Agent Instructions Set up Steltic for Claude Code — USA edition Version 2026-11-06. For Claude Code in the Claude desktop app, running locally on the user's computer (environment "This computer"). Complete text: https://stelticai.com/agent-instructions-claude-code.txt 0. BEFORE YOU START Work from the complete text of these instructions. If what you have is a summary of the stelticai.com page rather than this full text, download this file verbatim and read the whole file from disk before doing anything else: curl -fsSL https://stelticai.com/agent-instructions-claude-code.txt -o STELTIC_SETUP.txt Windows PowerShell: curl.exe -fsSL https://stelticai.com/agent-instructions-claude-code.txt -o STELTIC_SETUP.txt You are Claude Code, running locally on the user's computer in the folder the user opened for this session. This session is the Main session: the user's own Claude Code conversation. Set up a reusable Steltic team in this folder, using the US repositories and the instructions below. Complete the installation, configure the agents, test their communication and software, and report the actual readiness of each role. Work through routine setup autonomously. Request only genuinely missing inputs or platform-required access, while continuing independent work. ALL LLM WORK MUST USE CLAUDE CODE'S OWN MODEL. Do not connect to an external LLM, request an LLM API key, configure an LLM provider endpoint, or start a repository's API-backed model loop. The Main session and its subagents supply the reasoning, planning, engineering calculations, interpretation and narrative. Steltic supplies the numerical engines, data contracts, retrieval utilities, validators and renderers. Installing a Steltic web app or the Steltic Hub is not a substitute for creating this agent team. The Claude Code behaviour described below is as documented at https://code.claude.com/docs when this version was written. Check the current documentation and the tools actually available in this session before relying on it, and adapt where it has changed. 1. CREATE THE TEAM Create these nine agents with exactly these display names, capitalisation and spelling. The numbers in this document are not part of their names: Admin Query file manager HR Steel CFS Steel Design variations Probabilistic analysis User-designed building Steltic Nonlinear openBIM drawings The underlying software module for Probabilistic analysis is named steltic_probabilistic. Do not add "App", "Bot", "SNL" or another suffix to any name. The Main session is the user's own Claude Code session, not another agent to create or rename. Reporting structure: User <-> Main session <-> Admin |-- Query file manager |-- HR Steel |-- CFS Steel |-- Design variations |-- Probabilistic analysis |-- User-designed building |-- Steltic Nonlinear `-- openBIM drawings The Main session creates or resumes Admin and delegates Steltic setup and subsequent Steltic work to Admin. Admin creates or resumes the eight specialists and owns their coordination. Every specialist reports to Admin. Admin reports to the Main session. Among the Steltic team, only Admin may communicate directly with the Main session. The user reads and writes in the Main session, so the Main session relays Admin's messages to the user without changing their substance, and relays the user's answers back to Admin. Specialists send all progress, questions, blockers, decisions and final reports to Admin, including anything an upstream instruction calls "ask the user" or "tell the user". The Main session routes Steltic requests through Admin rather than messaging specialists. All nine Steltic agents may communicate directly with one another about assigned work. In particular, every specialist and Admin may query Query file manager directly. Peer collaboration does not change reporting lines or authorise a change to the user's scope. Send Admin a concise record of consequential peer decisions and completed handoffs. The Main session requests any retrieval it needs through Admin. How to build the team in Claude Code: a. Agent definitions. Write one project subagent file per role in .claude/agents/ under STELTIC_ROOT, with YAML frontmatter and the role's instructions as the body. Use these definition ids as the file names and as the frontmatter name: admin, query-file-manager, hr-steel, cfs-steel, design-variations, probabilistic-analysis, user-designed-building, steltic-nonlinear, openbim-drawings. State the exact display name in each description and in the first line of each body. Set model: inherit, so every role uses this session's model, and memory: project. Preload each role's skills with the skills field (section 2). Keep the tools each role needs: the shell tools (Bash and/or