Blame for scripts/build.py

#!/usr/bin/env python3
"""Clean Codex rollout logs into a tool-use training dataset.

Run:  python3 -I dataset-clean/scripts/build.py [--out DIR] [--files-list FILE] [--workers 8]

Reads  : /home/israel/Downloads/sessions/05/**/rollout-*.jsonl   (read-only, streamed line by line)
Writes : <out>/trajectories.{train,eval}.jsonl, pairs.{train,eval}.jsonl, redactions.jsonl,
         manifest.json   (REPORT.md is produced by report.py after validate.py)
"""
import argparse
import datetime
import glob
import hashlib
import json
import multiprocessing as mp
import os
import random
import re
import sys
import time
from collections import Counter, defaultdict

HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, HERE)  # python -I does not add the script dir to sys.path
import extract  # noqa: E402
import redact  # noqa: E402

ROOT = "/home/israel/Downloads/sessions"
RAW = ROOT + "/05"
DEFAULT_OUT = ROOT + "/dataset-clean"

SYSTEM_TMPL = ("You are a coding agent working in a local Windows workspace. Use the available tools to inspect "
               "files, edit code, build and run tests. Working directory: {cwd}.")
COMPACT_NOTE = "[Earlier context in this task was compacted. Its summary is not available.]"
GOAL_PREFIX = "Continue working on this goal:\n"
TRUNC_LIMIT, TRUNC_HEAD, TRUNC_TAIL = 4000, 2000, 2000
LARGE_ARG = 20000
SPLIT_SEED = 42
EXIT_RE = re.compile(r"\s*Exit code:\s*(-?\d+)")
UUID_RE = re.compile(r"([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})\.jsonl$")

SELECTED = []  # filled in the parent before forking Stage 2 workers
RECOVERY = {}  # turn_id -> manual recovery entry (scripts/manual_recoveries.json)
MANUAL_FILE = os.path.join(HERE, "manual_recoveries.json")


# ============================================================================ helpers
def norm_output(o, stats):
    """Tool output -> plain string."""
    if isinstance(o, str):
        return o
    stats["non_string_output"] += 1
    if o is None:
        return ""
    if isinstance(o, list):
        parts = []
        for it in o:
            if isinstance(it, dict):
                t = it.get("text")
                if isinstance(t, str):
                    parts.append(t)
                elif "image" in str(it.get("type", "")):
                    stats["image_output_omitted"] += 1
                    parts.append("[image omitted]")
                else:
                    parts.append(json.dumps(it, ensure_ascii=False))
            else:
                parts.append(str(it))
        return "\n".join(parts)
    if isinstance(o, dict):
        c = o.get("content")
        if isinstance(c, (str, list)):
            return norm_output(c, stats)
        return json.dumps(o, ensure_ascii=False)
    return str(o)


def is_failed(output_text):
    """Non-zero exit code, or a failed apply_patch."""
    m = EXIT_RE.match(output_text[:200])
    if m and int(m.group(1)) != 0:
        return True
    return output_text.lstrip()[:60].startswith("apply_patch verification failed")


def to_line(obj):
    s = json.dumps(obj, ensure_ascii=False)
    try:
        s.encode("utf-8")
    except UnicodeEncodeError:  # lone surrogates from the raw log: keep them escaped, stay valid JSON
        s = json.dumps(obj, ensure_ascii=True)
    return s


def truncate_output(text, events, stats):
    """4000-char rule: first 2000 + last 2000 with a marker. Redaction events that fall in the removed
    middle are dropped from the log (they no longer exist in the data)."""
    if len(text) <= TRUNC_LIMIT:
        return text, events
    n = len(text) - TRUNC_LIMIT
    new = text[:TRUNC_HEAD] + f"\n[... {n} chars truncated ...]\n" + text[-TRUNC_TAIL:]
    kept = []
    for ev in events:
        s, e = ev[0], ev[1]
        if s < 0 or s < TRUNC_HEAD or e > len(text) - TRUNC_TAIL:
            kept.append(ev)
        else:
            stats["redactions_in_truncated_middle"] += 1
    stats["outputs_truncated"] += 1
    stats["chars_removed_by_truncation"] += n
    return new, kept


