Experiment Iterative Coder

Iterative Coder

Iterative code refinement through structured plan → code → evaluate → refine cycles. Each cycle runs objective checks (lint, tests) and self-evaluation, then diagnoses failures and plans targeted improvements. Reaches production quality in 3-8 iterations.

When to Use This Skill

• Main agent delegates a code task prefixed with "MODE: MORE_EFFORT"
• User selected "More Effort" mode for code generation
• Task requires high code quality with verified correctness
• Task involves complex implementation (5+ files, multiple modules)
• You want to iterate on code quality rather than submit first-pass code
• You mention "iterative refinement", "code quality loop", "plan-code-evaluate"

The Iteration Mindset

Code quality comes from fast feedback loops, not careful first attempts. A fast plan → code → evaluate → fix cycle beats spending 30 minutes on a "perfect" first implementation. The evaluate step reveals problems you cannot predict by thinking alone — lint errors, import failures, test regressions, and missing edge cases all surface immediately when you actually run the code.

Before Starting: Load Context

1. Read /memory/experiment-memory.md for proven strategies from past cycles (skip if it doesn't exist)
2. Identify existing tests, linting config (pyproject.toml, ruff.toml), or CI setup in the workspace
3. Check available tools:

   ruff --version 2>&1; echo "---"; python -m pytest --version 2>&1

   If either is missing, you will skip that check during evaluation (do not fail the iteration).

Phase Decomposition

Before iterating, analyze the task and break it into sequential phases:

For each phase, define:

• Name: concise label (e.g., "Data loading pipeline")
• Goal: what "done" looks like for this phase
• Verification signal: how to confirm the phase is complete (specific test, lint clean, output matches)

Order phases by dependency — later phases may build on earlier ones.

The Iteration Loop

For each phase, iterate up to 3 times. Global maximum: 10 iterations across all phases.

Step 1: Plan

Read current code and previous evaluation feedback (if any). Write a concise improvement plan.

First iteration of a phase: Write an initial implementation plan based on the phase goal.

Subsequent iterations: Analyze the last evaluation's feedback and diagnose the root cause of failures before planning changes. Do not repeat the same approach that already failed.

Adapt your plan based on the failure mode from the last evaluation:

Step 2: Code

Implement the plan. Keep changes focused on what the plan specifies.

• Do not rewrite working files unless the plan explicitly requires it
• After writing code, do a quick sanity read of the changed files

Step 3: Evaluate

CRITICAL: You MUST run these commands every iteration. Do not skip evaluation.

# 1. Lint check
ruff check . 2>&1 | tail -20
echo "LINT_EXIT: $?"

# 2. Format check
ruff format --check . 2>&1 | tail -10
echo "FORMAT_EXIT: $?"

# 3. Run tests (only if test files exist in workspace)
python -m pytest -x -q --tb=short 2>&1 | tail -30
echo "TEST_EXIT: $?"

If ruff is not installed, skip checks 1-2. If pytest is not installed or no test files exist, skip check 3. Record which checks were skipped.

Step 4: Score

Compute a composite score from objective signals and self-assessment.

Objective signals (from Step 3 exit codes):

• LINT_EXIT=0 → lint_score = 1.0, else lint_score = 0.0
• FORMAT_EXIT=0 → format_score = 1.0, else format_score = 0.0
• TEST_EXIT=0 → test_score = 1.0, else parse pass ratio from pytest output (e.g., "3 passed, 1 failed" → 0.75)

Self-assessment (rate 0.0 – 1.0): Evaluate on: correctness (does the code do what was asked?), completeness (all requirements addressed?), error handling (reasonable edge cases covered?), readability (clear names, structure).

Composite score — dynamic weighting based on available signals:

• Lint + tests available: 0.2 × lint + 0.1 × format + 0.3 × test + 0.4 × self
• Lint only (no tests): 0.3 × lint + 0.1 × format + 0.6 × self
• Tests only (no ruff): 0.4 × test + 0.6 × self
• Neither available: 1.0 × self

Self-assessment hard caps — prevent score inflation from self-assessment:

• If lint check FAILED → composite capped at 0.4, regardless of self-assessment
• If any test FAILED → composite capped at 0.6
• Only claim composite ≥ 0.85 if BOTH lint and tests pass AND implementation is complete
• Deductions: missing error handling for obvious cases (−0.1), hardcoded absolute paths (−0.05)

See references/evaluation-protocol.md for detailed scoring edge cases.