data/tasks.jsonl

{"task_id": "candidate-0036-software-data-engineering", "source_id": "candidate-0036-software-data-engineering", "domain": "terminal", "task_path": "tasks/candidate-0036-software-data-engineering", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:3db4bdcb19504df7b6a9ad1de3ed33b7ec32340dcb4ff108b6b947af39cd5bee", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair wildcard output flags in the workflow engine\n\nYou are given a frozen, pre-fix Snakemake source tree at\n`/app/vendor/snakemake`. A data workflow declares several output patterns with\nwildcard constraints. Some outputs are marked `touch`, `temp`, `protected`,\n`pipe`, or `service`. After a rule is expanded for a concrete wildcard value,\nthe resulting job must retain the flags attached to that concrete output.\n\nDiagnose the interaction between output declaration in\n`src/snakemake/rules.py` and concrete job construction in\n`src/snakemake/jobs.py`. Repair the existing production modules so flag\nclassification is based on the expanded output object and remains correct for\nall of the flags above. Do not special-case the sample names, remove wildcard\nconstraints, or add shell commands that manufacture marker files.\n\nRun the three-stage offline integration replay:\n\n```sh\ncd /app\npython3 workflow_probe.py\n```\n\nIt must write `/app/output.json` with schema version `workflow-repair-1`, two\nconcrete wildcard jobs, one touch output per job, preserved temp/protected and\npipe/service categories, and a completed downstream aggregate. The replay is\nonly valid when both existing source modules have been repaired.\n"}
{"task_id": "candidate-0109-science-robotics", "source_id": "candidate-0109-science-robotics", "domain": "terminal", "task_path": "tasks/candidate-0109-science-robotics", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:92e1760ed72a11016f8792e3f7c8ec97943ad4f16c591164e8d5d4e0b81c9ab5", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair numerical IK restart validation\n\nWork in the frozen source tree at `/app/vendor/robotics-toolbox-python`.\n\nThe numerical inverse-kinematics solvers use random restarts when no explicit initial\nconfiguration is available. That restart operation exists in both the Python solver\nhelper and the compiled ETS fast path. Models containing non-finite joint-limit\nmetadata currently reach those samplers and fail inconsistently or contaminate the\nrestart state.\n\nRepair the existing implementation with these requirements:\n\n1. Modify both `src/roboticstoolbox/robot/IK.py` and\n   `src/roboticstoolbox/ets/cpp-extensions/ik.cpp`; a one-backend repair is incomplete.\n2. Reject `NaN`, positive infinity, or negative infinity in any lower or upper joint\n   limit before consuming random-number state or producing a restart sample.\n3. Expose `ValueError` to Python callers. The diagnostic must identify joint-limit\n   validation and the offending joint index and values sufficiently to locate bad\n   robot metadata.\n4. Preserve existing valid behavior: seeded sampling remains deterministic, scalar\n   and batched requests retain shape `(count, joint_count)`, and every sample remains\n   inside its closed finite interval. Zero-width finite intervals remain valid.\n5. Do not clamp, replace, or invent finite limits, and do not bypass the native path,\n   replace the library with a standalone implementation, or modify tests to hide the\n   defect.\n\nAfter repairing the source, run the offline integration check:\n\n```sh\ncd /app/vendor/robotics-toolbox-python\npython3 /app/tools/run_validation.py > /app/output.json\n```\n\nThe command must succeed and produce `/app/output.json` with schema version\n`robotics.ik_restart_validation.v1`. Network access and package installation are not\navailable.\n"}
{"task_id": "candidate-0260-security-appsec", "source_id": "candidate-0260-security-appsec", "domain": "terminal", "task_path": "tasks/candidate-0260-security-appsec", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:22f5efb3150202c6e0894f9838c064c2a88e758bd9d4e62d795ad95d9986913a", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the offline security scan pipeline\n\nYou are given a vendored Bandit source tree under `/app/vendor/pycqa-bandit` and\nan offline fixture set under `/app/workload/fixtures`. Repair the existing source\nso the normal file-level Trojan Source check can be used by an integration\ncaller, then produce `/app/output.json` by running the supplied scan entrypoint.\nDo not replace the scanner with a new standalone evaluator.\n\nThe producer must preserve the scanner's security semantics:\n\n- bidirectional Unicode controls are reported as B613 with HIGH severity and\n  MEDIUM confidence;\n- the finding location is 1-based and points to the physical source line and\n  character column containing the control character;\n- file-level plugins may provide their own non-empty `linerange`; generic tester\n  enrichment must not erase it, while issues without a custom range still get\n  the framework context range;\n- filename and original file bytes remain attached to each issue, and the\n  non-UTF-8 fixture must be decoded according to its Python encoding declaration;\n- clean files must remain clean.\n\nUse the existing modules and integration path. The output contract is documented\nin `/app/CONTRACT.md`; its schema version must be exactly\n`bandit_trojan_repair.v1`. Include every `*.py` fixture in lexical filename order,\nincluding clean files, and use the supplied `/app/run_scan.py` after repairing\nthe source. Keep the workspace offline and do not add dependencies.\n"}
{"task_id": "candidate-0308-security-forensics", "source_id": "candidate-0308-security-forensics", "domain": "terminal", "task_path": "tasks/candidate-0308-security-forensics", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:c1b2d7e5807fef4065b79fe92e1dfef047503159349ced9b19941b48f684c4e6", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the Windows forensic plugin integration\n\nThe vendored Volatility3 tree contains a partially completed API migration in\nthe Windows forensic plugins. The shared `Handles` plugin now exposes its\nhandle-table helpers as context-explicit classmethods and has a newer plugin\nversion. Several consumers still use the old contract.\n\nWork in `/app/vendor/volatility3`. Repair the integration in all four existing\nconsumer modules: `callbacks.py`, `dumpfiles.py`, `poolscanner.py`, and\n`psxview.py`. Preserve the upstream filtering and invalid-memory behavior.\n\nRequirements:\n\n1. Every consumer must declare the current `Handles` requirement version.\n2. Calls to `get_type_map`, `find_cookie`, and `handles` must use the\n   context-explicit classmethod interface, passing the active kernel module\n   name and handle table where applicable.\n3. Do not replace the forensic traversal with constants, a new evaluator, or\n   a separate script. Keep the existing plugin dependency graph intact.\n4. Run a compile smoke check that does not leave generated bytecode in the\n   source tree, then write `/app/output.json` with exactly\n   this schema:\n\n```json\n{\"schema_version\":\"volatility-repair-report.v1\",\"modules\":[\"callbacks\",\"dumpfiles\",\"poolscanner\",\"psxview\"],\"checks\":{\"requirements\":4,\"context_calls\":true,\"compile\":true}}\n```\n\nThe report is a summary of the repaired source tree, not a substitute for the\nsource repair. Runtime is offline and no packages or files may be downloaded.\n"}
{"task_id": "candidate-0390-security-appsec", "source_id": "candidate-0390-security-appsec", "domain": "terminal", "task_path": "tasks/candidate-0390-security-appsec", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:2e60ef07848d3fdb80ec45dd1afa9651a2d66629e2acbbba401104287d43f507", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the static-analysis pipeline\n\nThe workspace contains a frozen, offline Bandit source tree under `/app/vendor/bandit` and two representative inputs under `/app/inputs`. A regression affects the line-oriented bidirectional-control-character check when its finding passes through the shared tester.\n\nRepair the existing source modules so that:\n\n- Running `python3 -m bandit -q -f json -o /app/output.json /app/inputs/suspicious.py /app/inputs/latin1.py` completes and writes valid JSON.\n- The suspicious input produces exactly one B613 finding, with high severity, medium confidence, CWE 838, and the finding's manually identified line and column preserved in the serialized result.\n- The Latin-1 input completes without an analyzer traceback and produces no B613 finding.\n- The shared tester still supplies context-derived location data for findings that do not provide their own line range, while never replacing a non-empty plugin-provided range.\n- Existing source, formatter, manager, and functional-test modules remain usable; do not replace the analyzer with a new standalone evaluator or disable checks.\n\nYou may inspect and modify the existing Bandit modules. Keep the solution offline and CPU-only. Your final deliverable is `/app/output.json`; do not include explanations in that file.\n"}
{"task_id": "candidate-0461-ml-evaluation", "source_id": "candidate-0461-ml-evaluation", "domain": "terminal", "task_path": "tasks/candidate-0461-ml-evaluation", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:80a528e6d2c5a0c311dd24246c9db89f2f5eaa175390aa00dd0a5252e60d518d", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the fairness feature path\n\nYou are working in an offline checkout of a Fairlearn development package. A downstream evaluation pipeline uses `MetricFrame` to compute selection rates and accuracy by sensitive and control groups.\n\nThe current feature-processing path has an unsafe edge case: a missing sensitive or control value can be treated as an ordinary subgroup instead of being rejected. This can silently change fairness conclusions. Repair the existing package rather than replacing it with a new evaluator.\n\n## Required workflow\n\n1. Inspect the `MetricFrame` feature-processing path and the `GroupFeature` representation. Reproduce the failure using `environment/cases/metricframe_cases.json` and `environment/run_case.py`.\n2. Modify the existing production modules so missing values in every supported feature container (Python scalar lists, NumPy arrays, pandas Series/DataFrames, and control features) raise a clear `ValueError` before grouping.\n3. Preserve valid behavior: length checks, feature names, multi-column grouping, aggregate metrics, and sample weights must remain unchanged.\n4. Run the focused upstream feature-processing test and the supplied integration runner. Write `/app/repair_report.json` with the exact schema below.\n\nThe report must contain `output_schema_version: \"fairlearn-repair-report.v1\"`, `valid_metrics` for the valid fixture, `rejected_cases` naming all invalid cases, and `focused_tests_passed` as a boolean. Do not hard-code a report without repairing and exercising the package.\n\nThe workspace must remain offline and CPU-only. Do not install packages, download data, modify tests to bypass the repair, or delete the existing source tree.\n\n"}
{"task_id": "candidate-0534-ml-training", "source_id": "candidate-0534-ml-training", "domain": "terminal", "task_path": "tasks/candidate-0534-ml-training", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:9b39c1530ad2f0d047b802b1ad0de1f483131d448efb2a8b2f224c6832b086b9", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair incremental training after model persistence\n\nThe frozen scikit-learn source slice under `/app/vendor/scikit-learn` has a regression in a stateful CPU training workflow. The public harness trains an `MLPRegressor`, serializes and reloads it, then changes the target and performs repeated `partial_fit` calls with both Adam and momentum SGD.\n\nRun:\n\n```bash\npython3 /app/run_regression.py\n```\n\nThe current implementation advances optimizer bookkeeping but leaves the reloaded estimator's live weights and predictions unchanged. Diagnose the parameter ownership across the MLP fit loop and stochastic optimizers, then repair these existing production modules:\n\n- `/app/vendor/scikit-learn/sklearn/neural_network/_multilayer_perceptron.py`\n- `/app/vendor/scikit-learn/sklearn/neural_network/_stochastic_optimizers.py`\n\nAcceptance requirements:\n\n1. The optimizer update contract must apply gradients to the estimator's current coefficient and intercept arrays, including after serialization.\n2. Existing Adam moments/step count and SGD momentum state must continue across reload and fine-tuning; do not recreate the optimizer on each incremental call.\n3. Both public Adam and SGD cases must move materially closer to the changed target while preserving the fitted layer shapes and finite numeric state.\n4. `python3 /app/run_regression.py` must complete and write `/app/results/finetune_report.json` with `output_schema_version` equal to `tbench.mlp_continuation.v1`.\n\nDo not replace or edit `run_regression.py`, `runtime_loader.py`, `cases.json`, immutable vendored modules, or the license. Do not hard-code predictions, bypass serialization, change the workload, disable a solver, or install/download anything. The repair must be in both listed source modules and must remain compatible with both stochastic optimizers.\n"}
{"task_id": "candidate-0628-media-music", "source_id": "candidate-0628-media-music", "domain": "terminal", "task_path": "tasks/candidate-0628-media-music", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:c41c260588f06ccd0e4b0f8c8ba6b29ef497cc0c180ddf891a4e5d214f7c68aa", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Task\n\nRepair the frozen Partitura source tree in `/app/vendor/partitura` and produce `/app/output.mid` by running `/app/run_boundary_workflow.py`.\n\nThe current tree fails at several connected boundaries in its real MusicXML-to-MIDI workflow. Implement the repair in the existing production modules `partitura/io/importmusicxml.py` and `partitura/io/exportmidi.py`.\n\n## Required behavior\n\n1. A MusicXML note marked as a chord member normally inherits the preceding note's onset and duration. If such a marker appears before any anchor note in its measure, loading must emit a warning and keep that note as a standalone event instead of raising an assertion or silently dropping it.\n2. Valid chord members must still share the anchor note's onset and duration; the malformed-input fallback must not disable normal chord semantics.\n3. MIDI export must accept `Score`, `Part`, `PartGroup`, ordinary iterables, and one-shot iterators of parts without consuming an iterator during preliminary inspection.\n4. Exporting a score with no parts must raise a clear domain-level `ValueError` indicating that the score has no parts, before NumPy concatenation or MIDI serialization fails.\n5. Compute PPQ from the score's actual quarter-duration values. Preserve an exact computed PPQ when it is within the Standard MIDI File range; do not substitute a conventional constant.\n6. Standard MIDI File ticks-per-beat is limited to 32767. If the computed PPQ is larger, emit a `RuntimeWarning` explaining that timing is rounded, cap PPQ at 32767, and still serialize a readable MIDI file with integer event ticks.\n7. Run `python3 /app/run_boundary_workflow.py` after the repair. The resulting `/app/output.mid` must be a readable MIDI file generated through the repaired public import/export APIs.\n\nKeep the solution offline and modify the existing implementation rather than replacing the package, altering fixtures, or adding a parallel exporter.\n"}
{"task_id": "candidate-0674-ml-evaluation", "source_id": "candidate-0674-ml-evaluation", "domain": "terminal", "task_path": "tasks/candidate-0674-ml-evaluation", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:5ca4cd941a5a7509de9686894e4c8cb5f3c9519e5de687b4b6592903951329e7", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Task\n\nRepair the frozen TorchMetrics retrieval source tree under `/app/vendor/torchmetrics` so all four public evaluation paths enforce one consistent `top_k` contract:\n\n- `retrieval_average_precision`\n- `retrieval_reciprocal_rank`\n- `RetrievalMAP`\n- `RetrievalMRR`\n\n## Required behavior\n\n1. `top_k=None` evaluates the complete ranking for each query.\n2. A positive integer evaluates only the highest-scored `k` documents **within each query**.\n3. Zero, negative integers, non-integral numbers, and other non-integer values must raise `ValueError` at the public API boundary.\n4. Ranking remains descending by prediction score, preserving each score's relevance label.\n5. Stateful metrics must group by query index before truncation and preserve `empty_target_action` plus `mean`, `median`, `min`, and `max` aggregation behavior.\n6. Functional and stateful APIs must agree on valid inputs.\n\nModify the existing production modules rather than adding a replacement evaluator. The intended repair spans the two functional retrieval modules and the two stateful retrieval modules. Do not delete or rewrite the supplied source tree, public cases, or workflow helper.\n\nAfter repairing the modules, run:\n\n```bash\npython3 /app/tools/run_retrieval_workflow.py\n```\n\nThe command must finish successfully and write `/app/evaluation_report.json` with schema version `retrieval_topk_eval.v1`. Runtime networking and package installation are not allowed.\n"}
