id
stringlengths
16
26
scenario
stringclasses
8 values
messages
listlengths
30
150
token_budget
int64
350
900
token_count_estimate
int64
417
2.54k
token_count_is_estimate
bool
1 class
strategy
stringclasses
3 values
pinned_message_ids
listlengths
1
1
protected_tool_groups
listlengths
1
19
evicted_message_ids
listlengths
0
136
retained_message_ids
listlengths
14
51
expected_behavior
stringclasses
8 values
notes
stringclasses
1 value
coding-agent-001
coding-agent
[ { "id": "msg_coding-agent_0", "role": "system", "content": "You are a coding assistant. Follow the user's requirements and preserve important project constraints (the app uses JWTs with refresh tokens, not sessions).", "pinned": true }, { "id": "msg_coding-agent_1", "role": "assistant", ...
800
464
true
priority
[ "msg_coding-agent_0" ]
[ { "call_id": "msg_coding-agent_13", "result_id": "msg_coding-agent_14" }, { "call_id": "msg_coding-agent_25", "result_id": "msg_coding-agent_26" } ]
[]
[ "msg_coding-agent_0", "msg_coding-agent_1", "msg_coding-agent_2", "msg_coding-agent_3", "msg_coding-agent_4", "msg_coding-agent_5", "msg_coding-agent_6", "msg_coding-agent_7", "msg_coding-agent_8", "msg_coding-agent_9", "msg_coding-agent_10", "msg_coding-agent_11", "msg_coding-agent_12", "...
Recent, high-priority implementation details (JWT/refresh-token decisions) survive; low-value filler acknowledgements are evicted first.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
coding-agent-002
coding-agent
[ { "id": "msg_coding-agent_0", "role": "system", "content": "You are a coding assistant. Follow the user's requirements and preserve important project constraints (the app uses JWTs with refresh tokens, not sessions).", "pinned": true }, { "id": "msg_coding-agent_1", "role": "user", "...
800
1,251
true
priority
[ "msg_coding-agent_0" ]
[ { "call_id": "msg_coding-agent_13", "result_id": "msg_coding-agent_14" }, { "call_id": "msg_coding-agent_22", "result_id": "msg_coding-agent_23" }, { "call_id": "msg_coding-agent_28", "result_id": "msg_coding-agent_29" }, { "call_id": "msg_coding-agent_36", "result_id": "...
[ "msg_coding-agent_1", "msg_coding-agent_2", "msg_coding-agent_4", "msg_coding-agent_5", "msg_coding-agent_6", "msg_coding-agent_7", "msg_coding-agent_9", "msg_coding-agent_10", "msg_coding-agent_11", "msg_coding-agent_12", "msg_coding-agent_15", "msg_coding-agent_16", "msg_coding-agent_17", ...
[ "msg_coding-agent_0", "msg_coding-agent_3", "msg_coding-agent_8", "msg_coding-agent_13", "msg_coding-agent_14", "msg_coding-agent_19", "msg_coding-agent_20", "msg_coding-agent_22", "msg_coding-agent_23", "msg_coding-agent_24", "msg_coding-agent_28", "msg_coding-agent_29", "msg_coding-agent_3...
Recent, high-priority implementation details (JWT/refresh-token decisions) survive; low-value filler acknowledgements are evicted first.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
coding-agent-003
coding-agent
[ { "id": "msg_coding-agent_0", "role": "system", "content": "You are a coding assistant. Follow the user's requirements and preserve important project constraints (the app uses JWTs with refresh tokens, not sessions).", "pinned": true }, { "id": "msg_coding-agent_1", "role": "user", "...
800
2,400
true
priority
[ "msg_coding-agent_0" ]
[ { "call_id": "msg_coding-agent_13", "result_id": "msg_coding-agent_14" }, { "call_id": "msg_coding-agent_31", "result_id": "msg_coding-agent_32" }, { "call_id": "msg_coding-agent_40", "result_id": "msg_coding-agent_41" }, { "call_id": "msg_coding-agent_45", "result_id": "...
[ "msg_coding-agent_2", "msg_coding-agent_3", "msg_coding-agent_4", "msg_coding-agent_5", "msg_coding-agent_7", "msg_coding-agent_8", "msg_coding-agent_9", "msg_coding-agent_10", "msg_coding-agent_11", "msg_coding-agent_15", "msg_coding-agent_16", "msg_coding-agent_17", "msg_coding-agent_19", ...
