feat(grok): close the Claude-parity divergences (reasoning, subagents, web, bash-guard)

Bring the grok CLI to parity with the Claude path on the four deliberate
differences:

- Reasoning: drop the per-role `--effort low` default — Claude sets no per-role
  thinking budget, so grok now uses the model default for every role. The
  fleet-wide ROBOCO_GROK_REASONING_EFFORT override stays as a cost lever. (This
  also un-caps intake-draft quality, the one that actually mattered.)
- Subagents: the intake interviewer may now fan out to subagents (parity with the
  Claude intake's `Task` allowance); every other role still has `Agent` removed.
- Web: `--disable-web-search` for every role — no agent gets direct web (Claude's
  tool set has none either); the roles that get web reach it through the gated
  roboco-search MCP, unaffected.
- Bash command filtering: full parity, split by deny semantics. Verified live that
  a grok PreToolUse hook deny CANCELS the run, while native `--deny` denies
  GRACEFULLY (the agent gets a permission error and recovers). So:
    * git network/branch/history ops -> native `--deny` (operational reflex; the
      agent must recover, not drop the task). Expanded to the full bash-guard set.
    * credential-exfil / identity-forgery / internal-API / env-dump patterns ->
      the SAME bash-guard the Claude path runs, wired as a grok PreToolUse hook
      (ROBOCO_GUARD_SKIP_GIT=1 so it leaves git to `--deny`). A hard cancel is the
      right response there — no legitimate agent reads ~/.netrc or forges an
      X-Agent-ID. One tolerance line (accept grok's camelCase `toolInput`) makes
      the one tested script guard both runtimes; +5 grok cases (50/50 green).

Also cleaned stale internal task-number / smoke labels out of bash-guard-hook.sh.
This commit is contained in:
Renn F
2026-06-19 05:35:19 +02:00
parent 414b3492c4
commit f6aa590370
5 changed files with 218 additions and 55 deletions
+9 -6
View File
@@ -101,13 +101,16 @@ ROBOCO_DEFAULT_EMBEDDING_MODEL=qwen3-embedding:0.6b
# ROBOCO_GROK_AGENT_IMAGE=roboco-agent-grok:latest # ROBOCO_GROK_AGENT_IMAGE=roboco-agent-grok:latest
# ROBOCO_GROK_CLI_MODEL=grok-build # ROBOCO_GROK_CLI_MODEL=grok-build
# Per-role tool permissions are computed as native grok flags (subagents off; # Per-role tool permissions are computed as native grok flags (subagents off
# edit/shell removed for non-coding roles; raw git mutation + rm -rf denied for # except intake; edit/shell removed for non-coding roles; git network/branch/
# coding roles) — there is nothing to set here. # history mutation + rm -rf denied for coding roles; web search off for all —
# gated web is via the roboco-search MCP). Credential-exfil / identity-forgery /
# internal-API shell patterns are blocked by the same bash-guard the Claude path
# runs, wired as a grok PreToolUse hook. Nothing to set here.
# Force one reasoning effort for ALL Grok agents: low | medium | high | xhigh | # Reasoning effort for ALL Grok agents: low | medium | high | xhigh | max. Empty
# max (or empty for the per-role default — coordination/docs/board roles request # keeps grok's model default for every role (parity with Claude — no per-role
# "low" to cut reasoning cost, code roles keep full reasoning). # cut); set this to trade quality for cost across the whole fleet.
