's.
2215
+ // what its relative paths resolve against (replay/builtins.ts). The
2216
+ // session's own cwd from the hook payloads, never the daemon's.
2217
2217
  state.sessionCwd = session.cwd ?? this.opts.cwd;
2218
2218
  // The service offered to record this turn as the recording's newest run:
2219
2219
  // should the agent redo the task after the plan, the plan's own calls go
@@ -2372,8 +2372,8 @@ export class ControlServer {
2372
2372
  }
2373
2373
  state.sessionId = session.sessionId;
2374
2374
  // The project directory an allow-listed built-in step executes in, and
2375
- // the jail it may not leave (replay/builtins.ts). The session's own cwd
2376
- // from the hook payloads, never the daemon's.
2375
+ // what its relative paths resolve against (replay/builtins.ts). The
2376
+ // session's own cwd from the hook payloads, never the daemon's.
2377
2377
  state.sessionCwd = session.cwd ?? this.opts.cwd;
2378
2378
  run.replay = state;
2379
2379
  run.replays.push(state);
@@ -4715,8 +4715,8 @@ export class ControlServer {
4715
4715
  }
4716
4716
  state.sessionId = session.sessionId;
4717
4717
  // The project directory an allow-listed built-in step executes in, and
4718
- // the jail it may not leave (replay/builtins.ts). The session's own cwd
4719
- // from the hook payloads, never the daemon's.
4718
+ // what its relative paths resolve against (replay/builtins.ts). The
4719
+ // session's own cwd from the hook payloads, never the daemon's.
4720
4720
  state.sessionCwd = session.cwd ?? this.opts.cwd;
4721
4721
  // The request this plan serves, for a note at its decision (SB-18): the
4722
4722
  // turn's prompt. `arm` was given this step's intent instead.
@@ -5,7 +5,7 @@
5
5
  * A built-in lives inside the host session, not behind any MCP server, so there
6
6
  * is no proxy to relay it to: historically every built-in step threw `not_run`
7
7
  * and the plan handed the task back to the model. This module is the other
8
- * option: execute the step here, against the *session's own project directory*,
8
+ * option: execute the step here, from the *session's own project directory*,
9
9
  * so a chain that starts with a `Read` no longer costs a hand-over.
10
10
  *
11
11
  * THE RULES, because this bypasses the host's permission prompts entirely:
@@ -17,10 +17,17 @@
17
17
  * already has — its pinned call is answered `allow` too. What running it
18
18
  * here changes is that no model sees the call before it runs, so an
19
19
  * operator who wants one in the loop takes the tool off the list.
20
- * - **Path jail.** Every path is resolved against, and confined to, the
21
- * session's project directory ({@link jailed}), symlinks included. The
22
- * daemon's own cwd is never used: it is `~/.baseinstrunner`, not the
23
- * project (§5.1 of the change plan).
20
+ * - **Any path the agent could reach, unless `BIR_PATH_ALLOW` narrows it.**
21
+ * A relative path resolves against the session's project directory, as
22
+ * the host's tools resolve it; the daemon's own cwd is never used: it is
23
+ * `~/.baseinstrunner`, not the project (§5.1 of the change plan). An
24
+ * absolute path is taken as it is, inside the project or not: the agent's
25
+ * own tools reach any folder its session allows — the recorded call was
26
+ * one of them — and a runner-side `Bash` reaches any folder already. Up
27
+ * to 0.3.20 every path was confined to the project directory;
28
+ * `BIR_PATH_ALLOW=project` brings that back, and a list of folders
29
+ * ({@link parsePathAllow}) confines the file tools to those, symlinks
30
+ * included ({@link allowedPath}).
24
31
  * - **Bash runs any command unless `BIR_BASH_ALLOW` narrows it.** Set, a
25
32
  * command must match one of its prefixes ({@link bashAllowed}); `none`
26
33
  * allows no command, and then `Bash` counts as off the list, so its steps
@@ -48,6 +55,12 @@ export declare const EXECUTABLE_BUILTINS: ReadonlySet<string>;
48
55
  * set (`Read,Glob,Grep`).
49
56
  */
50
57
  export declare const DEFAULT_BUILTIN_TOOLS: readonly string[];
58
+ /**
59
+ * The built-ins that reach files by a path, the ones `BIR_PATH_ALLOW`
60
+ * governs. `Bash` is not among them: what a command touches is
61
+ * `BIR_BASH_ALLOW`'s business.
62
+ */
63
+ export declare const FILE_BUILTINS: ReadonlySet<string>;
51
64
  /**
52
65
  * Parse `BIR_BUILTIN_TOOLS`. Unset means {@link DEFAULT_BUILTIN_TOOLS}; set,
53
66
  * it is the whole list — `none` (or an empty value) turns built-ins off
@@ -63,6 +76,19 @@ export declare function parseBuiltinTools(value: string | undefined): ReadonlySe
63
76
  * value) allows no command.
64
77
  */
65
78
  export declare function parseBashAllow(value: string | undefined): readonly string[];
79
+ /**
80
+ * Parse `BIR_PATH_ALLOW`: the folders a runner-side `Read`, `Glob`, `Grep`,
81
+ * `Edit` or `Write` may touch, comma-separated, `*` for anywhere. Unset means
82
+ * anywhere, as the agent's own tools reach any folder its session allows (up
83
+ * to 0.3.20 it was the project directory alone). In a list, `project` is the
84
+ * session's project directory, `~` the home folder (`~/.claude` a folder in
85
+ * it), and any other relative folder resolves against the project directory.
86
+ * Set, it is the whole list, as `BIR_BASH_ALLOW` is: `none` (or an empty
87
+ * value) allows no folder, and the file tools then count as off
88
+ * `BIR_BUILTIN_TOOLS`, so their steps are steered in the session
89
+ * (controller.ts).