[ "msg_coding-agent_0", "msg_coding-agent_1", "msg_coding-agent_6", "msg_coding-agent_12", "msg_coding-agent_13", "msg_coding-agent_14", "msg_coding-agent_18", "msg_coding-agent_26", "msg_coding-agent_28", "msg_coding-agent_37", "msg_coding-agent_38", "msg_coding-agent_42", "msg_coding-agent_4...
Recent, high-priority implementation details (JWT/refresh-token decisions) survive; low-value filler acknowledgements are evicted first.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
customer-support-001
customer-support
[ { "id": "msg_customer-support_0", "role": "system", "content": "You are a support agent for Acme Cloud. Be concise, and never share account credentials in plain text.", "pinned": true }, { "id": "msg_customer-support_1", "role": "assistant", "content": "Understood.", "priority": ...
500
430
true
slidingWindow
[ "msg_customer-support_0" ]
[ { "call_id": "msg_customer-support_10", "result_id": "msg_customer-support_11" }, { "call_id": "msg_customer-support_14", "result_id": "msg_customer-support_15" }, { "call_id": "msg_customer-support_18", "result_id": "msg_customer-support_19" } ]
[ "msg_customer-support_1", "msg_customer-support_2", "msg_customer-support_3", "msg_customer-support_4", "msg_customer-support_5", "msg_customer-support_6", "msg_customer-support_7", "msg_customer-support_8", "msg_customer-support_9", "msg_customer-support_10", "msg_customer-support_11", "msg_c...
[ "msg_customer-support_0", "msg_customer-support_17", "msg_customer-support_18", "msg_customer-support_19", "msg_customer-support_20", "msg_customer-support_21", "msg_customer-support_22", "msg_customer-support_23", "msg_customer-support_24", "msg_customer-support_25", "msg_customer-support_26", ...
Recent useful context (the actual issue and its resolution) remains; repetitive greeting/filler turns are evicted.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
customer-support-002
customer-support
[ { "id": "msg_customer-support_0", "role": "system", "content": "You are a support agent for Acme Cloud. Be concise, and never share account credentials in plain text.", "pinned": true }, { "id": "msg_customer-support_1", "role": "user", "content": "How long will the refund take to ap...
500
1,143
true
slidingWindow
[ "msg_customer-support_0" ]
[ { "call_id": "msg_customer-support_6", "result_id": "msg_customer-support_7" }, { "call_id": "msg_customer-support_9", "result_id": "msg_customer-support_10" }, { "call_id": "msg_customer-support_19", "result_id": "msg_customer-support_20" }, { "call_id": "msg_customer-suppor...
[ "msg_customer-support_1", "msg_customer-support_2", "msg_customer-support_3", "msg_customer-support_4", "msg_customer-support_5", "msg_customer-support_6", "msg_customer-support_7", "msg_customer-support_8", "msg_customer-support_9", "msg_customer-support_10", "msg_customer-support_11", "msg_c...
[ "msg_customer-support_0", "msg_customer-support_67", "msg_customer-support_68", "msg_customer-support_69", "msg_customer-support_70", "msg_customer-support_71", "msg_customer-support_72", "msg_customer-support_73", "msg_customer-support_74", "msg_customer-support_75", "msg_customer-support_76", ...
Recent useful context (the actual issue and its resolution) remains; repetitive greeting/filler turns are evicted.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
customer-support-003
customer-support
[ { "id": "msg_customer-support_0", "role": "system", "content": "You are a support agent for Acme Cloud. Be concise, and never share account credentials in plain text.", "pinned": true }, { "id": "msg_customer-support_1", "role": "user", "content": "Hi, I was double-charged on my last...
500
2,216
true
slidingWindow
[ "msg_customer-support_0" ]
[ { "call_id": "msg_customer-support_5", "result_id": "msg_customer-support_6" }, { "call_id": "msg_customer-support_11", "result_id": "msg_customer-support_12" }, { "call_id": "msg_customer-support_32", "result_id": "msg_customer-support_33" }, { "call_id": "msg_customer-suppo...
[ "msg_customer-support_1", "msg_customer-support_2", "msg_customer-support_3", "msg_customer-support_4", "msg_customer-support_5", "msg_customer-support_6", "msg_customer-support_7", "msg_customer-support_8", "msg_customer-support_9", "msg_customer-support_10", "msg_customer-support_11", "msg_c...
[ "msg_customer-support_0", "msg_customer-support_137", "msg_customer-support_138", "msg_customer-support_139", "msg_customer-support_140", "msg_customer-support_141", "msg_customer-support_142", "msg_customer-support_143", "msg_customer-support_144", "msg_customer-support_145", "msg_customer-supp...
