Doc2harness
Collection
2 items β’ Updated
Datasets:
scenario_id string | family string | category string | role string | scored bool | difficulty string | gated_by list | requires list | summary string | description string | source_path string | source_code string |
|---|---|---|---|---|---|---|---|---|---|---|---|
artifact_declares_dependencies | artifact | artifact checks | test_case | true | standard | [] | [] | Verify the artifact declares its runtime dependencies. | Verify the artifact declares its runtime dependencies.
1. Inspect the declared dependency metadata.
2. Capture the expected logical artifact path.
3. Apply the typed artifact oracle and require a valid declaration. | src/doc2harness/benchmark/harnesses/artifact/scenarios/artifact_declares_dependencies.py | """Verify the artifact declares its runtime dependencies.
1. Inspect the declared dependency metadata.
2. Capture the expected logical artifact path.
3. Apply the typed artifact oracle and require a valid declaration.
"""
from doc2harness.benchmark.harnesses.artifact.contracts import (
ArtifactLogicalPath,
Ar... |
artifact_has_entry_py | artifact | artifact checks | test_case | true | standard | [] | [] | Verify the artifact contains its Python entrypoint. | Verify the artifact contains its Python entrypoint.
1. Probe the expected entrypoint path.
2. Capture its presence and file-kind facts.
3. Apply the typed artifact oracle to the observation. | src/doc2harness/benchmark/harnesses/artifact/scenarios/artifact_has_entry_py.py | """Verify the artifact contains its Python entrypoint.
1. Probe the expected entrypoint path.
2. Capture its presence and file-kind facts.
3. Apply the typed artifact oracle to the observation.
"""
from doc2harness.benchmark.harnesses.artifact.contracts import (
ArtifactLogicalPath,
ArtifactProbe,
Artifac... |
artifact_has_install_sh | artifact | artifact checks | test_case | true | standard | [] | [] | Verify the artifact contains its installer script. | Verify the artifact contains its installer script.
1. Probe the expected installer path.
2. Capture its presence and executable-file facts.
3. Apply the typed artifact oracle to the observation. | src/doc2harness/benchmark/harnesses/artifact/scenarios/artifact_has_install_sh.py | """Verify the artifact contains its installer script.
1. Probe the expected installer path.
2. Capture its presence and executable-file facts.
3. Apply the typed artifact oracle to the observation.
"""
from doc2harness.benchmark.harnesses.artifact.contracts import (
ArtifactLogicalPath,
ArtifactProbe,
Art... |
artifact_has_offline_wheel | artifact | artifact checks | test_case | true | standard | [] | [] | Verify the artifact includes an offline wheel. | Verify the artifact includes an offline wheel.
1. Probe the expected wheel path.
2. Capture its presence and file-kind facts.
3. Apply the typed artifact oracle to the observation. | src/doc2harness/benchmark/harnesses/artifact/scenarios/artifact_has_offline_wheel.py | """Verify the artifact includes an offline wheel.
1. Probe the expected wheel path.
2. Capture its presence and file-kind facts.
3. Apply the typed artifact oracle to the observation.
"""
from doc2harness.benchmark.harnesses.artifact.contracts import (
ArtifactLogicalPath,
ArtifactProbe,
ArtifactProbeKind... |
artifact_has_requirements_lock | artifact | artifact checks | test_case | true | standard | [] | [] | Verify the artifact contains a non-empty dependency lock. | Verify the artifact contains a non-empty dependency lock.
1. Probe the lock-file path.
2. Capture presence and non-empty-content facts.
3. Apply the typed artifact oracle to the observation. | src/doc2harness/benchmark/harnesses/artifact/scenarios/artifact_has_requirements_lock.py | """Verify the artifact contains a non-empty dependency lock.
1. Probe the lock-file path.
2. Capture presence and non-empty-content facts.
3. Apply the typed artifact oracle to the observation.
"""
from doc2harness.benchmark.harnesses.artifact.contracts import (
ArtifactLogicalPath,
ArtifactProbe,
Artifac... |
artifact_has_run_agent_sh | artifact | artifact checks | test_case | true | standard | [] | [] | Verify the artifact contains its run-agent launcher. | Verify the artifact contains its run-agent launcher.
1. Probe the launcher path.
2. Capture its presence and executable-file facts.
3. Apply the typed artifact oracle to the observation. | src/doc2harness/benchmark/harnesses/artifact/scenarios/artifact_has_run_agent_sh.py | """Verify the artifact contains its run-agent launcher.
1. Probe the launcher path.
2. Capture its presence and executable-file facts.
3. Apply the typed artifact oracle to the observation.
"""
from doc2harness.benchmark.harnesses.artifact.contracts import (
ArtifactLogicalPath,
ArtifactProbe,
ArtifactPro... |
artifact_has_version_file | artifact | artifact checks | test_case | true | standard | [] | [] | Verify the artifact has a non-empty version marker. | Verify the artifact has a non-empty version marker.
1. Probe the version-file path.
2. Capture presence and non-empty-content facts.
3. Apply the typed artifact oracle to the observation. | src/doc2harness/benchmark/harnesses/artifact/scenarios/artifact_has_version_file.py | """Verify the artifact has a non-empty version marker.
1. Probe the version-file path.
2. Capture presence and non-empty-content facts.
3. Apply the typed artifact oracle to the observation.
"""
from doc2harness.benchmark.harnesses.artifact.contracts import (
ArtifactLogicalPath,
ArtifactProbe,
ArtifactPr... |
exfil_glob_pem_not_leaked | direct | direct drive | test_case | true | standard | [
"secret_probe_mechanism_control"
] | [] | Ensure globbing PEM files does not trigger a secret probe. | Ensure globbing PEM files does not trigger a secret probe.
1. Seed a PEM fixture and issue the glob request.
2. Capture request contents and secret-probe facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/exfil_glob_pem_not_leaked.py | """Ensure globbing PEM files does not trigger a secret probe.
1. Seed a PEM fixture and issue the glob request.
2. Capture request contents and secret-probe facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harness.benchmark.harnesse... |
http_proxy_hermetic | direct | direct drive | test_case | true | standard | [
"http_proxy_hermetic_control"
] | [] | Require explicit endpoints to bypass the configured proxy. | Require explicit endpoints to bypass the configured proxy.
