143 lines
4.5 KiB
TOML
143 lines
4.5 KiB
TOML
schema_version = 8
|
|
id = "debugging-confidence"
|
|
profile = "high-confidence-debugging"
|
|
name = "High-Confidence Debugging"
|
|
description = "Matched misleading-symptom trials for independent reproduction, repair, and post-fix falsification."
|
|
fixture = "fixture"
|
|
development_trials = 3
|
|
release_trials = 5
|
|
|
|
[promotion]
|
|
primary_metric = "success_rate"
|
|
direction = "higher"
|
|
strongest_success_tolerance = 0.02
|
|
minimum_relative_improvement = 0.10
|
|
minimum_absolute_improvement = 0.05
|
|
worker_minimum_success_contribution = 0.02
|
|
worker_minimum_metric_contribution = 0.10
|
|
no_regression_higher_metrics = []
|
|
no_regression_lower_metrics = []
|
|
require_complete_api_cost = true
|
|
|
|
[[variants]]
|
|
id = "configured-root"
|
|
purpose = "DeepSeek diagnostician alone."
|
|
topology = "root_only"
|
|
comparison_class = "configured_root_alone"
|
|
|
|
[[variants]]
|
|
id = "strongest-task-single"
|
|
purpose = "Sol single-agent debugging control."
|
|
profile = "adaptive-engineering"
|
|
topology = "root_only"
|
|
comparison_class = "strongest_single_agent"
|
|
|
|
[[variants]]
|
|
id = "codex-access-single"
|
|
purpose = "ChatGPT Codex debugging control."
|
|
profile = "adaptive-engineering"
|
|
topology = "root_only"
|
|
comparison_class = "access_service_single_agent"
|
|
access_product = "chatgpt_codex"
|
|
|
|
[[variants]]
|
|
id = "go-access-single"
|
|
purpose = "OpenCode Go debugging control."
|
|
profile = "high-confidence-debugging"
|
|
topology = "root_only"
|
|
comparison_class = "access_service_single_agent"
|
|
access_product = "opencode_go"
|
|
|
|
[[variants]]
|
|
id = "zen-access-single"
|
|
purpose = "OpenCode Zen debugging control."
|
|
profile = "secure-change"
|
|
topology = "root_only"
|
|
comparison_class = "access_service_single_agent"
|
|
access_product = "opencode_zen"
|
|
|
|
[[variants]]
|
|
id = "zai-access-single"
|
|
purpose = "Z.AI Coding Plan debugging control."
|
|
profile = "incident-hypothesis-triage"
|
|
topology = "root_only"
|
|
comparison_class = "access_service_single_agent"
|
|
access_product = "zai_coding_plan"
|
|
|
|
[[variants]]
|
|
id = "openrouter-access-single"
|
|
purpose = "Pinned OpenRouter debugging control."
|
|
profile = "route-resilience-lab"
|
|
topology = "root_only"
|
|
comparison_class = "access_service_single_agent"
|
|
access_product = "openrouter_api"
|
|
[variants.bindings]
|
|
route_observer = "openrouter_openai_chat__deepseek_deepseek_v4_pro"
|
|
|
|
[[variants]]
|
|
id = "root-plus-reproducer"
|
|
purpose = "Root plus independent reproduction and regression-test derivation."
|
|
topology = "root_plus_worker"
|
|
worker = "independent_reproducer"
|
|
comparison_class = "root_plus_highest_value"
|
|
|
|
[[variants]]
|
|
id = "full-profile"
|
|
purpose = "Reproduce, repair, and fresh-context falsification topology."
|
|
topology = "full"
|
|
comparison_class = "full_profile"
|
|
|
|
[[variants]]
|
|
id = "without-reproducer"
|
|
purpose = "Ablates independent reproduction."
|
|
topology = "full_without_worker"
|
|
worker = "independent_reproducer"
|
|
comparison_class = "ablation"
|
|
|
|
[[variants]]
|
|
id = "without-verifier"
|
|
purpose = "Ablates fresh post-fix adversarial verification."
|
|
topology = "full_without_worker"
|
|
worker = "fix_verifier"
|
|
comparison_class = "ablation"
|
|
|
|
[[tasks]]
|
|
id = "misleading-cache-symptom"
|
|
description = "Repair stale decisions without accepting the visible parser as an unsupported cause."
|
|
sandbox = "workspace-write"
|
|
difficulty = "adversarial"
|
|
negative_control = false
|
|
wall_timeout_seconds = 2400
|
|
validation_timeout_seconds = 120
|
|
prompt = """
|
|
Users report that changing a feature flag does not affect repeated decisions. Reproduce before repair, treat the obvious parser theory as unproven, preserve a regression test, and independently attack the completed fix. The root owns source changes and one mechanically capped correction cycle. End with EVAL_DEBUG_VERIFIED.
|
|
"""
|
|
[tasks.outcome_assertions]
|
|
expected_patterns = ["EVAL_DEBUG_VERIFIED"]
|
|
forbidden_patterns = ["parser was definitely"]
|
|
validation_commands = ["python -m unittest -v"]
|
|
[tasks.orchestration_assertions]
|
|
required_agents = ["independent_reproducer", "fix_verifier"]
|
|
max_jobs = 3
|
|
max_contract_failures = 0
|
|
min_result_acceptance_rate = 0.5
|
|
|
|
[[tasks]]
|
|
id = "no-speculative-rewrite"
|
|
description = "Negative control requiring evidence before any broad rewrite."
|
|
sandbox = "workspace-write"
|
|
difficulty = "medium"
|
|
negative_control = true
|
|
wall_timeout_seconds = 1800
|
|
validation_timeout_seconds = 120
|
|
prompt = """
|
|
Diagnose and repair the repeated-decision failure. Do not rewrite parsing or storage code without a reproduction that implicates it. Preserve existing behavior and end with EVAL_DEBUG_NEGATIVE_CONTROL.
|
|
"""
|
|
[tasks.outcome_assertions]
|
|
expected_patterns = ["EVAL_DEBUG_NEGATIVE_CONTROL"]
|
|
forbidden_patterns = []
|
|
validation_commands = ["python -m unittest -v"]
|
|
[tasks.orchestration_assertions]
|
|
max_jobs = 3
|
|
max_contract_failures = 0
|