mini-swe-agent-main · Business Workflow Map

Workflow area 1

Agent workflows

1 workflow

Workflow 1

agent orchestration and trajectory persistence / task execution loop and limit enforcement

What this workflow does

DefaultAgent renders prompts, queries the model, executes returned bash actions in the environment, and saves the resulting trajectory or exit state. / DefaultAgent processes model-driven shell actions, captures observations, and stops on wall-time, cost, or repeated format-error limits.

Purpose

solve a task by repeatedly querying an LLM, running the model-selected shell actions, and recording the session / carry out a task until submission or a stopping limit is reached

What it produces

agent trajectory and submission payload / agent run state, conversation history, and accumulated cost/step budgets

Who relies on it

local execution environment and downstream consumer of saved trajectory / model caller / task runner

Why this workflow matters

Produces a saved run record with model/environment metadata
Converts model actions into local command execution
Stops on limits, time, or task completion
produces a submitted final output when completion is reached
terminates cleanly on time, cost, or format-error budget conditions

Implementation references

DefaultAgent src/minisweagent/agents/default.py · lines 38–190
DefaultAgent.run() src/minisweagent/agents/default.py · lines 88–124
LocalEnvironment src/minisweagent/environments/local.py · lines 19–69
LocalEnvironment.execute() src/minisweagent/environments/local.py · lines 24–43
LocalEnvironment._check_finished() src/minisweagent/environments/local.py · lines 45–56
AgentConfig src/minisweagent/agents/default.py · lines 19–35
DefaultAgent.query() src/minisweagent/agents/default.py · lines 130–152
test_agent_save_includes_class_names() tests/run/test_save.py · lines 10–48
DefaultAgent._render_template() src/minisweagent/agents/default.py · lines 66–67
DefaultAgent.step() src/minisweagent/agents/default.py · lines 126–128
DefaultAgent.add_messages() src/minisweagent/agents/default.py · lines 69–72
DefaultAgent.execute_actions() src/minisweagent/agents/default.py · lines 154–157
DefaultAgent.save() src/minisweagent/agents/default.py · lines 182–190
DefaultAgent.__init__() src/minisweagent/agents/default.py · lines 39–50
DefaultAgent.get_template_vars() src/minisweagent/agents/default.py · lines 52–64
DefaultAgent.handle_uncaught_exception() src/minisweagent/agents/default.py · lines 74–86
DefaultAgent.serialize() src/minisweagent/agents/default.py · lines 159–180
OpenRouterModel src/minisweagent/models/openrouter_model.py · lines 55–169
format_toolcall_observation_messages() src/minisweagent/models/utils/actions_toolcall.py · lines 79–113
parse_toolcall_actions() src/minisweagent/models/utils/actions_toolcall.py · lines 30–76
InterruptAgentFlow.__init__() src/minisweagent/exceptions.py · lines 4–6
test_message_history_tracking() tests/agents/test_default.py · lines 334–362
test_wall_time_limit_enforcement() tests/agents/test_default.py · lines 410–426
test_format_errors_count_against_cost_limit() tests/agents/test_default.py · lines 530–544
test_repeated_format_errors_terminate_cleanly() tests/agents/test_default.py · lines 481–493
test_step_adds_messages() tests/agents/test_default.py · lines 365–385

Workflow area 2

Command workflows

2 workflows

Workflow 2

human-in-the-loop command execution / interactive task execution and mode switching

What this workflow does

InteractiveAgent lets a human switch modes, confirm or reject LM commands, and intercept submission or interruption before execution. / InteractiveAgent routes user confirmations and mode switches (confirm, yolo, human) while allowing continuation after completion.