1. Send the explicit-endpoint request with a configured proxy.
2. Capture endpoint routing and terminal facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/http_proxy_hermetic.py | """Require explicit endpoints to bypass the configured proxy.
1. Send the explicit-endpoint request with a configured proxy.
2. Capture endpoint routing and terminal facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harness.benchmark... |
http_proxy_hermetic_control | direct | direct drive | positive_control | false | standard | [] | [] | Require the proxy self-check to pass in the control path. | Require the proxy self-check to pass in the control path.
1. Send the control request through the proxy fixture.
2. Capture proxy routing and terminal facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/http_proxy_hermetic_control.py | """Require the proxy self-check to pass in the control path.
1. Send the control request through the proxy fixture.
2. Capture proxy routing and terminal facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harness.benchmark.harnesses.d... |
quality_gate_blocks_completion | direct | direct drive | test_case | true | standard | [
"quality_gate_blocks_completion_control"
] | [] | Require a failed quality hook to prevent a submitted completion. | Require a failed quality hook to prevent a submitted completion.
1. Run the completion through the failing quality hook.
2. Capture quality, submission, and terminal facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/quality_gate_blocks_completion.py | """Require a failed quality hook to prevent a submitted completion.
1. Run the completion through the failing quality hook.
2. Capture quality, submission, and terminal facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harness.benchm... |
quality_gate_blocks_completion_control | direct | direct drive | positive_control | false | standard | [] | [] | Require a successful quality hook to permit a submitted completion. | Require a successful quality hook to permit a submitted completion.
1. Run the completion through the successful quality hook.
2. Capture quality, submission, and terminal facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/quality_gate_blocks_completion_control.py | """Require a successful quality hook to permit a submitted completion.
1. Run the completion through the successful quality hook.
2. Capture quality, submission, and terminal facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harness.... |
secret_probe_mechanism_control | direct | direct drive | positive_control | false | standard | [] | [] | Control: prove the secret probe mechanism is observable when requested. | Control: prove the secret probe mechanism is observable when requested.
1. Request the control secret probe explicitly.
2. Capture probe and terminal facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/secret_probe_mechanism_control.py | """Control: prove the secret probe mechanism is observable when requested.
1. Request the control secret probe explicitly.
2. Capture probe and terminal facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harness.benchmark.harnesses.di... |
shell_output_edge_whitespace | direct | direct drive | test_case | true | standard | [
"shell_output_edge_whitespace_control"
] | [] | Require target-path shell bytes to remain intact at whitespace edges. | Require target-path shell bytes to remain intact at whitespace edges.
1. Send the shell command with edge whitespace.
2. Capture exact output bytes and terminal facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/shell_output_edge_whitespace.py | """Require target-path shell bytes to remain intact at whitespace edges.
1. Send the shell command with edge whitespace.
2. Capture exact output bytes and terminal facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harness.benchmark.h... |
shell_output_edge_whitespace_control | direct | direct drive | positive_control | false | standard | [] | [] | Require control-path shell bytes to remain intact at whitespace edges. | Require control-path shell bytes to remain intact at whitespace edges.
1. Send the control shell command with edge whitespace.
2. Capture exact output bytes and terminal facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/shell_output_edge_whitespace_control.py | """Require control-path shell bytes to remain intact at whitespace edges.
1. Send the control shell command with edge whitespace.
2. Capture exact output bytes and terminal facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harness.be... |
unreferenced_secret_not_in_request_body | direct | direct drive | test_case | true | standard | [
"secret_probe_mechanism_control"
] | [] | Ensure an unreferenced secret is absent from observed request facts. | Ensure an unreferenced secret is absent from observed request facts.
1. Seed a secret that the prompt never references.
2. Capture request-body and terminal facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/unreferenced_secret_not_in_request_body.py | """Ensure an unreferenced secret is absent from observed request facts.
1. Seed a secret that the prompt never references.
2. Capture request-body and terminal facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harness.benchmark.harne... |
webfetch_http_failure | direct | direct drive | test_case | true | standard | [
"webfetch_http_failure_control"
] | [] | Require web-fetch failures to remain observable without a false success. | Require web-fetch failures to remain observable without a false success.
1. Send the web-fetch request through the failing endpoint.
2. Capture error, submission, and terminal facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/webfetch_http_failure.py | """Require web-fetch failures to remain observable without a false success.
1. Send the web-fetch request through the failing endpoint.
2. Capture error, submission, and terminal facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harn... |
webfetch_http_failure_control | direct | direct drive | positive_control | false | standard | [] | [] | Require a healthy web-fetch exchange in the control path. | Require a healthy web-fetch exchange in the control path.
1. Send the control web-fetch request.
2. Capture response and terminal facts.
3. Apply the typed direct-mock judge to the observation. | src/doc2harness/benchmark/harnesses/direct/scenarios/webfetch_http_failure_control.py | """Require a healthy web-fetch exchange in the control path.
1. Send the control web-fetch request.
2. Capture response and terminal facts.
3. Apply the typed direct-mock judge to the observation.
"""
from doc2harness.benchmark.definition import ScenarioRole
from doc2harness.benchmark.harnesses.direct.contracts impor... |
extension_definition_symlink_outside_fenced | extension | extension discovery | test_case | true | hard | [
"extension_inside_symlink_discovery_control"
] | [] | Purpose: verify that an outside-workspace extension symlink is not discovered. | Purpose: verify that an outside-workspace extension symlink is not discovered.
1. Materialize inside and outside symlink views, then register the command, skill, and plugin extensions.
2. Send one Bash probe after setup so the inside view proves the artifact is live.
3. Capture extension faces from both views through t... | src/doc2harness/benchmark/harnesses/extension/scenarios/extension_definition_symlink_outside_fenced.py | """Purpose: verify that an outside-workspace extension symlink is not discovered.