Recent useful context (the actual issue and its resolution) remains; repetitive greeting/filler turns are evicted.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
long-running-agent-001
long-running-agent
[ { "id": "msg_long-running-agent_0", "role": "system", "content": "You are a long-running autonomous agent. Track the current task state and never lose sight of the original objective.", "pinned": true }, { "id": "msg_long-running-agent_1", "role": "user", "content": "Current objectiv...
700
502
true
priority
[ "msg_long-running-agent_0" ]
[ { "call_id": "msg_long-running-agent_16", "result_id": "msg_long-running-agent_17" } ]
[]
[ "msg_long-running-agent_0", "msg_long-running-agent_1", "msg_long-running-agent_2", "msg_long-running-agent_3", "msg_long-running-agent_4", "msg_long-running-agent_5", "msg_long-running-agent_6", "msg_long-running-agent_7", "msg_long-running-agent_8", "msg_long-running-agent_9", "msg_long-runnin...
The current task-state messages survive across many turns; the original objective is never lost because it is high priority.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
long-running-agent-002
long-running-agent
[ { "id": "msg_long-running-agent_0", "role": "system", "content": "You are a long-running autonomous agent. Track the current task state and never lose sight of the original objective.", "pinned": true }, { "id": "msg_long-running-agent_1", "role": "user", "content": "Reminder: the or...
700
1,358
true
priority
[ "msg_long-running-agent_0" ]
[ { "call_id": "msg_long-running-agent_12", "result_id": "msg_long-running-agent_13" }, { "call_id": "msg_long-running-agent_17", "result_id": "msg_long-running-agent_18" }, { "call_id": "msg_long-running-agent_19", "result_id": "msg_long-running-agent_20" }, { "call_id": "msg_...
[ "msg_long-running-agent_1", "msg_long-running-agent_2", "msg_long-running-agent_4", "msg_long-running-agent_5", "msg_long-running-agent_6", "msg_long-running-agent_9", "msg_long-running-agent_10", "msg_long-running-agent_11", "msg_long-running-agent_14", "msg_long-running-agent_16", "msg_long-ru...
[ "msg_long-running-agent_0", "msg_long-running-agent_3", "msg_long-running-agent_7", "msg_long-running-agent_8", "msg_long-running-agent_12", "msg_long-running-agent_13", "msg_long-running-agent_15", "msg_long-running-agent_17", "msg_long-running-agent_18", "msg_long-running-agent_19", "msg_long-...
The current task-state messages survive across many turns; the original objective is never lost because it is high priority.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
long-running-agent-003
long-running-agent
[ { "id": "msg_long-running-agent_0", "role": "system", "content": "You are a long-running autonomous agent. Track the current task state and never lose sight of the original objective.", "pinned": true }, { "id": "msg_long-running-agent_1", "role": "user", "content": "Step 3 of 7 comp...
700
2,500
true
priority
[ "msg_long-running-agent_0" ]
[ { "call_id": "msg_long-running-agent_10", "result_id": "msg_long-running-agent_11" }, { "call_id": "msg_long-running-agent_14", "result_id": "msg_long-running-agent_15" }, { "call_id": "msg_long-running-agent_16", "result_id": "msg_long-running-agent_17" }, { "call_id": "msg_...
[ "msg_long-running-agent_1", "msg_long-running-agent_2", "msg_long-running-agent_3", "msg_long-running-agent_6", "msg_long-running-agent_7", "msg_long-running-agent_8", "msg_long-running-agent_9", "msg_long-running-agent_12", "msg_long-running-agent_13", "msg_long-running-agent_14", "msg_long-run...
[ "msg_long-running-agent_0", "msg_long-running-agent_4", "msg_long-running-agent_5", "msg_long-running-agent_10", "msg_long-running-agent_11", "msg_long-running-agent_27", "msg_long-running-agent_38", "msg_long-running-agent_39", "msg_long-running-agent_40", "msg_long-running-agent_44", "msg_long...
The current task-state messages survive across many turns; the original objective is never lost because it is high priority.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
pinned-instruction-001
pinned-instruction
[ { "id": "msg_pinned-instruction_0", "role": "system", "content": "You are a secure production coding assistant. Never expose credentials, API keys, or secrets in any response.", "pinned": true }, { "id": "msg_pinned-instruction_1", "role": "user", "content": "I can't share credential...
