This commit is contained in:
2026-08-24 08:11:59 -07:00
commit 53df0eed10
275 changed files with 133056 additions and 0 deletions
@@ -0,0 +1,7 @@
# Access-Efficient Escalation Lab
This lab tests whether cheap or local work can reduce scarce-tier use without unacceptable quality loss. Luna remains the responsible root; local Qwen receives only runtime-generated literal operations; DeepSeek Flash handles one objective isolated patch; GLM is a deliberate read-only escalation.
The local route is optional. Weak-model tasks have no free-form prompt surface, no network, no writes, no spawning, a single GPU slot, and hash/line verification. The lab is not promoted by lower nominal cost alone: matched trials must show at least 50% lower API-equivalent or scarce-tier consumption with no more than two points of success loss.
All three workers use persistent supervised Unix app-server threads. The root can inspect, trace, steer, interrupt, pause, continue, detach, finalize, fully stop, or cancel each one. Goal token ceilings or clock-free turns replace per-spawn model timeouts; stall warnings never interrupt, and suspension retains terminal history, trace, partial evidence, and any isolated patch. The Qwen role remains low-trust and cannot control or spawn anything.
@@ -0,0 +1 @@
Own the critical path and canonical workspace. Use the local scout only for one structured literal operation, never free-form reasoning. Use DeepSeek Flash for one bounded routine patch with objective tests. Escalate to GLM only when recorded evidence shows unresolved ambiguity, security risk, or failed routine verification. Continue useful root work while independent tasks run, explicitly accept or reject results, and integrate isolated patches only after tests. Record which scarce-tier calls were avoided; this lab makes no quality claim until matched trials satisfy its gate.
@@ -0,0 +1 @@
Address only the explicitly documented escalation trigger. Analyze the unresolved high-impact question read-only, cite repository evidence, challenge the routine conclusion, and return a decision recommendation with falsification tests. Do not edit or manufacture certainty.
@@ -0,0 +1 @@
The runtime supplies one literal operation. Perform only that operation and return exact path, hash, line-range, and excerpt evidence or a hash-linked summary of supplied text. Do not infer intent, diagnose, design, recommend, execute arbitrary commands, edit files, browse, or spawn.
@@ -0,0 +1 @@
Implement only the bounded routine change in the assigned isolated scope. Run objective tests and return correlated command evidence. Do not expand the task or assume the root will integrate the patch.
@@ -0,0 +1,13 @@
{
"title": "Economical routine engineering patch",
"type": "object",
"additionalProperties": false,
"properties": {
"summary": {"type": "string", "minLength": 1},
"files": {"type": "array", "items": {"type": "string"}, "maxItems": 30, "uniqueItems": true},
"commands": {"type": "array", "maxItems": 30, "items": {"type": "object", "additionalProperties": false, "properties": {"command": {"type": "string", "minLength": 1}, "exit_code": {"type": "integer"}, "observation": {"type": "string", "minLength": 1}}, "required": ["command", "exit_code", "observation"]}},
"risks": {"type": "array", "items": {"type": "string"}, "maxItems": 20},
"blockers": {"type": "array", "items": {"type": "string"}, "maxItems": 20}
},
"required": ["summary", "files", "commands", "risks", "blockers"]
}
@@ -0,0 +1,13 @@
{
"title": "Evidence-triggered difficult-reasoning escalation",
"type": "object",
"additionalProperties": false,
"properties": {
"trigger": {"type": "string", "minLength": 1},
"recommendation": {"type": "string", "minLength": 1},
"evidence": {"type": "array", "items": {"type": "string"}, "minItems": 1, "maxItems": 50},
"falsification_tests": {"type": "array", "items": {"type": "string"}, "minItems": 1, "maxItems": 30},
"uncertainties": {"type": "array", "items": {"type": "string"}, "maxItems": 30}
},
"required": ["trigger", "recommendation", "evidence", "falsification_tests", "uncertainties"]
}
@@ -0,0 +1,31 @@
{
"title": "Mechanically checked literal evidence",
"type": "object",
"additionalProperties": false,
"properties": {
"operation": {"type": "string", "enum": ["locate", "references", "extract", "summarize_supplied"]},
"input_sha256": {"type": ["string", "null"], "pattern": "^[0-9a-f]{64}$"},
"evidence": {
"type": "array",
"maxItems": 200,
"items": {
"type": "object",
"additionalProperties": false,
"properties": {
"path": {"type": "string", "minLength": 1},
"sha256": {"type": "string", "pattern": "^[0-9a-f]{64}$"},
"start_line": {"type": "integer", "minimum": 1},
"end_line": {"type": "integer", "minimum": 1},
"excerpt": {"type": "string"}
},
"required": ["path", "sha256", "start_line", "end_line", "excerpt"]
}
},
"summary_points": {"type": "array", "items": {"type": "string"}, "maxItems": 50}
},
"required": ["operation", "input_sha256", "evidence", "summary_points"],
"allOf": [
{"if": {"properties": {"operation": {"enum": ["locate", "references", "extract"]}}, "required": ["operation"]}, "then": {"properties": {"input_sha256": {"const": null}, "evidence": {"minItems": 1}, "summary_points": {"maxItems": 0}}}},
{"if": {"properties": {"operation": {"const": "summarize_supplied"}}, "required": ["operation"]}, "then": {"properties": {"input_sha256": {"type": "string", "pattern": "^[0-9a-f]{64}$"}, "evidence": {"maxItems": 0}, "summary_points": {"minItems": 1}}}}
]
}
@@ -0,0 +1,111 @@
schema_version = 8
id = "access-efficient-escalation-lab"
version = "8.0.0"
display_name = "Access-Efficient Escalation Lab"
description = "An experimental Luna root routes literal reconnaissance to a contained local model, routine patches to economical DeepSeek Flash, and difficult read-only analysis to GLM only when justified."
