Datasets:

FWeindel
/

validated-tasks

"""Convert the pinned MiMo-V2.6-RL-oss Code rows to Harbor tasks and archive the Cyber rows.
    python data/mimo/patch.py --source SOURCE_DIR --output NEW_DIR
SOURCE_DIR holds the pinned code.parquet, cyber.parquet and image-mapping.jsonl
and is only read. NEW_DIR receives tasks.parquet (2,600 Code tasks whose image tag still exists and
without reward-hack residue), tasks.archive.parquet (1,000 Cyber rows, 96 Code rows with
removed images and 2 with residue, original payloads),
tasks.manifest.json and the patcher snapshot.
Code semantics follow OpenSourceCodeEnvironment (opensource_code.py and
datasets/base.py, MiMo-Agent commit 467f0a19016f0ac4d63b8d17a1f0da9ba07f232c):
capture the base ref before the agent starts, reset every path the test patch
touches to that ref, git apply the test patch, run test_command in cwd, reward 1
if and only if it exits 0, then reset the touched paths again.
"""

# Support both direct execution and python -m data.<source>.patch.
import sys
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parents[2]))


import argparse
import hashlib
import io
import json
import re
import shlex
import tarfile

import pyarrow.parquet as pq

from data.utils.full_source.harbor_parquet import pack_task, write_tasks
from data.utils.patch_reporting import write_conversion_report


DATASET = "XiaomiMiMo/MiMo-V2.6-RL-oss"
REVISION = "639865fd3374018d6cb29b9fb82dd531406fcf5f"
MIMO_AGENT_REVISION = "467f0a19016f0ac4d63b8d17a1f0da9ba07f232c"
SUBSETS = {"code": ("code.parquet", 2698), "cyber": ("cyber.parquet", 1000)}
SHA256 = {
    "code.parquet": "e15733cf2451cfbc5492a4120f7f8cfddbad818aa9f0b324c79888dd1fece161",
    "cyber.parquet": "9e6e69034f4f3e6cfb5c901de642dddc7f4bacf35ab886b0e8eb53eb03084ba1",
    "image-mapping.jsonl": "704ad2716d746430a805b9b1ac7c6725aea2117fd8e4823ebe9d339d0313ad58",
}
REQUIRED_KEYS = {
    # _REQUIRED_FIELDS of opensource_code.py plus dataset_type; the rows have exactly these keys.
    "code": {"cwd", "dataset_type", "docker_image", "instance_id", "problem_statement",
             "test_command", "test_patch", "verifier_timeout_sec"},
    "cyber": {"cwd", "dataset_type", "description", "docker_image", "instance_id", "problem_statement"},
}
TASK_ID = re.compile(r"[A-Za-z0-9._-]+")
WORKDIR = re.compile(r"/[A-Za-z0-9._/-]+")

# agent_loop.py default trajectory_timeout at the pinned MiMo-Agent commit.
AGENT_TIMEOUT_SEC = 1800
# Kept outside the repository so the agent's own commits cannot move it.
BASE_REF_FILE = "/var/lib/mimo/base_ref"

CONVERSION_LABEL = "converted-to-harbor"
CONVERSION_REASON = (
    "The Code subset is published as rows of a training Parquet (problem statement, prebuilt "
    "image, hidden test patch and test command) and is graded by the MiMo-Agent environment "
    "OpenSourceCodeEnvironment, not as Harbor tasks. Each row becomes one Harbor task with its "
    "original instance ID. The instruction is the problem statement, unchanged. The environment is "
    "the published image with WORKDIR set to the row's cwd, so the repository and its dependencies "
    "are exactly the upstream ones and there is one image per task. The only addition is tmux, which "
    "the agent harness needs for its terminal: one image build step installs it from the image's own "
    "package manager when the published image has none, and does nothing otherwise. On Debian and "
    "Ubuntu that step registers root ownership for the utempter helper of libutempter0 first, "
    "because the single-user build namespace cannot assign the utmp group and the package would stay "
    "half installed (stage 3, job 997001: format-code-task-002953 could not start for that reason). "
    "The upstream reward "
    "steps are split between two verifier scripts: tests/setup.sh resets every path the test patch "
    "touches to the base commit and applies the test patch with git apply, and tests/test.sh runs "
    "the published test command in cwd, gives reward 1 if and only if it exits 0, and resets the "
    "touched paths again. A patch that cannot be reset or applied gives reward 0, as upstream. "
    "The verifier timeout is the row's verifier_timeout_sec (1800 s) and the agent timeout is the "
    "upstream default trajectory timeout (1800 s). The source has no resource or network fields, so "
    "4 CPUs, 4096 MiB and internet access are adapter defaults. Upstream records the base commit "
    "before the agent starts. Here an environment healthcheck does that before the agent runs. If "
    "that record is missing, the verifier uses the current HEAD, which differs from upstream only "
    "when an agent commits changes to files the test patch modifies. The rows contain no reference "
    "solution, so the tasks have none."
)

# Pinned image digests of the 2026-10-16 re-upload (see image-digests.json). Every Code task's
# Dockerfile uses FROM <repository>@<digest>, so a later re-push cannot change a validated task.
IMAGE_DIGESTS = json.loads((Path(__file__).parent / "image-digests.json").read_text())
REMOVED_CATEGORY = "upstream-image-removed"
REMOVED_REASON = (
    "The published image tag was deleted when xiaomimimo/mimo-v2.6-rl-oss re-pushed all tags on "
    "2026-10-16 (the registry and Docker Hub return 404), so the task environment can no longer be "
    "built. Not a task-quality finding; the original payload is preserved."
)


def pinned(image):
    """(FROM reference pinned by digest, digest), or (None, None) when the tag was removed."""
    repository, _, tag = image.rpartition(":")
    if repository != IMAGE_DIGESTS["repository"]:
        raise ValueError(f"unexpected image repository: {image}")
    if tag in IMAGE_DIGESTS["removed_tags"]:
        return None, None
    digest = IMAGE_DIGESTS["digests"].get(tag)
    if not digest:
        raise ValueError(f"no pinned digest for {image}")
    return f"{repository}@{digest}", digest