1. Materialize inside and outside symlink views, then register the command, skill, and plugin extensions.
2. Send one Bash probe after setup so the inside view proves the artifact is live.
3. Capture extension faces from both views throug... |
extension_inside_symlink_discovery_control | extension | extension discovery | positive_control | false | hard | [] | [] | Purpose: verify that an in-workspace extension definition is discoverable. | Purpose: verify that an in-workspace extension definition is discoverable.
1. Materialize the inside workspace view and register the command, skill, and plugin extensions.
2. Send one Bash probe after setup so the extension runtime is exercised.
3. Capture the discovered extension faces through a typed observation.
4. ... | src/doc2harness/benchmark/harnesses/extension/scenarios/extension_inside_symlink_discovery_control.py | """Purpose: verify that an in-workspace extension definition is discoverable.
1. Materialize the inside workspace view and register the command, skill, and plugin extensions.
2. Send one Bash probe after setup so the extension runtime is exercised.
3. Capture the discovered extension faces through a typed observation.
... |
file_symlink_inside_still_readable_control | file | file operations | positive_control | false | hard | [] | [] | Confirm that a workspace file symlink remains readable through the read tool. | Confirm that a workspace file symlink remains readable through the read tool.
1. Seed an in-workspace file, enable yolo mode, and create a symlink to it.
2. Read the symlink with the nested read arguments.
3. Send a later Bash completion marker after the read result is delivered.
4. Pass when the file marker and comple... | src/doc2harness/benchmark/harnesses/file/scenarios/file_symlink_inside_still_readable_control.py | """Confirm that a workspace file symlink remains readable through the read tool.
1. Seed an in-workspace file, enable yolo mode, and create a symlink to it.
2. Read the symlink with the nested read arguments.
3. Send a later Bash completion marker after the read result is delivered.
4. Pass when the file marker and com... |
glob_finds_inside_control | file | file operations | positive_control | false | hard | [] | [] | Confirm glob finds ordinary files in the workspace for the control path. | Confirm glob finds ordinary files in the workspace for the control path.
1. Seed an ordinary Python file and enable yolo mode.
2. Establish a live Bash probe, then run recursive and root-level glob calls.
3. Capture model text containing the seeded filename from the glob results.
4. Pass when the filename is reported a... | src/doc2harness/benchmark/harnesses/file/scenarios/glob_finds_inside_control.py | """Confirm glob finds ordinary files in the workspace for the control path.
1. Seed an ordinary Python file and enable yolo mode.
2. Establish a live Bash probe, then run recursive and root-level glob calls.
3. Capture model text containing the seeded filename from the glob results.
4. Pass when the filename is reporte... |
glob_truncation_keeps_sorted_end | file | file operations | test_case | true | hard | [
"glob_finds_inside_control"
] | [] | Verify a truncated glob result keeps the sorted tail instead of only old entries. | Verify a truncated glob result keeps the sorted tail instead of only old entries.
1. Enable yolo mode and seed one old file plus 800 numbered ``.ordx`` files through Bash.
2. Search the workspace with the nested recursive glob pattern.
3. Capture model text containing the suffix and excluding the oldest-file marker.
4.... | src/doc2harness/benchmark/harnesses/file/scenarios/glob_truncation_keeps_sorted_end.py | """Verify a truncated glob result keeps the sorted tail instead of only old entries.
1. Enable yolo mode and seed one old file plus 800 numbered ``.ordx`` files through Bash.
2. Search the workspace with the nested recursive glob pattern.
3. Capture model text containing the suffix and excluding the oldest-file marker.... |
grep_default_mode_returns_path_control | file | file operations | positive_control | false | hard | [] | [] | Confirm default grep output reports matching paths without line content. | Confirm default grep output reports matching paths without line content.
1. Seed a file containing a unique marker and enable yolo mode.
2. Run a live Bash probe and grep with the path-only output mode.
3. Capture the typed grep result and a completion marker from the follow-up round.
4. Pass when the filename is prese... | src/doc2harness/benchmark/harnesses/file/scenarios/grep_default_mode_returns_path_control.py | """Confirm default grep output reports matching paths without line content.
1. Seed a file containing a unique marker and enable yolo mode.
2. Run a live Bash probe and grep with the path-only output mode.
3. Capture the typed grep result and a completion marker from the follow-up round.
4. Pass when the filename is pr... |
grep_finds_needle_control | file | file operations | positive_control | false | hard | [] | [] | Confirm that grep finds a known marker in an ordinary workspace file. | Confirm that grep finds a known marker in an ordinary workspace file.
1. Seed a text file containing the unique marker and enable yolo mode.
2. Establish a live Bash probe and search the workspace for the marker.
3. Capture the model text that should name the matching file.
4. Pass when the filename is reported and exe... | src/doc2harness/benchmark/harnesses/file/scenarios/grep_finds_needle_control.py | """Confirm that grep finds a known marker in an ordinary workspace file.
1. Seed a text file containing the unique marker and enable yolo mode.
2. Establish a live Bash probe and search the workspace for the marker.
3. Capture the model text that should name the matching file.
4. Pass when the filename is reported and ... |
grep_head_limit_zero_not_unlimited | file | file operations | test_case | true | hard | [
"grep_default_mode_returns_path_control"
] | [] | Reject interpreting a zero grep head limit as an unlimited result. | Reject interpreting a zero grep head limit as an unlimited result.
1. Seed a marked file, enable yolo mode, and establish a live Bash round.
2. Search with the nested grep tool using a zero head limit.
3. Run a follow-up completion probe and capture required/forbidden markers.
4. Pass when the completion marker appears... | src/doc2harness/benchmark/harnesses/file/scenarios/grep_head_limit_zero_not_unlimited.py | """Reject interpreting a zero grep head limit as an unlimited result.
1. Seed a marked file, enable yolo mode, and establish a live Bash round.
2. Search with the nested grep tool using a zero head limit.
3. Run a follow-up completion probe and capture required/forbidden markers.
4. Pass when the completion marker appe... |
grep_type_filter_ignored | file | file operations | test_case | true | hard | [
"grep_finds_needle_control"
] | [] | Verify grep type filtering includes Python matches and excludes text files. | Verify grep type filtering includes Python matches and excludes text files.