{"task_id": "candidate-0688-hardware-rtl", "source_id": "candidate-0688-hardware-rtl", "domain": "terminal", "task_path": "tasks/candidate-0688-hardware-rtl", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:396bc53d212898043d303c8b275d4b16d51a73f37024f7377d9b240fb435e981", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\nRepair the vendored Ibex RTL so an in-flight multi-cycle RV32Zcmp compressed expansion is discarded when the IF stage redirects to an exception handler. Preserve normal Zcmp expansion and the parameterized decoder interface.\n\nWork in `/app/vendor/project`. The source tree is intentionally pre-fix. Add the necessary decoder flush input, connect it from the IF-stage exception redirect, and give the flush priority over the state-machine transition. Then run:\n\n    python3 /app/vendor/project/rtl_regression.py --output /app/output.json\n\nThe command must exit successfully and produce `/app/output.json` matching the public schema in `/app/vendor/project/output_contract.json`. Do not add a replacement evaluator or bypass the existing RTL modules.\n"}
{"task_id": "candidate-0758-ml-inference", "source_id": "candidate-0758-ml-inference", "domain": "terminal", "task_path": "tasks/candidate-0758-ml-inference", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:93f1160b8e5c20c4d0071744d49d53f64386c57b402db3404fc26590a58df47e", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\nThe workspace under `/app/vendor/onnx` is a frozen slice of a tensor operator used for bounded\nattention KV-cache updates. Diagnose and repair the circular-mode indexing bug in the existing\nproduction files:\n\n- `vendor/onnx/onnx/reference/ops/op_tensor_scatter.py`\n- `vendor/onnx/onnx/defs/tensor/defs.cc`\n\nKeep the operator's public contract intact. For each batch/prefix coordinate, `write_indices[b]`\nselects the sequence start. In `circular` mode, wrap only the selected sequence coordinate modulo\nthe cache length. Never modulo batch/head prefix coordinates. Preserve untouched cache values,\nupdate ordering, axis normalization, and the existing shape and mode errors. The same sequence-only\nrule must be visible in the C++ operator pseudocode.\n\nRun the real source through the supplied workflow:\n\n```sh\ncd /app\npython3 run_kv_inference.py --input fixtures/kv_cases.json --output /app/output.json\n```\n\nThe output must be deterministic JSON with `output_schema_version` equal to\n`tensor_scatter_inference.v1` and one result for each input case. Do not modify the fixture, runner,\nor supporting source files to bypass the repair.\n"}
{"task_id": "candidate-0803-ml-training", "source_id": "candidate-0803-ml-training", "domain": "terminal", "task_path": "tasks/candidate-0803-ml-training", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:07630d0920ee4055051a97efb49b681e1dc83b0fe24c10ce2fd2d312ba620741", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair effective-batch normalization\n\nA deterministic language-model training reproduction in `/app/vendor/transformers/tools/effective_batch_case.py` exposes a regression in the existing Trainer label-smoothing path. With gradient accumulation, microbatches can contain different numbers of non-ignored target tokens. Repair the existing framework modules; do not rewrite or bypass the reproduction harness or its local tensor runtime.\n\nRequirements:\n\n1. Trace the call from `Trainer.compute_loss` into `LabelSmoother` and make the effective active-item count available to the smoothing calculation.\n2. Preserve the standalone fallback: when no effective count is supplied, normalization must still use the active non-`-100` labels in that call.\n3. Preserve causal-LM shifting: logits and labels must remain aligned after shifting, while the supplied count is forwarded.\n4. Keep ignored labels out of both NLL and smoothing mass. Do not alter the reproduction inputs or its expected semantics.\n5. Run the existing CPU/offline harness and write its JSON result to `/app/output.json`. The JSON must retain `output_schema_version` equal to `effective_batch.v1` and contain the computed loss, finite-difference gradient, and parameter-update fields produced by the harness.\n\nThe repair is expected to involve the existing `trainer.py` and `trainer_pt_utils.py` modules. The task intentionally uses a small vendored tensor runtime because the execution image has no external ML framework installed. Do not add dependencies, use the network, modify the harness/runtime/tests, or solve by hard-coding the report.\n"}
{"task_id": "candidate-0847-software-languages", "source_id": "candidate-0847-software-languages", "domain": "terminal", "task_path": "tasks/candidate-0847-software-languages", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:f09a84c006a3e88d6e559f2b6cd46a902bbe1fc5f60ae50b515543ec3c6f2937", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair independent LALR interactive-parser clones\n\nThe frozen Lark tree in `/app/vendor/lark` has a clone-ownership regression. Copying an `InteractiveParser` after partially consuming a source stream creates multiple lexer objects whose references do not agree. A clone may therefore consume the original parser's remaining input, or a superficially independent implementation may leave `parser_state.lexer` disconnected from `lexer_thread`.\n\nRepair the existing implementation in both files:\n\n- `/app/vendor/lark/lark/parsers/lalr_parser_state.py`\n- `/app/vendor/lark/lark/parsers/lalr_interactive_parser.py`\n\n## Required behavior\n\n1. `ParserState.copy(deepcopy_values=...)` must create an independent lexer thread and independent parser stack while preserving its current shallow-versus-deep value-stack policy.\n2. `InteractiveParser.copy(deepcopy_values=...)` must use one coherent cloned lexer thread: the clone's `parser_state.lexer` must be the clone's `lexer_thread`, and neither may be the original parser's lexer thread.\n3. After partially consuming input, clone and original must both be resumable to the same complete parse tree in either execution order. Advancing either continuation must not change the other's cursor.\n4. Existing manual `feed_token`, `exhaust_lexer`, EOF, and ordinary LALR parsing behavior must remain compatible.\n\nDo not replace the parser with a task-specific implementation, hard-code the sample streams, disable tests, or edit the supplied upstream test asset.\n\n## Validation and artifact\n\nFrom `/app/vendor/lark`, run:\n\n```bash\npython3 -m compileall -q lark\npython3 tests/test_parser.py TestLalrBasic.test_parser_interactive_parser -q\n```\n\nThen generate the required artifact:\n\n```bash\ncd /app\npython3 tools/run_clone_matrix.py --output /app/repair_report.json\n```\n\nThe artifact must use `output_schema_version` **`lark-clone-repair-v1`** and the driver must report `summary.all_cases_pass: true` after the source repair.\n\n"}
{"task_id": "candidate-0867-software-frontend", "source_id": "candidate-0867-software-frontend", "domain": "terminal", "task_path": "tasks/candidate-0867-software-frontend", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:71f6e6d61f5ec8b0134e51456e24d5b32e70f6c7ae567c738c54722b24408cfc", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Restore reliable file selection\n\nYou are repairing the existing React dropzone source under\n`/app/vendor/project`; do not replace it with a new script or bypass its\nruntime path.\n\nWhen `useFsAccessApi` is enabled, opening the dropzone first uses the browser\nFile System Access picker. Some browsers reject that picker with a\n`DOMException` named `NotAllowedError` even though the API was detected and the\nuser did not cancel. Treat that outcome as an unavailable capability:\n\n1. Extend the shared exception helpers in `src/utils/index.ts` with a precise\n   predicate for this failure.\n2. Update the existing picker branch in `src/index.tsx` so this failure is\n   handled like the existing security fallback: disable the FS Access path for\n   future opens, reset the native input, and click it only after the rejected\n   promise. Keep `AbortError` as cancellation and preserve the existing error\n   path for unrelated exceptions.\n3. Do not fabricate files. The native input change event must continue through\n   the existing extraction, accept/type, size, and `maxFiles` validation and\n   callback pipeline.\n\nCheck both files and the surrounding tests before editing. Run the focused\noffline regression contract after the repair:\n\n```sh\ncd /app/vendor/project\npython3 integration/run_regression.py --output /app/output.json\n```\n\nThe command must exit successfully and create `/app/output.json` with the\ncontract's schema and all checks true. Do not install packages or use network\naccess.\n"}
{"task_id": "candidate-0938-science-chemistry", "source_id": "candidate-0938-science-chemistry", "domain": "terminal", "task_path": "tasks/candidate-0938-science-chemistry", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:1d41382d0752848d54ff88e2cbbf0f7336d041c6001ddee2ae69face1a1d800c", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the orbital-analysis integration\n\nYou are working in `/app/vendor/project`, a small frozen slice of a computational-chemistry library. The included `data/casebook.json` contains six `ccData` scenarios covering restricted, open-shell, unrestricted, and incomplete molecular-orbital data.\n\nThe integration command is:\n\n```sh\npython3 /app/tools/orbital_check.py --input /app/data/casebook.json --output /app/output.json\n```\n\nRepair the existing library modules under `vendor/project/src/cclib` so this command completes and writes the declared JSON artifact. Preserve the public APIs and use the real `ccData` and `Orbitals` classes.\n\nRequirements:\n- Orbital analysis must reject incomplete data before analysis, identifying each missing required attribute. `mocoeffs`, `moenergies`, and `homos` are all required for `Orbitals`.\n- The closed-shell convenience property must instantiate the public `Orbitals` class and return its computed boolean result.\n- Preserve the scientific behavior: a single MO set with differing HOMO indices is open-shell; two MO sets are closed-shell only when their energy arrays agree within numerical tolerance.\n- Exercise all six cases, inspect the generated JSON, and ensure the schema version is `orbital-report.v1` with one record per input case.\n\nDo not edit the casebook or integration script to hard-code expected answers. Do not add network dependencies or bypass the library source.\n"}
{"task_id": "candidate-0993-software-systems", "source_id": "candidate-0993-software-systems", "domain": "terminal", "task_path": "tasks/candidate-0993-software-systems", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:7dbdb84be548a95ca247523d3e4cbe7d07294491ed3d33f6c3d30030cae76dc1", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair wheel tag validation\n\nYou are working in a frozen checkout of the Python `packaging` library under\n`/app/vendor/packaging`. A compatibility regression is present at the boundary\nbetween compressed wheel-tag parsing and wheel filename parsing.\n\nRepair the existing production code, rather than adding a parallel evaluator.\nThe public APIs involved are `packaging.tags.parse_tag` and\n`packaging.utils.parse_wheel_filename`.\n\nRequirements:\n\n1. A compressed tag has interpreter, ABI, and platform components. Every\n   member of every compressed component must be validated before expansion.\n2. Interpreter members must be valid Python identifiers. Reject a numeric\n   dotted value such as `2.7.6` as one malformed interpreter member; do not\n   expand it into `2`, `7`, and `6`.\n3. Preserve valid compressed tags such as `py3.py2-none-any`, and preserve\n   custom identifier forms such as `graalpy311` and `_custom`.\n4. Direct parsing must expose the existing `InvalidTag` API. Wheel filename\n   parsing must translate malformed tag validation into its existing\n   `InvalidWheelFilename` API without changing valid filename normalization,\n   versions, build tags, or tag sets.\n5. Keep the existing source tree and tests intact. Update the two production\n   modules that own this behavior: `src/packaging/tags.py` and\n   `src/packaging/utils.py`.\n6. Run the focused regression tests in the repository, compile the package,\n   and create `/app/output.json` using the repaired public APIs.\n\nThe output must be JSON with this exact top-level contract:\n\n```json\n{\n  \"output_schema_version\": \"packaging-tag-audit.v1\",\n  \"cases\": [\n    {\n      \"name\": \"valid-compressed\",\n      \"direct\": {\"status\": \"ok\", \"tag_count\": 2, \"interpreters\": [\"py2\", \"py3\"]},\n      \"wheel\": {\"status\": \"ok\", \"name\": \"demo\", \"version\": \"1.0\", \"build\": [], \"tag_count\": 2, \"interpreters\": [\"py2\", \"py3\"]}\n    },\n    {\n      \"name\": \"custom-identifiers\",\n      \"direct\": {\"status\": \"ok\", \"tag_count\": 2, \"interpreters\": [\"_custom\", \"graalpy311\"]},\n      \"wheel\": {\"status\": \"ok\", \"name\": \"demo\", \"version\": \"1.0\", \"build\": [], \"tag_count\": 2, \"interpreters\": [\"_custom\", \"graalpy311\"]}\n    },\n    {\n      \"name\": \"malformed-interpreter\",\n      \"direct\": {\"status\": \"invalid-tag\", \"tag_count\": 0, \"interpreters\": []},\n      \"wheel\": {\"status\": \"invalid-wheel\", \"tag_count\": 0, \"interpreters\": []}\n    }\n  ]\n}\n```\n\nFor the malformed wheel case, use a filename whose tag is\n`2.7.6-none-any`; do not hard-code an answer without exercising both public\nentry points. JSON arrays must be deterministically sorted as shown.\n"}
{"task_id": "candidate-1001-security-forensics", "source_id": "candidate-1001-security-forensics", "domain": "terminal", "task_path": "tasks/candidate-1001-security-forensics", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:549cd81822ede4b3cfdbd34e5e06f9e34cbc36aed6ee245e2370c8a9064639d9", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the Windows Registry forensic workflow\n\nThe vendored project under `/app/vendor/plaso` is a pre-fix slice of a Windows Registry forensic parser. A recent input can contain one registry value whose typed decoder raises `dfwinreg.errors.WinRegistryValueError`. Repair the real plugin code, not the runner, so extraction remains useful and the caller integration remains consistent.\n\nRequirements:\n\n1. Update the shared Windows Registry value-formatting helper to receive the parser mediator and handle only the decoder failure represented by `WinRegistryValueError`. The affected value must remain in the formatted event as `[TYPE] (N bytes)`, where `N` is the raw payload length.\n2. Emit exactly one extraction warning for that recovery. The warning must include the value type, value name, registry key path, and decoder error text.\n3. Preserve existing formatting for strings, integers, multi-strings, binary values, and empty values, including their type labels and byte/empty representations.\n4. Update the Services plugin caller to pass the mediator into the shared helper while keeping its skip list. The service event must still be produced with its typed `Start` and `Type` fields and remaining values.\n5. Keep the existing source modules and runner intact. Run the offline integration command:\n\n```sh\ncd /app/vendor/plaso\nPYTHONPATH=/app/vendor/plaso python3 tools/run_case.py /app/output.json\n```\n\nThe output file must be JSON with `output_schema_version` equivalent to `schema_version` `registry-forensics.v1`, two events, one warning, a recovered byte-count value, and the service event fields. Do not use network access or add dependencies.\n"}