90
+ */
91
+ export declare function parsePathAllow(value: string | undefined): readonly string[];
66
92
  /**
67
93
  * Whether `command` matches the `BIR_BASH_ALLOW` prefixes. A prefix that ends
68
94
  * mid-word must end the command's word too: `git` allows `git status` and
@@ -72,8 +98,9 @@ export declare function parseBashAllow(value: string | undefined): readonly stri
72
98
  export declare function bashAllowed(command: string, allow: readonly string[] | undefined): boolean;
73
99
  /**
74
100
  * The step cannot run here — policy, not breakage: the tool is not
75
- * implemented, its path leaves the project, `Bash` is not allowed, or the
76
- * input is not even the right shape. Maps to `not_run`: nothing happened, the
101
+ * implemented, its path is outside the folders `BIR_PATH_ALLOW` names, the
102
+ * command is not one `BIR_BASH_ALLOW` allows, or the input is not even the
103
+ * right shape. Maps to `not_run`: nothing happened, the
77
104
  * session may still do it, a recorded output may stand in.
78
105
  */
79
106
  export declare class BuiltinNotRunnable extends Error {
@@ -90,8 +117,9 @@ export declare class BuiltinTimeout extends Error {
90
117
  }
91
118
  export interface BuiltinContext {
92
119
  /**
93
- * The session's project directory — the jail every path resolves against.
94
- * The caller (controller.ts) refuses to execute at all without one; the
120
+ * The session's project directory — what a relative path resolves
121
+ * against, and the folder `project` names in {@link pathAllow}. The
122
+ * caller (controller.ts) refuses to execute at all without one; the
95
123
  * daemon's own cwd is never a substitute.
96
124
  */
97
125
  cwd: string;
@@ -99,6 +127,14 @@ export interface BuiltinContext {
99
127
  timeoutMs?: number;
100
128
  /** Parsed `BIR_BASH_ALLOW` ({@link parseBashAllow}). Absent or empty, no command runs. */
101
129
  bashAllow?: readonly string[];
130
+ /**
131
+ * Parsed `BIR_PATH_ALLOW` ({@link parsePathAllow}): the folders a file
132
+ * step may touch, `*` for any. Absent means the project directory alone —
133
+ * a caller that says nothing gets the narrow answer, as with
134
+ * {@link bashAllow}; the controller always passes the environment's, which
135
+ * is `*` when the variable is unset. Empty, no file step runs.
136
+ */
137
+ pathAllow?: readonly string[];
102
138
  /**
103
139
  * Where this plan's *previous* `Bash` step left its shell — the host's own
104
140
  * semantics, which a recorded chain was written against: the working
@@ -5,7 +5,7 @@
5
5
  * A built-in lives inside the host session, not behind any MCP server, so there
6
6
  * is no proxy to relay it to: historically every built-in step threw `not_run`
7
7
  * and the plan handed the task back to the model. This module is the other
8
- * option: execute the step here, against the *session's own project directory*,
8
+ * option: execute the step here, from the *session's own project directory*,
9
9
  * so a chain that starts with a `Read` no longer costs a hand-over.
10
10
  *
11
11
  * THE RULES, because this bypasses the host's permission prompts entirely:
@@ -17,10 +17,17 @@
17
17
  * already has — its pinned call is answered `allow` too. What running it
18
18
  * here changes is that no model sees the call before it runs, so an
19
19
  * operator who wants one in the loop takes the tool off the list.
20
- * - **Path jail.** Every path is resolved against, and confined to, the
21
- * session's project directory ({@link jailed}), symlinks included. The
22
- * daemon's own cwd is never used: it is `~/.baseinstrunner`, not the
23
- * project (§5.1 of the change plan).
20
+ * - **Any path the agent could reach, unless `BIR_PATH_ALLOW` narrows it.**
21
+ * A relative path resolves against the session's project directory, as
22
+ * the host's tools resolve it; the daemon's own cwd is never used: it is
23
+ * `~/.baseinstrunner`, not the project (§5.1 of the change plan). An
24
+ * absolute path is taken as it is, inside the project or not: the agent's
25
+ * own tools reach any folder its session allows — the recorded call was
26
+ * one of them — and a runner-side `Bash` reaches any folder already. Up
27
+ * to 0.3.20 every path was confined to the project directory;
28
+ * `BIR_PATH_ALLOW=project` brings that back, and a list of folders
29
+ * ({@link parsePathAllow}) confines the file tools to those, symlinks
30
+ * included ({@link allowedPath}).
24
31
  * - **Bash runs any command unless `BIR_BASH_ALLOW` narrows it.** Set, a
25
32
  * command must match one of its prefixes ({@link bashAllowed}); `none`
26
33
  * allows no command, and then `Bash` counts as off the list, so its steps
@@ -43,7 +50,8 @@ import { execFile } from "node:child_process";
43
50
  import { randomBytes } from "node:crypto";
44
51
  import { existsSync } from "node:fs";
45
52
  import { readdir, readFile, realpath, stat, writeFile } from "node:fs/promises";
46
- import { dirname, join, resolve, sep } from "node:path";
53
+ import { homedir } from "node:os";
54
+ import { basename, dirname, join, relative, resolve, sep } from "node:path";
47
55
  // ── configuration ────────────────────────────────────────────────────────────
48
56
  /** The built-ins {@link runBuiltin} actually implements. */
49
57
  export const EXECUTABLE_BUILTINS = new Set([
@@ -61,8 +69,24 @@ export const EXECUTABLE_BUILTINS = new Set([
61
69
  * set (`Read,Glob,Grep`).
62
70
  */
63
71
  export const DEFAULT_BUILTIN_TOOLS = [...EXECUTABLE_BUILTINS];
72
+ /**
73
+ * The built-ins that reach files by a path, the ones `BIR_PATH_ALLOW`
74
+ * governs. `Bash` is not among them: what a command touches is
75
+ * `BIR_BASH_ALLOW`'s business.
