shiro-neko/docs/tools.md

Tools

The approval model

Every call resolves to allow, ask, or deny through a rule matched against the call's subject — the command for bash, the path for a file tool. Permissions is the full reference; the short version:

Allowed by default. Read-only tools and anything touching the agent's own state: read_file, read_many_files, glob, grep, list_dir, task, the whole git set, todo_write, remember, recall, forget, skill, ask, and anything a plugin marks auto-approved.

Denied by default. *.env, *.env.*, and *.pem on read. Not gated, refused: a secret that reaches the context is on the wire and in the session file, and there is no taking it back. *.env.example is allowed.

Asked by default. write_file, edit_file, multi_edit, apply_patch, move_file, delete_file, bash, bash_stop, web_fetch, and every mcp__* tool.

bash wants to run
git status --porcelain
y allow once | a always allow bash git * | n deny

a whitelists the pattern, not the tool: approving git status runs git log unprompted and still asks about npm publish. n tells the model it was denied and to ask what to do instead.

A rule turns the common cases off entirely:

{ "permission": { "bash": { "*": "ask", "git *": "allow", "bun test*": "allow" } } }

Three more things sit around the rules:

• The guard plugin refuses first. It is not an approval, and --yolo does not reach it. See plugins.
• A repeated call asks anyway. The same tool with identical input three times in one turn stops for approval even when allowed — a model repeating itself is not making progress.
• The SDK enforces the decision. A denied call provably never executes, because the SDK never reaches the tool's execute. A tool cannot forget to honour a denial. See architecture.

Tool sets

Each tool costs its name, its description, and its JSON schema on every request. The current registry has forty-one built-ins. /tools shows the live set; disabling an optional set removes its schemas from both the request and the system prompt.

Tool	Bytes	Tool	Bytes
read_many_files	972	git_blame	499
multi_edit	934	git_log	484
edit_file	618	git_diff	473
grep	595	bash	466
list_dir	594	git_show	432
read_file	526	git_status	292
glob	499	write_file	289

Selection accuracy also falls as the list grows: a model choosing between six tools picks better than one choosing between twenty.

Sets let you switch off what a project does not need:

Set	Tools	Cost
core	read_file read_many_files write_file edit_file glob grep bash bash_status bash_stop	~3,200 B
edit-plus	multi_edit list_dir apply_patch move_file delete_file	patch and file ops
nav	find_symbol json_query	navigation and structured reads
extra	20 tools: line edits, fs inspect, git extensions, code/env reads	on by default
git	git_status git_diff git_log git_show git_blame git_branch git_commit_message	~2,180 B + message
net	web_fetch, web_search	opt in

{ "toolSets": ["edit-plus"] }

Omit toolSets for the default sets. Add net when the agent should fetch public pages. core is always on — without read, edit, and bash the agent is not an agent. A disabled set reaches neither the wire nor the system prompt, since a prompt that names an absent tool teaches the model to attempt calls that cannot succeed. Session, plugin, and MCP tools are not part of this budget and are never gated here.

An unrecognised set name is dropped silently. The header line at startup shows which sets actually loaded, so a typo reads as "that set is off" rather than as an error — worth checking if a tool you expected is missing.

/tools shows which set each live tool came from:

tools
20 offered this turn of 22 registered
- `bash`             core
- `git_diff`         git
- `list_dir`         edit-plus
- `remember`

A tool with no set is a session, plugin, or MCP tool.

Which sets to keep

Both extra sets earn their place in most projects, but not all:

• No git in the repo? git is 2,180 bytes the model can never use. Switch it off.
• A model that handles many tools badly? { "toolSets": [] } trims to six, which is the smallest set that still lets the agent work.
• Reading a lot, editing rarely? Keep edit-plus for list_dir and read_many_files alone; they pay for themselves in round trips saved.

The extra set

Twenty-one tools across five families, on by default. Each follows the same rules as the core tools: writes are jailed to the workspace, reads honour .gitignore, and every git call spawns the binary with a fixed argument array, never a shell string.

Line edits

Precise edits by line number, for changes that need no full-file rewrite and no exact-string match. All refuse a path outside the workspace.

Tool	Does
insert_lines	Insert a block before a 1-based line, pushing the rest down. One past the end appends.
delete_lines	Delete an inclusive line range. Refuses the whole file — that is delete_file's job.
replace_lines	Replace an inclusive line range with new text in one write.
append_file	Add text to the end of a file.
prepend_file	Add text to the top of a file, e.g. a header or import block.
count_lines	Line count for one file, or per file across a glob. A size read before opening something large.

