Files
DECNET/decnet/profiler/behave_shell/_features/environmental.py
anti cd7c7ea5a2 feat(profiler/behave_shell): emit environmental.keyboard_layout
ANTI authorised dropping the PII boundary for this primitive. ctx
gains typed_unigram_counts / typed_bigram_counts / typed_letter_count
populated during the existing single-pass input walk (paste-class
events excluded).

Two-axis classifier:
* layout-artefact unigrams take priority — q rate above floor with
  low English saturation → azerty; z above floor with y below → qwertz
* fallback to English-bigram saturation: ≥ floor → qwerty, else other

Sample-size floor 200 typed letters; bigram histogram capped at
top-64 to bound memory. Confidence cap stays moderate (0.40-0.55) —
heuristic discriminator.
2026-05-04 00:38:24 -04:00

300 lines
10 KiB
Python

"""``environmental.*`` feature functions.
Phase F ships the five environmental primitives plus F.0's shared
prompt-line detector. F.0 itself emits no primitive — it populates
``SessionContext.prompt_lines`` and ``Command.followed_by_prompt``
which F.1 / F.3 / E.4 read.
Step F.1: ``environmental.shell_type``.
Step F.2: ``environmental.terminal_multiplexer``.
Step F.3: ``environmental.locale``.
Step F.4: ``environmental.keyboard_layout``.
"""
from __future__ import annotations
import collections
import re
from typing import Iterator
from decnet_behave_core.spec.envelope import Observation
from decnet.profiler.behave_shell._ctx import SessionContext
from decnet.profiler.behave_shell._features._emit import make_observation
from decnet.profiler.behave_shell._parse import PromptLine, strip_ansi
from decnet.profiler.behave_shell._thresholds import (
LAYOUT_AZERTY_ENG_MAX,
LAYOUT_AZERTY_Q_MIN,
LAYOUT_MIN_TYPED_LETTERS,
LAYOUT_QWERTY_ENG_MIN,
LAYOUT_QWERTZ_Y_MAX,
LAYOUT_QWERTZ_Z_MIN,
LAYOUT_TOP_ENG_BIGRAMS,
LOCALE_MIN_VALUE_LENGTH,
SHELL_TYPE_MIN_PROMPTS,
)
# Multiplexer fingerprints scanned over RAW output (multiplexer escapes
# ARE ANSI sequences, so we must NOT strip-ANSI before searching).
# Sources:
# tmux DCS passthrough: ESC P tmux ;
# tmux focus reporting: ESC [ ? 1004 (set/reset)
# tmux window-title with explicit tmux marker
# screen DCS: ESC P =
# screen-specific OSC: ESC ] 83 ;
_TMUX_MARKERS: tuple[str, ...] = (
"\x1bPtmux;",
"\x1b[?1004",
"\x1b]2;tmux",
)
_SCREEN_MARKERS: tuple[str, ...] = (
"\x1bP=",
"\x1b]83;",
)
# Locale envvar regex: matches `KEY=VALUE` where KEY is one of the
# three locale envvars and VALUE is a POSIX locale name. The value
# pattern is intentionally restrictive — letters, underscore for the
# territory delimiter, optional codeset (.UTF-8 / .utf8), optional
# modifier (@euro). The trailing `(?=[\s'\"\\$]|$)` anchors the
# match against shell quoting and end-of-line.
_LOCALE_VALUE_RE = re.compile(
r"(?P<key>LC_ALL|LANG|LC_CTYPE)=(?P<val>[A-Za-z]{2,3}"
r"(?:_[A-Za-z]{2,3})?(?:\.[A-Za-z0-9-]+)?(?:@[A-Za-z0-9]+)?|C|POSIX)"
)
_LOCALE_KEY_PRIORITY: dict[str, int] = {"LC_ALL": 3, "LANG": 2, "LC_CTYPE": 1}
def _to_bcp47(posix_value: str) -> str | None:
"""Normalise a POSIX locale value to a BCP-47 tag.
