Agentlas Core Engine Meta-Agent Team

September 12, 2026 · View on GitHub

Asked to install this repo, not to work on it? This file is the contributor constitution and will not help you. Read the install block at the top of README.md: it points at scripts/install-all-runtimes.sh, which writes only under ~/.agentlas, ~/.local/bin, and this host's own plugin/command-adapter directories.

Repository Constitution: Local Main Only

This repository uses one canonical development line: the local main branch in the canonical Agentlas OS checkout.

Make every source change and commit directly on local main.
Do not create feature, release, backup, agent-named, or temporary branches.
Do not create Git worktrees. Use an external recovery directory or a verified Git bundle when a safety snapshot is required.
Before editing, confirm the canonical checkout is on main, inspect status, fetch remote refs, and inspect main...origin/main. A GUI "Pull origin" button is not permission to pull or merge blindly.
If local main is dirty, preserve it with a reviewed checkpoint commit before reconciling remote changes on that same branch.
Push only main and intentional release tags. Do not publish side branches.

Public Release Allowlist (Hard Rule)

Before any commit, push, tag, release, or upload to a public GitHub repository, construct and review an explicit allowlist. Public source and release artifacts may contain only files required for an end user to install and run the product, plus public-facing README, LICENSE, and CHANGELOG material.

Never publish internal design or research documents, plans, benchmarks, benchmark prompts or results, tests, fixtures, test data, scores, logs, screenshots, signing material, certificates, credentials, environment files, operator notes, private paths, local memory, or recovery artifacts. Run tests and benchmarks only in local/private temporary storage or private CI; publish neither their inputs nor their artifacts. Existing unrelated local work must remain unstaged.

For a public release, do not use a blanket git add -A. Stage the allowlist explicitly, inspect git diff --cached --name-only and the staged archive manifest, scan the staged content for secrets/private paths, and stop the release if any excluded class appears. A passing test does not authorize publishing the test.

This repository is a portable four-agent meta-agent team. Use it to create or package Agentlas-compatible single agents and multi-agent teams for Codex, Claude Code, Gemini CLI, Antigravity, Cursor, OpenCode, OpenClaw, Hermes Agent, Ollama-served local models (Gemma, DeepSeek — see docs/local-models.md), and AGENTS.md-compatible runtimes.

Source Of Truth

Canonical entry point: AGENTS.md.
Architecture ownership rule: docs/source-of-truth.md.
Runtime split and sync boundary: docs/runtime-sync-boundaries.md.
Third-party plugin contribution boundary: CONTRIBUTING.md and PLUGIN_CONTRIBUTIONS.md.
Global command contract: docs/global-command-contract.md.
Production Ontology Runtime: docs/ontology-runtime.md, ontology/, bin/ontology, and scripts/verify-ontology-runtime.sh.
Agentlas Cloud runtime contract: docs/agentlas-cloud-runtime.md, agentlas_cloud/, schemas/agentlas-manifest.schema.json, and templates/agentlas.json.tpl.
Hephaestus Network 2.0 contract: docs/hephaestus-network-2.0.md, docs/runtime-fallback-adapters.md, agentlas_cloud/networking/, schemas/routing-card.schema.json, .agentlas/routing-card.json, and scripts/verify-routing-cards.sh.
Stormbreaker robust execution contract: docs/robustness-protocol.md, docs/robustness-eval.md, and schemas/robustness-eval-result.schema.json.
Canonical Stormbreaker Goal + UltraCode harness: docs/stormbreaker-goal-ultracode-harness.md, agentlas_cloud/networking/stormbreaker_harness.py, and schemas/stormbreaker-goal-ultracode-harness.schema.json.
Builder quality gate: contracts/builder-interview-research-gate.md, docs/builder-quality-research-basis.md, templates/builder-interview.md.tpl, templates/research-sources.md.tpl, templates/tool-selection.md.tpl, templates/domain-expert-synthesis.md.tpl, templates/prompt-performance-contract.md.tpl, and templates/capability-eval-plan.json.tpl.
Portable support contracts: docs/mode-classifier.md, docs/clarify-question-loop.md, docs/agentlas-auto-activation.md, docs/local-credential-store.md, and docs/skill-lifecycle-promotion.md.
Agent experience and MCP contracts: docs/agent-experience-assets.md, docs/mcp-build-resolution.md, the matching schemas/*.schema.json files, .agentlas/mcp-policy.json, and scripts/verify-experience-assets-contract.sh.
Team members: agents/10-single-agent-builder/agent.md, agents/20-multi-agent-team-builder/agent.md, and agents/30-agentlas-packager/agent.md, plus agents/40-session-agent-builder/agent.md.
Mode contracts: modes/single-agent-creator.md, modes/team-builder.md, modes/agentlas-packager.md, and modes/session-agent-builder.md.
Portable runtime core: .agents/agentlas-core-engine-meta-agent/agent.md.
Reusable skills: .agents/skills/*/SKILL.md and root skills/*/SKILL.md.
Agentlas contracts: .agentlas/mode-map.json, .agentlas/agent-card.json, .agentlas/company-blueprint.json, .agentlas/sitemap.json, .agentlas/global-commands.json, .agentlas/memory-map.json, .agentlas/memory-tickets.jsonl, .agentlas/vault-references.json, .agentlas/local-credentials.map.json, .agentlas/mcp-policy.json, and skill lifecycle files emitted in generated packages.
Public install surfaces: codex/, .claude/, .gemini/, antigravity/, and scripts/.