{"task_id": "candidate-1005-software-systems", "source_id": "candidate-1005-software-systems", "domain": "terminal", "task_path": "tasks/candidate-1005-software-systems", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:56b2cf332e1f88c7d2e85b3efa7bcc132fec61717fe8876d693e4d72f11069b0", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\nA frozen Requests source tree is available at `/app/vendor/requests`.\n\nSome file-like wrappers expose `read` only through `__getattr__`. Runtime protocol checks do not reliably recognize that delegated capability, so request preparation uses inconsistent rules in different branches.\n\nRepair the existing implementation. Your submission must satisfy all of the following:\n\n1. Add one reusable internal predicate in `/app/vendor/requests/src/requests/_types.py` that recognizes an object as readable only when resolving its `read` attribute yields a callable.\n2. Use that predicate consistently in `/app/vendor/requests/src/requests/models.py` for parameter encoding, multipart file-content reading, and the decision not to synthesize `application/x-www-form-urlencoded`.\n3. A wrapper that hides `__iter__` but delegates `read` must remain a raw request body, and request preparation must not add a form content type.\n4. The same delegated wrapper must be read correctly when supplied as a multipart file, for both byte and text streams.\n5. Preserve existing behavior for strings, bytes, `None`, mappings, ordered form pairs, and objects whose `read` attribute exists but is not callable. In particular, mappings must still be form-encoded with their normal content type.\n6. Modify both existing modules named above. Do not replace the Requests package, delete neighboring source/tests, disable checks, or add network/package-install steps.\n7. Compile the two repaired modules and run:\n\n```bash\nPYTHONPATH=/app/vendor/requests/src python3 /app/tools/run_regression.py --output /app/repair_report.json\n```\n\nThe required artifact is `/app/repair_report.json` with `output_schema_version` equal to `requests.stream_proxy.repair.v1` and `passed` equal to `true`. The verifier also imports and exercises the repaired source independently; a forged report is insufficient.\n"}
{"task_id": "candidate-1048-operations-virtualization", "source_id": "candidate-1048-operations-virtualization", "domain": "terminal", "task_path": "tasks/candidate-1048-operations-virtualization", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:529269bf9f03ab9cf9c1ceeb75fb243f8ba906c6854ea19fdae42ec0ddc47a59", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\nThe vendored Vagrant VirtualBox provider has a regression in private-network\nconfiguration. Repair the existing provider source, preserving its public\nAPIs and version-specific VBoxManage parsing.\n\nA private network's `name` may be either the stable host-only interface name or\nthe display name reported by VBoxManage. Make the common network action resolve\nthe display form to the canonical interface name before adapter setup. Ensure\nthe legacy driver parsers and the VirtualBox 7 host-only-network adapter expose\nthe same normalized `display_name` field while retaining their native address,\nmask, IPv6, and status behavior. The generic action must use the portable\nhost-only interface inventory; the host-only-network inventory is only a\nDarwin/VirtualBox-7 implementation detail. Keep the shared driver facade's\ndelegation surface synchronized with the versioned inventory APIs.\n\nCanonical interface names must continue to resolve unchanged. An unknown name\nmust remain unknown rather than selecting the first available interface.\n\nWork directly in the existing files under `vendor/vagrant`. Do not add a\nreplacement evaluator or hard-code the fixture's interface name. The vendored\nupstream tests are available for code-reading context, but do not install gems\nor attempt the full upstream suite. Finish by running:\n\n```sh\npython3 /app/contract_runner.py > /app/contract_result.json\n```\n\nThe command must succeed offline and leave the declared artifact at that exact\npath. The JSON artifact must declare\n`output_schema_version = \"vagrant.virtualbox.network-contract.v1\"` and report the\ncross-version inventories, the three name-resolution outcomes, portable API\ncall counts, Meta delegation status, and the sorted result-key list produced by\nthe supplied runner.\n"}
{"task_id": "candidate-1155-software-systems", "source_id": "candidate-1155-software-systems", "domain": "terminal", "task_path": "tasks/candidate-1155-software-systems", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:b69825011aa7bc37ee9e9564fa42a581f5f4621aff19455b6a3b7286ad12ced9", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair configurable routing behavior\n\nThe workspace contains a frozen source tree for a Python URL-routing library and `routing_workflow.py`. Diagnose and repair the routing implementation in the vendored source tree, then run:\n\n```sh\npython3 /app/routing_workflow.py\n```\n\nThe runner must create `/app/output.json` with `output_schema_version` set to `werkzeug-routing-repair-v1`.\n\nThe public `Map.merge_slashes` setting controls whether repeated separators in a request are canonicalized to a redirect. It must work both when supplied at construction and when changed on an existing map. A rule-level opt-out must continue to reject the normalized spelling while accepting its literal spelling. Preserve strict-trailing-slash redirects, integer conversion, path converters (including repeated separators inside a variable), host binding, and URL building.\n\nWork only in the vendored library source. Do not replace the runner or hard-code its JSON, remove tests or source modules, install packages, access the network, or rely on external services.\n"}
{"task_id": "candidate-1165-ml-evaluation", "source_id": "candidate-1165-ml-evaluation", "domain": "terminal", "task_path": "tasks/candidate-1165-ml-evaluation", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:5bdaefdf97293fca3f552b581ee9ce96c42c0d1d3a3c54a88748fe285c43a08e", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the evaluation metric boundary\n\nThe workspace contains a frozen slice of a Hugging Face Evaluate repository\nunder `/app/vendor/project`. Its precision, recall, and F1 metric scripts call\nscikit-learn and then normalize the returned score for downstream JSON\nserialization. With current scikit-learn, binary averages can return a native\nPython scalar while per-label (`average=None`) calls return an array. The\nexisting normalization assumes every result has the same array API.\n\nRepair the existing metric modules so the integration runner works for every\ncase in `/app/cases.json` and for unseen case manifests supplied by the\nverifier. Keep sklearn's metric values and options intact: binary and other\naggregate averages must serialize as a Python `float`, while `average=None`\nmust remain an ordered per-label array. Apply the compatibility boundary\nconsistently to `metrics/f1/f1.py`, `metrics/precision/precision.py`, and\n`metrics/recall/recall.py`.\n\nRun the integration command:\n\n```sh\npython3 /app/run_metrics.py --input /app/cases.json --output /app/output.json\n```\n\nThe output must be exactly one JSON object with `schema_version`\n`ml-eval-output.v1` and a `results` array in input order. Each row contains its\ncase `id` and `f1`, `precision`, and `recall` values (a number or an array).\nDo not add dependencies, access the network, or replace the vendored metric\nimplementations with a new evaluator.\n"}
{"task_id": "candidate-1216-software-frontend", "source_id": "candidate-1216-software-frontend", "domain": "terminal", "task_path": "tasks/candidate-1216-software-frontend", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:d2c208c7f2b8c183b142ee8e900d10baf2fa7c778a785c227eeaf818fde5a176", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair bundled development runtime ordering\n\nWork in `/app/vendor/vite`. The frozen source implements an experimental bundled\ndevelopment mode, but its client runtime is currently attached at the wrong layer.\n\nRepair the existing production source so all of these behaviors hold:\n\n1. Define one exported canonical filename for the bundled development client in the\n   shared node constants module, and use it consistently for both the built client entry\n   path and the in-memory server filename.\n2. During `serve` for a bundled **client** environment, generated HTML must contain a\n   `<script type=\"module\">` for that client. Its `src` must honor the configured `base`.\n3. The client script must be the first element injected into `<head>`, before generated\n   application chunk scripts. Existing application script order must otherwise remain\n   unchanged.\n4. The in-memory bundled server must continue serving the client runtime file, but entry\n   chunk source must be returned unchanged: it must not prepend another client import.\n5. Normal production builds, server/SSR consumers, and unbundled development behavior\n   must remain unchanged.\n\nModify the existing production modules; do not solve this by replacing the repository,\nadding a parallel evaluator, changing upstream tests, or hard-coding only the root-base\ncase. At minimum, the repair is expected to involve:\n\n- `/app/vendor/vite/packages/vite/src/node/constants.ts`\n- `/app/vendor/vite/packages/vite/src/node/plugins/html.ts`\n- `/app/vendor/vite/packages/vite/src/node/server/bundledDev.ts`\n\nWhen finished, run:\n\n```bash\ncd /app\npython3 tools/run_regression.py\n```\n\nThe command must write `/app/regression-report.json` using the schema documented in\n`/app/REPAIR_CONTRACT.md`. The private verifier independently recomputes the behaviors\nfrom the repaired production source; the report alone is not sufficient.\n"}
{"task_id": "candidate-1247-security-reverse-engineering", "source_id": "candidate-1247-security-reverse-engineering", "domain": "terminal", "task_path": "tasks/candidate-1247-security-reverse-engineering", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:32c0341d894be70881dd01a0602e027894827d08481285a87b11662795cc14ea", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\nThe frozen project at `/app/vendor/dissect-cstruct` parses C-like structure definitions and uses them to decode binary data. Its current lexer and parser report only a line number for malformed definitions, which is insufficient when diagnosing a long forensic schema.\n\nRepair the implementation in these existing production modules:\n\n- `/app/vendor/dissect-cstruct/dissect/cstruct/lexer.py`\n- `/app/vendor/dissect-cstruct/dissect/cstruct/parser.py`\n\nDo not replace the parser, disable malformed-input failures, or edit the bundled tests and diagnostic runner to hide a regression.\n\n## Required behavior\n\n1. `LexerError` and `ParserError` messages must retain their current first line: `line N: <message>`.\n2. When source text is available, append a numbered excerpt containing the failing logical line plus up to two existing lines before and after it.\n3. Format every excerpt row exactly as ` N: <source>` for a surrounding line and `>N: <source>` for the failing line. Line numbers are one-based and are not padded. Preserve blank source lines.\n4. Lexer excerpts must come from the actual input being tokenized and must be calculated dynamically for any failing line, not from fixed examples.\n5. Parser excerpts must use the parser's normalized source after backslash-newline continuations have been joined. An error formatted before any parse has started must remain safe and line-only.\n6. Route parser failures through the contextual error path where a source location is available; do not leave a special malformed-array path with an unnumbered bare error.\n7. Valid definitions must continue to compile and decode with unchanged field widths, array values, byte order, and serialization.\n8. Keep the bundled focused lexer/parser tests passing.\n\n## Integration workflow\n\nAfter repairing both modules, run the focused tests from the project root:\n\n```sh\ncd /app/vendor/dissect-cstruct\nPYTHONPATH=. python3 -m pytest -q tests/test_lexer.py tests/test_parser.py\n```\n\nThen generate the required diagnostic artifact from the repaired source tree:\n\n```sh\nPYTHONPATH=/app/vendor/dissect-cstruct \\\n  python3 /app/tools/run_diagnostics.py --output /app/repair_report.json\n```\n\n`/app/repair_report.json` must use schema version `tbench.cstruct_repair_report.v1`. Do not use the network or install packages.\n"}
{"task_id": "candidate-1257-software-systems", "source_id": "candidate-1257-software-systems", "domain": "terminal", "task_path": "tasks/candidate-1257-software-systems", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:5dde74728d7d458d80c166752b6e0afa5e870a4d5a4c5e79976c75c5e8b92ed5", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair compressed wheel-tag parsing\n\nYou are given a frozen Python packaging library under `/app/vendor/packaging`.\nThe public tag parser currently accepts malformed interpreter components in a\ncompressed tag. For example, a dotted numeric component can be expanded into\nseveral bogus interpreter tags. The wheel filename utility then exposes the\nsame defect through its public API.\n\nRepair the existing production source in both `src/packaging/tags.py` and\n`src/packaging/utils.py`.\n\nRequirements:\n\n1. Every interpreter member of a compressed tag must be a Python identifier.\n   Reject invalid members before creating any `Tag` objects. Digits are valid\n   when the complete value is an identifier: preserve values such as `py3`,\n   `graalpy311`, and `_custom`.\n2. Keep existing empty-component, three-component, ordering, and tag-limit\n   checks intact.\n3. `parse_wheel_filename()` must translate tag validation failures to its own\n   `InvalidWheelFilename` exception while retaining normal project-name,\n   version, build, and valid-tag normalization behavior.\n4. Run the offline integration audit at `/app/run_audit.py`. It reads\n   `/app/inputs/tag_cases.json` and writes the required artifact\n   `/app/output.json` using the schema in `/app/contract.json`.\n\nDo not add dependencies or access the network. The deliverable is the modified\nsource tree plus the generated JSON artifact.\n\n"}
{"task_id": "candidate-1271-media-games", "source_id": "candidate-1271-media-games", "domain": "terminal", "task_path": "tasks/candidate-1271-media-games", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:37034a49d0e3b9840cf385f178466f3db6777e2a656c2765c6f711b5c7894f96", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair Arcade's camera and batch rendering integration\n\nA dependency upgrade changed how window cameras and batch drawing interact.\nThe frozen Arcade tree in `/app/vendor/project` still uses the old integration:\ncamera activation writes matrices to obsolete window attributes, the default\nprojector does not implement the batch-camera scope protocol, and both graphics\nbackends bypass Arcade's buffer abstraction when binding the window matrix\nblock.\n\nRepair the existing source tree. Modify these six production modules:\n\n- `arcade/camera/camera_2d.py`\n- `arcade/camera/default.py`\n- `arcade/camera/orthographic.py`\n- `arcade/camera/perspective.py`\n- `arcade/gl/backends/opengl/context.py`\n- `arcade/gl/backends/webgl/context.py`\n\nThe repaired behavior must satisfy all of the following:\n\n1. `Camera2D.use()`, `OrthographicProjector.use()`, and\n   `PerspectiveProjector.use()` make themselves current, preserve their\n   viewport/scissor behavior, and write generated projection and view matrices\n   to `window.ctx.projection_matrix` and `window.ctx.view_matrix`. They must not\n   create or rely on legacy `window.projection` or `window.view` attributes.\n2. `DefaultProjector` exposes live `projection` and `view` properties backed by\n   its Arcade context. It implements batch camera `begin()` and `end()` hooks as\n   no-ops and `get_group_scissor()` returns `None`, so a batch does not inherit\n   an Arcade scissor rectangle as an extra pyglet group.\n3. Camera activation scopes continue to restore the previous camera, matrices,\n   viewport, scissor, and framebuffer state even when drawing raises an\n   exception.\n4. Both `OpenGLArcadeContext.bind_window_block()` and\n   `WebGLArcadeContext.bind_window_block()` bind the Arcade-owned buffer through\n   `buffer.bind_to_uniform_block(binding=0, offset=0, size=128)`. Do not call a\n   raw backend binding API or use the underlying buffer id.\n5. Keep the public API and the rest of the source tree intact. Do not replace\n   the six modules, modify `/app/tools/check_camera_integration.py`, or change\n   the camera data/projection helpers or Arcade buffer abstraction.\n\nAfter repairing the modules, run:\n\n```bash\npython3 /app/tools/check_camera_integration.py\n```\n\nThis must exit successfully and create `/app/repair_report.json`. The output\ncontract is `/app/output_contract.json`; its `output_schema_version` is\n`arcade.camera-repair-report.v1`. The report must show `status: \"pass\"`, no\nerrors, all six compiled modules, context-routed state for all three camera\nfamilies, a complete default-projector protocol result, and identical recorded\nuniform-buffer calls for both backends.\n\nThe final submission is the repaired source tree plus\n`/app/repair_report.json`. Runtime work must remain offline and CPU-only.\n"}
