Files
threshing-atlas/collection/etl/transform/bot.py
T
ekcSchianoandClaude Opus 5.5 43066ad67f Bot import: one folder per account, run records with steps
- the importer reads data/accounts/*/runs.jsonl and logs/accounts/*/encounters.jsonl;
  with account folders, the top-level files (a mirror of the first account) are skipped
- run records that carry their steps (schema 2) become trajectories directly, keeping
  what the bot saw (shadow, meadow dragons), propensities, decision mode and policy id;
  older runs are still rebuilt from the decision log; bonded runs without steps are kept
- account names become salted hashes (DK_ACCOUNT_SALT, or a private salt in data/raw)
- runs_clean.csv: new bot_mode column; saw_shadow and saw_blue_meadow from what the bot
  saw; deliberate burns say why
- crosswalk: the valley Red, leaving squadmates, squad direction, hold your ground,
  hide/approach reactions, the wounded cadet's other two choices

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
2026-11-10 02:08:57 +00:00

368 lines
17 KiB
Python

"""The DragonKind bot's private data -> anonymised trajectories and a scene catalogue (Python port of bot.js).
Raw input (from the private bot repo, packed by extract.local_dir): data/knowledge.json (scene atlas: prose,
choices, tint, backdrop), data/runs.jsonl (one record per attempt) and logs/encounters.jsonl (every decision the
bot made, with the options on screen). Nothing of the prose or the account ids leaves this module: the output
holds scene fingerprints, short labels (collection/specs/bot.yaml), choice text and outcomes, and check_public()
refuses anything else.
Output: {"scenes": {sceneId: {stage, tint, backdrop, label, options: [{id, text}]}},
"trajectories": [{id, source, meta, steps, outcome}], "unlabeled": [...]}
"""
import hashlib
import json
import re
from datetime import datetime
from ..common import read_yaml
SPEC = read_yaml("collection/specs/bot.yaml")
_NORMALIZE = [(re.compile(a), b) for a, b in SPEC["normalize"]]
CHOICE_EVENTS = set(SPEC["choice_events"])
def sha1(s, n=12):
return hashlib.sha1(str(s).encode("utf-8")).hexdigest()[:n]
def normalize(text):
"""Must match normalize() in the bot's src/strategy.js, so fingerprints line up."""
t = str(text or "").lower()
for rx, rep in _NORMALIZE:
t = rx.sub(rep, t)
return t.strip()
def scene_id(prose, choices):
"""Must match computeSceneFingerprint() in the bot's src/learner.js."""
prose_norm = normalize(prose)[:500]
opts = sorted(c for c in (normalize(c) for c in choices) if c)
return sha1(f"{prose_norm} ### {' || '.join(opts)}")
def option_id(text):
return sha1(normalize(text), 10)
def label_for(sid, node):
if sid in SPEC["labels"]:
return SPEC["labels"][sid]
bits = [node.get("stage") or "Scene"]
if node.get("tint"):
bits.append(node["tint"])
return " · ".join(bits)
def _jsonl(text):
out = []
for line in (text or "").splitlines():
if line.strip():
try:
out.append(json.loads(line))
except json.JSONDecodeError:
pass
return out
def build_scenes(knowledge, encounters):
scenes = {}
for sid, n in ((knowledge or {}).get("nodes") or {}).items():
scenes[sid] = {
"stage": n.get("stage") or "UNKNOWN", "tint": n.get("tint"), "backdrop": n.get("backdrop"),
"label": label_for(sid, n),
"options": [{"id": option_id(c.get("label") or c.get("choiceNorm")), "text": c.get("label") or c.get("choiceNorm")}
for c in (n.get("choices") or {}).values()],
}
for e in encounters:
if not isinstance(e.get("choices"), list) or not e.get("prose"):
continue
sid = e.get("nodeId") or scene_id(e["prose"], e["choices"])
if sid in scenes:
continue
stage = (e.get("decision") or {}).get("stage") or e.get("stage") or "UNKNOWN"
scenes[sid] = {"stage": stage, "tint": None, "backdrop": None, "label": label_for(sid, {"stage": stage}),
"options": [{"id": option_id(t), "text": t} for t in e["choices"]]}
return scenes
def trajectories_from_encounters(encounters, scenes):
"""Group the log by attempt, keep the decisions acted on, and close each attempt at its first BURNED or
DRAGON_BONDED event. Repeated BURNED events with no decision between them are ignored."""
