kavirubc commited on
Commit
b374b66
·
verified ·
1 Parent(s): bb1d2b6

Upload folder using huggingface_hub

Browse files
DATASHEET.md ADDED
@@ -0,0 +1,197 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Datasheet for AgentDisruptBench
2
+
3
+ _Following the [Datasheets for Datasets](https://arxiv.org/abs/1803.09010) framework (Gebru et al., 2021)._
4
+
5
+ ---
6
+
7
+ ## Motivation
8
+
9
+ ### For what purpose was the dataset created?
10
+
11
+ AgentDisruptBench was created to evaluate AI agent resilience under runtime tool-call disruptions. Existing benchmarks measure *whether* agents can use tools correctly, but assume tools behave perfectly — an unrealistic assumption in production environments where APIs time out, return malformed responses, enforce rate limits, and cascade failures. This benchmark fills the gap by providing structured tasks paired with configurable disruption profiles to measure agent reliability, recovery strategies, and graceful degradation.
12
+
13
+ ### Who created the dataset and on behalf of which entity?
14
+
15
+ AgentDisruptBench Contributors, as part of academic research into AI agent robustness.
16
+
17
+ ### Who funded the creation of the dataset?
18
+
19
+ <!-- Fill in: grant numbers, institutional support, etc. -->
20
+ [To be specified]
21
+
22
+ ---
23
+
24
+ ## Composition
25
+
26
+ ### What do the instances that comprise the dataset represent?
27
+
28
+ Each instance is a **benchmark task** defined in YAML with:
29
+ - A natural language task description (the prompt given to the agent)
30
+ - Required tools (from a set of 30 deterministic mock tools)
31
+ - Ground truth: expected outcome, required tool calls, evaluation rubric with weighted criteria
32
+ - Metadata: domain, difficulty (1-5), task type
33
+
34
+ ### How many instances are there in total?
35
+
36
+ **100 tasks** organized as follows:
37
+
38
+ | Domain | Standard | Adversarial | Impossible | Handover | Total |
39
+ |------------|:--------:|:-----------:|:----------:|:--------:|:-----:|
40
+ | Retail | 20 | 2 | 2 | 1 | 25 |
41
+ | Travel | 20 | 2 | 2 | 1 | 25 |
42
+ | Finance | 20 | 2 | 2 | 1 | 25 |
43
+ | DevOps | 20 | 2 | 2 | 1 | 25 |
44
+ | **Total** | **80** | **8** | **8** | **4** | **100** |
45
+
46
+ ### Does the dataset contain all possible instances or is it a sample?
47
+
48
+ The dataset is a curated collection of synthetic tasks. It is not a sample from a larger population. Task difficulty levels are distributed as D1=8, D2=16, D3=24, D4=20, D5=12 tasks across standard tasks, following a roughly bell-shaped distribution centered on medium difficulty.
49
+
50
+ ### What data does each instance consist of?
51
+
52
+ Each task instance contains:
53
+ - `task_id`: Unique identifier (e.g., `retail_001`)
54
+ - `title`: Short human-readable title
55
+ - `description`: Full prompt given to the agent (natural language)
56
+ - `domain`: One of {retail, travel, finance, devops}
57
+ - `difficulty`: Integer 1-5
58
+ - `task_type`: One of {standard, adversarial, impossible}
59
+ - `required_tools`: List of tool names needed
60
+ - `expected_tool_call_depth`: Expected number of tool calls under clean conditions
61
+ - `ground_truth`: Structured evaluation criteria including:
62
+ - `expected_outcome`: Description of success
63
+ - `required_tool_calls`: Tools that must be called
64
+ - `forbidden_tool_calls`: Tools that must NOT be called
65
+ - `correct_final_answer`: Exact expected answer (if applicable)
66
+ - `evaluation_rubric`: Criterion → weight mapping (sums ≈ 1.0)
67
+ - `disruption_sensitive_tools`: Tools where failure is most impactful
68
+ - `recovery_actions`: Expected recovery behaviours
69
+ - `trap_description`: (adversarial only) the trap the agent should avoid
70
+ - `impossibility_reason`: (impossible only) why no valid solution exists
71
+
72
+ ### Is any information missing from individual instances?
73
+
74
+ No. All fields are populated for every task instance.
75
+
76
+ ### Are relationships between individual instances made explicit?
77
+
78
+ Tasks within the same domain share tools and mock data (e.g., retail tasks all use the same product catalog and customer database via deterministic mock tools). Cross-domain relationships do not exist.
79
+
80
+ ### Are there recommended data splits?
81
+
82
+ No train/test split is provided because this is a benchmark, not a training dataset. All 100 tasks are intended for evaluation. However, researchers may partition by domain, difficulty, or task type for focused analysis.
83
+
84
+ ### Does the dataset contain data that might be considered confidential?
85
+
86
+ No. All data is synthetic. No real personal data, API keys, or proprietary information is included.
87
+
88
+ ### Does the dataset contain data that, if viewed directly, might be offensive, insulting, threatening, or might otherwise cause anxiety?
89
+
90
+ No. Tasks involve routine operations (e-commerce, travel booking, financial transactions, DevOps operations) with no offensive content.
91
+
92
+ ---
93
+
94
+ ## Collection Process
95
+
96
+ ### How was the data associated with each instance acquired?
97
+
98
+ All tasks were authored manually by the benchmark contributors. They are synthetic, designed to cover a range of realistic tool-use scenarios across difficulty levels.
99
+
100
+ ### What mechanisms or procedures were used to collect the data?
101
+
102
+ Task design followed a structured process:
103
+ 1. Define 4 domains based on common LLM agent use cases
104
+ 2. Design 8 mock tools per domain (30 total)
105
+ 3. Author 20 standard tasks per domain with increasing difficulty
106
+ 4. Author adversarial tasks with designed trap patterns
107
+ 5. Author impossible tasks with clear impossibility reasons
108
+ 6. Author handover tasks requiring human escalation
109
+ 7. Validate all ground truth against mock tool outputs
110
+
111
+ ### Who was involved in the data collection process?
112
+
113
+ The benchmark authors.
114
+
115
+ ### Over what timeframe was the data collected?
116
+
117
+ March-April 2026.
118
+
119
+ ### Were any ethical review processes conducted?
120
+
121
+ <!-- Fill in if applicable -->
122
+ [To be specified]
123
+
124
+ ---
125
+
126
+ ## Uses
127
+
128
+ ### What are the intended uses of the dataset?
129
+
130
+ 1. **Primary**: Evaluate AI agent resilience under runtime tool-call disruptions using the provided disruption profiles (clean → hostile_environment)
131
+ 2. **Secondary**: Compare agent frameworks (LangChain, OpenAI, AutoGen, CrewAI) on reliability metrics
132
+ 3. **Tertiary**: Study recovery strategies and failure patterns in LLM agents
133
+
134
+ ### What are some tasks/uses that the dataset should not be used for?
135
+
136
+ - **Not for training**: The benchmark should be used for evaluation only. Training on the tasks would compromise benchmark validity.
137
+ - **Not for real tool testing**: Mock tools are deterministic simulations, not real API integrations.
138
+ - **Not for safety/security evaluation**: The disruptions model reliability failures, not adversarial attacks.
139
+
140
+ ### Is there anything about the composition of the dataset or the way it was collected that might impact future uses?
141
+
142
+ - Tasks are in English only
143
+ - Mock tools simulate US-centric services (USD currency, US dates, etc.)
144
+ - Ground truth rubrics use string-matching heuristics which may not capture all valid agent responses
145
+
146
+ ---
147
+
148
+ ## Distribution
149
+
150
+ ### How will the dataset be distributed?
151
+
152
+ - **Primary**: GitHub repository at [https://github.com/Kavirubc/AgentDisruptBench](https://github.com/Kavirubc/AgentDisruptBench)
153
+ - **Persistent**: HuggingFace Datasets at [To be created]
154
+ - **Metadata**: Croissant JSON-LD file included in the repository
155
+
156
+ ### When will the dataset be distributed?
157
+
158
+ The dataset is publicly available as of the initial release.
159
+
160
+ ### Will the dataset be distributed under a copyright or other intellectual property (IP) license?
161
+
162
+ MIT License.
163
+
164
+ ### Have any third parties imposed IP-based or other restrictions on the data?
165
+
166
+ No.
167
+
168
+ ---
169
+
170
+ ## Maintenance
171
+
172
+ ### Who is supporting/hosting/maintaining the dataset?
173
+
174
+ AgentDisruptBench Contributors via the GitHub repository.
175
+
176
+ ### How can the owner/curator/manager of the dataset be contacted?
177
+
178
+ Via GitHub Issues at the repository.
179
+
180
+ ### Will the dataset be updated?
181
+
182
+ Yes. Planned updates include:
183
+ - Additional task variants for ε-robustness testing
184
+ - New domains
185
+ - Track B: Network-layer disruption injection via Envoy proxy
186
+
187
+ ### If the dataset relates to people, are there applicable limits on the retention of the data?
188
+
189
+ N/A — the dataset is entirely synthetic with no real personal data.
190
+
191
+ ### Will older versions of the dataset continue to be available?
192
+
193
+ Yes, via Git tags and versioned releases.
194
+
195
+ ### If others want to extend/augment/build on/contribute to the dataset, is there a mechanism for doing so?
196
+
197
+ Yes. See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines on contributing new tasks, domains, tools, and disruption types.
README.md ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ license: mit
5
+ task_categories:
6
+ - text-generation
7
+ tags:
8
+ - agents
9
+ - benchmarks
10
+ - tool-calling
11
+ - resilience
12
+ - disruption
13
+ - fault-injection
14
+ - LLM
15
+ - evaluation
16
+ pretty_name: AgentDisruptBench
17
+ size_categories:
18
+ - n<1K
19
+ ---
20
+
21
+ # AgentDisruptBench
22
+
23
+ A benchmark for evaluating AI agent resilience under runtime tool-call disruptions.
24
+
25
+ ## Overview
26
+
27
+ AgentDisruptBench provides **100 benchmark tasks** across **4 domains** with a systematic **20-type disruption taxonomy** and **9 disruption severity profiles** to measure how well LLM-based agents handle real-world tool failures.
28
+
29
+ ## Task Statistics
30
+
31
+ | Domain | Standard | Adversarial | Impossible | Handover | Total |
32
+ |----------|:--------:|:-----------:|:----------:|:--------:|:-----:|
33
+ | Retail | 20 | 2 | 2 | 1 | 25 |
34
+ | Travel | 20 | 2 | 2 | 1 | 25 |
35
+ | Finance | 20 | 2 | 2 | 1 | 25 |
36
+ | DevOps | 20 | 2 | 2 | 1 | 25 |
37
+ | **Total**| **80** | **8** | **8** | **4** | **100** |
38
+
39
+ ## Disruption Taxonomy (20 Types)
40
+
41
+ | Category | Types |
42
+ |----------|-------|
43
+ | Timing | timeout, latency |
44
+ | HTTP Status | http_429, http_401, http_403, http_500, http_502, http_503 |
45
+ | Response Content | malformed_json, truncated, null_response, missing_fields, type_mismatch, schema_drift, wrong_data |
46
+ | Behavioral | intermittent, flapping, quota_exhausted, auth_expiry, cascading |
47
+
48
+ ## Key Metric: R(k, ε, λ) Reliability Surface
49
+
50
+ - **k-consistency**: Pass rate across repeated seeds (same task, same profile)
51
+ - **ε-robustness**: Pass rate across task-wording variants
52
+ - **λ-fault-tolerance**: Pass rate across disruption profiles
53
+
54
+ ## Files
55
+
56
+ - `tasks/retail.yaml` — 20 retail domain tasks
57
+ - `tasks/travel.yaml` — 20 travel domain tasks
58
+ - `tasks/finance.yaml` — 20 finance domain tasks
59
+ - `tasks/devops.yaml` — 20 DevOps domain tasks
60
+ - `tasks/adversarial.yaml` — 8 adversarial trap tasks
61
+ - `tasks/impossible.yaml` — 8 impossible tasks
62
+ - `tasks/handover.yaml` — 4 handover tasks
63
+ - `tasks/variants.yaml` — 18 ε-robustness task variants
64
+ - `profiles/` — 9 disruption profile definitions
65
+
66
+ ## Usage
67
+
68
+ ```python
69
+ pip install agentdisruptbench
70
+
71
+ from agentdisruptbench import TaskRegistry, DisruptionEngine
72
+ registry = TaskRegistry.from_builtin()
73
+ tasks = registry.filter(domain="retail", max_difficulty=3)
74
+ ```
75
+
76
+ ## Citation
77
+
78
+ ```bibtex
79
+ @inproceedings{agentdisruptbench2026,
80
+ title={AgentDisruptBench: Evaluating AI Agent Resilience Under Runtime Tool-Call Disruptions},
81
+ author={AgentDisruptBench Contributors},
82
+ year={2026}
83
+ }
84
+ ```
85
+
86
+ ## License
87
+
88
+ MIT
croissant.json ADDED
@@ -0,0 +1,208 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "@context": {
3
+ "@language": "en",
4
+ "@vocab": "https://schema.org/",
5
+ "citeAs": "cr:citeAs",
6
+ "column": "cr:column",
7
+ "conformsTo": "dct:conformsTo",
8
+ "cr": "http://mlcommons.org/croissant/",
9
+ "data": {
10
+ "@id": "cr:data",
11
+ "@type": "@json"
12
+ },
13
+ "dataBiases": "cr:dataBiases",
14
+ "dataCollection": "cr:dataCollection",
15
+ "dataType": {
16
+ "@id": "cr:dataType",
17
+ "@type": "@vocab"
18
+ },
19
+ "dct": "http://purl.org/dc/terms/",
20
+ "extract": "cr:extract",
21
+ "field": "cr:field",
22
+ "fileProperty": "cr:fileProperty",
23
+ "fileSet": "cr:fileSet",
24
+ "format": "cr:format",
25
+ "includes": "cr:includes",
26
+ "isLiveDataset": "cr:isLiveDataset",
27
+ "jsonPath": "cr:jsonPath",
28
+ "key": "cr:key",
29
+ "md5": "cr:md5",
30
+ "parentField": "cr:parentField",
31
+ "path": "cr:path",
32
+ "personalSensitiveInformation": "cr:personalSensitiveInformation",
33
+ "recordSet": "cr:recordSet",
34
+ "references": "cr:references",
35
+ "regex": "cr:regex",
36
+ "repeated": "cr:repeated",
37
+ "replace": "cr:replace",
38
+ "sc": "https://schema.org/",
39
+ "separator": "cr:separator",
40
+ "source": "cr:source",
41
+ "subField": "cr:subField",
42
+ "transform": "cr:transform"
43
+ },
44
+ "@type": "sc:Dataset",
45
+ "name": "AgentDisruptBench",
46
+ "description": "A benchmark for evaluating AI agent resilience under runtime tool-call disruptions. Contains 100 tasks across 4 domains (Retail, Travel, Finance, DevOps) with 20 disruption types and 9 disruption profiles.",
47
+ "conformsTo": "http://mlcommons.org/croissant/1.0",
48
+ "url": "https://github.com/Kavirubc/AgentDisruptBench",
49
+ "license": "https://opensource.org/licenses/MIT",
50
+ "version": "0.1.0",
51
+ "datePublished": "2026-04-03",
52
+ "creator": {
53
+ "@type": "sc:Organization",
54
+ "name": "AgentDisruptBench Contributors"
55
+ },
56
+ "keywords": [
57
+ "AI agents",
58
+ "benchmarks",
59
+ "resilience",
60
+ "tool-calling",
61
+ "disruptions",
62
+ "evaluation",
63
+ "LLM",
64
+ "fault-injection",
65
+ "reliability"
66
+ ],
67
+ "isLiveDataset": false,
68
+ "distribution": [
69
+ {
70
+ "@type": "cr:FileObject",
71
+ "@id": "retail-tasks",
72
+ "name": "retail.yaml",
73
+ "description": "20 standard retail domain tasks (difficulty 1-5)",
74
+ "contentUrl": "python/agentdisruptbench/tasks/builtin/retail.yaml",
75
+ "encodingFormat": "application/x-yaml",
76
+ "sha256": "5c00f518f411c868ab8ee81c35c664e443734716ce38944833deffc9ea49b98d"
77
+ },
78
+ {
79
+ "@type": "cr:FileObject",
80
+ "@id": "travel-tasks",
81
+ "name": "travel.yaml",
82
+ "description": "20 standard travel domain tasks (difficulty 1-5)",
83
+ "contentUrl": "python/agentdisruptbench/tasks/builtin/travel.yaml",
84
+ "encodingFormat": "application/x-yaml",
85
+ "sha256": "2e06442d72fb8d02d098bb1bb08106abce521d0eccf6d044b3354c69689913b7"
86
+ },
87
+ {
88
+ "@type": "cr:FileObject",
89
+ "@id": "finance-tasks",
90
+ "name": "finance.yaml",
91
+ "description": "20 standard finance domain tasks (difficulty 1-5)",
92
+ "contentUrl": "python/agentdisruptbench/tasks/builtin/finance.yaml",
93
+ "encodingFormat": "application/x-yaml",
94
+ "sha256": "3b5e30eec5b047e2e02b862f7f3e4e1672d16e491aea4990e17440f614bd0594"
95
+ },
96
+ {
97
+ "@type": "cr:FileObject",
98
+ "@id": "devops-tasks",
99
+ "name": "devops.yaml",
100
+ "description": "20 standard DevOps domain tasks (difficulty 1-5)",
101
+ "contentUrl": "python/agentdisruptbench/tasks/builtin/devops.yaml",
102
+ "encodingFormat": "application/x-yaml",
103
+ "sha256": "4f1f6f83275c6b02802b13dd5a67b1760047376354bf0bb6740489877abde947"
104
+ },
105
+ {
106
+ "@type": "cr:FileObject",
107
+ "@id": "adversarial-tasks",
108
+ "name": "adversarial.yaml",
109
+ "description": "8 adversarial trap tasks (2 per domain)",
110
+ "contentUrl": "python/agentdisruptbench/tasks/builtin/adversarial.yaml",
111
+ "encodingFormat": "application/x-yaml",
112
+ "sha256": "f089191acdc97f69fb689cad337a68eee1c5460c391bb94f98443cbefd6b414d"
113
+ },
114
+ {
115
+ "@type": "cr:FileObject",
116
+ "@id": "impossible-tasks",
117
+ "name": "impossible.yaml",
118
+ "description": "8 impossible tasks where no valid solution exists (2 per domain)",
119
+ "contentUrl": "python/agentdisruptbench/tasks/builtin/impossible.yaml",
120
+ "encodingFormat": "application/x-yaml",
121
+ "sha256": "efc67f7b35e356d5ab33cd84f16063a906dd6f3932065f2a127731d2f8f21953"
122
+ },
123
+ {
124
+ "@type": "cr:FileObject",
125
+ "@id": "handover-tasks",
126
+ "name": "handover.yaml",
127
+ "description": "4 handover tasks requiring human escalation (1 per domain)",
128
+ "contentUrl": "python/agentdisruptbench/tasks/builtin/handover.yaml",
129
+ "encodingFormat": "application/x-yaml",
130
+ "sha256": "fb6d522ca219f0f3634914f7534c45c8bbd5ffcf0088019279b8a468788e6de0"
131
+ }
132
+ ],
133
+ "recordSet": [
134
+ {
135
+ "@type": "cr:RecordSet",
136
+ "@id": "tasks",
137
+ "name": "Benchmark Tasks",
138
+ "description": "The 100 benchmark tasks across 4 domains and 3 task types.",
139
+ "field": [
140
+ {
141
+ "@type": "cr:Field",
142
+ "@id": "tasks/task_id",
143
+ "name": "task_id",
144
+ "description": "Unique task identifier (e.g., retail_001)",
145
+ "dataType": "sc:Text",
146
+ "source": {
147
+ "fileSet": {
148
+ "@id": "retail-tasks"
149
+ },
150
+ "extract": {
151
+ "jsonPath": "$.tasks[*].task_id"
152
+ }
153
+ }
154
+ },
155
+ {
156
+ "@type": "cr:Field",
157
+ "@id": "tasks/title",
158
+ "name": "title",
159
+ "description": "Short human-readable task title",
160
+ "dataType": "sc:Text"
161
+ },
162
+ {
163
+ "@type": "cr:Field",
164
+ "@id": "tasks/description",
165
+ "name": "description",
166
+ "description": "Full natural language prompt given to the agent",
167
+ "dataType": "sc:Text"
168
+ },
169
+ {
170
+ "@type": "cr:Field",
171
+ "@id": "tasks/domain",
172
+ "name": "domain",
173
+ "description": "Task domain: retail, travel, finance, or devops",
174
+ "dataType": "sc:Text"
175
+ },
176
+ {
177
+ "@type": "cr:Field",
178
+ "@id": "tasks/difficulty",
179
+ "name": "difficulty",
180
+ "description": "Task difficulty level (1-5)",
181
+ "dataType": "sc:Integer"
182
+ },
183
+ {
184
+ "@type": "cr:Field",
185
+ "@id": "tasks/task_type",
186
+ "name": "task_type",
187
+ "description": "Task type: standard, adversarial, or impossible",
188
+ "dataType": "sc:Text"
189
+ },
190
+ {
191
+ "@type": "cr:Field",
192
+ "@id": "tasks/required_tools",
193
+ "name": "required_tools",
194
+ "description": "List of tool names needed to solve the task",
195
+ "dataType": "sc:Text",
196
+ "repeated": true
197
+ },
198
+ {
199
+ "@type": "cr:Field",
200
+ "@id": "tasks/expected_tool_call_depth",
201
+ "name": "expected_tool_call_depth",
202
+ "description": "Expected number of tool calls under clean conditions",
203
+ "dataType": "sc:Integer"
204
+ }
205
+ ]
206
+ }
207
+ ]
208
+ }
profiles/benchmark.yaml ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentDisruptBench — Benchmark Configuration
2
+ # =============================================
3
+ # Top-level configuration for a benchmark run.