Runtime-specific folders are adapters. They must mirror the canonical core, not become separate sources of truth.

Agent Ontology (AO), Agent Workforce Ontology, the local semantic ontology, Context Map, and Career Graph are independent active systems. Super Ontology is retired and must not be generated, loaded, validated, packaged, or presented as a source of truth. Its retained documentation is an archived design record only.

Third-Party Plugin Boundary

Third-party services and optional provider integrations belong in independently installable Agentlas plugins, not in Agentlas Core. This includes search APIs, model providers, hosted browsers, data sources, SaaS connectors, payment providers, and vendor-specific MCP servers.

Reject a provider contribution when it requires provider-specific changes to Core registries, CLI provider maps, default or full loadouts, credential maps, doctor or profile logic, ranking policy, network allowlists, or generated runtime mirrors. An existing built-in integration is legacy precedent, not permission to duplicate that wiring for another provider.

An acceptable third-party plugin must be independently installable, upgradeable, disableable, and removable; activate only by explicit user choice; declare capabilities, authentication, network hosts, operations, and privacy effects without embedding credential values; and fail safely when unavailable. Installing or removing it must not require a Core patch.

Core changes are considered only for a provider-neutral, versioned extension contract that serves multiple independent plugins or enforces a Core invariant. Propose that contract separately from any vendor implementation. Do not use a vendor PR to introduce both the extension point and its first provider. Follow PLUGIN_CONTRIBUTIONS.md for the submission and review checklist.

Generated Instruction Language

All generated or repaired agent instruction files must be written in English. This includes AGENTS.md, CLAUDE.md, GEMINI.md, agent.md, role cards, skills, workflow/command adapters, runtime prompts, handoff contracts, return contracts, and package docs that are read as operating instructions.

Korean and other languages are allowed only for user-facing public copy, localized marketplace fields, trigger examples, and sample user inputs. If the source material or the user's request is written in Korean, translate the agent behavior into English before writing runtime instructions.

Source Code & Commit Language

This is a public, open-source repository. Code comments and commit messages must be written in English going forward, regardless of the language the requesting session or the user's instructions were written in. This applies to new commits only — do not rewrite existing git history to relabel past commit messages.

Exempt from this rule (leave as-is, do not force-translate): localized marketplace/product copy, Korean sample inputs used as routing or i18n test fixtures (e.g. "사업계획서 만들어줘" as a query example), and any string that is itself the user-facing product output rather than an explanation of the code. When translating an explanatory comment, preserve its meaning and any file:line/command/evidence citations exactly — do not summarize away the specifics that made the comment worth writing.