# ROBOCO_GROK_REASONING_EFFORT= # ROBOCO_GROK_REASONING_EFFORT=
# Hard ceiling on agentic turns per run (loop guard). # Hard ceiling on agentic turns per run (loop guard).
+23 -11
View File
@@ -10,7 +10,16 @@
# #
# Claude Code passes the PreToolUse event on stdin as JSON: # Claude Code passes the PreToolUse event on stdin as JSON:
# { "tool_name": "Bash", "tool_input": { "command": "...", "description": "..." } } # { "tool_name": "Bash", "tool_input": { "command": "...", "description": "..." } }
# Exit 0 = allow. Exit 2 = deny with message on stdout. # The grok CLI passes the same event with camelCase keys (toolName / toolInput);
# the extractor accepts either, so the one tested script guards both runtimes.
# Exit 0 = allow. Exit 2 = deny.
#
# ROBOCO_GUARD_SKIP_GIT=1 skips the git-ops category. The grok path sets it because
# grok handles git via NATIVE --deny rules, which deny GRACEFULLY (the agent gets
# a permission error and recovers) — whereas a grok hook deny CANCELS the whole
# run. So grok keeps git on --deny (operational reflex → recoverable) and uses
# this hook only for the exfil categories (no legit use → a hard cancel is the
# right response). Claude has no such --deny, so it keeps the git block here.
# #
# Deny categories: # Deny categories:
# - Network git ops (require token injection only done by the MCP layer) # - Network git ops (require token injection only done by the MCP layer)
@@ -27,7 +36,7 @@ cmd=$(printf '%s' "$input" | python3 -c '
import json, sys import json, sys
try: try:
d = json.loads(sys.stdin.read()) d = json.loads(sys.stdin.read())
ti = d.get("tool_input", {}) or {} ti = d.get("tool_input") or d.get("toolInput") or {}
print(ti.get("command", "")) print(ti.get("command", ""))
except Exception: except Exception:
print("") print("")
@@ -36,7 +45,7 @@ except Exception:
low=$(printf '%s' "$cmd" | tr "[:upper:]" "[:lower:]") low=$(printf '%s' "$cmd" | tr "[:upper:]" "[:lower:]")
# Skeletonize the command for the git-ops check ONLY (#165): strip heredoc # Skeletonize the command for the git-ops check ONLY: strip heredoc
# bodies and echo/printf literal arguments. Those are data the shell writes # bodies and echo/printf literal arguments. Those are data the shell writes
# to a file, never commands the shell executes — so a README/heredoc that # to a file, never commands the shell executes — so a README/heredoc that
# merely documents `git commit` must not be mistaken for invoking git. # merely documents `git commit` must not be mistaken for invoking git.
@@ -82,7 +91,10 @@ fi
git_skel_low=$(printf '%s' "$git_skel" | tr "[:upper:]" "[:lower:]") git_skel_low=$(printf '%s' "$git_skel" | tr "[:upper:]" "[:lower:]")
# --- git network / auth ops --------------------------------------------------- # --- git network / auth ops ---------------------------------------------------
if echo "$git_skel_low" | grep -qE '(^|[[:space:];&|])git[[:space:]]+(fetch|pull|push|clone|remote|ls-remote|checkout|commit|merge|rebase|reset|cherry-pick|revert|tag[[:space:]]+-d|update-ref|reflog[[:space:]]+delete)'; then # Skipped on grok (handled by native --deny so a blocked git op is recoverable,
# not a run-cancelling hook deny). See the header.
if [[ "${ROBOCO_GUARD_SKIP_GIT:-}" != "1" ]] && \
echo "$git_skel_low" | grep -qE '(^|[[:space:];&|])git[[:space:]]+(fetch|pull|push|clone|remote|ls-remote|checkout|commit|merge|rebase|reset|cherry-pick|revert|tag[[:space:]]+-d|update-ref|reflog[[:space:]]+delete)'; then
echo "Denied: shell git for network / auth / branch-mutating ops is blocked." >&2 echo "Denied: shell git for network / auth / branch-mutating ops is blocked." >&2
echo "Use the verb listed in your role's State→Verb table (e.g. commit, complete, i_am_done)." >&2 echo "Use the verb listed in your role's State→Verb table (e.g. commit, complete, i_am_done)." >&2
exit 2 exit 2
@@ -119,20 +131,20 @@ fi
# - scheme-less: `curl roboco-orchestrator:8000/api` # - scheme-less: `curl roboco-orchestrator:8000/api`
# - protocol-relative: `curl //roboco-orchestrator:8000/api` # - protocol-relative: `curl //roboco-orchestrator:8000/api`
# - any flag ordering: `curl -s -X POST http://localhost:8000/x -d ...` # - any flag ordering: `curl -s -X POST http://localhost:8000/x -d ...`
# Interpreter / library-driven HTTP is handled by the #175 rule below. # Interpreter / library-driven HTTP is handled by the rule below.
