anti
07ff5ff0c9
Per-prompt classification mode over ctx.prompt_lines. $/# → bash; % → zsh; > with 'PS ' prefix → powershell; > with 'C:\' substring → cmd.exe; > otherwise → fish. New _features/environmental.py module opens Phase F.
324 lines
17 KiB
Python
324 lines
17 KiB
Python
"""Numeric thresholds for BEHAVE-SHELL primitive classification.
|
|
|
|
Each constant cites its calibration source. When the registry's
|
|
``notes:`` field disagrees with a constant here, the registry is
|
|
authoritative — fix the constant, re-run the calibration grid.
|
|
|
|
Empirical thresholds inherited from the BEHAVE prototype extractor
|
|
(``BEHAVE/prototype_extractors/shell/extract.py``); see lines 40-90 of
|
|
that file for the calibration history. Any change here must keep the
|
|
five-class grid green.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
# ── paste-burst detection (Step 1) ──────────────────────────────────────────
|
|
# A single input event with ≥ PASTE_MIN_CHARS_PER_EVENT chars is the
|
|
# paste-class proxy used by the prototype; xterm-kitty / iTerm / VS Code
|
|
# pastes arrive as one bulk write.
|
|
PASTE_MIN_CHARS_PER_EVENT: int = 4
|
|
|
|
# Consecutive paste-class events arriving within this IAT collapse into
|
|
# one PasteBurst record. 200ms is the prototype's IKI burst cap.
|
|
PASTE_BURST_MAX_IAT_S: float = 0.20
|
|
|
|
# ── motor.input_modality (Step 2) ───────────────────────────────────────────
|
|
# Paste-event ratio thresholds. ≥ 40% paste events → "pasted" (LLM-driven);
|
|
# ≤ 5% → "typed" (human at the keyboard); in between → "mixed".
|
|
# Lowered from 0.5 after the 47.6% case in sessions-2026-05-02-with-llm.jsonl
|
|
# was clearly LLM-driven but missed the 0.5 floor.
|
|
MODALITY_PASTED_MIN: float = 0.40
|
|
MODALITY_TYPED_MAX: float = 0.05
|
|
|
|
# ── motor.paste_burst_rate (Step 3) ─────────────────────────────────────────
|
|
# Same paste-event ratio re-bucketed for the "how often does the operator
|
|
# paste" axis. Coarser than input_modality on purpose: this primitive is the
|
|
# habit signal, input_modality is the dominant-channel signal.
|
|
PASTE_RATE_HABITUAL_MIN: float = 0.50
|
|
PASTE_RATE_OCCASIONAL_MIN: float = 0.10
|
|
|
|
# ── cognitive.inter_command_latency_class (Step 5) ──────────────────────────
|
|
# Bucket edges (seconds) for the median inter-command IAT. Prototype
|
|
# values; v0.2 splits the original llm_roundtrip 2-8s band into
|
|
# llm_lightweight (orchestrated agents w/ small models / terse prompts) and
|
|
# llm_heavyweight (reasoning-class agents in tool loops with text
|
|
# generation between calls). Empirical anchor: Claude Opus driving recon
|
|
# via tmux send-keys produced a median of 15.5s.
|
|
INTER_CMD_INSTANT_MAX: float = 0.30
|
|
INTER_CMD_TYPING_MAX: float = 1.50
|
|
INTER_CMD_DELIBERATE_MAX: float = 2.00
|
|
INTER_CMD_LLM_LIGHTWEIGHT_MAX: float = 8.00
|
|
INTER_CMD_LLM_HEAVYWEIGHT_MAX: float = 30.00
|
|
|
|
# Sample-size floor for inter-command IAT primitives. Below this we
|
|
# halve the confidence per BEHAVE-EXTRACTOR.md "sample-size honesty".
|
|
MIN_COMMANDS_FOR_FULL_CONFIDENCE: int = 5
|
|
|
|
# ── cognitive.command_branch_diversity (Step 6) ─────────────────────────────
|
|
# unique_first_tokens / total_commands ratio. Prototype's empirical
|
|
# split (sessions-2026-05-02-* corpus): CLAUDE-CL chasing one finding
|
|
# ≈ 0.55-0.60 (adaptive), HUMAN exploring filesystem ≈ 0.65-0.70
|
|
# (adaptive), YOU-sim / CLAUDE-FF scripted recon ≈ 0.75+ (linear).
|
|
BRANCH_DIVERSITY_LINEAR_MIN: float = 0.70 # >= → linear_playbook
|
|
|
|
# ── cognitive.feedback_loop_engagement (Step 7) ─────────────────────────────
|
|
# Pearson r threshold for "the operator's pause grew with the volume of
|
|
# preceding output". |r| > this → significant; sign carries direction.
|
|
FEEDBACK_CORRELATION_MIN: float = 0.30
|
|