# Installed copies in the 2026-10-16 images that still differ from the repository (checked
# 2026-10-17, job 1039963, ~/autonomous-20261017/mimo-residue/probe-1039963.out: every differing
# file's blob was looked up in the repository's history with git log --all --find-object).
# Kept after the same check: format-code-task-000407 (all 3 differing airflow files are older
# repository versions), format-code-task-002070 (node_modules/nopt 7.2.1 differs only in
# package.json and the bin/ tester; lib/ equals the repository), format-code-task-002517 (no
# /testbed; the Go module cache holds the go-streams v0.12.0 release the test module requires;
# the requested S3 connector, NewS3Source, exists nowhere outside the repository).
RESIDUE_CATEGORY = "reward-hack-residue"
RESIDUE = {
    "format-code-task-000092": (
        "The image installs the repository's tests package in /usr/local/lib/python3.10/site-packages/tests; "
        "3 of its 99 files differ from the repository and none of the 3 blobs is in the repository's history "
        "(post-base content). They contain the test cases of the requested feature: test_checker.py adds "
        "var_rank_multi_1..5 and var_rank_multi_broadcast, test_parser.py and test_specs.py add specs with "
        "several variable-rank dimensions ([ds1..., ds2...]), the behaviour the hidden test "
        "tests/gpflow/experimental/check_shapes/test_multi_varrank.py checks. An agent can read the expected "
        "behaviour from the image."),
    "format-code-task-001397": (
        "The image's virtual environment /testbed/.venv/lib/python3.12/site-packages/grai_schemas has "
        "v1/metadata/edges.py with TableToTableMetadata typed by TableToTableAttributes and included in the "
        "Metadata union; the repository copy has the defect (TableToColumnAttributes, union without "
        "TableToTableMetadata) and the installed blob is not in the repository's history (post-base content). "
        "grai_client/schemas/__init__.py likewise imports an edge module the repository lacks. The installed "
        "copy is the fix for the relationship-metadata contract the hidden test checks."),
}