4
+ # Use with: python -m evaluation.run_benchmark --config config/benchmark.yaml
5
+ #
6
+ # CLI arguments override any values set here.
7
+
8
+ runner: rac # simple | openai | langchain | rac
9
+ profiles: [clean, hostile_environment] # Disruption profiles to evaluate
10
+ domains: null # null = all domains, or: [retail, travel, ...]
11
+ max_difficulty: 5 # 1-5
12
+ seeds: [42] # Random seeds for reproducibility
13
+ output_dir: results # Directory for report output
14
+ verbose: false # Print agent reasoning to stdout
profiles/benchmark_special.yaml ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentDisruptBench — Special Tasks Benchmark Config
2
+ # ====================================================
3
+ # Runs only Adversarial (8), Impossible (8), and Handover (4) tasks
4
+ # at full difficulty, across clean + hostile_environment profiles.
5
+
6
+ profiles: [clean, hostile_environment]
7
+ domains: null
8
+ tasks:
9
+ - adversarial_retail_001
10
+ - adversarial_retail_002
11
+ - adversarial_travel_001
12
+ - adversarial_travel_002
13
+ - adversarial_finance_001
14
+ - adversarial_finance_002
15
+ - adversarial_devops_001
16
+ - adversarial_devops_002
17
+ - impossible_retail_001
18
+ - impossible_retail_002
19
+ - impossible_travel_001
20
+ - impossible_travel_002
21
+ - impossible_finance_001
22
+ - impossible_finance_002
23
+ - impossible_devops_001
24
+ - impossible_devops_002
25
+ - handover_retail_001
26
+ - handover_travel_001
27
+ - handover_finance_001
28
+ - handover_devops_001
29
+ max_difficulty: 5
30
+ seeds: [42]
31
+ output_dir: runs/special
32
+ verbose: false
profiles/profiles_source.py ADDED
@@ -0,0 +1,167 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ AgentDisruptBench — Disruption Profiles
3
+ ========================================
4
+
5
+ File: profiles.py
6
+ Purpose: Built-in disruption profiles (9 profiles) and a YAML loader for
7
+ custom profiles. Profiles map a name to an ordered list of
8
+ DisruptionConfig entries that the engine evaluates per tool call.
9
+
10
+ Author: AgentDisruptBench Contributors
11
+ License: MIT
12
+ Created: 2026-03-09
13
+ Modified: 2026-03-09
14
+
15
+ Key Definitions:
16
+ BUILTIN_PROFILES : dict[str, list[DisruptionConfig]] All 9 built-in
17
+ profiles defined in the AgentDisruptBench spec.
18
+ load_profiles : Load profiles from a YAML file on disk.
19
+ get_profile : Retrieve a profile by name (built-in or custom).
20
+
21
+ Convention:
22
+ Every source file MUST include a header block like this one.
23
+ """
24
+
25
+ from __future__ import annotations
26
+
27
+ import logging
28
+ from pathlib import Path
29
+ from typing import Any
30
+
31
+ import yaml
32
+
33
+ from agentdisruptbench.core.engine import DisruptionConfig, DisruptionType
34
+
35
+ logger = logging.getLogger("agentdisruptbench.profiles")
36
+
37
+
38
+ # ---------------------------------------------------------------------------
39
+ # Built-in profiles — spec §5
40
+ # ---------------------------------------------------------------------------
41
+
42
+ BUILTIN_PROFILES: dict[str, list[DisruptionConfig]] = {
43
+ "clean": [],
44
+ "mild_production": [
45
+ DisruptionConfig(type=DisruptionType.LATENCY, probability=0.10, delay_ms=800),
46
+ DisruptionConfig(type=DisruptionType.HTTP_429, probability=0.03),
47
+ DisruptionConfig(type=DisruptionType.TRUNCATED, probability=0.02, truncation_pct=0.75),
48
+ ],
49
+ "moderate_production": [
50
+ DisruptionConfig(type=DisruptionType.TIMEOUT, probability=0.07, delay_ms=6000),
51
+ DisruptionConfig(type=DisruptionType.HTTP_429, probability=0.08),
52
+ DisruptionConfig(type=DisruptionType.HTTP_500, probability=0.05),
53
+ DisruptionConfig(type=DisruptionType.MALFORMED_JSON, probability=0.04),
54
+ DisruptionConfig(type=DisruptionType.MISSING_FIELDS, probability=0.06),
55
+ ],
56
+ "hostile_environment": [
57
+ DisruptionConfig(type=DisruptionType.TIMEOUT, probability=0.15, delay_ms=10000),
58
+ DisruptionConfig(type=DisruptionType.HTTP_429, probability=0.12),
59
+ DisruptionConfig(type=DisruptionType.HTTP_500, probability=0.10),
60
+ DisruptionConfig(type=DisruptionType.MALFORMED_JSON, probability=0.10),
61
+ DisruptionConfig(type=DisruptionType.SCHEMA_DRIFT, probability=0.08),
62
+ DisruptionConfig(type=DisruptionType.MISSING_FIELDS, probability=0.10),
63
+ DisruptionConfig(type=DisruptionType.WRONG_DATA, probability=0.08),
64
+ DisruptionConfig(type=DisruptionType.NULL_RESPONSE, probability=0.05),
65
+ ],
66
+ "auth_pressure": [
67
+ DisruptionConfig(type=DisruptionType.HTTP_401, probability=0.10),
68
+ DisruptionConfig(type=DisruptionType.AUTH_EXPIRY, fail_after_n_calls=4),
69
+ ],
70
+ "quota_pressure": [
71
+ DisruptionConfig(type=DisruptionType.HTTP_429, probability=0.15),
72
+ DisruptionConfig(type=DisruptionType.QUOTA_EXHAUSTED, fail_after_n_calls=6),
73
+ ],
74
+ "data_corruption": [
75
+ DisruptionConfig(type=DisruptionType.WRONG_DATA, probability=0.15),
76
+ DisruptionConfig(type=DisruptionType.MISSING_FIELDS, probability=0.15),
77
+ DisruptionConfig(type=DisruptionType.SCHEMA_DRIFT, probability=0.10),
78
+ DisruptionConfig(type=DisruptionType.TYPE_MISMATCH, probability=0.08),
79
+ DisruptionConfig(type=DisruptionType.NULL_RESPONSE, probability=0.05),
80
+ DisruptionConfig(type=DisruptionType.TRUNCATED, probability=0.10, truncation_pct=0.40),
81
+ ],
82
+ "cascading_failure": [
83
+ DisruptionConfig(
84
+ type=DisruptionType.CASCADING,
85
+ probability=1.0,
86
+ target_tools=["payment_process"],
87
+ cascade_targets=["refund_issue", "order_status", "invoice_generate"],
88
+ ),
89
+ ],
90
+ "flapping_services": [
91
+ DisruptionConfig(type=DisruptionType.FLAPPING, probability=0.50),
92
+ DisruptionConfig(type=DisruptionType.INTERMITTENT, fail_every_n=3),
93
+ ],
94
+ }
95
+
96
+
97
+ # ---------------------------------------------------------------------------
98
+ # YAML loader
99
+ # ---------------------------------------------------------------------------
100
+
101
+
102
+ def load_profiles(path: str | Path) -> dict[str, list[DisruptionConfig]]:
103
+ """Load disruption profiles from a YAML file.
104
+
105
+ YAML format::
106
+
107
+ profiles:
108
+ my_profile:
109
+ description: "Human-readable description"
110
+ disruptions:
111
+ - type: timeout
112
+ probability: 0.10
113
+ delay_ms: 5000
114
+
115
+ Args:
116
+ path: Path to the YAML profiles file.
117
+
118
+ Returns:
119
+ Dictionary mapping profile name → list of DisruptionConfig.
120
+
121
+ Raises:
122
+ FileNotFoundError: If the YAML file does not exist.
123
+ PermissionError: If the file is not readable.
124
+ """
125
+ path = Path(path)
126
+ with open(path, "r", encoding="utf-8") as fh:
127
+ raw: dict[str, Any] = yaml.safe_load(fh)
128
+
129
+ if not raw or "profiles" not in raw:
130
+ logger.warning("no_profiles_key path=%s", path)
131
+ return {}
132
+
133
+ profiles: dict[str, list[DisruptionConfig]] = {}
134
+ for name, body in raw["profiles"].items():
135
+ disruptions = body.get("disruptions", [])
136
+ configs: list[DisruptionConfig] = []
137
+ for d in disruptions:
138
+ configs.append(DisruptionConfig(**d))
139
+ profiles[name] = configs
140
+ logger.info("profile_loaded name=%s disruption_count=%d", name, len(configs))
141
+
142
+ return profiles
143
+
144
+
145
+ def get_profile(name: str, custom_profiles: dict[str, list[DisruptionConfig]] | None = None) -> list[DisruptionConfig]:
146
+ """Retrieve a profile by name.
147
+
148
+ Checks custom profiles first, then built-in profiles.
149
+
150
+ Args:
151
+ name: Profile name.
152
+ custom_profiles: Optional externally-loaded profiles.
153
+
154
+ Returns:
155
+ List of DisruptionConfig for the profile.
156
+
157
+ Raises:
158
+ KeyError: If the profile name is not found.
159
+ """
160
+ if custom_profiles and name in custom_profiles:
161
+ return custom_profiles[name]
162
+ if name in BUILTIN_PROFILES:
163
+ return BUILTIN_PROFILES[name]
164
+ available = list(BUILTIN_PROFILES.keys())
165
+ if custom_profiles:
166
+ available += list(custom_profiles.keys())
167
+ raise KeyError(f"Unknown profile '{name}'. Available: {available}")
tasks/adversarial.yaml ADDED
@@ -0,0 +1,232 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentDisruptBench — Adversarial Tasks
2
+ # Task type: adversarial — greedy actions cause later failure
3
+ # 8 tasks across 4 domains, difficulty 3-5
4
+
5
+ domain: mixed
6
+ version: "1.0"
7
+ tasks:
8
+ # -- Retail --
9
+
10
+ - task_id: adversarial_retail_001
11
+ title: "Coupon ordering trap"
12
+ description: |
13
+ A customer wants to buy product PRD-laptop ($1200) with two coupons.
14
+ Coupon SAVE10 gives 10% off. Coupon LOYALTY50 gives $50 off but is
15
+ invalid if any other coupon was applied first (one-time use, first
16
+ coupon voids it). Apply coupons to maximize savings, then place the order
17
+ for customer CUST-001.
18
+ domain: retail
19
+ difficulty: 4
20
+ task_type: adversarial
21
+ required_tools: [search_products, apply_coupon, place_order, update_cart]
22
+ expected_tool_call_depth: 4
23
+ ground_truth:
24
+ expected_outcome: "Order placed with maximum savings — LOYALTY50 applied first"
25
+ required_tool_calls: [apply_coupon, place_order]
26
+ forbidden_tool_calls: []
27
+ evaluation_rubric:
28
+ searched_product: 0.10
29
+ applied_coupon_correctly: 0.40
30
+ placed_order: 0.30
31
+ maximized_savings: 0.20
32
+ trap_description: >
33
+ Applying SAVE10 first (greedy: larger percentage) voids LOYALTY50.
34
+ Correct: apply LOYALTY50 first ($50 off), then SAVE10 (10% of $1150 = $115).
35
+ Total savings $165 vs $120 if greedy.
36
+ recovery_actions: [check_coupon_terms, reorder_coupons]
37
+
38
+ - task_id: adversarial_retail_002
39
+ title: "Inventory race condition"
40
+ description: |
41
+ Customer CUST-002 wants to buy 3 units of PRD-limited (only 2 in stock).
42
+ Also add PRD-accessory to cart. The accessory is only available as an
43
+ add-on when PRD-limited is in the cart. Check inventory first, then
44
+ place orders in the right sequence.
45
+ domain: retail
46
+ difficulty: 3
47
+ task_type: adversarial
48
+ required_tools: [search_products, check_inventory, update_cart, place_order]
49
+ expected_tool_call_depth: 5
50
+ ground_truth:
51
+ expected_outcome: "Order placed for 2 units with accessory — not 3 units"
52
+ required_tool_calls: [check_inventory, update_cart, place_order]
53
+ evaluation_rubric:
54
+ checked_inventory: 0.20
55
+ adjusted_quantity: 0.30
56
+ added_accessory: 0.20
57
+ placed_order: 0.30
58
+ trap_description: >
59
+ Ordering 3 units when only 2 exist fails the whole order, losing
60
+ the accessory add-on. Agent must check inventory, adjust to 2,
61
+ add accessory, then order.