Purpose

supervise or override LM-issued commands before they run / inspect, approve, reject, or directly issue commands before submission

What it produces

command list and task completion handoff / interactive agent mode, pending action, and submission continuation task

Who relies on it

interactive user and local environment / human operator using the agent

Why this workflow matters

Prevents undesired command execution unless confirmed
Allows user to switch between human, confirm, and yolo modes
Supports new-task handoff or clean submission when the agent finishes
supports safer confirmation before shell commands
permits yolo mode for fast execution

Implementation references

InteractiveAgent src/minisweagent/agents/interactive.py · lines 33–209
InteractiveAgent.execute_actions() src/minisweagent/agents/interactive.py · lines 124–139
InteractiveAgentConfig src/minisweagent/agents/interactive.py · lines 24–30
InteractiveAgent._prompt_and_handle_slash_commands() src/minisweagent/agents/interactive.py · lines 184–209
InteractiveAgent._check_for_new_task_or_submit() src/minisweagent/agents/interactive.py · lines 144–160
InteractiveAgent._ask_confirmation_or_interrupt() src/minisweagent/agents/interactive.py · lines 165–182
InteractiveAgent._stdin_is_interactive() src/minisweagent/agents/interactive.py · lines 96–107
InteractiveAgent._interrupt() src/minisweagent/agents/interactive.py · lines 40–41
InteractiveAgent._add_observation_messages() src/minisweagent/agents/interactive.py · lines 141–142
InteractiveAgent._should_ask_confirmation() src/minisweagent/agents/interactive.py · lines 162–163
InteractiveAgent.__init__() src/minisweagent/agents/interactive.py · lines 36–38
InteractiveAgent.add_messages() src/minisweagent/agents/interactive.py · lines 43–56
InteractiveAgent.step() src/minisweagent/agents/interactive.py · lines 109–122
InteractiveAgent.query() src/minisweagent/agents/interactive.py · lines 58–94
LocalEnvironment src/minisweagent/environments/local.py · lines 19–69
UserInterruption src/minisweagent/exceptions.py · lines 21–22
FormatError src/minisweagent/exceptions.py · lines 25–26
test_successful_completion_with_confirmation() tests/agents/test_interactive.py · lines 133–150
test_human_mode_basic_functionality() tests/agents/test_interactive.py · lines 391–414
test_yolo_mode_activation() tests/agents/test_interactive.py · lines 183–206
test_human_mode_switch_to_confirm() tests/agents/test_interactive.py · lines 450–480
test_action_rejection_and_recovery() tests/agents/test_interactive.py · lines 153–180
test_non_whitelisted_action_requires_confirmation() tests/agents/test_interactive.py · lines 363–385
test_whitelisted_actions_skip_confirmation() tests/agents/test_interactive.py · lines 238–260
test_confirmation_mode_switch_to_human_with_rejection() tests/agents/test_interactive.py · lines 483–513
test_human_mode_switch_to_yolo() tests/agents/test_interactive.py · lines 417–447
test_continue_after_completion_multiple_cycles() tests/agents/test_interactive.py · lines 1020–1053
test_continue_after_completion_with_new_task() tests/agents/test_interactive.py · lines 950–986

Workflow 3

containerized command execution for agent tasks

What this workflow does

Caller supplies a command to DockerEnvironment; the environment runs it in a container and returns captured output/returncode, with a special submission marker raising Submitted.

Purpose

run a command in an isolated container and receive the result back

What it produces

DockerEnvironment / DockerEnvironmentConfig

Who relies on it

agent runtime / task runner

Why this workflow matters

Provides the sandbox used by agent workflows
Converts command completion into structured observation output
Ends the environment cleanly after use

Implementation references

DockerEnvironment src/minisweagent/environments/docker.py · lines 45–161
DockerEnvironment.__init__() src/minisweagent/environments/docker.py · lines 46–59
DockerEnvironment._start_container() src/minisweagent/environments/docker.py · lines 74–99
DockerEnvironment.execute() src/minisweagent/environments/docker.py · lines 101–138
DockerEnvironment._check_finished() src/minisweagent/environments/docker.py · lines 140–151
DockerEnvironment.cleanup() src/minisweagent/environments/docker.py · lines 153–157
test_docker_environment_command_failure() tests/environments/test_docker.py · lines 195–205
test_docker_environment_basic_execution() tests/environments/test_docker.py · lines 56–67
test_docker_environment_forward_nonexistent_env_variables() tests/environments/test_docker.py · lines 115–126
test_docker_environment_custom_cwd() tests/environments/test_docker.py · lines 167–178
test_docker_environment_cwd_parameter_override() tests/environments/test_docker.py · lines 181–192
test_docker_environment_combined_env_and_forward() tests/environments/test_docker.py · lines 129–143
test_docker_environment_env_override_forward() tests/environments/test_docker.py · lines 146–164
test_docker_environment_set_env_variables() tests/environments/test_docker.py · lines 70–89
test_docker_environment_forward_env_variables() tests/environments/test_docker.py · lines 92–112
test_docker_environment_custom_container_timeout() tests/environments/test_docker.py · lines 208–230
test_docker_environment_config_defaults() tests/environments/test_docker.py · lines 43–53

Workflow area 3

Local workflows

2 workflows

Workflow 4

local task execution

What this workflow does

CLI user provides a task and config, then the mini CLI builds runtime components and runs the agent in a local environment.