Operating Loop

Before applying the package-building loop below, preserve the product runtime authority model:

Agentlas One is the sole controller of every One session. Explicit @agent targets are turn-scoped sub-agents and never change session ownership.
Desktop and Mobile Work create only project-bound tasks. The first agent in the project's ordered agent pool is the controller; remaining members are eligible task-scoped workers selected through validated WorkOrders.
Plan, Goal, Network, Live, permission, and explicit agent targets are optional overrides. Omission means the controller decides, not that the capability is disabled.
Do not recreate global chat, Agent Group, generated group orchestrator, persistent hired-agent, or session agent-switch contracts in adapters, templates, examples, or generated packages.
Do not make semantic decisions with regex gates, keyword lists, glossaries, deterministic lexical scores, or default-agent substitution. Leave a decision unavailable when the connected model cannot make a valid judgment.
Protocols may define state, evidence, capability, CTA slots, and question layouts. One or the project controller supplies user-facing recovery content from current context; do not expose internal codes or operator vocabulary.

Run the public mode classifier (skills/mode-classification/SKILL.md) and classify the request by independent ownership boundaries, not keywords. If existing material is being converted, repaired, cleaned, imported, or released, route to packager first. Otherwise count roles that independently own memory/context, tools/permissions, and success criteria; one boundary is single-agent, two or more plus routing/synthesis/produces-consumes handoff is team-builder. If the boundary is unclear, clarify before generation instead of guessing. Classify the request as one of:
- create one single agent;
- create a multi-agent team;
- package or repair an existing local/external agent or team;
- compile an explicitly exported session into a reviewed agent or team.
If required details would change the package, run the clarify loop (skills/clarify-question-loop/SKILL.md) before generating files.
Route to exactly one core team member:
- 10-single-agent-builder;
- 20-multi-agent-team-builder;
- 30-agentlas-packager;
- 40-session-agent-builder for an explicitly exported session source.
Run the Builder Interview and Research Gate (contracts/builder-interview-research-gate.md) before writing substantial generated package files. Ask an 8-12 question first batch when the request is vague, continue follow-ups until the functional brief is clear, research official sources, similar agent repositories or comparables, academic/professional theory, and plugin documentation, then select tools/plugins only after checking permissions, secrets, alternatives, and smoke-test paths. If single-agent vs team shape is unclear, the first batch must ask in plain language: "이 일을 한 명의 전문가가 처음부터 끝까지 맡으면 되나요, 아니면 조사/분석/검토처럼 여러 전문가가 나눠 맡고 마지막에 합쳐야 하나요?" Do not expose internal labels like ownership boundary, memory/context, synthesis, or produces/consumes to non-technical users.
Inspect current files before making claims. Prefer real files over remembered assumptions.
Emit or repair the smallest useful Agentlas package.
Add the required architecture contracts for the selected mode.
Add docs/builder-interview.md, docs/research-sources.md, docs/tool-selection.md, docs/domain-expert-synthesis.md, docs/prompt-performance-contract.md, and .agentlas/capability-eval-plan.json unless the task is explicitly a minimal private scaffold or trivial adapter repair.
Before writing generated agent instructions, enforce the Generated Instruction Language policy: write runtime instructions in English, even when the source brief is Korean.
Add thin adapters for Codex, Claude Code, Gemini CLI, Antigravity, and optional Cursor.
Assign one canonical global command during creation, write .agentlas/global-commands.json, add runtime command files or aliases, and keep team worker roles routed through the orchestrator/HQ command unless the user explicitly asks for direct worker commands.
Add .agentlas seed files when the generated or packaged output needs local continuity; local runtimes may auto-activate them using skills/agentlas-auto-activation/SKILL.md.
Add or repair agentlas.json so Agentlas Cloud can compile a runtime bundle, gate lazy file reads, and separate private sync from public clean copies. Seed .agentlas/mcp-policy.json only when missing: resolve the system-global registry first, ask once for the selected set, load selected tool schemas and triggered skills only, and isolate each MCP failure.
Add skill lifecycle metadata using skills/skill-lifecycle-promotion/SKILL.md when the package contains reusable skills.
For generated or repaired packages, run scripts/verify-team-package.sh <generated-package-root> before the final report. If it fails, do not report completed; fix the shape by collapsing to a valid single-agent package or adding orchestrator/HQ plus company topology.
Also run scripts/verify-generated-package.sh <generated-package-root>. This is the gate that checks the package against package-contract.json — every required artifact present, parseable, schema-valid, and with no {{PLACEHOLDER}} left unfilled. It is mandatory for the same reason the contract exists at all: while nothing compared a package to that list, the live catalogue drifted to capability-eval-plan.json on 4% of published releases and mcp-policy.json on 50%, both marked required since the day the contract was written, and 142 of 144 forge packages failed the benchmark artifact on its PATH alone while the file sat on disk. Do not report completed on a FAIL; the named artifact is missing or malformed, and a package that ships without it cannot be routed to.
Verify with scripts/verify-package.sh.
For ontology runtime changes, also verify with scripts/verify-ontology-runtime.sh.
For long-running or multi-file execution work, apply docs/robustness-protocol.md: scope lock, plan lock, evidence loop, review gate, and final gate before claiming completion.

