fischer-agentkit/src/agentkit/core/phase.py

252 lines
10 KiB
Python

"""Phase state machine for PLAN_EXEC mode (G6, R24/R25).
Four sequential phases enforce per-step tool whitelists:
PLANNING → BUILDING → VERIFICATION → DELIVERY
KTD3 (Wave 3 plan): state machine lives in ReActEngine, not skill config.
KTD5: default whitelist matches brainstorm R24 (Planning: think/search;
Building: write_file; etc.).
KTD6: transitions are LLM-driven via AdvancePhaseTool; auto-advance is opt-in.
U3 (R3): ``default_policy()`` accepts an optional ``workspace_root`` and
populates ``PhasePolicy.verification_commands`` via coding-task detection
(``pyproject.toml`` / ``.py`` presence) — coding tasks force pytest/ruff;
non-coding tasks leave the list empty for Spec-declared commands.
"""
from __future__ import annotations
import enum
import logging
import re
from dataclasses import dataclass, field, replace
from pathlib import Path
from typing import Any, Callable
from agentkit.tools.shell import ShellTool
logger = logging.getLogger(__name__)
class PhaseState(enum.Enum):
"""Phases of the SOLO state machine (extends ExecutionMode.PLAN_EXEC)."""
PLANNING = "planning"
BUILDING = "building"
VERIFICATION = "verification"
DELIVERY = "delivery"
@classmethod
def next_of(cls, current: "PhaseState") -> "PhaseState | None":
"""Return the phase after `current`, or None if `current` is the last."""
order = [cls.PLANNING, cls.BUILDING, cls.VERIFICATION, cls.DELIVERY]
try:
idx = order.index(current)
except ValueError:
return None
if idx + 1 >= len(order):
return None
return order[idx + 1]
@classmethod
def from_string(cls, value: str) -> "PhaseState":
"""Parse from string (case-insensitive). Raises ValueError on unknown."""
try:
return cls(value.lower())
except ValueError as e:
valid = ", ".join(p.value for p in cls)
raise ValueError(f"Invalid phase name {value!r}. Valid: {valid}") from e
# Wildcard token meaning "all tools allowed in this phase".
WILDCARD = "*"
@dataclass(slots=True)
class PhasePolicy:
"""Per-phase tool whitelist + bash command filter for PLAN_EXEC mode.
The policy is enforced by ReActEngine._execute_loop before each tool
dispatch. A tool not in the current phase's whitelist is rejected with
a structured error returned to the LLM (the loop continues — the LLM
gets to react to the rejection and either switch tools or call
AdvancePhaseTool).
Wildcard ``"*"`` in a phase's whitelist means "all tools allowed"
(used by DELIVERY by default).
`bash_command_filter` values accept either:
- `Callable[[str], bool]`: returns True if the command is dangerous
(matches `ShellTool._is_dangerous` semantics); allowed = not dangerous.
- `re.Pattern`: pattern matches dangerous substrings; allowed = no match.
Kept for backward compat with Wave 3 configs.
- `None`: no restriction for this phase.
"""
whitelist: dict[PhaseState, frozenset[str]]
bash_command_filter: dict[PhaseState, Callable[[str], bool] | re.Pattern | None] = field(
default_factory=dict
)
auto_advance_after_steps: int | None = None # None = manual (LLM calls advance_phase)
start_phase: PhaseState = PhaseState.PLANNING
# U3/R3: verification commands to run at the VERIFICATION phase's final-answer
# point. Populated by default_policy() via coding-task detection. None = no
# opinion (ReActEngine falls back to its own verification_commands param or
# VerificationLoop defaults). An empty list means "no commands" (verification
# passes trivially — for non-coding tasks using Spec-declared commands instead).
verification_commands: list[str] | None = None
# U4/R11: total step budget for the plan (sum of think+verify+reflect).
# None = use ReActEngine's max_steps. Provides a checkpoint-reconstructable
# record of the plan's total step budget (KTD-7).
step_budget: int | None = None
def __post_init__(self) -> None:
# Fail-fast: empty whitelist for a non-wildcard phase = bug.
for phase, tools in self.whitelist.items():
if not tools:
raise ValueError(
f"Phase {phase.value!r} has an empty whitelist — set ['*'] for "
f"'all tools allowed' or list specific tool names."
)
def is_tool_allowed(self, tool_name: str, phase: PhaseState) -> bool:
"""Return True if `tool_name` is allowed in `phase`."""
allowed = self.whitelist.get(phase, frozenset())
if WILDCARD in allowed:
return True
return tool_name in allowed
def is_bash_command_allowed(self, command: str, phase: PhaseState) -> bool:
"""Return True if `command` passes the bash filter for `phase`.
A None filter = no restriction. An empty command is allowed (ShellTool
separately rejects empty commands) — short-circuited here so the
ShellTool path emits a clearer "empty command" error instead of a
phase-violation noise injected back to the LLM.