# ============================================================================ stage 2: one turn -> rows
def build_segment_messages(seg_events, stats):
    """Convert one segment of raw events into chat messages (pre-redaction).
    Returns (messages, n_calls, n_failed)."""
    call_ids = {e[1] for e in seg_events if e[0] == "C"}
    out_first = {}
    for e in seg_events:
        if e[0] == "O" and e[1] in call_ids and e[1] not in out_first:
            out_first[e[1]] = e
    msgs = []
    pend_text = None
    pend_calls = []
    n_calls = 0
    n_failed = 0
    seen_out = set()
    seen_calls_in_pending = set()

    def flush():
        nonlocal pend_text, pend_calls
        if pend_text is not None or pend_calls:
            m = {"role": "assistant", "content": pend_text}
            if pend_calls:
                m["tool_calls"] = pend_calls
            msgs.append(m)
        pend_text = None
        pend_calls = []

    for e in seg_events:
        k = e[0]
        if k == "A":
            if pend_calls:
                stats["text_after_unanswered_calls"] += 1
            flush()
            pend_text = e[1]
        elif k == "C":
            cid = e[1]
            if cid not in out_first:
                stats["tool_calls_dropped_no_output"] += 1
                continue
            args = e[3]
            kind = e[4]
            if kind == "f" and not isinstance(args, str):
                stats["non_string_arguments"] += 1
                args = json.dumps(args, ensure_ascii=False)
            if kind == "c" and not isinstance(args, str):
                stats["non_string_arguments"] += 1
                args = "" if args is None else json.dumps(args, ensure_ascii=False)
            pend_calls.append({"id": cid, "type": "function",
                               "function": {"name": e[2] or "", "arguments": args, "_kind": kind}})
            n_calls += 1
        elif k == "O":
            cid = e[1]
            if cid not in call_ids:
                stats["tool_outputs_dropped_orphan"] += 1
                continue
            if cid in seen_out:
                stats["tool_outputs_dropped_duplicate"] += 1
                continue
            if out_first.get(cid) is not e:
                continue
            seen_out.add(cid)
            flush()
            text = norm_output(e[2], stats)
            if is_failed(text):
                n_failed += 1
            msgs.append({"role": "tool", "tool_call_id": cid, "content": text})
        elif k == "U":
            flush()
            msgs.append({"role": "user", "content": e[1]})
    flush()
    return msgs, n_calls, n_failed


def build_turn(idx):
    sel = SELECTED[idx]
    inst = sel["inst"]
    meta = sel["file"]
    stats = Counter()
    logs = []          # (row_id, field, kind, sha, replacement)
    res = {"idx": idx, "rows": [], "logs": logs, "stats": stats, "drop": None}

    ctx = inst["ctx"] or {}
    session = meta["session"] or {}
    cwd = ctx.get("cwd") or session.get("cwd") or "unknown"
    events = [e for e in inst["events"]]
    rec = RECOVERY.get(inst["turn_id"])
    rec_kind = rec["kind"] if rec else None

    # ---- task message (step 3) -------------------------------------------------
    task = None
    task_source = None
    if rec_kind == "task_reconstructed":
        task, task_source = rec["task"], "reconstructed"
    for k, e in enumerate(events if task is None else []):
        if e[0] == "U":
            if isinstance(e[1], str) and e[1].strip():
                task, task_source = e[1], "user"
                if any(x[0] in ("A", "C", "O") for x in events[:k]):
                    stats["content_before_first_user_message"] += 1
                del events[k]
                break
    if task is None:
        goal = inst["goal_in_turn"] if rec_kind == "goal_set_inside_turn" else inst["goal"]
        if goal:
            task, task_source = GOAL_PREFIX + goal, "goal"
        else:
            res["drop"] = "no_task_message_and_no_goal"
            if inst["goal_in_turn"]:
                stats["no_task_but_goal_inside_turn"] += 1
            return res
    # drop empty mid-turn user messages
    n0 = len(events)
    events = [e for e in events if not (e[0] == "U" and not (isinstance(e[1], str) and e[1].strip()))]
    stats["empty_user_events_dropped"] += n0 - len(events)

    # ---- segmentation at compaction points (step 7) ---------------------------------
    segs = [[]]
    for e in events:
        if e[0] == "K":
            segs.append([])
        else:
            segs[-1].append(e)
    stats["compaction_markers"] += len(segs) - 1
    cleaned = []
    for si, sg in enumerate(segs):
        empty = not any(x[0] in ("A", "C", "O", "U") for x in sg)
        if empty and si < len(segs) - 1:
            stats["empty_segment_before_compaction_ignored"] += 1   # leading / back-to-back compactions
            continue
        cleaned.append(sg)
    segs = cleaned
    split = len(segs) > 1

    # ---- messages per segment ---------------------------------------------------
    seg_msgs = []
    total_calls = 0
    for sg in segs:
        m, nc, nf = build_segment_messages(sg, stats)
        seg_msgs.append([m, nc, nf])
        total_calls += nc
    if total_calls == 0:
        res["drop"] = "no_tool_calls_with_output"
        return res