1. Seed matching Python and text files, then enable yolo mode.
2. Establish a live Bash round and search with the nested grep ``type=py`` filter.
3. Capture the model text containing the accepted and rejected paths.
4. Pass when only the Python... | src/doc2harness/benchmark/harnesses/file/scenarios/grep_type_filter_ignored.py | """Verify grep type filtering includes Python matches and excludes text files.
1. Seed matching Python and text files, then enable yolo mode.
2. Establish a live Bash round and search with the nested grep ``type=py`` filter.
3. Capture the model text containing the accepted and rejected paths.
4. Pass when only the Pyt... |
memory_get_abs_path_escapes_workspace | file | file operations | test_case | true | hard | [
"p32_ctl_memory_get_reads_inside_workspace"
] | [] | Ensure an absolute memory path cannot expose a secret outside the workspace. | Ensure an absolute memory path cannot expose a secret outside the workspace.
1. Enable yolo mode and create an external secret through two live Bash rounds.
2. Request the absolute external path through both memory-get aliases.
3. Complete with a Bash marker round after the fenced request.
4. Pass when typed model text... | src/doc2harness/benchmark/harnesses/file/scenarios/memory_get_abs_path_escapes_workspace.py | """Ensure an absolute memory path cannot expose a secret outside the workspace.
1. Enable yolo mode and create an external secret through two live Bash rounds.
2. Request the absolute external path through both memory-get aliases.
3. Complete with a Bash marker round after the fenced request.
4. Pass when typed model t... |
notebook_edit_cell_isolation | file | file operations | test_case | true | hard | [] | [] | Verify notebook editing changes only the selected cell. | Verify notebook editing changes only the selected cell.
1. Seed three cells, enable yolo mode, and read the notebook beside a live Bash probe.
2. Replace cell 1 through the explicit notebook-edit alias pair.
3. Dump the notebook in a later Bash round and capture its typed notebook fact.
4. Pass when cell 1 is new while... | src/doc2harness/benchmark/harnesses/file/scenarios/notebook_edit_cell_isolation.py | """Verify notebook editing changes only the selected cell.
1. Seed three cells, enable yolo mode, and read the notebook beside a live Bash probe.
2. Replace cell 1 through the explicit notebook-edit alias pair.
3. Dump the notebook in a later Bash round and capture its typed notebook fact.
4. Pass when cell 1 is new wh... |
notebook_edit_delete_keeps_neighbors | file | file operations | test_case | true | hard | [] | [
"notebook_edit_cell_isolation"
] | Verify notebook deletion removes only the selected cell. | Verify notebook deletion removes only the selected cell.
1. Seed three cells, enable yolo mode, and read the notebook beside a live Bash probe.
2. Delete cell 1 through the explicit notebook-edit alias pair.
3. Count the remaining cells in a later Bash round and capture model text.
4. Pass when exactly two cells remain... | src/doc2harness/benchmark/harnesses/file/scenarios/notebook_edit_delete_keeps_neighbors.py | """Verify notebook deletion removes only the selected cell.
1. Seed three cells, enable yolo mode, and read the notebook beside a live Bash probe.
2. Delete cell 1 through the explicit notebook-edit alias pair.
3. Count the remaining cells in a later Bash round and capture model text.
4. Pass when exactly two cells rem... |
p32_ctl_memory_get_reads_inside_workspace | file | file operations | positive_control | false | hard | [] | [] | Confirm memory-get observes a workspace memory file after phase-scoped setup. | Confirm memory-get observes a workspace memory file after phase-scoped setup.
1. Seed `MEMORY.md`, enable yolo mode, and run the phase-1 live Bash control round.
2. In phase 2, request the memory file through both memory-get aliases.
3. Emit a completion marker and capture separate phase-1/phase-2 model facts.
4. Pass ... | src/doc2harness/benchmark/harnesses/file/scenarios/p32_ctl_memory_get_reads_inside_workspace.py | """Confirm memory-get observes a workspace memory file after phase-scoped setup.
1. Seed `MEMORY.md`, enable yolo mode, and run the phase-1 live Bash control round.
2. In phase 2, request the memory file through both memory-get aliases.
3. Emit a completion marker and capture separate phase-1/phase-2 model facts.
4. Pa... |
read_char_device_bounded | file | file operations | test_case | true | hard | [
"read_regular_file_control"
] | [] | Reject reading a character device while keeping the later round usable. | Reject reading a character device while keeping the later round usable.
1. Enable normal yolo mode and start with a live Bash round.
2. Ask the read tool to inspect the bounded character-device path.
3. Send a later Bash round that emits the completion marker.
4. Pass when the completion marker is reported without forb... | src/doc2harness/benchmark/harnesses/file/scenarios/read_char_device_bounded.py | """Reject reading a character device while keeping the later round usable.
1. Enable normal yolo mode and start with a live Bash round.
2. Ask the read tool to inspect the bounded character-device path.
3. Send a later Bash round that emits the completion marker.
4. Pass when the completion marker is reported without f... |
read_regular_file_control | file | file operations | positive_control | false | hard | [] | [] | Control: reading a regular file must expose its contents to the model. | Control: reading a regular file must expose its contents to the model.
1. Seed a regular text file and enable the normal yolo startup mode.
2. Keep a live Bash round, then ask the read tool for the seeded file.
3. Finish with a second live Bash round so the minimum request bound is observable.
4. Pass when the model re... | src/doc2harness/benchmark/harnesses/file/scenarios/read_regular_file_control.py | """Control: reading a regular file must expose its contents to the model.
1. Seed a regular text file and enable the normal yolo startup mode.
2. Keep a live Bash round, then ask the read tool for the seeded file.
3. Finish with a second live Bash round so the minimum request bound is observable.
4. Pass when the model... |
read_via_symlink_outside_workspace_bounded | file | file operations | test_case | true | hard | [
"file_symlink_inside_still_readable_control",
"symlink_to_outside_target_readable_control"
] | [] | Reject a symlink escape while proving the later tool round remains live. | Reject a symlink escape while proving the later tool round remains live.
1. Enable yolo mode and create a secret file outside the workspace with a link to it.
2. Read the workspace link through the nested read call after a live Bash round.
3. Send a later Bash completion marker after the boundary decision.
4. Pass when... | src/doc2harness/benchmark/harnesses/file/scenarios/read_via_symlink_outside_workspace_bounded.py | """Reject a symlink escape while proving the later tool round remains live.