Purpose

run a coding/problem-solving task with selected model, agent, and environment settings

What it produces

task/problem statement and saved agent trajectory

Who relies on it

CLI user

Why this workflow matters

starts an agent session
persists the resulting trajectory to the configured output path
shows the saved output location when available

Implementation references

main() src/minisweagent/run/mini.py · lines 54–105
recursive_merge() src/minisweagent/utils/serialize.py · lines 6–29
get_agent() src/minisweagent/agents/__init__.py · lines 25–28
get_agent_class() src/minisweagent/agents/__init__.py · lines 14–22
get_environment() src/minisweagent/environments/__init__.py · lines 30–33
get_environment_class() src/minisweagent/environments/__init__.py · lines 19–27
get_model() src/minisweagent/models/__init__.py · lines 45–62

Workflow 5

Run a local agent task from CLI

What this workflow does

CLI user supplies config/model/task; mini main constructs the agent and runs the task, optionally writing an output trajectory file.

Purpose

Launch the agent against a user-provided task with selected model/environment settings

What it produces

task and agent configuration

Who relies on it

CLI user

Why this workflow matters

Produces an agent submission/result
Can persist a trajectory/output file for later inspection
Supports confirm vs yolo mode and exit-immediately behavior

Implementation references

test_configure_if_first_time_called() tests/run/test_cli_integration.py · lines 18–52
test_output_file_is_created() tests/run/test_cli_integration.py · lines 563–649
test_exit_immediately_flag_sets_confirm_exit_false() tests/run/test_cli_integration.py · lines 452–488
test_no_exit_immediately_flag_sets_confirm_exit_true() tests/run/test_cli_integration.py · lines 491–526
test_mini_calls_prompt_when_no_task_provided() tests/run/test_cli_integration.py · lines 97–138
test_cost_limit_zero_is_preserved() tests/run/test_cli_integration.py · lines 271–298
RunBatchProgressManager._save_overview_data_yaml() src/minisweagent/run/benchmarks/utils/batch_progress.py · lines 175–178

Workflow area 4

Benchmark workflows

3 workflows

Workflow 6

Run benchmark evaluation batches

What this workflow does

CLI benchmark operator launches SWE-bench batch processing, which runs agent instances in sandboxed environments and persists per-instance submissions and trajectories.

Purpose

Evaluate many SWE-bench instances and collect results

What it produces

SWE-bench instance, preds.json entry, trajectory JSON

Who relies on it

benchmark runner operator

Why this workflow matters

Produces per-instance model patches for evaluation
Persists execution traces for later inspection
Tracks batch progress and completion status

Implementation references

main() src/minisweagent/run/benchmarks/swebench.py · lines 201–271
update_preds_file() src/minisweagent/run/benchmarks/swebench.py · lines 97–108
remove_from_preds_file() src/minisweagent/run/benchmarks/swebench.py · lines 111–119
process_instance() src/minisweagent/run/benchmarks/swebench.py · lines 122–177
get_sb_environment() src/minisweagent/run/benchmarks/swebench.py · lines 79–94
SingularityEnvironment._build_sandbox() src/minisweagent/environments/singularity.py · lines 46–65
SingularityEnvironmentConfig src/minisweagent/environments/singularity.py · lines 18–34
BubblewrapEnvironment.execute() src/minisweagent/environments/extra/bubblewrap.py · lines 81–117
SingularityEnvironment.execute() src/minisweagent/environments/singularity.py · lines 80–119
BubblewrapEnvironment src/minisweagent/environments/extra/bubblewrap.py · lines 69–151
SingularityEnvironment src/minisweagent/environments/singularity.py · lines 37–139
SwerexModalEnvironment src/minisweagent/environments/extra/swerex_modal.py · lines 37–124

Workflow 7

benchmark batch execution

What this workflow does

Batch benchmark runner loads SWE-bench instances, creates per-instance environments and agents, executes them in parallel, and archives submissions and trajectories.