by_attempt = {}
for e in encounters:
key = f"{e['account']}:{e.get('attempt')}" if e.get("account") else str(e.get("attempt"))
by_attempt.setdefault(key, []).append(e)
out = []
for key, events in by_attempt.items():
events = sorted(events, key=lambda e: str(e.get("ts")))
steps, outcome = [], None
for e in events:
if outcome:
break
if e.get("event") in CHOICE_EVENTS:
d = e.get("decision") or {}
text = d.get("choiceText") or e.get("choiceText")
if d.get("action") and d["action"] != "choose":
continue # paused, nothing happened
if not text or not isinstance(e.get("choices"), list):
continue
scene = e.get("nodeId") or scene_id(e.get("prose") or "", e["choices"])
manual = e["event"] != "DECISION"
sc = scenes.get(scene) or {}
steps.append({
"scene": scene, "stage": d.get("stage") or sc.get("stage") or "UNKNOWN", "tint": sc.get("tint"),
"option": option_id(text), "optionText": text, "offered": [option_id(c) for c in e["choices"]],
"rule": d.get("rule"), "decisionMode": "manual" if manual else d.get("source"),
"manual": manual, "at": e.get("ts"),
})
elif e.get("event") == "BURNED" and steps:
outcome = {"type": "BURNED", "colour": None, "intentional": bool(e.get("intentional"))}
elif e.get("event") == "DRAGON_BONDED" and steps:
outcome = {"type": "BONDED", "colour": str(e["dragon"]).lower() if e.get("dragon") else None, "intentional": False}
# a decision repeated on the same scene (the bot re-detecting the screen) counts once
dedup = [s for i, s in enumerate(steps)
if not (i > 0 and steps[i - 1]["scene"] == s["scene"] and steps[i - 1]["option"] == s["option"])]
for i, s in enumerate(dedup):
s["to"] = dedup[i + 1]["scene"] if i + 1 < len(dedup) else None
if not dedup:
continue
if outcome:
dedup[-1]["to"] = "BURNED" if outcome["type"] == "BURNED" else f"BONDED:{outcome['colour'] or 'unknown'}"
try:
attempt = int(float(key.split(":")[-1])) or None
except ValueError:
attempt = None
started = dedup[0]["at"]
out.append({
"id": "bot-" + sha1(f"{key}:{started}", 10),
"source": "bot",
"meta": {"attempt": attempt, "startedAt": dedup[0]["at"], "accountKey": None, "target": None, "decisionMode": None},
"steps": dedup,
"outcome": outcome or {"type": "ABORTED", "colour": None, "intentional": False},
})
return out
def _ms(ts):
try:
return datetime.fromisoformat(str(ts).replace("Z", "+00:00")).timestamp() * 1000
except ValueError:
return float("nan")
def attach_run_meta(trajs, runs, salt):
"""Run-level metadata (target colour, mode) from runs.jsonl, where an attempt is in both."""
window = SPEC["match_hours"] * 3600e3
for r in runs:
t = next((x for x in trajs if x["meta"]["attempt"] == r.get("attempt")
and abs(_ms(x["meta"]["startedAt"]) - _ms(r.get("startedAt"))) < window), None)
if not t:
continue
m = t["meta"]
m["target"] = r.get("targetDragon")
m["decisionMode"] = r.get("decisionMode")
m["policyId"] = r.get("policyId")
m["experimentId"] = r.get("experimentId")
m["arm"] = r.get("arm")
if r.get("accountKey"):
m["accountKey"] = r["accountKey"]
elif r.get("accountId") and salt:
m["accountKey"] = sha1(f"{salt}:{r['accountId']}", 10)
if r.get("outcome") == "BONDED" and r.get("dragonColor"):
t["outcome"] = {"type": "BONDED", "colour": str(r["dragonColor"]).lower(), "intentional": False}
if r.get("intentionalBurn"):
t["outcome"]["intentional"] = True
return trajs
def import_bot(knowledge, encounters, runs, salt=None):
scenes = build_scenes(knowledge, encounters)
trajectories = attach_run_meta(trajectories_from_encounters(encounters, scenes), runs, salt)
unlabeled = [f"{sid} ({s['stage']}{', ' + s['tint'] if s['tint'] else ''})" for sid, s in scenes.items() if sid not in SPEC["labels"]]
return scenes, trajectories, unlabeled
def check_public(obj, refuse, max_string):
"""Nothing that looks like game prose or an account identifier may leave this module."""