tags = ["cost", "local-first", "escalation", "lab"]
maturity = "lab"
root = "access_router"
smoke = "smoke.toml"
[agents.access_router]
kind = "root"
description = "Productive routing root, canonical writer, and escalation authority"
model = "codex_chatgpt_builtin__gpt_5_6_luna"
reasoning = "medium"
allowed_reasoning_efforts = ["low", "medium", "high"]
execution_mode = "goal"
goal_token_budget = 250000
max_goal_token_budget = 400000
stall_warning_seconds = 3600
finalization_grace_seconds = 1200
plan_reasoning = "medium"
permissions = "workspace-write"
can_spawn = ["literal_scout", "routine_engineer", "flagship_escalation"]
max_active = 1
trust = "high"
verification = "always"
instructions = "agents/access-router.md"
[agents.literal_scout]
description = "Mechanically contained local model for literal repository evidence only"
model = "llama_cpp_local_openai_chat__qwen3_5_9b"
reasoning = "none"
allowed_reasoning_efforts = ["none"]
execution_mode = "turn"
stall_warning_seconds = 1800
finalization_grace_seconds = 600
permissions = "read-only"
backends = ["mcp"]
can_spawn = []
max_active = 1
max_task_chars = 2500
allowed_task_kinds = ["locate", "references", "extract", "summarize_supplied"]
trust = "low"
verification = "always"
network_access = false
web_search = "disabled"
contract_enforcement = "strict"
resource_group = "local_gpu_0"
output_contract = "contracts/literal-evidence.json"
instructions = "agents/literal-scout.md"
[agents.routine_engineer]
description = "Economical isolated worker for one bounded routine patch"
model = "opencode_go_openai_chat__deepseek_v4_flash"
reasoning = "high"
allowed_reasoning_efforts = ["low", "high", "max"]
execution_mode = "turn"
stall_warning_seconds = 1800
finalization_grace_seconds = 1200
permissions = "workspace-write"
backends = ["mcp"]
can_spawn = []
max_active = 1
allowed_task_kinds = ["implement", "debug", "test", "refactor"]
trust = "high"
verification = "always"
contract_enforcement = "strict"
output_contract = "contracts/engineering.json"
instructions = "agents/routine-engineer.md"
[agents.flagship_escalation]
description = "Read-only difficult-reasoning escalation used only after explicit trigger evidence"
model = "zai_coding_responses__glm_5_3"
reasoning = "max"
allowed_reasoning_efforts = ["high", "max"]
execution_mode = "turn"
stall_warning_seconds = 1800
finalization_grace_seconds = 900
permissions = "read-only"
backends = ["mcp"]
can_spawn = []
max_active = 1
allowed_task_kinds = ["analysis", "review", "security_review"]
trust = "high"
verification = "root_adjudication"
contract_enforcement = "strict"
output_contract = "contracts/escalation.json"
instructions = "agents/flagship-escalation.md"
[agents.access_router.controls.literal_scout]
actions = ["inspect", "trace", "steer", "interrupt", "pause", "continue", "detach", "stop", "finalize", "compact", "respond", "set_effort", "fork"]
[agents.access_router.controls.routine_engineer]
actions = ["inspect", "trace", "steer", "interrupt", "pause", "continue", "detach", "stop", "finalize", "compact", "respond", "set_effort", "fork"]
[agents.access_router.controls.flagship_escalation]
actions = ["inspect", "trace", "steer", "interrupt", "pause", "continue", "detach", "stop", "finalize", "compact", "respond", "set_effort", "fork"]
[coordination]
mode = "access_efficient_escalation_experiment"
orchestration = "mcp"
max_active_agents = 3
max_depth = 1
max_children_per_agent = 3
max_active_writers = 1
reject_ancestor_role = true
wait_policy = "dependency_only"
write_conflict_policy = "reject"
contradiction_policy = "primary_evidence"
result_visibility = "ancestors"
@@ -0,0 +1,36 @@
schema_version = 8
[[tasks]]
agent = "access_router"
task_kind = "analysis"
task = "Inspect this lab read-only and state the mechanical escalation and weak-model containment rules."
mode = "read-only"
wall_timeout_seconds = 600
[[tasks]]
agent = "literal_scout"
backend = "mcp"
mode = "read-only"
wall_timeout_seconds = 600
[tasks.literal_task]
operation = "locate"
needle = "MMO_SCHEMA_VERSION"
paths = ["libexec/mmo_profiles.py"]
max_results = 10
[[tasks]]
agent = "routine_engineer"
task_kind = "test"
task = "Inspect one existing test read-only, run one focused validation command, and return the engineering JSON with no file changes."
backend = "mcp"
mode = "read-only"
wall_timeout_seconds = 900
[[tasks]]
agent = "flagship_escalation"
task_kind = "analysis"
task = "Treat uncertainty about whether schema v1 remains accepted as the explicit escalation trigger. Analyze it read-only and return only the escalation JSON with exact evidence and a falsification test."
backend = "mcp"
mode = "read-only"
wall_timeout_seconds = 900