76
+ */
77
+ export const FILE_BUILTINS = new Set(["Read", "Glob", "Grep", "Edit", "Write"]);
64
78
  /** `read` → `Read`: the vocabulary is small enough to forgive casing. */
65
79
  const CANONICAL_BUILTIN = new Map([...EXECUTABLE_BUILTINS].map((name) => [name.toLowerCase(), name]));
80
+ /**
81
+ * A switch's list once it is set: comma-separated, trimmed, empty entries
82
+ * dropped, and `none` / `off` / `0` meaning nothing at all.
83
+ */
84
+ function parseList(value) {
85
+ return value
86
+ .split(",")
87
+ .map((s) => s.trim())
88
+ .filter((s) => s.length > 0 && !["none", "off", "0"].includes(s.toLowerCase()));
89
+ }
66
90
  /**
67
91
  * Parse `BIR_BUILTIN_TOOLS`. Unset means {@link DEFAULT_BUILTIN_TOOLS}; set,
68
92
  * it is the whole list — `none` (or an empty value) turns built-ins off
@@ -73,12 +97,7 @@ const CANONICAL_BUILTIN = new Map([...EXECUTABLE_BUILTINS].map((name) => [name.t
73
97
  export function parseBuiltinTools(value) {
74
98
  if (value === undefined)
75
99
  return new Set(DEFAULT_BUILTIN_TOOLS);
76
- const names = value
77
- .split(",")
78
- .map((s) => s.trim())
79
- .filter((s) => s.length > 0 && !["none", "off", "0"].includes(s.toLowerCase()))
80
- .map((s) => CANONICAL_BUILTIN.get(s.toLowerCase()) ?? s);
81
- return new Set(names);
100
+ return new Set(parseList(value).map((s) => CANONICAL_BUILTIN.get(s.toLowerCase()) ?? s));
82
101
  }
83
102
  /**
84
103
  * Parse `BIR_BASH_ALLOW`: comma-separated command prefixes, `*` for
@@ -89,10 +108,24 @@ export function parseBuiltinTools(value) {
89
108
  export function parseBashAllow(value) {
90
109
  if (value === undefined)
91
110
  return ["*"];
92
- return value
93
- .split(",")
94
- .map((s) => s.trim())
95
- .filter((s) => s.length > 0 && !["none", "off", "0"].includes(s.toLowerCase()));
111
+ return parseList(value);
112
+ }
113
+ /**
114
+ * Parse `BIR_PATH_ALLOW`: the folders a runner-side `Read`, `Glob`, `Grep`,
115
+ * `Edit` or `Write` may touch, comma-separated, `*` for anywhere. Unset means
116
+ * anywhere, as the agent's own tools reach any folder its session allows (up
117
+ * to 0.3.20 it was the project directory alone). In a list, `project` is the
118
+ * session's project directory, `~` the home folder (`~/.claude` a folder in
119
+ * it), and any other relative folder resolves against the project directory.
120
+ * Set, it is the whole list, as `BIR_BASH_ALLOW` is: `none` (or an empty
121
+ * value) allows no folder, and the file tools then count as off
122
+ * `BIR_BUILTIN_TOOLS`, so their steps are steered in the session
123
+ * (controller.ts).
124
+ */
125
+ export function parsePathAllow(value) {
126
+ if (value === undefined)
127
+ return ["*"];
128
+ return parseList(value);
96
129
  }
97
130
  /**
98
131
  * Whether `command` matches the `BIR_BASH_ALLOW` prefixes. A prefix that ends
@@ -117,8 +150,9 @@ export function bashAllowed(command, allow) {
117
150
  // ── the failure contract ─────────────────────────────────────────────────────
118
151
  /**
119
152
  * The step cannot run here — policy, not breakage: the tool is not
120
- * implemented, its path leaves the project, `Bash` is not allowed, or the
121
- * input is not even the right shape. Maps to `not_run`: nothing happened, the
153
+ * implemented, its path is outside the folders `BIR_PATH_ALLOW` names, the
154
+ * command is not one `BIR_BASH_ALLOW` allows, or the input is not even the
155
+ * right shape. Maps to `not_run`: nothing happened, the
122
156
  * session may still do it, a recorded output may stand in.
123
157
  */
124
158
  export class BuiltinNotRunnable extends Error {
@@ -186,7 +220,9 @@ export async function runBuiltin(toolName, input, ctx) {
186
220
  throw new BuiltinNotRunnable(`no runner-side implementation for ${toolName}`);
187
221
  }
188
222
  }
189
- // ── the path jail ────────────────────────────────────────────────────────────
223
+ // ── which paths a file step may touch ────────────────────────────────────────
224
+ /** The word in `BIR_PATH_ALLOW` that stands for the session's project directory. */
225
+ const PROJECT_WORD = "project";
190
226
  /** Case-blind on Windows, where `C:\Repo` and `c:\repo` are one directory. */
191
227
  function samePath(a, b) {
192
228
  return process.platform === "win32" ? a.toLowerCase() === b.toLowerCase() : a === b;
@@ -199,54 +235,87 @@ function underRoot(root, abs) {
199
235
  : abs.startsWith(prefix)));
200
236
  }
201
237
  /**
202
- * Resolve `p` against the session root and refuse anything outside it — first
203
- * textually, then through `realpath` so a symlink inside the project cannot
204
- * point the step at a file beyond it. The nearest existing ancestor is what
205
- * gets resolved, so a `Write` to a new file is judged by its real directory.
238
+ * The folders the context's `BIR_PATH_ALLOW` list names, as absolute paths,
239
+ * or `anywhere` for `*`. `project` is the session's project directory, `~`
240
+ * the home folder, and any other relative entry resolves against the project.
206
241
  */
207
- async function jailed(root, p, what = "file_path") {
208
- if (typeof p !== "string" || p.trim() === "") {
209
- throw new BuiltinNotRunnable(`${what} is required`);
210
- }
211
- const r = resolve(root);
212
- const abs = resolve(r, p);
213
- const outside = () => {
214
- throw new BuiltinNotRunnable(`${p} is outside the session's project directory, so bir does not touch it unattended`);
215
- };
216
- if (!underRoot(r, abs))
217
- outside();
218
- // The textual check passed; now the physical one. Walk up to what exists,
219
- // realpath it, and require *that* to sit under the root's own realpath.