"""
if not command:
return True
filter_value = self.bash_command_filter.get(phase)
if filter_value is None:
return True
if callable(filter_value):
# Callable contract: returns True if dangerous.
return not filter_value(command)
# re.Pattern contract: search() returns a Match if dangerous.
return not filter_value.search(command)
def to_dict(self) -> dict[str, Any]:
"""Serialize for logging/telemetry. Not round-trippable (regex/callable → str)."""
return {
"whitelist": {phase.value: sorted(tools) for phase, tools in self.whitelist.items()},
"bash_command_filter": {
phase.value: ("<callable>" if callable(p) else (p.pattern if p else None))
for phase, p in self.bash_command_filter.items()
},
"auto_advance_after_steps": self.auto_advance_after_steps,
"start_phase": self.start_phase.value,
"verification_commands": self.verification_commands,
"step_budget": self.step_budget,
}
def default_policy(workspace_root: str | Path | None = None) -> PhasePolicy:
"""Return the KTD5 default PhasePolicy.
Whitelist (R24):
- PLANNING: search, tool_search, read_file, shell (read-only)
- BUILDING: write_file, shell (full), read_file, search
- VERIFICATION: shell (test commands), read_file, search
- DELIVERY: all tools (wildcard)
Bash filter:
- PLANNING/VERIFICATION: reuse `ShellTool._is_dangerous` (Wave 4 U1).
Closes the regex ceiling — catches `:>file`, `dd of=/dev/sda`, chain
operators, and the full danger taxonomy shared with the ShellTool
confirmation path.
- BUILDING/DELIVERY: no filter (full bash)
U3/R3: ``verification_commands`` is populated via coding-task detection on
``workspace_root``. Coding workspaces (``pyproject.toml`` or ``.py``
present) force ``pytest -x -q`` and ``ruff check src/``. Non-coding
workspaces get ``None`` (no opinion — Spec-declared commands are used).
"""
# U3/R3: coding-task detection. Local import avoids a circular dependency
# (sandbox.py is standalone, but keeping the import local makes the R3
# concern visually scoped to default_policy).
from agentkit.core.sandbox import detect_verification_commands
verification_cmds = detect_verification_commands(workspace_root)
# detect_verification_commands returns [] for non-coding workspaces.
# For non-coding workspaces, leave verification_commands as None so the
# caller knows "no coding-specific commands" and can substitute Spec-declared
# commands. For coding workspaces, set the forced pytest/ruff list.
return PhasePolicy(
whitelist={
# Tool name is "shell" (ShellTool default); bash_command_filter
# gates on the same name. Using "bash" here would make the filter
# dead code and block the LLM from shell access.
PhaseState.PLANNING: frozenset({"search", "tool_search", "read_file", "shell"}),
PhaseState.BUILDING: frozenset(
{"write_file", "shell", "read_file", "search", "tool_search"}
),
PhaseState.VERIFICATION: frozenset({"shell", "read_file", "search"}),
PhaseState.DELIVERY: frozenset({WILDCARD}),
},
bash_command_filter={
PhaseState.PLANNING: ShellTool._is_dangerous,
PhaseState.VERIFICATION: ShellTool._is_dangerous,
PhaseState.BUILDING: None,
PhaseState.DELIVERY: None,
},
auto_advance_after_steps=None, # manual by default
start_phase=PhaseState.PLANNING,
verification_commands=verification_cmds if verification_cmds else None,
)
def policy_from_config(config: dict[str, Any]) -> PhasePolicy | None:
"""Build a PhasePolicy from the `plan_exec` config section.
Returns None if `config` is empty or `enabled` is False (opt-out).
Config shape:
plan_exec:
enabled: true # default true if section present
auto_advance_after_steps: 5 # optional
start_phase: planning # optional, default planning
whitelist_override: # optional, merges with default
planning: [search, read_file]
building: [write_file, bash]
"""
if not config:
return None
if config.get("enabled", True) is False:
return None
policy = default_policy()
# Start phase
start_phase_str = config.get("start_phase")
if start_phase_str:
policy = replace(policy, start_phase=PhaseState.from_string(start_phase_str))
# Auto-advance override
if "auto_advance_after_steps" in config:
policy = replace(policy, auto_advance_after_steps=config["auto_advance_after_steps"])
# Whitelist override — merge with default (override wins on conflict)
override = config.get("whitelist_override") or {}
if override:
new_whitelist = dict(policy.whitelist)
for phase_name, tools in override.items():
phase = PhaseState.from_string(phase_name)
if not isinstance(tools, list):
raise ValueError(
f"whitelist_override[{phase_name!r}] must be a list, got {type(tools).__name__}"
)
new_whitelist[phase] = frozenset(str(t) for t in tools)
policy = replace(policy, whitelist=new_whitelist)
return policy