    # ---- final answer (step 5) -----------------------------------------------------
    final = inst["last_msg"]
    last_raw = segs[-1][-1] if segs[-1] else None
    if not (isinstance(final, str) and final.strip()):
        # task_complete.last_agent_message is null (plan-mode turns, interrupted goal turns).
        # If the turn really ends on an assistant text message, that text is the final answer.
        if last_raw is not None and last_raw[0] == "A" and last_raw[1].strip():
            final = last_raw[1]
            stats["final_answer_taken_from_last_assistant_text"] += 1
        elif rec_kind == "ends_on_tool_output":
            final = None   # manual decision: keep the turn as it is, ending on the last tool message; nothing is invented
            stats["manual_turn_kept_without_final_answer"] += 1
        else:
            res["drop"] = "no_final_answer_(null_last_agent_message,_turn_ends_on_tool_output)"
            return res
    if final is None:
        pass
    elif last_raw is not None and last_raw[0] == "A" and last_raw[1].strip() == final.strip():
        stats["final_answer_already_last"] += 1
    else:
        seg_msgs[-1][0].append({"role": "assistant", "content": final})
        stats["final_answer_appended"] += 1

    # ---- build rows: redact + truncate + flags -------------------------------------------
    turn_id = inst["turn_id"]
    cache = {}

    def red(text, row_id, field, is_output=False):
        key = text if len(text) < 200000 else None
        if key is not None and key in cache and not is_output:
            new, ev = cache[key]
        else:
            new, ev = redact.redact_text(text)
            if key is not None and not is_output:
                cache[key] = (new, ev)
        if is_output:
            new, ev = truncate_output(new, ev, stats)
        for (_s, _e, kind, sha, rep) in ev:
            logs.append((row_id, field, kind, sha, rep))
            stats["redact:" + kind] += 1
        return new

    def red_args(kind, raw, row_id, field):
        if kind == "c":
            new = red(raw, row_id, field)
            return json.dumps({"input": new})
        obj = None
        try:
            obj = json.loads(raw)
        except (ValueError, RecursionError):
            pass
        if isinstance(obj, (dict, list)):
            sink = []
            new_obj, changed = redact.redact_json_leaves(obj, sink)
            for (kd, sha, rep) in sink:
                logs.append((row_id, field, kd, sha, rep))
                stats["redact:" + kd] += 1
            if not changed:
                return raw
            return json.dumps(new_obj, ensure_ascii=False, separators=(",", ":"))
        return red(raw, row_id, field)

    ts_ctx = None
    for si, (msgs, nc, nf) in enumerate(seg_msgs):
        seg_no = si + 1
        row_id = f"{turn_id}#{seg_no}"
        head = [{"role": "system", "content": SYSTEM_TMPL.format(cwd=cwd)},
                {"role": "user", "content": task}]
        if seg_no > 1:
            head.append({"role": "user", "content": COMPACT_NOTE})
        allm = head + msgs
        out_msgs = []
        calls_seen = Counter()
        n_calls_row = 0
        large = False
        for mi, m in enumerate(allm):
            role = m["role"]
            if role == "tool":
                text = red(m["content"], row_id, f"messages[{mi}].content", is_output=True)
                out_msgs.append({"role": "tool", "tool_call_id": m["tool_call_id"], "content": text})
            elif role == "assistant":
                content = m.get("content")
                if content is not None:
                    content = red(content, row_id, f"messages[{mi}].content")
                om = {"role": "assistant", "content": content}
                if "tool_calls" in m:
                    tcs = []
                    for ci, tc in enumerate(m["tool_calls"]):
                        fn = tc["function"]
                        field = f"messages[{mi}].tool_calls[{ci}].function.arguments"
                        a = red_args(fn["_kind"], fn["arguments"], row_id, field)
                        if len(a) > LARGE_ARG:
                            large = True
                        calls_seen[(fn["name"], a)] += 1
                        n_calls_row += 1
                        tcs.append({"id": tc["id"], "type": "function", "function": {"name": fn["name"], "arguments": a}})
                    om["tool_calls"] = tcs
                out_msgs.append(om)
            else:
                text = red(m["content"], row_id, f"messages[{mi}].content")
                out_msgs.append({"role": role, "content": text})
        # flags (step 9)
        flags = []
        n_dup = sum(c - 1 for c in calls_seen.values())
        if n_calls_row and n_dup / n_calls_row > 0.10:
            flags.append("high_dup_calls")
        last = out_msgs[-1]
        if seg_no == len(seg_msgs) and last["role"] == "assistant" and len(last.get("content") or "") < 200:
            flags.append("short_final")
        if large:
            flags.append("large_patch")
        if rec and seg_no == len(seg_msgs):
            flags.append("manual_recovery")
            if final is None:
                flags.append("ends_on_tool_output")
        elif rec:
            flags.append("manual_recovery")
        if n_calls_row != nc:
            stats["internal_call_count_mismatch"] += 1
        row_cwd = red(cwd, row_id, "cwd")
        row = {
            "id": row_id, "session_id": meta["session_id"], "turn_id": turn_id, "segment": seg_no,
            "segments_total": len(seg_msgs), "compaction_summary_available": not split,
            "source_file": meta["rel"], "thread_source": session.get("thread_source"),
            "agent_role": session.get("agent_role"), "model": ctx.get("model"), "effort": ctx.get("effort"),
            "cwd": row_cwd, "timestamp": inst["ts"], "task_source": task_source,
            "n_tool_calls": n_calls_row, "n_failed_calls": nf, "flags": flags, "messages": out_msgs,
        }
        if split and seg_no < len(seg_msgs):
            stats["non_final_segments"] += 1
            if out_msgs[-1]["role"] != "assistant":
                stats["non_final_segments_ending_in_" + out_msgs[-1]["role"]] += 1
        res["rows"].append((meta["session_id"], row_id, to_line(row)))
    if split:
        stats["turns_split"] += 1
    stats["rows"] += len(seg_msgs)
    return res