# KNOWN GAP (still out of scope here): # KNOWN GAP (still out of scope here):
# - Variable expansion: `URL=http://orchestrator/x; curl $URL` — the guard # - Variable expansion: `URL=http://orchestrator/x; curl $URL` — the guard
# sees `curl $URL`, not the expanded URL, so this slips through. The # sees `curl $URL`, not the expanded URL, so this slips through. The
# X-Agent-Role check (task 4) is the second gate. # server-side X-Agent-Role check is the second gate.
if echo "$low" | grep -qE '(^|[[:space:];&|])(curl|wget|http|https|httpie)[[:space:]]' && \ if echo "$low" | grep -qE '(^|[[:space:];&|])(curl|wget|http|https|httpie)[[:space:]]' && \
echo "$low" | grep -qE '((http|https)://)?/?(roboco-[a-z0-9_-]+|localhost|127\.0\.0\.1|0\.0\.0\.0)[:/]'; then echo "$low" | grep -qE '((http|https)://)?/?(roboco-[a-z0-9_-]+|localhost|127\.0\.0\.1|0\.0\.0\.0)[:/]'; then
echo "Denied: internal API calls bypass the gateway. Use the MCP verbs (roboco-flow / roboco-do / roboco-git-readonly / roboco-optimal / roboco-docs) — they route through the orchestrator with the right auth and tracing." >&2 echo "Denied: internal API calls bypass the gateway. Use the MCP verbs (roboco-flow / roboco-do / roboco-git-readonly / roboco-optimal / roboco-docs) — they route through the orchestrator with the right auth and tracing." >&2
exit 2 exit 2
fi fi
# --- interpreter/library HTTP to an internal host (task #175) ----------------- # --- interpreter/library HTTP to an internal host ----------------------------
# The curl/wget rule above only fires when the FIRST token is an HTTP CLI. # The curl/wget rule above only fires when the FIRST token is an HTTP CLI.
# smoke-17 showed an agent reach the orchestrator with forged X-Agent-* # A live run showed an agent reach the orchestrator with forged X-Agent-*
# identity headers via: # identity headers via:
# python3 << 'EOF' # python3 << 'EOF'
# import httpx # import httpx
@@ -140,7 +152,7 @@ fi
# headers={"X-Agent-ID": "<self>", "X-Agent-Role": "developer"}) # headers={"X-Agent-ID": "<self>", "X-Agent-Role": "developer"})
# EOF # EOF
# The binary is python3 (slips the CLI check) and it imports httpx, not # The binary is python3 (slips the CLI check) and it imports httpx, not
# roboco.* (slips the #164 import check). Close it language-agnostically: # roboco.* (slips the roboco-internals import check). Close it language-agnostically:
# deny when the command pairs an HTTP-client token with a forbidden # deny when the command pairs an HTTP-client token with a forbidden
# internal host. The whole command (heredoc body included) is in $low, # internal host. The whole command (heredoc body included) is in $low,
# consistent with the curl/wget sibling above. Legitimate shell work does # consistent with the curl/wget sibling above. Legitimate shell work does
@@ -200,7 +212,7 @@ if echo "$low" | grep -qE '(^|[[:space:];&|])(python3?|perl|node|ruby|awk|sed)[[
exit 2 exit 2
fi fi
# --- gateway-internals import bypass (task #164) ------------------------------ # --- gateway-internals import bypass ------------------------------------------
# An agent must reach the orchestrator ONLY through its manifest-bound MCP # An agent must reach the orchestrator ONLY through its manifest-bound MCP
# verbs. Importing the server package directly # verbs. Importing the server package directly
# uv run python3 -c "from roboco.mcp.flow_server import open_pr; open_pr(...)" # uv run python3 -c "from roboco.mcp.flow_server import open_pr; open_pr(...)"