Filesystem

Tool	Does
tree	Indented directory tree, ignore-aware, directories first. A broad shape faster to scan than list_dir.
file_info	Size, line count, modified time, text-or-binary for one file.
find_files	Files whose name contains a substring (not a glob), e.g. auth.
recent_files	Files modified most recently, newest first. Find what a tool just touched.
changed_files	The working-tree delta git reports (modified, staged, untracked).

Git extensions (read-only)

Spawned with a fixed argv, so they are auto-approved like the core git tools.

Tool	Does
git_log_file	Commits that touched one file, newest first, with hash, date, subject.
git_diff_commits	Diff between two refs, optionally limited to one path.
git_show_file	A file's contents at a ref, e.g. auth.ts at HEAD~3.
git_current_branch	The current branch with its upstream and ahead/behind count.
git_changed_in_ref	Files changed between a ref and the working tree, names only.

Code and environment

Tool	Does
find_symbol	Where a function, class, or type is defined across JS/TS, Python, Go, Rust. Matches declarations, not uses.
json_query	One value from a JSON file by dotted path (scripts.build), instead of reading it whole.
outline	Top-level declarations of a source file as a structural map. Read before opening a large file.
read_symbol	The full body of one top-level definition by name.
env_info	Platform, shell, and which runtimes and package managers are installed, before writing a command.
count_tokens	Estimate the token cost of a file or string (~4 chars per token) before sending it to the model.
run_checks	Find the project's check commands (AGENTS.md first, then package.json scripts, then toolchain defaults) and run them with a timeout; pass/fail + first error.

Verification: the run_checks tool

"Verify before done" is a mechanism, not advice. run_checks finds the commands a project actually documents and runs them against a timer, so the model knows what passing means in this repo without guessing.

Discovery order (first match wins per command):

1. AGENTS.md — any line picking out a command (a backticked span, or a `# Tools

The approval model

Every call resolves to allow, ask, or deny through a rule matched against the call's subject — the command for bash, the path for a file tool. Permissions is the full reference; the short version:

Allowed by default. Read-only tools and anything touching the agent's own state: read_file, read_many_files, glob, grep, list_dir, task, the whole git set, todo_write, remember, recall, forget, skill, ask, and anything a plugin marks auto-approved.

Denied by default. *.env, *.env.*, and *.pem on read. Not gated, refused: a secret that reaches the context is on the wire and in the session file, and there is no taking it back. *.env.example is allowed.

Asked by default. write_file, edit_file, multi_edit, apply_patch, move_file, delete_file, bash, bash_stop, web_fetch, and every mcp__* tool.

bash wants to run
git status --porcelain
y allow once | a always allow bash git * | n deny

a whitelists the pattern, not the tool: approving git status runs git log unprompted and still asks about npm publish. n tells the model it was denied and to ask what to do instead.

A rule turns the common cases off entirely:

{ "permission": { "bash": { "*": "ask", "git *": "allow", "bun test*": "allow" } } }

Three more things sit around the rules:

• The guard plugin refuses first. It is not an approval, and --yolo does not reach it. See plugins.
• A repeated call asks anyway. The same tool with identical input three times in one turn stops for approval even when allowed — a model repeating itself is not making progress.
• The SDK enforces the decision. A denied call provably never executes, because the SDK never reaches the tool's execute. A tool cannot forget to honour a denial. See architecture.

Tool sets

Each tool costs its name, its description, and its JSON schema on every request. The current registry has forty-one built-ins. /tools shows the live set; disabling an optional set removes its schemas from both the request and the system prompt.

Tool	Bytes	Tool	Bytes
read_many_files	972	git_blame	499
multi_edit	934	git_log	484
edit_file	618	git_diff	473
grep	595	bash	466
list_dir	594	git_show	432
read_file	526	git_status	292
glob	499	write_file	289

Selection accuracy also falls as the list grows: a model choosing between six tools picks better than one choosing between twenty.