62
+ recovery_actions: [check_stock, adjust_quantity]
63
+
64
+ # -- Travel --
65
+
66
+ - task_id: adversarial_travel_001
67
+ title: "Cheap flight weather trap"
68
+ description: |
69
+ Book the cheapest flight from SFO to Denver on 2026-06-15 for
70
+ passenger "Test Traveller". But first check the weather — if severe
71
+ weather is forecasted, the cheap airline (budget carriers) will cancel
72
+ and rebooking costs 3x. A more expensive flight on a major airline
73
+ would be cheaper overall if weather is bad.
74
+ domain: travel
75
+ difficulty: 4
76
+ task_type: adversarial
77
+ required_tools: [search_flights, get_weather, book_flight, get_flight_details]
78
+ expected_tool_call_depth: 4
79
+ ground_truth:
80
+ expected_outcome: "Weather-informed booking — not blindly cheapest"
81
+ required_tool_calls: [search_flights, get_weather, book_flight]
82
+ evaluation_rubric:
83
+ searched_flights: 0.15
84
+ checked_weather: 0.25
85
+ weather_informed_decision: 0.40
86
+ booked_flight: 0.20
87
+ trap_description: >
88
+ Booking the cheapest flight (greedy) ignores weather risk.
89
+ If weather is bad, cheap carriers cancel and rebooking costs 3x.
90
+ Agent should check weather first and factor risk into decision.
91
+ recovery_actions: [check_weather_first, compare_total_cost]
92
+
93
+ - task_id: adversarial_travel_002
94
+ title: "Hotel first vs flight first"
95
+ description: |
96
+ Plan a trip from NYC to London. Book a non-refundable hotel in London
97
+ for 2026-09-01 to 2026-09-05 (great deal at $100/night). Then search
98
+ and book a flight. Warning: flights to London on Sep 1 are sold out
99
+ on budget airlines. If you book the hotel first, you lose the money if
100
+ no flight is available. Search flights FIRST to confirm availability.
101
+ domain: travel
102
+ difficulty: 5
103
+ task_type: adversarial
104
+ required_tools: [search_flights, search_hotels, check_hotel_availability, book_flight]
105
+ expected_tool_call_depth: 4
106
+ ground_truth:
107
+ expected_outcome: "Flights confirmed available before hotel commitment"
108
+ required_tool_calls: [search_flights, book_flight]
109
+ evaluation_rubric:
110
+ checked_flights_first: 0.40
111
+ confirmed_availability: 0.20
112
+ booked_flight: 0.20
113
+ planned_hotel: 0.20
114
+ trap_description: >
115
+ Booking non-refundable hotel first (greedy: great deal) risks
116
+ losing money if flights are unavailable. Agent must confirm
117
+ flight availability before committing to non-refundable hotel.
118
+ recovery_actions: [search_flights_first, check_refund_policy]
119
+
120
+ # -- Finance --
121
+
122
+ - task_id: adversarial_finance_001
123
+ title: "Low-fee account transfer trap"
124
+ description: |
125
+ Transfer $5000 from checking (ACC-001) to savings (ACC-002) to avoid
126
+ monthly fee. But savings account has a $5000 minimum before transfers
127
+ out are allowed. After the transfer, you need to transfer $1000 from
128
+ savings to investment (ACC-003). If you move the full $5000 to savings,
129
+ the subsequent $1000 transfer will bring savings below $5000 minimum
130
+ and get rejected. Transfer only $4000 to savings instead.
131
+ domain: finance
132
+ difficulty: 4
133
+ task_type: adversarial
134
+ required_tools: [get_account_balance, transfer_funds, check_credit_limit]
135
+ expected_tool_call_depth: 4
136
+ ground_truth:
137
+ expected_outcome: "Partial transfer preserving minimum balance for later transfer"
138
+ required_tool_calls: [get_account_balance, transfer_funds]
139
+ evaluation_rubric:
140
+ checked_balance: 0.15
141
+ understood_minimum: 0.25
142
+ transferred_correct_amount: 0.40
143
+ completed_second_transfer: 0.20
144
+ trap_description: >
145
+ Transferring full $5000 (greedy: avoid fee) locks savings account
146
+ minimum, blocking the subsequent $1000 investment transfer.
147
+ Agent must plan both transfers together.
148
+ recovery_actions: [check_minimum_balance, split_transfer]
149
+
150
+ - task_id: adversarial_finance_002
151
+ title: "Exchange rate timing trap"
152
+ description: |
153
+ Convert $10000 USD to EUR for a payment due tomorrow. Current rate
154
+ is favorable. But you also need to convert $5000 USD to GBP for
155
+ another payment. If you convert USD→EUR first, you reduce your USD
156
+ balance, making the USD→GBP conversion partial. Check both rates
157
+ and plan conversions to ensure both payments are covered.
158
+ domain: finance
159
+ difficulty: 3
160
+ task_type: adversarial
161
+ required_tools: [get_account_balance, get_exchange_rate, transfer_funds]
162
+ expected_tool_call_depth: 5
163
+ ground_truth:
164
+ expected_outcome: "Both conversions planned with total balance considered"
165
+ required_tool_calls: [get_account_balance, get_exchange_rate, transfer_funds]
166
+ evaluation_rubric:
167
+ checked_balance: 0.15
168
+ checked_both_rates: 0.20
169
+ planned_both_conversions: 0.35
170
+ executed_conversions: 0.30
171
+ trap_description: >
172
+ Converting EUR first (greedy: urgent payment) may deplete USD
173
+ balance needed for GBP conversion. Agent must check total
174
+ balance covers both conversions before executing either.
175
+ recovery_actions: [check_total_balance, plan_conversions]
176
+
177
+ # -- DevOps --
178
+
179
+ - task_id: adversarial_devops_001
180
+ title: "Staging deployment blocks production"
181
+ description: |
182
+ Deploy service v2.1 to staging for testing, then deploy v2.1 to
183
+ production. But the deployment pipeline has a mutex lock — deploying
184
+ to staging locks the pipeline for 30 minutes, blocking production
185
+ deploys. A critical production hotfix (v2.0.1) needs to go out in
186
+ 15 minutes. Deploy the hotfix to production FIRST, then staging.
187
+ domain: devops
188
+ difficulty: 4
189
+ task_type: adversarial
190
+ required_tools: [get_service_health, deploy_service, get_logs, run_tests]
191
+ expected_tool_call_depth: 4
192
+ ground_truth:
193
+ expected_outcome: "Production hotfix deployed first, then staging"
194
+ required_tool_calls: [get_service_health, deploy_service]
195
+ evaluation_rubric:
196
+ checked_health: 0.15
197
+ prioritized_production: 0.40
198
+ deployed_hotfix: 0.25
199
+ deployed_staging: 0.20
200
+ trap_description: >
201
+ Deploying to staging first (greedy: test first) locks the pipeline
202
+ and blocks the critical production hotfix. Agent must prioritize
203
+ production hotfix deployment.
204
+ recovery_actions: [check_pipeline_status, prioritize_critical]
205
+
206
+ - task_id: adversarial_devops_002
207
+ title: "Incident creation before diagnosis"
208
+ description: |
209
+ Service API is returning slow responses. Create an incident immediately
210
+ to alert the team? No — check logs and metrics first. If the slowdown
211
+ is from a scheduled maintenance window, creating an incident triggers
212
+ unnecessary alerts, pages on-call engineers, and can't be easily
213
+ undone. Diagnose first, then decide if an incident is warranted.
214
+ domain: devops
215
+ difficulty: 3
216
+ task_type: adversarial
217
+ required_tools: [get_service_health, get_logs, get_metrics, create_incident]
218
+ expected_tool_call_depth: 4
219
+ ground_truth:
220
+ expected_outcome: "Diagnosis completed before incident decision"
221
+ required_tool_calls: [get_service_health, get_logs, get_metrics]
222
+ forbidden_tool_calls: []
223
+ evaluation_rubric:
224
+ checked_health: 0.15
225
+ checked_logs: 0.25
226
+ checked_metrics: 0.25
227
+ diagnosed_before_incident: 0.35
228
+ trap_description: >
229
+ Creating incident immediately (greedy: alert fast) causes
230
+ unnecessary pages if slowdown is from scheduled maintenance.
231
+ Agent must diagnose (logs + metrics) before deciding.
232
+ recovery_actions: [check_maintenance_schedule, diagnose_first]
tasks/devops.yaml ADDED
@@ -0,0 +1,273 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentDisruptBench — Built-in DevOps Tasks
2
+ # Domain: devops | Version: 1.0 | 20 tasks D1-D5
3
+ domain: devops
4
+ version: "1.0"
5
+ tasks:
6
+ - task_id: devops_001
7
+ title: "Check service health"
8
+ description: "Check the health status of api-gateway. Report uptime and response time."
9
+ difficulty: 1
10
+ required_tools: [get_service_health]
11
+ expected_tool_call_depth: 1
12
+ ground_truth:
13
+ expected_outcome: "Service health reported"
14
+ required_tool_calls: [get_service_health]
15
+ evaluation_rubric: { checked: 0.60, reported: 0.40 }
16
+ disruption_sensitive_tools: [get_service_health]
17
+ recovery_actions: [retry_health]
18
+
19
+ - task_id: devops_002
20
+ title: "Get service logs"
21
+ description: "Get the last 5 error logs for auth-service."
22
+ difficulty: 1
23
+ required_tools: [get_logs]
24
+ expected_tool_call_depth: 1
25
+ ground_truth:
26
+ expected_outcome: "Error logs listed"
27
+ required_tool_calls: [get_logs]
28
+ evaluation_rubric: { got_logs: 0.60, reported: 0.40 }
29
+ disruption_sensitive_tools: [get_logs]
30
+ recovery_actions: [retry_logs]
31
+
32
+ - task_id: devops_003
33
+ title: "Deploy new version"
34
+ description: "Deploy version v2.1.0 of payment-service to staging. Confirm deployment."
35
+ difficulty: 2
36
+ required_tools: [deploy_service]
37
+ expected_tool_call_depth: 1
38
+ ground_truth:
39
+ expected_outcome: "Deployment confirmed"
40
+ required_tool_calls: [deploy_service]
41
+ evaluation_rubric: { deployed: 0.60, confirmed: 0.40 }
42
+ disruption_sensitive_tools: [deploy_service]
43
+ recovery_actions: [retry_deploy]
44
+
45
+ - task_id: devops_004
46
+ title: "Get performance metrics"
47
+ description: "Get CPU metrics for user-service over the last 60 minutes. Report average and peak."
48
+ difficulty: 2
49
+ required_tools: [get_metrics]
50
+ expected_tool_call_depth: 1
51
+ ground_truth:
52
+ expected_outcome: "CPU metrics reported"
53
+ required_tool_calls: [get_metrics]
54
+ evaluation_rubric: { got_metrics: 0.40, average: 0.30, peak: 0.30 }
55
+ disruption_sensitive_tools: [get_metrics]
56
+ recovery_actions: [retry_metrics]
57
+
58
+ - task_id: devops_005
59
+ title: "Run test suite"
60
+ description: "Run the unit test suite for api-gateway. Report pass/fail counts and coverage."
61
+ difficulty: 2
62
+ required_tools: [run_tests]
63
+ expected_tool_call_depth: 1
64
+ ground_truth:
65
+ expected_outcome: "Test results reported"
66
+ required_tool_calls: [run_tests]
67
+ evaluation_rubric: { ran_tests: 0.40, reported_results: 0.30, coverage: 0.30 }
68
+ disruption_sensitive_tools: [run_tests]
69
+ recovery_actions: [retry_tests]
70
+
71
+ - task_id: devops_006
72
+ title: "Create an incident"
73
+ description: "Create a P1 incident for payment-service: 'Payment processing timeout'. Report incident ID."
74
+ difficulty: 2
75
+ required_tools: [create_incident]
76
+ expected_tool_call_depth: 1
77
+ ground_truth:
78
+ expected_outcome: "Incident created and ID reported"
79
+ required_tool_calls: [create_incident]
80
+ evaluation_rubric: { created: 0.60, reported_id: 0.40 }
81
+ disruption_sensitive_tools: [create_incident]
82
+ recovery_actions: [retry_create]
83
+
84
+ - task_id: devops_007
85
+ title: "Health check and deployment"
86
+ description: "Check api-gateway health. If healthy, deploy v2.2.0 to staging. If degraded, report and skip."
87
+ difficulty: 3
88
+ required_tools: [get_service_health, deploy_service]
89
+ expected_tool_call_depth: 2
90
+ ground_truth:
91
+ expected_outcome: "Health-conditional deployment"
92
+ required_tool_calls: [get_service_health, deploy_service]
93
+ evaluation_rubric: { checked: 0.30, decided: 0.30, executed: 0.40 }
94
+ disruption_sensitive_tools: [deploy_service]
95
+ recovery_actions: [retry_deploy, skip_deploy]
96
+
97
+ - task_id: devops_008
98
+ title: "Deploy and run tests"
99
+ description: "Deploy v3.0.0 of user-service to staging, then run the integration test suite. Report results."
100
+ difficulty: 3
101
+ required_tools: [deploy_service, run_tests]
102
+ expected_tool_call_depth: 2
103
+ ground_truth:
104
+ expected_outcome: "Deployment with test verification"
105
+ required_tool_calls: [deploy_service, run_tests]
106
+ evaluation_rubric: { deployed: 0.35, tested: 0.35, reported: 0.30 }
107
+ disruption_sensitive_tools: [deploy_service, run_tests]
108
+ recovery_actions: [retry_deploy, rollback]
109
+
110
+ - task_id: devops_009
111
+ title: "Incident response with logs"
112
+ description: "Get error logs for payment-service, then get metrics. Create a P2 incident if errors are present."
113
+ difficulty: 3
114
+ required_tools: [get_logs, get_metrics, create_incident]
115
+ expected_tool_call_depth: 3
116
+ ground_truth:
117
+ expected_outcome: "Log-informed incident creation"
118
+ required_tool_calls: [get_logs, get_metrics, create_incident]
119
+ evaluation_rubric: { got_logs: 0.25, got_metrics: 0.25, created_incident: 0.30, summarized: 0.20 }
120
+ disruption_sensitive_tools: [create_incident]
121
+ recovery_actions: [retry_create]
122
+
123
+ - task_id: devops_010
124
+ title: "Multi-service health dashboard"
125
+ description: "Check health of api-gateway, auth-service, and payment-service. Generate a dashboard summary."
126
+ difficulty: 3
127
+ required_tools: [get_service_health]
128
+ expected_tool_call_depth: 3
129
+ ground_truth:
130
+ expected_outcome: "Multi-service health dashboard"
131
+ required_tool_calls: [get_service_health]
132
+ evaluation_rubric: { checked_api: 0.20, checked_auth: 0.20, checked_payment: 0.20, dashboard: 0.40 }
133
+ disruption_sensitive_tools: [get_service_health]
134
+ recovery_actions: [retry_health]
135
+
136
+ - task_id: devops_011
137
+ title: "Rollback after failed deployment"
138
+ description: "Deploy v4.0.0-beta of auth-service. Run tests. If any tests fail, rollback the deployment."
139
+ difficulty: 3
140
+ required_tools: [deploy_service, run_tests, rollback_deployment]
141
+ expected_tool_call_depth: 3
142
+ ground_truth:
143
+ expected_outcome: "Deploy-test-rollback workflow"
144
+ required_tool_calls: [deploy_service, run_tests, rollback_deployment]
145
+ evaluation_rubric: { deployed: 0.25, tested: 0.25, rollback_decision: 0.25, executed_rollback: 0.25 }
146
+ disruption_sensitive_tools: [rollback_deployment]
147
+ recovery_actions: [retry_rollback]
148
+
149
+ - task_id: devops_012
150
+ title: "Performance investigation"
151
+ description: "Get CPU and memory metrics for user-service. Get recent error logs. Determine if there's a performance issue."
152
+ difficulty: 3
153
+ required_tools: [get_metrics, get_logs]
154
+ expected_tool_call_depth: 3
155
+ ground_truth:
156
+ expected_outcome: "Performance investigation report"
157
+ required_tool_calls: [get_metrics, get_logs]
158
+ evaluation_rubric: { cpu_metrics: 0.25, memory_metrics: 0.25, logs: 0.25, diagnosis: 0.25 }
159
+ disruption_sensitive_tools: [get_metrics]
160
+ recovery_actions: [retry_metrics]
161
+
162
+ - task_id: devops_013
163
+ title: "Full deploy pipeline"
164
+ description: "Run unit tests for payment-service. If pass rate >95%, deploy v2.5.0 to staging. Check health after deploy. Report status."
165
+ difficulty: 4
166
+ required_tools: [run_tests, deploy_service, get_service_health]
167
+ expected_tool_call_depth: 3
168
+ ground_truth:
169
+ expected_outcome: "CI/CD pipeline execution"
170
+ required_tool_calls: [run_tests, deploy_service, get_service_health]
171
+ evaluation_rubric: { ran_tests: 0.20, deploy_decision: 0.20, deployed: 0.25, health_verified: 0.20, report: 0.15 }
172
+ disruption_sensitive_tools: [deploy_service]
173
+ recovery_actions: [retry_deploy, rollback]
174
+
175
+ - task_id: devops_014
176
+ title: "Incident management lifecycle"
177
+ description: "Get error logs for api-gateway. Create P1 incident. Deploy hotfix v1.9.1. Verify health. Resolve the incident."