{"task_id": "candidate-1305-software-data-engineering", "source_id": "candidate-1305-software-data-engineering", "domain": "terminal", "task_path": "tasks/candidate-1305-software-data-engineering", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:6af5f2e49a663c7fce492eb4d3e5d1d08fd3f004bb4e7608586504a072d4de9c", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\nYou are repairing the vendored petl ETL library in `/app/vendor/project`.\n\nThe frozen source has a cross-module fluent-API regression: some `Table` view\nclasses store constructor options in instance attributes named `cache`,\n`complement`, `index`, or `skip`. Those names are inherited operations on table\ncontainers, so composed pipelines can fail with `TypeError` when a fluent\noperation is called. Correct the existing source modules; do not replace the\nlibrary with a new evaluator or a one-off script.\n\nRequirements:\n\n1. Preserve the public fluent operations and their function-form equivalents.\n   A view created with `cache`, `complement`, `index`, or `skip` options must\n   still apply those options during lazy iteration and materialisation.\n2. Cover every affected view family in the vendored tree, including hash joins,\n   sorting, row/field selection, regex search, basic field operations,\n   materialisation, and the optional Avro and bcolz readers. Keep optional\n   backends import-safe; do not install or download their dependencies.\n3. Preserve exact headers, row values, duplicate-key join rows, ordering, and\n   the meaning of false/true option values. Do not eagerly consume a source\n   merely to make a method callable.\n4. Exercise a composed pipeline from `fixtures/pipeline.json` and write\n   `/app/repair_report.json` with this exact schema:\n\n```json\n{\n  \"schema_version\": \"petl-repair-report.v1\",\n  \"output_schema_version\": \"petl-repair-report.v1\",\n  \"source_tree\": \"vendor/project\",\n  \"modules_repaired\": [\"petl/...\"],\n  \"joined_rows\": [[...]],\n  \"complement_rows\": [[...]],\n  \"sorted_rows\": [[...]],\n  \"regex_rows\": [[...]],\n  \"index_position\": 0,\n  \"method_checks\": {\"cache\": true, \"complement\": true, \"index\": true, \"skip\": true}\n}\n```\n\n`modules_repaired` must name the existing source modules you actually changed.\nThe rows must come from running petl, not from hardcoded expected output. The\nreport must be valid JSON and written at the absolute path above. You may use\nstandard-library tooling and the vendored package only; leave the fixture\nunchanged.\n"}
{"task_id": "candidate-1399-security-reverse-engineering", "source_id": "candidate-1399-security-reverse-engineering", "domain": "terminal", "task_path": "tasks/candidate-1399-security-reverse-engineering", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:04f742598087c75a634ddf50808e42bb441f18d7a9af3072834626435e40e315", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair cross-backend address semantics\n\nYou are working in a frozen reverse-engineering feature extractor. Repair the existing source tree, not by replacing it with a new evaluator or modifying the regression runner.\n\nThe three backend file extractors receive virtual addresses from their analysis APIs. File-scope ASCII and UTF-16 strings must retain those values as `AbsoluteVirtualAddress`; they must not be relabeled as file offsets merely because downstream code also supports file offsets.\n\nEmbedded PE detection is different: when a backend can prove a byte-to-file mapping, emit `FileOffsetAddress`. Ghidra can report an unmapped address with `-1`; in that case emit the embedded feature at the original `AbsoluteVirtualAddress` instead of dropping it. Preserve the existing Binary Ninja and IDA mapped-offset behavior and keep feature values unchanged.\n\nWork in `vendor/capa/features/extractors/{ghidra,ida,binja}/file.py`. Then run:\n\n```sh\npython3 tools/run_regression.py\n```\n\nThe command executes the existing extractor functions against deterministic backend doubles whose virtual addresses deliberately differ from file offsets. It must produce `/app/output.json` with `output_schema_version` equal to `capa-address-regression.v1`, two string observations per backend, all four mapped/unmapped embedded-PE cases, and a compatibility section. Preserve the detected feature values and exact backend-provided coordinates.\n\nDo not install packages, use the network, edit `tools/run_regression.py`, or add a replacement extractor. The verifier protects the rest of the frozen project and independently executes the repaired functions with its own fixtures.\n"}
{"task_id": "candidate-1442-software-languages", "source_id": "candidate-1442-software-languages", "domain": "terminal", "task_path": "tasks/candidate-1442-software-languages", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:01477582487894e1982905b5a201d8eab0e426f2a426d7941cfadae6c672bc47", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Task: repair wheel interpreter-tag validation\n\nThe frozen library in `/app/vendor/packaging` accepts malformed wheel interpreter-tag\nmembers and expands them into `Tag` objects. Repair the existing parser pipeline.\n\n## Required behavior\n\n1. In `src/packaging/tags.py`, every member of the interpreter field in a direct\n   `parse_tag()` call must be a non-empty valid Python identifier. Invalid members must\n   raise `packaging.tags.InvalidTag` with an informative message that identifies the\n   interpreter problem and the offending member.\n2. In `src/packaging/utils.py`, `parse_wheel_filename()` must translate that failure to\n   `packaging.utils.InvalidWheelFilename`. Its message must preserve useful interpreter\n   validation context and identify the rejected wheel filename. Do not expose a chained\n   internal parser exception for this validation failure.\n3. Preserve valid behavior: compressed interpreter identifiers still expand as a set;\n   ABI and platform members retain their distinct syntax; version/name normalization,\n   build-tag parsing, order validation, empty-component validation, and tag-count limits\n   must continue to work.\n4. Make substantive repairs in both existing production modules above. Do not replace\n   the package, edit its tests to hide failures, disable validation, or special-case one\n   filename.\n\n## Required artifact\n\nAfter repairing and checking the source, write `/app/repair_report.json` with this public\ncontract:\n\n```json\n{\n  \"output_schema_version\": \"tbench.packaging_repair_report.v1\",\n  \"status\": \"repaired\",\n  \"source_files\": [\n    \"vendor/packaging/src/packaging/tags.py\",\n    \"vendor/packaging/src/packaging/utils.py\"\n  ],\n  \"checks\": {\n    \"direct_invalid_rejected\": true,\n    \"wheel_invalid_wrapped\": true,\n    \"valid_compressed_preserved\": true,\n    \"abi_platform_preserved\": true\n  }\n}\n```\n\nYou may include additional fields, but the declared fields and values must be present.\nThe verifier independently imports and exercises the repaired source; a fabricated\nreport does not pass.\n"}
{"task_id": "candidate-1559-ml-evaluation", "source_id": "candidate-1559-ml-evaluation", "domain": "terminal", "task_path": "tasks/candidate-1559-ml-evaluation", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:4b0334efbc88b88f371f109114d0a2eb2f90bbffec2e25a7d5bbf1fd3c4e1a4c", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair adversarial backend feature-matrix validation\n\nThe vendored Fairlearn adversarial backend infers neural-network widths from\nfeature, target, and sensitive-feature matrices. Its generic backend currently\nindexes the feature shape without first enforcing the matrix contract. As a\nresult, malformed feature data can either fail with an incidental exception or\nsilently construct a model with the wrong feature semantics.\n\nRepair the existing source tree under `/app/vendor/fairlearn`.\n\nRequirements:\n\n1. In `fairlearn/adversarial/_constants.py`, define the shared `_X_NOT_2D`\n   error-message template used for this validation. It must format with the\n   received rank and clearly state that `X` must be two-dimensional.\n2. In `fairlearn/adversarial/_backend_engine.py`, import and use that shared\n   template. `BackendEngine` must reject every `X` whose `ndim` is not exactly\n   `2` by raising `ValueError` formatted with the actual rank.\n3. Perform this validation before inferring widths or constructing either\n   model. Do not flatten, reshape, squeeze, or otherwise reinterpret invalid\n   input.\n4. Preserve valid behavior: predictor input width is `X.shape[1]`; target and\n   sensitive-feature widths come from the fitted transformers; warm-start,\n   loss, optimizer, shuffling, evaluation, and training contracts remain\n   unchanged.\n5. Do not replace or delete the vendored package or diagnostic runner. The\n   repair must be implemented in both existing modules named above; changing\n   callers or generating a report without repairing the backend is not a\n   solution.\n\nAfter repairing the source, compile it and run the supplied integration\ndiagnostic:\n\n```bash\ncd /app/vendor/fairlearn\npython3 -m compileall -q fairlearn\ncd /app\npython3 tools/backend_diagnostics.py --output /app/repair_report.json\n```\n\nSubmit the repaired source tree and `/app/repair_report.json`. The report must\nuse schema version `fairlearn.backend-diagnostics.v1`, include rejected rank\ncases `0`, `1`, `3`, and `4`, and include the valid-path model widths, shuffle\nalignment, evaluation result, and training losses produced by the runner.\n\n"}
{"task_id": "candidate-1601-software-data-engineering", "source_id": "candidate-1601-software-data-engineering", "domain": "terminal", "task_path": "tasks/candidate-1601-software-data-engineering", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:2490cc50a00a3f3589109540e80053457897ec9e48d840f6d1a95712c45f17a1", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair uneven-row joins in petl\n\nAn ETL pipeline built on the frozen petl tree in `/app/vendor/petl` receives\niterable tables whose data-row widths do not always match their headers. The\ncurrent relational join implementations can raise `IndexError`, shift a value\nunder the wrong output field, or emit a row shorter or longer than the output\nheader.\n\nRepair the existing implementation in both:\n\n- `/app/vendor/petl/petl/transform/joins.py`\n- `/app/vendor/petl/petl/transform/hashjoins.py`\n\nThe public sort-based and hash-based join families must share these semantics:\n\n1. Before relational processing, each data row behaves as a row of exactly its\n   source header width. Missing positions use the operation's `missing` value\n   when that API has one, otherwise `None`. Values beyond the source header do\n   not leak into the joined schema.\n2. Key selection by field name, index, natural key, distinct `lkey`/`rkey`, or\n   compound key remains correct even when a key component is absent from a\n   short row.\n3. Inner, left, right, outer, lookup, anti, and cross joins preserve documented\n   match cardinality and duplicate-key behavior. The equivalent hash joins\n   must agree semantically while retaining their normal input-order behavior.\n4. Prefix handling, custom missing sentinels, lazy iteration, cached and\n   uncached execution, and repeated materialization continue to work. Do not\n   mutate caller-owned input rows.\n\nDo not modify the visible workflow, existing tests, or other library modules.\nThe verifier protects those assets and exercises additional private tables.\n\nAfter repairing the source, run:\n\n```sh\ncd /app\npython3 -m py_compile vendor/petl/petl/transform/joins.py vendor/petl/petl/transform/hashjoins.py\npython3 workflow.py /app/output.json\n```\n\n`/app/output.json` must use schema `petl-uneven-joins.v1` with the keys\n`case_order`, `cases`, `row_counts`, and `all_rectangular`; it must be produced\nby the supplied workflow from the repaired source tree.\n\n"}
{"task_id": "candidate-1634-software-databases", "source_id": "candidate-1634-software-databases", "domain": "terminal", "task_path": "tasks/candidate-1634-software-databases", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:0da98e97526d6de162e492f79afbd442ee35de47de62fb08734396b746e19858", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\nThe vendored ORM in `/app/vendor/peewee` has a transaction argument propagation regression. Repair the existing source modules, preserving their public API and backend behavior.\n\nRequirements:\n\n1. `SqliteDatabase.atomic(lock_type=...)` and `SqliteDatabase.transaction(lock_type=...)` must accept `DEFERRED`, `IMMEDIATE`, and `EXCLUSIVE` (case preserved or normalized by SQLite), and the selected mode must reach the actual `BEGIN` statement.\n2. Existing positional transaction-type calls must continue to work. Nested `atomic()` scopes must use savepoints and must not issue a second top-level `BEGIN`.\n3. The generic transaction implementation and the APSW adapter in `playhouse/apsw_ext.py` must retain compatible lock-mode forwarding; do not replace either module with a new standalone script.\n4. Run `/app/integration_check.py` and leave its JSON result at `/app/output.json` with `schema_version` `tbench.peewee-lock.v1`, a `checks` object, and all checks true.\n\nWork offline. Do not add dependencies or modify the integration checker.\n"}
{"task_id": "candidate-1669-operations-marketing", "source_id": "candidate-1669-operations-marketing", "domain": "terminal", "task_path": "tasks/candidate-1669-operations-marketing", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:4682f20691e854905196beb7597e364509d46dde42be74a71bf1ab2af04f0e4a", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair campaign-outcome validation in uplift evaluation\n\nWork in `/app/vendor/scikit-uplift`, a frozen source tree for a real marketing uplift library. Its treatment-aware metrics correctly require two binary campaign arms, but malformed non-binary response outcomes currently flow into incrementality calculations and plots.\n\nDiagnose and repair the existing implementation. The repair must be made in both `sklift/metrics/metrics.py` and `sklift/viz/base.py`; do not replace the library with a separate evaluator, coerce response values, or special-case the supplied fixture.\n\nThe public behavior is:\n\n- Every metric entry point that accepts `y_true` must reject response arrays containing values outside binary `0` and `1` with `ValueError` before computing cohort statistics.\n- Every plotting entry point that accepts `y_true` must perform the same validation locally, before delegating to metric helpers or constructing misleading diagnostics.\n- Existing consistent-length validation and binary treatment validation must remain in force.\n- Valid binary campaigns must retain the existing model, uplift/Qini curve, AUC, percentile, and plotting behavior.\n- Keep the frozen source and existing tests intact except for the two designated production modules.\n\nAfter repairing the source, run the supplied end-to-end workflow:\n\n```sh\npython3 /app/tools/run_campaign_evaluation.py --output /app/campaign_evaluation.json\n```\n\nThe command must complete offline. `/app/campaign_evaluation.json` must use `output_schema_version` `sklift.campaign-evaluation.v1`, record the three workflow stages, contain the valid campaign's model predictions, curves, scores, percentile table and plot titles, record each malformed campaign case as rejected with `ValueError`, and list the five source modules exercised by the workflow.\n\n"}
{"task_id": "candidate-1682-science-physics", "source_id": "candidate-1682-science-physics", "domain": "terminal", "task_path": "tasks/candidate-1682-science-physics", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:512c33b53f073a3ff35859d2def7135b5c39f89e9c38d1684e97c943ef6496fd", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 360.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair programmatic particle names across the domain model\n\nThe vendored Scikit-HEP `particle` source under `/app/vendor/particle` has a\nsemantic regression in its Python-safe identifier generation. A trailing `0`\nhas two meanings in the package's domain data:\n\n- for an ordinary neutral particle it is a charge suffix and must be separated\n  in the programmatic identifier, such as `H0` becoming `H_0`;\n- for a nucleus it can be the final digit of the atomic mass number and must\n  remain attached, such as `C10` and `Ag100` remaining `C10` and `Ag100`.\n\nRepair the existing implementation in place. Both of these source modules must\nbe changed:\n\n- `/app/vendor/particle/src/particle/particle/utilities.py`\n- `/app/vendor/particle/src/particle/particle/particle.py`\n\nThe lower-level `programmatic_name` utility must accept explicit nucleus\nclassification while remaining backward compatible for ordinary one-argument\ncallers. The public `Particle.programmatic_name` property must pass the domain\nobject's actual PDG nucleus classification into that utility. Direct utility\ncallers and the public property must therefore agree.\n\nPreserve all established normalization behavior for ordinary neutral names,\ncharged particles, multi-character charges, resonances, parentheses, stars,\nand antiparticle markers. In particular, fixing terminal-zero nuclei must not\nregress names such as `pi0`, `K*(892)0`, `Delta(1232)++`, or `Lambda~`, and an\nanti-nucleus must still receive normal antiparticle normalization.\n\nDo not replace the project with a new evaluator, hard-code only the visible\nexamples, remove bundled PDG/nucleus data, edit or disable tests, or bypass the\nexisting `Particle` lookup and property path. The repaired source tree must\nremain importable offline and the existing focused utility tests must continue\nto pass.\n"}
