Skip to content

AGENTS.md

Latest commit

 

History

History
168 lines (151 loc) · 12.8 KB

AGENTS.md

File metadata and controls

168 lines (151 loc) · 12.8 KB
Raw

NeqSim — Agent Instructions

Read automatically by VS Code Copilot, OpenAI Codex, Claude Code and other agents. This file is kept short on purpose because it is resent on every model call. It is already in your context — do not read_file it. Detail lives in:

  • docs/development/AGENT_REFERENCE.md — full workflow, results.json schema, Automation API, code patterns, key paths, JavaDoc and documentation rules. Search it for the heading you need; never read it whole.
  • Skills in .github/skills/ (find them with uv run --project devtools --locked devtools/skill_search.py "<topic>").
  • CONTEXT.md — 60-second repo map. docs/development/CODE_PATTERNS.md — starters.

Hard constraints (CI fails otherwise)

  • Java 8 everywhere, tests included. Never var, List.of/Set.of/Map.of, String.repeat (use StringUtils.repeat), isBlank/strip/lines (use trim()), text blocks, records, pattern-matching instanceof, Optional.isEmpty().
  • Logging: Log4j2 only (private static final Logger logger = LogManager.getLogger(X.class);, parameterized logger.info("{}", v)). Never System.out/err.println, also not in tests or examples.
  • Spotless: after creating or editing ANY .java file run .\mvnw.cmd spotless:apply (Linux/Mac ./mvnw), then git add. CI runs spotless:check. Never --no-verify.
  • JavaDoc on every class and method, private included: description, @param for each parameter, @return, @throws for every declared exception, class @author/@version. HTML5: <table> needs <caption>, no summary=; @see only with Java references; no </p> after </ul>/</ol>; no -> or raw > in JavaDoc (&gt;). Check with mvnw javadoc:javadoc. Spotless/Checkstyle/JavaDoc failures: neqsim-code-hygiene skill.
  • Serializable classes: non-serializable fields (Function, Consumer, Thread, JDBC, interpolators) must be private transient (SpotBugs SE_BAD_FIELD).
  • Verify APIs before use: find the class, read constructor/method signatures, never assume convenience overloads, enum locations or getter arguments.
  • Discover across all functionality: in a source checkout use uv run --project devtools --locked devtools/capability_search.py "<task or API>" --json. It returns current source types, workflow agents/skills, guides and test references; load only the relevant skill sections. In plugin/MCP-only sessions use runtime discovery. Read the discovery guide before declaring functionality absent. Routing coverage and test references do not establish qualification.
  • Agent/MCP coverage: src/main/resources/neqsim/mcp/engineering-coverage.json is a generated snapshot, not a source of truth. After Java formatting or agent/skill mapping changes, regenerate it with uv run --project devtools --locked devtools/build_engineering_coverage.py and stage the output. For merge conflicts, resolve source and registry conflicts first, then regenerate and git add this file. Never hand-merge its JSON or combine branch copies; neither produces a valid inventory of the merged sources. CI rebuilds and validates the snapshot, then commits changed output back to same-repository PR branches and master when permissions and branch protection allow. Fork PRs remain validation-only. Register new engineering operations with routes and evidence; source presence is not validation.
  • AI workflow names: GitHub Actions workflows for agent, skill, MCP, task, or research tooling MUST use ai-<purpose>.yml in lowercase kebab-case, directly in .github/workflows/. Categorize by primary responsibility: publishing MCP artifacts still uses publish-. Follow .github/WORKFLOWS.md; update filename references when renaming and preserve required-check names unless their consumers have been reviewed.
  • Documentation code must be test-verified: every Java/Python snippet in docs gets a JUnit test (e.g. DocExamplesCompilationTest) that is run and passes before the doc is done.
  • Notebooks must be executed end to end (NeqSim Runner by default); unexecuted = incomplete.
  • Linux-only engines (FluidMagic, OpenFOAM, OPM Flow) need Docker Desktop/WSL2 on Windows; run neqsim docker ensure --linux before container work to start installed Docker automatically and wait for readiness (source checkout: uv run --project devtools --locked devtools/docker_runtime.py ensure --linux). Never install/elevate, switch contexts or container modes, or stop Docker automatically. Report missing installations, startup timeouts and denied access as blockers, never skip silently.
  • FluidMagic is Equinor-internal, not on PyPI (enterprise agents/skills only): running it needs the private repo equinor/FluidMagic (own GitHub SSO; gh repo clone equinor/FluidMagic, pip install -e FluidMagic/fluidmagic). Never pip install fluidmagic; if import fluidmagic fails, tell the user how to get it and fall back to the offline hand-off.