# Need at least this many (output_bytes, next_pause) pairs to even
|
|
# attempt a correlation. Below this the answer is "unknown".
|
|
FEEDBACK_MIN_PAIRS: int = 5
|
|
|
|
# ── cognitive.inter_command_consistency (Step 8) ────────────────────────────
|
|
# CV (stdev / mean) of inter-command IATs. Empirical (this corpus):
|
|
# human session CV=0.94 → variable; LLM-simulated CV=0.24 → metronomic;
|
|
# anything beyond 1.5 is heuristically "bimodal" (real bimodal detection
|
|
# via Hartigan dip is filed for v0.2).
|
|
PAUSE_CV_METRONOMIC_MAX: float = 0.40
|
|
PAUSE_CV_BIMODAL_MIN: float = 1.50
|
|
|
|
# ── output error-signal helper (Step D.0) ──────────────────────────────────
|
|
# The canonical bash/sh error fingerprints live in ``_parse.py`` as
|
|
# ``_OUTPUT_ERROR_PATTERNS`` (compiled regexes). They're not threshold
|
|
# numbers, so they live next to the helper that uses them rather than
|
|
# here. This v0.1 heuristic will be subsumed by Phase F.0's prompt
|
|
# parser (PS1 echo + exit-code sniff), at which point this comment and
|
|
# the patterns block move to ``_parse.py``'s prompt section. Until then,
|
|
# any drift in registry value definitions for ``error_resilience.*`` or
|
|
# ``cognitive_load`` must be reflected by editing the patterns tuple
|
|
# (not a constant, so no boundary-band logic applies).
|
|
|
|
# ── cognitive.cognitive_load (Step D.1) ─────────────────────────────────────
|
|
# Composite ∈ [0, 1] over three sub-signals (each clipped to [0, 1]):
|
|
#
|
|
# A = chunking_load = median_intra_cmd_cv / CHUNKING_REF_CV
|
|
# B = error_load = errored_cmds / total_cmds
|
|
# C = pace_variability_load = (stdev / mean of inter_cmd_iats) / PACE_REF_CV
|
|
#
|
|
# load = mean(A, B, C); bucket:
|
|
# load < COGNITIVE_LOAD_LOW_MAX → low
|
|
# load < COGNITIVE_LOAD_MEDIUM_MAX → medium
|
|
# else → high
|
|
#
|
|
# v0.1 thresholds — D.8 re-tunes once D.1-D.7 are stable. The reference
|
|
# CVs (CHUNKING_REF_CV / PACE_REF_CV) are the value at which that single
|
|
# component saturates to a load contribution of 1.0; anything past
|
|
# saturates the term but doesn't double-count.
|
|
COGNITIVE_LOAD_CHUNKING_REF_CV: float = 1.00
|
|
COGNITIVE_LOAD_PACE_REF_CV: float = 1.50
|
|
COGNITIVE_LOAD_LOW_MAX: float = 0.33
|
|
COGNITIVE_LOAD_MEDIUM_MAX: float = 0.67
|
|
|
|
# ── cognitive.exploration_style (Step D.2) ─────────────────────────────────
|
|
# Two-axis classification over the first_token_hash sequence:
|
|
#
|
|
# repetition_rate (R) = 1 - (unique_first_tokens / total_commands)
|
|
# backtrack_rate (J) = transitions where commands[i+1].first_token_hash
|
|
# appeared anywhere in commands[0..i-1] but is NOT
|
|
# equal to commands[i].first_token_hash (jumping
|
|
# back to an older tool, not just repeating).
|
|
#
|
|
# J >= EXPLORATION_CHAOTIC_BACKTRACK_MIN → chaotic
|
|
# else if R >= EXPLORATION_TARGETED_REP_MIN → targeted
|
|
# else → methodical
|
|
#
|
|
# Methodical = low repetition, low backtracks (linear progression through
|
|
# novel tools). Targeted = high repetition (drilling the same tool).
|
|
# Chaotic = jumping between prior tools without a clear thread.
|
|
# v0.1; D.8 re-tunes.
|
|
EXPLORATION_TARGETED_REP_MIN: float = 0.50
|
|
EXPLORATION_CHAOTIC_BACKTRACK_MIN: float = 0.30
|
|
|
|
# ── cognitive.planning_depth (Step D.3) ────────────────────────────────────
|
|
# Distribution of inter-command IATs.
|
|
# deep_pause_fraction = (count of inter_cmd_iats > IKI_THINK_MAX_S) / N
|
|
# reactive_pause_fraction = (count of inter_cmd_iats <= INTER_CMD_INSTANT_MAX) / N
|
|
#
|
|
# deep_pause_fraction >= PLANNING_DEEP_MIN → deep
|
|
# reactive_pause_fraction >= PLANNING_REACTIVE_MIN → reactive
|
|
# otherwise → shallow
|
|
#
|
|
# v0.1; D.8 re-tunes once D.1-D.7 are stable.
|
|
PLANNING_DEEP_MIN: float = 0.40
|
|
PLANNING_REACTIVE_MIN: float = 0.50
|
|
|
|
# ── cognitive.tool_vocabulary (Step D.4) ───────────────────────────────────
|
