Files
NexusOS/synapse/self_edit.py
T
AthenaandClaude Sonnet 5 0bbe5e200e feat: self-alteration tools (edit_playbook, edit_settings, edit_source)
Gives the assistant three new ACTION tools to change its own playbooks,
runtime settings, and (source checkout only) its own source code, all
reusing the existing run_snippet/remember approval framework — but with
a hardcoded floor (self_edit.ALWAYS_ASK_TOOLS) so these three always pause
for per-call human approval regardless of the global action_tool_policy
setting. Flipping that policy for an unrelated tool must never silently
also unlock unattended self-modification.

synapse/self_edit.py is the new module doing the actual work, documented
in the same explicit "here's what is and isn't a security boundary" style
as code_run.py:
  - edit_source is confined to settings.project_root via the same
    realpath + Path.parents boundary check that just closed a sibling-
    directory bypass in /icons/image, plus a denylist of dangerous
    subtrees (.git, the venv, node_modules, build output, runtime state).
    Gated on settings.source_checkout — refuses cleanly in a wheel
    install, where there's no live repo to edit or commit into.
  - The model sends full file content, never a diff; the server computes
    the diff itself via difflib against what's actually on disk, so a
    human reviews ground truth, not a description the model wrote.
  - Every applied source edit best-effort commits to git as an audit
    trail — independent of, not a substitute for, the approval gate.
  - edit_playbook merges instead of replacing (main.py's prior
    _persist_playbook did a raw replace, which was only safe because the
    frontend form always sent a complete object — unsafe for a tool a
    model calls with a partial argument set, so this also fixes that
    latent bug). Becoming the active system prompt requires an explicit
    make_active flag, never a side effect of an ordinary edit.
  - edit_settings reuses the existing _SETTINGS_DEFAULTS allowlist.

The approval UI (Chatbot.jsx) previously rendered a tool call's arguments
as Object.values(args).join(", ") in a single-line badge — unusable for
reviewing a diff. It now renders a real, server-computed preview (diff
for source, before/after for playbook/settings) via a new shared
diff-view.js helper, with a loud banner when a change would become the
active system prompt or touch action_tool_policy/system_prompt. A new
nexus-edit fence (self-edit-langs.js + Markdown.jsx's EditBlock) shows
the same diff after an edit is applied, mirroring nexus-run.

Verified: 212 backend tests pass (29 new in test_self_edit.py; the 12
pre-existing C/C++/Rust toolchain failures are unrelated and unchanged),
57 frontend node:test cases pass (15 new), eslint and vite build clean,
and the full approval-preview render path was exercised against the real
built UI with a mocked SSE stream covering all four preview branches
(source diff, playbook becomes-main, settings policy-change, and a
rejected/failing preview).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-08-25 20:13:35 -05:00

