{"data":{"kind":"file","path":"README.md","version_id":"coa0c681p77y4ai1ligo8a2w","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2554,"modified_at":"2026-09-12T21:38:17.427000","content_hash":"fedf96198054f1cb0edac8d93956691100885f0509005dd93bfdf47b662839e8"},"entries":[],"content":"# Long-Horizon Data Pipeline Recovery V1\n\nAgentForge-Bench RL environment for long-horizon recovery of stateful data\npipelines.\n\n## Overview\n\nThe environment models a pipeline:\n\nInput → Schema Validation → Cleaning → Transformation →\nAggregation → Output Validation → Output\n\nThe agent encounters an incident and must inspect, diagnose, select a recovery\nbranch, apply changes, execute the pipeline, respond to downstream\nconsequences, recover or replan when necessary, and complete global validation.\n\n## Why This Task Exists\n\nTask 002 is designed to extend the long-horizon reasoning requirements of\nAgentForge-Bench.\n\nCompared with a short single-step repair task, this environment introduces:\n\n- sequential dependencies\n- multiple incident families\n- branch-specific recovery\n- nested recovery strategies\n- downstream consequences\n- delayed failure detection\n- recovery\n- replanning\n- global validation\n\n## Public Execution Model\n\nThe public environment exposes mechanics required for interaction.\n\nThe causal dependency graph is:\n\n```text\nInput\n  ↓\nSchema Validation\n  ↓\nCleaning\n  ↓\nTransformation\n  ↓\nAggregation\n  ↓\nOutput Validation\n  ↓\nOutput\nA component can be locally repaired while downstream components remain\ninvalid or blocked.\n\nActions\n\nThe current action contract includes:\n\ninspect_pipeline\ninspect_component\nanalyze_failure\nselect_branch\nselect_strategy\nmodify_component\nmodify_config\nrun_pipeline\ncheck_schema\ncheck_output\nrecover\nreplan\nfinal_verify\nBranch Families\n\nThe public branch model contains:\n\nschema recovery\ntransformation recovery\ndata quality recovery\nresource recovery\n\nEach family provides multiple recovery strategies.\n\nStatefulness\n\nThe environment tracks:\n\nlogical progress\nincident state\ndiagnosis\nselected branch\nselected strategy\ncomponent validity\npipeline execution\nfailure detection\nrecovery\nreplanning\ndownstream inconsistency\nglobal validation\nterminal success\nAnti-Cheat\n\nThe agent-facing package does not intentionally contain:\n\noracle answers\nhidden answers\nreference solutions\nhidden evaluation truth\nexpected optimal trajectories\n\nSee:\n\ndocs/ANTI_CHEAT.md\ndocs/ANTI_MEMORIZATION.md\nEvaluation Boundary\n\nPublic mechanics live in the task package.\n\nHidden scenario data and evaluator-only truth belong under the evaluator boundary\nand must not be imported by the agent-facing environment.\n\nCurrent Status\n\nDevelopment.\n\nThe execution engine and causal layer are under iterative validation.\n\nThis task must not be considered frozen until the complete release checklist\nhas passed.\n","encoding":"utf-8","truncated":false,"total_bytes":2554},"status":null}