|
# Absolute count of distinct first_token_hashes across the session.
|
|
#
|
|
# distinct <= TOOL_VOCAB_NARROW_MAX → narrow
|
|
# distinct >= TOOL_VOCAB_BROAD_MIN → broad
|
|
# otherwise → moderate
|
|
#
|
|
# Absolute, not normalised. A 3-command session with 3 unique tools is
|
|
# ``narrow`` not ``broad`` — the operator simply hasn't shown range yet.
|
|
# Sample-size honesty drops confidence below MIN_COMMANDS_FOR_FULL_CONFIDENCE.
|
|
# v0.1; D.8 re-tunes.
|
|
TOOL_VOCAB_NARROW_MAX: int = 3
|
|
TOOL_VOCAB_BROAD_MIN: int = 10
|
|
|
|
# ── cognitive.error_resilience.frustration_typing (Step D.6) ───────────────
|
|
# Compare the median within-command IAT of commands *following* an
|
|
# errored command against the same statistic for commands following a
|
|
# successful command. The relative absolute delta:
|
|
#
|
|
# delta = |median_post_error - median_post_success| / median_post_success
|
|
#
|
|
# delta < FRUSTRATION_LOW_MAX → low
|
|
# delta < FRUSTRATION_MODERATE_MAX → moderate
|
|
# else → high
|
|
#
|
|
# v0.1; D.8 re-tunes.
|
|
FRUSTRATION_LOW_MAX: float = 0.10
|
|
FRUSTRATION_MODERATE_MAX: float = 0.30
|
|
|
|
# ── temporal.session_duration (Step E.1) ───────────────────────────────────
|
|
# Bucket edges (seconds) for ``ctx.duration_s``:
|
|
#
|
|
# duration_s < SESSION_DURATION_SHORT_MAX → short
|
|
# duration_s < SESSION_DURATION_MEDIUM_MAX → medium
|
|
# duration_s < SESSION_DURATION_LONG_MAX → long
|
|
# else → marathon
|
|
#
|
|
# 60s / 600s / 3600s are the BEHAVE-EXTRACTOR.md defaults; D.8-equivalent
|
|
# re-tune for E lands when calibration corpus is run.
|
|
SESSION_DURATION_SHORT_MAX: float = 60.0
|
|
SESSION_DURATION_MEDIUM_MAX: float = 600.0
|
|
SESSION_DURATION_LONG_MAX: float = 3600.0
|
|
|
|
# ── temporal.escalation_pattern (Step E.2) ─────────────────────────────────
|
|
# Bin commands into non-overlapping windows. Width is dynamic:
|
|
#
|
|
# width = max(ESCALATION_WINDOW_MIN_S, duration_s / ESCALATION_WINDOW_TARGET)
|
|
#
|
|
# so a 30s session uses 10s windows (3 windows) and a 1h session uses
|
|
# 6min windows (10 windows). CV of per-window counts + zero-window
|
|
# fraction classify:
|
|
#
|
|
# zero_frac >= ESCALATION_BURSTY_ZERO_FRAC AND CV >= ESCALATION_BURSTY_CV
|
|
# → bursty (silences then spikes)
|
|
# CV < ESCALATION_SUSTAINED_CV
|
|
# → sustained (steady cadence throughout)
|
|
# else
|
|
# → erratic (variable but no real silence pattern)
|
|
#
|
|
# v0.1; corpus re-tune deferred. Sample-size honesty caps confidence
|
|
# below ESCALATION_MIN_WINDOWS or ESCALATION_MIN_COMMANDS.
|
|
ESCALATION_WINDOW_MIN_S: float = 10.0
|
|
ESCALATION_WINDOW_TARGET: int = 10
|
|
ESCALATION_BURSTY_ZERO_FRAC: float = 0.30
|
|
ESCALATION_BURSTY_CV: float = 1.00
|
|
ESCALATION_SUSTAINED_CV: float = 0.50
|
|
ESCALATION_MIN_WINDOWS: int = 5
|
|
ESCALATION_MIN_COMMANDS: int = 5
|
|
|
|
# ── temporal.lifecycle_markers.landing_ritual (Step E.3) ──────────────────
|
|
# How many of the first ``LANDING_RITUAL_FIRST_N`` commands must hit
|
|
# the recon-token set (uname / id / whoami / pwd / hostname / w / who)
|
|
# for the session to count as having a landing ritual.
|
|
LANDING_RITUAL_FIRST_N: int = 5
|
|
LANDING_RITUAL_HIT_MIN: int = 2
|
|
LANDING_RITUAL_MIN_COMMANDS: int = 3
|
|
|
|
# ── F.0 prompt-line detector ──────────────────────────────────────────────
|
|
# A prompt line in the output stream ends with one of these characters
|
|
# followed by a space or EOL. ``$`` and ``#`` are sh/bash; ``%`` is zsh;
|
|
# ``>`` is fish / cmd.exe / powershell (disambiguated by line content
|
|
# at F.1 time). Capped at 256 chars to bound memory; ANTI authorised
|
|
# retaining PS1 text on ctx (PII relaxation), but a malicious operator
|
|
# inflating the prompt buffer is still bounded.
|
|
PROMPT_SUFFIX_CHARS: frozenset[str] = frozenset({"$", "#", "%", ">"})