PowerShell), Read, Write, Edit, Glob, Grep, SendMessage, Monitor where available, and the Agent tool for Admin, Design variations and User-designed building. Each body contains the role instructions from section 4, the reporting rules above, the shared query protocol (section 3), the roster, the relevant source paths and the exact environment commands for that role. b. Project instructions. Write CLAUDE.md at the root of STELTIC_ROOT. It tells every session opened in this folder: this is a Steltic installation; read steltic_state.json; route every Steltic request to Admin under the rules above; do not message specialists directly; if setup is incomplete, resume it from the recorded state. Keep it short and point to the state file and the definitions for detail. c. Settings. In .claude/settings.json set the environment variable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS to "0" for this project. With agent teams enabled, named subagents launch as teammates, which cannot create teammates of their own and are not restored when a session is resumed; this team needs ordinary nested subagents. Do not reduce CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH below its default of 3 (Main session -> Admin -> specialist -> an HR Steel instance started by Design variations). Do not edit permission settings to approve your own actions. If approval prompts are frequent, tell the user which permission mode or allow rules would reduce them. Application-required approval prompts remain subject to Claude Code's controls. d. Loading the definitions. Claude Code watches only the agent and skill directories that existed when the session started. If .claude/agents/ or .claude/skills/ is new in this folder, first finish the bootstrap (section 2) and save steltic_state.json. Then ask the user to start a new Claude Code session in this same folder (Code -> New session -> This computer -> this folder) and type "Continue Steltic setup". CLAUDE.md and the state file tell the new session where to resume. Do not substitute general-purpose subagents carrying pasted role text for the defined agents. e. Starting the agents. Once the definitions are loaded, the Main session starts Admin with the Agent tool from the admin definition, naming the instance with its exact display name, "Admin", so it can be addressed with SendMessage. Admin starts Query file manager first, so later agents see it in their roster, then the other seven specialists, each from its own definition and named with its exact display name. If Claude Code rejects a name containing spaces, use the definition id as the address, record the mapping, and keep the exact display names in every report. Record every instance's agent ID in steltic_state.json. Admin may run further instances of a specialist for separate jobs, for example one HR Steel instance per design variation; name them " · ", record them, and keep each in its own job folder. f. What these agents are. Subagents start when work arrives and can be resumed by name or ID with their history; messaging a stopped subagent resumes it. They are not always-running processes. Persistence comes from the definitions, agent memory and the state files in STELTIC_ROOT: a new session re-creates each role from its definition and resumes its recorded work. Describe them that way. Do not represent saved prompt files or role-play inside one conversation as nine running agents. g. Long runs. Run commands longer than a couple of minutes as background shell commands, monitor them, and keep their logs in the job folder. The computer must stay awake and the Claude desktop app open while work runs; say so before starting any run expected to take more than about 15 minutes. Verify that Admin can delegate to its specialists and receive their results, and that specialists can address peers without going through the Main session. Use SendMessage between named agents, or a private inbox under STELTIC_ROOT/comms/ when that provides real delivery and acknowledgement. Check access to any shared path from the receiving agent before relying on it. If Claude Code cannot support the required hierarchy, persistence or communication, report the precise limitation through Admin, retain completed installation work, and mark the affected capability BLOCKED. Do not silently replace the team with a flat group that reports to the Main session. 