Changelog Archive - July-August 2026

sys.stdout and sys.stderr. Scripts had no way to write to stderr, so warnings and errors shared stdout with the report. Both streams are now available with write(), writelines(), flush(), isatty(), and with support, and print(..., file=...) now accepts any object with a write method, sys.stderr, sys.stdout, or your own class. The streams follow output capture, host writers, and sandbox discarding just like print() does, and Go hosts get a matching SetErrorWriter(io.Writer).

import sys

sys.stderr.write("warning: retrying (attempt 2)\n")
print("fatal: disk full", file=sys.stderr)
print('{"status": "ok"}')   # stdout stays report-only

ZeroDivisionError is now catchable. Division by zero raised an error that no except clause could name, only a bare except Exception caught it, even though ZeroDivisionError already existed as a builtin. /, //, and % now raise a ZeroDivisionError that except ZeroDivisionError: matches, for both integer and float operands. Uncaught behaviour is unchanged (same message, file, and line), and except Exception still catches it, so existing scripts are unaffected.

try:
    x = 1 / 0
except ZeroDivisionError as e:
    print("caught:", e)

Iterating a dict directly now yields its keys, matching Python. for k in d, comprehensions ([k for k in d], {k for k in d}), and *d unpacking all raised type error: expected iterable, got DICT; they now iterate the keys like CPython. list(d) and tuple(d) already worked. The view methods d.keys() / d.values() / d.items() are unchanged, and dict_keys still has no .sort() (use sorted(d.keys())), also matching CPython.

d = {"b": 2, "a": 1}
for k in d:          # "b", then "a"
    print(k, d[k])

Substantially faster evaluation, with far fewer allocations. Four changes to the evaluator, all measured against v0.19.0:

• Integer arithmetic is no longer boxed per operator. An expression such as total + i * 2 - 1 allocated an integer object for every operator, even though only the final value is ever observed. Side-effect-free integer arithmetic and comparisons now evaluate without boxing intermediates.
• Attribute reads no longer allocate. obj.attr is indexing by a string literal internally, so every field read was allocating its own copy of the field name. String literals now reuse one immutable value.
• Builtin calls no longer allocate a context. Each call built a throwaway context to carry the environment; it is now derived once per evaluation.
• a + b + c chains no longer allocate scratch slices. The chain path collected its operands into two slices before discovering whether they were strings, so numeric and list chains paid for a string path they never used. Operands are now folded as they are evaluated.

Measured on operation-dense workloads (loops doing real work inside a single evaluation), these average roughly a 30% reduction in run time, with per-operation allocations down 40–100% — integer chains and loops, which previously boxed every operator, now allocate almost nothing. The gains are largest for scripts that loop over arithmetic, attribute access, or string/list building. Short one-shot Eval calls see smaller gains because the per-call interpreter setup dominates them; parse and compile times are unchanged.

bytes type — a dedicated binary data type mirroring Python’s bytes: an immutable sequence of byte values (0–255) with the usual operator support (len(), indexing, slicing, +/*, comparison, in, iteration, truthiness). Mixing bytes with str in concatenation or comparison raises a TypeError — call .decode() to convert. It is a global builtin with the Python-compatible constructor bytes(source, encoding="utf-8"), plus bytes.fromhex() and bytes.frombase64() static constructors. Methods on values: .decode(), .hex(), .base64(), .length().

b = bytes("hi")          # b'hi'
b = bytes([104, 105])    # b'hi'
assert b.decode() == "hi"
assert b.hex() == "6869"
assert b + bytes("!") == bytes("hi!")

msgpack library — MessagePack binary serialisation, the compact counterpart to json. Mirrors Python’s msgpack module: packb(obj) returns bytes, unpackb(packed) accepts bytes, with pack/unpack aliases. bytes round-trips as msgpack bin; str as msgpack str. The default codec is shamaton-msgpack, also used by gossip.DefaultConfig(), so a fresh instance and cluster agree on the wire format without explicit wiring.

import msgpack

payload = msgpack.packb({"user": "alice", "id": 42})
data = msgpack.unpackb(payload)
print(data["user"])  # "alice"

Binary interop across file I/O, HTTP, and sockets — bytes is now a first-class participant in every I/O surface, mirroring Python 3’s split between str and bytes:

• Files — new os.read_bytes(path) returns bytes; os.read_lines(path) iterates lines lazily for large files; os.write_file and os.append_file accept bytes.
• HTTP — responses gain a .content field (bytes), matching Python’s requests; the request data= accepts bytes for binary bodies. .text and .body remain str.
• Sockets — scriptling.net.unicast and scriptling.net.multicast send() accepts bytes; receive() returns data as bytes (see breaking change above).
• Packages — new scriptling.package.read_bytes(name, path).

import os, msgpack

# Pack to bytes, persist, read back, unpack — zero corruption.
os.write_file("/tmp/data.msgpack", msgpack.packb({"k": "v"}))
data = msgpack.unpackb(os.read_bytes("/tmp/data.msgpack"))

scriptling.package library — read-only access to files inside loaded packages (app bundles and library bundles). Works identically in directory and zip mode. Every function takes the package name (from manifest.toml) as its first argument.

import scriptling.package as package

spec = package.read_file("myapp", "data/spec.md")
files = package.list("myapp", "tools/")
py = package.glob("myapp", "**/*.py")

Functions: read_file, file_exists, list, glob, exists, version, names. Package name uniqueness is enforced at load time.

additional_files manifest field — ship extra files or directories in a package zip. A trailing / includes the entire directory tree; a bare path includes a single file.

additional_files = ["data/", "LICENSE"]

Decorator-based HTTP and JSON-RPC handler registration — attach routes directly to handler functions with @http.get("/path"), @http.post(...), @http.route(...), @http.websocket(...), @http.middleware, @http.not_found, and @jsonrpc.method("name"), @jsonrpc.notification("name"), plus @plugin.register_function and @plugin.register_class for plugin servers. No separate setup-script registration block needed; the route lives right above the function.

import scriptling.runtime.http as http

@http.get("/health")
def health(request):
    return http.json(200, {"status": "ok"})

@http.post("/api/users")
def create_user(request):
    return http.json(201, {"name": request.json()["name"]})

The module name is auto-resolved from __name__ (falling back to __file__ for the setup script). The imperative API (runtime.http.get("/path", "lib.func")) continues to work; both forms coexist.

App bundles — ship MCP tools, HTTP routes, JSON-RPC methods and static assets in a single package (folder or zip). A manifest.toml with a serve field declares the app’s protocols; the CLI provides only the transport. Dev folders and production zips run the same code path.

name    = "myapp"
version = "1.0.0"
main    = "setup.py"
libs    = ["lib", "vendor"]
serve   = ["http", "mcp"]

scriptling --server :8000 --package ./myapp       # dev (folder)
scriptling --server :8000 --package myapp.zip     # prod (zip)

See examples/app-bundle/ for a complete working example.

Decorator-based MCP tool registration — define tools with @mcp.tool() in a single .py file, no .toml sidecar needed. Parameters, types and descriptions live alongside the implementation. Multiple tools per file are supported.

import scriptling.runtime.mcp as mcp

@mcp.tool("Calculate an expression", params={"expr": "Math expression"})
def calc(expr):
    return f"{expr} = {eval(expr)}"

The legacy .toml + .py format continues to work; both formats can coexist in the same tools/ folder.

pack build is manifest-driven — inclusion follows the manifest (libs dirs, main script, convention dirs). Unknown top-level entries produce warnings; missing declared entries are build errors.

Tool scripts no longer get implicit sibling imports — pass -L ./tools if sibling imports are needed.

HTTP ServeMux conflict between /mcp and GET / routes resolved by registering MCP/JSON-RPC endpoints with explicit methods.

Data race on reloadMCP — signal-triggered reload and file-watcher debounce reload are now serialized with a mutex, preventing concurrent mutation of MCP registration maps and entry-tracking slices.

Setup script panic recovery — the setup script goroutine now recovers panics, converting them to a clean startup error instead of blocking NewServer forever or crashing the process.

HTTP handler request cancellation — runHandler now passes r.Context() to the script evaluator so client disconnects cancel in-flight handler scripts.

MCP handler socket consistency — MCP tool/resource/prompt handlers now respect --docker-host / --podman-host overrides (previously hardcoded to default socket paths).

