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 +1,15 @@
# Research-Backed Engineering
Use this profile when implementation correctness depends on current external APIs, standards, advisories, provider behavior, or versioned specifications.
```text
MiniMax M3 research lead (active root, sole writer, built-in live web search)
├── Luna source researcher (MCP, read-only, built-in live web search)
└── Claude Sonnet 5 source auditor (MCP, read-only, built-in live web search)
```
Independent research can run in parallel while the root maps the repository, but an authoritative dated dossier is a real dependency before root-owned implementation. The auditor independently re-fetches pivotal sources and checks both the dossier and change. Exact models, live-search grants, contracts, goal/turn lifecycle, and sole-writer boundary are mechanical. Source authority and implementation ordering are instruction-driven and evaluated.
Cost and latency are high. Conflicting or unavailable primary sources produce a blocked or reduced-confidence result. Operators may clone the profile and grant a registered Firecrawl-like Tool MCP server only to the researcher; no deployment-specific server is bundled.
Researcher and auditor are persistent supervised Unix app-server workers. The root controls both; the auditor may steer, interrupt, pause, continue, detach, or return the researcher for a source correction while acceptance remains with normal lineage authority. Pending input remains durable, and transport suspension retains terminal history, the dated dossier, trace, and partial source evidence.
@@ -0,0 +1,3 @@
Own the repository work, canonical writes, and final source adjudication. Split independent current-source questions early and continue inspecting the repository while researchers browse. Require every consequential claim to carry a URL, publication or update date, retrieval timestamp, authority classification, and concise evidence. Re-fetch pivotal primary sources yourself before allowing them to govern implementation. Use the source auditor when sources conflict, the change is temporally sensitive, or source-to-code compliance is consequential.
Contracts validate shape and captured-event correlation, not truth. Resolve conflicts by authority, date, direct applicability, and reproducible behavior. Explicitly accept or reject dossiers; record unavailable routes and unresolved uncertainty. Never treat a search snippet or worker consensus as evidence.
@@ -0,0 +1 @@
Independently audit the source dossier and resulting code. Revisit primary sources when permitted, check dates and applicability, identify contradictions or unsupported leaps, and map material requirements to exact implementation evidence. A valid JSON result is not proof of truth. Return a pass only when pivotal claims and code are independently supported; otherwise identify the precise re-fetch or empirical test needed.
@@ -0,0 +1 @@
Research one bounded question using current primary sources first. Return URLs, source dates, retrieval timestamps, authority, directly supported claims, conflicts, and limitations. Separate source statements from your inference and keep excerpts short. Do not edit the repository or expand into the other researcher's question.
@@ -0,0 +1,44 @@
{
"title": "Dated current-source dossier",
"type": "object",
"additionalProperties": false,
"properties": {
"status": {"type": "string", "enum": ["supported", "conflict", "blocked"]},
"question": {"type": "string", "minLength": 1},
"claims": {
"type": "array",
"maxItems": 50,
"items": {
"type": "object",
"additionalProperties": false,
"properties": {
"claim": {"type": "string", "minLength": 1},
"source_url": {"type": "string", "format": "uri"},
"source_date": {"type": "string", "format": "date"},
"retrieved_at": {"type": "string", "format": "date-time"},
"authority": {"type": "string", "enum": ["primary", "supporting"]},
"evidence": {"type": "string", "minLength": 1},
"inference": {"type": "boolean"}
},
"required": ["claim", "source_url", "source_date", "retrieved_at", "authority", "evidence", "inference"]
}
},
"conflicts": {"type": "array", "items": {"type": "string"}, "maxItems": 30},
"limitations": {"type": "array", "items": {"type": "string"}, "maxItems": 30}
},
"required": ["status", "question", "claims", "conflicts", "limitations"],
"allOf": [
{
"if": {"properties": {"status": {"const": "supported"}}, "required": ["status"]},
"then": {"properties": {"claims": {"minItems": 1}, "conflicts": {"maxItems": 0}}}
},
{
"if": {"properties": {"status": {"const": "conflict"}}, "required": ["status"]},
"then": {"properties": {"conflicts": {"minItems": 1}}}
},
{
"if": {"properties": {"status": {"const": "blocked"}}, "required": ["status"]},
"then": {"properties": {"limitations": {"minItems": 1}}}
}
]
}
@@ -0,0 +1,37 @@
{
"title": "Independent source-to-code audit",
"type": "object",
"additionalProperties": false,
"properties": {
"verdict": {"type": "string", "enum": ["pass", "concerns", "fail", "blocked"]},
"checked_sources": {"type": "array", "items": {"type": "string", "format": "uri"}, "maxItems": 50, "uniqueItems": true},
"findings": {
"type": "array",
"maxItems": 50,
"items": {
"type": "object",
"additionalProperties": false,
"properties": {
"claim": {"type": "string", "minLength": 1},
"source_url": {"type": "string", "format": "uri"},
"code_location": {"type": "string", "minLength": 1},
"evidence": {"type": "string", "minLength": 1}
},
"required": ["claim", "source_url", "code_location", "evidence"]
}
},
"conflicts": {"type": "array", "items": {"type": "string"}, "maxItems": 30},
"blockers": {"type": "array", "items": {"type": "string"}, "maxItems": 30}
},
"required": ["verdict", "checked_sources", "findings", "conflicts", "blockers"],
"allOf": [
{
"if": {"properties": {"verdict": {"const": "pass"}}, "required": ["verdict"]},
"then": {"properties": {"checked_sources": {"minItems": 1}, "findings": {"maxItems": 0}, "blockers": {"maxItems": 0}}}
},
{
"if": {"properties": {"verdict": {"const": "blocked"}}, "required": ["verdict"]},
"then": {"properties": {"blockers": {"minItems": 1}}}
}
]
}
@@ -0,0 +1,95 @@
schema_version = 8
id = "research-backed-engineering"
version = "8.0.0"
display_name = "Research-Backed Engineering"
description = "A long-context MiniMax root integrates current primary-source research, re-fetches pivotal claims, and invites an independent source-to-code audit."