2. INSTALL AND SAVE THE CONFIGURATION STELTIC_ROOT is the folder the user opened for this session; record its absolute path. If it already contains a Steltic installation (steltic_state.json), reuse it: do not duplicate agents or overwrite jobs. If it contains unrelated files, create a Steltic subfolder inside it and use that. Specification PDFs and converted-corpus files are Steltic inputs, not unrelated files: leave them where they are and record their paths for Query file manager. On Windows, prefer a short local path outside OneDrive-synced folders. Layout, created as needed: STELTIC_ROOT/ CLAUDE.md routing rules for every session in this folder steltic_state.json roster, agent IDs, repository commits, environment paths, routes, job index, readiness .claude/agents/ the nine role definitions .claude/skills/ the packaged Steltic skills .claude/settings.json project settings (agent teams off) repos/ the six repositories envs/ one Python environment per module engineering_rag/ retrieval corpus: documents/, indexes/, search/, queue/in, queue/out, scripts/ standards/ the user's licensed PDFs, or a recorded pointer to their location jobs// one folder per job, with its owner recorded comms/ inbox records and acknowledgements, if used setup/ setup commands, logs, test outputs and the readiness report Bootstrap (Main session, before Admin exists): check the operating system, git, the Python launchers, uv, free disk, CPU count and memory, and the Claude Code tools available in this session. Clone the repositories into repos/ and record each commit. Write the definitions, skills, CLAUDE.md, settings and steltic_state.json. Everything after the bootstrap is Admin's work. Save enough state for the Main session to find Admin and for Admin to resume every role after this conversation ends. Reuse this installation on subsequent setup requests without duplicating agents or overwriting jobs. Obtain these official repositories, excluding the India editions: https://github.com/Steltic/steltic_grokbot https://github.com/Steltic/steltic https://github.com/Steltic/steltic_cfs https://github.com/Steltic/steltic_hub https://github.com/Steltic/steltic_nonlinear https://github.com/Steltic/steltic_openBIM Use the current default branch for a fresh installation and record each checked-out commit. Read the current README, dependency declarations, relevant agent contracts and module manifests before running commands. Keep the tested revisions fixed during a job. Reconcile version conflicts explicitly; do not mix an old tutorial's assumptions with a newer engine or silently weaken a validator. Use separate Python environments for the modules under envs/, normally Python 3.12 for the OpenSees modules, subject to the checked-out compatibility requirements. HR and CFS MUST have separate environments: both install packages named steltic, steel_engine and contract, with different implementations. Invoke the correct environment's interpreter explicitly by full path (envs\hr\Scripts\python.exe on Windows, envs/hr/bin/python elsewhere) and verify the imported engine's file location. Install the required native OpenSees libraries. Keep Docling in its own environment, pinned to 2.123.1 unless the repository requires a documented, fully revalidated replacement. Prefer uv to provision Python 3.12 where no suitable interpreter exists; the Hub's own installer uses uv. If Windows Smart App Control blocks Python ("Application Control policy has blocked this file" or os error 4551), use a signed python.org or uv-managed build as the HR README describes, and report it if the block persists. Discover the installed Tesseract data directory (on Windows usually under Program Files\Tesseract-OCR\tessdata) instead of assuming a Linux path. If Tesseract is missing, report what is needed and continue other work. Use steltic_hub as the source of the bundled modules, manifests and integration conventions. Admin, Design variations, Probabilistic analysis and User-designed building are bundled under: steltic_hub/catalog/steltic_admin/ steltic_hub/catalog/steltic_variations/ steltic_hub/catalog/steltic_probabilistic/ steltic_hub/catalog/steltic_userbuilding/ These paths are inside the steltic_hub repository. They are not separate GitHub repositories. Provision their dependencies and deterministic tools from the manifests. Do not start their servers: call their functions directly. The openBIM repository name is steltic_openBIM; its module ID is steltic_openbim. Keep these software IDs distinct from the exact agent display names. Adapt the upstream harness assumptions to Claude Code: old names "HR Steel App" and "CFS Steel App" mean HR Steel and CFS Steel here; Grok operations (CreateAgent, SendToAgent, CopyToBox, CopyFromBox) become the Claude Code Agent tool, SendMessage and local file paths; /workspace/... and other fixed paths become verified paths under STELTIC_ROOT. Install the packaged skills as project skills, one folder each under .claude/skills/. Each SKILL.md holds the packaged text unchanged, except that the frontmatter name is set to the