Wasted KV store allocation — ResetRuntime no longer opens a throwaway in-memory KV store that InitKVStore immediately replaces.

pack build zip close error — file close failures during pack build are now propagated instead of silently swallowed.

os.symlink and os.path.islink: create and detect symbolic links.

os.symlink(src, dst) creates a symbolic link at dst pointing to src. os.path.islink(path) returns True if the path is a symbolic link (using Lstat so the link itself is checked, not the target). Both respect the allowed-paths security configuration.

import os
import os.path

os.symlink("../target.js", "node_modules/.bin/tool")
os.path.islink("node_modules/.bin/tool")  # True

find.entries: new include_symlinks option and link_target field.

When include_symlinks=True is passed to find.entries, symlink entries are yielded as-is (not followed) with their target in the link_target field. Use this to mirror symlink trees without resolving them.

import scriptling.find as find

for e in find.entries("/site", include_symlinks=True):
    if e["link_target"]:
        print(e["path"], "->", e["link_target"])

find.entries: new include_hash option and hash field.

When include_hash=True is passed to find.entries, each file entry’s content is crc64-hashed and the result is returned in the hash field. Use this for definitive change detection — two files with the same hash have identical bytes.

import scriptling.find as find

for e in find.entries("/site", include_hash=True, type="file"):
    print(e["path"], e["hash"])

scriptling.find.entries: matching files and directories with size, mtime, and type.

find.path returns matching paths as a list of strings — convenient for most uses, but callers that need to compare trees (differential sync, change detection, build artefact caching) previously had to stat every match themselves. The new find.entries function applies the same filters as find.path and returns a list[dict] with path, size, mtime, and is_dir per entry, stat’ing each match inside the existing worker pool so the metadata arrives in a single pass.

import scriptling.find as find

# Build a {path: mtime} index for differential sync
mtimes = {e["path"]: e["mtime"]
          for e in find.entries("/site", type="file")}

# Walk the matches with their size readily available
for e in find.entries("/logs", name="*.log", size_min=10 * 1024 * 1024):
    print(e["path"], e["size"])

find.path is unchanged; use it when only the strings are needed, since it skips the per-entry stat in the no-filter common case. The Go API gains a matching extlibs.FindEntries returning []extlibs.FindEntry.

sorted(), sum(), min(), max(), and str.join() now accept any iterable.

These builtins previously accepted only lists and tuples (sum/min/max additionally took flat numeric arrays); anything else raised type error: expected LIST or TUPLE. They now iterate any iterable — dict views (dict_keys/dict_values/dict_items), sets, strings, dicts, and iterators — matching Python’s semantics. str.join likewise accepts any iterable of strings, not just lists and tuples.

d = {"a": 3, "c": 1, "b": 2}

sorted(d.items(), key=lambda x: x[1])   # [(c, 1), (b, 2), (a, 3)]
sum(d.values())                         # 6
min(d.keys())                           # "a"
max(set([5, 2, 8]))                     # 8
sorted("cab")                           # [a, b, c]
"-".join(("x", "y", "z"))               # "x-y-z"

The numeric FloatArray fast paths are preserved, and sorted() does not mutate its input.

Set algebra operators &, |, -, ^ on sets.

Sets already supported set algebra via methods (.intersection(), .union(), .difference(), .symmetric_difference()); they now support the operator syntax too, matching Python. Both operands must be sets — use the methods for arbitrary iterables.

a = set([1, 2, 3])
b = set([2, 3, 4])

a & b   # {2, 3}   intersection
a | b   # {1, 2, 3, 4}   union
a - b   # {1}   difference
a ^ b   # {1, 4}   symmetric difference

Integer bitwise operations are unchanged — the new behaviour only applies when the left operand is a set. The augmented-assignment forms (&=, |=, -=, ^=) work on sets too, rebinding the name to the resulting set.

Set value equality (== / !=) now compares contents.

Two sets with the same elements previously compared equal only if they were the same object; distinct sets always reported as unequal. == and != now compare contents order-independently, so set([1, 2, 3]) == set([3, 2, 1]) is true. Cross-type equality (set == list) returns false rather than erroring.

