Skip to content

Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.

feat(society): run agent-started coding threads with full access

rubenluetke10-beep2ce3f1070dd7
diff --git a/docs/agent-society/agent-definition.md b/docs/agent-society/agent-definition.md
index baba371fb5..1aed9c9cd1 100644
--- a/docs/agent-society/agent-definition.md
+++ b/docs/agent-society/agent-definition.md
@@ -130,13 +130,15 @@ Agentic IDE (`docs/agentic-ide-threads.md`): an ordinary chat session on the
`agent` surface that the person can open, read and continue. Actions:
`agents` (the coding CLIs on this computer and their models), `projects`,
`open` (agent, folder or project name, the brief the agent writes, optional
-model, effort and access `plan`/`build`/`full`), `send`, `read`, `answer` (the
+model and effort), `send`, `read`, `answer` (the
thread's question card, or `build`/`keep` on its plan card), `stop`, `threads`,
and `files` (read-only `ls`/`read`/`glob`/`grep` inside the thread's folder).
-Reads are safe-tier, the rest monitor-tier; `open` with `full` access (no
-sandbox, no approvals) is ask-tier. Grants, denies, approval rules and the
-ceiling apply through the normal ToolExecutor path; plan/read-only chats do
-not get the tool. Approvals of the CLI's own tool calls stay with the person.
+Reads are safe-tier, the rest monitor-tier. Every thread an agent opens runs
+in the runner's bypass mode (Claude Code `bypassPermissions`, Codex
+`full-access`, …), so it never stops for an approval (maintainer decision
+2026-10-06). Grants, denies, approval rules and the ceiling still gate the
+tool itself through the normal ToolExecutor path; plan/read-only chats do not
+get it.
The thread records who started it (`agent_chat_thread_owners`) and every
message the agent writes carries an `author`, so the thread shows the agent's
diff --git a/docs/agentic-ide-threads.md b/docs/agentic-ide-threads.md
index 66919b171b..b2401174b2 100644
--- a/docs/agentic-ide-threads.md
+++ b/docs/agentic-ide-threads.md
@@ -96,7 +96,8 @@ The agent's own chat shows a line that opens the thread. When a turn the
agent started finishes, asks a question, presents a plan or waits for an
approval, the agent is woken with the result: it checks the work, answers the
question or plan card, sends a follow-up, and tells the person the outcome.
-Approvals of the coding agent's own commands stay with the person. A turn the
+A thread an agent opens runs with full access (bypass permissions), so it does
+not stop for approvals; the person can lower that in the composer. A turn the
person types into the thread is theirs and wakes no agent. After 25 automatic
updates for one thread the agent stops following it until the person writes
in the agent's chat again.
diff --git a/jarvis/society/coding_threads.py b/jarvis/society/coding_threads.py
index b2e6f44650..309ed6abef 100644
--- a/jarvis/society/coding_threads.py
+++ b/jarvis/society/coding_threads.py
@@ -65,7 +65,6 @@ DUPLICATE_OPEN_S: Final[float] = 120.0
_READ_ACTIONS: Final[frozenset[str]] = frozenset(
{"agents", "projects", "threads", "read", "files"}
)
-_ACCESS: Final[tuple[str, ...]] = ("plan", "build", "full")
_FILE_OPS: Final[dict[str, str]] = {"ls": "Ls", "read": "Read", "glob": "Glob", "grep": "Grep"}
_TEXT_LIMIT: Final[int] = 4000
_WRITING_TOOLS: Final[frozenset[str]] = frozenset(
@@ -280,7 +279,8 @@ class CodingThreadTool:
"agents (the coding agents on this computer, their models), projects (known project "
"folders), open (start a thread: agent, folder = absolute project path or project "
"name, prompt = a complete, self-contained brief you write, optional model, effort, "
- "access plan|build|full, title), send (a follow-up into your thread), read (the "