Sets let you switch off what a project does not need:

Set	Tools	Cost
core	read_file read_many_files write_file edit_file glob grep bash bash_status bash_stop	~3,200 B
edit-plus	multi_edit list_dir apply_patch move_file delete_file	patch and file ops
nav	find_symbol json_query	navigation and structured reads
extra	20 tools: line edits, fs inspect, git extensions, code/env reads	on by default
git	git_status git_diff git_log git_show git_blame git_branch git_commit_message	~2,180 B + message
net	web_fetch, web_search	opt in

{ "toolSets": ["edit-plus"] }

Omit toolSets for the default sets. Add net when the agent should fetch public pages. core is always on — without read, edit, and bash the agent is not an agent. A disabled set reaches neither the wire nor the system prompt, since a prompt that names an absent tool teaches the model to attempt calls that cannot succeed. Session, plugin, and MCP tools are not part of this budget and are never gated here.

An unrecognised set name is dropped silently. The header line at startup shows which sets actually loaded, so a typo reads as "that set is off" rather than as an error — worth checking if a tool you expected is missing.

/tools shows which set each live tool came from:

tools
20 offered this turn of 22 registered
- `bash`             core
- `git_diff`         git
- `list_dir`         edit-plus
- `remember`

A tool with no set is a session, plugin, or MCP tool.

Which sets to keep

Both extra sets earn their place in most projects, but not all:

• No git in the repo? git is 2,180 bytes the model can never use. Switch it off.
• A model that handles many tools badly? { "toolSets": [] } trims to six, which is the smallest set that still lets the agent work.
• Reading a lot, editing rarely? Keep edit-plus for list_dir and read_many_files alone; they pay for themselves in round trips saved.

The extra set

Twenty-one tools across five families, on by default. Each follows the same rules as the core tools: writes are jailed to the workspace, reads honour .gitignore, and every git call spawns the binary with a fixed argument array, never a shell string.

Line edits

Precise edits by line number, for changes that need no full-file rewrite and no exact-string match. All refuse a path outside the workspace.

Tool	Does
insert_lines	Insert a block before a 1-based line, pushing the rest down. One past the end appends.
delete_lines	Delete an inclusive line range. Refuses the whole file — that is delete_file's job.
replace_lines	Replace an inclusive line range with new text in one write.
append_file	Add text to the end of a file.
prepend_file	Add text to the top of a file, e.g. a header or import block.
count_lines	Line count for one file, or per file across a glob. A size read before opening something large.

Filesystem

Tool	Does
tree	Indented directory tree, ignore-aware, directories first. A broad shape faster to scan than list_dir.
file_info	Size, line count, modified time, text-or-binary for one file.
find_files	Files whose name contains a substring (not a glob), e.g. auth.
recent_files	Files modified most recently, newest first. Find what a tool just touched.
changed_files	The working-tree delta git reports (modified, staged, untracked).

Git extensions (read-only)

Spawned with a fixed argv, so they are auto-approved like the core git tools.

Tool	Does
git_log_file	Commits that touched one file, newest first, with hash, date, subject.
git_diff_commits	Diff between two refs, optionally limited to one path.
git_show_file	A file's contents at a ref, e.g. auth.ts at HEAD~3.
git_current_branch	The current branch with its upstream and ahead/behind count.
git_changed_in_ref	Files changed between a ref and the working tree, names only.

Code and environment

Tool	Does
find_symbol	Where a function, class, or type is defined across JS/TS, Python, Go, Rust. Matches declarations, not uses.
json_query	One value from a JSON file by dotted path (scripts.build), instead of reading it whole.
outline	Top-level declarations of a source file as a structural map. Read before opening a large file.
read_symbol	The full body of one top-level definition by name.
env_info	Platform, shell, and which runtimes and package managers are installed, before writing a command.
-prefixed block).
This is the strongest source: it is written by the people who know what a cold agent
should run.
2. package.json scripts — test, typecheck, check, lint, build in that priority,
then the rest alphabetically. The runner matches the lockfile: bun run when bun.lock
exists, npm run otherwise.
3. Toolchain defaults — what the project's own build system says verify means: bun test
for a Bun project, cargo test for Rust, go test ./... for Go, pytest for Python.

The target argument selects one suggestion by name (or all). Output is capped, runs are killed at the timeout (SIGTERM reported, not a clean exit), and a 3-iteration fix ceiling is baked into the system prompt's verify loop — a failing check gets fixed and re-run, but a check that keeps failing stops grinding and reports instead.

Each candidate is gated by permission rules on its target name, the only part the model chooses.