# Images whose package installation upstream never finished (validation 2026-10-17, final run
# array 1047484, image build logs). Stage 1 refuses an image with unconfigured packages, and
# 001314 cannot even install tmux (apt stops on the interrupted dpkg). Reproduced from the image
# digest alone, so the image as published is incomplete; not a task-content or runtime finding.
INCOMPLETE_PACKAGES_CATEGORY = "upstream-image-incomplete-packages"
INCOMPLETE_PACKAGES = {
    "format-code-task-001302": (
        "dpkg in the published image lists mysql-server as unpacked but not configured and "
        "mysql-server-8.0 as half configured (shard 0032, image build log 5911a2363542)."),
    "format-code-task-000809": (
        "dpkg in the published image lists libevent-dev as unpacked but not configured "
        "(shard 0017, image build log 3dea9d582103)."),
    "format-code-task-001314": (
        "dpkg in the published image was interrupted: apt-get stops with \"dpkg was interrupted, you "
        "must manually run 'dpkg --configure -a'\", so no package (tmux) can be installed "
        "(shard 0041, image build log 4d5d8b7feef4)."),
    "format-code-task-002899": (
        "dpkg in the published image lists typesense-server as half configured; tmux was already "
        "present, so the build installed nothing (shard 0074, image build log da84e0f3c6d8)."),
}
# No-op stage review of the final run (array 1047484, 2026-10-18, user-approved): tasks whose
# reward-0 run did not come from the missing feature. runtime-user-emulation / runtime-incompatible:
# our single-uid Apptainer runtime (services as their own users, root browsers, host network or
# CPU count); upstream-environment-broken: the image cannot run the tests at all; upstream-task-defect:
# the repository lacks files or a base commit the task needs; nop-unverifiable: no output shows that a
# test ran. Per-task evidence: verdict files of the review (scratch nop/verdicts-*.jsonl).
NOP_REVIEW = {
    'format-code-task-000348': ('upstream-environment-broken', 'Memory measurement on CLAIX (2026-10-18, 16 GiB limit): needs more than 16 GiB (peak 15,674 MiB at the 16 GiB limit; no tests ran)'),
    'format-code-task-000707': ('upstream-environment-broken', 'Memory measurement on CLAIX (2026-10-18, 16 GiB limit): its JavaScript test runs out of Node\'s own heap at about 4 GB ("JavaScript heap out of memory"), which the test command does not raise; more container memory does not help'),
    'format-code-task-001633': ('upstream-environment-broken', 'Memory measurement on CLAIX (2026-10-18, 16 GiB limit): its JavaScript test runs out of Node\'s own heap ("Reached heap limit"), which the test command does not raise; more container memory does not help'),
    'format-code-task-002280': ('upstream-environment-broken', 'Memory measurement on CLAIX (2026-10-18, 16 GiB limit): needs more than 16 GiB (peak reached the 16 GiB limit)'),
    'format-code-task-000252': ('upstream-environment-broken', 'No-op rerun on CLAIX (job 5117303, 2026-10-18, direct internet): the StarRocks fe-core Maven build ran 19 minutes and then exceeded the task memory limit; too slow and too large for validation (user rule: Maven tasks over 20 minutes are archived).'),
    'format-code-task-001107': ('upstream-environment-broken', 'No-op rerun on CLAIX (job 5117303, 2026-10-18, direct internet): Gradle cannot resolve the com.gradle.enterprise 3.16.2 settings plugin from any plugin repository, so the build fails before any test.'),
    'format-code-task-002916': ('upstream-environment-broken', 'No-op rerun on CLAIX (job 5117303, 2026-10-18, direct internet): pnpm install --frozen-lockfile hangs without output until the 1800 s verifier timeout, so no test runs.'),
    'format-code-task-000020': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18, traced with bash -x): Julia segfaults in Pkg.instantiate before any test runs.'),
    'format-code-task-000111': ('runtime-incompatible', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Firefox never connected to karma under xvfb, so no spec ran; the likely cause is Firefox under single-uid fakeroot root without user namespaces for its sandbox (not proven by the log).'),
    'format-code-task-000130': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18, traced with bash -x): Julia segfaults before any test runs (same image setup as 000020).'),
    'format-code-task-000177': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Postgres not running; Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e.'),
    'format-code-task-000178': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): pg_ctlcluster start failed silently (2>/dev/null); Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e.'),
    'format-code-task-000179': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): pg_ctlcluster start failed (hidden); Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e.'),
    'format-code-task-000181': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Postgres did not start and sudo -u postgres used; Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e.'),
    'format-code-task-000182': ('runtime-incompatible', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Script writes /etc/hosts, which Apptainer binds read-only.'),
    'format-code-task-000209': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): pg_ctlcluster start fails under the single-uid runtime (root-owned PGDATA, cf. 001272); the first selected base test errors in DB setup.'),
    'format-code-task-000277': ('nop-unverifiable', 'No-op stage (final run array 1047484, reviewed 2026-10-18, traced with bash -x): forge test -q prints only compile lines; no test result is visible.'),
    'format-code-task-000279': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): pg_ctlcluster start failed silently (output sent to /dev/null) and the tests never ran; this is the same root-owned datadir/single-uid issue as the other postgres tasks (inferred).'),
    'format-code-task-000394': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18, traced with bash -x): usecase.sh stops at its precondition "FAIL: molecule CLI is not runnable": the project CLI does not start in the image.'),
    'format-code-task-000474': ('upstream-task-defect', 'The repository has no commit (all files only staged, the base commit exists only as an unreferenced object), so the base ref the task reset needs cannot be recorded; the environment healthcheck fails (CLAIX rerun 5116864, 2026-10-18).'),
    'format-code-task-000528': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): MySQL cannot open its datadir because under the single-uid Apptainer runtime the mysql user cannot own or access /var/lib/mysql, so the tests never ran.'),
    'format-code-task-000622': ('runtime-user-emulation', "No-op stage (final run array 1047484, reviewed 2026-10-18): HOME/GOMODCACHE belong to the image's testuser; under the single-uid runtime the directory is not accessible, so go cannot start."),
    'format-code-task-000723': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): MariaDB did not come up (mysql client errors hidden, script exits under -e); inferred same single-uid ownership problem as 001983.'),
    'format-code-task-000945': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Postgres not running (pg_ctlcluster start fails silently, same root-owned PGDATA problem as 001272); additionally our ot_pytest_execution plugin crashes on pytest 3.0.2 (no _pytest.outcomes).'),
    'format-code-task-000965': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Offline Maven repo lacks parent POM; no test compiled.'),
    'format-code-task-001030': ('upstream-task-defect', "No-op stage (final run array 1047484, reviewed 2026-10-18): Upstream 'task base' commit deleted 64 files (incl. internal/apischema/openai/openai.go, 60 of them mention LLM vendors) that the instruction calls the existing extension; the package cannot build regardless of the feature (task defective)."),
    'format-code-task-001050': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Repo has go.mod requiring testify but no go.sum at all (checked in image); `go test ./lib/lru` cannot load the module graph independent of the feature (new test imports only reflect/testing).'),
    'format-code-task-001059': ('upstream-task-defect', "No-op stage (final run array 1047484, reviewed 2026-10-18): Upstream 'task base' deleted common/openai/{chat,exports,msg}.go (all mention OpenAI), so yakgrpc no longer compiles; unrelated to the YaklangInspectInformation feature (task defective)."),
    'format-code-task-001069': ('upstream-task-defect', "No-op stage (final run array 1047484, reviewed 2026-10-18): Upstream 'task base' deleted 76 files incl. src/any_agent/config.py and tracing/agent_trace.py; the package cannot be imported at all although the instruction only asks to extend spans_to_messages (task defective)."),
    'format-code-task-001077': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18, traced with bash -x): a jar the test classpath needs (commons-lang 2.6) is missing from the offline Maven repository in the image.'),
    'format-code-task-001098': ('upstream-task-defect', 'No-op stage (final run array 1047484, reviewed 2026-10-18): The prerequisite lib-sourcify build fails on a missing ipfsHash module in existing code, which is unrelated to the custom-API feature, so the tests never run (the repo snapshot may be missing a file, which would make the task defective).'),
    'format-code-task-001137': ('upstream-task-defect', "No-op stage (final run array 1047484, reviewed 2026-10-18): Upstream 'task base' deleted doc_comments_ai/app.py, llm.py, utils.py (and an existing test response_parser_test.py also fails); the instruction presumes the CLI exists (task defective)."),
    'format-code-task-001168': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18): conftest import fails because mlflow imports as a namespace package (mlflow.__file__ is None), unrelated to log_table modes; environment/repo state broken.'),
    'format-code-task-001172': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Maven offline cache lacks junit-bom; build cannot even load the POM.'),
    'format-code-task-001183': ('upstream-task-defect', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Same griptape repo state as 001406: conftest imports an existing driver module that is missing; unrelated to retrieval fusion (likely defective task).'),
    'format-code-task-001188': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18, traced with bash -x): the test command starts mysqld as the mysql user, which our single-user runtime cannot provide.'),
    'format-code-task-001193': ('runtime-incompatible', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Flow server crashes in select(); most likely too many worker fds because flow sizes workers from the 384 host CPUs (select fd limit 1024). Not task-related.'),
    'format-code-task-001203': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Silent exit at first service start: pg_ctlcluster ... 2>/dev/null fails under -e; Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e.'),
    'format-code-task-001207': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Silent exit: service postgresql start / sudo -u postgres psql 2>/dev/null fail under -e; Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e.'),
    'format-code-task-001212': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Bench dir owned by frappe uid is unreadable for root under the single-uid fakeroot runtime.'),
    'format-code-task-001272': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Postgres refuses to start because under the single-uid Apptainer runtime PGDATA appears root-owned; base tests error in setup before any feature test runs.'),
    'format-code-task-001273': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): PostgreSQL refuses to start because under fakeroot/single-uid the datadir appears owned by root; the tests never ran.'),
    'format-code-task-001406': ('upstream-task-defect', 'No-op stage (final run array 1047484, reviewed 2026-10-18): conftest import chain fails on an existing griptape driver module unrelated to the RAG-provenance feature; repo snapshot appears to be missing that file (task likely defective).'),
    'format-code-task-001438': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): pg_ctlcluster start failed (errors hidden); Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e.'),