400
574
true
dropOldest
[ "msg_pinned-instruction_0" ]
[ { "call_id": "msg_pinned-instruction_10", "result_id": "msg_pinned-instruction_11" } ]
[ "msg_pinned-instruction_1", "msg_pinned-instruction_2", "msg_pinned-instruction_3", "msg_pinned-instruction_4", "msg_pinned-instruction_5", "msg_pinned-instruction_6", "msg_pinned-instruction_7", "msg_pinned-instruction_8", "msg_pinned-instruction_9" ]
[ "msg_pinned-instruction_0", "msg_pinned-instruction_10", "msg_pinned-instruction_11", "msg_pinned-instruction_12", "msg_pinned-instruction_13", "msg_pinned-instruction_14", "msg_pinned-instruction_15", "msg_pinned-instruction_16", "msg_pinned-instruction_17", "msg_pinned-instruction_18", "msg_pi...
The pinned system instruction survives eviction regardless of how small the budget is, even when hundreds of other messages do not.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
pinned-instruction-002
pinned-instruction
[ { "id": "msg_pinned-instruction_0", "role": "system", "content": "You are a secure production coding assistant. Never expose credentials, API keys, or secrets in any response.", "pinned": true }, { "id": "msg_pinned-instruction_1", "role": "user", "content": "Understood — use the sec...
400
1,314
true
dropOldest
[ "msg_pinned-instruction_0" ]
[ { "call_id": "msg_pinned-instruction_15", "result_id": "msg_pinned-instruction_16" }, { "call_id": "msg_pinned-instruction_22", "result_id": "msg_pinned-instruction_23" }, { "call_id": "msg_pinned-instruction_27", "result_id": "msg_pinned-instruction_28" }, { "call_id": "msg_...
[ "msg_pinned-instruction_1", "msg_pinned-instruction_2", "msg_pinned-instruction_3", "msg_pinned-instruction_4", "msg_pinned-instruction_5", "msg_pinned-instruction_6", "msg_pinned-instruction_7", "msg_pinned-instruction_8", "msg_pinned-instruction_9", "msg_pinned-instruction_10", "msg_pinned-ins...
[ "msg_pinned-instruction_0", "msg_pinned-instruction_60", "msg_pinned-instruction_61", "msg_pinned-instruction_62", "msg_pinned-instruction_63", "msg_pinned-instruction_64", "msg_pinned-instruction_65", "msg_pinned-instruction_66", "msg_pinned-instruction_67", "msg_pinned-instruction_68", "msg_pi...
The pinned system instruction survives eviction regardless of how small the budget is, even when hundreds of other messages do not.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
pinned-instruction-003
pinned-instruction
[ { "id": "msg_pinned-instruction_0", "role": "system", "content": "You are a secure production coding assistant. Never expose credentials, API keys, or secrets in any response.", "pinned": true }, { "id": "msg_pinned-instruction_1", "role": "user", "content": "Can you show me the curr...
400
2,380
true
dropOldest
[ "msg_pinned-instruction_0" ]
[ { "call_id": "msg_pinned-instruction_11", "result_id": "msg_pinned-instruction_12" }, { "call_id": "msg_pinned-instruction_19", "result_id": "msg_pinned-instruction_20" }, { "call_id": "msg_pinned-instruction_22", "result_id": "msg_pinned-instruction_23" }, { "call_id": "msg_...
[ "msg_pinned-instruction_1", "msg_pinned-instruction_2", "msg_pinned-instruction_3", "msg_pinned-instruction_4", "msg_pinned-instruction_5", "msg_pinned-instruction_6", "msg_pinned-instruction_7", "msg_pinned-instruction_8", "msg_pinned-instruction_9", "msg_pinned-instruction_10", "msg_pinned-ins...
[ "msg_pinned-instruction_0", "msg_pinned-instruction_128", "msg_pinned-instruction_129", "msg_pinned-instruction_130", "msg_pinned-instruction_131", "msg_pinned-instruction_132", "msg_pinned-instruction_133", "msg_pinned-instruction_134", "msg_pinned-instruction_135", "msg_pinned-instruction_136", ...
The pinned system instruction survives eviction regardless of how small the budget is, even when hundreds of other messages do not.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
priority-based-context-001
priority-based-context
[ { "id": "msg_priority-based-context_0", "role": "system", "content": "You are an assistant prioritizing the current task above historical discussion.", "pinned": true }, { "id": "msg_priority-based-context_1", "role": "user", "content": "Focus on the deploy issue, not the earlier for...