@@ -217,7 +229,7 @@ if echo "$low" | grep -qE '(python3?|uv[[:space:]]+run|poetry[[:space:]]+run|pip
exit 2 exit 2
fi fi
# --- agent-identity forgery (task #164) -------------------------------------- # --- agent-identity forgery --------------------------------------------------
# ROBOCO_AGENT_ID is the agent's identity. It is injected by the orchestrator # ROBOCO_AGENT_ID is the agent's identity. It is injected by the orchestrator
# at spawn and the agent process must never rewrite it — doing so lets one # at spawn and the agent process must never rewrite it — doing so lets one
# agent act as another (forged audit trail, bypassed ownership checks). No # agent act as another (forged audit trail, bypassed ownership checks). No
+23
View File
@@ -106,6 +106,29 @@ run_case "allow declare -a arr" 0 "declare -a arr=(a b c)"
run_case "allow cat README" 0 "cat README.md" run_case "allow cat README" 0 "cat README.md"
run_case "allow curl non-github" 0 "curl https://example.com/info" run_case "allow curl non-github" 0 "curl https://example.com/info"
# ---------- grok variant: camelCase input + ROBOCO_GUARD_SKIP_GIT ----------
# The grok CLI sends `toolInput` (camelCase) and the grok hook runs with
# ROBOCO_GUARD_SKIP_GIT=1 (git is handled by graceful native --deny). Exfil
# categories must STILL deny; git must now pass through to --deny.
run_case_grok() {
local label="$1" expected="$2" cmd="$3" json actual
# shellcheck disable=SC2016
json=$(python3 -c 'import json, sys; print(json.dumps({"toolName":"run_terminal_command","toolInput":{"command":sys.argv[1]}}))' "$cmd")
echo "$json" | ROBOCO_GUARD_SKIP_GIT=1 $HOOK >/dev/null 2>&1
actual=$?
if [[ "$actual" == "$expected" ]]; then
PASS=$((PASS + 1))
else
FAIL=$((FAIL + 1))
FAILS+=("[$label] expected $expected, got $actual | cmd: $cmd")
fi
}
run_case_grok "grok skip-git: git push allowed" 0 "git push origin main"
run_case_grok "grok camelCase: netrc denied" 2 "cat ~/.netrc"
run_case_grok "grok camelCase: env denied" 2 "env"
run_case_grok "grok camelCase: identity forgery" 2 "export ROBOCO_AGENT_ID=other"
run_case_grok "grok camelCase: allow ls" 0 "ls -la /workspace"
# ---------- Report ---------- # ---------- Report ----------
echo echo
echo "===== bash-guard-hook tests =====" echo "===== bash-guard-hook tests ====="
+99 -31
View File
@@ -51,39 +51,53 @@ GROK_AGENTS_PATH = Path.home() / ".grok" / "AGENTS.md"
SYSTEM_PROMPT_PATH = Path( SYSTEM_PROMPT_PATH = Path(
os.environ.get("ROBOCO_SYSTEM_PROMPT", "/app/system-prompt.md") os.environ.get("ROBOCO_SYSTEM_PROMPT", "/app/system-prompt.md")
) )