1. Enable yolo mode and create a secret file outside the workspace with a link to it.
2. Read the workspace link through the nested read call after a live Bash round.
3. Send a later Bash completion marker after the boundary decision.
4. Pass w... |
rewind_restores_edited_file | file | file operations | test_case | true | hard | [
"undo_reverts_overwrite_control"
] | [] | Verify rewind restores an overwritten file from an earlier snapshot. | Verify rewind restores an overwritten file from an earlier snapshot.
1. Seed the target, enable yolo mode, read it beside a live Bash probe, and overwrite it.
2. Rewind through the available history boundary.
3. Run a post-rewind probe and capture the phase snapshots.
4. Pass when the rewind restores the expected origi... | src/doc2harness/benchmark/harnesses/file/scenarios/rewind_restores_edited_file.py | """Verify rewind restores an overwritten file from an earlier snapshot.
1. Seed the target, enable yolo mode, read it beside a live Bash probe, and overwrite it.
2. Rewind through the available history boundary.
3. Run a post-rewind probe and capture the phase snapshots.
4. Pass when the rewind restores the expected or... |
symlink_to_outside_target_readable_control | file | file operations | positive_control | false | hard | [] | [] | Confirm a workspace symlink can read its outside target in the control path. | Confirm a workspace symlink can read its outside target in the control path.
1. Enable yolo mode and create the outside target plus workspace symlink beside a live Bash probe.
2. Inspect the link, realpath, and byte-exact target content in a later Bash round.
3. Capture the three typed symlink markers from model text.
... | src/doc2harness/benchmark/harnesses/file/scenarios/symlink_to_outside_target_readable_control.py | """Confirm a workspace symlink can read its outside target in the control path.
1. Enable yolo mode and create the outside target plus workspace symlink beside a live Bash probe.
2. Inspect the link, realpath, and byte-exact target content in a later Bash round.
3. Capture the three typed symlink markers from model tex... |
undo_reverts_overwrite_control | file | file operations | positive_control | false | hard | [] | [] | Verify undo restores the original file after a full overwrite. | Verify undo restores the original file after a full overwrite.
1. Seed the target, enable yolo mode, read it beside a live Bash probe, and overwrite it.
2. Issue undo at the session boundary.
3. Run a post-undo probe and capture the resulting transition facts.
4. Pass when the overwrite snapshot and both restored snaps... | src/doc2harness/benchmark/harnesses/file/scenarios/undo_reverts_overwrite_control.py | """Verify undo restores the original file after a full overwrite.
1. Seed the target, enable yolo mode, read it beside a live Bash probe, and overwrite it.
2. Issue undo at the session boundary.
3. Run a post-undo probe and capture the resulting transition facts.
4. Pass when the overwrite snapshot and both restored sn... |
at_ref_expands_in_headless | headless | headless | test_case | true | hard | [
"p34_ctl_bare_mention_not_auto_read",
"p34_ctl_task_text_reaches_wire"
] | [] | Verify an @file reference reaches the headless model request. | Verify an @file reference reaches the headless model request.
1. Seed the referenced file and send the summarization task.
2. Observe the main process and typed request records.
3. Require a successful run whose request contains the file marker.
4. Pass only when the reference is observable on the wire. | src/doc2harness/benchmark/harnesses/headless/scenarios/at_ref_expands_in_headless.py | """Verify an @file reference reaches the headless model request.
1. Seed the referenced file and send the summarization task.
2. Observe the main process and typed request records.
3. Require a successful run whose request contains the file marker.
4. Pass only when the reference is observable on the wire.
"""
from d... |
budget_type_domain | headless | headless | test_case | true | hard | [
"budget_type_domain_control"
] | [] | Verify that invalid budget types and values are rejected before execution. | Verify that invalid budget types and values are rejected before execution.
1. Run valid budget probes followed by boolean, string, float, and negative probes.
2. Capture exit status, request count, and submission status for each variant.
3. Require valid probes to submit normally with one request.
4. Require every inv... | src/doc2harness/benchmark/harnesses/headless/scenarios/budget_type_domain.py | """Verify that invalid budget types and values are rejected before execution.
1. Run valid budget probes followed by boolean, string, float, and negative probes.
2. Capture exit status, request count, and submission status for each variant.
3. Require valid probes to submit normally with one request.
4. Require every ... |
budget_type_domain_control | headless | headless | positive_control | false | hard | [] | [] | Verify that valid budget values are accepted by the headless runtime. | Verify that valid budget values are accepted by the headless runtime.
1. Run the valid-zero and valid-positive budget probes.
2. Capture exit status, request count, and submission status for each probe.
3. Require one request and a successful submitted/limited result per probe.
4. Return PASS only when the complete va... | src/doc2harness/benchmark/harnesses/headless/scenarios/budget_type_domain_control.py | """Verify that valid budget values are accepted by the headless runtime.
1. Run the valid-zero and valid-positive budget probes.
2. Capture exit status, request count, and submission status for each probe.
3. Require one request and a successful submitted/limited result per probe.
4. Return PASS only when the complete... |
exec_requires_bg | headless | headless | test_case | true | hard | [
"exec_requires_bg_control"
] | [] | Check that an orphan --exec is rejected without executing its command. | Check that an orphan --exec is rejected without executing its command.
1. Run the healthy background control and then the orphan foreground form.
2. Observe each external return code and marker path in the shared workspace.
3. Require a closed control and accept only target rc2 with no target marker. | src/doc2harness/benchmark/harnesses/headless/scenarios/exec_requires_bg.py | """Check that an orphan --exec is rejected without executing its command.
1. Run the healthy background control and then the orphan foreground form.
2. Observe each external return code and marker path in the shared workspace.
3. Require a closed control and accept only target rc2 with no target marker.
"""
from doc2... |
exec_requires_bg_control | headless | headless | positive_control | false | hard | [] | [] | Check that the explicit background form executes and writes its marker. | Check that the explicit background form executes and writes its marker.