# ============================================================================ pairs (step 10)
def build_pairs(file_results, stats):
    """Brief/report pairs for subagent sessions: the own turn(s) whose task is the session's first user message."""
    pairs = []
    pair_logs = []
    seen_briefs = {}
    for fr in file_results:
        sess = fr["session"]
        if not sess or sess.get("thread_source") != "subagent":
            continue
        stats["subagent_files"] += 1
        own = [i for i in fr["instances"] if i["own"]]
        first = None
        for i in own:
            if isinstance(i["first_user"], str) and i["first_user"].strip():
                first = i["first_user"]
                break
        if first is None:
            stats["subagent_files_without_first_user_message"] += 1
            continue
        got_one = False
        for i in own:
            if i["first_user"] != first:
                continue
            if not i["complete"]:
                stats["pair_candidate_turn_not_complete"] += 1
                continue
            report = i["last_msg"]
            if not (isinstance(report, str) and report.strip()):
                stats["pair_candidate_empty_report"] += 1
                continue
            row_id = "pair:" + i["turn_id"]
            brief_r, ev1 = redact.redact_text(first)
            report_r, ev2 = redact.redact_text(report)
            cwd = (i["ctx"] or {}).get("cwd") or sess.get("cwd")
            cwd_r, ev3 = redact.redact_text(cwd) if cwd else (cwd, [])
            if brief_r in seen_briefs:
                stats["pairs_dropped_duplicate_brief"] += 1
                continue
            seen_briefs[brief_r] = i["turn_id"]
            for field, ev in (("brief", ev1), ("report", ev2), ("cwd", ev3)):
                for (_s, _e, kind, sha, rep) in ev:
                    pair_logs.append((row_id, field, kind, sha, rep))
                    stats["pair_redact:" + kind] += 1
            pairs.append({
                "session_id": fr["session_id"],
                "line": to_line({"id": i["turn_id"], "session_id": fr["session_id"], "turn_id": i["turn_id"],
                                 "agent_role": sess.get("agent_role"), "cwd": cwd_r, "brief": brief_r,
                                 "report": report_r}),
                "rel": fr["rel"], "start_line": i["line"], "tool_calls": i["n_calls"],
            })
            got_one = True
        if not got_one:
            stats["subagent_files_without_pair"] += 1
    return pairs, pair_logs


# ============================================================================ main
def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--out", default=DEFAULT_OUT)
    ap.add_argument("--files-list", default=None, help="dev: only process the files listed in this text file")
    ap.add_argument("--workers", type=int, default=8)
    ap.add_argument("--no-manual", action="store_true", help="ignore scripts/manual_recoveries.json (automatic rules only)")
    args = ap.parse_args()
    out = os.path.abspath(args.out)
    for forbidden in (RAW, ROOT + "/dataset-analysis"):
        if out == forbidden or out.startswith(forbidden + "/"):
            sys.exit("refusing to write into read-only input: " + out)
    os.makedirs(out, exist_ok=True)
    t0 = time.time()

    if args.files_list:
        files = [l.strip() for l in open(args.files_list) if l.strip()]
    else:
        files = glob.glob(RAW + "/*/rollout-*.jsonl")
    files = sorted(set(files))
    jobs = [(p, os.path.relpath(p, ROOT)) for p in files]
    jobs_by_size = sorted(jobs, key=lambda j: -os.path.getsize(j[0]))
    print(f"[stage1] {len(jobs)} files, {sum(os.path.getsize(j[0]) for j in jobs)/1e9:.2f} GB", flush=True)

