feat(profiler/behave_shell): emit cognitive.error_resilience.retry_tactic

Modal response across Command.errored=True commands:
* same first_token_hash on next command → rerun
* different first_token_hash         → switch
* no next command                    → abort
Tiebreak in registry order. The fourth registry value 'modify'
requires within-command arg diffing (PII boundary); deferred to v0.2.
This commit is contained in:
2026-05-03 23:58:00 -04:00
parent f286c84d95
commit b704352783
3 changed files with 140 additions and 0 deletions

View File

@@ -14,6 +14,7 @@ from decnet.profiler.behave_shell._ctx import SessionContext
from decnet.profiler.behave_shell._features.cognitive import (
cognitive_load,
command_branch_diversity,
error_resilience_retry_tactic,
exploration_style,
feedback_loop_engagement,
planning_depth,
@@ -53,4 +54,5 @@ FEATURES: tuple[FeatureFn, ...] = (
exploration_style,
planning_depth,
tool_vocabulary,
error_resilience_retry_tactic,
)

View File

@@ -186,6 +186,59 @@ def feedback_loop_engagement(ctx: SessionContext) -> Iterator[Observation]:
)
def error_resilience_retry_tactic(ctx: SessionContext) -> Iterator[Observation]:
"""Emit ``cognitive.error_resilience.retry_tactic``.
For each command with ``Command.errored=True``, classify the
operator's response by the *next* command:
* **rerun** — same first_token_hash as the errored command. The
operator re-invoked the same tool (often after fixing args
mid-edit, but we can't see args).
* **switch** — different first_token_hash. Pivoted to a different
tool.
* **abort** — no next command. Session ended after the error.
The session's reported tactic is the **modal** response across all
errored commands (with ties broken in registry order: rerun >
modify > switch > abort). Skip emission entirely when no commands
errored — the registry has no ``unknown`` here, and silence is the
most honest answer.
The ``modify`` value (edit-and-retry) requires within-command
diffing of arg tokens, which crosses the PII boundary the engine
holds (only ``first_token_hash`` is retained per command). v0.1
therefore never emits ``modify``; v0.2 will once the PII trade-off
is revisited against a real attacker corpus.
"""
errored = [(i, c) for i, c in enumerate(ctx.commands) if c.errored]
if not errored:
return
counts = {"rerun": 0, "switch": 0, "abort": 0}
for i, cmd in errored:
if i + 1 >= len(ctx.commands):
counts["abort"] += 1
elif ctx.commands[i + 1].first_token_hash == cmd.first_token_hash:
counts["rerun"] += 1
else:
counts["switch"] += 1
# Registry-order tiebreak (rerun > modify > switch > abort).
# `modify` deferred — never increments here.
order = ("rerun", "switch", "abort")
value = max(order, key=lambda k: counts[k])
if len(errored) < MIN_COMMANDS_FOR_FULL_CONFIDENCE:
confidence = 0.40
else:
confidence = 0.65
yield make_observation(
ctx,
primitive="cognitive.error_resilience.retry_tactic",
value=value,
confidence=confidence,
)
def tool_vocabulary(ctx: SessionContext) -> Iterator[Observation]:
"""Emit ``cognitive.tool_vocabulary`` ∈ {narrow, moderate, broad}.

View File

@@ -0,0 +1,85 @@
"""Step D.5: ``cognitive.error_resilience.retry_tactic``."""
from __future__ import annotations
from decnet.profiler.behave_shell import extract_session
from decnet.profiler.behave_shell._parse import AsciinemaEvent
PRIMITIVE = "cognitive.error_resilience.retry_tactic"
def _of(observations: list, primitive: str):
obs = [o for o in observations if o.primitive == primitive]
assert len(obs) == 1, f"expected exactly one {primitive}, got {len(obs)}"
return obs[0]
def _typed(text: str, t0: float = 0.0, dt: float = 0.05) -> list[AsciinemaEvent]:
return [(t0 + i * dt, "i", c) for i, c in enumerate(text)]
def _err_then(token: str, next_token: str | None, t0: float = 0.0) -> list[AsciinemaEvent]:
"""``token`` errors; ``next_token`` is the operator's response (or None).
Output event lands after the ``\\r`` so it falls inside the
command's post-execution window.
"""
events: list[AsciinemaEvent] = []
events.extend(_typed(f"{token}\r", t0=t0))
cmd_end = t0 + len(token) * 0.05 # \r is the last char
events.append((cmd_end + 0.10, "o", f"bash: {token}: command not found\n"))
if next_token is not None:
events.extend(_typed(f"{next_token}\r", t0=t0 + 1.5))
return events
def test_no_errors_no_emission() -> None:
events: list[AsciinemaEvent] = _typed("ls\r") + [(0.5, "o", "file1\n")]
out = list(extract_session(events, sid="rt-noerr"))
assert [o for o in out if o.primitive == PRIMITIVE] == []
def test_majority_rerun_emits_rerun() -> None:
"""Operator re-invokes the same tool after each error → rerun."""
events: list[AsciinemaEvent] = []
for i in range(5):
events.extend(_err_then("foo", "foo", t0=i * 2.0))
out = list(extract_session(events, sid="rt-rerun"))
assert _of(out, PRIMITIVE).value == "rerun"
def test_majority_switch_emits_switch() -> None:
"""Operator pivots to a different tool after each error → switch."""
events: list[AsciinemaEvent] = []
for i in range(5):
events.extend(_err_then("foo", f"bar{i}", t0=i * 2.0))
out = list(extract_session(events, sid="rt-switch"))
assert _of(out, PRIMITIVE).value == "switch"
def test_terminal_error_emits_abort() -> None:
"""Single errored command at session end → abort (only candidate)."""
events = _err_then("foo", None, t0=0.0)
out = list(extract_session(events, sid="rt-abort"))
assert _of(out, PRIMITIVE).value == "abort"
def test_low_error_count_reduces_confidence() -> None:
short_events: list[AsciinemaEvent] = []
for i in range(2):
short_events.extend(_err_then("foo", "foo", t0=i * 2.0))
full_events: list[AsciinemaEvent] = []
for i in range(6):
full_events.extend(_err_then("foo", "foo", t0=i * 2.0))
s = _of(list(extract_session(short_events, sid="rt-short")), PRIMITIVE)
f = _of(list(extract_session(full_events, sid="rt-full")), PRIMITIVE)
assert s.confidence < f.confidence
def test_pii_no_command_bodies_in_observation() -> None:
events: list[AsciinemaEvent] = []
for i in range(5):
events.extend(_err_then("supersecret", "supersecret", t0=i * 2.0))
out = list(extract_session(events, sid="rt-pii"))
obs = _of(out, PRIMITIVE)
assert "supersecret" not in obs.model_dump_json()