Purpose

run many benchmark instances and collect artifacts for evaluation

What it produces

benchmark instance, submission archive, trajectory file, exit status

Who relies on it

benchmark operator / evaluation pipeline

Why this workflow matters

produces evaluation-ready outputs
records per-instance status
archives workspace submission artifacts

Implementation references

main() src/minisweagent/run/benchmarks/programbench.py · lines 121–189
process_instance() src/minisweagent/run/benchmarks/programbench.py · lines 63–117
copy_submission() src/minisweagent/run/benchmarks/programbench.py · lines 46–60

Workflow 8

Run benchmark instances and persist predictions

What this workflow does

Benchmark runner filters SWEBench instances, executes them in workers, records predictions and trajectories, and can skip or redo existing results.

Purpose

Process a selected SWEBench subset and write benchmark outputs for each instance

What it produces

benchmark instance, preds.json, trajectory file

Who relies on it

benchmark operator

Why this workflow matters

Produces deterministic benchmark predictions
Persists per-instance trajectory artifacts
Allows rerun control via redo_existing and filtering

Implementation references

test_swebench_end_to_end() tests/run/test_swebench.py · lines 34–77
test_exception_handling_in_agent_run() tests/run/test_swebench.py · lines 461–501
test_update_preds_file_new_file() tests/run/test_swebench.py · lines 218–232
test_filter_instances_no_filters() tests/run/test_swebench.py · lines 135–139
test_redo_existing_false_skips_existing() tests/run/test_swebench.py · lines 341–373
test_redo_existing_true_overwrites_existing() tests/run/test_swebench.py · lines 376–414
test_filter_instances_regex_filter() tests/run/test_swebench.py · lines 142–152
test_filter_instances_slice_only() tests/run/test_swebench.py · lines 155–160
test_different_exception_types() tests/run/test_swebench.py · lines 504–533
test_filter_instances_filter_and_slice() tests/run/test_swebench.py · lines 179–190
test_get_sb_environment_runs_startup_command_as_dict() tests/run/test_swebench.py · lines 107–118

Workflow area 5

Llm workflows

3 workflows

Workflow 9

LLM provider response handling

What this workflow does

Model wrapper calls a Responses API, extracts bash tool calls, and preserves the raw response payload when parsing fails.

Purpose

Send a conversation to a model provider, recover tool-call actions, and keep the raw response for later inspection if parsing fails

What it produces

model response / tool-call output

Who relies on it

agent runtime / trajectory logger

Why this workflow matters

Returns a structured message enriched with actions, cost, and timestamp
Preserves the raw provider response in FormatError extra data for debugging and replay
Supports stateless Responses-API conversation handling

Implementation references

LitellmResponseModel src/minisweagent/models/litellm_response_model.py · lines 25–98
OpenRouterResponseModel src/minisweagent/models/openrouter_response_model.py · lines 31–133
LitellmResponseModel._prepare_messages_for_api() src/minisweagent/models/litellm_response_model.py · lines 29–38
OpenRouterResponseModel._prepare_messages_for_api() src/minisweagent/models/openrouter_response_model.py · lines 70–83
LitellmResponseModel.query() src/minisweagent/models/litellm_response_model.py · lines 52–78
OpenRouterResponseModel.query() src/minisweagent/models/openrouter_response_model.py · lines 85–103
parse_toolcall_actions_response() src/minisweagent/models/utils/actions_toolcall_response.py · lines 58–104
finish_reason_from_responses_api() src/minisweagent/models/utils/actions_toolcall_response.py · lines 45–55
LitellmResponseModel._query() src/minisweagent/models/litellm_response_model.py · lines 40–50
OpenRouterResponseModel._query() src/minisweagent/models/openrouter_response_model.py · lines 44–68
_format_error_message() src/minisweagent/models/utils/actions_toolcall_response.py · lines 28–35
test_litellm_model_format_error_persists_response() tests/models/test_format_error_response_persistence.py · lines 82–106
test_persisted_response_round_trips_through_json() tests/models/test_format_error_response_persistence.py · lines 303–328
TestHasToolCallsDistinction.test_response_parser_passes_has_tool_calls() tests/models/test_truncation_finish_reason.py · lines 112–120
format_toolcall_observation_messages() src/minisweagent/models/utils/actions_toolcall_response.py · lines 107–141

Workflow 10

LLM provider orchestration through Portkey

What this workflow does

Portkey-backed chat/completions and Responses-API wrappers send stateless requests, parse bash actions, and serialize response metadata with cost tracking.