    ctx = mp.get_context("fork")
    results = {}
    with ctx.Pool(args.workers) as pool:
        for k, r in enumerate(pool.imap_unordered(extract.scan_file, jobs_by_size, chunksize=1)):
            results[r["rel"]] = r
            if (k + 1) % 50 == 0 or r["size"] > 3e8:
                print(f"[stage1] {k+1}/{len(jobs)} done ({time.time()-t0:.0f}s) last={r['rel'][-40:]} size={r['size']/1e6:.0f}MB", flush=True)
    file_results = [results[j[1]] for j in jobs]
    file_results.sort(key=lambda r: r["rel"])
    print(f"[stage1] finished in {time.time()-t0:.0f}s", flush=True)

    man = {"manifest_version": 1}
    gstats = Counter()
    for fr in file_results:
        gstats.update(fr["stats"])
        # session id: first session_meta, fallback to the uuid in the file name
        m = UUID_RE.search(fr["rel"])
        fname_id = m.group(1) if m else None
        sess = fr["session"]
        fr["session_id"] = (sess or {}).get("id") or fname_id
        if sess and fname_id and sess.get("id") != fname_id:
            gstats["session_id_differs_from_filename"] += 1
        if not sess and fr["size"] > 0:
            gstats["files_without_session_meta"] += 1
        if fr["stats"].get("malformed_lines") or fr["stats"].get("scalar_lines"):
            gstats["files_with_malformed_or_scalar_lines"] += 1
        if sess and fr["sm_count"] > 1:
            gstats["files_with_embedded_parent_context"] += 1
    man["files_read"] = len(file_results)
    man["files_empty_0_bytes"] = gstats["empty_file"]
    man["files_without_session_meta_nonempty"] = gstats["files_without_session_meta"]
    man["files_with_embedded_parent_context"] = gstats["files_with_embedded_parent_context"]
    man["files_with_malformed_or_scalar_lines"] = gstats["files_with_malformed_or_scalar_lines"]
    man["lines_total"] = gstats["lines"]
    man["blank_lines"] = gstats["blank_lines"]
    man["malformed_lines"] = gstats["malformed_lines"]
    man["scalar_lines"] = gstats["scalar_lines"]
    man["malformed_lines_breakdown"] = {k.split(":", 1)[1]: v for k, v in sorted(gstats.items()) if k.startswith("malformed:")}
    man["records_parsed"] = gstats["records"]
    man["bytes_read"] = sum(fr["size"] for fr in file_results)

    # ---- turn accounting (steps 1-2) ---------------------------------------------------
    inst_all = []
    for fr in file_results:
        for i in fr["instances"]:
            inst_all.append((fr, i))
    reasons = Counter()
    passing = []
    for fr, i in inst_all:
        if i["turn_id"] is None:
            reasons["instance_without_turn_id"] += 1
        elif not i["complete"] and i["n_calls"] == 0:
            reasons["no_task_complete_and_no_tool_calls"] += 1
        elif not i["complete"]:
            reasons["no_task_complete_but_has_tool_calls"] += 1
        elif i["n_calls"] == 0:
            reasons["has_task_complete_but_no_tool_calls"] += 1
        else:
            passing.append((fr, i))
    man["turn_instances_found"] = len(inst_all)
    man["unique_turn_ids_found"] = len({i["turn_id"] for _, i in inst_all})
    man["turn_instances_passing_step1_filter"] = len(passing)
    by_tid = defaultdict(list)
    for fr, i in passing:
        by_tid[i["turn_id"]].append((fr, i))
    selected = []
    dup_dropped = 0
    for tid, lst in by_tid.items():
        lst.sort(key=lambda x: (-x[1]["n_calls"], x[0]["size"], x[0]["rel"], x[1]["line"]))
        selected.append(lst[0])
        dup_dropped += len(lst) - 1
    selected.sort(key=lambda x: (x[0]["rel"], x[1]["line"]))
    reasons["duplicate_turn_id_lower_priority_copy"] = dup_dropped
    # Sensitivity of the step order (spec: filter, then dedup). "Dedup first" would pick the copy with the most calls
    # among ALL copies and then filter; count turn_ids where that gives a different outcome.
    all_by_tid = defaultdict(list)
    for fr, i in inst_all:
        if i["turn_id"] is not None:
            all_by_tid[i["turn_id"]].append((fr, i))
    order_diff = 0
    for tid, lst in all_by_tid.items():
        best = sorted(lst, key=lambda x: (-x[1]["n_calls"], x[0]["size"], x[0]["rel"], x[1]["line"]))[0]
        best_passes = best[1]["complete"] and best[1]["n_calls"] >= 1
        any_pass = any(i["complete"] and i["n_calls"] >= 1 for _, i in lst)
        if any_pass and not best_passes:
            order_diff += 1
    man["dedup_order_sensitivity_turn_ids_that_would_differ_if_dedup_ran_before_filter"] = order_diff
    orphan_with_calls = sum(1 for fr in file_results for o in fr["orphan_turns"] if o["orphan_calls"] > 0)
    orphan_all = sum(len(fr["orphan_turns"]) for fr in file_results)
    man["orphan_turns_task_complete_without_start"] = {"total": orphan_all, "with_tool_calls": orphan_with_calls}
    man["turns_after_dedup"] = len(selected)          # the spec's step-2 number; manual recoveries are added AFTER it
    man["selected_turns_not_belonging_to_files_own_session"] = sum(1 for _, i in selected if not i["own"])
    print(f"[dedup] instances={len(inst_all)} passing={len(passing)} unique_after_dedup={len(selected)}", flush=True)
    selected_core = list(selected)