File tools

read_file

path    file path relative to the workspace root
offset  first line, 1-based
limit   max lines, default 2000

Returns contents with 1-based line numbers. Refuses binaries: a NUL byte in the first 8 KB means the file is not text, and a model that reads a 90 MB executable has burned its whole context on nothing.

read_many_files

files  [{ path, offset?, limit? }], at most 20

One round trip for several files, each with its own window. Reads run concurrently and the blocks come back in the order given, labelled:

===== src/app.ts =====
1: export const port = 8080;

===== src/gone.ts =====
[unreadable: No such file: src/gone.ts]

A path that cannot be read is reported in its own block rather than throwing, so one wrong guess costs a line instead of the whole call. Numbering and binary refusal are the same code path as read_file, so a batch read cannot drift from a single one.

write_file

path     file path
content  full contents

New files and full rewrites only. Creates parent directories.

A rewrite that collapses whitespace is flagged in the result: similar character count, a fraction of the lines. A model writing a large file under output pressure squeezes newlines and indentation before it cuts markup — the bytes survive, the layout does not — so the result names the collapse and the turn fixes it in place:

Wrote 139 chars to index.blade.php, but it collapsed 9 lines into 1. If that was not
intended, re-send the content with its original newlines and indentation.

edit_file

path        file path
oldString   exact text to find, whitespace and indentation included
newString   replacement
replaceAll  replace every occurrence instead of requiring exactly one

oldString must match byte-for-byte and appear exactly once unless replaceAll is set. An ambiguous match is an error naming the count, which pushes the model to add surrounding context rather than guessing which occurrence it meant:

oldString appears 3 times in src/users.ts. Add surrounding context or set replaceAll.

That error is deliberately specific. edit failed would leave the model to retry blind; the count tells it what to do next.

multi_edit

path   file path
edits  [{ oldString, newString, replaceAll? }], in the order to apply them

Several edits to one file in one call, one approval, one write. Each edit sees the result of the previous one, so edits may build on each other.

Atomic: every edit is validated and applied in memory first, so a failure on the third edit leaves the file exactly as it was rather than half-changed. The same uniqueness rule as edit_file applies per edit, and the error names which edit failed:

edit 2: oldString not found in src/users.ts. No edits were applied.

The last sentence matters. Without it a model reading the error has to guess whether edit 1 landed, and its next move — retry the whole batch, or only what failed — depends on the answer.

apply_patch

patch  one envelope containing Add, Update, Move, and Delete file markers

All operations are validated before anything is written, so a failure leaves every file unchanged. Use it when one change spans files that must land together; use multi_edit for several edits to one file and edit_file for one edit. Paths stay inside the workspace and the call asks for approval.

move_file

from  existing file path
to    new path, including the filename

Renames or relocates one file, creating the target directory. Refuses a missing source and an occupied target, so a rename cannot silently overwrite work. Permission rules match both ends, so denying src/generated/* catches a move that lands there as well as one that starts there.

For a rename plus its callers in one atomic step, apply_patch is the better tool: it lands the move and the edits together or not at all.

delete_file

path  file to delete

Deletes one file and reports its size. A directory is refused: removing a tree is exactly what the guard plugin blocks in bash, and it is not something to do implicitly through a tool whose name says "file". Delete the files you mean, one call each.

list_dir

path            directory, relative to the workspace root, default the root
depth           levels to descend, 1-6, default 2
includeIgnored  also show files git ignores

Tree view honouring .gitignore. Directories end with /, files show their size:

.
README.md  2B
src/
  app.ts  2K
  ui/

Past the depth limit the containing directory is still listed, so the shape of the tree stays visible without its contents — src/ui/ above appears at depth: 2 even though its files do not. Capped at 300 entries.

glob

pattern         e.g. "src/**/*.ts"
limit           max paths, default 200
includeIgnored  also return files git ignores

Walks the tree honouring .gitignore and .shiroignore, skipping .git and node_modules unconditionally. Nested ignore files apply only within their own directory, as git does. Returns posix paths relative to the workspace root.

A symlinked directory is classified as a directory and not descended into. Both halves matter: readdir reports a junction as a non-directory, so without the extra stat a symlinked directory leaked past dir/ ignore rules and was yielded as a file with a nonsense size. Not descending is separate — a link can point anywhere, including back into the tree.

grep

pattern         regex source
include         glob limiting the search, default "**/*"
ignoreCase      case-insensitive
includeIgnored  also search files git ignores

Shells out to ripgrep when it is on PATH — roughly 15x faster on a real repo — and falls back to a JavaScript walker otherwise. Output is path:line: text either way, so the model sees one format regardless. Skips binaries. Caps at 200 hits.