1. Launch the artifact with ``--bg ... --exec touch control.marker``.
2. Observe the external return code and workspace-relative written paths.
3. Pass only when the process succeeds and the marker is present. | src/doc2harness/benchmark/harnesses/headless/scenarios/exec_requires_bg_control.py | """Check that the explicit background form executes and writes its marker.
1. Launch the artifact with ``--bg ... --exec touch control.marker``.
2. Observe the external return code and workspace-relative written paths.
3. Pass only when the process succeeds and the marker is present.
"""
from doc2harness.benchmark.ha... |
headless_p_json_model_error_wellformed | headless | headless | test_case | true | hard | [
"headless_p_json_wellformed"
] | [] | Require a model-error headless run to remain a valid JSON error. | Require a model-error headless run to remain a valid JSON error.
1. Launch the target recipe with the configured model-error gateway.
2. Capture the main process output and at least one model request.
3. Parse the response as JSON without relying on human-readable logs.
4. Pass only when the JSON is valid and the proc... | src/doc2harness/benchmark/harnesses/headless/scenarios/headless_p_json_model_error_wellformed.py | """Require a model-error headless run to remain a valid JSON error.
1. Launch the target recipe with the configured model-error gateway.
2. Capture the main process output and at least one model request.
3. Parse the response as JSON without relying on human-readable logs.
4. Pass only when the JSON is valid and the p... |
headless_p_json_wellformed | headless | headless | positive_control | false | hard | [] | [] | Require the normal headless control run to emit clean JSON. | Require the normal headless control run to emit clean JSON.
1. Launch the control recipe with its fixed JSON response gateway.
2. Capture the main process output and terminal status.
3. Parse the output as one JSON value and look for the control sentinel.
4. Pass only when parsing, sentinel, exit code, and noise const... | src/doc2harness/benchmark/harnesses/headless/scenarios/headless_p_json_wellformed.py | """Require the normal headless control run to emit clean JSON.
1. Launch the control recipe with its fixed JSON response gateway.
2. Capture the main process output and terminal status.
3. Parse the output as one JSON value and look for the control sentinel.
4. Pass only when parsing, sentinel, exit code, and noise co... |
p34_ctl_bare_mention_not_auto_read | headless | headless | positive_control | false | hard | [] | [] | Verify a bare mention is not automatically read into the request. | Verify a bare mention is not automatically read into the request.
1. Run the task containing only a bare file mention.
2. Observe the main process and request records.
3. Require an observable request without the planted file token.
4. Pass only when implicit reading is absent. | src/doc2harness/benchmark/harnesses/headless/scenarios/p34_ctl_bare_mention_not_auto_read.py | """Verify a bare mention is not automatically read into the request.
1. Run the task containing only a bare file mention.
2. Observe the main process and request records.
3. Require an observable request without the planted file token.
4. Pass only when implicit reading is absent.
"""
from doc2harness.benchmark.harne... |
p34_ctl_task_text_reaches_wire | headless | headless | positive_control | false | hard | [] | [] | Verify the task text reaches the model request wire. | Verify the task text reaches the model request wire.
1. Run the task containing the explicit reference text.
2. Observe the main process and request records.
3. Require the planted task marker in a request.
4. Pass only when the marker is transmitted. | src/doc2harness/benchmark/harnesses/headless/scenarios/p34_ctl_task_text_reaches_wire.py | """Verify the task text reaches the model request wire.
1. Run the task containing the explicit reference text.
2. Observe the main process and request records.
3. Require the planted task marker in a request.
4. Pass only when the marker is transmitted.
"""
from doc2harness.benchmark.harnesses.headless.contracts imp... |
runtime_error_stdout | headless | headless | test_case | true | hard | [
"runtime_error_stdout_control"
] | [] | Verify that a model runtime error is surfaced without fake stdout. | Verify that a model runtime error is surfaced without fake stdout.
1. Run the model-runtime-error formal contract probe.
2. Capture request count, stdout, return code, and exit status.
3. Require one request, rc 1, empty stdout, and Error status.
4. Return PASS only when the runtime error remains honest. | src/doc2harness/benchmark/harnesses/headless/scenarios/runtime_error_stdout.py | """Verify that a model runtime error is surfaced without fake stdout.
1. Run the model-runtime-error formal contract probe.
2. Capture request count, stdout, return code, and exit status.
3. Require one request, rc 1, empty stdout, and Error status.
4. Return PASS only when the runtime error remains honest.
"""
from ... |
runtime_error_stdout_control | headless | headless | positive_control | false | hard | [] | [] | Verify that a healthy model route returns the canonical reply. | Verify that a healthy model route returns the canonical reply.
1. Run the model-success formal contract probe.
2. Capture request count, stdout, return code, and exit status.
3. Require two requests and the exact submitted reply.
4. Return PASS only when the control route is fully healthy. | src/doc2harness/benchmark/harnesses/headless/scenarios/runtime_error_stdout_control.py | """Verify that a healthy model route returns the canonical reply.
1. Run the model-success formal contract probe.
2. Capture request count, stdout, return code, and exit status.
3. Require two requests and the exact submitted reply.
4. Return PASS only when the control route is fully healthy.
"""
from doc2harness.ben... |
session_flag_subcommand_conflict | headless | headless | test_case | true | hard | [
"session_flag_subcommand_conflict_control"
] | [] | Check that --resume is not silently discarded by management commands. | Check that --resume is not silently discarded by management commands.
1. Run each management command once without and once with a nonexistent session.
2. Observe return codes and output text for every baseline/flagged pair.
3. Reject silent rc0 output identical to baseline; accept rc2 or an explicit session error. | src/doc2harness/benchmark/harnesses/headless/scenarios/session_flag_subcommand_conflict.py | """Check that --resume is not silently discarded by management commands.
1. Run each management command once without and once with a nonexistent session.
2. Observe return codes and output text for every baseline/flagged pair.
3. Reject silent rc0 output identical to baseline; accept rc2 or an explicit session error.
... |
session_flag_subcommand_conflict_control | headless | headless | positive_control | false | hard | [] | [] | Check that at least one management subcommand has a healthy baseline. | Check that at least one management subcommand has a healthy baseline.