    # ---- manual recoveries (scripts/manual_recoveries.json) ------------------------------------------
    RECOVERY.clear()
    n_orphans_recovered = 0
    manual_cfg = {"recoveries": [], "left_out": []}
    if not args.no_manual and os.path.exists(MANUAL_FILE):
        manual_cfg = json.load(open(MANUAL_FILE, encoding="utf-8"))
        core_ids = {i["turn_id"] for _, i in selected_core}
        orphan_index = {}
        for fr in file_results:
            for o in fr["orphan_turns"]:
                if o.get("events"):
                    orphan_index[o["turn_id"]] = (fr, o)
        for entry in manual_cfg["recoveries"]:
            tid, kind = entry["turn_id"], entry["kind"]
            RECOVERY[tid] = entry
            if kind == "task_reconstructed":
                if tid not in orphan_index:
                    sys.exit(f"manual recovery {tid}: no orphan turn with that id")
                fr, o = orphan_index[tid]
                started = None
                if o.get("completed_at") is not None and o.get("duration_ms") is not None:
                    st_epoch = int(o["completed_at"]) - round(int(o["duration_ms"]) / 1000)
                    started = datetime.datetime.fromtimestamp(st_epoch, datetime.timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
                inst = {"turn_id": tid, "ts": started, "started_at": None, "line": -1,
                        "ctx": {"model": None, "effort": None, "cwd": entry["cwd"]}, "complete": True,
                        "last_msg": o["last_msg"], "n_calls": o["orphan_calls"], "n_compact": 0, "own": True,
                        "goal": None, "goal_in_turn": None, "first_user": None, "n_user_events": 0, "aborted": False,
                        "drop": {}, "passing": True, "events": o["events"], "user_images": 0, "dup_calls": 0,
                        "manual_orphan": True}
                fr2 = dict(fr)
                fr2["session"] = {"thread_source": entry.get("thread_source"), "agent_role": entry.get("agent_role"),
                                  "cwd": entry["cwd"], "id": fr["session_id"]}
                selected.append((fr2, inst))
                n_orphans_recovered += 1
            elif tid not in core_ids:
                sys.exit(f"manual recovery {tid}: not among the {len(core_ids)} turns that passed step 1")
        selected.sort(key=lambda x: (x[0]["rel"], x[1]["line"]))
    reasons["task_complete_without_task_started_(file_corruption)"] = orphan_all - n_orphans_recovered

    # ---- stage 2 ----------------------------------------------------------------------------
    SELECTED.clear()
    for fr, i in selected:
        SELECTED.append({"inst": i, "file": {"rel": fr["rel"], "session": fr["session"], "session_id": fr["session_id"]}})
    s2 = Counter()
    results2 = [None] * len(SELECTED)
    t1 = time.time()
    with ctx.Pool(args.workers) as pool:  # forked after SELECTED is filled: workers share it copy-on-write
        for k, r in enumerate(pool.imap_unordered(build_turn, range(len(SELECTED)), chunksize=1)):
            results2[r["idx"]] = r
    print(f"[stage2] built in {time.time()-t1:.0f}s", flush=True)

    rows = []         # (rel, line_no, seg, session_id, row_id, json)
    logs = []
    turn_drop = Counter()
    for idx, r in enumerate(results2):
        s2.update(r["stats"])
        if r["drop"]:
            turn_drop[r["drop"]] += 1
            continue
        for (sid, row_id, line) in r["rows"]:
            rows.append((sid, row_id, line))
        logs.extend(r["logs"])
    reasons.update({"step3_" + k if k.startswith("no_task") else "step5_" + k: v for k, v in turn_drop.items()})