def walk(v, p):
if isinstance(v, list):
for i, x in enumerate(v):
walk(x, f"{p}[{i}]")
elif isinstance(v, dict):
for k, x in v.items():
if k in refuse:
raise ValueError(f'refusing to export field "{k}" at {p}')
walk(x, f"{p}.{k}")
elif isinstance(v, str) and len(v) > max_string:
raise ValueError(f"refusing to export a {len(v)}-character string at {p} (looks like prose)")
walk(obj, "$")
def merge_scenes(existing, scenes):
"""Add new scenes and new options to the published catalogue; never drop what is already there."""
out = json.loads(json.dumps(existing or {}))
for sid, s in scenes.items():
old = out.get(sid)
if not old:
out[sid] = s
continue
ids = {o["id"] for o in old["options"]}
old["options"] += [o for o in s["options"] if o["id"] not in ids]
old["tint"] = old.get("tint") or s["tint"]
old["backdrop"] = old.get("backdrop") or s["backdrop"]
if "·" in old["label"] and "·" not in s["label"]:
old["label"] = s["label"]
return dict(sorted(out.items()))
# ---------------------------------------------------------------- run records with steps (the bot's schema 2)
def _to(key):
"""The bot's toKey -> ours: a scene id, BURNED or BONDED:<colour>."""
key = str(key or "")
if key == "TERMINAL:BURNED":
return "BURNED"
if key.startswith("TERMINAL:BONDED"):
return "BONDED:" + (key.split("_", 1)[1].lower() if "_" in key else "unknown")
return key or None
def _seen(observed):
"""What the bot saw on a screen that the route model needs: the dragons (from the tint) and the shadow."""
o = observed or {}
out = {}
if o.get("dragons"):
out["dragons"] = [str(d).lower() for d in o["dragons"]]
if "shadow" in o:
out["shadow"] = bool(o["shadow"])
return out
def trajectory_from_run(r, account_key):
"""A run record that carries its own steps -> a trajectory. Nothing but choice text, ids and outcomes is kept."""
steps = []
for st in r.get("steps") or []:
text = st.get("choiceText") or st.get("choiceNorm")
if not text:
continue
steps.append({
"scene": st.get("nodeId"), "stage": st.get("stage") or "UNKNOWN", "tint": st.get("tint"),
"option": option_id(text), "optionText": text,
"offered": [option_id(o.get("text") or o.get("norm")) for o in st.get("offered") or [] if o.get("text") or o.get("norm")],
"rule": st.get("rule"), "decisionMode": "manual" if st.get("manual") else st.get("decisionMode"),
"manual": bool(st.get("manual")), "at": st.get("at"), "seen": _seen(st.get("observed")),
"propensity": st.get("propensity"), "to": _to(st.get("toKey")),
})
colour = str(r["dragonColor"]).lower() if r.get("dragonColor") else None
return {
"id": "bot-" + sha1(r["runId"], 10),
"source": "bot",
"meta": {"attempt": r.get("accountAttempt") or r.get("attempt"), "startedAt": r.get("startedAt"),
"accountKey": account_key, "accountRole": r.get("accountRole"), "target": r.get("targetDragon"),
"decisionMode": r.get("decisionMode"), "policyId": r.get("policyId"),
"experimentId": r.get("experimentId"), "arm": r.get("arm")},
"steps": steps,
"outcome": {"type": r.get("outcome") or "ABORTED", "colour": colour, "intentional": bool(r.get("intentionalBurn")),
"reason": (r.get("burn") or {}).get("reason")},
}
def _account_files(files):
"""{account: {"runs": [...], "encounters": [...]}} from the packed files. The top-level files are the bot's
original single account; data/accounts/<name>/ and logs/accounts/<name>/ hold one account each."""
out = {}
for rel, text in files.items():
if text is None:
continue
parts = rel.split("/")
acct = parts[2] if len(parts) == 4 and parts[1] == "accounts" else "(default)"
kind = "runs" if rel.endswith("runs.jsonl") else "encounters" if rel.endswith("encounters.jsonl") else None
if kind:
out.setdefault(acct, {"runs": [], "encounters": []})[kind] += _jsonl(text)
# since the bot keeps one folder per account, its top-level files mirror its first account (the same runs,
# logged a moment apart): counting both would count those runs twice
if len(out) > 1:
out.pop("(default)", None)
return out
def account_salt(env, privacy):
"""The salt that turns account names into anonymous keys: the environment variable if set, otherwise one made
once and kept in the private raw store (data/raw/bot/account-salt.txt, never committed)."""