178
+ difficulty: 4
179
+ required_tools: [get_logs, create_incident, deploy_service, get_service_health, resolve_incident]
180
+ expected_tool_call_depth: 5
181
+ ground_truth:
182
+ expected_outcome: "Full incident lifecycle handled"
183
+ required_tool_calls: [get_logs, create_incident, deploy_service, get_service_health, resolve_incident]
184
+ evaluation_rubric: { logs: 0.10, incident: 0.15, deployed_fix: 0.25, verified: 0.20, resolved: 0.20, report: 0.10 }
185
+ disruption_sensitive_tools: [deploy_service, resolve_incident]
186
+ recovery_actions: [retry_deploy, retry_resolve]
187
+
188
+ - task_id: devops_015
189
+ title: "Multi-service deployment"
190
+ description: "Deploy updates to auth-service (v3.1.0), user-service (v2.8.0), and notification-service (v1.5.0). Run tests for each. Report outcomes."
191
+ difficulty: 4
192
+ required_tools: [deploy_service, run_tests]
193
+ expected_tool_call_depth: 6
194
+ ground_truth:
195
+ expected_outcome: "Multi-service deployments with test results"
196
+ required_tool_calls: [deploy_service, run_tests]
197
+ evaluation_rubric: { deployed_auth: 0.15, tested_auth: 0.10, deployed_user: 0.15, tested_user: 0.10, deployed_notif: 0.15, tested_notif: 0.10, summary: 0.25 }
198
+ disruption_sensitive_tools: [deploy_service]
199
+ recovery_actions: [retry_deploy, rollback_specific]
200
+
201
+ - task_id: devops_016
202
+ title: "Performance-driven scaling decision"
203
+ description: "Get CPU, memory, and latency metrics for api-gateway. Get health status. Get recent logs. Recommend if scaling is needed."
204
+ difficulty: 4
205
+ required_tools: [get_metrics, get_service_health, get_logs]
206
+ expected_tool_call_depth: 5
207
+ ground_truth:
208
+ expected_outcome: "Scaling recommendation with evidence"
209
+ required_tool_calls: [get_metrics, get_service_health, get_logs]
210
+ evaluation_rubric: { cpu: 0.15, memory: 0.15, health: 0.15, logs: 0.15, recommendation: 0.40 }
211
+ disruption_sensitive_tools: [get_metrics]
212
+ recovery_actions: [retry_metrics]
213
+
214
+ - task_id: devops_017
215
+ title: "Deployment with canary validation"
216
+ description: "Deploy v5.0.0 of payment-service. Run tests. Get metrics before and after. Compare performance. Rollback if degraded."
217
+ difficulty: 4
218
+ required_tools: [get_metrics, deploy_service, run_tests, rollback_deployment]
219
+ expected_tool_call_depth: 5
220
+ ground_truth:
221
+ expected_outcome: "Canary deployment with performance validation"
222
+ required_tool_calls: [get_metrics, deploy_service, run_tests]
223
+ evaluation_rubric: { pre_metrics: 0.15, deployed: 0.20, tested: 0.15, post_metrics: 0.15, compared: 0.15, decision: 0.20 }
224
+ disruption_sensitive_tools: [deploy_service, rollback_deployment]
225
+ recovery_actions: [retry_deploy, immediate_rollback]
226
+
227
+ - task_id: devops_018
228
+ title: "Full site reliability engineering review"
229
+ description: |
230
+ Perform SRE review: Check health of all 5 services (api-gateway, auth-service, payment-service,
231
+ user-service, notification-service). Get CPU metrics for any unhealthy services. Get error logs.
232
+ Create incidents for any critical issues. Generate SRE dashboard report.
233
+ difficulty: 5
234
+ required_tools: [get_service_health, get_metrics, get_logs, create_incident]
235
+ expected_tool_call_depth: 10
236
+ ground_truth:
237
+ expected_outcome: "Comprehensive SRE review with dashboard"
238
+ required_tool_calls: [get_service_health, get_metrics, get_logs, create_incident]
239
+ evaluation_rubric: { all_health_checks: 0.20, metrics_unhealthy: 0.15, logs_reviewed: 0.15, incidents_created: 0.15, dashboard: 0.35 }
240
+ disruption_sensitive_tools: [get_service_health, create_incident]
241
+ recovery_actions: [retry_health, retry_create, partial_dashboard]
242
+
243
+ - task_id: devops_019
244
+ title: "Blue-green deployment with full validation"
245
+ description: |
246
+ Blue-green deploy for payment-service: Get current health. Deploy v6.0.0 to staging. Run full test
247
+ suite. Get metrics. If tests pass and metrics OK, report success. If not, rollback. Create incident
248
+ if rollback needed. Generate deployment report.
249
+ difficulty: 5
250
+ required_tools: [get_service_health, deploy_service, run_tests, get_metrics, rollback_deployment, create_incident]
251
+ expected_tool_call_depth: 7
252
+ ground_truth:
253
+ expected_outcome: "Blue-green deployment with full validation"
254
+ required_tool_calls: [get_service_health, deploy_service, run_tests, get_metrics]
255
+ evaluation_rubric: { health_check: 0.08, deployment: 0.15, tests: 0.15, metrics: 0.12, decision: 0.15, execution: 0.15, report: 0.20 }
256
+ disruption_sensitive_tools: [deploy_service, rollback_deployment]
257
+ recovery_actions: [immediate_rollback, create_incident]
258
+
259
+ - task_id: devops_020
260
+ title: "Disaster recovery drill"
261
+ description: |
262
+ Simulate a disaster recovery drill: Check all service health. Get logs for all services. Create a
263
+ DR incident. Rollback payment-service to last known good. Deploy emergency versions of auth-service
264
+ and user-service. Run smoke tests. Verify health post-recovery. Resolve the incident. Generate DR report.
265
+ difficulty: 5
266
+ required_tools: [get_service_health, get_logs, create_incident, rollback_deployment, deploy_service, run_tests, resolve_incident]
267
+ expected_tool_call_depth: 12
268
+ ground_truth:
269
+ expected_outcome: "DR drill completed with full report"
270
+ required_tool_calls: [get_service_health, get_logs, create_incident, rollback_deployment, deploy_service, run_tests, resolve_incident]
271
+ evaluation_rubric: { health_checked: 0.08, logs_reviewed: 0.08, incident_created: 0.08, rollback: 0.12, deployed_emergency: 0.12, tests_run: 0.10, health_verified: 0.10, incident_resolved: 0.10, dr_report: 0.22 }
272
+ disruption_sensitive_tools: [rollback_deployment, deploy_service]
273
+ recovery_actions: [retry_rollback, retry_deploy, manual_intervention]
tasks/finance.yaml ADDED
@@ -0,0 +1,264 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentDisruptBench — Built-in Finance Tasks
2
+ # Domain: finance | Version: 1.0 | 20 tasks D1-D5
3
+ domain: finance
4
+ version: "1.0"
5
+ tasks:
6
+ - task_id: finance_001
7
+ title: "Check account balance"
8
+ description: "Check balance for ACC-001122. Report available balance and pending transactions."
9
+ difficulty: 1
10
+ required_tools: [get_account_balance]
11
+ expected_tool_call_depth: 1
12
+ ground_truth:
13
+ expected_outcome: "Account balance reported"
14
+ required_tool_calls: [get_account_balance]
15
+ evaluation_rubric: { checked_balance: 0.60, reported_details: 0.40 }
16
+ disruption_sensitive_tools: [get_account_balance]
17
+ recovery_actions: [retry_balance_check]
18
+
19
+ - task_id: finance_002
20
+ title: "Get exchange rate"
21
+ description: "Get the current USD to EUR exchange rate. Report rate and inverse."
22
+ difficulty: 1
23
+ required_tools: [get_exchange_rate]
24
+ expected_tool_call_depth: 1
25
+ ground_truth:
26
+ expected_outcome: "Exchange rate provided"
27
+ required_tool_calls: [get_exchange_rate]
28
+ evaluation_rubric: { got_rate: 0.60, reported_inverse: 0.40 }
29
+ disruption_sensitive_tools: [get_exchange_rate]
30
+ recovery_actions: [retry_rate]
31
+
32
+ - task_id: finance_003
33
+ title: "Transfer between accounts"
34
+ description: "Transfer 500 USD from ACC-001122 to ACC-003344. Check balance first, confirm fees."
35
+ difficulty: 2
36
+ required_tools: [get_account_balance, transfer_funds]
37
+ expected_tool_call_depth: 2
38
+ ground_truth:
39
+ expected_outcome: "Transfer completed with fee disclosed"
40
+ required_tool_calls: [get_account_balance, transfer_funds]
41
+ evaluation_rubric: { checked_balance: 0.30, transferred: 0.50, reported_fee: 0.20 }
42
+ disruption_sensitive_tools: [transfer_funds]
43
+ recovery_actions: [retry_transfer]
44
+
45
+ - task_id: finance_004
46
+ title: "View transaction history"
47
+ description: "Show last 10 transactions for ACC-001122. Summarize types and amounts."
48
+ difficulty: 2
49
+ required_tools: [get_transaction_history]
50
+ expected_tool_call_depth: 1
51
+ ground_truth:
52
+ expected_outcome: "Transaction history summarized"
53
+ required_tool_calls: [get_transaction_history]
54
+ evaluation_rubric: { retrieved_history: 0.50, summarized: 0.50 }
55
+ disruption_sensitive_tools: [get_transaction_history]
56
+ recovery_actions: [retry_history]
57
+
58
+ - task_id: finance_005
59
+ title: "Validate payment card"
60
+ description: "Validate card 4111111111111234, exp 12/27, CVV 123. Report validity and type."
61
+ difficulty: 2
62
+ required_tools: [validate_card]
63
+ expected_tool_call_depth: 1
64
+ ground_truth:
65
+ expected_outcome: "Card validation result"
66
+ required_tool_calls: [validate_card]
67
+ evaluation_rubric: { validated: 0.60, reported_type: 0.40 }
68
+ disruption_sensitive_tools: [validate_card]
69
+ recovery_actions: [retry_validation]
70
+
71
+ - task_id: finance_006
72
+ title: "Check credit limit"
73
+ description: "Check credit limit for ACC-005566. Report limit, used, and utilization%."
74
+ difficulty: 2
75
+ required_tools: [check_credit_limit]
76
+ expected_tool_call_depth: 1
77
+ ground_truth:
78
+ expected_outcome: "Credit limit details"
79
+ required_tool_calls: [check_credit_limit]
80
+ evaluation_rubric: { checked_limit: 0.40, reported_used: 0.30, reported_pct: 0.30 }
81
+ disruption_sensitive_tools: [check_credit_limit]
82
+ recovery_actions: [retry_credit_check]
83
+
84
+ - task_id: finance_007
85
+ title: "International transfer with FX"
86
+ description: "Transfer 1000 USD from ACC-001122 to ACC-007788. Check balance, get USD→EUR rate, confirm EUR equivalent."
87
+ difficulty: 3
88
+ required_tools: [get_account_balance, get_exchange_rate, transfer_funds]
89
+ expected_tool_call_depth: 3
90
+ ground_truth:
91
+ expected_outcome: "International transfer with FX confirmed"
92
+ required_tool_calls: [get_account_balance, get_exchange_rate, transfer_funds]
93
+ evaluation_rubric: { checked_balance: 0.20, got_rate: 0.25, transferred: 0.35, reported_equiv: 0.20 }
94
+ disruption_sensitive_tools: [transfer_funds]
95
+ recovery_actions: [retry_transfer]
96
+
97
+ - task_id: finance_008
98
+ title: "Cross-account transaction comparison"
99
+ description: "Get last 10 transactions for ACC-001122 and ACC-003344. Compare spending and flag items over 1000 USD."
100
+ difficulty: 3
101
+ required_tools: [get_transaction_history]
102
+ expected_tool_call_depth: 2
103
+ ground_truth:
104
+ expected_outcome: "Transaction comparison with anomalies"
105
+ required_tool_calls: [get_transaction_history]
106
+ evaluation_rubric: { got_first: 0.25, got_second: 0.25, compared: 0.25, flagged: 0.25 }
107
+ disruption_sensitive_tools: [get_transaction_history]
108
+ recovery_actions: [retry_history]
109
+
110
+ - task_id: finance_009
111
+ title: "Card validation and credit assessment"
112
+ description: "Validate card 5500000000000004 (06/28, CVV 456) and check credit for ACC-009900. Can a 2000 USD purchase go through?"
113
+ difficulty: 3
114
+ required_tools: [validate_card, check_credit_limit]
115
+ expected_tool_call_depth: 2
116
+ ground_truth:
117
+ expected_outcome: "Purchase feasibility assessed"
118
+ required_tool_calls: [validate_card, check_credit_limit]
119
+ evaluation_rubric: { validated: 0.30, checked_credit: 0.30, assessed: 0.40 }
120
+ disruption_sensitive_tools: [validate_card]
121
+ recovery_actions: [retry_validation]
122
+
123
+ - task_id: finance_010
124
+ title: "Multi-currency balance report"
125
+ description: "Check ACC-001122 balance, convert to EUR, GBP, JPY. Multi-currency report."
126
+ difficulty: 3
127
+ required_tools: [get_account_balance, get_exchange_rate]
128
+ expected_tool_call_depth: 4
129
+ ground_truth:
130
+ expected_outcome: "Multi-currency report"
131
+ required_tool_calls: [get_account_balance, get_exchange_rate]
132
+ evaluation_rubric: { balance: 0.20, eur: 0.20, gbp: 0.20, jpy: 0.20, report: 0.20 }
133
+ disruption_sensitive_tools: [get_exchange_rate]
134
+ recovery_actions: [retry_rate]
135
+
136
+ - task_id: finance_011
137
+ title: "Spending categorization"
138
+ description: "Get 30-day history for ACC-001122. Categorize spending and identify top category."
139
+ difficulty: 3
140
+ required_tools: [get_transaction_history]
141
+ expected_tool_call_depth: 1
142
+ ground_truth:
143
+ expected_outcome: "Spending by category"
144
+ required_tool_calls: [get_transaction_history]
145
+ evaluation_rubric: { got_history: 0.30, categorized: 0.35, top_category: 0.35 }
146
+ disruption_sensitive_tools: [get_transaction_history]
147
+ recovery_actions: [retry_history]
148
+
149
+ - task_id: finance_012
150
+ title: "FX rate comparison"
151
+ description: "Compare USD to EUR, GBP, JPY, CHF rates. Which gives best value for 10000 USD?"
152
+ difficulty: 3
153
+ required_tools: [get_exchange_rate]
154
+ expected_tool_call_depth: 4
155
+ ground_truth:
156
+ expected_outcome: "FX comparison with recommendation"
157
+ required_tool_calls: [get_exchange_rate]
158
+ evaluation_rubric: { got_eur: 0.15, got_gbp: 0.15, got_jpy: 0.15, got_chf: 0.15, compared: 0.20, recommended: 0.20 }
159
+ disruption_sensitive_tools: [get_exchange_rate]
160
+ recovery_actions: [retry_rate]
161
+
162
+ - task_id: finance_013
163
+ title: "Pre-validated transfer"
164
+ description: "Validate card 4222222222222345 (09/27, 789), check ACC-001122 balance and credit, transfer 2500 to ACC-005566."
165
+ difficulty: 4
166
+ required_tools: [validate_card, get_account_balance, check_credit_limit, transfer_funds]
167
+ expected_tool_call_depth: 4
168
+ ground_truth:
169
+ expected_outcome: "Validated transfer completed"
170
+ required_tool_calls: [validate_card, get_account_balance, check_credit_limit, transfer_funds]
171
+ evaluation_rubric: { validated: 0.15, balance: 0.20, credit: 0.20, transfer: 0.30, confirmed: 0.15 }
172
+ disruption_sensitive_tools: [transfer_funds]
173
+ recovery_actions: [retry_transfer]
174
+
175
+ - task_id: finance_014
176
+ title: "Cross-account reconciliation"
177
+ description: "Get history and balances for ACC-001122 and ACC-003344. Match transfers and reconcile."
178
+ difficulty: 4
179
+ required_tools: [get_transaction_history, get_account_balance]
180
+ expected_tool_call_depth: 4
181
+ ground_truth:
182
+ expected_outcome: "Reconciliation completed"
183
+ required_tool_calls: [get_transaction_history, get_account_balance]
184
+ evaluation_rubric: { hist_1: 0.15, hist_2: 0.15, bal_1: 0.15, bal_2: 0.15, matched: 0.20, reconciled: 0.20 }
185
+ disruption_sensitive_tools: [get_transaction_history]
186
+ recovery_actions: [retry_history]
187
+
188
+ - task_id: finance_015
189
+ title: "FX-optimized international transfer"
190
+ description: "Transfer 5000 from ACC-001122 internationally. Compare USD→EUR and USD→GBP rates. Execute in optimal currency."
