Evo Exprmn Optmzr

Goal

• Let Hermes run evo-style optimization without requiring Claude Code plugin surfaces.
• Use repo-local git worktrees as experiment branches.
• Keep only changes that improve a measurable benchmark and pass gates.
• Preserve traces, diffs, and branch history for later inspection.

Use when

• User wants benchmark-driven autonomous improvement.
• A repo has a clear target file/module and a benchmark command that emits a parsable score.
• The task benefits from branching search instead of one linear edit loop.
• You want Hermes to explore multiple hypotheses with rollback discipline.

Do not use when

• There is no executable benchmark.
• Success is subjective and cannot be scored.
• The repo is too dirty to establish a clean baseline.

Core model absorbed from evo

• Workspace init with explicit target, benchmark, metric, optional gate.
• Each experiment gets its own git worktree from a parent node.
• Benchmark output is parsed into a numeric score.
• Candidate is kept only if score improves over parent and gates pass.
• Failed branches are discarded; good branches become new parents.
• Traces, diff, and notes are saved per experiment.
• Prefer shallow tree search over blind repeated edits on one branch.

Hermes-native workflow

1. Preconditions

   • Verify repo is git-controlled and clean enough to branch.
   • Verify benchmark command runs from repo root.
   • Verify score extraction rule before any code edits.
   • Verify any required gate commands.

2. Define experiment contract in chat or repo note

   • target path(s)
   • benchmark command
   • metric direction: max or min
   • gate command(s)
   • score parse rule
   • budget: number of branches / rounds

3. Create baseline

   • Run benchmark on current branch.
   • Save score and command output.
   • Treat this as root parent.

4. Branch experiments with worktrees

   • For each hypothesis, create a git worktree on a branch like evo/<topic>/<n>.
   • Make one focused change per branch.
   • Run benchmark.
   • Run gates.
   • Save benchmark output, gate output, diff, and short hypothesis note.

5. Keep or discard

   • If score improves and gates pass: keep branch as candidate parent.
   • Else: discard branch or leave as explicit failed trace if user wants auditability.

6. Iterate

   • Spawn more experiments from the best surviving parent.
   • Use traces from failed runs to propose new hypotheses.
   • Stop on budget exhaustion, stall, or user interrupt.

7. Finalize

   • Report best branch, best score delta, and gate status.
   • Present exact diff or commit for human decision.
   • Clean stale worktrees if user scope permits.

Suggested repo artifact layout

• .hermes/evo/<run-id>/baseline.log
• .hermes/evo/<run-id>/experiments/<exp-id>/benchmark.log
• .hermes/evo/<run-id>/experiments/<exp-id>/gate.log
• .hermes/evo/<run-id>/experiments/<exp-id>/diff.patch
• .hermes/evo/<run-id>/experiments/<exp-id>/notes.md
• .hermes/evo/<run-id>/summary.md

Recommended commands

• baseline: use terminal() from repo root
• branches/worktrees: git worktree add ...
• diff capture: git diff <parent>...<branch>
• traces/logs: write to repo-local files, not memory

Score parsing

• Prefer benchmark commands that print one numeric score plainly.
• If output is noisy, extract with a stable regex or small Python parser via execute_code.
• Never guess score direction; make max/min explicit.

Gates

• Gates are pass/fail safety rails separate from benchmark score.
• Typical gates: tests, lint, typecheck, smoke run, invariant checks.
• A higher score that fails a gate is rejected.

Heuristics

• Change one variable per branch when possible.
• Keep hypotheses short and falsifiable.
• Prefer 2-5 high-signal branches over many low-quality ones.
• If benchmark is slow, reduce branch fanout and increase hypothesis quality.
• Use delegate_task only when the user already approved a multi-branch coding run and subtasks are independent.

Pitfalls

• Do not start before baseline is reproducible.
• Do not compare scores across changed benchmark definitions.
• Do not keep ungated wins.
• Do not let multiple branches edit overlapping files without clear isolation.
• Do not store experiment state only in chat; keep repo-local artifacts.

Output pattern

• baseline score
• best branch
• score delta
• gate result
• kept vs discarded branches
• next best experiment directions