291 lines
13 KiB
Python

"""The assistant's ability to change what it is: its own playbooks, its own
runtime settings, and (in a source checkout only) its own source files.
WHAT THIS IS NOT
----------------
Not a way to skip human review. Every function here is either read-only
(the `preview_*`/`diff_text` functions, safe to call before anything is
approved) or is only ever reached after the per-call approval round-trip in
chat.py — and `edit_playbook`/`edit_settings`/`edit_source` are *always*
gated that way, regardless of the global `action_tool_policy` setting (see
`ALWAYS_ASK_TOOLS`). This module is the mechanism the approval actually acts
on, layered:
1. Consent edit_* tools are ACTION tools with a hardcoded approval
floor — chat.py pauses for Approve/Deny even when the
global policy is "allow", so flipping that setting for an
unrelated tool (web_search, say) can never silently unlock
unattended self-modification too.
2. Real diff the human reviews a diff computed here, server-side, from
what is actually on disk versus the model's proposed
`new_content` — never a diff or description the model wrote
itself. A model can misdescribe a change; it cannot make
difflib misreport one.
3. Boundary edit_source is confined to one root (settings.project_root)
via the same realpath + `Path.parents` check that closed a
sibling-directory bypass in main.py's /icons/image, plus a
denylist of dangerous subtrees inside that root (.git, the
venv, node_modules, build output, runtime state).
4. Audit trail every applied source edit is committed to git (best-effort;
a missing git binary or a non-repo root degrades the result
to `commit: None`, it never blocks the write). This is a
reversibility net, not a substitute for layer 1 — the
approval already happened before anything is written.
5. Size caps MAX_FILE_CHARS/MAX_DIFF_CHARS bound what a single call can
submit or what the approval UI has to render, mirroring
code_run.py's MAX_SOURCE/MAX_OUTPUT.
A file edit does not hot-reload the running process. Python does not re-import
a changed module on its own, and the frontend's production build is the static
`dist/` the backend serves — editing interface/web/src/*.jsx only affects a
developer's own `npm run dev` session, if one happens to be running. Every
successful edit_source result says so explicitly, because both the model and
the human reviewing it will otherwise reasonably expect an instant effect that
does not happen.
"""
from __future__ import annotations
import difflib
import os
import subprocess
from pathlib import Path
from typing import Any
from .memory.store import store
from .nexus_config import settings
from .playbooks.store import playbook_store
from . import playbook_manager
MAX_FILE_CHARS = 200_000 # chars of proposed file content accepted
MAX_DIFF_CHARS = 20_000 # chars of diff text shown/returned, clipped like code_run.MAX_OUTPUT
# Subdirectories of settings.project_root that edit_source must never touch,
# even though they're inside the one allowed root. `data`/`runtime` land here
# in a source checkout (see nexus_config.py DATA_DIR/RUNTIME_DIR) and are owned
# by their own tools (edit_settings, the memory store), not raw file writes.
_DENYLIST_SUBDIRS = frozenset({
".git", "Promethean", "node_modules", "dist", "__pycache__",
"runtime", "data", ".venv", "venv",
})
# Tools that must always pause for human approval, regardless of the global
# action_tool_policy setting. Shared by tools.py (ACTION_TOOLS membership) and
# chat.py (the approval-gating condition) so there is one definition of the
# floor, not two that could drift apart.
ALWAYS_ASK_TOOLS = frozenset({"edit_playbook", "edit_settings", "edit_source"})
# Tool names whose approval payload gets a computed, human-readable preview
# attached before the human ever sees the Approve/Deny prompt.
PREVIEWABLE = frozenset({"edit_source", "edit_playbook", "edit_settings"})
WINDOWS = os.name == "nt"
_NO_WINDOW = subprocess.CREATE_NO_WINDOW if WINDOWS else 0
class PathError(ValueError):
pass
def _resolve_source_path(rel_path: str) -> Path:
"""A path the model gave us, resolved and boundary-checked against
settings.project_root. Raises PathError with a human-readable reason on
any rejection — the same message is shown in the pre-approval preview and
returned as the tool's error, so both audiences see exactly why."""
raw = (rel_path or "").strip()
if not raw:
raise PathError("path is required")
# Cheap rejection before ever touching the filesystem: an absolute path or
# a Windows drive prefix is never a legitimate "file in this project" path.
if raw.startswith(("/", "\\")) or (len(raw) > 1 and raw[1] == ":"):
raise PathError(f"path must be relative to the project root, not absolute: {raw!r}")
root = Path(os.path.realpath(str(settings.project_root)))
candidate = root / raw
real = Path(os.path.realpath(str(candidate)))
# Same real == root or root in real.parents pattern as
# icons/compositor.py::_is_allowed_path — a bare str.startswith() here
# would let a sibling directory that merely shares a prefix pass, exactly
# the bug just fixed in main.py's /icons/image.
if not (real == root or root in real.parents):
raise PathError(f"path escapes the project root: {raw!r}")
try:
top = real.relative_to(root).parts[0]
except (ValueError, IndexError):
top = ""
if top in _DENYLIST_SUBDIRS:
raise PathError(f"{top}/ is off-limits to edit_source — use its own tool if there is one")
return real
def source_checkout_required() -> None:
"""Raise if this install has no live, editable source tree to write into.
In a wheel install, synapse/ lives inside site-packages with no sibling
repo — settings.project_root would just be the installed package dir, and
there is nothing to commit into. Same precedent as nexusos_cli/ncp.py
refusing to start the Vite dev server in a wheel install."""
if not settings.source_checkout:
raise RuntimeError(
