This guide describes the checked-in configuration and source reviewed on 2026-10-11. It accompanies the README. The May agent blueprint preserves the earlier design plan; its proposed defaults and completion targets are not a statement of the current runtime.
- The bot receives a player message or an execution event. Player chat and command access are configured through
only_chat_withandpermissionsin settings.js. - With
use_orchestrator_v2anduse_tool_use_protocolenabled, the orchestrator coordinates model turns and structured tool calls. The live-state block supplies observations such as inventory and supported task-progress information. - The LLM plans and chooses actions. The prompter also invokes the coding model for generated actions. If a profile omits
code_model, it uses the chat model for coding. - The task queue, action manager, and modes manage work and interruptions. Preconditions and completion checks cover particular supported actions and task criteria; they do not validate every possible plan.
The bot-side verification code parses supported inventory criteria: absolute counts, gains or losses relative to the start, and sets of required items. Unknown criteria can return programmatic: false. A check being implemented in code does not make every task measurable.
The external evaluation scorer takes its own measurements. It shares the inventory-criterion parser with the bot, and can use blockscan.py for supported square-platform criteria. Unsupported criteria, missing snapshots, or unavailable measurements can fall back to the reported task outcome with label_source: honor_system. Analysis must distinguish those labels from scorer-measured outcomes.
The experiment’s harness switches control perception additions, preconditions, reflexes, bot-side completion verification, and automatic task termination. The external scorer remains enabled when those components are disabled. Player-safety mechanisms also remain enabled. The comparison is between configurations of this fork.
The checked-in settings enable the rate-limit wrapper, tool-use protocol, and v2 orchestrator. They disable use_background_handles and use_subagent_isolation. The historical blueprint’s proposed background execution and context-isolation behavior should not be inferred from the presence of those source files.
The starter Claude profile selects Claude Sonnet 4.6 and OpenAI embeddings. The prompter uses that embedding selection separately from the chat/coding model. Configure the corresponding credentials when using the profile unchanged.
The supplied settings contain development-bot and player configuration. They also enable MCP and allow_insecure_coding; the README explains what to review for a new instance. Generated JavaScript execution is a runtime capability, not evidence that the harness is an operating-system security boundary.
eval_db.py defines the local SQLite ledger and its PINCER_EVAL_DB override. campaign_report.py analyzes existing attempts and applies recorded exclusions; it does not collect a new experiment. A fresh clone has no populated ledger.
The results document the existing 356-attempt primary set, exclusions, calibration limits, and experimental deviations. The runbook documents collection. The reported experiment provides evidence for that task set; it does not establish general long-horizon reliability.