Empty tuples, sets, and dict views are now falsy.

Truthiness checks (if, and, or, bool()) previously treated any value without an explicit rule as truthy, so empty (), set(), dict_keys(), dict_values(), and dict_items() were all truthy — unlike Python, where empty containers are falsy. They now follow Python’s rule (empty → falsy, non-empty → truthy), so set() and f() short-circuits and if my_set: works as expected. Lists, dicts, strings, and numbers were already correct.

scriptling.template.html / scriptling.template.text: custom action delimiters.

Set() now accepts optional left and right keyword arguments to override Go’s default {{ }} template markers. This lets you render templates that contain literal {{ }} — for client-side frameworks (Vue, Handlebars), CSS, JSON, or upstream config files that already use those characters.

import scriptling.template.text as text

# Use {% %} so {{ }} survives into the output untouched
tmpl = text.Set(left="{%", right="%}")
tmpl.add("Hello, {%.Name%}! Upstream var: {{ service.tag }}")
print(tmpl.render({"Name": "Alice"}))
# Output: Hello, Alice! Upstream var: {{ service.tag }}

Both arguments default to the standard delimiters ({{ and }}); pass an empty string to fall back to the default for that side.

scriptling.mcp.Client: environment variables for stdio servers.

mcp.Client() now accepts an env keyword argument (stdio servers only) — a list of KEY=value strings applied to the launched subprocess. Variables are merged on top of the inherited environment, so PATH, HOME, and other defaults remain available. Passing env with an HTTP URL raises an error.

import scriptling.mcp as mcp

client = mcp.Client("npx",
    args=["-y", "@modelcontextprotocol/server-filesystem", "/data"],
    env=["FS_ROOT=/data", "LOG_LEVEL=debug"],
    namespace="fs",
)

sorted() and list.sort() now order tuples and lists lexicographically.

Previously, sorting a list of tuples or lists was a silent no-op (list.sort()) or raised unsupported type for sorting: TUPLE (sorted()), because the comparator only handled numbers and strings. Tuples and lists now compare element-by-element (Python-style), so nested structures sort correctly too.

sorted([(3, "c"), (1, "a"), (2, "b")])  # [(1, "a"), (2, "b"), (3, "c")]

rows = [(1, 9), (1, 3), (1, 7)]
rows.sort()                              # in-place: [(1, 3), (1, 7), (1, 9)]

scriptling.nomad: dynamic host volume support.

The nomad library now covers Nomad’s dynamic host volumes (introduced in Nomad 1.8+):

• host_volumes_list — list dynamic host volumes, with optional filters for namespace, node_id, node_pool, and plugin_id.
• host_volume_get — fetch full details for a single host volume.
• host_volume_register — register a pre-existing host volume (e.g. a pre-mounted CephFS path) with Nomad.
• host_volume_create — provision storage via a host volume plugin and register it with Nomad.
• host_volume_delete — destroy backing storage via the plugin and deregister from Nomad.

import scriptling.nomad as nomad

c = nomad.Client("https://nomad.example.com:4646", token="secret")

# List all host volumes
for v in c.host_volumes_list():
    print(v["name"], v["node_id"], v["state"])

# Create a new host volume via plugin
c.host_volume_create("vol-new-01", {
    "Name": "app-data",
    "PluginID": "mkdir",
    "NodePool": "production",
    "RequestedCapacityMinBytes": 50 * 1024 * 1024 * 1024,
    "RequestedCapabilities": [{"AccessMode": "single-node-writer", "AttachmentMode": "file-system"}],
})

The nomad library now covers the complete CSI volume lifecycle:

• csi_volume_create — provisions new backing storage via the CSI controller plugin (e.g. creates a Ceph RBD image) and registers it in Nomad.
• csi_volume_delete — destroys backing storage via the CSI controller plugin and deregisters the volume from Nomad.

These complement the existing csi_volume_register (register pre-existing storage) and csi_volume_deregister (un-track without destroying data).