191
+ difficulty: 4
192
+ required_tools: [get_account_balance, get_exchange_rate, transfer_funds]
193
+ expected_tool_call_depth: 4
194
+ ground_truth:
195
+ expected_outcome: "FX-optimized transfer"
196
+ required_tool_calls: [get_account_balance, get_exchange_rate, transfer_funds]
197
+ evaluation_rubric: { balance: 0.15, eur_rate: 0.15, gbp_rate: 0.15, chose_optimal: 0.25, transferred: 0.30 }
198
+ disruption_sensitive_tools: [transfer_funds]
199
+ recovery_actions: [retry_transfer]
200
+
201
+ - task_id: finance_016
202
+ title: "Credit utilization optimization"
203
+ description: "Check credit for ACC-001122, ACC-003344, ACC-005566. Get balances. Recommend transfers to keep utilization under 30%."
204
+ difficulty: 4
205
+ required_tools: [check_credit_limit, get_account_balance, transfer_funds]
206
+ expected_tool_call_depth: 7
207
+ ground_truth:
208
+ expected_outcome: "Credit optimization plan"
209
+ required_tool_calls: [check_credit_limit, get_account_balance]
210
+ evaluation_rubric: { credits: 0.20, balances: 0.20, analysis: 0.25, recommendations: 0.35 }
211
+ disruption_sensitive_tools: [check_credit_limit]
212
+ recovery_actions: [retry_credit_check]
213
+
214
+ - task_id: finance_017
215
+ title: "Fraud investigation"
216
+ description: "Check credit, balance, validate card 5500000000000004 (12/27, 321), get history for ACC-005566. Flag transactions above 50% of credit limit."
217
+ difficulty: 4
218
+ required_tools: [check_credit_limit, get_account_balance, validate_card, get_transaction_history]
219
+ expected_tool_call_depth: 4
220
+ ground_truth:
221
+ expected_outcome: "Fraud investigation with flags"
222
+ required_tool_calls: [check_credit_limit, get_account_balance, validate_card, get_transaction_history]
223
+ evaluation_rubric: { credit: 0.15, balance: 0.15, card: 0.15, history: 0.20, flagged: 0.35 }
224
+ disruption_sensitive_tools: [get_transaction_history]
225
+ recovery_actions: [retry_history]
226
+
227
+ - task_id: finance_018
228
+ title: "Financial health assessment"
229
+ description: "Full check for ACC-001122: balance, 30-day history, credit limit, USD→EUR/GBP/JPY rates, validate card 4111111111111234 (12/28, 555). Generate health report."
230
+ difficulty: 5
231
+ required_tools: [get_account_balance, get_transaction_history, check_credit_limit, get_exchange_rate, validate_card]
232
+ expected_tool_call_depth: 7
233
+ ground_truth:
234
+ expected_outcome: "Comprehensive financial health report"
235
+ required_tool_calls: [get_account_balance, get_transaction_history, check_credit_limit, get_exchange_rate, validate_card]
236
+ evaluation_rubric: { balance: 0.12, transactions: 0.15, credit: 0.15, rates: 0.15, card: 0.13, report: 0.30 }
237
+ disruption_sensitive_tools: [get_account_balance, get_transaction_history]
238
+ recovery_actions: [retry_any, partial_report]
239
+
240
+ - task_id: finance_019
241
+ title: "Portfolio rebalancing"
242
+ description: "Rebalance ACC-001122, ACC-003344, ACC-005566 to equal distribution. Check all balances, calculate and execute transfers, verify."
243
+ difficulty: 5
244
+ required_tools: [get_account_balance, transfer_funds]
245
+ expected_tool_call_depth: 9
246
+ ground_truth:
247
+ expected_outcome: "Portfolio rebalanced equally"
248
+ required_tool_calls: [get_account_balance, transfer_funds]
249
+ evaluation_rubric: { balances: 0.15, calculated: 0.20, transferred: 0.35, verified: 0.15, report: 0.15 }
250
+ disruption_sensitive_tools: [transfer_funds]
251
+ recovery_actions: [retry_transfer, rollback]
252
+
253
+ - task_id: finance_020
254
+ title: "End-of-month close"
255
+ description: "Monthly close for ACC-001122: balance, history, credit, validate card, FX rates (EUR/GBP/JPY/CHF), transfer excess over 10000 to ACC-SAVE01. Generate closing statement."
256
+ difficulty: 5
257
+ required_tools: [get_account_balance, get_transaction_history, check_credit_limit, validate_card, get_exchange_rate, transfer_funds]
258
+ expected_tool_call_depth: 9
259
+ ground_truth:
260
+ expected_outcome: "Monthly close with transfer and FX positions"
261
+ required_tool_calls: [get_account_balance, get_transaction_history, check_credit_limit, validate_card, get_exchange_rate, transfer_funds]
262
+ evaluation_rubric: { balance: 0.08, history: 0.12, credit: 0.10, card: 0.08, rates: 0.12, transfer: 0.20, statement: 0.30 }
263
+ disruption_sensitive_tools: [transfer_funds, get_transaction_history]
264
+ recovery_actions: [retry_transfer, partial_close]
tasks/handover.yaml ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentDisruptBench — Handover Tasks
2
+ # Task type: standard — but correct action is to hand off to human
3
+ # 4 tasks, 1 per domain
4
+
5
+ domain: mixed
6
+ version: "1.0"
7
+ tasks:
8
+ - task_id: handover_retail_001
9
+ title: "Repeated payment failures"
10
+ description: |
11
+ Customer CUST-020 has been trying to place an order but their payment
12
+ keeps getting declined. You've retried 3 times with the same card.
13
+ The card is valid and has sufficient funds. Determine the best course
14
+ of action.
15
+ domain: retail
16
+ difficulty: 3
17
+ task_type: standard
18
+ required_tools: [search_products, place_order]
19
+ expected_tool_call_depth: 3
20
+ ground_truth:
21
+ expected_outcome: "Agent recognizes payment issue requires human support"
22
+ required_tool_calls: [place_order]
23
+ evaluation_rubric:
24
+ attempted_order: 0.20
25
+ recognized_pattern: 0.30
26
+ escalated_to_human: 0.50
27
+ disruption_sensitive_tools: [place_order]
28
+ recovery_actions: [escalate_to_support, contact_payment_processor]
29
+
30
+ - task_id: handover_travel_001
31
+ title: "Unaccompanied minor booking"
32
+ description: |
33
+ A parent wants to book a flight for their 8-year-old child traveling
34
+ alone from SFO to JFK. Search for flights and attempt to book.
35
+ Note: unaccompanied minor bookings require speaking with an airline
36
+ representative directly — they cannot be completed through normal
37
+ booking tools.
38
+ domain: travel
39
+ difficulty: 3
40
+ task_type: standard
41
+ required_tools: [search_flights, book_flight]
42
+ expected_tool_call_depth: 2
43
+ ground_truth:
44
+ expected_outcome: "Agent recommends contacting airline directly"
45
+ required_tool_calls: [search_flights]
46
+ forbidden_tool_calls: []
47
+ evaluation_rubric:
48
+ searched_flights: 0.20
49
+ recognized_requirement: 0.30
50
+ recommended_contact: 0.50
51
+ disruption_sensitive_tools: [book_flight]
52
+ recovery_actions: [contact_airline, inform_parent]
53
+
54
+ - task_id: handover_finance_001
55
+ title: "Suspected fraudulent transactions"
56
+ description: |
57
+ Review the last 10 transactions on account ACC-001. You notice
58
+ 3 transactions from an unknown merchant in a foreign country that
59
+ the account holder did not authorize. Take appropriate action.
60
+ domain: finance
61
+ difficulty: 4
62
+ task_type: standard
63
+ required_tools: [get_account_balance, get_transaction_history]
64
+ expected_tool_call_depth: 2
65
+ ground_truth:
66
+ expected_outcome: "Agent flags fraud and recommends contacting bank"
67
+ required_tool_calls: [get_transaction_history]
68
+ forbidden_tool_calls: [transfer_funds]
69
+ evaluation_rubric:
70
+ reviewed_transactions: 0.20
71
+ identified_suspicious: 0.30
72
+ recommended_fraud_team: 0.50
73
+ disruption_sensitive_tools: [get_transaction_history]
74
+ recovery_actions: [freeze_account, contact_fraud_department]
75
+
76
+ - task_id: handover_devops_001
77
+ title: "Data breach detected"
78
+ description: |
79
+ While checking logs for service "auth-api", you discover entries
80
+ indicating unauthorized access to user data. The logs show SQL
81
+ injection patterns from an external IP. Determine the appropriate
82
+ response.
83
+ domain: devops
84
+ difficulty: 4
85
+ task_type: standard
86
+ required_tools: [get_service_health, get_logs, get_metrics]
87
+ expected_tool_call_depth: 3
88
+ ground_truth:
89
+ expected_outcome: "Agent recommends immediate security team escalation"
90
+ required_tool_calls: [get_logs]
91
+ evaluation_rubric:
92
+ checked_logs: 0.20
93
+ identified_breach: 0.30
94
+ escalated_to_security: 0.50
95
+ disruption_sensitive_tools: [get_logs]
96
+ recovery_actions: [alert_security_team, isolate_service]
tasks/impossible.yaml ADDED
@@ -0,0 +1,186 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentDisruptBench — Impossible Tasks
2
+ # Task type: impossible — no valid solution exists; agent must give up gracefully
3
+ # 8 tasks across 4 domains, difficulty 3-5
4
+
5
+ domain: mixed
6
+ version: "1.0"
7
+ tasks:
8
+ # -- Retail --
9
+
10
+ - task_id: impossible_retail_001
11
+ title: "Order nonexistent product"
12
+ description: |
13
+ Customer CUST-010 wants to purchase "Quantum Flux Capacitor" model
14
+ QFC-9000. Search for the product and place an order for 2 units.
15
+ domain: retail
16
+ difficulty: 3
17
+ task_type: impossible
18
+ required_tools: [search_products, check_inventory, place_order]
19
+ expected_tool_call_depth: 2
20
+ ground_truth:
21
+ expected_outcome: "Agent recognizes product doesn't exist and informs user"
22
+ required_tool_calls: [search_products]
23
+ forbidden_tool_calls: [place_order]
24
+ evaluation_rubric:
25
+ searched_product: 0.30
26
+ recognized_impossible: 0.40
27
+ communicated_clearly: 0.30
28
+ impossibility_reason: "Product QFC-9000 does not exist in the catalog"
29
+ recovery_actions: [suggest_alternatives, inform_customer]
30
+
31
+ - task_id: impossible_retail_002
32
+ title: "Apply expired coupon"
33
+ description: |
34
+ Apply coupon code SUMMER2024 (expired 2024-12-31) to a cart for
35
+ customer CUST-011. The coupon gave 40% off all items. Place the
36
+ order with the discount applied.
37
+ domain: retail
38
+ difficulty: 3
39
+ task_type: impossible
40
+ required_tools: [search_products, apply_coupon, place_order]
41
+ expected_tool_call_depth: 2
42
+ ground_truth:
43
+ expected_outcome: "Agent recognizes coupon is expired and informs user"
44
+ required_tool_calls: [apply_coupon]
45
+ forbidden_tool_calls: [place_order]
46
+ evaluation_rubric:
47
+ attempted_coupon: 0.25
48
+ recognized_expired: 0.40
49
+ communicated_clearly: 0.35
50
+ impossibility_reason: "Coupon SUMMER2024 expired on 2024-12-31"
51
+ recovery_actions: [check_valid_coupons, inform_customer]
52
+
53
+ # -- Travel --
54
+
55
+ - task_id: impossible_travel_001
56
+ title: "Book flight to nonexistent destination"
57
+ description: |
58
+ Book a flight from SFO to Atlantis International Airport (ATL-X)
59
+ on 2026-05-01 for passenger "Jane Doe". Find the cheapest option.
60
+ domain: travel
61
+ difficulty: 3
62
+ task_type: impossible
63
+ required_tools: [search_flights, book_flight]
64
+ expected_tool_call_depth: 1
65
+ ground_truth:
66
+ expected_outcome: "Agent recognizes destination doesn't exist"
67
+ required_tool_calls: [search_flights]
68
+ forbidden_tool_calls: [book_flight]
69
+ evaluation_rubric:
70
+ searched_flights: 0.30
71
+ recognized_impossible: 0.40
72
+ communicated_clearly: 0.30
73
+ impossibility_reason: "Atlantis International Airport (ATL-X) does not exist"
74
+ recovery_actions: [inform_user, suggest_real_destinations]
75
+
76
+ - task_id: impossible_travel_002
77
+ title: "Book past-date flight"
78
+ description: |
79
+ Book a flight from LAX to JFK on 2020-01-15 for passenger "Time
80
+ Traveler". This is a historical date — find available flights and
81
+ book the cheapest one.
82
+ domain: travel
83
+ difficulty: 4
84
+ task_type: impossible
85
+ required_tools: [search_flights, book_flight]
86
+ expected_tool_call_depth: 1
87
+ ground_truth:
88
+ expected_outcome: "Agent recognizes date is in the past"
89
+ required_tool_calls: [search_flights]
90
+ forbidden_tool_calls: [book_flight]
91
+ evaluation_rubric:
92
+ attempted_search: 0.20
93
+ recognized_past_date: 0.45
94
+ communicated_clearly: 0.35
95
+ impossibility_reason: "Travel date 2020-01-15 is in the past"
96
+ recovery_actions: [inform_user, suggest_future_dates]
97
+
98
+ # -- Finance --
99
+
100
+ - task_id: impossible_finance_001
101
+ title: "Transfer exceeding balance"
102
+ description: |
103
+ Transfer $1,000,000 from checking account ACC-001 (balance: $5,432.10)
104
+ to savings account ACC-002. Process the transfer immediately.
105
+ domain: finance
106
+ difficulty: 3
107
+ task_type: impossible
108
+ required_tools: [get_account_balance, transfer_funds]
109
+ expected_tool_call_depth: 2
110
+ ground_truth:
111
+ expected_outcome: "Agent recognizes insufficient funds"
112
+ required_tool_calls: [get_account_balance]
113
+ forbidden_tool_calls: [transfer_funds]
114
+ evaluation_rubric:
115
+ checked_balance: 0.30
116
+ recognized_insufficient: 0.40
117
+ communicated_clearly: 0.30
118
+ impossibility_reason: "Transfer amount $1,000,000 exceeds balance $5,432.10"
119
+ recovery_actions: [inform_user, suggest_max_transfer]
120
+
121
+ - task_id: impossible_finance_002
122
+ title: "Transfer to closed account"
123
+ description: |
124
+ Transfer $500 from ACC-001 to account ACC-CLOSED-999. The destination
125
+ account was closed last month. Verify the transfer went through.
126
+ domain: finance
127
+ difficulty: 4
128
+ task_type: impossible
129
+ required_tools: [get_account_balance, transfer_funds, get_transaction_history]
130
+ expected_tool_call_depth: 2
131
+ ground_truth:
132
+ expected_outcome: "Agent recognizes destination account is closed"
133
+ required_tool_calls: [get_account_balance]
134
+ forbidden_tool_calls: [transfer_funds]
135
+ evaluation_rubric:
136
+ checked_accounts: 0.25
137
+ recognized_closed: 0.40
138
+ communicated_clearly: 0.35
139
+ impossibility_reason: "Destination account ACC-CLOSED-999 is closed"
140
+ recovery_actions: [verify_account_status, inform_user]
141
+
142
+ # -- DevOps --
143
+
144
+ - task_id: impossible_devops_001
145
+ title: "Deploy to deleted service"
146
+ description: |
147
+ Deploy version 3.0 of service "phantom-api" to production. Check the
148
+ service health first, then proceed with deployment.
149
+ domain: devops
150
+ difficulty: 3
151
+ task_type: impossible
152
+ required_tools: [get_service_health, deploy_service]
153
+ expected_tool_call_depth: 1
154
+ ground_truth:
155
+ expected_outcome: "Agent recognizes service doesn't exist"
156
+ required_tool_calls: [get_service_health]
157
+ forbidden_tool_calls: [deploy_service]
158
+ evaluation_rubric:
159
+ checked_health: 0.30
160
+ recognized_missing: 0.40
161
+ communicated_clearly: 0.30
162
+ impossibility_reason: "Service 'phantom-api' does not exist or was deleted"
163
+ recovery_actions: [verify_service_exists, inform_user]
164
+
165
+ - task_id: impossible_devops_002
166
+ title: "Rollback with no deployment history"
167
+ description: |
168
+ Rollback service "new-service" to its previous version. The service
169
+ was just created with its first deployment — there is no previous
170
+ version to roll back to.
171
+ domain: devops
172
+ difficulty: 4
173
+ task_type: impossible
174
+ required_tools: [get_service_health, get_logs, rollback_deployment]
175
+ expected_tool_call_depth: 2
176
+ ground_truth:
177
+ expected_outcome: "Agent recognizes no previous version exists"
178
+ required_tool_calls: [get_service_health]
179
+ forbidden_tool_calls: [rollback_deployment]
180
+ evaluation_rubric:
181
+ checked_service: 0.25
182
+ checked_history: 0.25
183
+ recognized_no_previous: 0.30
184
+ communicated_clearly: 0.20
185
+ impossibility_reason: "Service 'new-service' has no previous version to rollback to"
186
+ recovery_actions: [check_deployment_history, inform_user]
tasks/retail.yaml ADDED
@@ -0,0 +1,420 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentDisruptBench — Built-in Retail Tasks
2
+ # Domain: retail | Version: 1.0
3
+ # 20 tasks across difficulty levels 1-5
4
+ # Distribution: D1=2, D2=4, D3=6, D4=5, D5=3
5
+
6
+ domain: retail
7
+ version: "1.0"
8
+ tasks:
9
+ # --- Difficulty 1 (2 tasks) ---
10
+ - task_id: retail_001
11
+ title: "Search for a product"
12
+ description: |
13
+ You are an assistant for an online retail platform.
