feat(realism): EditAction read-modify-write of planted files

Stage 3b of the realism migration. A TODO.md planted on Monday gets a
checkbox flipped on Tuesday; a notes file grows a follow-up line; a
cron log gets a fresh entry tacked on. The synthetic_files row's
edit_count, last_modified, and content_hash advance.

New surface:

- EditAction dataclass (peer of FileAction in scheduler.py): carries
  decky, path, persona, content_class, previous_body, mtime, and
  synthetic_file_uuid for the worker's update path.
- realism.bodies.next_iteration(cls, persona, prev, rng): per-class
  deterministic mutators. TODO flips an unchecked box and/or appends;
  notes/drafts/scripts append; logs are append-only (mirroring real
  log behaviour). Canary, cache_tmp, email raise KeyError —
  unsupported.
- realism.planner.pick gains an edit branch: 60% create, 30% edit
  (when an edit_candidate is supplied), 10% leave-alone. Returns
  None on leave-alone — quiet ticks are realism too.
- scheduler.pick_file pre-fetches a single edit candidate via
  repo.pick_random_synthetic_file_for_edit ~50% of ticks; the
  planner decides whether to use it.
- SSHDriver._run_edit: turns next_iteration output into a
  plant_file call (mtime-bumped, mode 0o644). Stashes new_body in
  result.payload so the worker can hash it for synthetic_files.
- worker._bump_synthetic_file_after_edit: patches edit_count + 1,
  last_modified=now, content_hash, last_body for the row UUID.
  No-op when the row was pruned mid-flight.
- events.to_row / topic_for / event_type_for now recognise
  EditAction (kind="file", action="file:edit").
This commit is contained in:
2026-04-27 16:38:17 -04:00
parent 32eeb0c813
commit b321e29002
7 changed files with 484 additions and 29 deletions

View File

@@ -231,3 +231,120 @@ def make_body(
f"no body generator registered for content_class={content_class!r}"
)
return gen(persona, rng)
# ── Edit-in-place mutators ─────────────────────────────────────────────────
# Stage 3b: deterministic per-class mutations. The contract: take the
# previous body bytes, return a plausible *next* iteration (append a
# line, flip a checkbox, fix a typo). Append-only for logs; small
# in-place edits for user content. LLM enrichment in stage 6 wires
# next_iteration to ask "what would <persona> write next" with the
# previous body in the prompt; the deterministic path stays as the
# fallback.
def _edit_todo(
prev: str, persona: str, rng: secrets.SystemRandom,
) -> str:
"""Flip an unchecked box, append a new item, or both.
Real TODO files evolve: items get checked off as work happens, new
items get added, occasionally a sub-bullet appears under an
existing one. We pick one of those mutations per call.
"""
lines = prev.splitlines()
unchecked_indices = [
i for i, ln in enumerate(lines) if ln.startswith("- [ ]")
]
op = rng.choice(("flip", "append", "both") if unchecked_indices else ("append",))
if op in ("flip", "both") and unchecked_indices:
idx = rng.choice(unchecked_indices)
lines[idx] = lines[idx].replace("- [ ]", "- [x]", 1)
if op in ("append", "both"):
new_item = rng.choice(_TODO_VERBS)
marker = "[x]" if rng.random() < 0.15 else "[ ]"
lines.append(f"- {marker} {new_item}")
return "\n".join(lines) + ("" if prev.endswith("\n") else "\n")
def _edit_note(
prev: str, persona: str, rng: secrets.SystemRandom,
) -> str:
"""Append one new note line or insert a follow-up under an existing one."""
new_line = rng.choice(_NOTE_TEMPLATES)
if prev.endswith("\n"):
return prev + new_line + "\n"
return prev + "\n" + new_line + "\n"
def _edit_draft(
prev: str, persona: str, rng: secrets.SystemRandom,
) -> str:
"""Append a new short paragraph to the existing draft."""
addition = (
"\nFollow-up: I'll send the deck once finance signs off on the numbers.\n",
"\nP.S.: Looping in ops on the rollout sequence — they have context I don't.\n",
"\nLet me know if any of this needs another pass.\n",
)
return prev.rstrip() + "\n" + rng.choice(addition)
def _edit_script(
prev: str, persona: str, rng: secrets.SystemRandom,
) -> str:
"""Append a comment line — scripts evolve via comments and small fixes."""
comments = (
"# TODO: handle the empty-input case\n",
"# 2026-04-27: hardened error path after the prod incident\n",
"# noqa: shellcheck disagrees but this is what the runbook says\n",
)
return prev.rstrip() + "\n" + rng.choice(comments)
def _edit_log_cron(
prev: str, persona: str, rng: secrets.SystemRandom,
) -> str:
"""Append one new cron syslog line — logs only ever grow."""
extra = _body_log_cron(persona, rng)
return prev.rstrip() + "\n" + extra.splitlines()[-1] + "\n"
def _edit_log_daemon(
prev: str, persona: str, rng: secrets.SystemRandom,
) -> str:
extra = _body_log_daemon(persona, rng)
return prev.rstrip() + "\n" + extra.splitlines()[-1] + "\n"
_EDITORS: dict[ContentClass, Callable[[str, str, secrets.SystemRandom], str]] = {
ContentClass.NOTE: _edit_note,
ContentClass.TODO: _edit_todo,
ContentClass.DRAFT: _edit_draft,
ContentClass.SCRIPT: _edit_script,
ContentClass.LOG_CRON: _edit_log_cron,
ContentClass.LOG_DAEMON: _edit_log_daemon,
}
def next_iteration(
content_class: ContentClass,
persona: str,
previous_body: str,
*,
rand: Optional[secrets.SystemRandom] = None,
) -> str:
"""Return the next-iteration body for an edit-in-place mutation.
Raises :class:`KeyError` for content classes that don't support
editing (canary blobs, cache-tmp scratch files, email). The
planner filters those out before producing an :class:`EditAction`,
so reaching this branch with an unsupported class is a bug worth
surfacing loudly.
"""
rng = rand or secrets.SystemRandom()
editor = _EDITORS.get(content_class)
if editor is None:
raise KeyError(
f"content_class={content_class!r} does not support edits"
)
return editor(previous_body, persona, rng)