242
+ function allowedFolders(ctx) {
243
+ const allow = ctx.pathAllow ?? [PROJECT_WORD];
244
+ if (allow.includes("*"))
245
+ return "anywhere";
246
+ const project = resolve(ctx.cwd);
247
+ return allow.map((entry) => {
248
+ if (entry.toLowerCase() === PROJECT_WORD)
249
+ return project;
250
+ if (entry === "~")
251
+ return homedir();
252
+ if (entry.startsWith("~/") || entry.startsWith("~\\"))
253
+ return resolve(homedir(), entry.slice(2));
254
+ return resolve(project, entry);
255
+ });
256
+ }
257
+ /** Why a path was refused, naming the switch that refused it. */
258
+ function outsideMessage(p, allow) {
259
+ return allow.length === 1 && allow[0].toLowerCase() === PROJECT_WORD
260
+ ? `${p} is outside the session's project directory, and BIR_PATH_ALLOW=project keeps bir inside it`
261
+ : `${p} is outside the folders BIR_PATH_ALLOW names (${allow.join(", ")}), so bir does not touch it unattended`;
262
+ }
263
+ /**
264
+ * `abs` as it is on disk: its longest existing ancestor replaced by that
265
+ * ancestor's realpath (symlinks resolved, and on macOS `/var` read as
266
+ * `/private/var`), the rest kept as written. A path none of which exists
267
+ * comes back as it is.
268
+ */
269
+ async function physical(abs) {
220
270
  let probe = abs;
221
- let real;
271
+ const rest = [];
222
272
  for (;;) {
223
273
  try {
224
- real = await realpath(probe);
225
- break;
274
+ const real = await realpath(probe);
275
+ return rest.length > 0 ? join(real, ...rest.reverse()) : real;
226
276
  }
227
277
  catch {
228
278
  const parent = dirname(probe);
229
279
  if (samePath(parent, probe))
230
- break; // hit the filesystem root: nothing exists
280
+ return abs; // the filesystem root: nothing exists
281
+ rest.push(basename(probe));
231
282
  probe = parent;
232
283
  }
233
284
  }
234
- if (real !== undefined) {
235
- let realRoot;
236
- try {
237
- realRoot = await realpath(r);
238
- }
239
- catch {
240
- throw new BuiltinNotRunnable(`the session's project directory does not exist: ${root}`);
241
- }
242
- if (!underRoot(realRoot, real))
243
- outside();
285
+ }
286
+ /**
287
+ * Resolve `p` as the host does — against the session's project directory
288
+ * when it is relative — and allow it anywhere, as the agent's own tools
289
+ * would, unless `BIR_PATH_ALLOW` names folders. Then anything outside them is
290
+ * refused: first as written, then physically ({@link physical}, both sides),
291
+ * so a symlink inside an allowed folder cannot point the step at a file
292
+ * beyond it, and a `Write` to a new file is judged by its real folder.
293
+ */
294
+ async function allowedPath(ctx, p, what = "file_path") {
295
+ if (typeof p !== "string" || p.trim() === "") {
296
+ throw new BuiltinNotRunnable(`${what} is required`);
244
297
  }
298
+ const abs = resolve(ctx.cwd, p);
299
+ const folders = allowedFolders(ctx);
300
+ if (folders === "anywhere")
301
+ return abs;
302
+ if (folders.length === 0) {
303
+ throw new BuiltinNotRunnable("BIR_PATH_ALLOW allows no folder, so bir touches no file unattended");
304
+ }
305
+ const outside = () => {
306
+ throw new BuiltinNotRunnable(outsideMessage(p, ctx.pathAllow ?? [PROJECT_WORD]));
307
+ };
308
+ if (!folders.some((folder) => underRoot(folder, abs)))
309
+ outside();
310
+ const real = await physical(abs);
311
+ const realFolders = await Promise.all(folders.map(physical));
312
+ if (!realFolders.some((folder) => underRoot(folder, real)))
313
+ outside();
245
314
  return abs;
246
315
  }
247
316
  // ── Read ─────────────────────────────────────────────────────────────────────
248
317
  async function runRead(input, ctx) {
249
- const filePath = await jailed(ctx.cwd, input.file_path);
318
+ const filePath = await allowedPath(ctx, input.file_path);
250
319
  const size = (await stat(filePath)).size;
251
320
  if (size > MAX_FILE_BYTES) {
252
321
  throw new Error(`file is ${size} bytes, more than the ${MAX_FILE_BYTES} this runner reads`);
@@ -276,7 +345,7 @@ async function runGlob(input, ctx, deadline) {
276
345
  if (typeof pattern !== "string" || pattern.trim() === "") {
277
346
  throw new BuiltinNotRunnable("pattern is required");
278
347
  }
279
- const root = await jailed(ctx.cwd, typeof input.path === "string" ? input.path : ".", "path");
348
+ const root = await allowedPath(ctx, typeof input.path === "string" ? input.path : ".", "path");
280
349
  const startedAt = Date.now();
281
350
  const re = globToRegExp(pattern);
282
351
  const found = [];
@@ -324,7 +393,7 @@ async function runGrep(input, ctx, deadline) {
324
393
  catch (err) {
325
394
  throw new BuiltinNotRunnable(`the pattern is not a regular expression this runner can run: ${err.message}`);
326
395
  }
327
- const target = await jailed(ctx.cwd, typeof input.path === "string" ? input.path : ".", "path");
396
+ const target = await allowedPath(ctx, typeof input.path === "string" ? input.path : ".", "path");
328
397
  // ripgrep's `-g '*.ts'` matches the basename at any depth; only a pattern
329
398
  // with a `/` in it speaks about the path.