Two details keep the two paths in agreement. ripgrep is passed --no-require-git, because it otherwise ignores .gitignore outside a repository while the JavaScript fallback always honours it. And an rg exit code above 1 means rg could not run the search at all, so the fallback takes over; exit 1 is simply "no matches" and is reported as such.

Regex syntax differs between the two: ripgrep is Rust regex, the fallback is JavaScript. A pattern using look-around works in the fallback and fails under rg. An invalid pattern is reported as Invalid regex: <reason> rather than returning an empty result set.

bash

command     shell command
timeout     ms, default 120000, max 600000 (ignored when background is true)
background  detach the command and return immediately with a handle
name        label for a background command, shown in /bash and bash_status

Runs in the workspace root through bash -lc or cmd /c. Output streams live to the panel above the input rather than appearing all at once when the command exits — a two-minute test run is otherwise indistinguishable from a hang. Both pipes are drained concurrently, since a command that fills one while you block on the other deadlocks.

Returns exit code, stdout, stderr, and a note if a signal killed it.

ctrl-c interrupts the command, not the turn. The shell and everything it started are killed — on Windows through taskkill /T, because killing cmd alone leaves the real command holding both pipes open and the read never ends. The call then fails rather than returning, so the model cannot mistake a killed command for one that ran and failed on its own:

The user interrupted this command. It did not finish, so its effects are unknown.
stdout:
[whatever it printed first]

The turn continues from there. esc still aborts everything, and ctrl-c with nothing running quits as usual. When a background command is running and nothing foreground is in flight, ctrl-c stops the most recently started one instead of quitting — an accidental quit must not kill a dev server the user still wants.

bash background mode (dev servers, watchers)

A command that does not exit — bun dev, a watcher, a test suite that never returns — blocks the tool until its timeout, which looks like the agent is stuck. Set background: true instead:

command:  bun dev
background: true
name:  dev server

The tool returns immediately with a handle:

background 1: running (pid 4821) — poll with bash_status handle=1, stop with bash_stop handle=1
command: bun dev

The command runs detached (its own process group), so it keeps running while the agent works, ctrl-c in the agent does not signal it, and the model can poll it and keep going:

• bash_status handle — whether it is still running, its exit code when finished, and any output produced since the last status. Poll this while working.
• bash_stop handle — kill it, awaited so the process really is gone.

$ bash_status  handle: 1
handle 1: dev server
status: running
new output:
  VITE ready in 312 ms

Background commands are reaped when the agent exits (killed and removed), and anything left over from a crashed session is killed at the next boot, so a dev server an agent started cannot linger unnoticed. /bash lists what is running, /bash stop <id> and /bash stop all stop them. Like any bash, they never go through the file-snapshot/undo system.

web_fetch

url       absolute HTTP(S) URL
maxChars  returned characters, default 30,000, max 30,000

Fetches a public text page and converts HTML to markdown. HTTPS is required for public hosts; private and loopback addresses are refused, redirects are checked one hop at a time, and the body is capped. The result is untrusted page content, not an instruction, and the call asks for approval. It belongs to the opt-in net set.

web_search

query     what to search for

Searches the web (DuckDuckGo Lite — no API key) and returns up to five results with title, URL, and snippet. Results are re-checked against the same private-address rule as web_fetch, so a result cannot point the model at localhost or a cloud metadata endpoint. Same net set, same approval, same "untrusted text" framing: search results are a stranger's claims, and a web_fetch on one of them is the right follow-up.

Git tools

All five are read-only and therefore approval-free. Each spawns git with a fixed argument array rather than a shell string, so an argument like --author="; rm -rf /" can only ever be a literal argument — which is what makes auto-approval safe. A test asserts exactly that: git_log with the path ; touch pwned.txt creates no file.

Output is described rather than raw porcelain. git_status names the branch and says staged modified or untracked per file instead of leaving the model to decode porcelain's two leading columns:

On main, 2 changed:
src/app.ts  (staged modified, modified)
new.ts  (untracked)

That file has a staged change and a later unstaged one, which the raw MM prefix conveys only to a reader who knows the format.

Outside a repository they fail with <cwd> is not a git repository. rather than passing git's own error text through. Other git failures do pass through, on purpose: git_show no-such-ref reports what git said, because git's own message is the most useful thing available.

git_status                                  branch, staged, modified, untracked
git_diff    staged?  path?                  unified diff of uncommitted changes
git_log     limit?   path?                  hash, date, author, subject; newest first
git_show    ref      path?                  one commit: message, author, diff
git_blame   path     startLine?  endLine?   who last changed each line
git_branch  remote?                         branches, newest commit first, current marked

git_commit_message

Generates one commit message from the staged changes. The nested model call sees two things: the staged diff, and the fifteen most recent commit subjects, because a message that ignores the repository's established style reads as foreign however accurate it is. An oversized diff is truncated before it reaches the model.