Hephaestus Network Commands

Agentlas-native surfaces are commandless: Agentlas Terminal and the Agentlas app should accept plain language and dispatch through native Agentlas/ Hephaestus tools without requiring a /hep-* command. External LLM hosts registered through MCP, plugins, prompts, or command files expose the core commands /hep-build, /hep-network, /hep-local, /hep-cloud, /hep-hub, /hep-search, /hep-browser, /hep-call, /hep-upload, /hep-storm, and /hep-graph. Current Codex exposes build, Network, Cloud, upload, Stormbreaker, and Graph as $hephaestus-* plugin skills; the remaining typed MCP surfaces are requested in plain language. Custom /prompts:hep-* commands are legacy only. Antigravity is an independent runtime target, not a Gemini CLI mode, and is the default Gemini-family target where the operator did not explicitly request Gemini CLI.

/hep-build <request> is the creation, repair, memory, playbook, and diagnostics surface. /hep-network <request> (alias @Hephaestus <request>, terminal hep-network "<request>") is the Agent Workforce Ontology staffing surface across Local, Cloud, and Hub. /hep-local, /hep-cloud, and /hep-hub use exactly one source each and never widen scope. The active host LLM creates a redacted work order, calls the local hephaestus-network MCP tool workforce.search_candidates with sourceScope, preserves its source receipts and selection session, chooses the exact roster, then calls workforce.validate_selection with only the exact WorkOrder and host-authored Selection. Core resolves the full pinned federation result by selectionSessionId; the default projected menu must not be echoed as federationResult. It passes the accepted Selection and federatedSelection plus mandatory projectDir to workforce.prepare_execution. It then runs distinct worker/synthesis/verifier invocations. Do not expose a second remote agentlas MCP on hosts that can run local Core: Core reaches Cloud and Hub through its internal upstream client, and duplicate workforce.* tools would bypass the federation and privacy boundary. Deterministic code enforces hard constraints but never picks the final team. The standalone shell aliases can only report host_llm_required; the old card router is an explicit HEPHAESTUS_LEGACY_ROUTER=1 compatibility/debug surface. The legacy hephaestus_route and hephaestus_cloud_search MCP tools have the same explicit opt-in gate and are never substitutes for typed staffing. /hep-search <request> compares Cloud and Hub candidates without invoking. /hep-call <slugs> <context> prepares explicitly named agents. /hep-upload <agent-folder> is the upload gate: before any package, publish, register, add-source, reindex, or upload API call, ask whether the destination is private Agentlas Cloud or public Agentlas Hub.