330
399
  const globRe = typeof input.glob === "string" && input.glob
@@ -417,7 +486,7 @@ async function matchFile(path, re, firstOnly) {
417
486
  }
418
487
  // ── Edit / Write ─────────────────────────────────────────────────────────────
419
488
  async function runEdit(input, ctx) {
420
- const filePath = await jailed(ctx.cwd, input.file_path);
489
+ const filePath = await allowedPath(ctx, input.file_path);
421
490
  const oldString = input.old_string;
422
491
  const newString = input.new_string;
423
492
  if (typeof oldString !== "string" || oldString === "") {
@@ -452,7 +521,7 @@ async function runEdit(input, ctx) {
452
521
  };
453
522
  }
454
523
  async function runWrite(input, ctx) {
455
- const filePath = await jailed(ctx.cwd, input.file_path);
524
+ const filePath = await allowedPath(ctx, input.file_path);
456
525
  const content = input.content;
457
526
  if (typeof content !== "string")
458
527
  throw new BuiltinNotRunnable("content is required");
@@ -579,9 +648,10 @@ function takePwdMarker(stderr, mark) {
579
648
  const SKIP_DIRS = new Set([".git", "node_modules", ".hg", ".svn"]);
580
649
  /**
581
650
  * Walk `root`, calling `visit(relativePath, absolutePath, mtimeMs)` for every
582
- * regular file. Never follows directory symlinks — the jail was checked at
583
- * the root, and a link out of it must not re-open the door. Returns whether
584
- * the walk was cut short (entry cap, deadline, or `visit` saying stop).
651
+ * regular file. Never follows directory symlinks — the folder was checked at
652
+ * the root, and under a `BIR_PATH_ALLOW` list a link out of it must not
653
+ * re-open the door (nor may a link loop the walk). Returns whether the walk
654
+ * was cut short (entry cap, deadline, or `visit` saying stop).
585
655
  */
586
656
  async function walk(root, deadline, visit) {
587
657
  let seen = 0;
@@ -617,7 +687,9 @@ async function walk(root, deadline, visit) {
617
687
  catch {
618
688
  continue;
619
689
  }
620
- const rel = abs.slice(resolve(root).length + 1).split(sep).join("/");
690
+ // `relative`, not a slice: a walk may now start at a drive's or the
691
+ // disk's root (`C:\`, `/`), which already ends in a separator.
692
+ const rel = relative(root, abs).split(sep).join("/");
621
693
  if (!visit(rel, abs, mtime))
622
694
  return true;
623
695
  }
@@ -266,6 +266,16 @@ export interface ReplayOptions {
266
266
  * rather than armed `direct` and stopped at the first one.
267
267
  */
268
268
  bashAllow?: readonly string[];
269
+ /**
270
+ * `BIR_PATH_ALLOW` — the folders a runner-side `Read`, `Glob`, `Grep`,
271
+ * `Edit` or `Write` may touch (`*` for any; `project` is the session's
272
+ * project directory). Omitted, the environment decides; the default is
273
+ * `*`, any path the agent's own tools could reach (up to 0.3.20 the
274
+ * project directory alone). Empty allows no folder, and then the file tools
275
+ * count as off {@link ReplayOptions.builtinTools}, as `Bash` does under an
276
+ * empty {@link ReplayOptions.bashAllow}.
277
+ */
278
+ pathAllow?: readonly string[];
269
279
  budgets?: Partial<ReplayBudgets>;
270
280
  /**
271
281
  * The service: the lazy source-run fetch, and the parameter derivation this
@@ -418,8 +428,9 @@ export interface ReplayState {
418
428
  /**
419
429
  * That session's project directory, set by the control server from the
420
430
  * hook payloads — never the daemon's own cwd, which is `~/.baseinstrunner`.
421
- * It is where an allow-listed built-in step executes and the jail it may
422
- * not leave (replay/builtins.ts). Absent — `bir replay`, an older server —
431
+ * It is where an allow-listed built-in step executes and what its relative
432
+ * paths resolve against (replay/builtins.ts; `BIR_PATH_ALLOW=project`
433
+ * confines the file tools to it). Absent — `bir replay`, an older server —
423
434
  * no built-in executes: the step is `not_run`, exactly as before.
424
435
  */
425
436
  sessionCwd?: string;
@@ -689,6 +700,8 @@ export declare class ReplayController {
689
700
  private readonly builtinTools;
690
701
  /** See {@link ReplayOptions.bashAllow}. */
691
702
  private readonly bashAllow;
703
+ /** See {@link ReplayOptions.pathAllow}. */
704
+ private readonly pathAllow;
692
705
  /** Calls waiting for a plan's pins to change (see `awaitPins`). */
693
706
  private readonly pinWaiters;
694
707
  /**
@@ -18,7 +18,7 @@ import { capText, DEFAULT_MAX_PAYLOAD_BYTES, REPORT_PAYLOAD_MAX_BYTES, serialize
18
18
  import { logDetail, logLine, errText } from "../util/log.js";
19
19
  import { serviceHealth, serviceTimeoutMs, timeoutSignal, } from "../util/service-health.js";
20
20
  import { bashCommandFor } from "./bash-channel.js";
21
- import { BuiltinNotRunnable, BuiltinTimeout, parseBashAllow, parseBuiltinTools, runBuiltin, } from "./builtins.js";
21
+ import { BuiltinNotRunnable, BuiltinTimeout, FILE_BUILTINS, parseBashAllow, parseBuiltinTools, parsePathAllow, runBuiltin, } from "./builtins.js";
22
22
  import { BASH_CHANNEL_MAX, MAX_REPLAY_REASON, testWords, truncate } from "./bundle.js";
23
23
  import { coverageOf, modeFor, reachOf } from "./coverage.js";
24
24
  import { deriveParameters, localDay, } from "./derive.js";
@@ -216,6 +216,8 @@ export class ReplayController {
216
216
  builtinTools;
217
217
  /** See {@link ReplayOptions.bashAllow}. */
218
218
  bashAllow;
219
+ /** See {@link ReplayOptions.pathAllow}. */
220
+ pathAllow;
219
221
  /** Calls waiting for a plan's pins to change (see `awaitPins`). */
220
222
  pinWaiters = new WeakMap();
221
223
  /**
@@ -277,13 +279,17 @@ export class ReplayController {
277
279
  this.recordedFallback =
278
280
  opts.recordedFallback ?? process.env.BIR_REPLAY_RECORDED_OUTPUTS === "1";
279
281
  this.bashAllow = opts.bashAllow ?? parseBashAllow(process.env.BIR_BASH_ALLOW);
282
+ this.pathAllow = opts.pathAllow ?? parsePathAllow(process.env.BIR_PATH_ALLOW);
280
283
  const builtins = opts.builtinTools ?? parseBuiltinTools(process.env.BIR_BUILTIN_TOOLS);
281
- // A `Bash` that may run no command is not reachable here. Left on the
282
- // list, a chain of Bash steps would arm `direct` and stop at its first.