Purpose

Call a model provider via Portkey, recover tool-call actions, and keep structured response metadata for later use

What it produces

Portkey request/response payload

Who relies on it

agent runtime and trajectory logger

Why this workflow matters

Returns structured assistant message output with parsed actions
Persists serialized response metadata in message.extra
Tracks cost and adds it to GLOBAL_MODEL_STATS
Supports both chat-completions and Responses-API provider modes

Implementation references

PortkeyModel.__init__() src/minisweagent/models/portkey_model.py · lines 68–89
PortkeyResponseAPIModel src/minisweagent/models/portkey_response_model.py · lines 45–182
PortkeyResponseAPIModel._prepare_messages_for_api() src/minisweagent/models/portkey_response_model.py · lines 82–94
PortkeyModel src/minisweagent/models/portkey_model.py · lines 65–213
PortkeyResponseAPIModel.__init__() src/minisweagent/models/portkey_response_model.py · lines 54–72
PortkeyModel.query() src/minisweagent/models/portkey_model.py · lines 104–128
PortkeyResponseAPIModel.query() src/minisweagent/models/portkey_response_model.py · lines 96–121
PortkeyModel._parse_actions() src/minisweagent/models/portkey_model.py · lines 130–137
parse_toolcall_actions_response() src/minisweagent/models/utils/actions_toolcall_response.py · lines 58–104
test_portkey_model_missing_api_key() tests/models/test_portkey_model.py · lines 12–17
test_portkey_model_query() tests/models/test_portkey_model.py · lines 43–89
test_portkey_response_model_format_error_persists_response_with_model_dump() tests/models/test_format_error_response_persistence.py · lines 245–270
test_response_api_model_stateless_flattens_response() tests/models/test_portkey_response_model.py · lines 43–90
PortkeyModel.format_message() src/minisweagent/models/portkey_model.py · lines 139–140
PortkeyModel.get_template_vars() src/minisweagent/models/portkey_model.py · lines 155–156
PortkeyModel._prepare_messages_for_api() src/minisweagent/models/portkey_model.py · lines 99–102
PortkeyModel._query() src/minisweagent/models/portkey_model.py · lines 91–97
PortkeyModel.serialize() src/minisweagent/models/portkey_model.py · lines 158–166
PortkeyModel.format_observation_messages() src/minisweagent/models/portkey_model.py · lines 142–153
PortkeyModel._calculate_cost() src/minisweagent/models/portkey_model.py · lines 168–213
PortkeyResponseAPIModel.get_template_vars() src/minisweagent/models/portkey_response_model.py · lines 171–172
PortkeyResponseAPIModel._query() src/minisweagent/models/portkey_response_model.py · lines 74–80
PortkeyResponseAPIModel._parse_actions() src/minisweagent/models/portkey_response_model.py · lines 123–130
PortkeyResponseAPIModel.format_message() src/minisweagent/models/portkey_response_model.py · lines 148–156
PortkeyResponseAPIModel.serialize() src/minisweagent/models/portkey_response_model.py · lines 174–182
PortkeyResponseAPIModel.format_observation_messages() src/minisweagent/models/portkey_response_model.py · lines 158–169
PortkeyResponseAPIModel._calculate_cost() src/minisweagent/models/portkey_response_model.py · lines 132–146
test_portkey_model_config() tests/models/test_portkey_model.py · lines 20–24
test_portkey_model_initialization() tests/models/test_portkey_model.py · lines 27–40

Workflow 11

LLM provider adapter

What this workflow does

The Litellm model adapter sends completion requests, retries on transient failures, computes cost, parses tool calls, and packages the response for the agent loop.