Workforce candidate retrieval uses roles, skills, knowledge, MCP/tool capabilities, artifact contracts, runtimes, languages, authority, entity kind, and evidence levels. It never uses installs, ratings, revenue, invocation history, or local callability as semantic fit. Selection and preparation pin immutable release, package hash, and content digest; unavailable posts remain explicit and substitution requires a new host-LLM decision. Never send raw prompts or local memory to Hub. A selection or prepared BYOM bundle is not an execution receipt: completion requires planner parse success without fallback, distinct child invocations, artifact handoffs, synthesis, and an independent passing verifier. Validate the host-produced receipt locally with workforce.validate_execution_receipt; that read-only validator neither runs workers nor creates execution evidence. Generated and packaged repos must include .agentlas/routing-card.json (see schemas/routing-card.schema.json); cards compile into immutable workforce profiles and lifecycle events.

hep-global install is a shell utility, not a task command: it installs a managed Hephaestus Global Router marker block into ~/.codex/AGENTS.md and ~/.claude/CLAUDE.md, and ~/.gemini/GEMINI.md for Antigravity/Gemini. That block makes ordinary host prompts use Network federation unless an exact Local, Cloud, or Hub scope was requested. Exact-scope commands never silently widen when a source is unavailable, a workforce MCP call is rejected, or no qualified slot coverage exists. If a source is blocked by credits, entitlement, availability, or fit, the host reports that exact boundary. Status lines must name final workers, not router commands: 사용 에이전트:/Agents used: for agents and 사용 스킬:/Skills used: only for final skill fallbacks. hep-global status reports the managed block state; hep-global remove removes only the marker block and leaves the user's surrounding prompt file intact.

Hephaestus Network chooses the agent, team, plugin, or Hub bundle. Hephaestus Stormbreaker governs execution after that route is selected: it requires scope locking, issue-contract extraction, failure-memory checks, verifier-first planning, a parallel session fabric for pipeline work, bounded evidence loops, review gates, outcome ledgers, and a final completion gate for substantial work. Pipeline success is blocked until every required execution packet is passing or the run is explicitly reported as blocked.

Team Roles

10-single-agent-builder: one installable self-evolving worker package.
20-multi-agent-team-builder: multi-role team package with orchestrator/HQ, PM Soul, Memory Curator, Policy Gate, eval, QA, handoffs, memory, and runtime adapters.
30-agentlas-packager: package or repair existing local/external agents and teams into Agentlas architecture for local use, Agentlas import, Codex plugin use, Claude adapter use, or open-source release.
40-session-agent-builder: convert explicitly exported host sessions into a reviewed reusable agent or team candidate, with no raw transcript carryover.

PM Soul, Memory Curator, runtime adapters, sitemap/task-bias, policy, eval, and verification are generated architecture components. They are not extra members of this meta-agent team.

Output Contract

When asked to create, repair, or package an agent repo, return:

status: completed, blocked, or needs_clarification.
evidence: files inspected, commands run, and verification result.
output: generated path, changed files, or exact design.
global_commands: Claude Code, Codex, Gemini CLI, Antigravity, generic AGENTS.md, and terminal commands from .agentlas/global-commands.json.
interview_research: interview status, unresolved assumptions, similar agent/repository research, academic or professional theory basis, selected/rejected tools or plugins, domain-expert synthesis, prompt-performance contract, and capability-eval plan.
blockers: only blockers that require the user or external state.

Generated or packaged repos must include the relevant subset of:

visible agents/ and skills/ folders;
AGENTS.md, CLAUDE.md, GEMINI.md;
.agentlas/mode-map.json;
.agentlas/agent-card.json;
.agentlas/company-blueprint.json;
.agentlas/global-commands.json;
agentlas.json;
docs/builder-interview.md;
docs/research-sources.md;
docs/tool-selection.md;
docs/domain-expert-synthesis.md;
docs/prompt-performance-contract.md;
.agentlas/capability-eval-plan.json;
.agentlas/memory-map.json;
.agentlas/memory-tickets.jsonl;
.agentlas/vault-references.json;
.agentlas/local-credentials.map.json plus .env.example, signing/README.md, and credentials/README.md when local credentials are required;
.agentlas/mcp-policy.json with catalog-id-only requirements, value-free credential metadata, and no server command, args, endpoint, or key value;
.agentlas/skill-registry.json;
.agentlas/skill-trials.jsonl;
.agentlas/curator-decisions.jsonl;
runtime adapters and smoke-test docs;
global command adapter files such as .claude/commands/<slug>.md, codex/plugins/<package-id>/commands/<slug>.md, and gemini/extension/commands/<slug>.toml with optional .gemini/commands/<slug>.toml fallback;
production ontology runtime code, CLI, tests, and sample corpus when the package needs local-first knowledge storage, search, GraphRAG, Memory Curator bridge tickets, or Agent Working Memory;
install or verification scripts;
public-safety notes if the output is meant for GitHub.

Cloud-ready packages must pass the local setup wizard contract: agentlas.json exists, the Hephaestus security scan is PASS or reviewable WARN, the runtime bundle compiles from manifest allowlists, and denied paths never expose secret-like values.

Mode Rules

single-agent-creator produces one installable worker package. It can include multiple skills, setup guides, memory architecture, research refresh, and self-evolution proposals, but it must not inflate the request into a company, HQ roster, or swarm. It must expose one global command for the worker and report that command to the user after creation. It must not emit multiple loose worker agent.md files; that is a team shape and requires HQ/topology.

team-builder produces a multi-role team package. It must include a root orchestrator/HQ, PM Soul or project owner, Memory Curator, Policy Gate, worker roles, eval judge, QA/evidence gate, handoff rules, runtime adapters, memory architecture, and release checks. It must expose the orchestrator/HQ global command as the public entry point; worker roles stay behind HQ routing unless direct worker commands were explicitly requested. Team packages must pass scripts/verify-team-package.sh <package-root> before completion.

agentlas-packager repairs or converts an existing agent/team into the Agentlas shape. It adds canonical core files, .agentlas contracts, runtime adapters, schemas, manifest, install scripts, global command registry, verification, and public/private cleanup.

session-agent-builder accepts only explicitly named JSON/JSONL session exports. It normalizes and redacts them into a reviewed Work Brief and declarative session IR, then requires an explicit single or team choice and owner approval before joining the normal package flow. It never captures a recent session implicitly, carries raw transcript or hidden prompt content, widens permissions, activates MCP, promotes a Skill, uploads, or publishes.

Memory Preflight

Before substantial work:

Read .agentlas/memory-map.json if present.
Load only task-relevant memory.
Label remembered facts as verified, memory_derived, inferred, or stale_check_needed.
Re-check drift-prone facts against current files or sources.

Workers may emit ## Memory Events. The runtime or orchestrator wraps those events into Memory Tickets before durable curation.

Skills may also emit ## Skill Trial Events. Those events are lifecycle evidence for candidate/trial/first-class review and must not be mixed with Memory Events. Generated packages keep all skills at candidate tier until a local Curator approves promotion.

Safety Rules

Never store secrets, API keys, tokens, key material, credential file contents, raw logs, or full transcripts in generated repos or memory.
Never copy a base agent prompt, skill, package file, raw prompt, customer record, or credential into an Experience Pack. Variant bindings use exact release references only, and official success accepts verified replay-safe RunReceipts only.
Real credential values may be saved only in local gitignored project stores (.env, .env.local, signing/, credentials/) or a local keychain/vault. Generated public packages may include only placeholders, guide files, and a value-free .agentlas/local-credentials.map.json.
Generated project memory must keep a Local Credential Index (read first) section near the top. For deploy, release, store, billing, auth, API, or cloud work, check that section and .agentlas/local-credentials.map.json before reporting missing credentials.
Do not copy private local research folders into public output.
Do not call runtime adapters canonical.
Ask for explicit approval before destructive file actions, production deploys, paid API spend, permission widening, or public publishing.
Public packages must pass scripts/verify-package.sh and scripts/public_safety_check.sh before release.