{"task_id": "candidate-1702-ml-inference", "source_id": "candidate-1702-ml-inference", "domain": "terminal", "task_path": "tasks/candidate-1702-ml-inference", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:6b1a826116bf2f23e68ae41c36cb74a4173a25b563d3eb80a0e7e207bb41446b", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair context-length resolution\n\nYou are given a frozen slice of a Python language-model evaluation harness under\n`/app/vendor/project`. A regression causes automatic context-window detection to ignore the text\nconfiguration nested inside composite model configs. Repair the existing production modules so\nthe integration runner resolves the usable text context correctly.\n\nRequirements:\n\n1. Prefer a nested `text_config` length (`n_positions`, then `max_position_embeddings`, then\n   `n_ctx`) over unrelated top-level modality limits.\n2. Preserve the same attribute precedence for text-only configs.\n3. Use a finite tokenizer `model_max_length` only when config lengths are absent. The Hugging Face\n   infinity sentinel means \"unset\" and must return the default of 2048.\n4. An explicit `max_length` supplied to `HFLM` must override automatic resolution.\n5. Keep resolution logic shared in `lm_eval.models.utils` and make `HFLM.max_length` use it.\n\nRun the offline regression:\n\n```sh\ncd /app\npython3 tools/run_fixture.py inputs.json\n```\n\nIt must produce `/app/output.json` with schema `lm-eval-context-repair.v1` and one result per\ncase, retaining input order. Do not add dependencies, download models, or access the network.\n"}
{"task_id": "candidate-1736-ml-inference", "source_id": "candidate-1736-ml-inference", "domain": "terminal", "task_path": "tasks/candidate-1736-ml-inference", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:725b0369357804a893331bb31905dfa7ac973d86c315499b2ff46ea40174c5de", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair PoissonRegressor mixed Array API handling\n\nWork in `/app`. The frozen scikit-learn source tree is under `/app/vendor/scikit-learn`.\n\nA Poisson generalized linear model currently validates targets and sample weights before moving them to the feature matrix's Array API namespace/device. This breaks fitting and scoring when `X` uses one compatible namespace while `y` or `sample_weight` originates in another. The estimator-check registry also still declares the mixed-input behavior unsupported.\n\nRepair the existing source tree so that:\n\n1. `sklearn/linear_model/_glm/glm.py` moves both `y` and `sample_weight` to `X`'s discovered namespace/device before dtype-sensitive validation in both `fit` and `score`.\n2. Validation, Poisson target-domain checks, weighting semantics, sample order, solver behavior, and prediction mathematics remain unchanged after conversion. Do not replace namespace handling with unconditional NumPy coercion.\n3. `sklearn/utils/_test_common/instance_generator.py` enables the mixed-input Array API estimator check for `PoissonRegressor`, while retaining its separate same-namespace limitation.\n4. Existing source and test assets remain intact. Do not modify `/app/run_regression.py` or bypass the source methods.\n5. Run `python3 /app/run_regression.py`. It must succeed offline and produce `/app/output.json` with `output_schema_version` equal to `tbench.poisson-array-api-report.v1`, successful fit/score mixed-namespace checks, and the updated estimator-check status.\n"}
{"task_id": "candidate-1789-security-appsec", "source_id": "candidate-1789-security-appsec", "domain": "terminal", "task_path": "tasks/candidate-1789-security-appsec", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:7859c93e06fb105b5f6237da814dc7386cb49cce4d2e8dd5102109f77f36d957", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair explicit issue locations in Bandit\n\nThe frozen Bandit tree in `/app/vendor/bandit` has an integration defect in its\nfile-level Trojan Source check. The plugin finds the correct source line and\ncolumn, but the central test runner replaces part of the location metadata with\nthe synthetic file context used to invoke file-level plugins.\n\nRepair the existing implementation. You must modify both:\n\n- `/app/vendor/bandit/bandit/core/tester.py`\n- `/app/vendor/bandit/bandit/plugins/trojansource.py`\n\nRequired behavior:\n\n1. A plugin-provided, non-empty `Issue.linerange` is authoritative and survives\n   `BanditTester.run_tests` unchanged.\n2. An issue that leaves `linerange` empty still receives the existing context\n   fallback. This must remain generic; do not special-case B613 or filenames.\n3. B613 remains a high-severity, medium-confidence finding. For the bundled\n   UTF-8 example it reports line 4, column 25, and line range `[4]`.\n4. The bundled Latin-1 example remains decodable and produces no B613 finding.\n5. Preserve the real Bandit package, plugin registration, fixtures, and normal\n   issue processing. Do not modify tests, suppress findings, or replace the\n   scanner with a new standalone implementation.\n\nUse the focused offline workflow:\n\n```bash\ncd /app/vendor/bandit\npython3 /app/tools/check_syntax.py\nPYTHONPATH=/app/vendor/bandit python3 /app/tools/run_location_regression.py \\\n  --output /app/repair_report.json\n```\n\nThe final artifact must be `/app/repair_report.json` with schema version\n`tbench.bandit_location_report.v1`. The supplied runner writes the required\nshape and exits nonzero until all three focused cases pass.\n"}
{"task_id": "candidate-1824-media-design", "source_id": "candidate-1824-media-design", "domain": "terminal", "task_path": "tasks/candidate-1824-media-design", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:4737707ca641f27b4ec8e30f76f4cdee82ae34e467d67107616af0a4a5575c85", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the offline SVG sanitizer\n\nYou are working in `/app/vendor/project`, a frozen production slice of the SVGO optimizer. The supplied command is the integration entrypoint:\n\n```sh\nnode /app/vendor/project/optimize-design.mjs\n```\n\nIt reads `/app/input/design.svg` and must write `/app/output.json`. Repair the existing parser/plugin integration rather than replacing it with a new standalone sanitizer.\n\nThe `removeScripts` plugin must safely sanitize authored SVG design assets while preserving renderable artwork:\n\n- Remove executable `<script>` elements only when their effective namespace is SVG or XHTML; preserve elements in unrelated namespaces.\n- Treat an element as an SVG anchor when it is unprefixed in SVG or uses a prefix whose in-scope declaration resolves to the SVG namespace. Do not treat unrelated-namespace anchors as SVG anchors.\n- Detect `javascript:` and `vbscript:` URLs after trimming leading whitespace, lowercasing, and removing ASCII tab, line-feed, and carriage-return controls. Entity-decoded values must be handled too.\n- Treat only `application/xhtml+xml`, `image/svg+xml`, and `text/html` data media types as executable. Preserve inert image data URLs.\n- When an executable SVG anchor is removed, splice its non-text children into the parent so visible vector artwork remains.\n- Preserve the deterministic SVG stringification and emit exactly this JSON contract:\n\n```json\n{\"output_schema_version\":\"svgo-design-sanitized.v1\",\"optimized_svg\":\"...\"}\n```\n\nDo not modify tests or remove source modules. Keep the workflow offline and CPU-only.\n"}
{"task_id": "candidate-1836-hardware-cad", "source_id": "candidate-1836-hardware-cad", "domain": "terminal", "task_path": "tasks/candidate-1836-hardware-cad", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:9d983ebebcedba3546f09fe474b7b2a41b9eb1079eda0bbf0de433010b558a59", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair ACIS copy isolation\n\nWork in `/app/vendor/ezdxf`, an offline frozen `ezdxf` source tree. A real\n3DSOLID drawing is available at\n`/app/vendor/ezdxf/examples/acistools/3dsolids.dxf`.\n\nRepair the existing implementation so that temporary transformations behave\ncorrectly when ACIS BODY-family entities are copied:\n\n1. Run the existing ACIS transformation test and reproduce the issue with a\n   non-identity pending `Matrix44`.\n2. Make a copied entity preserve the complete pending matrix and own an\n   independent mutable temporary-transformation helper. Composing a matrix on\n   the copy must not change the source.\n3. Exercise the real pipeline: put a copied 3DSOLID in a named block, create a\n   translated INSERT, run `ezdxf.bbox.extents()` over the INSERT, and verify\n   that virtual-copy transforms do not attach to the block definition entity.\n4. Save and read the drawing back. Preserve the ACIS payload and named-block\n   structure.\n\nKeep the change in the existing source modules and preserve the common\n`TemporaryTransformation` abstraction. Do not clear or bypass transformations\ninside the bounding-box code, discard an existing matrix, or modify tests to\nhide the regression.\n\nWhen complete, run the workflow below; it must write the required artifact:\n\n```sh\nPYTHONPATH=/app/vendor/ezdxf/src python3 /app/run_repair.py\n```\n\nThe output must be `/app/output.json` with `output_schema_version`\n`cad-copy-isolation.v1`.\n"}
{"task_id": "candidate-1866-ml-inference", "source_id": "candidate-1866-ml-inference", "domain": "terminal", "task_path": "tasks/candidate-1866-ml-inference", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:ca717187d4b54ed4fbebf05da09687913b454797af81d9eb59aef849854f2554", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\nRepair the frozen Hugging Face Transformers source tree under\n`/app/vendor/transformers`. A pretrained inference path currently loses the\nadapter keyword-argument state at the boundary between adapter discovery and\n`from_pretrained`.\n\nComplete the following workflow:\n\n1. Inspect the adapter integration helper and its caller in the pretrained model\n   loader. Repair both existing production modules so their return/unpack\n   contract remains synchronized.\n2. Preserve adapter keyword arguments through local adapter discovery. A private\n   `_adapter_model_path` hint must be consumed by discovery and must not be\n   forwarded later, while ordinary adapter options remain unchanged.\n3. Preserve offline cache behavior: `local_files_only`, revision, subfolder, and\n   resolved commit metadata belong to discovery and must not be mixed into the\n   adapter kwargs passed to the later adapter load.\n4. Keep early-return behavior valid when PEFT is unavailable or the model path is\n   absent, including the caller-visible adapter kwargs value.\n5. Run the included integration workflow and write `/app/output.json`:\n\n   ```sh\n   python3 /app/tools/run_adapter_workflow.py \\\n     --cases /app/fixtures/adapter_cases.json \\\n     --output /app/output.json\n   ```\n\nThe output must use schema version `transformers.adapter_contract.v1`, include\none result for every fixture case, describe the repaired caller contract, and\nset every reported check to `true`. Compile the two modified modules before\nfinishing. Do not modify the fixture, runner, unrelated vendored files, or add a\nreplacement evaluator. Runtime network access and package installation are not\navailable.\n"}
{"task_id": "candidate-1890-ml-training", "source_id": "candidate-1890-ml-training", "domain": "terminal", "task_path": "tasks/candidate-1890-ml-training", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:7ba1a43b030037d71b95e452487761c5659a6d40caa614c812b8698c418bb200", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the training-stop lifecycle\n\nYou are given a source-focused, offline PyTorch Lightning workspace. A CPU\nregression driver in `/app/run_regression.py` exercises a stateful\n`EarlyStopping` callback together with the fit loop's minimum-training gates.\nThe current source contains a lifecycle defect: a callback decision made before\n`min_epochs` can remain latched and fire later even though the metric has\nrecovered.\n\nRepair the existing source tree; do not replace it with a new evaluator. Your\nfix must modify both of these production modules:\n\n- `vendor/pytorch-lightning/src/lightning/pytorch/callbacks/early_stopping.py`\n- `vendor/pytorch-lightning/src/lightning/pytorch/loops/fit_loop.py`\n\nRequirements:\n\n1. A patience/threshold decision made before `min_epochs` must be deferred,\n   leave `trainer.should_stop` false, and clear the callback's stopping reason.\n2. After the epoch gate is met, ordinary early stopping must still set\n   `should_stop`, `stopped_epoch`, and `PATIENCE_EXHAUSTED` correctly.\n3. `min_steps` is an independent batch-granularity gate. Do not suppress a\n   callback decision merely because `min_steps` has not been reached.\n4. A metric improvement after a deferred decision must reset the wait counter\n   and must not inherit a stale stop decision.\n5. Keep the existing framework structure and preserve the distributed boolean\n   reduction call.\n\nRun the supplied CPU regression driver and write the required artifact to\n`/app/output.json`. The artifact must retain\n`output_schema_version = tbench.early_stopping.repair.v1` and include both the\ncallback `cases` and the `fit_loop_gates` matrix produced by the driver; do not\nhard-code or hand-edit the expected results.\n"}
{"task_id": "candidate-1939-software-systems", "source_id": "candidate-1939-software-systems", "domain": "terminal", "task_path": "tasks/candidate-1939-software-systems", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:d7066c22ca55bab53a6d469b6a2a8a5492a3367b7417441257cf400792295bb8", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair compressed wheel-tag validation\n\nWork in `/app`, an offline checkout of a Python packaging compatibility library. A recent compatibility regression is exposed at two public entry points: direct compressed-tag parsing and wheel filename parsing. Diagnose the behavior from the existing source and focused tests, then repair the existing implementation rather than replacing it with a new standalone evaluator.\n\nRequirements:\n\n1. A compressed tag has interpreter, ABI, and platform components. Dots inside a component represent compression, but interpreter members must be valid identifier-like names; numeric-looking members such as `2.7.6` must not silently expand into three interpreters.\n2. Preserve valid identifiers containing digits and underscores, including `py3`, `graalpy311`, and `_custom`, as well as valid compressed tags such as `py3.py2-none-any`.\n3. Keep the public exception boundary: direct tag parsing must raise `InvalidTag`, while wheel filename parsing must raise `InvalidWheelFilename` for malformed tag semantics. Do not break valid filename normalization, versions, build tags, or sorted-tag validation.\n4. Modify the existing production modules that own these behaviors (`src/packaging/tags.py` and `src/packaging/utils.py`). Keep the existing package and regression tests intact.\n5. Run the provided workflow from `/app/tools/run_compatibility_workflow.py` and leave its result at `/app/repair_report.json`. The report must retain the declared `packaging-compatibility-report.v1` schema.\n\nSuggested workflow: inspect the call chain, reproduce both public failures, implement the shared validation and utility-level error translation, run focused tests, and then run the workflow to produce the report. No network, package installation, or external service is available or required.\n"}