folder name: steltic-querying <- steltic_grokbot/skills/Skill_querying_PACKAGED.md steltic-query-file-manager <- steltic_grokbot/skills/Skill_query_file_manager_PACKAGED.md steltic-check-and-clean <- steltic_grokbot/skills/CHECK-AND-CLEAN-CONVERSIONS.md steltic-snl <- steltic_nonlinear/skills/Skill_SNL_PACKAGED.md steltic-pushover <- steltic_nonlinear/skills/Skill_pushover_analyst_PACKAGED.md steltic-nlrha <- steltic_nonlinear/skills/Skill_nonlinear_dynamic_PACKAGED.md steltic-ddm <- steltic_nonlinear/skills/Skill_ddm_PACKAGED.md Record each source path and commit. Read the packaged skills as each agent's operating instructions and preserve their engineering and validation requirements. Contract and reference files (AGENT_START.md, BIM_EXPORT.md and the rest) stay in the checked-out repositories; the definitions point to them. Drive the repositories directly using Claude Code's own reasoning and shell/code tools. Where a bundled module normally asks an external LLM to plan variations, extract a description, fill calculations, interpret a report or write a review, the corresponding Claude Code agent performs that step and supplies the repository's expected structured input. Reuse existing engines, schemas, scoring, validation and rendering functions. A thin, tested local adapter is acceptable; replacing the engineering engine or bypassing its completion gates is not. If a step cannot yet be driven without an external model connection, report that specific integration gap and continue other setup. Never request an API key as the remedy. Require each agent to acknowledge that its instructions and paths are accessible. Keep job files separate, record which agent owns each output, and avoid concurrent writes to the same calculation package or model revision. Keep the user's licensed specifications and everything converted from them on this computer. Do not upload them, copy them into a repository or synchronised folder, or send them to another environment without the user's authorisation. 3. SHARED QUERY PROTOCOL Preload the steltic-querying skill (steltic_grokbot/skills/Skill_querying_PACKAGED.md) into Admin and every specialist. Query file manager must also understand this client protocol and execute incoming plans with its own retrieval tools. Only Query file manager owns PDF conversion, corpus repair and index rebuilding. Send one JSON retrieval plan per wave directly to Query file manager with SendMessage. The surrounding message identifies the sender and reply address; keep the plan compatible with the repository's schema: { "plan_id": "setup-HR-001", "source_job": "setup-demo", "queries": [ { "qid": "q1", "doc": "opensees", "type": "command", "query": "forceBeamColumn", "purpose": "Verify retrieval of the documented element command", "want_commentary": false, "context_neighbors": 1 }, { "qid": "q2", "doc": "ASCE7", "type": "exact_table", "query": "12.2-1", "purpose": "Retrieve the applicable seismic system provisions", "want_commentary": false, "context_neighbors": 1 } ] } Use unique plan_id values and qid values unique within each plan. Supported query types are exact_section, exact_equation, exact_table, fts, keyword, command and id. Each query names one document. Exact queries contain the identifier alone. Follow the skill's canonical IDs, aliases, lookup traps and standard/commentary distinctions. Query file manager writes each plan to queue/in/.json under engineering_rag/, runs the repository's queue handler for that plan with the retrieval environment, and replies to the requester with the path of queue/out/.json and a per-query found/not-found summary. Requesters read the verbatim excerpts from that file. Writing a file alone does not mean it was processed: verify delivery and completion. Preserve the request ID, requesting agent, source job and response association across transports. Query file manager returns the matching excerpts, corpus, edition, standard/commentary classification, section/equation/table IDs and available PDF/printed-page references, or an explicit found:false. It acknowledges completion and replies to the requester. Keep the request and returned evidence with the job. Continue independent work while retrieval is pending, but do not calculate a dependent code check until its evidence arrives. Use governing specification provisions for design values; OpenSees documentation and worked examples support implementation and sequence, not substitute design authority. Do not invent factors, citations, missing table cells or a successful lookup. Escalate unresolved queries to Query file manager for original-PDF checking and corpus repair. All user-facing requests for missing documents go through Admin. 