Returns:
``None`` when the value is malformed (caller skips emission).
``"und"`` for ``C`` / ``POSIX`` (BCP-47 'undetermined').
Otherwise ``language-REGION`` (e.g. ``en-US``, ``pt-BR``);
codeset / modifier suffixes dropped (BCP-47 doesn't carry them).
"""
if posix_value in ("C", "POSIX"):
return "und"
base = posix_value.split(".", 1)[0].split("@", 1)[0]
parts = base.split("_")
if not parts or not parts[0].isalpha() or len(parts[0]) < 2:
return None
lang = parts[0].lower()
if len(parts) == 1:
return lang
region = parts[1]
if not region.isalpha():
return None
return f"{lang}-{region.upper()}"
def _classify_shell_from_prompt(p: PromptLine) -> str:
"""Map one prompt line to a shell-type label."""
suffix = p.suffix_char
line = p.raw_line
if suffix in ("$", "#"):
# bash / sh / dash all share these — collapsed to "bash" per
# registry's bash-family stance. zsh CAN be configured to use
# $/# but that's the user's PS1 override; default zsh is %.
return "bash"
if suffix == "%":
return "zsh"
if suffix == ">":
# Disambiguate by line content. powershell's PS1 starts with
# "PS "; cmd.exe's prompt typically contains a Windows path
# like "C:\". Everything else is fish.
if line.lstrip().startswith("PS "):
return "powershell"
if "C:\\" in line or "c:\\" in line:
return "cmd.exe"
return "fish"
return "bash" # defensive — _detect_prompt_suffix only emits one of $#%>
def shell_type(ctx: SessionContext) -> Iterator[Observation]:
"""Emit ``environmental.shell_type``.
Mode of per-prompt-line classification across
``ctx.prompt_lines``. Skip emission when no prompts detected —
the registry's enum doesn't admit ``unknown`` and emitting
``bash`` from no observation at all would be dishonest.
Confidence drops below ``SHELL_TYPE_MIN_PROMPTS`` (3 prompts);
above that threshold the vote is solid.
"""
if not ctx.prompt_lines:
return
votes = collections.Counter(
_classify_shell_from_prompt(p) for p in ctx.prompt_lines
)
value, _ = votes.most_common(1)[0]
if len(ctx.prompt_lines) < SHELL_TYPE_MIN_PROMPTS:
confidence = 0.40
else:
confidence = 0.75
yield make_observation(
ctx,
primitive="environmental.shell_type",
value=value,
confidence=confidence,
)
def terminal_multiplexer(ctx: SessionContext) -> Iterator[Observation]:
"""Emit ``environmental.terminal_multiplexer`` ∈ {none, tmux, screen}.
Scans raw output (NOT ANSI-stripped — multiplexer escapes ARE ANSI
sequences) for tmux/screen-specific fingerprints. If both detected,
prefer tmux (more common in 2026 nested-mux setups). Even one
escape is conclusive — no sample-size floor.
Confidence 0.85 when a fingerprint matches; 0.55 for ``none`` (a
bare PTY genuinely has no multiplexer, but a hidden multiplexer
that suppresses its escapes would also yield ``none``).
Skip emission when the session has no commands — without operator
interaction the engine should not emit operator-derived primitives.
The smoke gates (``test_extract_session_empty_stream_yields_no_observations``,
``test_extract_session_zero_inputs_yields_nothing``) bind this:
no commands, no observations.
"""
if not ctx.commands:
return
has_tmux = False
has_screen = False
for _t, _k, data in ctx.output_events:
if not has_tmux and any(m in data for m in _TMUX_MARKERS):
has_tmux = True
if not has_screen and any(m in data for m in _SCREEN_MARKERS):
has_screen = True
if has_tmux and has_screen:
break
if has_tmux:
value = "tmux"
confidence = 0.85
elif has_screen:
value = "screen"
confidence = 0.85
else:
value = "none"
confidence = 0.55
yield make_observation(
ctx,
primitive="environmental.terminal_multiplexer",
value=value,
confidence=confidence,
)
def locale(ctx: SessionContext) -> Iterator[Observation]:
"""Emit ``environmental.locale`` (free-string BCP-47 tag).