Running calculations

  • MCP first: for a single flash/PVT/process/standards/sizing calculation use a curated mcp_neqsim_* tool (confirm fields with getSchema/validateInput), then runCapability (search → inspectApi → invoke). Write Python/Java only for loops, plotting, state, notebooks, reports, or inside /solve-task. If NeqSim lacks a capability, add it in Java + tests.
  • Python: use the selected interpreter, else C:\appl\neqsim-venv\Scripts\python.exe. Never bare python/pip/pytest, never create/activate/select environments; children reuse sys.executable. Missing package or interpreter = report the blocker.
  • Task notebooks load workspace classes through devtools/neqsim_dev_setup.py (ns = neqsim_init(project_root=...), then ns.* / ns.JClass); never from neqsim import jneqsim. Java not found → uv run --project devtools --locked devtools/java_locator.py / ensure_java_home() before declaring it missing.
  • Build/test: mvnw.cmd install; single test mvnw.cmd test -Dtest=Class#method; static analysis checkstyle:check spotbugs:check pmd:check.
  • "Package and update python": .\mvnw.cmd package -DskipTests, then copy target\neqsim-<version>.jar into ...\site-packages\neqsim\lib\ (flat, not a java11 subfolder).

Physics and API gotchas

  • EJML-only numerical code: Use primitive arrays at public numerical boundaries and EJML or LinearAlgebraOperations for calculations. Do not restore a JAMA dependency or reference backend. Former matrix-specific hooks and accessors now use EJML SimpleMatrix; downstream callers must migrate. Preserve dimensions, orientation, ownership and subclass-hook dispatch. EJML LU setA can accept a singular matrix: reject non-finite results. Pure-component phase-envelope points must satisfy the EOS pressure equation for both volume roots as well as fugacity equilibrium; a spinodal-stalled volume is not a saturation root.
  • Always setMixingRule(...) (numeric for CPA); createDatabase(true) for new components. Temperatures in K and pressures in bara by default. clone() fluids before branching.
  • After any flash call fluid.initProperties() before reading density/viscosity/conductivity; init(3) alone leaves transport properties at zero. Choosing init(...) levels in Java: neqsim-thermodynamic-initialization skill.
  • Phase envelopes: classify branches physically (the branch holding the cricondentherm is the dew curve); do not trust getter/label names alone — see neqsim-phase-envelope.
  • Multi-area plants: one ProcessSystem per area, combined in a ProcessModel (plant.add("area", ps)); never add a ProcessModel/ProcessModule to a ProcessSystem.
  • Loops: prefer process.makeRecycles() / setAutoRecycles(true) to hand-written tears. runUntilConverged(n) self-tunes tolerances — do not set a tolerance "to be helpful".
  • Agents drive models through process.getAutomation() (cached facade): string addresses, getVariableValueSafe/setVariableValueSafe, evaluate(...) as the never-throwing optimizer step gated on feasible, getUtilizationSnapshot() as the observation, newOptimizer() for AgenticProcessOptimizer. Details: neqsim-agentic-process-optimization skill.
  • Separator/vessel geometry and internals go through SeparatorMechanicalDesign, not Separator.
  • New equipment extends ProcessEquipmentBaseClass, implements validateSetup() with remediation hints.