# grok loads blocking ``PreToolUse`` hooks from ``$HOME/.grok/hooks/*.json``
# (always trusted). We install the SAME bash-guard the Claude path runs as a
# PreToolUse hook to get its full exfil-pattern analysis (credential files,
# /proc/environ, internal-API forgery, identity forgery, …) — far beyond the
# glob ``--deny`` rules. It runs with ``ROBOCO_GUARD_SKIP_GIT=1``: a grok hook
# deny CANCELS the run, which is the right response for an exfil attempt (no
# legit use) but wrong for a routine git op, so git stays on the graceful
# ``--deny`` rules. The script is baked into the agent base image.
GROK_HOOKS_DIR = Path.home() / ".grok" / "hooks"
BASH_GUARD_HOOK = os.environ.get(
"ROBOCO_BASH_GUARD_HOOK", "/app/scripts/bash-guard-hook.sh"
)
# The entrypoint reads the computed flags (one token per line) from this file. # The entrypoint reads the computed flags (one token per line) from this file.
GROK_ARGS_PATH = Path(os.environ.get("ROBOCO_GROK_ARGS_FILE", "/tmp/roboco-grok-args")) GROK_ARGS_PATH = Path(os.environ.get("ROBOCO_GROK_ARGS_FILE", "/tmp/roboco-grok-args"))
# Hard ceiling on agentic turns (loop guard). Operator-tunable. # Hard ceiling on agentic turns (loop guard). Operator-tunable.
_DEFAULT_MAX_TURNS = 200 _DEFAULT_MAX_TURNS = 200
# Roles that request reduced reasoning (grok bills reasoning at the output rate, # Reasoning effort is left at grok's model default for every role (parity with
# so it dominates cost). Code-quality roles keep full reasoning; coordination / # the Claude path, which sets no per-role thinking budget). An operator can still
# docs / board roles ask for ``low``. # trade quality for cost across the fleet with ``ROBOCO_GROK_REASONING_EFFORT``.
_MINIMAL_REASONING_ROLES = frozenset(
{
"cell_pm",
"main_pm",
"documenter",
"product_owner",
"head_marketing",
"auditor",
"prompter",
"secretary",
}
)
_FULL_REASONING_OVERRIDES = frozenset({"default", "full", "none", ""}) _FULL_REASONING_OVERRIDES = frozenset({"default", "full", "none", ""})
# Roles that legitimately run a shell. Review / board roles never do. # Roles that legitimately run a shell. Review / board roles never do.
_BASH_ROLES = frozenset({"developer", "documenter", "cell_pm", "main_pm"}) _BASH_ROLES = frozenset({"developer", "documenter", "cell_pm", "main_pm"})
# The intake interviewer reads the codebase to draft a task and may fan out
# exploration to subagents (parity with the Claude intake's ``Task`` allowance);
# every other role drives work through the gateway verbs, never CLI subagents.
_SUBAGENT_ALLOWED_ROLES = frozenset({"prompter"})
# Grok CLI tool IDs (from the CLI's --tools/--disallowed-tools reference). # Grok CLI tool IDs (from the CLI's --tools/--disallowed-tools reference).
_TOOL_SHELL = "run_terminal_cmd" _TOOL_SHELL = "run_terminal_cmd"
_TOOL_EDIT = "search_replace" _TOOL_EDIT = "search_replace"
_TOOL_SUBAGENT = "Agent" _TOOL_SUBAGENT = "Agent"
# Raw git mutation is gateway-mediated (the commit / open_pr verbs); agents never # Raw git network / branch / history mutation is gateway-mediated (the commit /
# push or commit via raw bash. Denied for every bash-capable role. # open_pr verbs); agents never run these via raw bash. Denied for every