"edit_source is unavailable — this is not a source checkout, so "
"there is no live project tree to edit or commit into."
)
def diff_text(path_display: str, before: str, after: str) -> str:
lines = difflib.unified_diff(
before.splitlines(keepends=True),
after.splitlines(keepends=True),
fromfile=path_display,
tofile=path_display,
)
text = "".join(lines)
if len(text) <= MAX_DIFF_CHARS:
return text
return text[:MAX_DIFF_CHARS] + f"\n... [truncated at {MAX_DIFF_CHARS} characters, {len(text)} total]"
def preview_source_edit(path: str, new_content: str) -> dict:
"""Read-only: never raises. Computes the real diff between what's on disk
and the proposed new_content, so the approval UI shows ground truth before
anything is written. Degrades to {"ok": False, "error": ...} on any
rejection (bad path, wheel install, oversized content) rather than
crashing the approval payload — the human still sees why it would fail."""
try:
source_checkout_required()
real = _resolve_source_path(path)
new_content = new_content or ""
if len(new_content) > MAX_FILE_CHARS:
return {"ok": False, "error": f"new_content is over {MAX_FILE_CHARS} characters"}
before = real.read_text(encoding="utf-8") if real.is_file() else ""
display = str(real.relative_to(Path(os.path.realpath(str(settings.project_root)))))
return {
"ok": True,
"path": display,
"is_new_file": not real.is_file(),
"diff": diff_text(display, before, new_content),
}
except (PathError, RuntimeError) as e:
return {"ok": False, "error": str(e)}
except Exception as e:
return {"ok": False, "error": f"could not compute preview: {e}"}
def _git_commit(real_path: Path, summary: str) -> str | None:
"""Best-effort audit-trail commit. Never raises, never undoes the write
that already happened — a missing git binary or a project root that isn't
a repo just means commit stays None."""
root = str(settings.project_root)
try:
rel = str(real_path.relative_to(Path(os.path.realpath(root))))
message = f"self-edit: {(summary or rel)[:180]}"
subprocess.run(
["git", "add", "--", rel], cwd=root, check=True,
capture_output=True, creationflags=_NO_WINDOW,
)
subprocess.run(
["git", "commit", "-m", message, "--", rel], cwd=root, check=True,
capture_output=True, creationflags=_NO_WINDOW,
)
sha = subprocess.run(
["git", "rev-parse", "--short", "HEAD"], cwd=root, check=True,
capture_output=True, text=True, creationflags=_NO_WINDOW,
)
return sha.stdout.strip() or None
except Exception:
return None
def apply_source_edit(path: str, new_content: str, summary: str = "") -> dict:
"""Write an approved edit_source call. Re-validates everything preview did
— never trust that nothing changed between preview and approval — then
writes, diffs against the pre-write content, and commits."""
try:
source_checkout_required()
real = _resolve_source_path(path)
new_content = new_content or ""
if len(new_content) > MAX_FILE_CHARS:
return {"ok": False, "error": f"new_content is over {MAX_FILE_CHARS} characters"}
except (PathError, RuntimeError) as e:
return {"ok": False, "error": str(e)}
before = real.read_text(encoding="utf-8") if real.is_file() else ""
display = str(real.relative_to(Path(os.path.realpath(str(settings.project_root)))))
real.parent.mkdir(parents=True, exist_ok=True)
real.write_text(new_content, encoding="utf-8")
return {
"ok": True,
"path": display,
"diff": diff_text(display, before, new_content),
"commit": _git_commit(real, summary),
}
def preview_playbook_edit(args: dict) -> dict:
"""Read-only before/after preview for edit_playbook. Mirrors the merge
semantics of playbook_manager.persist_playbook(merge=True) without writing
anything, so the approval UI shows exactly what will actually change."""
try:
pb_id = str(args.get("id") or "")
existing = playbook_store.get_playbook(pb_id) if pb_id else None
make_active = bool(args.get("make_active"))
fields = {}
for key in ("title", "goal", "instructions", "tags", "tools", "model"):
if key in args and args[key] is not None:
fields[key] = args[key]
elif existing is not None:
fields[key] = getattr(existing, key)
else:
fields[key] = [] if key in ("tags", "tools") else ""
if existing is None and not (fields["title"] and fields["goal"] and fields["instructions"]):
return {"ok": False, "error": "title, goal, and instructions are required to create a new playbook"}
is_new = existing is None
becomes_main = make_active or (is_new and not playbook_store.all_playbooks())
return {
"ok": True,
"is_new": is_new,
"before": existing.model_dump() if existing else None,
"after": fields,
"becomes_main_playbook": becomes_main,
}
except Exception as e:
return {"ok": False, "error": f"could not compute preview: {e}"}
def preview_settings_edit(args: dict) -> dict:
"""Read-only before/after preview for edit_settings, split into keys that
will actually apply versus ones update_settings would silently ignore —
the approval UI must never imply an unknown key will take effect."""
try:
changes = args.get("changes") or {}
current = store.get_settings()
applied: dict[str, dict[str, Any]] = {}
ignored_unknown: list[str] = []
for key, value in changes.items():
if key in store._SETTINGS_DEFAULTS:
applied[key] = {"before": current.get(key), "after": value}
else:
ignored_unknown.append(key)
return {
"ok": True,
"applied": applied,
"ignored_unknown": ignored_unknown,
"policy_change": "action_tool_policy" in applied,
"system_prompt_change": "system_prompt" in applied,
}
except Exception as e:
return {"ok": False, "error": f"could not compute preview: {e}"}
def preview_for(name: str, args: dict) -> dict:
"""Single dispatch entry point chat.py calls to enrich an approval payload."""
if name == "edit_source":
return preview_source_edit(args.get("path", ""), args.get("new_content", ""))
if name == "edit_playbook":
return preview_playbook_edit(args)
if name == "edit_settings":
return preview_settings_edit(args)
return {"ok": False, "error": f"no previewer for {name}"}