Searches the ANSI-stripped output stream for ``LANG=``,
``LC_ALL=``, or ``LC_CTYPE=`` substrings — emitted when the
operator runs ``env``, ``locale``, or ``printenv``. Highest-priority
key wins (``LC_ALL`` > ``LANG`` > ``LC_CTYPE``); the POSIX value is
normalised to BCP-47:
* ``en_US.UTF-8`` → ``en-US``
* ``pt_BR.UTF-8`` → ``pt-BR``
* ``C`` / ``POSIX`` → ``und``
* malformed → skip emission
Skip emission when no envvar dump is found in the output —
silence rather than fabricating a default.
"""
if not ctx.commands:
return
# Concatenate output; strip ANSI once (locale values aren't escape
# sequences themselves so the strip is safe).
raw = "".join(d for _t, _k, d in ctx.output_events)
if not raw:
return
text = strip_ansi(raw)
best_priority = 0
best_value: str | None = None
for m in _LOCALE_VALUE_RE.finditer(text):
prio = _LOCALE_KEY_PRIORITY[m.group("key")]
if prio <= best_priority:
continue
bcp47 = _to_bcp47(m.group("val"))
if bcp47 is None or len(bcp47) < LOCALE_MIN_VALUE_LENGTH:
continue
best_priority = prio
best_value = bcp47
if best_value is None:
return
yield make_observation(
ctx,
primitive="environmental.locale",
value=best_value,
confidence=0.80,
)
def keyboard_layout(ctx: SessionContext) -> Iterator[Observation]:
"""Emit ``environmental.keyboard_layout``.
Two independent signals over the typed-only character histograms:
1. **English-bigram saturation** — fraction of typed bigrams that
hit the top-10 English bigrams. High → presumed QWERTY.
2. **Layout-artefact unigrams** — letters that are rare in English
but frequent on operators using a different layout:
* ``q`` rate above floor AND English saturation low → ``azerty``
(AZERTY's `a` is on QWERTY's `q` position; mistypes bleed `q`)
* ``z`` rate above floor AND ``y`` rate below floor → ``qwertz``
(QWERTZ swaps `y`/`z`)
* Else: English saturation above floor → ``qwerty``
* Else: → ``other``
Threshold ordering matters — layout-artefact checks fire before
QWERTY because AZERTY/QWERTZ operators may still hit some English
bigrams.
Skip emission when typed letter count below
``LAYOUT_MIN_TYPED_LETTERS`` (200) — the histograms are too thin
to discriminate honestly.
"""
if ctx.typed_letter_count < LAYOUT_MIN_TYPED_LETTERS:
return
uni = ctx.typed_unigram_counts
bi = ctx.typed_bigram_counts
total_letters = ctx.typed_letter_count
total_bigrams = sum(bi.values())
eng_saturation = (
sum(bi.get(b, 0) for b in LAYOUT_TOP_ENG_BIGRAMS) / total_bigrams
if total_bigrams > 0 else 0.0
)
q_rate = uni.get("q", 0) / total_letters
z_rate = uni.get("z", 0) / total_letters
y_rate = uni.get("y", 0) / total_letters
if q_rate > LAYOUT_AZERTY_Q_MIN and eng_saturation < LAYOUT_AZERTY_ENG_MAX:
value = "azerty"
elif z_rate > LAYOUT_QWERTZ_Z_MIN and y_rate < LAYOUT_QWERTZ_Y_MAX:
value = "qwertz"
elif eng_saturation >= LAYOUT_QWERTY_ENG_MIN:
value = "qwerty"
else:
value = "other"
if total_letters < 500:
confidence = 0.40
else:
confidence = 0.55
yield make_observation(
ctx,
primitive="environmental.keyboard_layout",
value=value,
confidence=confidence,
)