# bash-capable role — the same set the Claude bash-guard blocks.
#
# Git ops are GRACEFUL native ``--deny`` denials (a blocked command returns a
# permission error to the model, which adapts to the gateway verb — the run
# continues), deliberately NOT routed through the bash-guard hook: verified live
# that a grok hook deny CANCELS the whole run, which would turn one reflexive git
# op into a dropped task. The exfil categories (credential reads, /proc/environ,
# internal-API forgery, …) DO run through the hook (see GROK_HOOKS_DIR) — there a
# hard cancel is the right response, since no legitimate agent triggers them.
_GIT_MUTATE_DENY = ( _GIT_MUTATE_DENY = (
"Bash(git push*)", "Bash(git push*)",
"Bash(git fetch*)", "Bash(git fetch*)",
@@ -92,6 +106,13 @@ _GIT_MUTATE_DENY = (
"Bash(git commit*)", "Bash(git commit*)",
"Bash(git remote*)", "Bash(git remote*)",
"Bash(git reset*)", "Bash(git reset*)",
"Bash(git ls-remote*)",
"Bash(git checkout*)",
"Bash(git merge*)",
"Bash(git rebase*)",
"Bash(git cherry-pick*)",
"Bash(git revert*)",
"Bash(git update-ref*)",
) )
_DESTRUCTIVE_DENY = ("Bash(rm -rf*)",) _DESTRUCTIVE_DENY = ("Bash(rm -rf*)",)
@@ -126,7 +147,9 @@ def _allows_write(role: str) -> bool:
def _disallowed_tools(role: str) -> str: def _disallowed_tools(role: str) -> str:
"""Comma-separated ``--disallowed-tools`` value for a role.""" """Comma-separated ``--disallowed-tools`` value for a role."""
tools = [_TOOL_SUBAGENT] tools: list[str] = []
if role not in _SUBAGENT_ALLOWED_ROLES:
tools.append(_TOOL_SUBAGENT)
if role not in _BASH_ROLES: if role not in _BASH_ROLES:
tools.append(_TOOL_SHELL) tools.append(_TOOL_SHELL)
if not _allows_write(role): if not _allows_write(role):
@@ -141,18 +164,17 @@ def _deny_rules(role: str) -> list[str]:
return [*_DESTRUCTIVE_DENY, *_GIT_MUTATE_DENY] return [*_DESTRUCTIVE_DENY, *_GIT_MUTATE_DENY]
def _effort_for(role: str) -> str | None: def _effort() -> str | None:
"""Resolve ``--effort`` for a role; ``None`` keeps grok's default reasoning. """Resolve ``--effort`` from the fleet override; ``None`` = grok's default.
A global ``ROBOCO_GROK_REASONING_EFFORT`` override wins over the per-role No per-role reduction (parity with Claude). A global
default (``default`` / ``full`` / empty disables the reduction). ``ROBOCO_GROK_REASONING_EFFORT`` lets an operator dial cost vs quality;
``default`` / ``full`` / empty keeps the model default.
""" """
override = os.environ.get("ROBOCO_GROK_REASONING_EFFORT", "").strip() override = os.environ.get("ROBOCO_GROK_REASONING_EFFORT", "").strip()
if override: if override and override.lower() not in _FULL_REASONING_OVERRIDES:
return ( return override.lower()
None if override.lower() in _FULL_REASONING_OVERRIDES else override.lower() return None
)
return "low" if role in _MINIMAL_REASONING_ROLES else None
def grok_cli_args_for_role( def grok_cli_args_for_role(
@@ -160,14 +182,17 @@ def grok_cli_args_for_role(
) -> list[str]: ) -> list[str]:
"""The per-role ``grok -p`` flag tokens (excludes ``-p``/model/cwd). """The per-role ``grok -p`` flag tokens (excludes ``-p``/model/cwd).
Order: tool removal, turn cap, deny rules, then effort. Each token is a Order: tool removal, web off, turn cap, deny rules, then effort. Each token is
separate list element so callers can splice them without shell quoting. a separate list element so callers can splice them without shell quoting.