It never commits — it returns the message only, approval-free, because generating text cannot mutate anything. Running the commit stays on the gated bash path, where the user sees the message and the command together.

$ git_commit_message
bump the server port to 9090

Nothing staged is a stated error rather than an empty message, so the model's next move is to stage, not to guess.

git_log defaults to 15 commits and caps at 40. git_blame without a range blames the whole file; with startLine and no endLine it covers 40 lines from there. git_branch sorts by last commit and marks the current branch with *, which is what makes an already-taken branch name obvious before proposing one.

Everything here is also reachable through bash. The reason the set exists anyway is the approval boundary: bash git diff stops for a decision on every call, while git_diff cannot mutate anything and so never needs one.

Agent tools

task

description  short label shown to you
prompt       self-contained instructions
kind         "explore" (default), "review", or "worker"

Spawns a subagent with its own context window. It returns one report, so the parent pays for the findings rather than the whole search transcript, and it sees none of the parent conversation, so its prompt has to stand alone.

explore finds and reports; review critiques code in severity order. Both are structurally read-only — they hold no gated tool at all, so they cannot trigger an approval prompt whatever the config says. The parent's context holds the conclusion instead of the search: a subagent reading forty files to answer one question costs the parent the answer, not the forty files.

worker holds the write tools as well, and every write and command routes through the parent's approval gate — the same rules and the same prompt as a direct call. The full delegation trade-offs are in agents.

Progress streams to the subagent panel, with each call's outcome. Capped at 20 steps, and it shares the parent's model — an explore run pays reasoning rates for what is really a search, which is on the list to fix.

Not worth delegating a single grep: the subagent is a whole extra model loop, so it wins on a search spanning many files and loses on anything you could answer in one call.

ask

question  one specific question
options   choices, recommendation first, each with an optional detail
multiple  allow more than one

Stops the turn and puts the question on screen. With options it is a picker; without, free text. esc skips, which returns "the user dismissed this; use your best judgement" — so a dismissal is an instruction to decide, not a dead end.

Withheld entirely in headless mode — a question with no one to answer it would hang. The system prompt says so, and tells the model to decide and state its assumption instead.

todo_write

todos  the complete list: content, status, optional note

Statuses: pending, in_progress, done, blocked. Send the whole list each time; it replaces the previous one. Warns when more than one task is in_progress, when nothing is in_progress while work remains, or when a blocked task has no note.

The list lives in the system prompt, rebuilt every step, so it survives both pruning and /compact. See memory.

remember, recall, forget

Durable per-project notes. See memory.

skill

name  skill name from the catalogue

Loads the body of a skill. See skills.

Path safety

Every path a tool receives goes through a jail: resolved against the workspace root, then checked that it did not escape.

export function jail(p: string, root = process.cwd()): string {
  const abs = isAbsolute(p) ? resolve(p) : resolve(root, p);
  const rel = relative(resolve(root), abs);
  if (rel.startsWith('..') || isAbsolute(rel)) throw new Error(`Path escapes workspace: ${p}`);
  return abs;
}

../../etc/passwd, a/../../secret, and absolute paths outside the root are all refused before any filesystem call. Resolving first and comparing after is what catches the middle case: string-prefix checks on the raw input miss a/../../secret entirely.

The model's output is a trust boundary. It can emit any string, so the check happens on every call rather than being assumed. That includes read_many_files, where a bad path is reported in its block like any other unreadable file.

jail guards the workspace, not the shell. bash runs whatever it is given, which is why every call needs approval and why the guard plugin exists — see plugins.

Output caps

Any single tool result is truncated at 30,000 characters with a note saying how much was cut:

... [truncated 41,233 chars]

Tool	Cap
any result	30,000 characters
grep	200 hits, each line cut at 300 chars
glob	200 paths
list_dir	300 entries
read_many_files	20 files
read_file	2,000 lines by default
bash	120 s default timeout, 600 s max (background mode has no timeout)

Without caps one grep for function can end a session. The caps are per call, so a model that needs more can narrow and ask again — which is cheaper than one call that fills the context and forces compaction.