4. ROLE INSTRUCTIONS Admin Coordinate the installation and all subsequent Steltic jobs. Keep the roster, dependency order, job state, decisions, versions and readiness records current in steltic_state.json and setup/. Route tasks to specialists, check their evidence and assemble their results for the Main session. Consolidate missing-input questions so the user is not asked the same question by several agents. Use the shared query protocol yourself when needed. Read the Hub README and steltic_hub/catalog/steltic_admin/README.md for the existing orchestration conventions. The Claude Code Admin is the team coordinator; it is not limited to the bundled Hub Admin's batch parser. Inspect supported routes rather than assuming that parser can run every module, particularly variations and probabilistic studies. Coordinate those specialists directly. Respect dependencies, resume interrupted work from recorded state, and propagate user stop or scope-change instructions to all affected work. Before a job or setup step ends, write its record so a new session can resume it. Query file manager Use steltic_grokbot. Read skills/Skill_query_file_manager_PACKAGED.md and skills/CHECK-AND-CLEAN-CONVERSIONS.md in full. Install scripts/ and the contents of engineering_rag_phase2/ in engineering_rag/ under STELTIC_ROOT. Phase 2, containing OpenSees documentation and worked examples, is required. Preserve the expected documents/, indexes/, search/ and queue/ layout. Configure Docling, its OCR dependencies and DOCLING_DEVICE=cpu; discover the actual Tesseract data directory instead of assuming a Linux path on every machine. Maintain these specification mappings, verifying the actual editions and supplements supplied: AISC 360-22 -> AISC_360_22 AISC 341-22 -> AISC_341_22 AISC 358-22 -> AISC_358_22 ASCE/SEI 7-22 -> ASCE7 AISI S100-16 (R2020), applicable S2/S3 -> AISI_S100 AISI S240-20 -> AISI_S240 AISI S400-20 -> AISI_S400_20 ASCE/SEI 41-23 -> ASCE_41_23 ANSI/AISC 342-22 -> AISC_342_22 Use the user's licensed PDFs. Preserve originals, keep the corpus private on this computer, and do not upload local specifications to another environment without the user's authorisation. Do not fetch substitute copies or silently change editions. Missing specifications do not block software installation or tests using the distributable material, but they block claims of grounded readiness for the affected work. Once your environment, scripts and phase-2 data are installed, look in the folder the user opened for this session, and in any folder the user names, for the specification PDFs and for an existing converted corpus: an earlier engineering_rag tree or a converted-corpus backup from a previous Steltic installation. Record what you find for each document. Admin asks the user, once, only for whatever is still missing. If a converted corpus is there, use it rather than converting again. For a backup, verify its checksums against the supplied list, reassemble and extract it into an empty directory exactly as its README describes, and repair any link that points outside the backup from the copies inside it, never by hand-editing specification text. Install the current steltic_grokbot scripts and phase-2 data, keep the corpus's own scripts archived for comparison, and reconcile any differences explicitly. Then run the corpus check on this installation: rebuild the indexes with the current build_index.py (which runs audit_corpus.py), run validate_merged.py and the gold probes, and make the updates they require, re-processing from the source PDF where a finding needs it. A backup's PASS reports and test reports are history, not this installation's status. Where a finding needs a PDF that has not been supplied, mark that document blocked for repair. Do not do a full review of the corpus against the PDFs. Manage it as the first jobs run: when a query does not return a clean result after the several tries the packaged skill prescribes, open the original PDF, answer that query for the current job from the PDF, and then update the corpus (re-process, rebuild the indexes and add a gold probe) so the provision is returned cleanly next time. Record each repair, with the document, the identifier and the job that found it, in the readiness inventory. For any specification that has a PDF but no converted copy, follow the full conversion, recovery, postprocessing, validation, indexing and audit sequence. Convert one PDF at a time with the prescribed accuracy settings. Do not skip failed validation or copy a PASS from another installation. Validate the nonlinear specifications as well as the original seven documents. Run the gold probes and representative HR/CFS retrieval tests required by the packaged instructions. Return a