    'format-code-task-001561': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Existing repository code imports dict2schema, which the installed webargs version no longer has, so the import precheck fails before pytest runs (a dependency version mismatch, not the feature).'),
    'format-code-task-001574': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): pg_ctlcluster start failed silently (output sent to /dev/null), so ALTER ROLE failed and no tests ran; this matches the postgres datadir-owned-by-root failure under fakeroot/single-uid (inferred).'),
    'format-code-task-001756': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Composer resolved latest deps (no lock) that need PHP 8 while PHP 7.4 runs; errors come from vendor code.'),
    'format-code-task-001757': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18): PHP 8.0 parse error from a typed class constant in an installed dependency (needs PHP 8.3); unrelated to onSuccess/onFailure. Other 3 tests pass.'),
    'format-code-task-001866': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): PostgreSQL never became ready, so no go test ran; this matches the root-owned datadir/single-uid problem of the other postgres tasks (inferred, the start output is suppressed).'),
    'format-code-task-001969': ('nop-unverifiable', 'No-op stage (final run array 1047484, reviewed 2026-10-18, traced with bash -x): npm aborts with exit code 243 after the install step with its output hidden; no test runs visibly.'),
    'format-code-task-001975': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): RabbitMQ cannot read its cookie (owned by rabbitmq uid) under the single-uid runtime; tests never ran.'),
    'format-code-task-001980': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Postgres cannot run as root and the single-uid runtime has no postgres user to switch to.'),
    'format-code-task-001982': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): MariaDB did not start, mysql error hidden by 2>/dev/null; inferred single-uid ownership problem as in 001983.'),
    'format-code-task-001983': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): mysqld_safe cannot manage its pid file: uid mapping / overlay under fakeroot.'),
    'format-code-task-002001': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): su to a non-root user cannot use its home under single-uid runtime.'),
    'format-code-task-002022': ('runtime-user-emulation', "No-op stage (final run array 1047484, reviewed 2026-10-18): Silent exit at 'pg_ctlcluster 16 main start 2>/dev/null' under -e; Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e."),
    'format-code-task-002023': ('runtime-user-emulation', "No-op stage (final run array 1047484, reviewed 2026-10-18): Silent exit at 'pg_ctlcluster 16 main start 2>/dev/null' under -e; Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e."),
    'format-code-task-002024': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Postgres not running; Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e.'),
    'format-code-task-002035': ('runtime-user-emulation', "No-op stage (final run array 1047484, reviewed 2026-10-18): Silent exit at 'pg_ctlcluster 16 main start 2>/dev/null' under -e; Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e."),
    'format-code-task-002036': ('runtime-user-emulation', "No-op stage (final run array 1047484, reviewed 2026-10-18): postgres never comes up (the start command's stderr is sent to /dev/null); same class as the known DB-ownership failures under our single-uid fakeroot runtime. Per user decision archive as runtime-user-emulation only after confirming the ownership error (start it once by hand without 2>/dev/null)."),
    'format-code-task-002037': ('runtime-user-emulation', "No-op stage (final run array 1047484, reviewed 2026-10-18): postgres never comes up (the start command's stderr is sent to /dev/null); same class as the known DB-ownership failures under our single-uid fakeroot runtime. Per user decision archive as runtime-user-emulation only after confirming the ownership error (start it once by hand without 2>/dev/null)."),
    'format-code-task-002038': ('runtime-user-emulation', "No-op stage (final run array 1047484, reviewed 2026-10-18): postgres never comes up (the start command's stderr is sent to /dev/null); same class as the known DB-ownership failures under our single-uid fakeroot runtime. Per user decision archive as runtime-user-emulation only after confirming the ownership error (start it once by hand without 2>/dev/null). Django aborts while creating the test DB, so the runner never reports results."),
    'format-code-task-002092': ('runtime-user-emulation', "No-op stage (final run array 1047484, reviewed 2026-10-18): postgres never comes up (the start command's stderr is sent to /dev/null); same class as the known DB-ownership failures under our single-uid fakeroot runtime. Per user decision archive as runtime-user-emulation only after confirming the ownership error (start it once by hand without 2>/dev/null)."),
    'format-code-task-002144': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18, traced with bash -x): the junit 3.8.1 jar the test command needs is missing from the offline Maven repository in the image.'),
    'format-code-task-002148': ('upstream-task-defect', "No-op stage (final run array 1047484, reviewed 2026-10-18): Upstream 'task base' deleted api/api.gen.go (plus clients/openapi specs), so package api cannot compile; unrelated to the subscription phase-patch feature (task defective)."),
    'format-code-task-002169': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Neo4j cannot write its log file (owned by the neo4j uid, not writable under single-uid fakeroot), so the database never starts and pytest never runs.'),
    'format-code-task-002180': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Postgres refuses to start because the single-uid fakeroot runtime makes its data dir root-owned; go tests never ran.'),
    'format-code-task-002293': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Old pnpm on newer Node fails every registry fetch (ERR_INVALID_THIS); dependencies never installed.'),
    'format-code-task-002294': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Silent exit in postgres setup (pg_ctlcluster ... || true, then sudo -u postgres psql with errors hidden); Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e.'),
    'format-code-task-002309': ('runtime-incompatible', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Karma never ran a spec because Chrome refuses to start as uid 0 without --no-sandbox; under fakeroot the verifier is root (Chrome would also need --no-sandbox in a root Docker container, so the task config is fragile too).'),
    'format-code-task-002338': ('runtime-user-emulation', "No-op stage (final run array 1047484, reviewed 2026-10-18): postgres never comes up (the start command's stderr is sent to /dev/null); same class as the known DB-ownership failures under our single-uid fakeroot runtime. Per user decision archive as runtime-user-emulation only after confirming the ownership error (start it once by hand without 2>/dev/null). Also uses su - postgres, which cannot switch uid with empty subuid."),
    'format-code-task-002396': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): The setup cannot read the postgres-owned pg_hba.conf because single-uid fakeroot "root" lacks access to files owned by another image uid, so the tests never ran.'),
    'format-code-task-002557': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Silent exit in postgres setup (su - postgres psql 2>/dev/null under -e); Postgres data dir owned by root under single-uid fakeroot (seen explicitly in 002180/002732); start errors hidden by 2>/dev/null, script exits under -e.'),
    'format-code-task-002558': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): test_command.sh runs `pg_ctlcluster 16 main start 2>/dev/null || true`; Postgres cannot start under the single-uid runtime (cf. 001272).'),
    'format-code-task-002559': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): postgresql init script fails on chmod under fakeroot/overlay uid mapping.'),
    'format-code-task-002613': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18): The pinned old PHPUnit cannot run on the installed PHP 8.1+, so it dies on startup before any test executes (unrelated to the feature).'),
    'format-code-task-002620': ('runtime-incompatible', 'No-op stage (final run array 1047484, reviewed 2026-10-18): The new sent_time test fails genuinely, but read_routes6 also fails because the Apptainer container shares the host network namespace (host ens13f0np0/ibs5 routes), so even a correct solution would score 0 here.'),
    'format-code-task-002629': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): MariaDB cannot be reached because of a permission error (errno 13) on the socket or datadir; under single-uid fakeroot the mysql user cannot use root-owned dirs, so the tests never ran.'),
    'format-code-task-002689': ('runtime-incompatible', 'No-op stage (final run array 1047484, reviewed 2026-10-18): The setup step chmods a symlink into the read-only image tree (/opt/node-22/...), and under Apptainer fakeroot this gives EINVAL, so the script exits before jest.'),
    'format-code-task-002693': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): The test_command.sh runs under bash -e and starts postgres with stderr sent to /dev/null; a failed pg_ctlcluster start ends the script silently before go test, which matches the postgres-under-fakeroot failures (inferred from the script, as the output is empty).'),
    'format-code-task-002732': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Postgres refuses root-owned data dir under single-uid fakeroot.'),
    'format-code-task-002740': ('runtime-incompatible', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Taskpools sizes from host CPU count (384 visible) and aborts in the regression test before the new test; our node exposes all host cores.'),
    'format-code-task-002850': ('runtime-user-emulation', 'No-op stage (final run array 1047484, reviewed 2026-10-18, traced with bash -x): pip install -e . fails with "Permission denied: /usr/local/bin/sqlmesh" under our single-user runtime, so no test runs.'),
    'format-code-task-003020': ('upstream-task-defect', 'No-op stage (final run array 1047484, reviewed 2026-10-18): Base app/reducers/links.js imports react-addons-update, which is not installed; bundle fails before tests. Dependency missing in image (possible task defect).'),
    'format-code-task-003039': ('upstream-environment-broken', 'No-op stage (final run array 1047484, reviewed 2026-10-18): mongod failed to start (log not captured, possibly runtime-related); tests never ran.'),
    'format-code-task-003053': ('runtime-user-emulation', "No-op stage (final run array 1047484, reviewed 2026-10-18): postgres never comes up (the start command's stderr is sent to /dev/null); same class as the known DB-ownership failures under our single-uid fakeroot runtime. Per user decision archive as runtime-user-emulation only after confirming the ownership error (start it once by hand without 2>/dev/null)."),
}
# Leftover from the residue audit that is still in the re-uploaded image (checked 2026-10-18): the
# new API names of the hidden tests exist only in that leftover. Six other leftovers of the audit
# (000046, 000283, 000556, 000699, 000114, 001372) contain none of them and stay.
RESIDUE_UNTESTED = {
    'format-code-task-000253': ("The image's .build_env/venv installs a pipeline package whose "
        "pipeline/contrib/node_timeout (handlers.py and its bytecode) already defines deal_with_timeout_node, "
        "parse_node_timeout_configs, forced_fail_and_skip and node_timeout_handler, the API the hidden tests check "
        "and the repository lacks (residue audit CharlieFRuan/rl-env-cleaner@836c6061, checked 2026-10-18)."),
}
# Every Code task archived by name: (category, reason).
EXCLUDED = {**{task: (RESIDUE_CATEGORY, reason) for task, reason in RESIDUE.items()},
            **{task: (RESIDUE_CATEGORY, reason) for task, reason in RESIDUE_UNTESTED.items()},
            **{task: (INCOMPLETE_PACKAGES_CATEGORY, reason) for task, reason in INCOMPLETE_PACKAGES.items()},
            **NOP_REVIEW}

