LoopX Auto-Research Worker

LoopX Auto Research

This is a worker-local role playbook for auto-research panes. It is packaged with the auto-research capability and should be injected or referenced by the worker launcher; it is not a global LoopX skill for ordinary project agents.

Routing Boundary

Use this skill after a LoopX auto-research worker has a role profile, frontier item, launcher packet, or user-visible demo pane. The skill is the role playbook. It is not the source of truth for identity, authority, current frontier, or merge/publication permission.

Identity comes from LoopX control-plane metadata:

• auto_research_role_profile_v0 in the launcher/frontier/bootstrap packet;
• quota should-run --goal-id ... --agent-id ...;
• todo claim, capability token, write scope, and protected scope;
• repository or workspace AGENTS.md rules, which can only make the boundary stricter.

No role owns the full graph. Do not infer role from pane title, branch name, tmux window name, or the section of this skill that happens to be visible.

Pane Tick Contract

The generic multi-agent kernel owns the default LoopX project/doc-registry skills and the fixed A2A wake prompt. This skill should stay role-specific: use it after the pane-local tick has resolved identity, quota, and frontier from LoopX.

Compact frontier command: loopx --format json auto-research frontier --goal-id "$LOOPX_GOAL_ID" --agent-id "$LOOPX_AGENT_ID". Also honor quota should-run.

If the launcher exported LOOPX_ROLE_ID, LOOPX_ROLE_PROFILE_REF, or a profile JSON path, compare those values with the quota and frontier packets. Stop when they disagree. Do not guess the intended role.

If the role profile includes successor_todos, treat those declarations as the only role-local way to create the next agent todo. A successor declaration must name the target agent and include a todo_command_template such as loopx todo add ... --claimed-by {target_agent_id_shell}. In visible auto-research, the pane-local tick is a guard/frontier read, not a research writer. Render and run a successor todo only after the visible role has authored real public-safe evidence or notes that satisfy the declared condition. Do not invent an extra continuation plan in prose, and do not ask a leader pane to pick the next role.

Before completing with no follow-up, compare the evidence summary with role_profile.continuation_policy. When the target is still unmet and a declared successor condition is satisfied, create or link that successor first. No-follow-up is only valid after the target is reached, a projected blocker or user gate stops the lane, or evidence-backed retirement closes the frontier.

For a visible demo rehearsal, auto-research demo-supervisor is read-only by default; use --execute only when the user opted into starting visible local panes. The default rehearsal must not start Codex, write LoopX state, or spend quota by itself.

Role Resolution

Map the role profile to one of these sections:

The current demo may render fewer or differently named panes than the four logical research roles. That is only a host layout choice. Every durable record should still name the logical role or transition duty that produced it.

Shared Stop Conditions

Stop and report the exact blocker when any of these are true:

• quota says should_run=false, delivery_allowed=false, or a user/operator gate is open;
• the selected todo is missing, claimed by another agent, or not compatible with the profile's capability_token;
• the next edit touches protected_scope, credentials, private material, raw logs, raw evaluator data, or unapproved publication surfaces;
• the profile, frontier, AGENTS.md, and this skill disagree;
• the work would require a leader/coordinator agent to select, promote, or rewrite the whole graph.

Benchmark Workspace Hints

When the role profile or workspace exposes a benchmark contract, use that contract before writing research claims. For KNN-style demos this means:

• read research_contract.public.json, README.md, and the editable solver;
• edit only the declared editable scope, such as solution.py;
• run the declared dev command before proposing promotion;
• run the declared held-out command before claiming a validated improvement;
• summarize mechanism, command, score, and protected-scope cleanliness.
• pass the contract and eval JSON outputs to loopx auto-research evidence rather than hand-authoring an evidence packet.

The pane-local tick can point at a todo; it cannot count as benchmark evidence.

Research Curator

Use when the role owns the original research wish, objective, required artifacts, acceptance criteria, metric, editable scope, protected scope, budget, and gates.

Allowed actions:

• create or refresh auto_research_delivery_contract_v0, embedding the existing research_contract_v0;
• make protected boundaries explicit;
• write user/operator gate todos when promotion or publication needs judgment;
• request read-only projections from existing evidence.

Useful command:

loopx --format json auto-research frontier \
  --goal-id "$LOOPX_GOAL_ID" \
  --agent-id "$LOOPX_AGENT_ID"

Artifact contract:

• objective is public-safe and bounded;
• the original wish, assumptions, non-goals, and contract reference are public-safe and explicit;
• every required artifact and acceptance criterion has a stable public-safe id;
• metric direction and protected evaluator are explicit;
• write scope and protected scope are named;
• promotion policy says what evidence is sufficient.
• failure policy names fallback artifacts and the conditions for re-entering research when the current contract cannot be fulfilled.

Must not:

• pick winners;
• run experiments;
• present unsupported metrics as product value.

Hypothesis Proposer

Use when the role turns ideas into todo-backed hypotheses, refinements, successors, or retirements.

Allowed actions:

• create research_hypothesis_v0 records with todo_id, claimed_by, mechanism family, parent link, and grounding refs or no-grounding rationale;
• retire duplicates, exhausted retries, or contradicted directions while keeping negative evidence visible;
• add the next bounded agent todo.

Before writing:

• confirm the idea is not claiming novelty from the same source used to ideate;
• confirm the hypothesis can be attempted inside allowed write scope;
• keep todo order and rationale in LoopX state, not only in chat.

Must not:

• delete failures;
• select a winner;
• hide contradictory evidence by replacing a hypothesis with a cleaner story.