View File

@@ -26,7 +26,7 @@ from typing import Any, Optional, Sequence
from decnet.realism import bodies, naming
from decnet.realism.diurnal import in_work_hours, sample_mtime
from decnet.realism.personas import EmailPersona
from decnet.realism.taxonomy import ContentClass, Plan
from decnet.realism.taxonomy import ContentClass, Plan, PlanAction # noqa: F401
# Stage-3 weighted sampling:
@@ -83,16 +83,23 @@ def pick(
deckies: Sequence[dict[str, Any]],
now: datetime,
*,
edit_candidate: Optional[dict[str, Any]] = None,
rand: Optional[secrets.SystemRandom] = None,
) -> Optional[Plan]:
"""Return a single :class:`Plan` for the orchestrator's tick.
Stage-3 policy: create-only. Stage 3b extends with the
create/edit/leave roll and the synthetic_files lookup for edits.
Stage-3b policy: weighted action roll — 60% create, 30% edit, 10%
"leave alone" (planner returns ``None`` to skip). When the roll
is "edit" and *edit_candidate* is set (a row from
:meth:`BaseRepository.pick_random_synthetic_file_for_edit`), we
return an edit Plan; otherwise we fall through to create.
Returns ``None`` when no eligible (decky, persona) pair exists —
the orchestrator treats that as "skip this tick" the same way the
pre-realism scheduler did.
The orchestrator scheduler is responsible for fetching the edit
candidate before calling — keeps this function pure-of-DB and
test-friendly.
Returns ``None`` when no eligible (decky, persona) pair exists or
when the action roll lands on "leave alone."
"""
rng = rand or secrets.SystemRandom()
@@ -100,12 +107,18 @@ def pick(
if not eligible:
return None
# Action roll. Edit only fires when there's a candidate from the
# repo — otherwise we either re-roll to create or skip.
roll = rng.random()
if roll < 0.10:
return None # "leave alone" — quiet tick is realism too
if roll < 0.40 and edit_candidate is not None:
return _edit_plan(edit_candidate, now, rng)
decky, persona = rng.choice(eligible)
# User vs system content — biased toward user (realism wins are
# bigger there). Once stage 3b ships edit-in-place, the edit
# branch will reuse the same content_class as the existing row;
# the create branch picks fresh here.
# bigger there).
if rng.random() < 0.7:
content_class = _weighted_pick(_USER_CLASS_WEIGHTS, rng)
else:
@@ -130,3 +143,48 @@ def pick(
f"window={persona.active_hours}",
),
)
def _edit_plan(
candidate: dict[str, Any],
now: datetime,
rng: secrets.SystemRandom,
) -> Optional[Plan]:
"""Build an edit-action :class:`Plan` from a synthetic_files row.
The candidate dict is the shape :meth:`BaseRepository.list_synthetic_files`
returns — we only need ``decky_uuid``, ``path``, ``persona``,
``content_class``, ``last_body``, ``uuid``. Returns ``None`` if
the candidate's content_class is somehow not editable (defensive
— the repo query already filters those out).
"""
try:
cls = ContentClass(candidate["content_class"])
except (KeyError, ValueError):
return None
if cls.is_canary() or cls == ContentClass.CACHE_TMP:
return None
# mtime: edits bump forward by ~hours-to-days, but never past now.
# We model as "the file was edited some time after creation but
# before now" — sample_mtime with a tighter cap keeps it recent.
edit_mtime = sample_mtime(
"00:00-00:00", now, rand=rng,
backdate_min_hours=1.0, backdate_max_days=2.0,
)
return Plan(
decky_uuid=candidate["decky_uuid"],
decky_name=candidate.get("decky_name", ""),
persona=candidate.get("persona", ""),
content_class=cls,
action="edit",
target_path=candidate["path"],
mtime=edit_mtime,
body_hint=None, # edit uses previous_body, not a fresh hint
previous_body=candidate.get("last_body", ""),
notes=(
f"persona={candidate.get('persona', '')}",
f"class={cls.value}",
"action=edit",
f"synthetic_file_uuid={candidate.get('uuid', '')}",
),
)