""" """
args: list[str] = ["--disallowed-tools", _disallowed_tools(role)] args: list[str] = ["--disallowed-tools", _disallowed_tools(role)]
# No direct web for any role (parity with the Claude path's tool set); the
# roles that get web reach it through the gated roboco-search MCP server.
args += ["--disable-web-search"]
args += ["--max-turns", str(max_turns)] args += ["--max-turns", str(max_turns)]
for rule in _deny_rules(role): for rule in _deny_rules(role):
args += ["--deny", rule] args += ["--deny", rule]
effort = _effort_for(role) effort = _effort()
if effort: if effort:
args += ["--effort", effort] args += ["--effort", effort]
return args return args
@@ -207,8 +232,50 @@ def write_agents_md(
return True return True
def bash_guard_hook_config(hook_path: str = BASH_GUARD_HOOK) -> dict[str, Any]:
"""The grok hooks JSON installing the bash-guard as a blocking PreToolUse hook.
Matcher ``Bash`` covers grok's ``run_terminal_command`` alias too. The hook
runs with ``ROBOCO_GUARD_SKIP_GIT=1`` so it only blocks the exfil categories
(git ops stay on the graceful ``--deny`` rules); it denies via exit 2.
"""
return {
"hooks": {
"PreToolUse": [
{
"matcher": "Bash",
"hooks": [
{
"type": "command",
"command": hook_path,
"env": {"ROBOCO_GUARD_SKIP_GIT": "1"},
}
],
}
]
}
}
def write_grok_hooks(
*, hooks_dir: Path = GROK_HOOKS_DIR, hook_path: str = BASH_GUARD_HOOK
) -> bool:
"""Install the bash-guard PreToolUse hook into ``~/.grok/hooks/`` (best-effort).
Skips writing (returns False) when the guard script is absent, so a missing
hook never fails the render.
"""
if not Path(hook_path).is_file():
return False
hooks_dir.mkdir(parents=True, exist_ok=True)
(hooks_dir / "roboco-bash-guard.json").write_text(
json.dumps(bash_guard_hook_config(hook_path), indent=2), encoding="utf-8"
)
return True
def main() -> int: def main() -> int:
"""Entrypoint: write ``~/.grok/config.toml`` + AGENTS.md + the per-role args.""" """Entrypoint: write ``~/.grok/config.toml`` + AGENTS.md + hooks + per-role args."""
agent_id = os.environ.get("ROBOCO_AGENT_ID", "") agent_id = os.environ.get("ROBOCO_AGENT_ID", "")
mcp_path = os.environ.get("ROBOCO_MCP_CONFIG", "/app/mcp-config.json") mcp_path = os.environ.get("ROBOCO_MCP_CONFIG", "/app/mcp-config.json")
try: try:
@@ -223,6 +290,7 @@ def main() -> int:
render_config_toml(_load_mcp_config(mcp_path)), encoding="utf-8" render_config_toml(_load_mcp_config(mcp_path)), encoding="utf-8"
) )
write_agents_md() write_agents_md()
write_grok_hooks()
GROK_ARGS_PATH.write_text( GROK_ARGS_PATH.write_text(
"\n".join(grok_cli_args(agent_id, max_turns=max_turns)) + "\n", encoding="utf-8" "\n".join(grok_cli_args(agent_id, max_turns=max_turns)) + "\n", encoding="utf-8"
) )
@@ -2,6 +2,7 @@
from __future__ import annotations from __future__ import annotations
import json
import tomllib import tomllib
from typing import TYPE_CHECKING from typing import TYPE_CHECKING
@@ -91,20 +92,76 @@ def test_main_pm_keeps_shell_but_denies_git(monkeypatch: pytest.MonkeyPatch) ->
assert "run_terminal_cmd" not in dis # PM keeps a shell assert "run_terminal_cmd" not in dis # PM keeps a shell
assert "search_replace" in dis # but does not write code assert "search_replace" in dis # but does not write code
assert "Bash(git push*)" in args # raw git mutation denied assert "Bash(git push*)" in args # raw git mutation denied
assert args[args.index("--effort") + 1] == "low" # Parity with Claude: model-default reasoning for every role, no per-role cut.
