Files
microsoft-SkillOpt/docs/sleep/plugin_load_test.md
2026-06-15 10:21:30 +00:00

3.4 KiB

SkillOpt-Sleep — plugin load-test (fresh examples)

This records an actual end-to-end load-test of all three plugin shells on a brand-new example (not the gbrain benchmark seeds), run on 2026-06-08.

The fresh persona

A data analyst whose SQL queries must always include a LIMIT clause — built from scratch for this test. Two forms were used:

  1. Real transcripts — crafted Claude Code session JSONL where the analyst asks for SQL, the agent forgets LIMIT, and the user complains ("you forgot a LIMIT again", "always cap results"). This exercises the real harvest → mine pipeline.
  2. Checkable tasks — the same intent with a rule judge (regex: (?i)LIMIT\s+100), so the optimizer can be scored on whether future SQL follows the house rule.

Results

Shell plumbing (all three drive the engine)

Shell What was run Result
Claude Code (scripts/sleep.sh) harvest, full run, adopt harvest found 2 sessions → 2 tasks; run staged a proposal; adopt honored the safety contract (no live change when nothing was accepted)
Codex (install.sh + shared runner) install.sh into a temp HOME placed the user-level ~/.agents/skills/skillopt-sleep/SKILL.md skill correctly and moved any legacy custom prompt aside instead of installing one
Copilot (mcp_server.py) initializetools/listtools/call sleep_harvest 5 tools listed; sleep_harvest returned real engine output (2 sessions → 2 tasks)

Genuine improvement (real model, fresh persona)

Optimizer Claude Sonnet 4.6 → target Claude Haiku 4.5, 3-way split (5 train / 2 val / 5 test), scored on the held-out test queries; and the same fresh persona self-optimized on Codex:

Backend Held-out test (fraction of SQL with LIMIT 100) before → after
Claude (Sonnet → Haiku) 0.00 → 1.00
Codex 0.00 → 1.00

In one night each optimizer wrote, into the protected learned block, a rule like:

"OVERRIDE: Every SQL query you generate MUST include LIMIT 100 …" (Claude) "Hard requirement: every SQL query response must include …" (Codex)

and the target then applied it to the unseen test queries. This is the whole claim on a task family the engine had never seen: it learned the user's house rule from their failures and generalized it — confirmed on both backends.

An honest finding from load-testing

The first attempt used val_fraction=0.34, test_fraction=0.34, which left only 1 train task for an 8-task set — too little signal — so reflect produced nothing and the night was a no-op (val already 0.75). Re-balancing the split to a real train pool (5 train) fixed it and produced the 0 → 1.00 result above. This is exactly the kind of issue that only surfaces when you actually run the thing, and it motivates a future guardrail: warn when the train pool is too small for the chosen split fractions.

Reproduce

The checkable persona run (real Claude):

# see the snippet in docs/sleep/plugin_load_test.md history, or run:
python -m skillopt_sleep.experiments.run_experiment --persona programmer --assert-improves  # deterministic

Shell checks:

# Copilot MCP server
printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' \
  | SKILLOPT_SLEEP_REPO="$(pwd)" python3 plugins/copilot/mcp_server.py
# Codex skill installer (into a throwaway HOME)
HOME=$(mktemp -d) bash plugins/codex/install.sh