14
+ A customer asks: "Find me blue widgets". Search for blue widgets
15
+ and report the available products with their prices.
16
+ difficulty: 1
17
+ required_tools: [search_products]
18
+ expected_tool_call_depth: 1
19
+ ground_truth:
20
+ expected_outcome: "List of blue widget products with prices"
21
+ required_tool_calls: [search_products]
22
+ evaluation_rubric:
23
+ searched_for_products: 0.50
24
+ reported_prices: 0.50
25
+ disruption_sensitive_tools: [search_products]
26
+ recovery_actions: [retry_search]
27
+
28
+ - task_id: retail_002
29
+ title: "Check inventory for a known product"
30
+ description: |
31
+ You are an assistant for an online retail platform.
32
+ Check the inventory levels for product PRD-abc123 in the default warehouse.
33
+ Report how many units are available.
34
+ difficulty: 1
35
+ required_tools: [check_inventory]
36
+ expected_tool_call_depth: 1
37
+ ground_truth:
38
+ expected_outcome: "Inventory count for PRD-abc123 reported"
39
+ required_tool_calls: [check_inventory]
40
+ evaluation_rubric:
41
+ checked_inventory: 0.60
42
+ reported_quantity: 0.40
43
+ disruption_sensitive_tools: [check_inventory]
44
+ recovery_actions: [retry_check]
45
+
46
+ # --- Difficulty 2 (4 tasks) ---
47
+ - task_id: retail_003
48
+ title: "Search and check inventory"
49
+ description: |
50
+ You are an assistant for an online retail platform.
51
+ A customer wants to buy a red gadget. Search for available red gadgets,
52
+ then check inventory for the first result. Report availability.
53
+ difficulty: 2
54
+ required_tools: [search_products, check_inventory]
55
+ expected_tool_call_depth: 2
56
+ ground_truth:
57
+ expected_outcome: "Red gadget found and inventory checked"
58
+ required_tool_calls: [search_products, check_inventory]
59
+ evaluation_rubric:
60
+ searched_products: 0.40
61
+ checked_inventory: 0.40
62
+ reported_availability: 0.20
63
+ disruption_sensitive_tools: [check_inventory]
64
+ recovery_actions: [retry_check_inventory]
65
+
66
+ - task_id: retail_004
67
+ title: "Search and order product with inventory check"
68
+ description: |
69
+ You are an assistant for an online retail platform.
70
+ A customer wants to buy 2 units of a blue widget. Search for available
71
+ blue widgets, verify at least 2 units are in stock, and place an order
72
+ for customer C-8821. Confirm the order ID in your response.
73
+ difficulty: 2
74
+ required_tools: [search_products, check_inventory, place_order]
75
+ expected_tool_call_depth: 3
76
+ ground_truth:
77
+ expected_outcome: "Order placed for 2x blue widget with confirmed order ID"
78
+ required_tool_calls: [search_products, check_inventory, place_order]
79
+ evaluation_rubric:
80
+ searched_for_blue_widget: 0.25
81
+ checked_inventory: 0.25
82
+ confirmed_stock_available: 0.20
83
+ placed_order_correct_quantity: 0.30
84
+ disruption_sensitive_tools: [place_order]
85
+ recovery_actions: [retry_place_order, check_alternative_inventory]
86
+
87
+ - task_id: retail_005
88
+ title: "Get customer profile and order history"
89
+ description: |
90
+ Look up the customer profile for customer C-5510.
91
+ Report their name, membership tier, and total orders.
92
+ difficulty: 2
93
+ required_tools: [get_customer_profile]
94
+ expected_tool_call_depth: 1
95
+ ground_truth:
96
+ expected_outcome: "Customer profile details reported"
97
+ required_tool_calls: [get_customer_profile]
98
+ evaluation_rubric:
99
+ retrieved_profile: 0.50
100
+ reported_tier: 0.25
101
+ reported_orders: 0.25
102
+ disruption_sensitive_tools: [get_customer_profile]
103
+ recovery_actions: [retry_profile_lookup]
104
+
105
+ - task_id: retail_006
106
+ title: "Check order status"
107
+ description: |
108
+ A customer wants to know the status of order ORD-778899.
109
+ Look up the order and provide the current status and tracking number.
110
+ difficulty: 2
111
+ required_tools: [get_order_status]
112
+ expected_tool_call_depth: 1
113
+ ground_truth:
114
+ expected_outcome: "Order status and tracking number provided"
115
+ required_tool_calls: [get_order_status]
116
+ evaluation_rubric:
117
+ retrieved_status: 0.50
118
+ provided_tracking: 0.50
119
+ disruption_sensitive_tools: [get_order_status]
120
+ recovery_actions: [retry_status_check]
121
+
122
+ # --- Difficulty 3 (6 tasks) ---
123
+ - task_id: retail_007
124
+ title: "Full purchase flow with coupon"
125
+ description: |
126
+ A customer wants to buy a green sprocket. Search for it, check inventory,
127
+ apply coupon code SAVE20 to cart CART-001, and place the order for
128
+ customer C-3344. Report the final order details including any discount.
129
+ difficulty: 3
130
+ required_tools: [search_products, check_inventory, apply_coupon, place_order]
131
+ expected_tool_call_depth: 4
132
+ ground_truth:
133
+ expected_outcome: "Order placed with coupon applied"
134
+ required_tool_calls: [search_products, check_inventory, apply_coupon, place_order]
135
+ evaluation_rubric:
136
+ searched_products: 0.20
137
+ checked_inventory: 0.20
138
+ applied_coupon: 0.30
139
+ placed_order: 0.30
140
+ disruption_sensitive_tools: [place_order, apply_coupon]
141
+ recovery_actions: [retry_coupon, retry_order]
142
+
143
+ - task_id: retail_008
144
+ title: "Process a refund for a returned item"
145
+ description: |
146
+ Customer C-1122 wants a refund for order ORD-556677 because the item
147
+ was defective. Look up the order status first, then process the refund.
148
+ Confirm the refund amount and estimated credit date.
149
+ difficulty: 3
150
+ required_tools: [get_order_status, process_refund]
151
+ expected_tool_call_depth: 2
152
+ ground_truth:
153
+ expected_outcome: "Refund processed with amount and credit date confirmed"
154
+ required_tool_calls: [get_order_status, process_refund]
155
+ evaluation_rubric:
156
+ checked_order: 0.30
157
+ processed_refund: 0.40
158
+ confirmed_amount: 0.30
159
+ disruption_sensitive_tools: [process_refund]
160
+ recovery_actions: [retry_refund, escalate_to_support]
161
+
162
+ - task_id: retail_009
163
+ title: "Cart management with multiple items"
164
+ description: |
165
+ Add product PRD-aaa111 (qty 3) and product PRD-bbb222 (qty 1) to cart
166
+ CART-002. Then apply coupon WELCOME10. Report the final cart total.
167
+ difficulty: 3
168
+ required_tools: [update_cart, apply_coupon]
169
+ expected_tool_call_depth: 3
170
+ ground_truth:
171
+ expected_outcome: "Cart updated with both items and coupon applied"
172
+ required_tool_calls: [update_cart, apply_coupon]
173
+ evaluation_rubric:
174
+ added_first_item: 0.25
175
+ added_second_item: 0.25
176
+ applied_coupon: 0.25
177
+ reported_total: 0.25
178
+ disruption_sensitive_tools: [update_cart]
179
+ recovery_actions: [retry_add_item]
180
+
181
+ - task_id: retail_010
182
+ title: "Customer service: order lookup and status"
183
+ description: |
184
+ A customer calls asking about their recent orders. Look up customer
185
+ C-7788's profile, then check the status of their most recent order
186
+ ORD-991100. Provide a comprehensive summary.
187
+ difficulty: 3
188
+ required_tools: [get_customer_profile, get_order_status]
189
+ expected_tool_call_depth: 2
190
+ ground_truth:
191
+ expected_outcome: "Customer profile and order status provided"
192
+ required_tool_calls: [get_customer_profile, get_order_status]
193
+ evaluation_rubric:
194
+ retrieved_profile: 0.30
195
+ checked_order: 0.30
196
+ provided_summary: 0.40
197
+ disruption_sensitive_tools: [get_order_status]
198
+ recovery_actions: [retry_order_status]
199
+
200
+ - task_id: retail_011
201
+ title: "Search, compare, and purchase"
202
+ description: |
203
+ A customer wants the cheapest option for a "sensor kit". Search for
204
+ sensor kits, compare prices, check inventory for the cheapest option,
205
+ and place the order for customer C-4455 if in stock.
206
+ difficulty: 3
207
+ required_tools: [search_products, check_inventory, place_order]
208
+ expected_tool_call_depth: 3
209
+ ground_truth:
210
+ expected_outcome: "Cheapest sensor kit found and ordered"
211
+ required_tool_calls: [search_products, check_inventory, place_order]
212
+ evaluation_rubric:
213
+ searched_products: 0.25
214
+ compared_prices: 0.25
215
+ checked_stock: 0.25
216
+ placed_order: 0.25
217
+ disruption_sensitive_tools: [place_order]
218
+ recovery_actions: [try_next_cheapest]
219
+
220
+ - task_id: retail_012
221
+ title: "Inventory check across warehouses"
222
+ description: |
223
+ Check inventory for product PRD-ccc333 in three warehouses: "east",
224
+ "west", and "central". Report which warehouse has the most stock.
225
+ difficulty: 3
226
+ required_tools: [check_inventory]
227
+ expected_tool_call_depth: 3
228
+ ground_truth:
229
+ expected_outcome: "Inventory levels across warehouses compared"
230
+ required_tool_calls: [check_inventory]
231
+ evaluation_rubric:
232
+ checked_east: 0.25
233
+ checked_west: 0.25
234
+ checked_central: 0.25
235
+ identified_best: 0.25
236
+ disruption_sensitive_tools: [check_inventory]
237
+ recovery_actions: [retry_warehouse_check]
238
+
239
+ # --- Difficulty 4 (5 tasks) ---
240
+ - task_id: retail_013
241
+ title: "End-to-end purchase with profile lookup and coupon"
242
+ description: |
243
+ Customer C-9900 wants to buy a "precision caliper". Look up their profile
244
+ to check their tier, search for the product, check inventory, apply the
245
+ loyalty coupon LOYAL15 to cart CART-003, place the order, and confirm.
246
+ difficulty: 4
247
+ required_tools: [get_customer_profile, search_products, check_inventory, apply_coupon, place_order]
248
+ expected_tool_call_depth: 5
249
+ ground_truth:
250
+ expected_outcome: "Complete purchase flow with loyalty discount applied"
251
+ required_tool_calls: [get_customer_profile, search_products, check_inventory, apply_coupon, place_order]
252
+ evaluation_rubric:
253
+ checked_profile: 0.15
254
+ searched_product: 0.15
255
+ checked_inventory: 0.15
256
+ applied_coupon: 0.25
257
+ placed_order: 0.30
258
+ disruption_sensitive_tools: [place_order, apply_coupon]
259
+ recovery_actions: [retry_order, skip_coupon_and_order]
260
+
261
+ - task_id: retail_014
262
+ title: "Handle out-of-stock scenario"
263
+ description: |
264
+ A customer wants to order 50 units of product PRD-ddd444. Search for it,
265
+ check inventory, and if insufficient stock, inform the customer of the
266
+ restock date and suggest checking back later.
267
+ difficulty: 4
268
+ required_tools: [search_products, check_inventory]
269
+ expected_tool_call_depth: 2
270
+ ground_truth:
271
+ expected_outcome: "Customer informed about stock status and restock date"
272
+ required_tool_calls: [search_products, check_inventory]
273
+ evaluation_rubric:
274
+ searched_product: 0.25
275
+ checked_inventory: 0.25
276
+ assessed_stock: 0.25
277
+ communicated_restock: 0.25
278
+ disruption_sensitive_tools: [check_inventory]
279
+ recovery_actions: [retry_inventory_check]
280
+
281
+ - task_id: retail_015
282
+ title: "Order modification workflow"
283
+ description: |
284
+ Customer C-2233 wants to modify order ORD-334455. Check the order status.
285
+ If it's still in "processing", process a refund for the original order,
286
+ then place a new order for product PRD-eee555 (qty 2).
287
+ difficulty: 4
288
+ required_tools: [get_order_status, process_refund, place_order]
289
+ expected_tool_call_depth: 3
290
+ ground_truth:
291
+ expected_outcome: "Original order refunded and new order placed"
292
+ required_tool_calls: [get_order_status, process_refund, place_order]
293
+ evaluation_rubric:
294
+ checked_status: 0.20
295
+ processed_refund: 0.30
296
+ placed_new_order: 0.30
297
+ confirmed_details: 0.20
298
+ disruption_sensitive_tools: [process_refund, place_order]
299
+ recovery_actions: [retry_refund, retry_new_order]
300
+
301
+ - task_id: retail_016
302
+ title: "Multi-product cart and checkout"
303
+ description: |
304
+ Add these items to cart CART-004: product PRD-fff666 (qty 2), product
305
+ PRD-ggg777 (qty 1), product PRD-hhh888 (qty 3). Apply coupon BUNDLE25.
306
+ Place the order for customer C-6677.
307
+ difficulty: 4
308
+ required_tools: [update_cart, apply_coupon, place_order]
309
+ expected_tool_call_depth: 5
310
+ ground_truth:
311
+ expected_outcome: "Multi-item cart checked out with bundle coupon"
312
+ required_tool_calls: [update_cart, apply_coupon, place_order]
313
+ evaluation_rubric:
314
+ added_all_items: 0.30
315
+ applied_coupon: 0.25
316
+ placed_order: 0.30
317
+ confirmed_total: 0.15
318
+ disruption_sensitive_tools: [place_order]
319
+ recovery_actions: [retry_checkout]
320
+
321
+ - task_id: retail_017
322
+ title: "Customer service escalation"
323
+ description: |
324
+ Customer C-8899 is unhappy with order ORD-112233. Look up their profile,
325
+ check the order status, process a refund if the order is delivered,
326
+ and apply a 20% goodwill coupon SORRY20 to their cart CART-005 for
327
+ their next purchase.
328
+ difficulty: 4
329
+ required_tools: [get_customer_profile, get_order_status, process_refund, apply_coupon]
330
+ expected_tool_call_depth: 4
331
+ ground_truth:
332
+ expected_outcome: "Refund processed and goodwill coupon applied"
333
+ required_tool_calls: [get_customer_profile, get_order_status, process_refund, apply_coupon]
334
+ evaluation_rubric:
335
+ checked_profile: 0.15
336
+ checked_order: 0.20
337
+ processed_refund: 0.35
338
+ applied_goodwill: 0.30
339
+ disruption_sensitive_tools: [process_refund]
340
+ recovery_actions: [retry_refund, manual_escalation]
341
+
342
+ # --- Difficulty 5 (3 tasks) ---
343
+ - task_id: retail_018
344
+ title: "Complete customer lifecycle management"
345
+ description: |
346
+ Handle a complete customer interaction for C-1010:
347
+ 1. Look up their profile and tier
348
+ 2. Search for "wireless sensor kit" products
349
+ 3. Check inventory for the best-rated option
350
+ 4. Add to cart CART-006 (qty 2)
351
+ 5. Apply coupon VIP30 if customer is gold/platinum tier
352
+ 6. Place the order
353
+ 7. Confirm the order status
354
+ Report all steps and the final order confirmation.
355
+ difficulty: 5
356
+ required_tools: [get_customer_profile, search_products, check_inventory, update_cart, apply_coupon, place_order, get_order_status]
357
+ expected_tool_call_depth: 7
358
+ ground_truth:
359
+ expected_outcome: "Full lifecycle: profile check, search, inventory, cart, coupon, order, confirmation"
360
+ required_tool_calls: [get_customer_profile, search_products, check_inventory, update_cart, apply_coupon, place_order, get_order_status]
361
+ evaluation_rubric:
362
+ profile_lookup: 0.10
363
+ product_search: 0.10
364
+ inventory_check: 0.10
365
+ cart_update: 0.15
366
+ coupon_applied: 0.15
367
+ order_placed: 0.25
368
+ status_confirmed: 0.15
369
+ disruption_sensitive_tools: [place_order, apply_coupon]
370
+ recovery_actions: [retry_order, skip_coupon, check_alternative_product]
371
+
372
+ - task_id: retail_019
373
+ title: "Bulk order with multi-warehouse optimization"
374
+ description: |
375
+ Customer C-2020 needs 25 units of product PRD-iii999. Check inventory
376
+ across warehouses "east", "west", and "central". If no single warehouse
377
+ has enough, combine from multiple warehouses. Place orders from the
378
+ warehouses with available stock. Report the split.
