Prosperity research session

Use this skill to run a durable research session for IMC Prosperity work.

Public-release note: this skill is the workflow glue for the repo. It tells an agent how to use the local MCP server, how to recover from long-running work, and how to avoid mistaking stale or cached evidence for fresh research.

Core behavior:

• Treat invocation of this skill as an explicit request to use subagents for non-trivial work.
• For non-trivial tasks, spawn bounded subagents early for parallel exploration, review, or verification.
• Keep final synthesis and implementation decisions in main thread.
• Use prosperityResearch MCP server as primary interface for session orchestration when available.
• If MCP server is not configured yet, fall back to python scripts/prosperity_research_cli.py ....
• If skill discovery misses this repo-local skill, use this file at .agents/skills/prosperity-research-session/SKILL.md directly and note that user-level install may still be needed.

When this skill is active:

1. Smoke MCP with tools/list or prosperityResearch.start_or_resume_session; verify a session manifest exists before using CLI fallback.
2. Inspect current frontier with prosperityResearch.get_session_status and prosperityResearch.get_top_candidates.
3. Before long alpha work, read prosperityResearch://alpha_autoresearch_protocol; use it as source of truth for OOS splits, keep/discard policy, lane policy, and continuous agent budget loop.
4. If frontier is weak or stale, continue loop with prosperityResearch.continue_session or prosperityResearch.run_alpha_autoresearch_loop.
5. For alpha autoresearch, pass explicit budget controls:
   • iteration_runtime_ms
   • max_total_runtime_seconds
   • stale_operation_timeout_seconds
   • response_mode: "compact" or "artifact_only" for long sweeps
6. MCP auto-detaches continue_session and run_alpha_autoresearch_loop when detach is omitted. Poll with prosperityResearch.get_compact_session_status or prosperityResearch.get_session_status; inspect active_operation, output_path, cancel_path, and artifact links before relaunching.
7. For ROUND_3/ROUND_4, replay concrete generated candidate labels only. Do not spend research budget replaying current algorithm.py as implicit fallback.
8. If there is a strong candidate, request a patch plan with prosperityResearch.request_patch_plan.
9. Summarize:
   • what was tried
   • what looks promising
   • what was rejected
   • what should be implemented next
   • what should be verified before implementation

CLI fallback:

• Use python scripts/prosperity_research_cli.py <tool> --input-file <payload.json> --output <result.json>.
• Prefer explicit input/output files for long loops so transport timeouts do not hide completed artifacts.
• Include same budget fields as MCP payloads. Use stable input/output files and inspect session artifacts before retrying after timeout.

Autoresearch/time-budget behavior:

• Treat wall-clock budget as real. Do not stop early only because a current best candidate exists unless user budget is exhausted, user cancels, or an explicit stop rule fires.
• run_alpha_autoresearch_loop is fixed-budget orchestration. It persists iteration_runtime_ms as per-iteration subprocess budget and respects max_total_runtime_seconds.
• Generator lane should synthesize or evaluate fresh generated labels before replay.
• Evaluation lane should run R3/R4 replay batches against explicit candidate labels.
• Algorithm experiment lane is optional and bounded. Use run_algorithm_autoresearch_experiments with agent_recipes[], research_program, and fixed_time_budget_seconds when the agent proposes direct algorithm.py changes.
• Continuous Karpathy-style autoresearch requires the outer agent to repeatedly propose new agent_recipes[], validate under a fixed budget, read the artifact/ratchet, update the hypothesis, and repeat until remaining budget is below the next validation slice. A static recipe catalog is not enough.
• Minimum loop for any multi-hour budget:
  1. Set a deadline from user time budget.
  2. Draft fresh non-duplicate agent_recipes[] from current artifacts, losses, and lessons.
  3. Call prosperityResearch.run_algorithm_autoresearch_experiments with fixed_time_budget_seconds <= remaining_time.
  4. Read returned artifact_path; inspect ratchet, best_variant, errors, and discarded variants.
  5. Generate next recipe from evidence. Do not repeat recipe after a dedup/cache hit.
  6. Continue until remaining time is below one validation slice or user cancels.
• Stop/relaunch when generated labels are empty, replay fingerprints repeat, or algorithm-cache fingerprints repeat; widen parents, recipes, or candidate generation before spending more replay budget.

Subagent policy:

• Spawn one bounded subagent per clearly separable investigative workstream for non-trivial work.
• Good subagent roles:
  • history reviewer
  • candidate reviewer
  • verification reviewer
  • code impact reviewer
• Each subagent must return:
  • scope inspected
  • key findings
  • risks or caveats
  • recommended next step

Output contract:

• Always return concise research summary.
• If there is a strong candidate, include:
  • candidate name
  • why it is promising
  • expected edge or diversification contribution
  • major risks
  • implementation outline
• If there is not yet a strong candidate, include:
  • top rejected ideas and why
  • top open questions
  • recommended next experiment batch

Guardrails:

• Do not claim strategy is good without evaluation evidence from research session.
• Do not recommend implementation before checking prior attempts and current frontier.
• Prefer small, testable iterations over broad rewrites.
• Do not give a final answer while a user-provided research time budget remains and the MCP is still able to validate more agent_recipes[].
• Inspect session manifest/artifacts after MCP transport timeout before retrying; long runs may have completed and persisted evidence.
• Keep negative evidence. Discard means no robust frontier advance, not deletion.