|
|
PROMPT_LINE_MAX_CHARS: int = 256
|
|
|
|
# ── environmental.shell_type (Step F.1) ────────────────────────────────────
|
|
# Below this many detected prompt-lines, drop confidence (sample-size
|
|
# honesty). Above, the shell-type vote is robust.
|
|
SHELL_TYPE_MIN_PROMPTS: int = 3
|
|
|
|
# ── motor.keystroke_cadence (Step B.1) ──────────────────────────────────────
|
|
# Typing bursts split at gaps > IKI_THINK_MAX_S so think-pauses between
|
|
# commands don't inflate the within-burst CV. Mirrors the prototype's
|
|
# _split_into_bursts (BEHAVE/prototype_extractors/shell/extract.py:275-286).
|
|
IKI_THINK_MAX_S: float = 1.50
|
|
# Sub-human floor for the "machine" classification — only paired with a
|
|
# pathologically uniform CV, since real humans never produce sub-5ms IATs
|
|
# in a sustained burst.
|
|
IKI_MACHINE_MAX_S: float = 0.005
|
|
CV_MACHINE_MAX: float = 0.05
|
|
CV_STEADY_MAX: float = 0.50
|
|
CV_BURSTY_MAX: float = 1.50
|
|
# Need this many input events before we'll claim a cadence at all.
|
|
MIN_INPUTS_FOR_CADENCE: int = 5
|
|
|
|
# ── motor.motor_stability (Step B.2) ────────────────────────────────────────
|
|
# Tremor proxy: fraction of within-burst IATs below TREMOR_FAST_FLOOR_S
|
|
# (30 ms — physiologically implausible double-press floor; humans can't
|
|
# reliably produce IATs below ~50 ms in sustained typing). High rate
|
|
# of sub-floor IATs flags double-press / motor twitch / stuck-key.
|
|
TREMOR_FAST_FLOOR_S: float = 0.030
|
|
TREMOR_RATE_MIN: float = 0.10 # ≥10% sub-floor → tremor
|
|
|
|
# ── motor.error_correction (Step B.3) ───────────────────────────────────────
|
|
# Backspace within this many seconds of the preceding key = "caught the
|
|
# typo mid-keystroke" (immediate). Beyond this = the operator paused,
|
|
# noticed, then went back (deferred).
|
|
BACKSPACE_IMMEDIATE_MAX_S: float = 0.50
|
|
|
|
# ── motor.command_chunking (Step B.4) ───────────────────────────────────────
|
|
# Median CV of within-command IATs. Below this → fluent (steady within
|
|
# each command); above → fragmented (operator pauses mid-command).
|
|
CMD_CHUNKING_FLUENT_CV_MAX: float = 0.50
|
|
|
|
# ── motor.shell_mastery.* (Phase C) ─────────────────────────────────────────
|
|
# Readline control bytes counted toward ``shortcut_usage``. The seven
|
|
# pinned by BEHAVE-EXTRACTOR.md §Phase C (line 472):
|
|
# ^A start-of-line ^E end-of-line ^W kill-prev-word
|
|
# ^U kill-line ^R reverse-i-search ^B back-char ^F forward-char
|
|