283
- this.builtinTools =
284
- this.bashAllow.length > 0 || !builtins.has("Bash")
285
- ? builtins
286
- : new Set([...builtins].filter((name) => name !== "Bash"));
284
+ // A `Bash` that may run no command, or a file tool that may touch no
285
+ // folder, is not reachable here. Left on the list, a chain of such steps
286
+ // would arm `direct` and stop at its first.
287
+ const reachable = (name) => name === "Bash"
288
+ ? this.bashAllow.length > 0
289
+ : !FILE_BUILTINS.has(name) || this.pathAllow.length > 0;
290
+ this.builtinTools = [...builtins].every(reachable)
291
+ ? builtins
292
+ : new Set([...builtins].filter(reachable));
287
293
  }
288
294
  /**
289
295
  * Whether a handed-out segment may actually arm (segmented.md R-OUT-10).
@@ -2873,11 +2879,12 @@ export class ReplayController {
2873
2879
  }
2874
2880
  const mcp = parseQualifiedName(name);
2875
2881
  if (!mcp) {
2876
- // An allow-listed built-in runs *here*, against the session's own
2882
+ // An allow-listed built-in runs *here*, from the session's own
2877
2883
  // project directory (replay/builtins.ts). Everything else about it —
2878
2884
  // no session cwd (`bir replay`, an older server), off the list, no
2879
- // runner-side implementation, a path that leaves the project, a `Bash`
2880
- // command `BIR_BASH_ALLOW` does not allow — stays the old contract:
2885
+ // runner-side implementation, a path outside the folders
2886
+ // `BIR_PATH_ALLOW` names, a `Bash` command `BIR_BASH_ALLOW` does not
2887
+ // allow — stays the old contract:
2881
2888
  // `not_run`, so the session can do the step itself or its recorded
2882
2889
  // output can stand in (`BIR_REPLAY_RECORDED_OUTPUTS=1`).
2883
2890
  if (this.builtinTools.has(name) && state.sessionCwd) {
@@ -2939,6 +2946,9 @@ export class ReplayController {
2939
2946
  cwd,
2940
2947
  timeoutMs,
2941
2948
  bashAllow: this.bashAllow,
2949
+ // Any folder unless `BIR_PATH_ALLOW` names some, as the agent's own
2950
+ // tools would reach it (up to 0.3.20, the project directory alone).
2951
+ pathAllow: this.pathAllow,
2942
2952
  // The plan's Bash steps share a working directory, as the host's do.
2943
2953
  bashCwd: state.bashCwd,
2944
2954
  onBashCwd: (pwd) => {
@@ -45,7 +45,7 @@ this design has a ladder whose cheapest rung is free — and the more of a stack
45
45
  |---|---|---|
46
46
  | Runs MCP tools itself | No (`replay.md` §12: *"rrepeat never executes tools — not even MCP"*) | **Yes**, on the connection it already owns |
47
47
  | Divergence fallback | A second Agent SDK session — the dominant cost | Direct `tools/call` through the proxies — **$0** |
48
- | Reaches built-ins (`Bash`, `Read`) | Only by steering the live session | `Read`, `Glob`, `Grep`, `Edit`, `Write` and `Bash` run on the runner itself, jailed to the project (`BIR_BUILTIN_TOOLS`, `BIR_BASH_ALLOW`); any other built-in is steered. In a host that is not Claude Code a tool with one of these names is the application's own, and the runner runs none of them (§7.4) |
48
+ | Reaches built-ins (`Bash`, `Read`) | Only by steering the live session | `Read`, `Glob`, `Grep`, `Edit`, `Write` and `Bash` run on the runner itself, on any path the agent could reach unless `BIR_PATH_ALLOW` names folders (`BIR_BUILTIN_TOOLS`, `BIR_BASH_ALLOW`, `BIR_PATH_ALLOW`); any other built-in is steered. In a host that is not Claude Code a tool with one of these names is the application's own, and the runner runs none of them (§7.4) |
49
49
  | Needs `@anthropic-ai/*` at runtime | Yes | No (§16.3) |
50
50
 
51
51
  ---
@@ -265,7 +265,7 @@ export function modeFor(steps: SerializedScenarioStep[], wrapped: ReadonlySet<st
265
265
 
266
266
  | Every step | Mode | Where the tools run | Model spend |
267
267
  |---|---|---|---|
268
- | a wrapped MCP tool, or a built-in on `BIR_BUILTIN_TOOLS` | **`direct`** (§6) | The proxies, on connections already open; allow-listed built-ins on the runner itself, jailed to the session's project directory (replay/builtins.ts) | one Haiku derivation + one turn that reads the results |
268
+ | a wrapped MCP tool, or a built-in on `BIR_BUILTIN_TOOLS` | **`direct`** (§6) | The proxies, on connections already open; allow-listed built-ins on the runner itself, from the session's project directory (replay/builtins.ts) | one Haiku derivation + one turn that reads the results |
269
269
  | anything else, or mixed | **`steer`** (§7) | The live session, inputs pinned | one Haiku derivation + the live turn |
270
270
  | — | `none` | nowhere | declines at gate 5 |
271
271
 
@@ -275,11 +275,17 @@ so `Task`, `WebFetch`; up to 0.3.7 also `Bash`, `Edit`, `Write`), an MCP server
275
275
  `claude-in-chrome` — which is `scope: "dynamic"`, appears in no config file, and therefore can never
276
276
  be wrapped (v1 design §1). None of them can be reached except through the session.