600
440
true
priority
[ "msg_priority-based-context_0" ]
[ { "call_id": "msg_priority-based-context_4", "result_id": "msg_priority-based-context_5" }, { "call_id": "msg_priority-based-context_7", "result_id": "msg_priority-based-context_8" }, { "call_id": "msg_priority-based-context_11", "result_id": "msg_priority-based-context_12" }, { ...
[]
[ "msg_priority-based-context_0", "msg_priority-based-context_1", "msg_priority-based-context_2", "msg_priority-based-context_3", "msg_priority-based-context_4", "msg_priority-based-context_5", "msg_priority-based-context_6", "msg_priority-based-context_7", "msg_priority-based-context_8", "msg_prior...
Messages tagged high priority (current task) survive over low-priority ones (earlier, unrelated discussion), regardless of age.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
priority-based-context-002
priority-based-context
[ { "id": "msg_priority-based-context_0", "role": "system", "content": "You are an assistant prioritizing the current task above historical discussion.", "pinned": true }, { "id": "msg_priority-based-context_1", "role": "assistant", "content": "OK, let me check.", "priority": 0 }...
600
1,234
true
priority
[ "msg_priority-based-context_0" ]
[ { "call_id": "msg_priority-based-context_4", "result_id": "msg_priority-based-context_5" }, { "call_id": "msg_priority-based-context_8", "result_id": "msg_priority-based-context_9" }, { "call_id": "msg_priority-based-context_24", "result_id": "msg_priority-based-context_25" }, { ...
[ "msg_priority-based-context_1", "msg_priority-based-context_2", "msg_priority-based-context_3", "msg_priority-based-context_6", "msg_priority-based-context_7", "msg_priority-based-context_11", "msg_priority-based-context_12", "msg_priority-based-context_13", "msg_priority-based-context_14", "msg_p...
[ "msg_priority-based-context_0", "msg_priority-based-context_4", "msg_priority-based-context_5", "msg_priority-based-context_8", "msg_priority-based-context_9", "msg_priority-based-context_10", "msg_priority-based-context_16", "msg_priority-based-context_24", "msg_priority-based-context_25", "msg_p...
Messages tagged high priority (current task) survive over low-priority ones (earlier, unrelated discussion), regardless of age.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
priority-based-context-003
priority-based-context
[ { "id": "msg_priority-based-context_0", "role": "system", "content": "You are an assistant prioritizing the current task above historical discussion.", "pinned": true }, { "id": "msg_priority-based-context_1", "role": "user", "content": "Sounds good.", "priority": 0 }, { ...
600
2,413
true
priority
[ "msg_priority-based-context_0" ]
[ { "call_id": "msg_priority-based-context_7", "result_id": "msg_priority-based-context_8" }, { "call_id": "msg_priority-based-context_15", "result_id": "msg_priority-based-context_16" }, { "call_id": "msg_priority-based-context_25", "result_id": "msg_priority-based-context_26" }, ...
[ "msg_priority-based-context_1", "msg_priority-based-context_2", "msg_priority-based-context_3", "msg_priority-based-context_4", "msg_priority-based-context_5", "msg_priority-based-context_6", "msg_priority-based-context_7", "msg_priority-based-context_8", "msg_priority-based-context_9", "msg_prior...
[ "msg_priority-based-context_0", "msg_priority-based-context_17", "msg_priority-based-context_19", "msg_priority-based-context_20", "msg_priority-based-context_27", "msg_priority-based-context_34", "msg_priority-based-context_36", "msg_priority-based-context_37", "msg_priority-based-context_42", "m...
Messages tagged high priority (current task) survive over low-priority ones (earlier, unrelated discussion), regardless of age.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
research-agent-001
research-agent
[ { "id": "msg_research-agent_0", "role": "system", "content": "You are a research assistant. Always cite the source of any claim you make from retrieved documents.", "pinned": true }, { "id": "msg_research-agent_1", "role": "user", "content": "Understood.", "priority": 0 }, { ...
900
509
true
priority
[ "msg_research-agent_0" ]
[ { "call_id": "msg_research-agent_4", "result_id": "msg_research-agent_5" }, { "call_id": "msg_research-agent_8", "result_id": "msg_research-agent_9" }, { "call_id": "msg_research-agent_17", "result_id": "msg_research-agent_18" }, { "call_id": "msg_research-agent_19", "res...