# v0.2 may extend to ^K/^Y/^L/^D/^P/^N once corpus calibration justifies it.
|
|
# Note: ^U / ^W also feed ``motor.error_correction`` (Step B.3) via the
|
|
# ``kill_line_count`` channel — these are independent measurements over
|
|
# the same byte stream, not double-counting.
|
|
SHORTCUT_CTRL_BYTES: frozenset[str] = frozenset({
|
|
"\x01", "\x05", "\x17", "\x15", "\x12", "\x02", "\x06",
|
|
})
|
|
|
|
# motor.shell_mastery.tab_completion — fraction of commands containing
|
|
# at least one ``\t`` keystroke. Registry buckets per BEHAVE-EXTRACTOR.md
|
|
# line 471: ``none`` (0%), ``occasional`` (<30%), ``habitual`` (≥50%).
|
|
# The 30%-50% gap rounds down to ``occasional`` — the registry's own gap.
|
|
TAB_COMPLETION_OCCASIONAL_MAX: float = 0.30
|
|
TAB_COMPLETION_HABITUAL_MIN: float = 0.50
|
|
|
|
# motor.shell_mastery.shortcut_usage — total readline ctrl-byte
|
|
# keystrokes per command. Registry buckets are qualitative
|
|
# (``none / moderate / heavy``); v0.1 thresholds are best-guesses
|
|
# pinned for five-class corpus calibration. Re-tune once HUMAN /
|
|
# YOU-sim / LW-sim / CLAUDE-FF / CLAUDE-CL data lands.
|
|
# 0/cmd → none
|
|
# <0.05/cmd → none (counted shortcuts but rare; rounds down)
|
|
# 0.05-0.30 → moderate
|
|
# ≥0.30/cmd → heavy
|
|
SHORTCUT_USAGE_MODERATE_MIN: float = 0.05
|
|
SHORTCUT_USAGE_HEAVY_MIN: float = 0.30
|
|
|
|
# motor.shell_mastery.pipe_chaining_depth — median ``|`` count across
|
|
# commands. Pipes are counted on every byte (typed AND pasted) — a
|
|
# pasted pipeline still indicates pipeline fluency the operator chose
|
|
# to execute. Registry buckets per BEHAVE-EXTRACTOR.md line 473:
|
|
# median ≤ 1 → shallow (no pipeline at all, or one stage)
|
|
# median == 2 → moderate
|
|
# median ≥ 3 → deep
|
|
# Median is integer-valued (sum of ints over commands), so the
|
|
# boundaries here are integer step boundaries; the proximity-band
|
|
# logic uses integer equality.
|
|
PIPE_CHAINING_MODERATE_MEDIAN: int = 2
|
|
PIPE_CHAINING_DEEP_MEDIAN: int = 3
|
|
|
|
# Sample-size floor below which Phase C primitives drop confidence to
|
|
# 0.40 (sample-size honesty). Mirrors MIN_COMMANDS_FOR_FULL_CONFIDENCE
|
|
# but is named separately so a future tune can move them independently.
|
|
SHELL_MASTERY_MIN_COMMANDS: int = 5
|
|
|
|
# Width of the "near a bucket boundary" band (relative to the boundary)
|
|
# used by Phase C primitives. ±10% of the boundary value drops
|
|
# confidence by 0.20 per BEHAVE-EXTRACTOR.md §"Threshold proximity".
|
|
SHELL_MASTERY_BOUNDARY_BAND: float = 0.10
|