ARCHIVE_CATEGORY = "unsupported-user-isolation"
ARCHIVE_REASON = (
    "Archived for compatibility with the current Helma Apptainer runtime: upstream Cyber grading "
    "requires separate agent, verify and root identities. The tested Helma account configuration "
    "has no subordinate UID/GID mappings, and switching to the required users fails. The current "
    "runtime cannot preserve the upstream grading permissions. This is a runtime compatibility "
    "exclusion, not a task-quality or reward-hacking finding; the original payload is preserved "
    "for future re-enablement. No claim is made that cluster policy prohibits these mappings."
)

# _setup_dataset_specific: safe.directory (result ignored, as upstream), _ensure_work_tree,
# then _capture_base_ref (without overwriting an earlier capture). A base ref file that
# cannot be written is not fatal; the verifier then resolves HEAD itself.
CAPTURE_BASE_REF = (
    "git config --global --add safe.directory {cwd}; cd {cwd} && "
    "{{ git rev-parse --git-dir >/dev/null 2>&1 || "
    "{{ git init -q && git add -A && git commit -q -m baseline --allow-empty; }}; }} && "
    "ref=$(git rev-parse --verify HEAD) && [ ${{#ref}} -eq 40 ] && "
    "{{ [ -s {file} ] || {{ mkdir -p {directory} && printf '%s\\n' \"$ref\" > {file}; }} || true; }}"
)

# The harness terminal needs tmux. The shared builder adds it in most builds, but not when it has
# to import the published image as is (job 997001, format-code-task-002953: Debian 11), and an
# installation at task start cannot give libutempter0's helper its root:utmp ownership in a
# single-UID namespace. Without the helper's privileged utmp accounting tmux works the same.
UTEMPTER = "/usr/lib/x86_64-linux-gnu/utempter/utempter"
INSTALL_TMUX = (
    "if ! command -v tmux >/dev/null 2>&1; then "
    "if command -v apt-get >/dev/null 2>&1; then "
    f"{{ dpkg-statoverride --list {UTEMPTER} >/dev/null 2>&1 || "
    f"dpkg-statoverride --add root root 0755 {UTEMPTER}; }} && "
    "apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends tmux; "
    "elif command -v apk >/dev/null 2>&1; then apk add --no-cache tmux; "
    "elif command -v dnf >/dev/null 2>&1; then dnf install -y -q tmux; "
    "elif command -v yum >/dev/null 2>&1; then yum install -y -q tmux; "
    # An explicit exit: errexit does not stop a build script for a failure inside the if.
    "fi; fi; tmux -V || exit 1"
)

RESET_SH = r"""#!/bin/bash
# Upstream reset_test_files_to_base: give every path the test patch touches a
# clean starting point at the base commit. Sourced by setup.sh and test.sh.
mimo_base_ref() {
  local ref=""
  if [ -s __BASE_REF_FILE__ ]; then
    ref=$(head -n 1 __BASE_REF_FILE__)
    if [ ${#ref} -eq 40 ] && git cat-file -e "$ref^{commit}" 2>/dev/null; then
      echo "base ref $ref (recorded before the agent started)" >&2
      printf '%s\n' "$ref"
      return 0
    fi
  fi
  # No usable record (the healthcheck did not run): same steps as upstream setup.
  if ! git rev-parse --git-dir >/dev/null 2>&1; then
    git init -q && git add -A && git commit -q -m baseline --allow-empty || return 1
  fi
  ref=$(git rev-parse --verify HEAD) || return 1
  [ ${#ref} -eq 40 ] || return 1
  echo "base ref $ref (current HEAD, no usable record at __BASE_REF_FILE__)" >&2
  printf '%s\n' "$ref"
}
mimo_reset_test_files() {
  local base tf
  cd __CWD__ || return 1
  base=$(mimo_base_ref) || return 1
  while IFS= read -r tf; do
    [ -z "$tf" ] && continue
    if git cat-file -e "$base":"$tf" 2>/dev/null; then
      git checkout "$base" -- "$tf" 2>/dev/null || true
    else
      git rm -f --cached "$tf" >/dev/null 2>&1 || true
      rm -f "$tf"
    fi
  done < /tests/touched_paths.txt
}
"""

SETUP_SH = r"""#!/bin/bash
# Verifier preparation, upstream reward steps 1 and 2: reset the paths the test
# patch touches to the base commit, then apply the test patch.
fail() {
  echo "$1" >&2
  # Upstream scores a testbed that cannot be reset or patched as reward 0.
  mkdir -p /logs/verifier
  rm -f /logs/verifier/reward.json
  echo 0 > /logs/verifier/reward.txt
  exit 1
}
git config --global --add safe.directory __CWD__
. /tests/reset_test_files.sh
mimo_reset_test_files || fail "reset_tests_failed"
cd __CWD__ || fail "reset_tests_failed"
git apply --verbose /tests/test.patch || fail "apply_test_patch_failed"
"""

TEST_SH = r"""#!/bin/bash
# Upstream reward step 3: reward 1 if and only if the published test command exits 0.
mkdir -p /logs/verifier
if cd __CWD__; then
  bash -lc __TEST_COMMAND__
  status=$?
else
  status=1
fi
if [ "$status" -eq 0 ]; then
  printf '{"reward": 1.0}\n' > /logs/verifier/reward.json
else
  printf '{"reward": 0.0}\n' > /logs/verifier/reward.json
fi
# Upstream leaves the test files as they were found.
. /tests/reset_test_files.sh
mimo_reset_test_files
exit 0
"""


def get_patch_touched_files(patch):
    """Upstream DatasetEnvironment.get_patch_touched_files, unchanged."""
    files = []
    for line in patch.split("\n"):
        m = re.match(r"^diff --git a/(.+?) b/(.+)$", line)
        if m:
            for fp in (m.group(1), m.group(2)):
                if fp not in files:
                    files.append(fp)
    return files


def sha256(path):
    with Path(path).open("rb") as stream:
        return hashlib.file_digest(stream, "sha256").hexdigest()