    # ---- pairs ------------------------------------------------------------------------------
    pstats = Counter()
    pairs, pair_logs = build_pairs(file_results, pstats)
    logs.extend(pair_logs)

    # ---- split by session id (seed 42) ----------------------------------------------------------
    sessions = sorted({r[0] for r in rows} | {p["session_id"] for p in pairs})
    rng = random.Random(SPLIT_SEED)
    shuffled = list(sessions)
    rng.shuffle(shuffled)
    n_eval = max(1, round(0.1 * len(shuffled))) if shuffled else 0
    eval_sessions = set(shuffled[:n_eval])
    outputs = {"trajectories.train.jsonl": [], "trajectories.eval.jsonl": [],
               "pairs.train.jsonl": [], "pairs.eval.jsonl": []}
    for sid, row_id, line in rows:
        outputs["trajectories.eval.jsonl" if sid in eval_sessions else "trajectories.train.jsonl"].append(line)
    for p in pairs:
        outputs["pairs.eval.jsonl" if p["session_id"] in eval_sessions else "pairs.train.jsonl"].append(p["line"])
    sizes = {}
    for name, lines in outputs.items():
        with open(os.path.join(out, name), "w", encoding="utf-8", newline="\n") as f:
            for ln in lines:
                f.write(ln + "\n")
        sizes[name] = os.path.getsize(os.path.join(out, name))
    with open(os.path.join(out, "redactions.jsonl"), "w", encoding="utf-8", newline="\n") as f:
        for (row_id, field, kind, sha, rep) in logs:
            f.write(json.dumps({"row_id": row_id, "field": field, "kind": kind, "sha256_of_original": sha,
                                "replacement": rep}, ensure_ascii=False) + "\n")
    sizes["redactions.jsonl"] = os.path.getsize(os.path.join(out, "redactions.jsonl"))