Purpose

obtain a structured model response for the agent loop

What it produces

model completion response and tool-call actions

Who relies on it

agent runtime

Why this workflow matters

returns assistant message plus metadata
tracks per-call cost
preserves response data when formatting fails

Implementation references

LitellmModel src/minisweagent/models/litellm_model.py · lines 49–164
_reorder_anthropic_thinking_blocks() src/minisweagent/models/utils/anthropic_utils.py · lines 11–28
GlobalModelStats.cost() src/minisweagent/models/__init__.py · lines 33–35
GlobalModelStats.n_calls() src/minisweagent/models/__init__.py · lines 37–39
GlobalModelStats.add() src/minisweagent/models/__init__.py · lines 25–31
GlobalModelStats.__init__() src/minisweagent/models/__init__.py · lines 16–23
LitellmModel.query() src/minisweagent/models/litellm_model.py · lines 81–106
GlobalModelStats src/minisweagent/models/__init__.py · lines 13–39
LitellmModel._prepare_messages_for_api() src/minisweagent/models/litellm_model.py · lines 76–79
LitellmModel._query() src/minisweagent/models/litellm_model.py · lines 64–74

Workflow area 6

legacy text action parsing and observation formatting

1 workflow

Workflow 12

legacy text action parsing and observation formatting

What this workflow does

Legacy text-based model utilities extract a single bash action from model text and render observation messages from tool outputs.

Purpose

Turn a text-only model reply into one executable bash action and then render execution output back into chat messages

What it produces

action text / observation message

Who relies on it

model runner and downstream prompt loop

Why this workflow matters

Produces a single normalized command for execution
Raises a structured FormatError when output does not match the required shape
Formats tool output into a user-visible observation message for the next turn

Implementation references

LitellmTextbasedModel src/minisweagent/models/litellm_textbased_model.py · lines 16–48
parse_regex_actions() src/minisweagent/models/utils/actions_text.py · lines 15–40
format_observation_messages() src/minisweagent/models/utils/actions_text.py · lines 43–70
TestRegexActionsTemplateKwargs tests/models/test_truncation_finish_reason.py · lines 54–69
TestRegexActionsTemplateKwargs.test_without_template_kwargs_still_works() tests/models/test_truncation_finish_reason.py · lines 66–69
TestResponseActionsTemplateKwargs.test_finish_reason_reported_on_no_tool_calls() tests/models/test_truncation_finish_reason.py · lines 73–78
TestHasToolCallsDistinction.test_empty_payload_with_tool_calls_reason_reports_truncation() tests/models/test_truncation_finish_reason.py · lines 95–100
TestHasToolCallsDistinction.test_malformed_toolcall_with_tool_calls_reason_shows_real_error() tests/models/test_truncation_finish_reason.py · lines 86–93
TestHasToolCallsDistinction.test_malformed_toolcall_with_length_reason_still_reports_truncation() tests/models/test_truncation_finish_reason.py · lines 102–110
TestHasToolCallsDistinction.test_response_parser_passes_has_tool_calls() tests/models/test_truncation_finish_reason.py · lines 112–120
TestRegexActionsTemplateKwargs.test_finish_reason_reported_on_zero_actions() tests/models/test_truncation_finish_reason.py · lines 55–64

Workflow area 7

global configuration file management

1 workflow

Workflow 13

global configuration file management

What this workflow does

CLI user sets, unsets, edits, or initializes the global config file, which updates the .env-backed model/API-key state and reloads the runtime environment.

Purpose

initialize and maintain model/API-key settings

What it produces

global_config_file and its keys

Who relies on it

local user operating mini-swe-agent

Why this workflow matters

Persists model selection and credential names/values
Allows later runs to reuse configured defaults
Supports editing and removal of config entries

Implementation references

configure_if_first_time() src/minisweagent/run/utilities/config.py · lines 62–66
_reload_config() src/minisweagent/run/utilities/config.py · lines 21–22
edit() src/minisweagent/run/utilities/config.py · lines 122–127
unset() src/minisweagent/run/utilities/config.py · lines 113–119
set() src/minisweagent/run/utilities/config.py · lines 99–110
setup() src/minisweagent/run/utilities/config.py · lines 69–96
TestConfigSet tests/run/test_extra_config.py · lines 117–246
TestConfigSet.test_set_overwrites_existing_key() tests/run/test_extra_config.py · lines 219–231
TestConfigUnset tests/run/test_extra_config.py · lines 249–372
TestTyperAppIntegration tests/run/test_extra_config.py · lines 436–479
TestConfigSetup tests/run/test_extra_config.py · lines 14–114
prompt() src/minisweagent/run/utilities/config.py · lines 55–59

Workflow area 8

Trajectory workflows

1 workflow

Workflow 14

trajectory browsing and inspection

What this workflow does

User opens a trajectory file or directory in the TrajectoryInspector UI, browses steps and trajectories, toggles reasoning, reloads files, and may open content in jless.