277
277
 
278
- A built-in **on** the allow-list runs on the runner, in the session's own project directory and
279
- confined to it — a path that resolves outside the project, symlinks included, is refused and the
280
- step is `not_run`, the clean hand-over. `Bash` has a second switch on top (`BIR_BASH_ALLOW`,
281
- command prefixes; unset means any command, `none` keeps `Bash` in the session), because a
282
- runner-side step never reaches the host's permission prompts. The
278
+ A built-in **on** the allow-list runs on the runner, from the session's own project directory: a
279
+ relative path resolves against it, as the host's tools resolve it, and an absolute path is taken
280
+ as it is, inside the project or not — the agent's own tools reach any folder its session allows,
281
+ and the recorded call was one of them (up to 0.3.20 every path was confined to the project, and a
282
+ plan that read the agent's memory file under `~/.claude` handed the whole task back). The file
283
+ tools have a switch of their own (`BIR_PATH_ALLOW`, folders; unset means any, `project` brings
284
+ back the confinement, `project,~/.claude` adds a folder, `none` keeps them in the session); under
285
+ a list, a path outside it, symlinks included, is refused and the step is `not_run`, the clean
286
+ hand-over. `Bash` has its switch too (`BIR_BASH_ALLOW`, command prefixes; unset means any command,
287
+ `none` keeps `Bash` in the session), because a runner-side step never reaches the host's
288
+ permission prompts. The
283
289
  results mimic the hook's view of each tool, which is what the recorder stored for built-in steps
284
290
  and what their `toolOutputLogic` was authored against; they pass through the same
285
291
  `serializeCapped(redact(…))` as every recorded payload. Bash keeps the host's own semantics a
@@ -241,8 +241,10 @@ When replay arms:
241
241
  §7) — and never in plan mode.
242
242
  - **Direct steps are not visible to the permission system at all.** They are dispatched straight to a
243
243
  `bir-proxy`, which calls the upstream on a connection it already holds — or, for Claude Code's own
244
- `Read`, `Glob`, `Grep`, `Edit`, `Write` and `Bash`, run by the recorder itself in the project
245
- folder, any `Bash` command included (`BIR_BUILTIN_TOOLS`, `BIR_BASH_ALLOW`; up to 0.3.7 only the
244
+ `Read`, `Glob`, `Grep`, `Edit`, `Write` and `Bash`, run by the recorder itself from the project
245
+ folder, on any file the agent could reach (from 0.3.21; up to 0.3.20 only files inside the
246
+ project) and any `Bash` command included (`BIR_BUILTIN_TOOLS`, `BIR_PATH_ALLOW`,
247
+ `BIR_BASH_ALLOW`; up to 0.3.7 only the
246
248
  first three). Claude Code never sees them.
247
249
  The `run_scenario` call that starts them is answered `allow` too, while the runner has announced a
248
250
  direct plan; your own deny or ask rules for it still apply.
@@ -253,7 +255,8 @@ That is what makes replay fast, and it is the whole risk. Two things follow:
253
255
  2. Run `bir replay allow <servers>` (§5.3). An allowlist of the servers you are comfortable having
254
256
  called unattended is cheap insurance, and it is the difference between "replay reads a dashboard"
255
257
  and "replay can do anything any wrapped server can do". The same goes for built-ins:
256
- `BIR_BUILTIN_TOOLS=Read,Glob,Grep` keeps writes and commands with the model, and
258
+ `BIR_BUILTIN_TOOLS=Read,Glob,Grep` keeps writes and commands with the model,
259
+ `BIR_PATH_ALLOW=project` keeps the file tools inside the project folder, and
257
260
  `BIR_BASH_ALLOW=git,npm test` (for example) limits the commands the recorder runs itself.
258
261
 
259
262
  Nothing else is approved on your behalf. Outside a matched scenario's own steps, a wrapped MCP call
@@ -368,8 +371,9 @@ so there is exactly one switch and it cannot get out of step with itself.