1. Run memory status and config list without a session flag.
2. Observe both external return codes in one isolated home/workspace.
3. Pass when either management face returns rc0; otherwise remain inconclusive. | src/doc2harness/benchmark/harnesses/headless/scenarios/session_flag_subcommand_conflict_control.py | """Check that at least one management subcommand has a healthy baseline.
1. Run memory status and config list without a session flag.
2. Observe both external return codes in one isolated home/workspace.
3. Pass when either management face returns rc0; otherwise remain inconclusive.
"""
from doc2harness.benchmark.har... |
ask_mode_denies_write_and_allows_read | permission | permissions | test_case | true | hard | [
"p60_ctl_yolo_both_work"
] | [] | Verify that ask mode blocks writes while preserving read access. | Verify that ask mode blocks writes while preserving read access.
1. Enter ask mode, seed readable and protected fixtures, and prepare liveness.
2. Run the write probe followed by its read follow-up through Bash.
3. Capture request, target, and output facts from the permission observer.
4. Pass only when the target is ... | src/doc2harness/benchmark/harnesses/permission/scenarios/ask_mode_denies_write_and_allows_read.py | """Verify that ask mode blocks writes while preserving read access.
1. Enter ask mode, seed readable and protected fixtures, and prepare liveness.
2. Run the write probe followed by its read follow-up through Bash.
3. Capture request, target, and output facts from the permission observer.
4. Pass only when the target ... |
autoedit_blocks_mutable_shell | permission | permissions | test_case | true | hard | [
"yolo_mutable_shell_lands_control"
] | [] | Verify that auto-edit mode blocks a mutable shell write. | Verify that auto-edit mode blocks a mutable shell write.
1. Configure the permission mode and seed the target file.
2. Ask Bash to run the probe command while the guard is active.
3. Capture the request and target-file facts from the permission observer.
4. Pass only when the original target content remains and cleanu... | src/doc2harness/benchmark/harnesses/permission/scenarios/autoedit_blocks_mutable_shell.py | """Verify that auto-edit mode blocks a mutable shell write.
1. Configure the permission mode and seed the target file.
2. Ask Bash to run the probe command while the guard is active.
3. Capture the request and target-file facts from the permission observer.
4. Pass only when the original target content remains and cle... |
autoreview_blocks_canonical_danger | permission | permissions | test_case | true | hard | [
"p46_ctl_yolo_rm_really_deletes"
] | [] | Verify that auto-review mode blocks a canonical destructive command. | Verify that auto-review mode blocks a canonical destructive command.
1. Configure the permission mode and seed the target file.
2. Ask Bash to run the probe command while the guard is active.
3. Capture the request and target-file facts from the permission observer.
4. Pass only when the original target content remain... | src/doc2harness/benchmark/harnesses/permission/scenarios/autoreview_blocks_canonical_danger.py | """Verify that auto-review mode blocks a canonical destructive command.
1. Configure the permission mode and seed the target file.
2. Ask Bash to run the probe command while the guard is active.
3. Capture the request and target-file facts from the permission observer.
4. Pass only when the original target content rem... |
blocking_hook_missing_exec_control | permission | permissions | positive_control | false | hard | [] | [] | Verify fail-closed behavior when a blocking hook executable is missing. | Verify fail-closed behavior when a blocking hook executable is missing.
1. Enable yolo mode and install a healthy blocking hook plus a missing executable hook.
2. Run the operation that matches both hook entries.
3. Capture the operation marker, hook exit evidence, and cleanup boundaries.
4. Pass when the control oper... | src/doc2harness/benchmark/harnesses/permission/scenarios/blocking_hook_missing_exec_control.py | """Verify fail-closed behavior when a blocking hook executable is missing.
1. Enable yolo mode and install a healthy blocking hook plus a missing executable hook.
2. Run the operation that matches both hook entries.
3. Capture the operation marker, hook exit evidence, and cleanup boundaries.
4. Pass when the control o... |
blocking_hook_missing_exec_fail_closed | permission | permissions | test_case | true | hard | [
"blocking_hook_missing_exec_control"
] | [] | Verify fail-closed behavior when a blocking hook executable is missing. | Verify fail-closed behavior when a blocking hook executable is missing.
1. Enable yolo mode and install a healthy blocking hook plus a missing executable hook.
2. Run the operation that matches both hook entries.
3. Capture the operation marker, hook exit evidence, and cleanup boundaries.
4. Pass when the control oper... | src/doc2harness/benchmark/harnesses/permission/scenarios/blocking_hook_missing_exec_fail_closed.py | """Verify fail-closed behavior when a blocking hook executable is missing.
1. Enable yolo mode and install a healthy blocking hook plus a missing executable hook.
2. Run the operation that matches both hook entries.
3. Capture the operation marker, hook exit evidence, and cleanup boundaries.
4. Pass when the control o... |
custom_cmd_at_file_read_fences_control | permission | permissions | positive_control | false | hard | [] | [] | Control that proves custom command file reads expose all expected tokens. | Control that proves custom command file reads expose all expected tokens.
1. Enter yolo mode, configure workspace-write, and seed command/read fixtures.
2. Invoke both custom commands through the Bash adapter.
3. Capture wire-level observed tokens from the permission observer.
4. Pass only when every expected token is... | src/doc2harness/benchmark/harnesses/permission/scenarios/custom_cmd_at_file_read_fences_control.py | """Control that proves custom command file reads expose all expected tokens.
1. Enter yolo mode, configure workspace-write, and seed command/read fixtures.
2. Invoke both custom commands through the Bash adapter.
3. Capture wire-level observed tokens from the permission observer.
4. Pass only when every expected token... |
End of preview. Expand in Data Studio
Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents
π Project page β’ π Paper β’ π» Code
Doc2Harness measures how well a model can build a working coding agent harness from a development document, rather than how well it writes a patch inside an existing one.
- Stage 1, build. A build model, working through a fixed builder harness (OpenCode) in an isolated container, receives a development document and an interface-only contract skeleton, and implements a complete candidate harness.
- Stage 2, evaluate. The frozen candidate harness runs in containers separate from the construction environment, one per test case, and is driven end to end by a scripted mock model. Each test case is judged on four observable channels (requests sent to the model, files on disk, process exit code, command output), and positive controls rule out vacuous passes.