Purpose

review saved agent conversations and their rendered tool/assistant content

What it produces

trajectory files and message steps

Who relies on it

developer or operator inspecting agent runs

Why this workflow matters

Provides visual review of agent steps and reasoning
Supports navigation across multiple trajectories and steps
Allows refreshing from disk after file changes

Implementation references

TrajectoryInspector src/minisweagent/run/utilities/inspector.py · lines 79–292
main() src/minisweagent/run/utilities/inspector.py · lines 295–312
TrajectoryInspector._load_current_trajectory() src/minisweagent/run/utilities/inspector.py · lines 153–176
TrajectoryInspector.__init__() src/minisweagent/run/utilities/inspector.py · lines 97–112
TrajectoryInspector.action_toggle_reasoning() src/minisweagent/run/utilities/inspector.py · lines 274–276
TrajectoryInspector.update_content() src/minisweagent/run/utilities/inspector.py · lines 195–225
test_trajectory_inspector_reload_clamps_step() tests/run/test_inspector.py · lines 465–481
test_trajectory_inspector_ansi_content() tests/run/test_inspector.py · lines 506–521
test_trajectory_inspector_reasoning_display() tests/run/test_inspector.py · lines 393–414
TrajectoryInspector.action_open_in_jless() src/minisweagent/run/utilities/inspector.py · lines 263–272
test_trajectory_inspector_load_trajectory_formats() tests/run/test_inspector.py · lines 270–303
test_trajectory_inspector_toggle_reasoning() tests/run/test_inspector.py · lines 417–435
test_trajectory_inspector_reload() tests/run/test_inspector.py · lines 438–462

Workflow area 9

Batch workflows

1 workflow

Workflow 15

Track and summarize batch execution progress

What this workflow does

Batch progress manager tracks per-instance lifecycle, counts exit statuses, computes ETA and cost, and optionally saves a YAML overview report.

Purpose

Provide live progress feedback and a persisted summary during benchmark runs

What it produces

instance exit statuses and overview report

Who relies on it

benchmark operator

Why this workflow matters

Shows live progress and ETA
Aggregates exit statuses for review
Persists a YAML summary of instance outcomes

Implementation references

RunBatchProgressManager src/minisweagent/run/benchmarks/utils/batch_progress.py · lines 36–178
RunBatchProgressManager.__init__() src/minisweagent/run/benchmarks/utils/batch_progress.py · lines 37–83
RunBatchProgressManager._save_overview_data_yaml() src/minisweagent/run/benchmarks/utils/batch_progress.py · lines 175–178
test_instance_lifecycle() tests/run/test_batch_progress.py · lines 44–51
test_multiple_instances() tests/run/test_batch_progress.py · lines 62–71
RunBatchProgressManager._get_eta_text() src/minisweagent/run/benchmarks/utils/batch_progress.py · lines 89–97
RunBatchProgressManager.on_instance_end() src/minisweagent/run/benchmarks/utils/batch_progress.py · lines 145–156
RunBatchProgressManager.on_uncaught_exception() src/minisweagent/run/benchmarks/utils/batch_progress.py · lines 158–159
test_manager_initialization() tests/run/test_batch_progress.py · lines 34–36
test_manager_with_yaml_path() tests/run/test_batch_progress.py · lines 39–41
test_yaml_report_generation() tests/run/test_batch_progress.py · lines 87–98
test_uncaught_exception() tests/run/test_batch_progress.py · lines 74–79
test_update_instance_status() tests/run/test_batch_progress.py · lines 82–84
test_get_overview_data() tests/run/test_batch_progress.py · lines 101–106
test_concurrent_operations() tests/run/test_batch_progress.py · lines 125–136
test_print_report() tests/run/test_batch_progress.py · lines 109–122