assert "--effort" not in args
def test_effort_override(monkeypatch: pytest.MonkeyPatch) -> None: def test_prompter_allows_subagents_but_no_shell_or_edit(
monkeypatch: pytest.MonkeyPatch,
) -> None:
monkeypatch.delenv("ROBOCO_GROK_REASONING_EFFORT", raising=False)
dis = _disallowed(gc.grok_cli_args_for_role("prompter"))
# The intake interviewer may fan out to subagents (parity with Claude's Task)…
assert "Agent" not in dis
# …but it is still a read-only conversational role: no shell, no editing.
assert "run_terminal_cmd" in dis
assert "search_replace" in dis
def test_web_search_disabled_for_every_role() -> None:
for role in ("developer", "prompter", "secretary", "main_pm", "pr_reviewer"):
assert "--disable-web-search" in gc.grok_cli_args_for_role(role)
def test_effort_is_fleet_override_only(monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setenv("ROBOCO_GROK_REASONING_EFFORT", "high") monkeypatch.setenv("ROBOCO_GROK_REASONING_EFFORT", "high")
assert ( args = gc.grok_cli_args("be-dev-1")
gc.grok_cli_args("be-dev-1")[gc.grok_cli_args("be-dev-1").index("--effort") + 1] assert args[args.index("--effort") + 1] == "high"
== "high" # "full" / "default" / empty keep grok's model default (no --effort).
)
# "full" disables the per-role reduction entirely.
monkeypatch.setenv("ROBOCO_GROK_REASONING_EFFORT", "full") monkeypatch.setenv("ROBOCO_GROK_REASONING_EFFORT", "full")
assert "--effort" not in gc.grok_cli_args("main-pm") assert "--effort" not in gc.grok_cli_args("main-pm")
monkeypatch.delenv("ROBOCO_GROK_REASONING_EFFORT", raising=False)
assert "--effort" not in gc.grok_cli_args("documenter")
def test_max_turns_is_emitted() -> None: def test_max_turns_is_emitted() -> None:
args = gc.grok_cli_args("be-dev-1", max_turns=7) args = gc.grok_cli_args("be-dev-1", max_turns=7)
assert args[args.index("--max-turns") + 1] == "7" assert args[args.index("--max-turns") + 1] == "7"
def test_bash_roles_deny_the_full_git_mutation_set() -> None:
# Graceful native --deny rules (the agent recovers) covering the same git
# network / branch / history ops the Claude bash-guard blocks.
args = gc.grok_cli_args_for_role("developer")
for op in ("push", "fetch", "clone", "checkout", "merge", "rebase", "revert"):
assert f"Bash(git {op}*)" in args
assert "Bash(rm -rf*)" in args
def test_bash_guard_hook_config_skips_git() -> None:
handler = gc.bash_guard_hook_config("/app/scripts/bash-guard-hook.sh")[
"hooks"
]["PreToolUse"][0]
assert handler["matcher"] == "Bash"
inner = handler["hooks"][0]
assert inner["command"] == "/app/scripts/bash-guard-hook.sh"
# Git is handled by graceful --deny, so the hook skips it (exfil only).
assert inner["env"]["ROBOCO_GUARD_SKIP_GIT"] == "1"
def test_write_grok_hooks_installs_when_script_present(tmp_path: Path) -> None:
script = tmp_path / "bash-guard-hook.sh"
script.write_text("#!/bin/bash\nexit 0\n", encoding="utf-8")
hooks_dir = tmp_path / ".grok" / "hooks"
assert gc.write_grok_hooks(hooks_dir=hooks_dir, hook_path=str(script)) is True
written = json.loads((hooks_dir / "roboco-bash-guard.json").read_text())
assert written["hooks"]["PreToolUse"][0]["matcher"] == "Bash"
def test_write_grok_hooks_noops_when_script_absent(tmp_path: Path) -> None:
hooks_dir = tmp_path / ".grok" / "hooks"
assert (
gc.write_grok_hooks(hooks_dir=hooks_dir, hook_path=str(tmp_path / "nope.sh"))
is False
)
assert not hooks_dir.exists()