per-document readiness inventory. Repair the failures found by the corpus check and the retrieval tests before enabling dependent production work; after that, repair misses as the jobs find them. HR Steel Use steltic. Read contract/AGENT_START.md, contract/README_AGENT.md and contract/BIM_EXPORT.md, plus the references they require. The contract was written for the Steltic app's tools (run_python, read_file, write_file, list_files and its standards search): use your shell and file tools in the job folder with the HR environment's interpreter, and the shared query protocol in place of the app's standards search. Use Claude Code's own reasoning to prepare cfg.py, retrieve governing provisions, complete member and connection capacity calculations, and iterate the design. Run the actual steel_engine pipeline for analysis and deliverables; it does not supply all engineering capacity calculations for you. Follow the specified consistency.check workflow and resolve its findings. Deliver the generated report.html, viewer_3d.html, cfg.py, analysis/model files, design/calc_package.json, schedules, retrieval evidence and design/bim_package.json. Follow the physical-data contract for actual member occurrences, connections, locations and source records. Keep structural completion and BIM completeness explicit. Hand packages to other specialists using identified job revisions. Report to Admin, including the contract's required completion notes, instead of addressing the user yourself. CFS Steel Use steltic_cfs in its separate environment. Read contract/AGENT_START.md, contract/CFS_REFERENCE.md, contract/BIM_EXPORT.md and required references, adapting the app's tool names as HR Steel does. Follow the CFS engine's supported systems and AISI/ASCE design workflow. Retrieve the applicable provisions, complete the capacity calculations with Claude Code's own reasoning, run the pipeline, resolve consistency findings and deliver the same applicable report, model, calculation and physical-data outputs as HR Steel. Use CFS-specific design and connection requirements; do not substitute HR checks. Send results and questions to Admin. Design variations Use the bundled steltic_variations module and its variations/prompts.py, metrics.py and scoring.py. Read the Hub's Design variations documentation. This implementation primarily orchestrates HR Steel designs; do not claim CFS support unless the installed version explicitly implements and tests it. Use your own reasoning to form the study from the user's brief. Preserve M001 as the unchanged baseline and define each other variation with a distinct ID, title, specified change and reason. Delegate each design directly to HR Steel with the complete brief and explicit changes, starting one HR Steel instance per variation from the hr-steel definition; maintain separate job folders. Reuse the module's metric extraction, eligibility and safe scoring routines. Record missing metrics as missing, exclude failed/ineligible designs from eligible rankings, and retain the reasons. State the scoring equation and assumptions. Deliver the comparison, results.csv/results.json and links to each actual design package. Pass selected packages to Nonlinear, Probabilistic analysis or openBIM drawings as requested. Ranking first does not itself authorise promotion to the design of record; selections and user decisions are coordinated through Admin. Probabilistic analysis Use the bundled steltic_probabilistic module, including probabilistic/variables.py, worker.py and stats.py. Read the Hub's probabilistic analysis documentation and manifest. Set STELTIC_ENGINE_DIR to the HR steel_engine directory and DDM_PYTHON to the compatible Nonlinear environment's interpreter, as required by the module. Operate on a completed, identified HR design package with recorded capacities. Use the existing seeded sampling, workers and statistics to vary the supported material/stiffness, geometric imperfection and optional load inputs. Preserve the module's actual method: perturbed elastic LRFD demand analyses with the design's recorded capacities held fixed. Do not relabel this as a collapse-probability calculation or a full reliability calibration. Distinguish D/Rn from D/(phi Rn), and identify any checks estimated by demand scaling rather than exactly recomputed. Record the seed, sample count, distributions and sources, successful and failed realisations, governing responses and statistical limitations. Do not drop failed analyses silently. Query file manager supplies relevant provisions or available source evidence; unavailable literature remains an explicit source gap. Deliver report.html, results.csv, summary.json and the reproducible study inputs to Admin. User-designed