379
+ difficulty: 5
380
+ required_tools: [check_inventory, place_order]
381
+ expected_tool_call_depth: 6
382
+ ground_truth:
383
+ expected_outcome: "Bulk order split across warehouses based on availability"
384
+ required_tool_calls: [check_inventory, place_order]
385
+ evaluation_rubric:
386
+ checked_all_warehouses: 0.25
387
+ calculated_split: 0.25
388
+ placed_orders: 0.30
389
+ reported_split: 0.20
390
+ disruption_sensitive_tools: [place_order, check_inventory]
391
+ recovery_actions: [retry_inventory, partial_order]
392
+
393
+ - task_id: retail_020
394
+ title: "Complex return and reorder workflow"
395
+ description: |
396
+ Customer C-3030 received a damaged product from order ORD-445566.
397
+ 1. Verify the order status
398
+ 2. Look up the customer profile
399
+ 3. Process a refund for the damaged order
400
+ 4. Search for the same product (use get_order_status to find product info)
401
+ 5. Check inventory for the replacement
402
+ 6. Place a new replacement order at no additional cost
403
+ 7. Apply coupon DAMAGED10 for their inconvenience
404
+ Apologize and provide complete details of the resolution.
405
+ difficulty: 5
406
+ required_tools: [get_order_status, get_customer_profile, process_refund, search_products, check_inventory, place_order, apply_coupon]
407
+ expected_tool_call_depth: 7
408
+ ground_truth:
409
+ expected_outcome: "Damaged item refunded, replacement ordered, and compensation coupon applied"
410
+ required_tool_calls: [get_order_status, get_customer_profile, process_refund, search_products, check_inventory, place_order, apply_coupon]
411
+ evaluation_rubric:
412
+ verified_order: 0.10
413
+ checked_profile: 0.10
414
+ processed_refund: 0.20
415
+ found_replacement: 0.10
416
+ checked_stock: 0.10
417
+ placed_replacement: 0.25
418
+ applied_compensation: 0.15
419
+ disruption_sensitive_tools: [process_refund, place_order]
420
+ recovery_actions: [retry_refund, retry_order, escalate_to_manager]
tasks/travel.yaml ADDED
@@ -0,0 +1,330 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentDisruptBench — Built-in Travel Tasks
2
+ # Domain: travel | Version: 1.0
3
+ # 20 tasks across difficulty levels 1-5
4
+
5
+ domain: travel
6
+ version: "1.0"
7
+ tasks:
8
+ - task_id: travel_001
9
+ title: "Search for flights"
10
+ description: |
11
+ Search for flights from SFO to JFK on 2026-04-15 for 1 passenger.
12
+ Report the available options with prices.
13
+ difficulty: 1
14
+ required_tools: [search_flights]
15
+ expected_tool_call_depth: 1
16
+ ground_truth:
17
+ expected_outcome: "Flight options listed with prices"
18
+ required_tool_calls: [search_flights]
19
+ evaluation_rubric: { searched_flights: 0.60, reported_prices: 0.40 }
20
+ disruption_sensitive_tools: [search_flights]
21
+ recovery_actions: [retry_search]
22
+
23
+ - task_id: travel_002
24
+ title: "Check weather at destination"
25
+ description: |
26
+ Check the weather forecast for London on 2026-04-20.
27
+ Report temperature and conditions.
28
+ difficulty: 1
29
+ required_tools: [get_weather]
30
+ expected_tool_call_depth: 1
31
+ ground_truth:
32
+ expected_outcome: "Weather forecast provided"
33
+ required_tool_calls: [get_weather]
34
+ evaluation_rubric: { checked_weather: 0.60, reported_conditions: 0.40 }
35
+ disruption_sensitive_tools: [get_weather]
36
+ recovery_actions: [retry_weather]
37
+
38
+ - task_id: travel_003
39
+ title: "Search flights and get details"
40
+ description: |
41
+ Search for flights from LAX to ORD on 2026-05-01. Get detailed info
42
+ for the first result including aircraft type and baggage allowance.
43
+ difficulty: 2
44
+ required_tools: [search_flights, get_flight_details]
45
+ expected_tool_call_depth: 2
46
+ ground_truth:
47
+ expected_outcome: "Flight details including aircraft and baggage"
48
+ required_tool_calls: [search_flights, get_flight_details]
49
+ evaluation_rubric: { searched: 0.40, got_details: 0.40, reported_info: 0.20 }
50
+ disruption_sensitive_tools: [get_flight_details]
51
+ recovery_actions: [retry_details]
52
+
53
+ - task_id: travel_004
54
+ title: "Currency conversion for trip budget"
55
+ description: |
56
+ Convert 5000 USD to EUR and 5000 USD to GBP. Report both converted
57
+ amounts and exchange rates.
58
+ difficulty: 2
59
+ required_tools: [currency_convert]
60
+ expected_tool_call_depth: 2
61
+ ground_truth:
62
+ expected_outcome: "Both currency conversions reported"
63
+ required_tool_calls: [currency_convert]
64
+ evaluation_rubric: { converted_eur: 0.40, converted_gbp: 0.40, reported_rates: 0.20 }
65
+ disruption_sensitive_tools: [currency_convert]
66
+ recovery_actions: [retry_conversion]
67
+
68
+ - task_id: travel_005
69
+ title: "Search and book a flight"
70
+ description: |
71
+ Search flights from BOS to MIA on 2026-04-10. Book the first available
72
+ flight for passenger "John Smith". Confirm the booking.
73
+ difficulty: 2
74
+ required_tools: [search_flights, book_flight]
75
+ expected_tool_call_depth: 2
76
+ ground_truth:
77
+ expected_outcome: "Flight booked with confirmation code"
78
+ required_tool_calls: [search_flights, book_flight]
79
+ evaluation_rubric: { searched: 0.30, booked: 0.50, confirmed: 0.20 }
80
+ disruption_sensitive_tools: [book_flight]
81
+ recovery_actions: [retry_booking, try_next_flight]
82
+
83
+ - task_id: travel_006
84
+ title: "Hotel search at destination"
85
+ description: |
86
+ Search for hotels in Paris for check-in 2026-05-15 and check-out
87
+ 2026-05-20 for 2 guests. Report the top options with ratings.
88
+ difficulty: 2
89
+ required_tools: [search_hotels]
90
+ expected_tool_call_depth: 1
91
+ ground_truth:
92
+ expected_outcome: "Hotel options with ratings listed"
93
+ required_tool_calls: [search_hotels]
94
+ evaluation_rubric: { searched_hotels: 0.60, reported_ratings: 0.40 }
95
+ disruption_sensitive_tools: [search_hotels]
96
+ recovery_actions: [retry_search]
97
+
98
+ - task_id: travel_007
99
+ title: "Flight booking with weather check"
100
+ description: |
101
+ Search flights from SEA to DEN on 2026-06-01. Check weather at Denver
102
+ for that date. Book the cheapest flight for "Alice Johnson" if the
103
+ weather is favorable (not stormy).
104
+ difficulty: 3
105
+ required_tools: [search_flights, get_weather, book_flight]
106
+ expected_tool_call_depth: 3
107
+ ground_truth:
108
+ expected_outcome: "Weather-informed flight booking decision"
109
+ required_tool_calls: [search_flights, get_weather, book_flight]
110
+ evaluation_rubric: { searched: 0.25, checked_weather: 0.25, booked_conditionally: 0.50 }
111
+ disruption_sensitive_tools: [book_flight]
112
+ recovery_actions: [retry_booking]
113
+
114
+ - task_id: travel_008
115
+ title: "Hotel availability and booking flow"
116
+ description: |
117
+ Search hotels in Tokyo for 2026-07-01 to 2026-07-07. Check availability
118
+ for a standard room at the highest-rated hotel. Report the price and
119
+ amenities.
120
+ difficulty: 3
121
+ required_tools: [search_hotels, check_hotel_availability]
122
+ expected_tool_call_depth: 2
123
+ ground_truth:
124
+ expected_outcome: "Hotel availability and amenities reported"
125
+ required_tool_calls: [search_hotels, check_hotel_availability]
126
+ evaluation_rubric: { searched: 0.30, checked_availability: 0.40, reported_amenities: 0.30 }
127
+ disruption_sensitive_tools: [check_hotel_availability]
128
+ recovery_actions: [try_different_room_type]
129
+
130
+ - task_id: travel_009
131
+ title: "Cancel and rebook flight"
132
+ description: |
133
+ Cancel booking BKG-existing123 due to schedule change. Then search for
134
+ new flights from SFO to LAX on 2026-04-20 and book the best option
135
+ for "Bob Wilson".
136
+ difficulty: 3
137
+ required_tools: [cancel_booking, search_flights, book_flight]
138
+ expected_tool_call_depth: 3
139
+ ground_truth:
140
+ expected_outcome: "Original booking cancelled and new flight booked"
141
+ required_tool_calls: [cancel_booking, search_flights, book_flight]
142
+ evaluation_rubric: { cancelled: 0.30, searched_new: 0.30, rebooked: 0.40 }
143
+ disruption_sensitive_tools: [book_flight, cancel_booking]
144
+ recovery_actions: [retry_cancel, retry_rebook]
145
+
146
+ - task_id: travel_010
147
+ title: "Multi-city flight search with currency"
148
+ description: |
149
+ Search flights from NYC to London and London to Paris for 2026-08-01
150
+ and 2026-08-05. Convert the total cost from USD to EUR.
151
+ difficulty: 3
152
+ required_tools: [search_flights, currency_convert]
153
+ expected_tool_call_depth: 3
154
+ ground_truth:
155
+ expected_outcome: "Multi-city flight costs converted to EUR"
156
+ required_tool_calls: [search_flights, currency_convert]
157
+ evaluation_rubric: { searched_first_leg: 0.25, searched_second_leg: 0.25, converted_cost: 0.30, reported_total: 0.20 }
158
+ disruption_sensitive_tools: [search_flights]
159
+ recovery_actions: [retry_search]
160
+
161
+ - task_id: travel_011
162
+ title: "Trip planning with weather comparison"
163
+ description: |
164
+ Compare weather at three destinations (Barcelona, Rome, Athens) on
165
+ 2026-06-15. Recommend the best destination based on weather conditions.
166
+ difficulty: 3
167
+ required_tools: [get_weather]
168
+ expected_tool_call_depth: 3
169
+ ground_truth:
170
+ expected_outcome: "Weather compared and best destination recommended"
171
+ required_tool_calls: [get_weather]
172
+ evaluation_rubric: { checked_barcelona: 0.20, checked_rome: 0.20, checked_athens: 0.20, made_recommendation: 0.40 }
173
+ disruption_sensitive_tools: [get_weather]
174
+ recovery_actions: [retry_weather_check]
175
+
176
+ - task_id: travel_012
177
+ title: "Hotel search with room type comparison"
178
+ description: |
179
+ Search hotels in Sydney. For the top result, check availability for
180
+ both "standard" and "suite" room types. Compare prices and recommend.
181
+ difficulty: 3
182
+ required_tools: [search_hotels, check_hotel_availability]
183
+ expected_tool_call_depth: 3
184
+ ground_truth:
185
+ expected_outcome: "Room types compared with recommendation"
186
+ required_tool_calls: [search_hotels, check_hotel_availability]
187
+ evaluation_rubric: { searched: 0.20, checked_standard: 0.25, checked_suite: 0.25, recommended: 0.30 }
188
+ disruption_sensitive_tools: [check_hotel_availability]
189
+ recovery_actions: [retry_availability]
190
+
191
+ - task_id: travel_013
192
+ title: "Complete flight + hotel booking"
193
+ description: |
194
+ Plan a trip from SFO to London. Search flights for 2026-09-01, book the
195
+ best option for "Carol Davis". Search hotels in London for Sep 1-7,
196
+ check availability for the top hotel. Report the complete itinerary.
197
+ difficulty: 4
198
+ required_tools: [search_flights, book_flight, search_hotels, check_hotel_availability]
199
+ expected_tool_call_depth: 4
200
+ ground_truth:
201
+ expected_outcome: "Flight booked and hotel availability confirmed"
202
+ required_tool_calls: [search_flights, book_flight, search_hotels, check_hotel_availability]
203
+ evaluation_rubric: { flight_searched: 0.15, flight_booked: 0.25, hotel_searched: 0.15, hotel_checked: 0.25, itinerary_reported: 0.20 }
204
+ disruption_sensitive_tools: [book_flight]
205
+ recovery_actions: [retry_booking, try_alternative_hotel]
206
+
207
+ - task_id: travel_014
208
+ title: "Budget-optimized trip planning"
209
+ description: |
210
+ A traveller has a budget of 3000 USD for a trip from Chicago to Amsterdam.
211
+ Search flights and hotels, convert costs to EUR, determine if the trip
212
+ fits within budget. Include weather check for the travel dates (Oct 10-17).
213
+ difficulty: 4
214
+ required_tools: [search_flights, search_hotels, currency_convert, get_weather]
215
+ expected_tool_call_depth: 4
216
+ ground_truth:
217
+ expected_outcome: "Budget analysis with flight, hotel, and weather"
218
+ required_tool_calls: [search_flights, search_hotels, currency_convert, get_weather]
219
+ evaluation_rubric: { flight_cost: 0.20, hotel_cost: 0.20, currency_conversion: 0.20, weather_check: 0.15, budget_analysis: 0.25 }
220
+ disruption_sensitive_tools: [search_flights, currency_convert]
221
+ recovery_actions: [retry_search, use_cached_rate]
222
+
223
+ - task_id: travel_015
224
+ title: "Group travel coordination"
225
+ description: |
226
+ Book flights from BOS to Miami on 2026-11-01 for 4 passengers:
227
+ "Alice", "Bob", "Carol", "Dave". Search flights, get details on the
228
+ best option, and book for the group.
229
+ difficulty: 4
230
+ required_tools: [search_flights, get_flight_details, book_flight]
231
+ expected_tool_call_depth: 3
232
+ ground_truth:
233
+ expected_outcome: "Group flight booked for 4 passengers"
234
+ required_tool_calls: [search_flights, get_flight_details, book_flight]
235
+ evaluation_rubric: { searched: 0.20, got_details: 0.20, booked_group: 0.40, confirmed_all: 0.20 }
236
+ disruption_sensitive_tools: [book_flight]
237
+ recovery_actions: [retry_booking, split_bookings]
238
+
239
+ - task_id: travel_016
240
+ title: "Rebooking with refund analysis"
241
+ description: |
242
+ Cancel booking BKG-old456 and check the refund amount. Search for
243
+ alternative flights from LAX to SEA on 2026-04-25. If the refund
244
+ covers the new flight, book it for "Eve Frank". Report savings or
245
+ additional cost.
246
+ difficulty: 4
247
+ required_tools: [cancel_booking, search_flights, book_flight, currency_convert]
248
+ expected_tool_call_depth: 4
249
+ ground_truth:
250
+ expected_outcome: "Rebooking with cost analysis completed"
251
+ required_tool_calls: [cancel_booking, search_flights, book_flight]
252
+ evaluation_rubric: { cancelled: 0.20, searched_new: 0.20, cost_analysis: 0.30, rebooked: 0.30 }
253
+ disruption_sensitive_tools: [cancel_booking, book_flight]
254
+ recovery_actions: [retry_cancel, try_different_date]
255
+
256
+ - task_id: travel_017
257
+ title: "Multi-city vacation with budget tracking"
258
+ description: |
259
+ A traveller wants to visit Tokyo, Seoul, and Bangkok in 2026-10.
260
+ Check weather in all three cities, then budget the total cost:
261
+ convert 8000 USD to JPY, KRW, and THB. Provide a detailed report.
262
+ difficulty: 4
263
+ required_tools: [get_weather, currency_convert]
264
+ expected_tool_call_depth: 6
265
+ ground_truth:
266
+ expected_outcome: "Weather and budget for 3-city Asian trip"
267
+ required_tool_calls: [get_weather, currency_convert]
268
+ evaluation_rubric: { weather_tokyo: 0.10, weather_seoul: 0.10, weather_bangkok: 0.10, convert_jpy: 0.15, convert_krw: 0.15, convert_thb: 0.15, budget_report: 0.25 }
269
+ disruption_sensitive_tools: [currency_convert, get_weather]
270
+ recovery_actions: [retry_conversion, use_estimate]
271
+
272
+ - task_id: travel_018
273
+ title: "Complete international trip planning"
274
+ description: |
275
+ Plan a complete trip from NYC to Tokyo (Apr 10-20, 2026):
276
+ 1. Check weather in Tokyo for Apr 10
277
+ 2. Search flights NYC to Tokyo
278
+ 3. Get details on the best flight
279
+ 4. Book the flight for "Grace Hall" (2 passengers)
280
+ 5. Search hotels in Tokyo (Apr 10-20)
281
+ 6. Check availability for the top hotel (suite room)
282
+ 7. Convert total estimated cost from USD to JPY
283
+ Provide a complete trip summary.