def check_source(source):
    """Pinned hashes are constants; the loop's input manifest must agree with them."""
    manifest = source.parent / "input-manifest.json"
    recorded = json.loads(manifest.read_text()) if manifest.is_file() else None
    if recorded is not None:
        if recorded.get("dataset") != DATASET or recorded.get("revision") != REVISION:
            raise ValueError(f"{manifest}: not {DATASET} at {REVISION}")
        if recorded.get("domains") != {name: count for name, (_, count) in SUBSETS.items()}:
            raise ValueError(f"{manifest}: unexpected subset counts {recorded.get('domains')}")
    for name, expected in SHA256.items():
        path = source / name
        if not path.is_file():
            raise FileNotFoundError(f"missing pinned MiMo source file: {path}")
        actual = sha256(path)
        if actual != expected:
            raise ValueError(f"{path}: sha256 {actual} differs from the pinned {expected}")
        if recorded is not None and recorded.get("sha256", {}).get(name) != expected:
            raise ValueError(f"{manifest}: sha256 of {name} differs from the pinned {expected}")


def read_mapping(path):
    mapping = {}
    for line in path.read_text().splitlines():
        entry = json.loads(line)
        if mapping.setdefault(entry["dataset_image"], entry["dockerhub_image"]) != entry["dockerhub_image"]:
            raise ValueError(f"conflicting image mapping: {entry['dataset_image']}")
    return mapping


def read_subset(source, subset, mapping):
    """Yield (row index, source row, decoded instance, published image) for every row."""
    filename, expected = SUBSETS[subset]
    rows = pq.read_table(source / filename).to_pylist()
    if len(rows) != expected:
        raise ValueError(f"{filename}: expected {expected} rows, found {len(rows)}")
    for index, row in enumerate(rows):
        instance = json.loads(row["extra_info"]["instance_json"])
        where = f"{filename} row {index}"
        if set(instance) != REQUIRED_KEYS[subset]:
            raise ValueError(f"{where}: unexpected instance keys {sorted(instance)}")
        empty = [key for key in sorted(instance) if not instance[key] or not isinstance(
            instance[key], (str, int) if key == "verifier_timeout_sec" else str)]
        if empty:
            raise ValueError(f"{where}: empty or mistyped field(s) {empty}")
        task_id = instance["instance_id"]
        if row["extra_info"]["instance_id"] != task_id:
            raise ValueError(f"{where}: instance ID mismatch {row['extra_info']['instance_id']!r} != {task_id!r}")
        if not TASK_ID.fullmatch(task_id):
            raise ValueError(f"{where}: unsafe instance ID {task_id!r}")
        if instance["docker_image"] not in mapping:
            raise ValueError(f"{where}: unmapped image {instance['docker_image']!r}")
        yield index, row, instance, mapping[instance["docker_image"]]


# Tasks that exceed the 4096 MiB default. Measured 2026-10-17 with a 16 GiB limit (job 1046057,
# cgroup peak of the no-op run, ~/OT_next_data validation of stages 1,3,4): jest starts worker
# processes and MiMo's /etc/profile.d/zz-autoinstall-oom.sh gives every node a 4 GiB heap, so
# 4096 MiB stopped the verifier ("Task memory limit exceeded"). Peak plus about 35 % headroom.
# Tasks that exceed the 4096 MiB default. Measured 2026-10-18 with a 16 GiB limit (CLAIX jobs
# 5117335, 5120667, 5120669; peak = sacct MaxRSS of the trial step). Limit = peak + 35 %, rounded up
# to the next size Terminal-Bench's check-resource-sizes.sh allows (8192 or 16384 MiB): all of them failed at
# 4096 MiB, where the limit also counts page cache.
MEMORY_MB = {
    "format-code-task-000035": 8192,  # peak 1,907 MiB
    "format-code-task-000036": 8192,  # 2026-10-17 jest profile
    "format-code-task-000037": 8192,  # peak 2,278 MiB
    "format-code-task-000197": 8192,  # peak 2,949 MiB
    "format-code-task-000268": 8192,  # peak 5,261 MiB
    "format-code-task-000283": 16384,  # peak 10,921 MiB
    "format-code-task-000319": 8192,  # peak 2,197 MiB
    "format-code-task-000475": 8192,  # peak 1,842 MiB
    "format-code-task-000479": 8192,  # peak 2,201 MiB
    "format-code-task-000576": 8192,  # peak 3,304 MiB
    "format-code-task-000583": 16384,  # peak 6,245 MiB
    "format-code-task-000604": 8192,  # peak 3,330 MiB
    "format-code-task-000609": 16384,  # peak 10,688 MiB
    "format-code-task-000632": 8192,  # peak 2,120 MiB
    "format-code-task-000656": 8192,  # peak 2,914 MiB
    "format-code-task-000756": 8192,  # peak 3,327 MiB
    "format-code-task-000772": 8192,  # peak 5,851 MiB
    "format-code-task-000776": 8192,  # peak 3,063 MiB
    "format-code-task-000867": 8192,  # peak 2,992 MiB
    "format-code-task-000890": 8192,  # peak 2,675 MiB
    "format-code-task-000891": 8192,  # peak 1,607 MiB
    "format-code-task-000928": 8192,  # peak 4,173 MiB
    "format-code-task-000943": 8192,  # peak 2,211 MiB
    "format-code-task-001085": 8192,  # peak 5,885 MiB
    "format-code-task-001239": 8192,  # peak 4,485 MiB
    "format-code-task-001379": 8192,  # peak 2,369 MiB
    "format-code-task-001388": 16384,  # peak 6,689 MiB
    "format-code-task-001403": 8192,  # peak 1,994 MiB
    "format-code-task-001408": 8192,  # peak 2,517 MiB
    "format-code-task-001409": 8192,  # peak 2,673 MiB
    "format-code-task-001549": 8192,  # peak 2,617 MiB
    "format-code-task-001572": 8192,  # peak 2,790 MiB
    "format-code-task-001625": 8192,  # peak 3.9 GiB
    "format-code-task-001632": 8192,  # peak 4,517 MiB
    "format-code-task-001656": 8192,  # peak 2,122 MiB
    "format-code-task-001698": 8192,  # peak 2,526 MiB
    "format-code-task-001856": 16384,  # peak 6,821 MiB
    "format-code-task-001883": 8192,  # peak 1,652 MiB
    "format-code-task-001890": 8192,  # peak 2,300 MiB
    "format-code-task-001907": 16384,  # peak 6,363 MiB
    "format-code-task-001954": 8192,  # peak 954 MiB
    "format-code-task-001955": 8192,  # peak 1,696 MiB
    "format-code-task-002057": 16384,  # peak 6,519 MiB
    "format-code-task-002080": 8192,  # peak 1,699 MiB
    "format-code-task-002141": 8192,  # peak 2,642 MiB
    "format-code-task-002242": 8192,  # peak 1,916 MiB
    "format-code-task-002317": 8192,  # peak 1,744 MiB
    "format-code-task-002318": 8192,  # peak 3,028 MiB
    "format-code-task-002319": 8192,  # peak 2,697 MiB
    "format-code-task-002320": 8192,  # peak 1,800 MiB
    "format-code-task-002322": 8192,  # peak 1,785 MiB
    "format-code-task-002323": 8192,  # peak 1,996 MiB
    "format-code-task-002324": 8192,  # 2026-10-17 jest profile
    "format-code-task-002325": 8192,  # peak 1,969 MiB
    "format-code-task-002326": 8192,  # peak 2,034 MiB
    "format-code-task-002327": 8192,  # peak 1,995 MiB
    "format-code-task-002328": 8192,  # peak 3,419 MiB
    "format-code-task-002329": 8192,  # peak 2,297 MiB
    "format-code-task-002330": 8192,  # peak 2,796 MiB
    "format-code-task-002331": 8192,  # peak 2,620 MiB
    "format-code-task-002332": 8192,  # peak 1,718 MiB
    "format-code-task-002414": 8192,  # peak 5,542 MiB
    "format-code-task-002432": 8192,  # peak 4,233 MiB
    "format-code-task-002433": 8192,  # peak 4,367 MiB
    "format-code-task-002435": 8192,  # peak 4,652 MiB
    "format-code-task-002436": 8192,  # peak 4,639 MiB
    "format-code-task-002437": 8192,  # peak 4,705 MiB
    "format-code-task-002438": 8192,  # peak 3,726 MiB
    "format-code-task-002439": 8192,  # peak 4,900 MiB
    "format-code-task-002440": 8192,  # peak 4,933 MiB
    "format-code-task-002441": 8192,  # peak 5,005 MiB
    "format-code-task-002491": 8192,  # peak 1,638 MiB
    "format-code-task-002492": 8192,  # peak 2,799 MiB
    "format-code-task-002711": 8192,  # peak 5,542 MiB
    "format-code-task-002728": 8192,  # peak 2,589 MiB
    "format-code-task-002730": 8192,  # peak 2,095 MiB
    "format-code-task-002741": 8192,  # peak 1,736 MiB
    "format-code-task-002751": 8192,  # peak 2,152 MiB
    "format-code-task-002752": 16384,  # peak 7,120 MiB
    "format-code-task-002753": 8192,  # peak 5,186 MiB
    "format-code-task-002764": 8192,  # peak 1,730 MiB
    "format-code-task-002766": 8192,  # peak 2,211 MiB
    "format-code-task-002767": 8192,  # peak 2,085 MiB
    "format-code-task-002768": 8192,  # peak 1,931 MiB
    "format-code-task-002777": 8192,  # peak 5,629 MiB
    "format-code-task-002797": 8192,  # peak 2,084 MiB
    "format-code-task-002820": 8192,  # peak 1,940 MiB
    "format-code-task-002821": 8192,  # peak 1,883 MiB
    "format-code-task-002896": 8192,  # peak 1,773 MiB
    "format-code-task-002905": 8192,  # peak 3,517 MiB
    "format-code-task-002911": 8192,  # peak 3.9 GiB
    "format-code-task-002945": 8192,  # peak 3,846 MiB
    "format-code-task-002967": 8192,  # peak 2,331 MiB
    "format-code-task-002988": 8192,  # peak 1,823 MiB
}