[]
[ "msg_research-agent_0", "msg_research-agent_1", "msg_research-agent_2", "msg_research-agent_3", "msg_research-agent_4", "msg_research-agent_5", "msg_research-agent_6", "msg_research-agent_7", "msg_research-agent_8", "msg_research-agent_9", "msg_research-agent_10", "msg_research-agent_11", "m...
Tool-call/tool-result pairs (search queries and their retrieved content) remain coherent; no orphaned tool result.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
research-agent-002
research-agent
[ { "id": "msg_research-agent_0", "role": "system", "content": "You are a research assistant. Always cite the source of any claim you make from retrieved documents.", "pinned": true }, { "id": "msg_research-agent_1", "role": "user", "content": "Thanks for confirming.", "priority": ...
900
1,336
true
priority
[ "msg_research-agent_0" ]
[ { "call_id": "msg_research-agent_3", "result_id": "msg_research-agent_4" }, { "call_id": "msg_research-agent_22", "result_id": "msg_research-agent_23" }, { "call_id": "msg_research-agent_24", "result_id": "msg_research-agent_25" }, { "call_id": "msg_research-agent_34", "r...
[ "msg_research-agent_1", "msg_research-agent_2", "msg_research-agent_6", "msg_research-agent_7", "msg_research-agent_8", "msg_research-agent_9", "msg_research-agent_10", "msg_research-agent_11", "msg_research-agent_12", "msg_research-agent_14", "msg_research-agent_15", "msg_research-agent_16", ...
[ "msg_research-agent_0", "msg_research-agent_3", "msg_research-agent_4", "msg_research-agent_5", "msg_research-agent_13", "msg_research-agent_18", "msg_research-agent_22", "msg_research-agent_23", "msg_research-agent_24", "msg_research-agent_25", "msg_research-agent_28", "msg_research-agent_33"...
Tool-call/tool-result pairs (search queries and their retrieved content) remain coherent; no orphaned tool result.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
research-agent-003
research-agent
[ { "id": "msg_research-agent_0", "role": "system", "content": "You are a research assistant. Always cite the source of any claim you make from retrieved documents.", "pinned": true }, { "id": "msg_research-agent_1", "role": "user", "content": "Will do.", "priority": 0 }, { ...
900
2,518
true
priority
[ "msg_research-agent_0" ]
[ { "call_id": "msg_research-agent_3", "result_id": "msg_research-agent_4" }, { "call_id": "msg_research-agent_21", "result_id": "msg_research-agent_22" }, { "call_id": "msg_research-agent_41", "result_id": "msg_research-agent_42" }, { "call_id": "msg_research-agent_56", "r...
[ "msg_research-agent_1", "msg_research-agent_2", "msg_research-agent_5", "msg_research-agent_6", "msg_research-agent_7", "msg_research-agent_9", "msg_research-agent_10", "msg_research-agent_11", "msg_research-agent_12", "msg_research-agent_14", "msg_research-agent_16", "msg_research-agent_17", ...
[ "msg_research-agent_0", "msg_research-agent_3", "msg_research-agent_4", "msg_research-agent_8", "msg_research-agent_13", "msg_research-agent_15", "msg_research-agent_21", "msg_research-agent_22", "msg_research-agent_24", "msg_research-agent_28", "msg_research-agent_33", "msg_research-agent_37"...
Tool-call/tool-result pairs (search queries and their retrieved content) remain coherent; no orphaned tool result.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
tool-call-atomicity-001
tool-call-atomicity
[ { "id": "msg_tool-call-atomicity_0", "role": "system", "content": "You are an assistant that calls external tools. Every tool call must be answered by exactly one tool result.", "pinned": true }, { "id": "msg_tool-call-atomicity_1", "role": "assistant", "content": "OK, let me check."...
350
417
true
dropOldest
[ "msg_tool-call-atomicity_0" ]
[ { "call_id": "msg_tool-call-atomicity_3", "result_id": "msg_tool-call-atomicity_4" }, { "call_id": "msg_tool-call-atomicity_6", "result_id": "msg_tool-call-atomicity_7" } ]
[ "msg_tool-call-atomicity_1", "msg_tool-call-atomicity_2", "msg_tool-call-atomicity_3", "msg_tool-call-atomicity_4", "msg_tool-call-atomicity_5", "msg_tool-call-atomicity_6", "msg_tool-call-atomicity_7" ]
[ "msg_tool-call-atomicity_0", "msg_tool-call-atomicity_8", "msg_tool-call-atomicity_9", "msg_tool-call-atomicity_10", "msg_tool-call-atomicity_11", "msg_tool-call-atomicity_12", "msg_tool-call-atomicity_13", "msg_tool-call-atomicity_14", "msg_tool-call-atomicity_15", "msg_tool-call-atomicity_16", ...