# Create a new volume with backing storage
c.csi_volume_create("qaprod-data-01", {
    "Name": "qaprod-data-01",
    "PluginID": "ceph-csi",
    "RequestedCapacityMin": 10 * 1024 * 1024 * 1024,
    "RequestedCapabilities": [{"AccessMode": "single-node-writer", "AttachmentMode": "file-system"}],
}, namespace="fortixqa")

# Delete volume and its Ceph backing store
c.csi_volume_delete("qaprod-orphaned-01", namespace="fortixqa")

See scriptling.nomad for the full method reference.

New scriptling.nomad library: manage HashiCorp Nomad CSI volumes and jobs.

A NomadClient obtained from nomad.Client(addr, token=...) talks directly to the Nomad HTTP API, covering CSI storage volumes and jobs. Requests default to a 10 second timeout, adjustable via timeout= on Client():

• CSI volumes: csi_volumes_list, csi_volume_get, csi_volume_register, csi_volume_deregister.
• Jobs: jobs_list, job_get, job_register, job_stop, wait_job_stopped, job_validate, job_plan, jobs_parse (HCL to JSON).

This is aimed at cluster housekeeping tasks, such as reconciling CSI volumes against a source of truth and removing orphaned ones after confirmation.

import scriptling.nomad as nomad

c = nomad.Client("https://nomad.example.com:4646", token="secret")

expected = set(open("expected_volumes.txt").read().split())
for v in c.csi_volumes_list(namespace="*"):
    if v["id"].startswith("qaprod") and v["id"] not in expected:
        c.csi_volume_deregister(v["id"], force=True)

See scriptling.nomad for the full method reference.

scriptling.container: wait_stopped() confirms a container has fully stopped.

ContainerClient.stop() already blocks until the container reports stopped for Docker, Podman, and Apple Containers, but there was no way to independently re-check that state afterwards. wait_stopped(name_or_id, timeout=30) polls the container’s running state until it stops or the timeout elapses, and treats a container that no longer exists as already stopped.

c.stop("web")
if not c.wait_stopped("web", timeout=15):
    print("container did not stop in time")

New scriptling.find library: locate files and directories by name, type, mtime, and size.

A new utility library inspired by the Unix find command, filling the gap between glob (name matching only) and a hand-rolled os.walk loop:

import scriptling.find as find
import time

# Markdown files modified in the last 24 hours
recent = find.path("/docs", name="*.md", type="file",
                   mtime_min=time.time() - 86400)

# Large log files
big = find.path("/var/log", name="*.log", type="file", size_min=104857600)

find.path(path, *, recursive=True, type="any", name="", mtime_min=None, mtime_max=None, size_min=None, size_max=None, include_hidden=False, follow_links=False, max_depth=None) returns matching paths as a list of strings. Recursive searches stat and filter entries concurrently using a bounded worker pool, the same model as scriptling.grep. See scriptling.find for the full reference.

New shlex, tempfile, and shutil standard libraries.

Three new standard libraries fill gaps in filesystem and command-line tooling:

All three enforce allowedPaths restrictions when registered with explicit paths (except shlex, which is pure string processing with no filesystem access). tempfile creates files with mode 0600 and directories with 0700, and falls back to the first allowed path when the system temp directory is outside the sandbox.

import shlex, subprocess
import tempfile, shutil

# Safe command construction
subprocess.run(shlex.split("echo " + shlex.quote(user_input)))

# Atomic write via temp file
tmp = tempfile.mkstemp(suffix=".tmp")
try:
    # ... write to tmp ...
    shutil.move(tmp, "output.final")
except:
    shutil.rmtree(tmp)

See shlex, tempfile, and shutil for the full references.

New zipfile and tarfile libraries: read and write compressed archives.

Two new standard libraries bring archive handling to Scriptling, matching Python’s zipfile and tarfile modules:

• zipfile — ZipFile(path, mode) with namelist(), read(), extract(), extractall(), write(), writestr(), is_zipfile().
• tarfile — TarFile(path, mode) with the same method set plus gzip support ("r:gz", "w:gz" modes), add(), addstr(), is_tarfile().

Both enforce allowedPaths and block zip/tar-slip (path traversal via crafted entry names). See zipfile and tarfile for the full references.

import tarfile

# Extract a .tar.gz
tf = tarfile.TarFile("release.tar.gz", "r:gz")
tf.extractall("/opt/app")
tf.close()

# Create a zip from files + inline data
import zipfile
zf = zipfile.ZipFile("config.zip", "w")
zf.writestr("version.txt", "1.0.0")
zf.write("/etc/app.conf")
zf.close()

scriptling.similarity: vector operations — cosine_similarity, most_similar, vectorize.