def code_files(row, instance, image):
    task_id, cwd = instance["instance_id"], instance["cwd"]
    reference, digest = pinned(image)
    if not WORKDIR.fullmatch(cwd):
        raise ValueError(f"{task_id}: unsupported cwd {cwd!r}")
    if [message["content"] for message in row["prompt"]] != [instance["problem_statement"]]:
        raise ValueError(f"{task_id}: prompt differs from problem_statement")
    touched = get_patch_touched_files(instance["test_patch"])
    if not touched or any(not path or path != path.strip() or "\r" in path for path in touched):
        raise ValueError(f"{task_id}: test patch paths cannot be listed one per line: {touched!r}")
    quoted = shlex.quote(cwd)
    health = CAPTURE_BASE_REF.format(cwd=quoted, file=BASE_REF_FILE, directory=str(Path(BASE_REF_FILE).parent))
    fill = lambda text: text.replace("__CWD__", quoted).replace("__BASE_REF_FILE__", BASE_REF_FILE)
    task_toml = (
        'version = "1.0"\n'
        f'[agent]\ntimeout_sec = {AGENT_TIMEOUT_SEC}\n'
        f'[verifier]\ntimeout_sec = {int(instance["verifier_timeout_sec"])}\n'
        f'[environment]\ncpus = 4\nmemory_mb = {MEMORY_MB.get(task_id, 4096)}\nallow_internet = true\n'
        '[environment.healthcheck]\n'
        f'command = {json.dumps(health)}\n'
        'timeout_sec = 300\nretries = 1\n'
        '[metadata]\n'
        f'source = {json.dumps(DATASET)}\n'
        'subset = "code"\n'
        f'source_revision = {json.dumps(REVISION)}\n'
        f'dataset_image = {json.dumps(instance["docker_image"])}\n'
        f'dockerhub_image = {json.dumps(image)}\n'
        f'image_digest = {json.dumps(digest)}\n'
    )
    return {
        "instruction.md": instance["problem_statement"].encode(),
        "environment/Dockerfile": f"FROM {reference}\nWORKDIR {cwd}\nRUN {INSTALL_TMUX}\n",
        "task.toml": task_toml,
        # Bytes as published: CRLF content and Git binary hunks must survive.
        "tests/test.patch": instance["test_patch"].encode(),
        "tests/touched_paths.txt": "".join(path + "\n" for path in touched),
        "tests/reset_test_files.sh": fill(RESET_SH),
        "tests/setup.sh": (fill(SETUP_SH), 0o755),
        "tests/test.sh": (fill(TEST_SH).replace("__TEST_COMMAND__", shlex.quote(instance["test_command"])), 0o755),
    }


def original_payload(filename, index, row, instance, image):
    """The complete source row, sufficient to rebuild it without the upstream Parquet."""
    dump = lambda value: json.dumps(value, sort_keys=True, ensure_ascii=False, indent=1) + "\n"
    return {
        "original/row.json": dump(row),
        "original/instance.json": dump(instance),
        "original/image.json": dump({
            "dataset": DATASET, "revision": REVISION, "source_file": filename, "row_index": index,
            "dataset_image": instance["docker_image"], "dockerhub_image": image}),
    }


def convert(source, output):
    check_source(source)
    mapping = read_mapping(source / "image-mapping.jsonl")
    records, code, archived = {}, {}, []