Every surviving tool-result message has a surviving tool-call it answers, and vice versa — never one without the other.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
tool-call-atomicity-002
tool-call-atomicity
[ { "id": "msg_tool-call-atomicity_0", "role": "system", "content": "You are an assistant that calls external tools. Every tool call must be answered by exactly one tool result.", "pinned": true }, { "id": "msg_tool-call-atomicity_1", "role": "assistant", "content": "Got it, thanks.", ...
350
1,147
true
dropOldest
[ "msg_tool-call-atomicity_0" ]
[ { "call_id": "msg_tool-call-atomicity_3", "result_id": "msg_tool-call-atomicity_4" }, { "call_id": "msg_tool-call-atomicity_21", "result_id": "msg_tool-call-atomicity_22" }, { "call_id": "msg_tool-call-atomicity_26", "result_id": "msg_tool-call-atomicity_27" }, { "call_id": "...
[ "msg_tool-call-atomicity_1", "msg_tool-call-atomicity_2", "msg_tool-call-atomicity_3", "msg_tool-call-atomicity_4", "msg_tool-call-atomicity_5", "msg_tool-call-atomicity_6", "msg_tool-call-atomicity_7", "msg_tool-call-atomicity_8", "msg_tool-call-atomicity_9", "msg_tool-call-atomicity_10", "msg_...
[ "msg_tool-call-atomicity_0", "msg_tool-call-atomicity_59", "msg_tool-call-atomicity_60", "msg_tool-call-atomicity_61", "msg_tool-call-atomicity_62", "msg_tool-call-atomicity_63", "msg_tool-call-atomicity_64", "msg_tool-call-atomicity_65", "msg_tool-call-atomicity_66", "msg_tool-call-atomicity_67",...
Every surviving tool-result message has a surviving tool-call it answers, and vice versa — never one without the other.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
tool-call-atomicity-003
tool-call-atomicity
[ { "id": "msg_tool-call-atomicity_0", "role": "system", "content": "You are an assistant that calls external tools. Every tool call must be answered by exactly one tool result.", "pinned": true }, { "id": "msg_tool-call-atomicity_1", "role": "user", "content": "Let me check the curren...
350
2,253
true
dropOldest
[ "msg_tool-call-atomicity_0" ]
[ { "call_id": "msg_tool-call-atomicity_5", "result_id": "msg_tool-call-atomicity_6" }, { "call_id": "msg_tool-call-atomicity_12", "result_id": "msg_tool-call-atomicity_13" }, { "call_id": "msg_tool-call-atomicity_28", "result_id": "msg_tool-call-atomicity_29" }, { "call_id": "...
[ "msg_tool-call-atomicity_1", "msg_tool-call-atomicity_2", "msg_tool-call-atomicity_3", "msg_tool-call-atomicity_4", "msg_tool-call-atomicity_5", "msg_tool-call-atomicity_6", "msg_tool-call-atomicity_7", "msg_tool-call-atomicity_8", "msg_tool-call-atomicity_9", "msg_tool-call-atomicity_10", "msg_...
[ "msg_tool-call-atomicity_0", "msg_tool-call-atomicity_130", "msg_tool-call-atomicity_131", "msg_tool-call-atomicity_132", "msg_tool-call-atomicity_133", "msg_tool-call-atomicity_134", "msg_tool-call-atomicity_135", "msg_tool-call-atomicity_136", "msg_tool-call-atomicity_137", "msg_tool-call-atomic...
Every surviving tool-result message has a surviving tool-call it answers, and vice versa — never one without the other.
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
tool-heavy-agent-001
tool-heavy-agent
[ { "id": "msg_tool-heavy-agent_0", "role": "system", "content": "You are a coding assistant. Follow the user's requirements and preserve important project constraints (the app uses JWTs with refresh tokens, not sessions).", "pinned": true }, { "id": "msg_tool-heavy-agent_1", "role": "assi...
900
540
true
priority
[ "msg_tool-heavy-agent_0" ]
[ { "call_id": "msg_tool-heavy-agent_6", "result_id": "msg_tool-heavy-agent_7" }, { "call_id": "msg_tool-heavy-agent_8", "result_id": "msg_tool-heavy-agent_9" }, { "call_id": "msg_tool-heavy-agent_11", "result_id": "msg_tool-heavy-agent_12" }, { "call_id": "msg_tool-heavy-agent...