The similarity library gains three vector functions for text-matching workflows:

• cosine_similarity(a, b) — compare two numeric vectors (-1.0 to 1.0). Now exposed in similarity (its natural home) in addition to scriptling.ai. The implementation is shared — no code duplication.
• most_similar(query, vectors, top_k=5) — rank a list of vectors by similarity to a query; returns [{"index": int, "score": float}, …].
• vectorize(text, dims=256) — generate a vector from text using the feature-hashing trick. CPU-only, deterministic, no model or API call required. Texts sharing words produce similar vectors.

import scriptling.similarity as sim

# CPU-only text matching — no embedding API needed
v1 = sim.vectorize("the quick brown fox")
v2 = sim.vectorize("the quick red fox")
print(sim.cosine_similarity(v1, v2))  # high — shares 3 of 4 words

# Rank documents against a query
docs = [sim.vectorize(d) for d in ["hello world", "quick fox", "goodbye"]]
for r in sim.most_similar(sim.vectorize("hi world"), docs, top_k=2):
    print(r["index"], r["score"])

New scriptling.csv library: CSV parsing and formatting (string-based).

A string-based CSV library backed by Go’s encoding/csv (RFC 4180 compliant). Unlike Python’s csv module (which requires file objects), this operates on strings — available in all environments including MCP (no filesystem access needed).

• loads(content, delimiter=",") — CSV text → list of lists.
• loads_dict(content, delimiter=",") — CSV text → list of dicts (first row = headers).
• dumps(rows, delimiter=",") — list of lists → CSV text (auto-quotes values with commas).
• dumps_dict(rows, delimiter=",", columns=None) — list of dicts → CSV text with header row.

import scriptling.csv as csv
import os

# Read and parse
people = csv.loads_dict(os.read_file("users.csv"))
for p in people:
    print(p["name"], p["email"])

# Write
os.write_file("output.csv", csv.dumps_dict(people, columns=["name", "email"]))

See scriptling.csv for the full reference.

New scriptling.xml library: XML parsing and formatting (dict-based).

A simple dict-based XML library (similar to Python’s xmltodict), using the loads/dumps convention consistent with json, yaml, toml, and csv. No filesystem access required — available in all environments including MCP.

• loads(content) — XML string → nested dict. Element tags become dict keys, attributes become @-prefixed keys, repeated elements become lists, text alongside attributes/children uses #text.
• dumps(data, indent="") — dict → XML string. Supports attributes, repeated elements (list values), and optional indentation.

import scriptling.xml as xml

data = xml.loads('<user id="123"><name>Alice</name></user>')
# {"user": {"@id": "123", "name": "Alice"}}

text = xml.dumps({"users": {"user": ["Alice", "Bob"]}}, indent="  ")
# <users>
#   <user>Alice</user>
#   <user>Bob</user>
# </users>

See scriptling.xml for the full reference.

glob: recursive and include_hidden keyword arguments, with bounded parallel recursive search.

glob.glob() and glob.iglob() gain two keyword-only parameters, both defaulting to False:

• recursive=True makes ** match files and directories recursively, descending into every subdirectory. When False (the default), ** is treated as *, matching Python’s glob module.
• include_hidden=True matches entries whose name starts with .; when False (the default) dot-files and dot-directories are skipped.

Recursive searches now run as a bounded parallel directory walk, using the same worker-pool model as scriptling.grep, so large trees are scanned concurrently rather than sequentially.

import glob

# Recursively find all Python files (descends into subdirectories)
py_files = glob.glob("**/*.py", recursive=True)

# Include dot-directories such as .github
all_py = glob.glob("**/*.py", recursive=True, include_hidden=True)

glob.iglob() accepts the same keyword arguments; the recursive path uses the same parallel walk internally.

Note that with the new default (include_hidden=False), dot-files are now skipped by wildcard patterns, matching Python’s behaviour. Previously Go’s glob included them; pass include_hidden=True to restore the old behaviour.