tags = ["research", "engineering", "current-sources", "featured"]
maturity = "featured"
root = "research_lead"
smoke = "smoke.toml"
[agents.research_lead]
kind = "root"
description = "Long-context source adjudicator, repository engineer, sole writer, and final authority"
model = "opencode_go_anthropic_messages__minimax_m3"
reasoning = "none"
allowed_reasoning_efforts = ["none"]
execution_mode = "goal"
goal_token_budget = 600000
max_goal_token_budget = 1000000
stall_warning_seconds = 3600
finalization_grace_seconds = 1800
permissions = "workspace-write"
can_spawn = ["source_researcher", "source_auditor"]
max_active = 1
trust = "high"
verification = "always"
network_access = false
web_search = "live"
instructions = "agents/research-lead.md"
[agents.source_researcher]
description = "Fast independent live-web researcher for one bounded current-source question"
model = "codex_chatgpt_builtin__gpt_5_6_luna"
reasoning = "medium"
allowed_reasoning_efforts = ["low", "medium", "high"]
execution_mode = "goal"
goal_token_budget = 200000
max_goal_token_budget = 300000
stall_warning_seconds = 1800
finalization_grace_seconds = 900
permissions = "read-only"
backends = ["mcp"]
can_spawn = []
max_active = 2
allowed_task_kinds = ["research", "source_verification"]
trust = "high"
verification = "always"
network_access = false
web_search = "live"
contract_enforcement = "strict"
output_contract = "contracts/research.json"
instructions = "agents/source-researcher.md"
[agents.source_auditor]
description = "Conditional independent audit of authority, currency, conflicts, and source-to-code compliance"
model = "opencode_zen_anthropic_messages__claude_sonnet_5"
reasoning = "high"
allowed_reasoning_efforts = ["medium", "high", "xhigh", "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 = ["source_audit", "review", "verification"]
trust = "adversarial_reviewer"
verification = "root_adjudication"
network_access = false
web_search = "live"
contract_enforcement = "strict"
output_contract = "contracts/review.json"
instructions = "agents/source-auditor.md"
[agents.research_lead.controls.source_researcher]
actions = ["inspect", "trace", "steer", "interrupt", "pause", "continue", "detach", "stop", "finalize", "compact", "respond", "set_effort", "fork"]
[agents.research_lead.controls.source_auditor]
actions = ["inspect", "trace", "steer", "interrupt", "pause", "continue", "detach", "stop", "finalize", "compact", "respond", "set_effort", "fork"]
[agents.source_auditor.controls.source_researcher]
actions = ["inspect", "trace", "steer", "interrupt", "pause", "continue", "detach", "stop", "finalize", "compact", "respond", "set_effort", "fork"]
[coordination]
mode = "research_implement_verify"
orchestration = "mcp"
max_active_agents = 3
max_depth = 1
max_children_per_agent = 2
max_active_writers = 0
reject_ancestor_role = true
wait_policy = "dependency_only"
write_conflict_policy = "reject"
contradiction_policy = "primary_evidence"
result_visibility = "ancestors"
@@ -0,0 +1,24 @@
schema_version = 8
[[tasks]]
agent = "research_lead"
task_kind = "analysis"
task = "Inspect the repository read-only and identify the mechanical gates that distinguish current-source evidence from unsupported claims."
mode = "read-only"
wall_timeout_seconds = 600
[[tasks]]
agent = "source_researcher"
task_kind = "research"
task = "Using a current primary source, verify one documented Python language behavior relevant to this repository. Return only the dated dossier JSON and separate inference from source evidence."
backend = "mcp"
mode = "read-only"
wall_timeout_seconds = 900
[[tasks]]
agent = "source_auditor"
task_kind = "source_audit"
task = "Audit docs/EXTERNAL-VERACITY.md against one directly cited primary source. Return only the source-to-code audit JSON, or blocked with a precise reason."
backend = "mcp"
mode = "read-only"
wall_timeout_seconds = 900