{"task_id": "candidate-1990-security-cryptography", "source_id": "candidate-1990-security-cryptography", "domain": "terminal", "task_path": "tasks/candidate-1990-security-cryptography", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:fac89f96a327916515da7475c8ad53783da8904613752fdd7709a6d6076c8fd3", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair nonce-derived AES-CCM message limits\n\nWork in `/app/vendor/pycryptodome`. The frozen source already includes its local Python modules and ABI3 native extensions, so the task must run offline with `PYTHONPATH=/app/vendor/pycryptodome/lib`.\n\nCCM encodes the message length in `Q = 15 - len(nonce)` bytes. The current implementation can initialize or process a message whose length cannot be represented in those `Q` bytes. Repair the existing source tree rather than adding a replacement cipher or bypassing its APIs.\n\n## Required source repairs\n\nModify both existing modules:\n\n- `/app/vendor/pycryptodome/lib/Crypto/Cipher/_mode_ccm.py`\n- `/app/vendor/pycryptodome/lib/Crypto/Cipher/_mode_ccm.pyi`\n\nThe repaired implementation must:\n\n1. Expose `CCMMessageTooLongError`, a `ValueError` subclass, from the CCM mode module and its typing stub.\n2. Reject a declared `msg_len` when `msg_len >= 2 ** (8 * Q)` before CCM MAC/counter state is initialized.\n3. When `msg_len` is omitted, reject an oversized first `encrypt()` or `decrypt()` input with the same exception before producing or accepting output. An `output=` buffer must remain unchanged on rejection.\n4. Derive the bound from the actual nonce length. For example, a 13-byte nonce cannot encode 65536 bytes, while a 12-byte nonce can.\n5. Preserve the public `AES.new(..., AES.MODE_CCM, ...)` API, valid associated-data handling, streaming behavior, ciphertext/tag values, and decrypt/verify behavior.\n\nDo not disable or replace the vendored tests, native modules, AES dispatch, or regression runner.\n\n## Artifact\n\nAfter repairing the source, run:\n\n```bash\ncd /app\nPYTHONPATH=/app/vendor/pycryptodome/lib python3 /app/regression/run_ccm_regression.py\n```\n\nThis must create `/app/ccm_repair_report.json` with schema version `tbench.ccm_repair_report.v1` and all checks passing. The private verifier independently recomputes the cryptographic and boundary behavior from trusted fixtures.\n\n"}
{"task_id": "candidate-2072-software-data-engineering", "source_id": "candidate-2072-software-data-engineering", "domain": "terminal", "task_path": "tasks/candidate-2072-software-data-engineering", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:f85d71ae7304ceabc856304d4e6dccbb3fcc0182ca81f7a013324a630a546730", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair redundant Parquet partition metadata\n\nA local data pipeline reads Hive-style partitioned Parquet datasets through two Dask backends. Some producers retain every partition key as a physical file column, while others omit every key and rely on directory metadata. The current frozen readers do not handle these layouts consistently.\n\nRepair both existing modules:\n\n- `/app/vendor/dask/dask/dataframe/io/parquet/arrow.py`\n- `/app/vendor/dask/dask/dataframe/io/parquet/fastparquet.py`\n\nImplement one policy in both engine paths:\n\n1. If **all** discovered Hive partition names are also physical file columns, treat the directory partitions as redundant. The physical columns must appear exactly once, in their original order. Clear associated virtual partition state before downstream part construction.\n2. If **none** of the partition names are physical columns, retain the directory partitions so they are materialized after the physical columns.\n3. If only a **proper subset** overlaps, raise `ValueError`; silently dropping either side is ambiguous.\n4. Preserve behavior when no partition metadata exists and keep the lazy metadata/part construction flow intact.\n\nThis is a source repair, not a report-only task. Do not replace the vendored modules or regression tool, disable checks, or add a parallel reader. After repairing both modules, run:\n\n```bash\ncd /app\npython3 tools/run_overlap_regression.py\n```\n\nThe command must succeed and create `/app/regression-report.json` with schema version `dask-parquet-overlap-report.v1`, both engines listed, all six case checks true, and `passed: true`.\n"}
{"task_id": "candidate-2160-software-frontend", "source_id": "candidate-2160-software-frontend", "domain": "terminal", "task_path": "tasks/candidate-2160-software-frontend", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:debd048e7ee0f778fe4f173e23b37c0a7684ed5f638d59064668fcc27a41e088", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\nThe existing event form in `/app/vendor/esn-frontend-calendar` incorrectly lets a calendar owner move an existing event even when that user is not the event organizer. Repair the production integration while preserving the frozen source tree and existing public APIs outside this workflow.\n\nWork in these existing modules:\n\n- `/app/vendor/esn-frontend-calendar/src/esn.calendar.libs/app/services/cal-ui-authorization-service.js`\n- `/app/vendor/esn-frontend-calendar/src/esn.calendar.libs/app/components/event/form/event-form.controller.js`\n\nRequired behavior:\n\n1. `calUIAuthorizationService.canMoveEvent` must evaluate the selected calendar, the existing event, and the current session user together. Movement is allowed only when the session user owns the calendar and is the event organizer.\n2. Calendar ownership must be checked with the session user's identifier, while organizer identity must use the full event and user objects through the existing `calEventUtils.isOrganizer` contract.\n3. Update both event-form controller call sites to use the same event-aware authorization signature: the initialization branch that decides whether to restrict the calendar list and the later authorization aggregation that sets `$scope.canMoveEvent`.\n4. Preserve the new-event branch, writable-calendar filtering, modification/recurrence checks, and the existing move guard. Do not hard-code fixture identities, replace the controller/service, remove tests, or write a parallel authorization implementation.\n5. Generate `/app/output.json` with schema version `calendar-authorization-repair.v1` by running:\n\n```sh\ncd /app\n./tools/check-calendar-repair.sh\n```\n\nThe command must complete successfully after the repair. Runtime network access and dependency installation are not available.\n"}
{"task_id": "candidate-2163-operations-marketing", "source_id": "candidate-2163-operations-marketing", "domain": "terminal", "task_path": "tasks/candidate-2163-operations-marketing", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:728c36759babf1a370503139431f25e7b8e32c6e118ffc23554f1f5ca7a95b1d", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair experiment exposure telemetry\n\nWork in `/app/vendor/flagsmith-python-client` and repair the existing Python SDK implementation. The current code can inflate experiment exposure counts and can attribute customers to a feature payload even when no experiment variant was assigned.\n\nModify both of these existing production modules:\n\n- `flagsmith/flagsmith.py`\n- `flagsmith/analytics.py`\n\nThe repaired integration must satisfy all of the following:\n\n1. `Flagsmith.get_experiment_flag(...)` must always return the resolved flag, but emit an exposure only when the result is a real `Flag`, the flag is enabled, and `flag.variant` is not `None`. Default-handler results, disabled flags, and variant-less flags must not emit exposure telemetry.\n2. A qualified exposure must use the experiment variant key as its event value, never the feature-state display payload.\n3. Exposure tracking requires a non-empty customer identifier. A direct exposure call without one must be ignored rather than buffered.\n4. Within one active event buffer, repeated exposures with the same feature name, customer identifier, and normalized value must collapse to one buffered record. Differences in any of those three fields must remain distinct.\n5. Deduplication applies only to `$flag_exposure` events. Repeated ordinary custom or conversion events must preserve their full multiplicity, even when their payloads are identical.\n6. A flush starts a new exposure-deduplication window, so the same exposure may be accepted again after the buffered batch is handed off. Existing failed-delivery requeue behavior must continue retaining the attempted events.\n7. Preserve the existing event payload shape, value stringification, automatic maximum-buffer flush behavior, timer lifecycle, and public method signatures.\n\nDo not delete or replace the vendored package or its tests, disable behavior, add network dependencies, or solve the task by creating a separate script outside the existing SDK modules.\n"}
{"task_id": "candidate-2207-security-appsec", "source_id": "candidate-2207-security-appsec", "domain": "terminal", "task_path": "tasks/candidate-2207-security-appsec", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:dfc9cc793105617f241685a543d863f8e69f46ab4b8b84002fab050e3f356a97", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the YAML-load security analysis\n\nYou are working in `/app/vendor/bandit`, a small frozen Bandit source tree. The\nworkspace file `/app/vendor/bandit/workspace/yaml_cases.py` contains safe and unsafe\nuses of `yaml.load`.\n\nRepair the existing implementation so B506 behaves correctly for both positional\nand keyword loader arguments:\n\n- Positional `SafeLoader`, `yaml.SafeLoader`, `CSafeLoader`, and `yaml.CSafeLoader`\n  are safe and must not produce B506 findings.\n- Missing loaders, `yaml.Loader`, arbitrary positional loader objects, and keyword\n  `Loader=yaml.Loader` remain findings.\n- Preserve the existing AST context behavior for names, literals, qualified\n  attributes, and keyword arguments; do not weaken unrelated callers.\n\nWork in the existing `bandit/core/context.py` and `bandit/plugins/yaml_load.py`\nmodules. Do not replace the analyzer with a standalone script or edit the fixture\nor tests to hide findings. Run the focused context regression if useful, then run\nthe real CLI integration path against the workspace and write exactly one JSON\nreport to `/app/output.json`.\n\nThe image already provides the offline compatibility modules needed by this frozen\nBandit tree. Bandit exits with status 1 when security findings remain; that status\nis expected for the final mixed safe/unsafe scan, provided the JSON report is\nwritten successfully.\n\nThe report must be Bandit JSON output with `results`, `metrics`, and `errors` keys.\nIt must contain only the B506 findings from the workspace scan, with the original\nfile paths and line numbers preserved. Ensure the output is valid JSON and does not\ninclude a trailing diagnostic stream in the artifact.\n"}
{"task_id": "candidate-2262-operations-compliance", "source_id": "candidate-2262-operations-compliance", "domain": "terminal", "task_path": "tasks/candidate-2262-operations-compliance", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:5ce02d560dd85c9d32a244ada2953e47aa5c449395866ff2d2746205a30e233c", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair XI VAT routing and canonicalization\n\nAn offline invoice-compliance pipeline uses the frozen python-stdnum source in\n`/app/vendor/project`. Its EU VAT router and United Kingdom VAT validator do\nnot correctly handle the `XI` prefix used for Northern Ireland movements.\n\nRepair the existing production modules:\n\n- `/app/vendor/project/stdnum/eu/vat.py`\n- `/app/vendor/project/stdnum/gb/vat.py`\n\nDo not replace the library with a new validator or weaken its existing United\nKingdom rules. The repaired behavior must satisfy all of these requirements:\n\n1. `XI` is a recognized EU VAT jurisdiction and dispatches to the existing GB\n   VAT implementation. `GB` itself remains outside `MEMBER_STATES` and is not\n   accepted by the EU router.\n2. The GB implementation accepts either `GB` or `XI` as a leading prefix and\n   removes exactly one such prefix during compaction. It must not strip an\n   arbitrary pair of letters or repeatedly remove duplicate prefixes.\n3. EU-layer compaction and validation preserve the canonical `XI` prefix,\n   while direct GB-layer validation returns the prefix-free GB payload.\n4. Existing 9/12-digit checksum rules and the `GD`/`HA` special-format\n   component rules remain enforced. Invalid checksums, malformed payloads,\n   duplicate prefixes, and invalid special-number ranges must stay invalid.\n\nThe deterministic workload is `/app/input/invoices.json`. After repairing the\nsource, run:\n\n```bash\npython3 /app/run_audit.py \\\n  --input /app/input/invoices.json \\\n  --output /app/output.json\n```\n\nThe artifact must follow `/app/output_contract.json` with\n`output_schema_version` equal to `vat-compliance-report.v1`. Include every\ninput row exactly once in input order. Do not modify the workload, contract,\nrunner, license, utility modules, or frozen upstream regression file.\n\nBefore finishing, run focused syntax and behavior checks against both repaired\nmodules and regenerate `/app/output.json` from the final source tree.\n"}
{"task_id": "candidate-2329-software-languages", "source_id": "candidate-2329-software-languages", "domain": "terminal", "task_path": "tasks/candidate-2329-software-languages", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:d25448699ef4e8e93599ca32c6c229dbe6fab54f3d9f29e101b186dfe7a11acd", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair interactive LALR parser copying\n\nThe vendored project in `/app/vendor/lark` is a pre-fix production source\ntree. An incremental LALR parse can be forked with `InteractiveParser.copy()`.\nWhen the fork is lexed and resumed, the two parser objects must own separate\nlexer cursors while retaining their already-built parse stacks and semantic\nvalues. The current implementation violates that ownership contract and can\nreplay input on one branch.\n\nDiagnose the ownership path across these existing modules:\n\n- `lark/parsers/lalr_interactive_parser.py`\n- `lark/parsers/lalr_parser_state.py`\n- `lark/lexer.py` (the `LexerThread` copy behavior is an important dependency)\n\nRepair the existing implementation, without replacing the parser with a new\nscript or changing tests. Preserve the public `copy`, `feed_token`,\n`exhaust_lexer`, `resume_parse`, and immutable-parser APIs.\n\nUse at least these workflow stages:\n\n1. Reproduce the bug by constructing a `Lark(..., parser=\"lalr\")`, copying an\n   interactive parser before lexing, exhausting the copy, and resuming it.\n2. Correct the coordinated state/lexer ownership in both mutable parser\n   modules. A copied parser state must retain independent stack/value storage,\n   and its `lexer` must be the same object exposed as that instance's\n   `lexer_thread`.\n3. Run independent branch checks for multiple grammars, then write the report\n   described below.\n\nCreate `/app/output.json` with exactly this top-level shape:\n\n```json\n{\n  \"output_schema_version\": \"interactive-parser-repair.v1\",\n  \"cases\": [\n    {\n      \"name\": \"...\",\n      \"copy_leaves\": [\"...\"],\n      \"original_leaves\": [\"...\"],\n      \"copy_owns_lexer\": true,\n      \"original_owns_lexer\": true\n    }\n  ],\n  \"all_passed\": true\n}\n```\n\nInclude one entry for each regression grammar exercised. `copy_leaves` and\n`original_leaves` must contain the token values in source order, exactly once\nper input token. Do not include private verifier files or network-derived\ndata. The verifier recomputes the parser behavior itself, so a fabricated\nreport cannot substitute for the source repair.\n\n"}
{"task_id": "candidate-2376-security-cryptography", "source_id": "candidate-2376-security-cryptography", "domain": "terminal", "task_path": "tasks/candidate-2376-security-cryptography", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:090c12b68da2bbe3d0d2872204b1e56a377f895090f0132aba20dbe51a8a60ca", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair JOSE key provenance\n\nYou are working in a frozen Authlib source tree under `/app/vendor/authlib`.\nAn application accepts JWS and JWE messages that may contain a `jwk` header.\nTreat that header as untrusted metadata: it must never become the verification\nor decryption key merely because the caller omitted a key.\n\nRepair the existing implementation, not just the report script. Trace the\ncompact and JSON deserialization paths into the algorithm key-preparation code\nand make the same trust-boundary rule hold for both JWS and JWE. Preserve the\nexisting algorithm allow-listing, explicit key handling, and callable key\nresolver behavior. Do not reject a message merely because it has a `jwk`\nheader when the caller explicitly supplies the trusted key.\n\nAfter editing the source modules, run:\n\n```sh\npython3 /app/run_repair_check.py\n```\n\nThe command must write `/app/output.json` using schema version\n`jose-trust-report.v1`. It must report successful explicit JWS, explicit JWE,\nand callback verification, while both omitted-key cases are rejected. Keep the\nvendored package and its existing tests intact; do not install dependencies or\nuse the network.\n"}