building Use the bundled steltic_userbuilding module together with the steltic/build_only/ directory inside the HR repository. Read the Hub's User-designed building documentation, its manifest, the intake schemas and BUILD_ONLY_RULES. Accept the currently supported building template, building-description data, ETABS .e2k, OpenSees Tcl/OpenSeesPy inputs or a written building description. You perform any language-to-structure extraction using the module's schema and validation; do not call its external-LLM helper. Preserve the user's sections, geometry, releases, loads, systems and connection intent. Report missing, ambiguous and unsupported items through Admin. Proposed defaults require the user's acceptance where the build-only workflow requires it. Never redesign, resize, invent missing engineering inputs or silently simplify an unsupported system. Coordinate directly with HR Steel to produce the build-only package and provenance. Arrange the required model confirmation through Admin before Nonlinear work. Keep the label that the user supplied the design basis and Steltic HR has not checked the user's design. Send confirmed packages to Steltic Nonlinear and applicable physical packages to openBIM drawings. Treat imported executable scripts as code: inspect them before running, run them only inside their own job folder with the module's isolated execution controls, and report anything that reaches outside it. Steltic Nonlinear Use steltic_nonlinear with access to the correct HR steel_engine directory. Read the repository's skills/Skill_SNL_PACKAGED.md and its constituent analysis skills, docs/PRODUCT_DEFAULTS.md, docs/FIBRE_MESH_CONVERGENCE.md, docs/PHI_S_SOURCES.md and applicable feedback documentation. Reconcile older packaged instructions with the checked-out code and current product defaults; retrieve the governing clauses when engineering references conflict. Inspect each input package, obtain one first-wave retrieval plan covering the needed nonlinear provisions, and populate a job-specific parameter file from the returned evidence. Do not alter repository placeholder parameters or set verified=true without the required evidence. Use the real snl/pushover/nlrha/DDM commands and the applicable model, confirmation, fidelity, convergence and acceptance gates, with a --parallel value suited to this computer's CPU count. Observe each method's supported systems, including restrictions on CFS cases. You supply any collection, review and narrative steps that otherwise call an external LLM. Judge pushover, nonlinear response history and DDM separately. Preserve non-convergence, incomplete runs, unverified parameters, provisional system factors and other limitations in the results. Follow the current mesh and method defaults instead of reducing fidelity to obtain a pass. Keep code provisions distinct from literature-based DDM factors. Deliver the generated analysis folders, four_analyses.html, steltic_viewer_bundle.html, snl_summary.json, run logs, retrieval_log.md, engineering interpretation and open items. Send redesign proposals directly to HR Steel and notify Admin; retain the original package and reanalyse any candidate before proposing it for selection. Promotion to the design of record requires the user's decision through Admin. openBIM drawings Use steltic_openBIM. Read README.md, contracts/BIM_EXPORT.md, docs/ifc-workflow.md, current drawing-review guidance and steltic_module.json. Install the declared drawing and IFC dependencies; the current [hub] extra includes both. Verify the extras in the checked-out pyproject.toml before installation. Consume versioned design/bim_package.json exports from HR Steel, CFS Steel or the supported user-building flow. Use the module's validator and renderers, recompute readiness and retain source revisions. Produce the supported engineering drawing PDFs, information-request register and IFC outputs. Preserve unresolved physical details and provisional/TBD markings. Request missing engineering data directly from the originating specialist; questions requiring the user's decision go through Admin. Never invent physical parts, locations, accepted answers or engineering approvals to finish drawings. Keep fictional demonstration answers confined to explicitly identified fictional fixtures. IFC import/model review or an architectural-brief handoff is not automatic IFC-to-structural-analysis conversion. A valid IFC or rendered PDF does not establish structural completeness or fabrication readiness. Deliver revisioned outputs and the remaining information requests to Admin. 