    # ---- manifest ---------------------------------------------------------------------------------
    n_rows = len(rows)
    kept_idx = [k for k, r in enumerate(results2) if not r["drop"]]
    kept_inst = [selected[k][1] for k in kept_idx]
    man["turns_funnel"] = {
        "turn_instances_found_in_all_files": len(inst_all),
        "  of_which_with_task_complete": sum(1 for _, i in inst_all if i["complete"]),
        "  of_which_with_task_complete_and_>=1_tool_call (step 1)": len(passing),
        "unique_turns_after_dedup (step 2)": len(selected_core),
        "  plus manually recovered turns whose task_started was destroyed": n_orphans_recovered,
        "  with_a_task_message (step 3)": len(selected) - turn_drop["no_task_message_and_no_goal"],
        "  with_a_final_answer_or_manual_exception_and_>=1_paired_call (step 5)": len(kept_idx),
    }
    reasons_clean = {k: v for k, v in reasons.items()}
    man["turns_dropped_per_reason"] = reasons_clean
    man["turns_dropped_per_reason_note"] = (
        "Counts are turn instances (one per task_started per file), AFTER manual recoveries. The first three reasons "
        "are step 1 (embedded parent-history copies dominate them); the orphan row is file corruption; the step3/step5 "
        "rows act on the 395 turns that passed step 1.")
    man["turns_kept"] = len(kept_idx)
    man["segments_total_rows"] = n_rows
    man["turns_split_at_compaction"] = s2["turns_split"]
    man["compaction_reconciliation"] = {
        "selected_turns_containing_a_compacted_record": sum(1 for _, i in selected_core if i["n_compact"] > 0),
        "  of_which_in_kept_turns": sum(1 for i in kept_inst if i["n_compact"] > 0),
        "  of_which_split_into_segments": s2["turns_split"],
        "compaction_markers_ignored_because_nothing_precedes_them": s2["empty_segment_before_compaction_ignored"],
        "extra_segments_created": n_rows - len(kept_idx),
    }
    man["compaction_markers_in_kept_turns"] = s2["compaction_markers"]
    rec_rows = Counter()
    for sid, row_id, line in rows:
        if '"manual_recovery"' in line[:200000] and json.loads(line)["turn_id"] in RECOVERY:
            rec_rows[json.loads(line)["turn_id"]] += 1
    man["manual_recoveries"] = {
        "source": "scripts/manual_recoveries.json" if RECOVERY else None,
        "turns_recovered": sum(1 for t in RECOVERY if rec_rows.get(t)),
        "by_kind": dict(Counter(e["kind"] for t, e in RECOVERY.items() if rec_rows.get(t))),
        "turn_ids": {t: {"kind": RECOVERY[t]["kind"], "rows": rec_rows[t]} for t in RECOVERY if rec_rows.get(t)},
        "requested_but_not_produced": [t for t in RECOVERY if not rec_rows.get(t)],
        "tool_outputs_dropped_because_their_call_was_destroyed": sum(
            results2[k]["stats"]["tool_outputs_dropped_orphan"] for k in range(len(results2))
            if SELECTED[k]["inst"]["turn_id"] in RECOVERY),
        "left_out": manual_cfg.get("left_out", []),
    }
    man["rows"] = {
        "train": len(outputs["trajectories.train.jsonl"]), "eval": len(outputs["trajectories.eval.jsonl"]),
        "pairs_train": len(outputs["pairs.train.jsonl"]), "pairs_eval": len(outputs["pairs.eval.jsonl"]),
    }
    man["split"] = {"unit": "session_id", "seed": SPLIT_SEED,
                    "algorithm": "sorted unique ids -> random.Random(42).shuffle -> first round(10%) to eval",
                    "sessions_total": len(sessions), "sessions_eval": len(eval_sessions),
                    "trajectory_rows_eval_share": round(man["rows"]["eval"] / n_rows, 4) if n_rows else None,
                    "pairs_eval_share": round(man["rows"]["pairs_eval"] / max(1, len(pairs)), 4)}
    man["tool_outputs"] = {"truncated": s2["outputs_truncated"], "chars_removed_by_truncation": s2["chars_removed_by_truncation"],
                           "non_string_outputs_normalised": s2["non_string_output"],
                           "image_parts_replaced_by_placeholder": s2["image_output_omitted"]}
    man["tool_calls"] = {"dropped_no_output": s2["tool_calls_dropped_no_output"],
                         "orphan_outputs_dropped": s2["tool_outputs_dropped_orphan"],
                         "duplicate_outputs_dropped": s2["tool_outputs_dropped_duplicate"],
                         "duplicate_call_ids_dropped": sum(i["dup_calls"] for i in kept_inst)}
    drop_cnt = Counter()
    for i in kept_inst:
        drop_cnt.update(i["drop"])
    man["dropped_records_in_kept_turns"] = dict(sorted(drop_cnt.items(), key=lambda kv: -kv[1]))
    man["user_messages_with_images_in_kept_turns(images_not_included)"] = sum(i["user_images"] for i in kept_inst)
    man["redactions"] = {k.split(":", 1)[1]: v for k, v in sorted(s2.items()) if k.startswith("redact:")}
    man["redactions_total_logged"] = len(logs)
    man["redactions_in_truncated_middle_not_logged"] = s2["redactions_in_truncated_middle"]
    man["pairs"] = {"subagent_files": pstats["subagent_files"], "pairs_written": len(pairs),
                    "dropped_duplicate_brief": pstats["pairs_dropped_duplicate_brief"],
                    "subagent_files_without_pair": pstats["subagent_files_without_pair"],
                    "no_pair_because_brief_turn_never_completed": pstats["pair_candidate_turn_not_complete"],
                    "no_pair_because_no_own_user_message": pstats["subagent_files_without_first_user_message"],
                    "no_pair_because_empty_report": pstats["pair_candidate_empty_report"]}
    man["kept_turn_diagnostics"] = {k: v for k, v in sorted(s2.items()) if not k.startswith("redact:")}
    man["raw_scan_counts_all_files_incl_embedded_replays"] = {k: v for k, v in sorted(gstats.items()) if not k.startswith("malformed:") and k not in (
        "lines", "records", "blank_lines", "malformed_lines", "scalar_lines", "empty_file", "files_without_session_meta",
        "files_with_embedded_parent_context", "files_with_malformed_or_scalar_lines")}
    man["headline"] = {
        "files_read": man["files_read"], "malformed_lines": man["malformed_lines"],
        "turns_found": man["turn_instances_found"], "unique_turn_ids_found": man["unique_turn_ids_found"],
        "turns_after_dedup": man["turns_after_dedup"], "turns_kept": man["turns_kept"],
        "segments": n_rows, "rows_per_split": dict(man["rows"]),
        "outputs_truncated": s2["outputs_truncated"], "chars_removed_by_truncation": s2["chars_removed_by_truncation"],
    }
    man["bytes_per_output_file"] = sizes
    man["bytes_total_output"] = sum(sizes.values())
    man["wall_seconds"] = round(time.time() - t0, 1)
    with open(os.path.join(out, "manifest.json"), "w") as f:
        json.dump(man, f, indent=2, sort_keys=False, ensure_ascii=False)
    print(json.dumps({k: man[k] for k in ("files_read", "turn_instances_found", "turns_after_dedup", "turns_kept",
                                           "segments_total_rows", "rows", "bytes_total_output", "wall_seconds")}, indent=1))


if __name__ == "__main__":
    main()