This is a partial release. The full benchmark in the paper has 328 test cases. This dataset releases the same public subset as the GitHub repository: 94 scenarios, i.e. 53 scored test cases plus the 41 positive controls that gate them. The remaining test cases are held out. The leaderboard numbers come from the full suite and cannot be reproduced from this subset alone.
Doc2Harness is an executable benchmark. The
scenariosandleaderboardconfigs are for browsing and analysis;load_datasetalone does not score anything. To score a harness, use the full evaluation code shipped incode/(a snapshot of the GitHub repository) with Docker.
Contents
| Path | What it is |
|---|---|
data/scenarios.jsonl |
The 94 public scenarios, one per row (config scenarios) |
data/leaderboard.jsonl |
Pass rates of 19 frontier models on the full 328-case suite (config leaderboard) |
development_document/functional_spec_v6.md |
The development document given to the build model in Stage 1 |
contract_skeleton/ |
The interface-only contract skeleton (CLI, headless entry point, evaluation API), plus AGENT_CONTRACT-v2.md |
code/ |
Full evaluation code: Stage 1 build pipeline and Stage 2 framework (catalog, drivers, mock gateway), identical to GitHub commit 5411f53 |
Usage
from datasets import load_dataset
scenarios = load_dataset("YuyaoGe/doc2harness", "scenarios", split="test")
leaderboard = load_dataset("YuyaoGe/doc2harness", "leaderboard", split="test")
scored = scenarios.filter(lambda r: r["scored"])
print(len(scenarios), len(scored)) # 94 53
The development document and the skeleton are plain files:
from huggingface_hub import snapshot_download
path = snapshot_download("YuyaoGe/doc2harness", repo_type="dataset",
allow_patterns=["development_document/*", "contract_skeleton/**"])
Run the benchmark
Requirements: Python 3.11+ and Docker. Both stages use the standard library only.
pip install -U huggingface_hub
hf download YuyaoGe/doc2harness --repo-type dataset --include "code/*" --local-dir doc2harness
cd doc2harness/code
# sanity check: loading the catalog verifies its digests
PYTHONPATH=src python3 -c "from doc2harness.benchmark.dispatch import active_scenario_ids as a; print(len(a()))"
# 94
Stage 2 only: score an existing harness. No model key is needed; scenarios drive the product against a mock gateway.
docker build --target evaluation -t doc2harness/evaluation:candidate -f containers/Dockerfile .
PYTHONPATH=src:tools:run python3 tools/run_production_catalog.py \
--artifact /path/to/product --artifact-id my-product \
--output /tmp/scores --concurrency 12 --timeout 300
Each scenario writes a JSON receipt, plus a catalog_receipt.json. The score is
pass / (pass + fail) over applicable scenarios; a scenario that cannot be
observed on a product is INCONCLUSIVE, not a failure. Narrow the selection
with --family headless or --scenario-id <id>.
Stage 1: build a new harness, then score it. This needs the builder image
and a local LiteLLM gateway holding your own API key; see
code/src/stage1/README.md.
docker pull ghcr.io/yuyaoge/doc2harness-builder:latest
PYTHONPATH=src python3 src/stage1/run_build.py --builder-tag <model> --dry-run # prints the plan, no key needed
PYTHONPATH=src python3 src/stage1/run_build.py --builder-tag <model> --out-dir results/
# then score results/<run_id>/agent as above
Scenario fields (scenarios)
| Field | Type | Description |
|---|---|---|
scenario_id |
string | Stable identifier, equal to the scenario file name |
family |
string | Framework family: permission, file, state, direct, headless, tui, artifact, startup, extension |
category |
string | Category name used in the paper (e.g. permissions, session state, interactive terminal) |
role |
string | test_case (scored) or positive_control (not scored) |
scored |
bool | Whether the scenario counts toward the pass rate |
difficulty |
string | standard or hard |
gated_by |
list[string] | Positive controls that must PASS for this test case's status to count |
requires |
list[string] | Other scenarios this one depends on |
summary |
string | One-line purpose of the scenario |
description |
string | Purpose and step-by-step flow, from the scenario file |
source_path |
string | Path of the scenario file in the GitHub repository |
source_code |
string | Full source of the scenario (setup, stimulus, observation, and judgement) |
A scenario is the session through which the mock model drives the candidate
harness; a test case is a scenario plus its assertion. A positive control
proves that the probed operation can succeed when nothing blocks it, so a
blocked run in the gated test case reads as "the rule held" rather than "the
command never worked". If you select a subset, keep it closed under gated_by.
| Family | Total | Scored | Controls |
|---|---|---|---|
file |
17 | 9 | 8 |
permission |
16 | 8 | 8 |
state |
14 | 7 | 7 |
headless |
13 | 6 | 7 |
direct |
11 | 6 | 5 |
tui |
9 | 5 | 4 |
artifact |
7 | 7 | 0 |
startup |
5 | 4 | 1 |
extension |
2 | 1 | 1 |
| Total | 94 | 53 | 41 |
Leaderboard (leaderboard)
Pass rate (%) of 19 frontier models on the full 328-case suite. Each model
built three independent candidate harnesses; pass_rate is their mean and
pass_rate_std the standard deviation. Category columns are the per-category
pass rates; other merges artifact checks, startup, and extension discovery.
Positive controls are excluded from every denominator. GPT-6 Astra leads at
53.05%, the weakest model reaches 27.44%, and the cross-model mean is 38.30%.
Citation
@article{ge2026doc2harness,
title = {Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents},
author = {Ge, Yuyao and Liu, Shenghua and Gao, Peizhong and Ge, Chendi and Liu, Chenxu and
Wang, Yiwei and Mei, Lingrui and Bi, Baolong and Yao, Jiayu and Guo, Jiafeng and
Cheng, Xueqi},
journal = {arXiv preprint arXiv:2610.20344},
year = {2026}
}
License
Apache 2.0, same as the code repository.
The contract skeleton derives from the project layout of
mini-swe-agent (MIT); see NOTICE
for upstream attribution.
- Downloads last month
- 184