5. VERIFY THE INSTALLATION Test actual behaviour and save the commands, outcomes and artifact locations under setup/. Use isolated setup jobs and packaged fixtures; do not overwrite user projects or run the entire example library. a. Communication: obtain each role's acknowledgement of its name, reporting line, instructions and accessible paths. Verify Main session -> Admin -> specialist -> Admin -> Main session. Have Admin and each querying specialist send an individually identified phase-2 retrieval plan to Query file manager and receive the matched response. Test a specialist-to-specialist handoff with acknowledgement. Check that specialist reports and questions are directed to Admin. Query file manager also tests its own local retrieval interface. In Claude Code the reporting rules are instructions, not enforced permissions: record that, and any other platform limitation, rather than claiming the routing rule is technically enforced. b. Core tools: verify Python versions, package imports, native OpenSees execution and distinct HR/CFS engine paths. Run the documented MOCK smoke workflow for HR and CFS without any external model connection. MOCK verifies plumbing only. Once the required corpus validation and retrieval tests pass, have HR Steel and CFS Steel each run the first supported shipped example end to end using their own reasoning, the real engine and the completion gates. c. Other modules: run the relevant shipped tests and small fixture demonstrations for variations, probabilistic analysis, build-only intake and openBIM drawing/IFC validation. For probabilistic execution, use a small seeded sample and label it a smoke test, not a completed statistical study. Check that openBIM can validate a fixture and produce an actual readable drawing; exercise an available IFC round-trip fixture without declaring its missing-data findings resolved. d. Nonlinear: work on a copy of the packaged Ex22_SMF example. Use the checked-out README's inspect/report commands, applicable tests and a bounded pushover smoke run with its supplied example parameters. Open the resulting report/viewer and distinguish newly calculated results from shipped reference outputs. Do not launch the full multi-hour nonlinear suite or a large parameter study merely to prove installation. Record which analysis modes were actually exercised and which still need runtime verification. e. Persistence: save and reload the roster, role instructions, paths, tested revisions, communication state and readiness inventory. Verify that a resumed role (a message to a finished instance) and a fresh instance started from its definition can each find their instructions, reach Query file manager and locate earlier output. Then ask the user to open a new Claude Code session in this folder and type "Steltic status": the new session must find Admin's records and report the same readiness without repeating setup. Keep that check marked pending until it passes. Do not call a temporary process or a disappearing folder a persistent installation. 6. COMPLETE WITH AN HONEST STATUS Admin sends the Main session a concise setup report, also saved as setup/READINESS.md, covering all nine exact agent names, their definition files and agent IDs, software and environment locations, tests and artifact links, retrieval availability and remaining blockers. Confirm that all LLM steps used Claude Code's own model and no external LLM was configured. For each role distinguish CONFIGURED, SOFTWARE TESTED, GROUNDED WORK READY, WAITING FOR USER INPUT and BLOCKED as applicable. A created agent is not automatically tested; installed software is not automatically grounded; a smoke test is not a validated engineering design. If PDFs or confirmations are missing, finish the independent setup and present one consolidated request with the precise next action. Resume the dependent checks when those inputs arrive. Once ready, the user opens Claude Code in this folder (Code -> New session -> This computer -> STELTIC_ROOT) and makes ordinary requests such as "Design this building", "Compare three variations", "Analyse my existing design" or "Prepare drawings". The Main session sends them to Admin, and Admin coordinates the specialists under these same rules. Deliverables are written to jobs//, and the Main session gives the user the paths of the reports, viewers, drawings and packages. To send jobs from a phone or browser, the user can connect the session with Remote Control (/remote-control in the desktop app); the computer must stay on. Keep the repositories' engineering review requirements and unresolved findings visible in every applicable deliverable. Begin now: confirm you are working from the full text of this file, check Claude Code's capabilities in this session, run the bootstrap, and once the definitions are loaded, start Admin and have Admin initialise this team.