150 lines
4.7 KiB
TOML
150 lines
4.7 KiB
TOML
schema_version = 8
|
|
id = "adaptive-change"
|
|
profile = "adaptive-engineering"
|
|
name = "Adaptive Engineering Change"
|
|
description = "Matched trials for selective delegation on decomposable and tightly coupled engineering work."
|
|
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 = "Configured Sol root with delegation mechanically disabled."
|
|
topology = "root_only"
|
|
comparison_class = "configured_root_alone"
|
|
|
|
[[variants]]
|
|
id = "strongest-task-single"
|
|
purpose = "Task-specific flagship Sol single-agent control."
|
|
profile = "adaptive-engineering"
|
|
topology = "root_only"
|
|
comparison_class = "strongest_single_agent"
|
|
|
|
[[variants]]
|
|
id = "codex-access-single"
|
|
purpose = "ChatGPT Codex access-service 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 economical DeepSeek 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 Claude 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 GLM 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 DeepSeek route 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-implementation"
|
|
purpose = "Configured root plus the objectively testable implementation branch."
|
|
topology = "root_plus_worker"
|
|
worker = "implementation_specialist"
|
|
comparison_class = "root_plus_highest_value"
|
|
|
|
[[variants]]
|
|
id = "full-profile"
|
|
purpose = "Complete selective adaptive topology."
|
|
topology = "full"
|
|
comparison_class = "full_profile"
|
|
|
|
[[variants]]
|
|
id = "without-scout"
|
|
purpose = "Ablates native reconnaissance."
|
|
topology = "full_without_worker"
|
|
worker = "repo_scout"
|
|
comparison_class = "ablation"
|
|
|
|
[[variants]]
|
|
id = "without-implementation"
|
|
purpose = "Ablates the economical isolated implementation branch."
|
|
topology = "full_without_worker"
|
|
worker = "implementation_specialist"
|
|
comparison_class = "ablation"
|
|
|
|
[[variants]]
|
|
id = "without-review"
|
|
purpose = "Ablates independent cross-maker review."
|
|
topology = "full_without_worker"
|
|
worker = "adversarial_reviewer"
|
|
comparison_class = "ablation"
|
|
|
|
[[tasks]]
|
|
id = "decomposable-ledger"
|
|
description = "Repair an idempotency defect with separable reconnaissance, implementation, and review work."
|
|
sandbox = "workspace-write"
|
|
difficulty = "hard"
|
|
negative_control = false
|
|
wall_timeout_seconds = 2400
|
|
validation_timeout_seconds = 120
|
|
prompt = """
|
|
Repair the ledger defect described in SPEC.md. Own diagnosis and canonical integration. Delegate only work with a concrete expected contribution, continue non-overlapping critical-path work while it runs, explicitly accept or reject results, and end with EVAL_ADAPTIVE_VERIFIED.
|
|
"""
|
|
[tasks.outcome_assertions]
|
|
expected_patterns = ["EVAL_ADAPTIVE_VERIFIED"]
|
|
forbidden_patterns = []
|
|
validation_commands = ["python -m unittest -v"]
|
|
[tasks.orchestration_assertions]
|
|
required_agents = ["implementation_specialist", "adversarial_reviewer"]
|
|
max_jobs = 3
|
|
max_contract_failures = 0
|
|
max_observed_mcp_wait_ratio = 0.8
|
|
|
|
[[tasks]]
|
|
id = "tight-small-change"
|
|
description = "Negative control where worker launch overhead should not be required."
|
|
sandbox = "workspace-write"
|
|
difficulty = "easy"
|
|
negative_control = true
|
|
wall_timeout_seconds = 1200
|
|
validation_timeout_seconds = 120
|
|
prompt = """
|
|
Make the smallest correct ledger repair. This task is intentionally compact: remain single-agent unless independent work has a concrete positive expected value. Run the tests and end with EVAL_ADAPTIVE_SMALL.
|
|
"""
|
|
[tasks.outcome_assertions]
|
|
expected_patterns = ["EVAL_ADAPTIVE_SMALL"]
|
|
forbidden_patterns = []
|
|
validation_commands = ["python -m unittest -v"]
|
|
[tasks.orchestration_assertions]
|
|
max_jobs = 1
|
|
max_contract_failures = 0
|