    known = set()
    for subset in SUBSETS:
        for index, row, instance, image in read_subset(source, subset, mapping):
            task_id = instance["instance_id"]
            if task_id.lower() in known:
                raise ValueError(f"duplicate instance ID: {task_id}")
            known.add(task_id.lower())
            records[task_id] = {"project": subset, "bug_id": task_id,
                                "dataset_image": instance["docker_image"], "dockerhub_image": image}
            if subset == "code" and task_id in EXCLUDED and pinned(image)[0] is not None:
                category, reason = EXCLUDED[task_id]
                archived.append({
                    "path": task_id,
                    "task_binary": pack_task(original_payload(SUBSETS[subset][0], index, row, instance, image)),
                    "archive_category": category,
                    "archive_reason": f"{category}: Code task {task_id} (published image {image}). " + reason})
            elif subset == "code" and pinned(image)[0] is None:
                archived.append({
                    "path": task_id,
                    "task_binary": pack_task(original_payload(SUBSETS[subset][0], index, row, instance, image)),
                    "archive_category": REMOVED_CATEGORY,
                    "archive_reason": f"{REMOVED_CATEGORY}: Code task {task_id} (published image {image}). "
                                      + REMOVED_REASON})
            elif subset == "code":
                code[task_id] = code_files(row, instance, image)
            else:
                archived.append({
                    "path": task_id,
                    "task_binary": pack_task(original_payload(SUBSETS[subset][0], index, row, instance, image)),
                    "archive_category": ARCHIVE_CATEGORY,
                    "archive_reason": f"{ARCHIVE_CATEGORY}: Cyber task {task_id} (published image {image}). "
                                      + ARCHIVE_REASON})
    archived.sort(key=lambda entry: entry["path"])
    output.mkdir(parents=True, exist_ok=True)
    write_tasks(sorted(code.items()), output / "tasks.parquet", expected_count=SUBSETS["code"][1] - len(removed_code(source, mapping)) - len(EXCLUDED))
    rows = [{"path": task_id, "task_binary": pack_task(files)} for task_id, files in sorted(code.items())]
    write_conversion_report(
        output, rows, archived, patcher=__file__, records=records,
        source={"dataset": DATASET, "url": f"https://huggingface.co/datasets/{DATASET}/tree/{REVISION}",
                "revision": REVISION, "task_count": len(records),
                "subsets": {name: count for name, (_, count) in SUBSETS.items()},
                "mimo_agent_revision": MIMO_AGENT_REVISION},
        # One label covers the whole adapter; the explanation appears once per PR table.
        change_labels={row["path"]: [CONVERSION_LABEL] for row in rows},
        change_reasons={row["path"]: CONVERSION_REASON for row in rows})


def removed_code(source, mapping):
    """Code task IDs whose published image tag was removed upstream."""
    return {instance["instance_id"] for _, _, instance, image in read_subset(source, "code", mapping)
            if pinned(image)[0] is None}


def self_check(source, output):
    """Re-read the written files and account for every source row."""
    mapping = read_mapping(source / "image-mapping.jsonl")
    expected = {subset: {instance["instance_id"]: (row, instance)
                         for _, row, instance, _ in read_subset(source, subset, mapping)}
                for subset in SUBSETS}
    kept = pq.read_table(output / "tasks.parquet", columns=["path"]).column("path").to_pylist()
    archive = pq.read_table(output / "tasks.archive.parquet").to_pylist()
    dropped = [row["path"] for row in archive]
    problems = []
    removed = removed_code(source, mapping)
    residue = set(EXCLUDED)
    if not residue <= set(expected["code"]) - removed:
        problems.append("archive by name names a task that is not a pinned Code task")
    if sorted(kept) != sorted(set(expected["code"]) - removed - residue) or len(set(kept)) != len(kept):
        problems.append("kept tasks differ from the Code source IDs with a pinned image and no residue")
    if sorted(dropped) != sorted(set(expected["cyber"]) | removed | residue) or len(set(dropped)) != len(dropped):
        problems.append("archived tasks differ from the Cyber source IDs plus removed-image and residue Code IDs")
    if set(kept) & set(dropped):
        problems.append("kept and archived IDs overlap")
    for row in archive:
        task_id = row["path"]
        category, subset, reason = ((REMOVED_CATEGORY, "Code", REMOVED_REASON) if task_id in removed
                                    else (EXCLUDED[task_id][0], "Code", EXCLUDED[task_id][1]) if task_id in residue
                                    else (ARCHIVE_CATEGORY, "Cyber", ARCHIVE_REASON))
        prefix = f"{category}: {subset} task {task_id} "
        if row["archive_category"] != category or not row["archive_reason"].startswith(prefix) \
                or not row["archive_reason"].endswith(reason):
            problems.append(f"{task_id}: unexpected archive category or reason")
        with tarfile.open(fileobj=io.BytesIO(row["task_binary"])) as payload:
            original = {name: json.load(payload.extractfile(name)) for name in payload.getnames()}
        source_row = expected["cyber"].get(task_id) or expected["code"].get(task_id)
        if (original.get("original/row.json"), original.get("original/instance.json")) != source_row:
            problems.append(f"{task_id}: archived payload differs from the source row")
    manifest = json.loads((output / "tasks.manifest.json").read_text())
    actions = {}
    for task in manifest["tasks"]:
        actions[task["action"]] = actions.get(task["action"], 0) + 1
    if manifest.get("complete") is not True or actions != {
            "changed": len(expected["code"]) - len(removed) - len(residue),
            "dropped": len(expected["cyber"]) + len(removed) + len(residue)}:
        problems.append(f"manifest is incomplete or has unexpected actions: {actions}")
    if any(not task.get("reason") for task in manifest["tasks"]):
        problems.append("manifest task without a reason")
    if problems:
        raise SystemExit("MiMo conversion self-check failed:\n" + "\n".join(problems[:20]))
    total = len(expected["code"]) + len(expected["cyber"])
    print(f"source rows read: {total} (code {len(expected['code'])}, cyber {len(expected['cyber'])})")
    print(f"tasks.parquet: {len(kept)} converted Code tasks pinned by image digest, label {CONVERSION_LABEL}")
    print(f"tasks.archive.parquet: {len(dropped)} archived rows ({len(expected['cyber'])} Cyber, category "
          f"{ARCHIVE_CATEGORY}; {len(removed)} Code, category {REMOVED_CATEGORY}; {len(RESIDUE) + len(RESIDUE_UNTESTED)} Code, category "
          f"{RESIDUE_CATEGORY}; {len(INCOMPLETE_PACKAGES)} Code, category {INCOMPLETE_PACKAGES_CATEGORY}; "
          f"{len(NOP_REVIEW)} Code, no-op review), "
          "original payloads equal the source rows")
    print(f"ID accounting: {len(set(kept) | set(dropped))} of {total} source IDs, no overlap")
    print(f"tasks.manifest.json: complete, {actions['changed']} changed, {actions['dropped']} dropped")


def main():
    parser = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
    parser.add_argument("--source", type=Path, required=True,
                        help="directory with the pinned code.parquet, cyber.parquet and image-mapping.jsonl")
    parser.add_argument("--output", type=Path, required=True, help="new output directory")
    args = parser.parse_args()
    if not args.source.is_dir():
        raise ValueError(f"missing pinned MiMo source directory: {args.source}")
    convert(args.source, args.output)
    self_check(args.source, args.output)


if __name__ == "__main__":
    main()