+ "title; the thread runs with full access), send (a follow-up into your thread), "
+ "read (the "
"thread's transcript; pass after=<cursor> for news only), answer (the thread's open "
"question card: answers = one string per question, an option label or your own "
"text; or its plan card: decision build|keep), stop (end the running turn), threads "
@@ -311,7 +311,6 @@ class CodingThreadTool:
"agent": {"type": "string", "description": "Coding agent id or name (open)."},
"model": {"type": "string"},
"effort": {"type": "string"},
- "access": {"type": "string", "enum": list(_ACCESS)},
"folder": {"type": "string", "description": "Project folder or project name."},
"title": {"type": "string"},
"prompt": {"type": "string"},
@@ -339,8 +338,6 @@ class CodingThreadTool:
action = str(args.get("action") or "")
if action in _READ_ACTIONS:
return "safe"
- if action == "open" and str(args.get("access") or "build") == "full":
- return "ask" # no sandbox, no approvals: the person decides
return "monitor"
def describe_args(self, args: dict[str, Any]) -> dict[str, str]:
@@ -410,11 +407,7 @@ class CodingThreadTool:
)
return {
"coding_agents": rows,
- "access": {
- "plan": "reads and plans only",
- "build": "edits in the folder; risky commands ask the person (default)",
- "full": "no sandbox, no approvals — only when the person asked for it",
- },
+ "access": "Threads you open run with full access: no approval prompts.",
}
async def _do_projects(self, service: Any, agent: Any, args: dict[str, Any]) -> dict[str, Any]:
@@ -446,9 +439,6 @@ class CodingThreadTool:
if not (await _installed_clis()).get(row.agent, True):
raise ValueError(f"{row.label} is not installed on this computer.")
folder = await asyncio.to_thread(resolve_folder, str(args.get("folder") or ""))
- access = str(args.get("access") or "build")
- if access not in _ACCESS:
- raise ValueError("access must be plan, build or full.")
fingerprint = hashlib.sha256(
f"{row.id}\n{folder}\n{prompt}".encode()
).hexdigest()[:24]
@@ -468,7 +458,9 @@ class CodingThreadTool:
model=match_model(row, str(args.get("model") or "")),
effort=(str(args["effort"]).strip() or None) if args.get("effort") else None,
cwd=folder,
- permission_mode=access_mode(runner, access),
+ # The maintainer's choice: an agent's coding threads never stop for
+ # approvals, so they always run in the runner's bypass mode.
+ permission_mode=access_mode(runner, "full"),
title=_clip(str(args.get("title") or ""), 80),
surface=THREAD_SURFACE,
)
diff --git a/tests/unit/society/test_coding_threads.py b/tests/unit/society/test_coding_threads.py
index 975e9191d1..0fd5af3a39 100644
--- a/tests/unit/society/test_coding_threads.py
+++ b/tests/unit/society/test_coding_threads.py
@@ -134,6 +134,7 @@ async def test_open_starts_a_thread_the_person_can_watch(society: Society, tmp_p
folder = await asyncio.to_thread(tmp_path.resolve)
assert session.cwd == str(folder)
assert session.model == "claude-opus-5-5"
+ assert session.permission_mode == "bypassPermissions"
assert rt.projects == [str(folder)]
first = rt_service.store.list_events(thread)[0]
assert first["kind"] == "user_message"
@@ -335,10 +336,10 @@ async def test_a_retried_open_is_the_same_thread(society: Society, tmp_path: Pat
assert first == second
-async def test_full_access_asks_the_person_first():
+async def test_threads_run_with_full_access_and_open_without_a_card():
t = CodingThreadTool(SimpleNamespace(), "nova")
- assert t.risk_tier_for_args({"action": "open", "access": "full"}) == "ask"
assert t.risk_tier_for_args({"action": "open"}) == "monitor"
+ assert "access" not in t.schema["properties"]
assert t.risk_tier_for_args({"action": "read"}) == "safe"