# Harness Engineering

## Purpose
This document defines how FESA uses long-running agent harnesses for planning, implementation, and evaluation.

The goal is not to maximize agent count. The goal is to keep long solver work coherent, testable, and reference-verified across context resets and independent sessions.

## Default Harness Shape
Use the smallest harness that can safely handle the task.

For meaningful solver implementation or phase execution, use:

```text
Planner -> Generator -> Evaluator
```

Roles:
- `Planner`: turns project docs and `PLAN.md` tasks into a testable sprint contract or phase step.
- `Generator`: implements exactly one accepted contract using TDD.
- `Evaluator`: independently checks the result against the contract, docs, tests, reference artifacts, and validation commands.

Do not use multi-agent ceremony for tiny documentation edits or obvious mechanical changes. Do use the full harness when a task touches solver behavior, numerical conventions, reference comparison, parser compatibility, result schema, or phase execution.

## Sprint Contract
Every implementation sprint must have a contract before code changes begin.

Recommended location:
- `phases/{phase}/stepN.md` for phase execution.
- `phases/{phase}/contracts/stepN-contract.md` only when a separate negotiation artifact is useful.

Required sections:

````markdown
# Sprint Contract: {name}

## Objective
{one concise outcome}

## Required Reading
- /AGENTS.md
- /PROGRESS.md
- /PLAN.md
- /docs/README.md
- /docs/HARNESS_ENGINEERING.md
- {topic docs}

## Scope
- {what may be changed}

## Allowed Files
- {paths or modules}

## Explicit Non-Goals
- {what must not be done}

## Tests To Write First
- {test files or test cases}

## Reference Artifacts
- {references/*.inp or references/*_displacements.csv, or "none"}

## Acceptance Commands
```bash
python scripts/validate_workspace.py
```

## Evaluator Checklist
- {contract-specific checks}

## Handoff Requirements
- Update PROGRESS.md for completed work.
- Update PLAN.md for future work or changed blockers.
````

Contract quality rules:
- The contract must be testable.
- The contract must identify unsupported Abaqus features rather than expanding support implicitly.
- The contract must state whether reference data is used.
- The contract must name file ownership boundaries to reduce conflicts.
- The contract must not prescribe formulas that are not present in `docs/MITC4_FORMULATION.md` or a cited source.

## Generator Rules
The Generator implements one contract at a time.

Required behavior:
- Read the contract and required docs before editing.
- Write or update tests before implementation.
- Keep changes inside allowed files unless the contract is updated first.
- Preserve architecture boundaries from `docs/ARCHITECTURE.md` and `docs/ADR.md`.
- Preserve numerical conventions from `docs/NUMERICAL_CONVENTIONS.md`.
- Run acceptance commands.
- Update `PROGRESS.md` and `PLAN.md` only for factual state changes.

Generator failure modes to avoid:
- Broad refactors outside the contract.
- Implementing parser support because a stored reference `.inp` contains unsupported Abaqus features.
- Comparing only reduced vectors when full-vector reaction recovery is required.
- Treating a passing compile as sufficient without tests or reference checks.

## Evaluator Rules
The Evaluator is independent from the Generator.

Evaluation order:
1. Read the sprint contract.
2. Read `AGENTS.md`, `PROGRESS.md`, `PLAN.md`, and the topic docs.
3. Inspect the changed files.
4. Run or review the acceptance commands.
5. Check tests, reference artifacts, and documented conventions.
6. Return pass/fail findings with concrete file references.

The Evaluator must fail the sprint if any of these are true:
- Required tests were not written first or are missing.
- `python scripts/validate_workspace.py` fails without explanation.
- A CRITICAL rule in `AGENTS.md` is violated.
- A change drifts from `docs/ARCHITECTURE.md`, `docs/ADR.md`, or `docs/NUMERICAL_CONVENTIONS.md`.
- `references/*_displacements.csv` comparison is required but not implemented or not checked.
- `RF` is computed from reduced quantities when full-vector recovery is required.
- Unsupported Abaqus features are silently accepted.
- Completed work is not recorded in `PROGRESS.md`, or future tasks are not recorded in `PLAN.md`.

If the sprint fails, the Evaluator should produce a concise feedback artifact:

```markdown
# Evaluation Feedback: {contract}

## Verdict
fail

## Findings
- {severity}: {file} - {risk}

## Required Fixes
- {minimal fix}

## Verification To Rerun
- {commands}
```

## FESA Evaluation Rubric
Use this rubric for implementation review.

| Criterion | Pass Condition |
|---|---|
| Contract compliance | Changes stay within scope and allowed files |
| Architecture | Domain, AnalysisModel, AnalysisState, DofManager, adapters, and factories follow documented ownership |
| Numerical conventions | DOF order, units, signs, double precision, int64 ids, constrained/free mapping, and full-vector reactions are preserved |
| Reference verification | Stored `references/` artifacts are used when required; CSV column mapping is correct |
| Tests | Tests exist before implementation and cover failure modes, not only happy paths |
| Diagnostics | Unsupported input and singular systems produce actionable diagnostics |
| Results schema | Outputs follow step/frame/field/history and HDF5 schema rules |
| Handoff | `PLAN.md` and `PROGRESS.md` reflect the new state |

## Harness Complexity Policy
Add harness complexity only when it catches real risk.

Use a single agent for:
- small wording changes.
- mechanical docs updates.
- metadata-only corrections.

Use Planner -> Generator -> Evaluator for:
- C++ solver implementation.
- parser behavior changes.
- result schema or HDF5 writer changes.
- reference comparator changes.
- MITC4 formulation-dependent work.
- phase generation or execution.

Review the harness periodically. If an agent role no longer adds value, simplify it.