{"task_id": "candidate-2487-security-cryptography", "source_id": "candidate-2487-security-cryptography", "domain": "terminal", "task_path": "tasks/candidate-2487-security-cryptography", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:a2d9c32ac607a8ab7ba95a6636188a76d1b8beeb0cf5ab7bd0e10d45f53f0ff1", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the offline JWT verification workflow\n\nYou are given a vendored PyJWT source tree at `/app/vendor/pyjwt`, frozen before a security-hardening release. Repair the existing production modules; do not replace the library with a new standalone implementation.\n\nThe integration workflow in `/app/run_workflow.py` must finish successfully and write `/app/output.json`. Preserve ordinary HS256 signing and verification while making the following behavior correct across the existing layers:\n\n- A JSON JWK or asymmetric public-key representation supplied as a raw HMAC secret must be rejected, even when the caller allows several algorithm families. A `PyJWK` must remain bound to its declared algorithm and reject a token advertising a different algorithm.\n- Implement RFC 7797 handling for `b64=false`: the protected header must declare `b64` in `crit`; detached payload verification must use the supplied bytes; malformed non-detached forms must be rejected; ordinary encoded payloads must continue working.\n- A per-call `options={\"enforce_minimum_key_length\": true}` passed through `jwt.decode` must enforce the HMAC minimum length without changing valid long-key behavior.\n- `PyJWKClient` may only accept HTTP(S) URI schemes. A failed refresh must raise its normal client error without deleting an already cached usable JWK set.\n\nUse the existing package APIs and regression tests as guides. Keep the changes limited to the real source modules and leave the output schema emitted by the workflow unchanged (`output_schema_version` `pyjwt-repair.v1` with a `cases` object).\n"}
{"task_id": "candidate-2500-hardware-cad", "source_id": "candidate-2500-hardware-cad", "domain": "terminal", "task_path": "tasks/candidate-2500-hardware-cad", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:864865850ff2964921719b70d2fe04c36c8cc6c6e2dacd12d2a3a7b626c77ecc", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair OpenSCAD range semantics\n\nYou are working in a frozen slice of the real `openscad_parser` source tree at\n`/app/vendor/openscad_parser`. Repair the existing implementation; do not\nreplace it with a standalone parser, modify vendored metadata, or add/download\ndependencies.\n\nOpenSCAD uses `[start:step:end]` for a three-expression range and `[start:end]`\nfor a range whose implicit step is `1.0`. The current AST path does not preserve\nthat contract consistently. Diagnose the PEG-to-AST flow and repair both\nexisting production modules involved in constructing and rendering\n`RangeLiteral` values:\n\n- `src/openscad_parser/ast/builder.py`\n- `src/openscad_parser/ast/nodes.py`\n\nThe repaired tree must satisfy all of these behaviors:\n\n1. Explicit positive, fractional, and descending steps occupy the AST `step`\n   field, while the final expression occupies `end`.\n2. Two-expression ranges retain `step = 1.0` and keep their second expression\n   as `end`.\n3. String/pretty-print output uses valid `[start : step : end]` order and can be\n   parsed again without changing the three semantic fields.\n4. AST dictionary/JSON serialization and deserialization preserve `start`,\n   `step`, and `end`, including a negative step expression.\n5. Existing scope propagation continues to reach all three range expressions.\n\nAfter repairing the source, parse `/app/cad_models.scad` with the public AST\nAPI and write `/app/output.json`. Use this exact top-level schema:\n\n```json\n{\n  \"output_schema_version\": \"openscad-range-repair.v1\",\n  \"source\": \"cad_models.scad\",\n  \"ranges\": {\n    \"layer_heights\": {\n      \"start\": 0.4,\n      \"step\": 0.6,\n      \"end\": 4.6,\n      \"rendered\": \"[0.4 : 0.6 : 4.6]\",\n      \"roundtrip_stable\": true\n    }\n  }\n}\n```\n\nInclude one entry for every assignment in the input file. Derive all values and\nrendered strings from the repaired AST; do not hard-code only the example.\nKeep the run CPU-only and offline.\n"}
{"task_id": "candidate-2511-security-cryptography", "source_id": "candidate-2511-security-cryptography", "domain": "terminal", "task_path": "tasks/candidate-2511-security-cryptography", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:e4780d105895dde1d5a392fefb9274f3abe2008f3273b0e4025d51d861005c86", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the HMAC key boundary\n\nYou are working in `/app/vendor/project`, a vendored JOSE implementation. A\nservice chooses the algorithm from a compact JWT header and then constructs a\nbackend key object. Reproduce and repair the unsafe behavior at that boundary.\n\n## Required workflow\n\n1. Trace compact JWT decoding through `jose/jwt.py`, `jose/jws.py`, and\n   `jose/jwk.py` into both HMAC backend implementations.\n2. Use `/app/cases.json` to exercise the behavior. Ordinary byte-string\n   secrets, including a secret containing dashes, must continue to work.\n3. Make the shared key classification recognize serialized asymmetric material\n   rather than relying on a short list of exact substrings. Cover PEM labels\n   used for public/private keys, certificates, CSRs, parameters, and CRLs, with\n   normal whitespace, and cover OpenSSH public and certificate key types.\n4. Enforce the same policy in the native and cryptography-backed HMAC paths.\n5. Run focused regression checks and write `/app/output.json` with exactly this\n   schema:\n\n```json\n{\n  \"output_schema_version\": \"jose-key-boundary.v1\",\n  \"results\": [\n    {\"id\": \"case id\", \"accepted\": true, \"claims\": {\"...\": \"...\"}}\n  ]\n}\n```\n\nInclude one result for every case, preserving input order. For accepted cases,\n`claims` must exactly equal the case claims. For rejected cases, set\n`accepted` to `false` and omit `claims`.\n\nThe result must be produced by running the repaired source tree, not by a\nhard-coded report. Do not install packages, access the network, or modify the\ntests. A successful repair preserves valid HS256/HS384/HS512 signing and\nverification while rejecting asymmetric key material before MAC verification.\n"}
{"task_id": "candidate-2517-software-data-engineering", "source_id": "candidate-2517-software-data-engineering", "domain": "terminal", "task_path": "tasks/candidate-2517-software-data-engineering", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:d0cff26f480dd639a9d0eb1e84a10c6a4b112fa6e730c80001c622fab3ae85a2", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Task: propagate pipeline environment into worker execution\n\nRepair the frozen `datapackage-pipelines` source tree under\n`/app/vendor/datapackage-pipelines` so a pipeline-level `environment` mapping\nsurvives discovery and is applied only to that pipeline's worker subprocess.\n\n## Required behavior\n\n1. `PipelineSpec` must expose an `environment` mapping. Direct construction and\n   pipeline specifications that omit the field must produce an independent\n   empty mapping without breaking existing callers.\n2. During pipeline discovery, remove the top-level `environment` entry from the\n   details sent through the existing schema-validation, processor-resolution,\n   schedule, dependency, and hash path, and retain it on the corresponding\n   `PipelineSpec`.\n3. `remote_execute_pipeline()` must launch the existing `dpp run --slave ...`\n   worker with a copy of the parent environment plus that spec's mapping.\n   Convert every YAML scalar value to `str` before passing it to the subprocess.\n4. Never mutate the parent process environment. A configured pipeline must not\n   leak values into a later sibling pipeline that omits `environment`.\n5. Preserve the existing command, cwd, stderr parsing, progress reporting,\n   schedule/dependency handling, and cache-hash behavior.\n\nThe intended production modules are:\n\n- `/app/vendor/datapackage-pipelines/datapackage_pipelines/specs/parsers/base_parser.py`\n- `/app/vendor/datapackage-pipelines/datapackage_pipelines/specs/specs.py`\n- `/app/vendor/datapackage-pipelines/datapackage_pipelines/manager/runner.py`\n\nDo not modify the public workload, schema, other production modules, or delete\nsource files. Do not add network or package-install steps.\n\n## Required artifact\n\nAfter repairing the source, run:\n\n```bash\npython3 /app/workload/check_repair.py\n```\n\nThe command must succeed and write `/app/workload/demo_result.json` with\n`output_schema_version` equal to `candidate2517.demo.v1`. The artifact is only a\npublic smoke result; private scoring independently imports the repaired source,\ndiscovers different YAML pipelines, launches a real local worker subprocess,\nand checks omission, scalar conversion, isolation, and repeatability.\n\n"}
{"task_id": "candidate-2535-security-cryptography", "source_id": "candidate-2535-security-cryptography", "domain": "terminal", "task_path": "tasks/candidate-2535-security-cryptography", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:eba881c4e34d3a8f9edc004f7b08f7c2589f1ca2b35146ded068f65d3ef4ca19", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the SSH lifecycle exception boundaries\n\nYou are working in the vendored Paramiko project at `/app/vendor/paramiko`. Diagnose and repair the production implementation, rather than adding a replacement wrapper or changing tests.\n\nSeveral SSH-agent, channel, client-socket, and transport-teardown paths currently use exception boundaries that are too broad. Preserve existing behavior for ordinary operational failures, including local SSH-agent discovery fallback, but do not let `KeyboardInterrupt` or `SystemExit` disappear during cleanup, forwarding, timeout setup, or transport shutdown.\n\nUpdate all four relevant production modules: `paramiko/agent.py`, `paramiko/channel.py`, `paramiko/client.py`, and `paramiko/transport.py`. Keep public APIs unchanged. Run a focused compile/regression check offline, then write `/app/output.json` using the schema shown below. Hashes must describe the final files.\n\n```json\n{\"schema_version\":\"paramiko-repair-report.v1\",\"status\":\"repaired\",\"compiled\":true,\"modules\":{\"paramiko/agent.py\":{\"bare_handlers\":[],\"sha256\":\"...\"},\"paramiko/channel.py\":{\"bare_handlers\":[],\"sha256\":\"...\"},\"paramiko/client.py\":{\"bare_handlers\":[],\"sha256\":\"...\"},\"paramiko/transport.py\":{\"bare_handlers\":[],\"sha256\":\"...\"}},\"control_flow_policy\":{\"ordinary_exception\":\"handled per lifecycle\",\"base_exception\":\"propagates\"}}\n```\n\nDo not use network access, install packages, or replace Paramiko with a standalone evaluator.\n"}
{"task_id": "candidate-2684-security-appsec", "source_id": "candidate-2684-security-appsec", "domain": "terminal", "task_path": "tasks/candidate-2684-security-appsec", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:5500f28f54d7b1440f51ccf8fde50ae3b9b278c1283894effe9c559474904f4a", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair precise security finding ranges\n\nYou are working in `/app/vendor/bandit`, a small frozen source tree from the\nBandit static analyzer. A regression currently causes findings produced by a\nfile-level security check to lose the source range they computed. This is\nespecially dangerous for Trojan Source detection: a report must identify the\nactual decoded source line containing the bidirectional control character, not\nthe surrounding raw context.\n\nRepair the existing Bandit modules, rather than replacing the scanner with a\nnew standalone implementation. Your changes must preserve all of the\nfollowing:\n\n1. The file-level Trojan Source check remains enabled and reports a B613 issue\n   for the first suspicious bidirectional control character.\n2. Its one-based `line_number`, one-based `col_offset`, and exact one-line\n   `line_range` survive the shared tester/result-normalization path.\n3. Ordinary AST checks that do not provide a range continue to receive the raw\n   context range.\n4. Existing `# nosec` filtering and source-encoding detection continue to\n   work.\n5. Run the real scanner workflow using the supplied fixture and write\n   `/app/output.json` with this exact top-level contract:\n\n```json\n{\n  \"output_schema_version\": \"bandit-range-report.v1\",\n  \"results\": [\n    {\n      \"filename\": \"...\",\n      \"test_id\": \"B613\",\n      \"line_number\": 4,\n      \"line_range\": [4]\n    }\n  ]\n}\n```\n\nThe report may contain the other fields emitted by Bandit, but it must be\nvalid JSON, have no scan errors, and include the B613 result for the supplied\nfixture. Use the repository's existing modules and `run_scan.py`; do not use\nthe network or install dependencies.\n"}
{"task_id": "candidate-2713-ml-inference", "source_id": "candidate-2713-ml-inference", "domain": "terminal", "task_path": "tasks/candidate-2713-ml-inference", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:8573abd9c68fe60b5f8cc517f4bf3bffd2d0babbe2a14fda30626d3bb9c331ba", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the binary evaluation plotting workflow\n\nThe workspace contains a small, frozen slice of a mature Python ML library. Repair the two public plotting entry points under `vendor/project/sklearn/metrics/_plot/` so they work correctly with fitted binary estimators that expose a response method through a preprocessing-style wrapper.\n\nUse the existing response-method helper and metric functions. The repaired workflow must:\n\n- accept raw feature rows through a fitted binary pipeline and preserve the estimator's positive-class ordering;\n- support both `predict_proba` and `decision_function` response methods, including the automatic fallback;\n- reject multiclass, regressors, unfitted estimators, and response arrays with more than two columns rather than silently selecting an arbitrary column;\n- pass `sample_weight` and ROC `drop_intermediate` through unchanged;\n- preserve the display numerical fields and matplotlib line/axes metadata.\n\nRun the repository's evaluation driver after repairing the source. It must create `/app/output.json` with schema version `ml-evaluation.v1`, containing the ROC and precision-recall arrays and their summary metrics.\n# Repair binary classifier evaluation plotting\n\nRepair the frozen scikit-learn source tree in `/app/vendor/project` so its ROC\nand precision-recall plotting helpers work with fitted preprocessing-style\nclassifiers.\n\nYou must modify both existing modules:\n\n- `/app/vendor/project/sklearn/metrics/_plot/roc_curve.py`\n- `/app/vendor/project/sklearn/metrics/_plot/precision_recall_curve.py`\n\nThe repaired helpers must satisfy all of the following:\n\n1. A fitted classifier pipeline may delegate `classes_` and its response\n   method to a final estimator. Raw feature inputs must be passed to that\n   response method without an eager fitted-state rejection at the plotting\n   entry point.\n2. Binary two-column responses use column 1, corresponding to the estimator's\n   second class. One-dimensional decision scores remain one-dimensional.\n3. A response with more than two columns is rejected as non-binary, even when\n   a malformed estimator advertises only two classes.\n4. Non-binary classifiers and regressors are rejected. Invalid or unavailable\n   response methods keep the existing response-method errors.\n5. `sample_weight`, ROC `drop_intermediate`, display names, plotting keyword\n   arguments, numerical display fields, and axes/line metadata remain intact.\n\nDo not replace the modules, bypass the public plotting helpers, modify\n`/app/run_evaluation.py`, or edit unrelated vendored source and upstream test\nassets. Run the offline integration workflow and leave its result at exactly\n`/app/output.json`.\n\nThe output JSON contract is `output_schema_version = \"ml-evaluation.v1\"` and\ncontains exactly these top-level keys:\n\n```text\noutput_schema_version, case_id, roc, precision_recall, display, metadata\n```\n\n`roc` contains `fpr`, `tpr`, and `auc`; `precision_recall` contains\n`precision`, `recall`, and `average_precision`; `display` contains estimator\nand line/axis metadata; and `metadata` records response method, weighting, and\nthe ROC `drop_intermediate` option.\n"}
