34 lines
4.3 KiB
Markdown
34 lines
4.3 KiB
Markdown
|
# Bundled profiles
|
||
|
|
|
||
|
|
The 8.0.0 portfolio contains eight featured workload profiles and four explicitly experimental labs. Every MMO-owned document in this development generation uses schema `8`; profiles also require `version = "8.0.0"`. Earlier generations, retired profile IDs, and old catalog keys are invalid, with no compatibility reader, alias, migration, fallback host, or automatic rewrite. External formats owned by Codex app-server, MCP, providers, and Switchyard keep their upstream identifiers. Spawn admission is based only on simultaneous active capacity: terminal, cold-paused, and suspended workers release their role, route, resource, child, and writer slots for later or resumed work, and continuation performs fresh admission.
|
||
|
|
|
||
|
|
Every root and Agent-MCP role is hosted by a Codex 0.149.0 app-server on a private Unix socket while executing. Goal roles use token budgets with compiled ceilings; turn roles have no task wall clock. Stall intervals are warnings, not interrupts, and the model is never told to measure time. Per-target `controls` grants expose only the admitted inspect/trace/steer/interrupt/pause/continue/detach/stop/finalize/compact/respond/effort/fork actions. Cold pause retires MCP hosts while retaining exact threads and evidence; detach leaves them live. Durable events, terminal history, partial results, and isolated patches survive client/controller/provider loss; they do not weaken the role's trust, permission, or verification policy.
|
||
|
|
|
||
|
|
## Featured
|
||
|
|
|
||
|
|
| Profile | Practical purpose | Composition | Engine |
|
||
|
|
|---|---|---|---|
|
||
|
|
| `adaptive-engineering` | General engineering with selective reconnaissance, isolated implementation, and independent review | Sol root; Luna, DeepSeek V4 Pro, Opus 5 | Hybrid |
|
||
|
|
| `codex-harness-team` | ChatGPT-only control showing the value of context isolation and role specialization | Sol root; Luna, Terra, fresh Sol critic | Hybrid |
|
||
|
|
| `high-confidence-debugging` | Reproduce first, repair second, then attempt to falsify the fix | DeepSeek V4 Pro root; Sonnet 5, Sol | Agent MCP |
|
||
|
|
| `research-backed-engineering` | Current-source implementation with claim-level citations and source-to-code audit | MiniMax M3 root; Luna, Sonnet 5 | Agent MCP |
|
||
|
|
| `secure-change` | Threat framing, independent vulnerability hunting, empirical reproduction, and remediation | Opus 5 root; GLM-5.3 via Z.AI Coding Plan, DeepSeek V4 Pro | Agent MCP |
|
||
|
|
| `visual-engineering` | Reference-to-code UI work with real browser rendering and fresh multimodal verification | Sol root; Terra, fresh Sol verifier | Hybrid |
|
||
|
|
| `contract-first-refactoring` | Freeze behavior, derive independent contract tests, refactor, and verify compatibility | Sonnet 5 root; MiniMax M3, Luna, DeepSeek V4 Pro | Agent MCP |
|
||
|
|
| `incident-hypothesis-triage` | Read-only incident analysis that separates evidence collection from causal challenge | GLM-5.3 Coding Plan root; DeepSeek V4 Flash, pinned Nemotron | Agent MCP |
|
||
|
|
|
||
|
|
## Labs
|
||
|
|
|
||
|
|
| Profile | Hypothesis under test | Composition | Engine |
|
||
|
|
|---|---|---|---|
|
||
|
|
| `access-efficient-escalation-lab` | Cheap/local bounded work plus deliberate escalation can reduce paid usage without losing quality | Luna root; local Qwen3.5-9B, DeepSeek V4 Flash, GLM-5.3 | Agent MCP |
|
||
|
|
| `route-resilience-lab` | The same maker/model through independent access routes can expose route drift and outage behavior | Terra root; DeepSeek V4 Pro through Go and pinned OpenRouter | Agent MCP |
|
||
|
|
| `bounded-research-organization-lab` | A strictly bounded research hierarchy can outperform flat fan-out on long corpora | Sol root; Kimi K3 lead, MiniMax M3 scouts | Agent MCP |
|
||
|
|
| `competing-implementations-lab` | Two isolated implementations selected by an objective contract can beat a single implementation | Terra judge; DeepSeek V4 Pro and Sonnet 5 candidates | Agent MCP |
|
||
|
|
|
||
|
|
Labs are packaged experiments, not recommended defaults. Their README and evaluation suite state the hypothesis, matched controls, and promotion gate.
|
||
|
|
|
||
|
|
Every pack contains only static TOML, Markdown instructions, and JSON result contracts. It does not contain credentials, endpoint definitions, Tool MCP server definitions, executable workflow code, or compatibility shims. Exact route-qualified model keys make the model maker, access route, and wire protocol inspectable through the catalog.
|
||
|
|
|
||
|
|
See [the full profile guide](../docs/PROFILES.md) for role graphs, availability behavior, trust boundaries, and selection guidance.
|