[]
[ "msg_tool-heavy-agent_0", "msg_tool-heavy-agent_1", "msg_tool-heavy-agent_2", "msg_tool-heavy-agent_3", "msg_tool-heavy-agent_4", "msg_tool-heavy-agent_5", "msg_tool-heavy-agent_6", "msg_tool-heavy-agent_7", "msg_tool-heavy-agent_8", "msg_tool-heavy-agent_9", "msg_tool-heavy-agent_10", "msg_to...
High-priority tool calls (the regression investigation) outlast low-priority ones (a stale, already-resolved lookup).
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
tool-heavy-agent-002
tool-heavy-agent
[ { "id": "msg_tool-heavy-agent_0", "role": "system", "content": "You are a coding assistant. Follow the user's requirements and preserve important project constraints (the app uses JWTs with refresh tokens, not sessions).", "pinned": true }, { "id": "msg_tool-heavy-agent_1", "role": "user...
900
1,377
true
priority
[ "msg_tool-heavy-agent_0" ]
[ { "call_id": "msg_tool-heavy-agent_10", "result_id": "msg_tool-heavy-agent_11" }, { "call_id": "msg_tool-heavy-agent_14", "result_id": "msg_tool-heavy-agent_15" }, { "call_id": "msg_tool-heavy-agent_19", "result_id": "msg_tool-heavy-agent_20" }, { "call_id": "msg_tool-heavy-a...
[ "msg_tool-heavy-agent_1", "msg_tool-heavy-agent_2", "msg_tool-heavy-agent_3", "msg_tool-heavy-agent_4", "msg_tool-heavy-agent_5", "msg_tool-heavy-agent_6", "msg_tool-heavy-agent_7", "msg_tool-heavy-agent_8", "msg_tool-heavy-agent_12", "msg_tool-heavy-agent_13", "msg_tool-heavy-agent_16", "msg_...
[ "msg_tool-heavy-agent_0", "msg_tool-heavy-agent_9", "msg_tool-heavy-agent_10", "msg_tool-heavy-agent_11", "msg_tool-heavy-agent_14", "msg_tool-heavy-agent_15", "msg_tool-heavy-agent_19", "msg_tool-heavy-agent_20", "msg_tool-heavy-agent_25", "msg_tool-heavy-agent_28", "msg_tool-heavy-agent_29", ...
High-priority tool calls (the regression investigation) outlast low-priority ones (a stale, already-resolved lookup).
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.
tool-heavy-agent-003
tool-heavy-agent
[ { "id": "msg_tool-heavy-agent_0", "role": "system", "content": "You are a coding assistant. Follow the user's requirements and preserve important project constraints (the app uses JWTs with refresh tokens, not sessions).", "pinned": true }, { "id": "msg_tool-heavy-agent_1", "role": "user...
900
2,541
true
priority
[ "msg_tool-heavy-agent_0" ]
[ { "call_id": "msg_tool-heavy-agent_4", "result_id": "msg_tool-heavy-agent_5" }, { "call_id": "msg_tool-heavy-agent_14", "result_id": "msg_tool-heavy-agent_15" }, { "call_id": "msg_tool-heavy-agent_16", "result_id": "msg_tool-heavy-agent_17" }, { "call_id": "msg_tool-heavy-age...
[ "msg_tool-heavy-agent_1", "msg_tool-heavy-agent_3", "msg_tool-heavy-agent_7", "msg_tool-heavy-agent_8", "msg_tool-heavy-agent_9", "msg_tool-heavy-agent_12", "msg_tool-heavy-agent_16", "msg_tool-heavy-agent_17", "msg_tool-heavy-agent_18", "msg_tool-heavy-agent_19", "msg_tool-heavy-agent_20", "m...
[ "msg_tool-heavy-agent_0", "msg_tool-heavy-agent_2", "msg_tool-heavy-agent_4", "msg_tool-heavy-agent_5", "msg_tool-heavy-agent_6", "msg_tool-heavy-agent_10", "msg_tool-heavy-agent_11", "msg_tool-heavy-agent_13", "msg_tool-heavy-agent_14", "msg_tool-heavy-agent_15", "msg_tool-heavy-agent_25", "m...
High-priority tool calls (the regression investigation) outlast low-priority ones (a stale, already-resolved lookup).
token_count_estimate uses the built-in heuristic estimator (~4 chars/token), the same default TokenBudget uses when no real tokenizer is configured — it is an estimate, not an exact count from any specific model's tokenizer.