salt = env.get(privacy.get("account_salt_env", ""))
if salt:
return salt
from ..common import path
import secrets
p = path("data/raw/bot/account-salt.txt")
if not p.exists():
p.parent.mkdir(parents=True, exist_ok=True)
p.write_text(secrets.token_hex(16), encoding="utf-8")
return p.read_text(encoding="utf-8").strip()
def transform(raw_path, spec, ctx):
with open(raw_path, encoding="utf-8") as f:
files = json.load(f)
k = files.get("data/knowledge.json")
knowledge = json.loads(k) if k else {"nodes": {}}
privacy = spec.get("privacy") or {}
salt = account_salt(ctx.get("env", {}), privacy)
accounts = _account_files(files)
all_encounters = [e for a in accounts.values() for e in a["encounters"]]
scenes = build_scenes(knowledge, all_encounters)
trajectories, seen_ids = [], set()
for acct, data in sorted(accounts.items()):
key = sha1(f"{salt}:{acct}", 10)
with_steps = [r for r in data["runs"] if r.get("steps") and r.get("runId")]
for r in with_steps:
t = trajectory_from_run(r, key)
if t["id"] not in seen_ids:
seen_ids.add(t["id"])
trajectories.append(t)
# older runs without steps: rebuild them from the decision log, as before
legacy = [r for r in data["runs"] if not r.get("steps")]
window = SPEC["match_hours"] * 3600e3
same_run = lambda t, r: (t["meta"]["attempt"] == (r.get("accountAttempt") or r.get("attempt")) # noqa: E731
and abs(_ms(t["meta"]["startedAt"]) - _ms(r.get("startedAt"))) < window)
old = attach_run_meta(trajectories_from_encounters(data["encounters"], scenes), legacy, salt)
kept = []
for t in old: # the decision log also holds the runs that have steps: keep only the others
if any(same_run(t, r) for r in with_steps) or t["id"] in seen_ids:
continue
t["meta"]["accountKey"] = key
seen_ids.add(t["id"])
trajectories.append(t)
kept.append(t)
# bonded runs logged without any steps still say which dragon was bonded
for r in legacy:
if r.get("outcome") == "BONDED" and r.get("runId") and not any(same_run(t, r) for t in kept):
t = trajectory_from_run({**r, "steps": []}, key)
if t["id"] not in seen_ids:
seen_ids.add(t["id"])
trajectories.append(t)
# scenes the run records show but the knowledge file does not have yet
for t in trajectories:
for st in t["steps"]:
if st["scene"] and st["scene"] not in scenes:
scenes[st["scene"]] = {"stage": st["stage"], "tint": st["tint"], "backdrop": None,
"label": label_for(st["scene"], {"stage": st["stage"], "tint": st["tint"]}), "options": []}
sc = scenes.get(st["scene"])
if sc is not None and st["option"] not in {o["id"] for o in sc["options"]}:
sc["options"].append({"id": st["option"], "text": st["optionText"]})
unlabeled = [f"{sid} ({s['stage']}{', ' + s['tint'] if s['tint'] else ''})" for sid, s in scenes.items() if sid not in SPEC["labels"]]
for obj in (scenes, trajectories):
check_public(obj, set(privacy.get("refuse_fields", [])), privacy.get("max_string", 200))
trajectories.sort(key=lambda t: (str(t["meta"].get("startedAt")), t["id"]))
return {"scenes": scenes, "trajectories": trajectories, "_unlabeled": unlabeled}
def describe(old, new):
outcomes = {}
for t in new["trajectories"]:
outcomes[t["outcome"]["type"]] = outcomes.get(t["outcome"]["type"], 0) + 1
lines = [f"{len(new['scenes'])} scenes, {len(new['trajectories'])} attempts {json.dumps(outcomes)}"]
if new.get("_unlabeled"):
lines.append(" - Scenes without a short label (add them to collection/specs/bot.yaml): " + ", ".join(new["_unlabeled"]))
return lines