{"task_id": "candidate-2728-software-data-engineering", "source_id": "candidate-2728-software-data-engineering", "domain": "terminal", "task_path": "tasks/candidate-2728-software-data-engineering", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:e14e756998ffafa1c484d0b5ac5807b60f3bd994eecee8a27da883d88fa71d5a", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair overlapping Parquet partition columns\n\nThe frozen Dask source tree is in `/app/vendor/dask`.\n\nSome producers write a Hive-style directory tree such as\n`region=west/kind=retail/data.parquet` **and** store `region` and `kind` as\nphysical columns in every Parquet file. The current reader treats this as a\nduplicate partition definition and fails. Repair the real metadata paths in\nboth backends:\n\n- `/app/vendor/dask/dask/dataframe/io/parquet/arrow.py`\n- `/app/vendor/dask/dask/dataframe/io/parquet/fastparquet.py`\n\nRequired behavior:\n\n1. When every discovered directory partition column is also present in the\n   physical file schema, use the physical columns and disable directory-based\n   partition injection for that read.\n2. When only a proper subset of directory partition columns is physically\n   stored, raise `ValueError`; the ambiguous partial-overlap case must not be\n   silently accepted.\n3. When there is no overlap, preserve the existing partition metadata and\n   normal read behavior.\n4. Apply the same contract to Arrow and fastparquet. The Arrow metadata result\n   must propagate any normalized partition state to the later part-construction\n   path.\n5. Modify the existing backend modules rather than adding a replacement reader,\n   bypassing Dask, deleting tests, or weakening unrelated behavior.\n\nYou may inspect the surrounding source and existing Parquet tests. Keep the\nrepair focused and compatible with the style of this frozen codebase.\n"}
{"task_id": "candidate-2760-ml-inference", "source_id": "candidate-2760-ml-inference", "domain": "terminal", "task_path": "tasks/candidate-2760-ml-inference", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:b5c4ca4523ec465784d4d4413f1b99e40525ddd508ea20a28418340145851426", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Task: repair River streaming predictive uncertainty\n\nThe frozen River source tree under `/app/vendor/river` has a coordinated inference regression.\nNumerically plausible values flow through the Bayesian regressor, Gaussian distribution,\nrunning variance, empirical covariance, and tree uncertainty helpers, but the state constructor\ncontract mixes variance and standard deviation. This makes learned predictive uncertainty too\nsmall while leaving point predictions apparently reasonable.\n\nRepair the existing production code. Do not replace the library, bypass its classes, disable\ntests, or solve the task only in the audit script.\n\n## Required behavior\n\n1. `Var._from_state` and `Gaussian._from_state` must accept the reconstructed **variance** as the\n   third state value, exposed by the keyword name `var`. `Var.get()` must return that variance,\n   and `Gaussian.sigma` must take its square root exactly once.\n2. `BayesianLinearRegression.predict_one(x, with_dist=True)` must pass its predictive variance\n   into the Gaussian state constructor without taking a premature square root. Its distribution\n   mean must remain equal to `predict_one(x)`, including after online updates.\n3. Preserve River's existing posterior semantics for sparse dictionaries, newly observed query\n   features, and the zero-update case. Unknown query features contribute prior uncertainty but\n   do not change the point prediction.\n4. Update the empirical-covariance and tree uncertainty callers to use the same variance keyword\n   contract. Reconstructed diagonal covariance entries and tree delta-loss variance must retain\n   their numeric values.\n5. Point predictions from `predict_many` must agree with the corresponding `predict_one` calls;\n   the repair must not alter point-inference behavior.\n\nThe expected production modules are:\n\n- `/app/vendor/river/river/linear_model/bayesian_lin_reg.py`\n- `/app/vendor/river/river/proba/gaussian.py`\n- `/app/vendor/river/river/stats/var.py`\n- `/app/vendor/river/river/covariance/emp.py`\n- `/app/vendor/river/river/tree/utils.py`\n\n## Integration artifact\n\nAfter repairing the source, run:\n\n```bash\npython3 /app/workflow/run_uncertainty_audit.py --output /app/output.json\n```\n\n`/app/output.json` must be valid JSON with top-level\n`output_schema_version = \"river-uncertainty-audit.v1\"`. It must contain the audit's dataset and\ntraining metadata, four prediction cases, three batch point predictions, and the six\n`state_contract` values produced by the repaired real code. All numeric values must be finite.\n\nThe task is fully offline. Do not install packages or access the network.\n"}
{"task_id": "candidate-2790-ml-evaluation", "source_id": "candidate-2790-ml-evaluation", "domain": "terminal", "task_path": "tasks/candidate-2790-ml-evaluation", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:851228618143212b846c6af699a339507f2c5e1aa80d514cedbb75e1e311ba05", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\nThe evaluator in `/app/vendor/lm-evaluation-harness` has a regression in hierarchical metric aggregation. Diagnose the interaction between `weight_by_size`, point estimates, and reported standard errors, and repair the existing production modules in `lm_eval/api/metrics.py` and `lm_eval/api/group.py`.\n\nYour submission must preserve filter selection, nested-group traversal, sample-count metadata, and explicit `N/A` propagation. Both weighting modes must remain correct: unweighted groups use the independent-estimates standard error for an arithmetic mean, while weighted groups retain the pooled sample-size standard error.\n\nRun `python3 /app/run_eval.py` after your repair. It must create `/app/output.json` with `schema_version` equal to `lm-eval-group-output.v1` and a `groups` object containing the `inner`, `outer`, and `weighted` records produced by the workflow. Keep the implementation offline and do not modify the supplied input, integration driver, frozen regression tests, or other vendored modules.\n"}
{"task_id": "candidate-2836-ml-kernels", "source_id": "candidate-2836-ml-kernels", "domain": "terminal", "task_path": "tasks/candidate-2836-ml-kernels", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:312cc57ccf115418f9a524484f4b9530d17c7e8c47f339cea17a33271a306b62", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Task\n\nRepair MLX's existing CPU backend handling for zero-length axes in scan, sort,\nand argsort.\n\n## Workspace\n\nThe frozen source tree is at `/app/vendor/mlx`. The production modules that\nmust be repaired are:\n\n- `/app/vendor/mlx/mlx/backend/cpu/scan.cpp`\n- `/app/vendor/mlx/mlx/backend/cpu/sort.cpp`\n\n## Required behavior\n\n1. `Scan::eval_cpu` must still establish a valid output allocation for an empty\n   result, but it must not register arrays or dispatch scan work when the output\n   has zero elements.\n2. `ArgSort::eval_cpu` must likewise allocate the correctly typed and shaped\n   empty index output, then return before command-encoder registration and\n   dispatch.\n3. `Sort::eval_cpu` must preserve its existing input-to-output copy/allocation\n   lifecycle, then return before command-encoder registration and in-place sort\n   dispatch for an empty result.\n4. Non-empty scan behavior, including dtype dispatch, inclusive/exclusive mode,\n   and reverse mode, must remain unchanged.\n5. Non-empty sort and argsort behavior, including stable ordering, axis\n   handling, NaN ordering, and `uint32` index output, must remain unchanged.\n6. Keep the repair inside the existing production modules. Do not modify tests,\n   build metadata, public wrappers, CUDA/Metal backends, or unrelated source.\n\n## Validation\n\nFrom `/app/vendor/mlx`, run:\n\n```bash\ng++ -std=c++20 -fsyntax-only -I. \\\n  mlx/backend/cpu/scan.cpp mlx/backend/cpu/sort.cpp\n```\n\nYour final submission is the repaired source tree in `/app`; no report or new\nartifact file is required.\n"}
{"task_id": "candidate-2838-ml-evaluation", "source_id": "candidate-2838-ml-evaluation", "domain": "terminal", "task_path": "tasks/candidate-2838-ml-evaluation", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:dcc1d2601c44e993b5e2563a37c52afa4c1ead60b4f12a4e93e865d4484d8222", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair scalar and vector classification metric results\n\nThe frozen Evaluate source tree under `/app/vendor/evaluate` is used by an\noffline classification evaluation workflow. Its F1, precision, and recall\nadapters correctly delegate metric calculation to scikit-learn, but their\nresult adaptation assumes every backend return value exposes an array `.size`.\nWith the installed backend, averaged scores can be native scalar objects while\nper-class scores are arrays.\n\nRepair the existing source tree in place.\n\nRequirements:\n\n1. Update all three production modules:\n   `vendor/evaluate/metrics/f1/f1.py`,\n   `vendor/evaluate/metrics/precision/precision.py`, and\n   `vendor/evaluate/metrics/recall/recall.py`.\n2. Preserve each scikit-learn backend call and its public arguments. This\n   includes `labels`, `pos_label`, `average`, `sample_weight`, and the existing\n   `zero_division` arguments where supported.\n3. Return an ordinary Python `float` for scalar backend scores, including\n   native Python and NumPy scalar objects.\n4. Preserve array results for `average=None`, including a one-element array\n   created by selecting a single label. Do not collapse per-class output and do\n   not replace it with a list.\n5. Keep the metric classes, result keys (`f1`, `precision`, and `recall`),\n   feature declarations, smoke inputs, schema, runner, and all other vendored\n   files intact.\n6. Compile the frozen source and run the supplied integration workflow to write\n   exactly `/app/repair_report.json`:\n\n```bash\ncd /app\npython3 -m compileall -q vendor/evaluate/metrics vendor/evaluate/src\npython3 /app/run_smoke.py --cases /app/smoke_cases.json --out /app/repair_report.json\n```\n\nThe report must conform to `/app/output_schema.json` and must be computed by\nexecuting the repaired adapters. The verifier independently runs additional\nbinary, multiclass, multilabel, label-order, sample-weight, scalar-type,\none-label-vector, and zero-division regressions.\n"}
{"task_id": "candidate-2916-ml-training", "source_id": "candidate-2916-ml-training", "domain": "terminal", "task_path": "tasks/candidate-2916-ml-training", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:16c35e1e16c7aa93593bb9749ee5e66c0b63871b9fa705c3a01a29dd8f8c4f69", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair the online AdaBoost training path\n\nThe vendored River source under `/app/vendor/project` contains a regression in its online AdaBoost classifier. Repair the existing implementation and run the frozen streaming-training workflow.\n\n## Required source repairs\n\nModify both of these existing modules:\n\n- `/app/vendor/project/river/ensemble/boosting.py`\n- `/app/vendor/project/river/utils/skmultiflow_utils.py`\n\nThe repaired behavior must satisfy all of the following:\n\n1. Derive each weak learner's error from its wrong prediction mass divided by its total observed mass.\n2. Convert that error into the OzaBoost learner weight, using a neutral weight for untrained learners and learners whose error exceeds one half.\n3. Normalize every weak learner probability dictionary before weighting it, aggregate all learners rather than selecting one, and normalize the final distribution.\n4. Add a reusable dictionary-value scaling utility. Its `inplace=False` mode must return an independent copy and must not mutate the input.\n5. Preserve graceful normalization for zero or NaN factors unless the caller requests an exception.\n\nDo not replace the source tree, bypass the streaming parser, batch-fit the data, or modify the frozen dataset and workload support files.\n\n## Produce the artifact\n\nAfter repairing the source, run:\n\n```bash\ncd /app/vendor/project\npython3 run_training.py --output /app/training_report.json\n```\n\nThe required artifact is `/app/training_report.json` with schema version `river.adaboost.training_report.v1`.\n"}
{"task_id": "candidate-2951-security-appsec", "source_id": "candidate-2951-security-appsec", "domain": "terminal", "task_path": "tasks/candidate-2951-security-appsec", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:a77795af4a4cdde92dfa7e8ab1c13d7cbfe73ea8c064061ab009f14c31c0af75", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Repair Bandit's Trojan Source result propagation\n\nThe frozen Bandit tree under `/app/vendor/bandit` detects bidirectional Unicode\ncontrol characters with rule B613. Its file-level finding can be created with a\ncorrect line and column, but the generic test runner later replaces part of the\nproducer-owned location metadata. When the SARIF formatter serializes a finding\nfar from the beginning of a file, it can index the wrong source context and\nraise instead of writing a report.\n\nRepair the existing production source. You must modify both:\n\n- `/app/vendor/bandit/bandit/core/tester.py`\n- `/app/vendor/bandit/bandit/plugins/trojansource.py`\n\nRequirements:\n\n1. A plugin-provided non-empty line range must survive generic result\n   attribution. The generic tester must still supply its context line range\n   when a plugin leaves that field empty.\n2. B613 must explicitly own the line range for the detected suspicious\n   character, consistent with its reported line number.\n3. Scanning a Python file containing a bidirectional control character with\n   the real `python3 -m bandit -f sarif -t B613 ...` path must write valid SARIF\n   without a formatter traceback.\n4. The SARIF result must retain rule `B613`, HIGH severity, MEDIUM confidence,\n   CWE-838, and a physical region that points to the actual suspicious line and\n   character column under Bandit's existing column convention.\n5. Clean files must remain clean, declared Python source encodings must remain\n   supported, and unrelated scanner/formatter modules must not be rewritten or\n   removed.\n\nDo not solve this by clamping or suppressing locations in the SARIF formatter,\ndisabling B613, downgrading the finding, replacing the CLI, or hard-coding one\nfixture. No network access or package installation is available.\n"}
{"task_id": "candidate-2996-ml-evaluation", "source_id": "candidate-2996-ml-evaluation", "domain": "terminal", "task_path": "tasks/candidate-2996-ml-evaluation", "docker_image": "docker.io/xiaomimimo/mimo-v2.6-rl-oss@sha256:89ab8bd4c2c192bf8c1d1583a74dd20169011d976345b1ce8b024da7e9df4b17", "step_limit": 500, "agent_timeout_sec": 900.0, "verifier_timeout_sec": 240.0, "instruction": "You are an agent, your current working directory is /app.\n\nYou can use the tools available to you to interact with the computer to assist the user in completing tasks.\n\n# Task\n\nRepair the frozen Hugging Face Evaluate source tree in `/app/vendor/evaluate`.\nThe precision, recall, and F1 adapters currently assume every value returned by\ntheir scikit-learn scorer has an array-style `size` attribute. In a supported\nruntime, scalar classification results can instead be ordinary Python numeric\nvalues, causing otherwise valid binary and averaged evaluations to fail.\n\n## Required changes\n\nModify all three existing modules:\n\n- `/app/vendor/evaluate/metrics/precision/precision.py`\n- `/app/vendor/evaluate/metrics/recall/recall.py`\n- `/app/vendor/evaluate/metrics/f1/f1.py`\n\nThe repaired adapters must satisfy the following contract:\n\n1. A scalar scorer result is returned under the module's stable key\n   (`precision`, `recall`, or `f1`) as a Python `float`, including when the\n   scorer returns an ordinary Python scalar.\n2. A genuine multi-value result is preserved as a vector; do not coerce every\n   result to `float` and do not discard classes.\n3. Existing scikit-learn semantics must remain intact for `average`, `labels`,\n   `pos_label`, `sample_weight`, and the supported zero-division behavior.\n4. The three adapters must handle the scalar/vector boundary consistently.\n5. Keep the repair inside the existing metric modules. Do not replace the\n   scorers, hard-code examples, edit repository support files, or add a new\n   evaluator that bypasses these adapters.\n\n## Verification workflow\n\nInspect the existing implementation, reproduce the failure, repair all three\nmodules, and run:\n\n```bash\npython3 /app/tools/check_metric_adapters.py\npython3 -m compileall -q /app/vendor/evaluate/metrics/precision/precision.py \\\n  /app/vendor/evaluate/metrics/recall/recall.py \\\n  /app/vendor/evaluate/metrics/f1/f1.py\n```\n\nThe private verifier uses additional binary, multiclass, selected-label,\nweighted, absent-class, and scalar-return cases. It also checks that the frozen\nsupporting source tree has not been replaced or weakened.\n"}