Solving engineering tasks

Use the solve-task agent; its workflow is the neqsim-task-workflow skill (load per phase). Non-negotiables:

  1. Create the task folder first: neqsim new-task "title" --type X --author "Name" --prompt "<verbatim request>" (neqsim not found → <python> -m neqsim_cli ..., same arguments). Root: neqsim --show-task-root (explicit > NEQSIM_TASK_ROOT > ~/.neqsim/task_defaults.json > task_solve/). Pass the absolute task path to every child agent/tool; resume existing tasks in place; never fall back silently.
  2. All output and downloaded documents stay in the task folder — documents under step1_scope_and_research/references/<source>/; run uv run --project devtools --locked devtools/generate_sources_md.py <task> --organize for SOURCES.md + manifest. Search the optional document root (neqsim documents PATTERN, read-only) and run neqsim fetch-docs <task> before calling a standard, datasheet or P&ID unavailable.
  3. Standard/Comprehensive: capability_assessment.md, analysis.md, skill + agent discovery (devtools/skill_search.py, devtools/agent_search.py), benchmark validation against independent data, Monte Carlo P10/P50/P90 + tornado, ISO 31000 risk register, devtools/consistency_checker.py before the report.
  4. results.json passes TaskResultValidator / devtools/validate_task_results.py, with assumptions and data_gaps always filled (schema: AGENT_REFERENCE.md Part A).
  5. Report: neqsim report <task> (Word+HTML, --pdf optional). Title from study.title, language from report.language. Hand-written prose goes in step3_report/report_sections.json — never open, edit or copy generate_report.py (~320 KB). The report run also writes WORK_RECORD.md; fill its narrative blocks.
  6. Improve the tooling every task (Java + JUnit for NeqSim gaps, SKILL.md/*.agent.md for tooling gaps), commit and push each fix to the repo that owns it, and record it in neqsim_improvements.md and results.json improvements (or state that nothing was needed). Never commit task output or company data to a code repo.
  7. Close the loop upstream. Prefer the connected GitHub capability when available or required. Honor existing user/campaign authorization for issues and PRs; otherwise ask once whether to file it on equinor/neqsim now: neqsim file-issue <task> (--list to preview, --pr to also propose opening a PR for an already-implemented gap). Every issue and PR is confirmed interactively (or needs --yes) and --dry-run never calls gh/git — never file or push without authorization. --yes can reuse existing authorization. Remove private task data. Without a connector, no gh/no local checkout (the common case for plugin users) falls back to a pre-filled github.com/.../issues/new link opened in a browser — a form is not a filed issue. Details: neqsim-task-workflow §6.4.
  8. Close with a privacy-safe docs/development/TASK_LOG.md entry.

Quick tasks skip the ceremony: answer with units, assumptions and a validation note.

Token budget (every call resends the whole context)

  • Search before reading: grep/grep_search then read the matching range. Never read whole TASK_LOG.md, AGENT_REFERENCE.md, generate_report.py or large skills end to end.
  • Load a skill only when its topic is in play; load neqsim-task-workflow sections per phase.
  • Documents: metadata first (STID get_doc_references() gives titles/types), extract text/OCR in a script, write a small JSON summary, read that. Use view_image only on cropped regions.
  • Keep tool output small (counts, Select-Object -First, no full JSON/notebook dumps).
  • Bulk work goes to scripts or subagents that write files; read only their summaries.

Documentation

Docs in docs/ need YAML front matter (title, description; quote values with colons), KaTeX $...$ / $$...$$ (never \( \[), no markdown inside HTML blocks, verified relative links, and entries in docs/REFERENCE_MANUAL_INDEX.md plus the section index.md. Full rules: AGENT_REFERENCE.md Part B.

Continuous improvement

Improving the agents and skills used — and their hand-offs — is part of every task. Fix missing patterns, API signatures and gotchas in the owning SKILL.md/*.agent.md, add cross-references both ways, and keep site-specific detail in the enterprise repos.