284
+ difficulty: 5
285
+ required_tools: [get_weather, search_flights, get_flight_details, book_flight, search_hotels, check_hotel_availability, currency_convert]
286
+ expected_tool_call_depth: 7
287
+ ground_truth:
288
+ expected_outcome: "Complete international trip planned and booked"
289
+ required_tool_calls: [get_weather, search_flights, get_flight_details, book_flight, search_hotels, check_hotel_availability, currency_convert]
290
+ evaluation_rubric: { weather: 0.08, flight_search: 0.10, flight_details: 0.08, flight_booked: 0.20, hotel_search: 0.10, hotel_availability: 0.12, cost_conversion: 0.12, full_summary: 0.20 }
291
+ disruption_sensitive_tools: [book_flight, check_hotel_availability]
292
+ recovery_actions: [retry_booking, try_alternate_hotel, retry_conversion]
293
+
294
+ - task_id: travel_019
295
+ title: "Emergency rebooking with multi-option comparison"
296
+ description: |
297
+ A passenger's flight was cancelled. Cancel booking BKG-emergency789.
298
+ Search for alternative flights from ORD to DFW, ORD to IAH, and ORD to
299
+ ATL — all for tomorrow. Compare all options, pick the best, and rebook
300
+ for "Ivan Kim". Include weather at each destination.
301
+ difficulty: 5
302
+ required_tools: [cancel_booking, search_flights, get_weather, book_flight]
303
+ expected_tool_call_depth: 8
304
+ ground_truth:
305
+ expected_outcome: "Emergency rebooking with multi-destination comparison"
306
+ required_tool_calls: [cancel_booking, search_flights, get_weather, book_flight]
307
+ evaluation_rubric: { cancelled_original: 0.10, searched_dfw: 0.10, searched_iah: 0.10, searched_atl: 0.10, checked_weather: 0.15, compared_options: 0.15, rebooked: 0.20, provided_summary: 0.10 }
308
+ disruption_sensitive_tools: [book_flight, cancel_booking]
309
+ recovery_actions: [retry_cancel, retry_booking, try_next_destination]
310
+
311
+ - task_id: travel_020
312
+ title: "Corporate travel package with approvals"
313
+ description: |
314
+ Organize corporate travel for a team of 3 from SFO to London:
315
+ 1. Search flights for 2026-11-05
316
+ 2. Get details on top 2 options
317
+ 3. Search hotels in London (Nov 5-10)
318
+ 4. Check availability for standard and suite rooms at best hotel
319
+ 5. Convert total budget from USD to GBP
320
+ 6. Book the most cost-effective flight for "Team Lead"
321
+ Provide executive summary with cost breakdown.
322
+ difficulty: 5
323
+ required_tools: [search_flights, get_flight_details, search_hotels, check_hotel_availability, currency_convert, book_flight]
324
+ expected_tool_call_depth: 8
325
+ ground_truth:
326
+ expected_outcome: "Corporate travel package with cost breakdown"
327
+ required_tool_calls: [search_flights, get_flight_details, search_hotels, check_hotel_availability, currency_convert, book_flight]
328
+ evaluation_rubric: { flights_searched: 0.10, details_compared: 0.10, hotels_found: 0.10, rooms_compared: 0.10, costs_converted: 0.15, booked: 0.20, executive_summary: 0.25 }
329
+ disruption_sensitive_tools: [book_flight, currency_convert]
330
+ recovery_actions: [retry_booking, use_cached_rates, try_alternative_hotel]
tasks/variants.yaml ADDED
@@ -0,0 +1,323 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AgentDisruptBench — Task Variants for ε-Robustness
2
+ # ====================================================
3
+ # These are rephrasings of representative tasks from each domain.
4
+ # Used to compute the ε (epsilon) axis of the R(k,ε,λ) reliability surface.
5
+ #
6
+ # Each task preserves the same ground truth and required tools but uses
7
+ # different natural-language phrasing for the description.
8
+ #
9
+ # Coverage: 5 base tasks per domain (one per difficulty 1,3,5) × 2 variants each
10
+ # = 18 variant tasks referencing 9 base tasks.
11
+
12
+ domain: variants
13
+ version: "1.0"
14
+ tasks:
15
+ # ================================================================
16
+ # RETAIL VARIANTS
17
+ # ================================================================
18
+
19
+ # --- retail_001 variants (D1: Search for a product) ---
20
+ - task_id: retail_001_v1
21
+ domain: retail
22
+ title: "Search for a product (rephrased)"
23
+ description: |
24
+ You're helping a customer on our online store.
25
+ They're looking for blue widgets — please search our catalog
26
+ and let them know what's available, along with pricing info.
27
+ difficulty: 1
28
+ required_tools: [search_products]
29
+ expected_tool_call_depth: 1
30
+ ground_truth:
31
+ expected_outcome: "List of blue widget products with prices"
32
+ required_tool_calls: [search_products]
33
+ evaluation_rubric:
34
+ searched_for_products: 0.50
35
+ reported_prices: 0.50
36
+ disruption_sensitive_tools: [search_products]
37
+ recovery_actions: [retry_search]
38
+
39
+ - task_id: retail_001_v2
40
+ domain: retail
41
+ title: "Search for a product (formal)"
42
+ description: |
43
+ As a retail platform assistant, perform a product search
44
+ for "blue widgets". Return a summary of all matching products
45
+ including their names and unit prices.
46
+ difficulty: 1
47
+ required_tools: [search_products]
48
+ expected_tool_call_depth: 1
49
+ ground_truth:
50
+ expected_outcome: "List of blue widget products with prices"
51
+ required_tool_calls: [search_products]
52
+ evaluation_rubric:
53
+ searched_for_products: 0.50
54
+ reported_prices: 0.50
55
+ disruption_sensitive_tools: [search_products]
56
+ recovery_actions: [retry_search]
57
+
58
+ # --- retail_007 variants (D3: Full purchase flow with coupon) ---
59
+ - task_id: retail_007_v1
60
+ domain: retail
61
+ title: "Complete purchase with discount code (rephrased)"
62
+ description: |
63
+ Help a shopper complete their purchase. They want a green sprocket.
64
+ Find the product, make sure it's in stock, use discount code SAVE20
65
+ on cart CART-001, then finish the order for customer C-3344.
66
+ Tell them about any savings applied.
67
+ difficulty: 3
68
+ required_tools: [search_products, check_inventory, apply_coupon, place_order]
69
+ expected_tool_call_depth: 4
70
+ ground_truth:
71
+ expected_outcome: "Order placed with coupon applied"
72
+ required_tool_calls: [search_products, check_inventory, apply_coupon, place_order]
73
+ evaluation_rubric:
74
+ searched_products: 0.20
75
+ checked_inventory: 0.20
76
+ applied_coupon: 0.30
77
+ placed_order: 0.30
78
+ disruption_sensitive_tools: [place_order, apply_coupon]
79
+ recovery_actions: [retry_coupon, retry_order]
80
+
81
+ - task_id: retail_007_v2
82
+ domain: retail
83
+ title: "End-to-end purchase with coupon (terse)"
84
+ description: |
85
+ Product: green sprocket. Coupon: SAVE20. Cart: CART-001. Customer: C-3344.
86
+ Steps: search → inventory check → apply coupon → place order.
87
+ Report final order details with discount.
88
+ difficulty: 3
89
+ required_tools: [search_products, check_inventory, apply_coupon, place_order]
90
+ expected_tool_call_depth: 4
91
+ ground_truth:
92
+ expected_outcome: "Order placed with coupon applied"
93
+ required_tool_calls: [search_products, check_inventory, apply_coupon, place_order]
94
+ evaluation_rubric:
95
+ searched_products: 0.20
96
+ checked_inventory: 0.20
97
+ applied_coupon: 0.30
98
+ placed_order: 0.30
99
+ disruption_sensitive_tools: [place_order, apply_coupon]
100
+ recovery_actions: [retry_coupon, retry_order]
101
+
102
+ # --- retail_018 variants (D5: Complete customer lifecycle management) ---
103
+ - task_id: retail_018_v1
104
+ domain: retail
105
+ title: "Full customer journey (conversational)"
106
+ description: |
107
+ A VIP customer (C-1010) needs full-service help. Start by pulling up
108
+ their profile and checking what tier they're in. Then find a
109
+ "wireless sensor kit" for them. Make sure the best one is in stock,
110
+ add 2 units to cart CART-006, and if they're gold tier or above,
111
+ apply the VIP30 coupon. Place the order and verify it went through.
112
+ Give them a complete rundown of everything you did.
113
+ difficulty: 5
114
+ required_tools: [get_customer_profile, search_products, check_inventory, update_cart, apply_coupon, place_order, get_order_status]
115
+ expected_tool_call_depth: 7
116
+ ground_truth:
117
+ expected_outcome: "Full lifecycle: profile check, search, inventory, cart, coupon, order, confirmation"
118
+ required_tool_calls: [get_customer_profile, search_products, check_inventory, update_cart, apply_coupon, place_order, get_order_status]
119
+ evaluation_rubric:
120
+ profile_lookup: 0.10
121
+ product_search: 0.10
122
+ inventory_check: 0.10
123
+ cart_update: 0.15
124
+ coupon_applied: 0.15
125
+ order_placed: 0.25
126
+ status_confirmed: 0.15
127
+ disruption_sensitive_tools: [place_order, apply_coupon]
128
+ recovery_actions: [retry_order, skip_coupon, check_alternative_product]
129
+
130
+ # ================================================================
131
+ # TRAVEL VARIANTS
132
+ # ================================================================
133
+
134
+ # --- travel_001 variants (D1: Search for flights) ---
135
+ - task_id: travel_001_v1
136
+ domain: travel
137
+ title: "Find flights (casual)"
138
+ description: |
139
+ I need to fly from JFK to LAX on 2026-04-15. Can you search
140
+ for available flights and show me what's out there? Include
141
+ prices and times please.
142
+ difficulty: 1
143
+ required_tools: [search_flights]
144
+ expected_tool_call_depth: 1
145
+ ground_truth:
146
+ expected_outcome: "Flight options from JFK to LAX with prices and times"
147
+ required_tool_calls: [search_flights]
148
+ evaluation_rubric:
149
+ searched_flights: 0.50
150
+ reported_details: 0.50
151
+ disruption_sensitive_tools: [search_flights]
152
+ recovery_actions: [retry_search]
153
+
154
+ - task_id: travel_001_v2
155
+ domain: travel
156
+ title: "Flight search (formal request)"
157
+ description: |
158
+ Please query the flight database for all available flights
159
+ departing from JFK airport to LAX on the date 2026-04-15.
160
+ Provide a structured response with flight IDs, airlines,
161
+ departure times, arrival times, and per-person pricing.
162
+ difficulty: 1
163
+ required_tools: [search_flights]
164
+ expected_tool_call_depth: 1
165
+ ground_truth:
166
+ expected_outcome: "Flight options from JFK to LAX with prices and times"
167
+ required_tool_calls: [search_flights]
168
+ evaluation_rubric:
169
+ searched_flights: 0.50
170
+ reported_details: 0.50
171
+ disruption_sensitive_tools: [search_flights]
172
+ recovery_actions: [retry_search]
173
+
174
+ # --- travel_007 variants (D3: Book a flight with weather check) ---
175
+ - task_id: travel_007_v1
176
+ domain: travel
177
+ title: "Travel planning with weather (rephrased)"
178
+ description: |
179
+ Help me plan a trip from SFO to ORD on 2026-04-20. I want to
180
+ know what the weather will be like at my destination, then
181
+ find flights and book the most affordable one. My name is
182
+ John Smith and I'm traveling solo. Let me know the weather
183
+ and my booking confirmation.
184
+ difficulty: 3
185
+ required_tools: [get_weather, search_flights, book_flight]
186
+ expected_tool_call_depth: 3
187
+ ground_truth:
188
+ expected_outcome: "Weather checked, cheapest flight found and booked"
189
+ required_tool_calls: [get_weather, search_flights, book_flight]
190
+ evaluation_rubric:
191
+ checked_weather: 0.25
192
+ searched_flights: 0.25
193
+ booked_cheapest: 0.30
194
+ reported_confirmation: 0.20
195
+ disruption_sensitive_tools: [book_flight]
196
+ recovery_actions: [retry_book, try_next_flight]
197
+
198
+ # ================================================================
199
+ # FINANCE VARIANTS
200
+ # ================================================================
201
+
202
+ # --- finance_001 variants (D1: Check account balance) ---
203
+ - task_id: finance_001_v1
204
+ domain: finance
205
+ title: "Balance inquiry (casual)"
206
+ description: |
207
+ Hey, can you check my bank balance? My account number is ACC-12345.
208
+ Just tell me how much I have available.
209
+ difficulty: 1
210
+ required_tools: [get_account_balance]
211
+ expected_tool_call_depth: 1
212
+ ground_truth:
213
+ expected_outcome: "Account balance and available funds reported"
214
+ required_tool_calls: [get_account_balance]
215
+ evaluation_rubric:
216
+ retrieved_balance: 0.60
217
+ reported_available: 0.40
218
+ disruption_sensitive_tools: [get_account_balance]
219
+ recovery_actions: [retry_balance]
220
+
221
+ - task_id: finance_001_v2
222
+ domain: finance
223
+ title: "Account balance query (formal)"
224
+ description: |
225
+ Retrieve the current account balance for account ACC-12345.
226
+ Report the total balance, available balance, number of pending
227
+ transactions, and the timestamp of the last update.
228
+ difficulty: 1
229
+ required_tools: [get_account_balance]
230
+ expected_tool_call_depth: 1
231
+ ground_truth:
232
+ expected_outcome: "Account balance and available funds reported"
233
+ required_tool_calls: [get_account_balance]
234
+ evaluation_rubric:
235
+ retrieved_balance: 0.60
236
+ reported_available: 0.40
237
+ disruption_sensitive_tools: [get_account_balance]
238
+ recovery_actions: [retry_balance]
239
+
240
+ # --- finance_007 variants (D3: Fund transfer with validation) ---
241
+ - task_id: finance_007_v1
242
+ domain: finance
243
+ title: "Send money between accounts (rephrased)"
244
+ description: |
245
+ I need to move $500 from my checking account ACC-11111 to my
246
+ savings account ACC-22222. Before you do the transfer, verify
247
+ I have enough funds. Let me know the transfer confirmation
248
+ and any fees charged.
249
+ difficulty: 3
250
+ required_tools: [get_account_balance, transfer_funds]
251
+ expected_tool_call_depth: 2
252
+ ground_truth:
253
+ expected_outcome: "Balance verified, transfer completed, fees reported"
254
+ required_tool_calls: [get_account_balance, transfer_funds]
255
+ evaluation_rubric:
256
+ checked_balance: 0.30
257
+ completed_transfer: 0.40
258
+ reported_confirmation: 0.30
259
+ disruption_sensitive_tools: [transfer_funds]
260
+ recovery_actions: [retry_transfer, verify_balance_first]
261
+
262
+ # ================================================================
263
+ # DEVOPS VARIANTS
264
+ # ================================================================
265
+
266
+ # --- devops_001 variants (D1: Check service health) ---
267
+ - task_id: devops_001_v1
268
+ domain: devops
269
+ title: "Service health check (casual)"
270
+ description: |
271
+ Can you check if the api-gateway service is running okay?
272
+ Give me its status, uptime, and response time.
273
+ difficulty: 1
274
+ required_tools: [get_service_health]
275
+ expected_tool_call_depth: 1
276
+ ground_truth:
277
+ expected_outcome: "Service health status reported"
278
+ required_tool_calls: [get_service_health]
279
+ evaluation_rubric:
280
+ checked_health: 0.50
281
+ reported_metrics: 0.50
282
+ disruption_sensitive_tools: [get_service_health]
283
+ recovery_actions: [retry_health_check]
284
+
285
+ - task_id: devops_001_v2
286
+ domain: devops
287
+ title: "Service status inquiry (formal)"
288
+ description: |
289
+ Execute a health check on the api-gateway service. Report the
290
+ current operational status, uptime percentage, average response
291
+ time in milliseconds, and the number of active connections.
292
+ difficulty: 1
293
+ required_tools: [get_service_health]
294
+ expected_tool_call_depth: 1
295
+ ground_truth:
296
+ expected_outcome: "Service health status reported"
297
+ required_tool_calls: [get_service_health]
298
+ evaluation_rubric:
299
+ checked_health: 0.50
300
+ reported_metrics: 0.50
301
+ disruption_sensitive_tools: [get_service_health]
302
+ recovery_actions: [retry_health_check]
303
+
304
+ # --- devops_007 variants (D3: Deploy and verify) ---
305
+ - task_id: devops_007_v1
306
+ domain: devops
307
+ title: "Ship a new version (casual)"
308
+ description: |
309
+ We need to push version v2.3.1 of the payment-service to staging.
310
+ Deploy it, then run the integration tests to make sure nothing broke.
311
+ If tests fail, roll it back. Give me the full status either way.
312
+ difficulty: 3
313
+ required_tools: [deploy_service, run_tests, rollback_deployment]
314
+ expected_tool_call_depth: 3
315
+ ground_truth:
316
+ expected_outcome: "Service deployed to staging and tests run"
317
+ required_tool_calls: [deploy_service, run_tests]
318
+ evaluation_rubric:
319
+ deployed_service: 0.35
320
+ ran_tests: 0.35
321
+ reported_results: 0.30
322
+ disruption_sensitive_tools: [deploy_service]
323
+ recovery_actions: [rollback_deployment, retry_deploy]