368
371
  | `BIR_STEP_TIMEOUT_MS` | `60000` | Per-step ceiling for one direct `tools/call`, clipped to what the plan has left |
369
372
  | `BIR_LOGIC_TIMEOUT_MS` | `1000` | How long one scenario logic body may run before it is stopped. At most 10 000 |
370
373
  | `BIR_REPLAY_RECORDED_OUTPUTS` | *(unset)* | `1` lets a step that could not be sent (a built-in off `BIR_BUILTIN_TOOLS`, no proxy, a server off the allowlist) stand on its recorded output instead of stopping the plan. Never for a timeout or an upstream error |
371
- | `BIR_BUILTIN_TOOLS` | `Read,Glob,Grep,Edit,Write,Bash` | Built-ins the runner executes itself during a replay, in the session's own project directory and confined to it. A chain whose only non-MCP steps are on this list arms `direct`. A runner-side step never reaches the host's permission prompts, so leave off any tool you want the model to call itself; `Read,Glob,Grep` is the read-only set (the default up to 0.3.7). `none`: every built-in step is steered in the session. Claude Code sessions only: in a LangChain application a tool with one of these names is the application's own, and the runner runs none (§6.5) |
374
+ | `BIR_BUILTIN_TOOLS` | `Read,Glob,Grep,Edit,Write,Bash` | Built-ins the runner executes itself during a replay, from the session's own project directory (which folders the file tools may touch: `BIR_PATH_ALLOW`). A chain whose only non-MCP steps are on this list arms `direct`. A runner-side step never reaches the host's permission prompts, so leave off any tool you want the model to call itself; `Read,Glob,Grep` is the read-only set (the default up to 0.3.7). `none`: every built-in step is steered in the session. Claude Code sessions only: in a LangChain application a tool with one of these names is the application's own, and the runner runs none (§6.5) |
372
375
  | `BIR_BASH_ALLOW` | *(unset: any command)* | Command prefixes a runner-side `Bash` step may run (`*` for all). Unset means any command; `none` means no command, and `Bash` steps are steered in the session. A plan's Bash steps share a working directory, exactly as the host's own Bash tool does — variables and functions do not persist, also as the host's. A call with `run_in_background`, or with its own `timeout` longer than the step's clock (`BIR_STEP_TIMEOUT_MS`), is never started on the runner: the plan hands over in front of it |
376
+ | `BIR_PATH_ALLOW` | *(unset: any folder, from 0.3.21)* | Folders a runner-side `Read`, `Glob`, `Grep`, `Edit` or `Write` may touch (`*` for all). Unset means any path, as the agent's own tools reach it — a plan that reads the agent's memory file under `~/.claude` runs. `project` is the session's project folder alone (the rule up to 0.3.20); add folders after a comma, `~` for the home folder: `project,~/.claude`. Under a list, a path outside it, symlinks included, is refused and the plan hands over in front of it. `none` means no folder, and the file steps are steered in the session. `Bash` is `BIR_BASH_ALLOW`'s |
373
377
  | `BIR_SEGMENT_ARM` | *(unset: on, from 0.3.19)* | A segment the service hands out runs mid-task: a recurring sub-task, or a half of a task after a plan's decision. `0` is observe-only: what would have armed is logged (`replay.segment_would_arm`) and nothing is replaced. Up to 0.3.18 observe-only was the default. A recorder started before an update keeps its old default until `bir up --restart` |
374
378
  | `BIR_INTENT_ARM` | *(unset)* | `1` lets a match found mid-turn arm a plan in the calling tool's place, which means waiting for it (`BIR_INTENT_ARM_BUDGET_MS`, 500); `0` never waits, and then nothing starts mid-task, segments included: what would have armed is logged; unset follows `BIR_SEGMENT_ARM`, so from 0.3.19 it is on unless `BIR_SEGMENT_ARM=0` |
375
379
  | `BIR_VERBOSE` | *(unset)* | `1` adds per-step `replay.*` detail lines |
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@basein/runner",
3
- "version": "0.3.20",
3
+ "version": "0.3.21",
4
4
  "description": "A recording MCP proxy: sits between any MCP client and its MCP servers, executes each call on the client's behalf, and records the run as a reusable BaseIn scenario.",
5
5
  "type": "module",
6
6
  "license": "MIT",
package/production.md CHANGED
@@ -1071,8 +1071,9 @@ Defaults were checked in `src/`.
1071
1071
  | `BIR_STEP_TIMEOUT_MS` | recorder | 60000 | One direct `tools/call`. |
1072
1072
  | `BIR_LOGIC_TIMEOUT_MS` | recorder | 1000 (capped at 10000) | Synchronous time limit for one scenario logic body. |
1073
1073
  | `BIR_REPLAY_RECORDED_OUTPUTS` | recorder | unset | `1`: a step that could not be sent may use its recorded output, labelled as such. |
1074
- | `BIR_BUILTIN_TOOLS` | recorder | `Read,Glob,Grep,Edit,Write,Bash` | Built-ins the runner executes itself during a replay, in the session's project directory and confined to it (up to 0.3.7 the default was `Read,Glob,Grep`). Runner-side steps bypass the host's permission prompts; leave a tool off to have the model call it. `none` turns it off. Not in a session of another host (§3.2): the runner runs no step itself there, whatever a tool is named. |
1074
+ | `BIR_BUILTIN_TOOLS` | recorder | `Read,Glob,Grep,Edit,Write,Bash` | Built-ins the runner executes itself during a replay, from the session's project directory; which folders the file tools may touch is `BIR_PATH_ALLOW` (up to 0.3.7 the default was `Read,Glob,Grep`). Runner-side steps bypass the host's permission prompts; leave a tool off to have the model call it. `none` turns it off. Not in a session of another host (§3.2): the runner runs no step itself there, whatever a tool is named. |
1075
1075
  | `BIR_BASH_ALLOW` | recorder | unset (any command) | Command prefixes a runner-side `Bash` step may run (`*` for all). Unset means any command; `none` means no command, and `Bash` steps are steered. A plan's Bash steps share a working directory, as the host's do; variables do not persist, also as the host's. |
1076
+ | `BIR_PATH_ALLOW` | recorder | unset (any folder) | Folders a runner-side `Read`, `Glob`, `Grep`, `Edit` or `Write` may touch (`*` for all). Unset means any path, as the agent's own tools (from 0.3.21; up to 0.3.20 the project directory alone). `project` is the session's project directory, `~` the home folder: `project,~/.claude`. Under a list, a path outside it, symlinks included, is `not_run` (hand-over). `none` means no folder, and file steps are steered. |
1076
1077
  | `BIR_DERIVE_RECENT_RESULTS` | recorder | 5 | How many of this prompt's tool results derivation may read (`0`: none). Never for plans with a command sink, and never for a plan started at the prompt: a prompt just submitted has no results yet, and an earlier prompt's are not its own. So in a session of another host, where a plan starts at the prompt only, none are read. |
1077
1078
  | `BIR_INTENT_MATCH` | recorder | on | `0` turns mid-turn intent probes off. |
1078
1079
  | `BIR_INTENT_MATCH_BUDGET_MS` | recorder | 4000 | One probe's own ceiling. It is not a hook wait. |