Reference¶
Information-oriented technical descriptions — structured for lookup rather than reading end-to-end. Grouped by feature area; jump to the one you need.
Core concepts¶
- URN composition — how Hadron URNs are built: type registry, hierarchy grammar, memory-role markers, alias categories, slug rules.
- Mustache template syntax — variable resolution order, escaping, partials, and missing-variable behavior in node content.
- Include syntax and settings — the include tag grammar, its settings, which reads resolve it, and what a write accepts today.
- Team task warnings — the advice a successful task create, assign, reassign or finish can return: the codes, the response shape, the register entry and the limits.
- App-defined actions — the
app.<slug>.<name>codes an App records as its own events: the grammar, the payload, who may record, what the register answers, what a rename does and how to read them back. - Team task handles — the short
AT<n>(.<n>)*handle a task carries: the grammar,AT0, the qualified form, permanence and the refusals. - Team task properties — the optional JSON object a team keeps on a task: who may write it, the bounds, the team's schema, the refusal codes and the change event.
- Team PR review protocol — recording a pull request's review on a team task: the five writes and one readiness query, the seat rules, the verdicts, the readiness rule, the refusal codes and the events.
- Boot-prompt pointers — the setters that point an Agent, an installed Agent or a worker at a compiled boot prompt, who may call them, the refusals, and how to read a pointer back.
- Compile a boot prompt — the three GraphQL mutations that compile a boot prompt from its source, who may call them, and every refusal.
- Serving a boot prompt — which compiled prompt a bind, chat or run is given, the six refusals when none can be served, and the immutable served-prompt record.
- Worker signature and trailer — the per-worker signature and commit trailer templates: defaults, variables, who may edit them, and when an edit takes effect.
Platform & operations¶
- Authorization — roles, grants, guardrails — the action vocabulary, the default role bundles, individual grants, the guardrail (ceiling) layers, and the exact order they compose in. Enforced server-side.
- App grants — who may create, list, revoke and withdraw an App grant, why a new grant is pending until a person signs it, the typed refusals, and the
hadron_create_grantMCP tool. - User profile visibility — which profile fields one user sees of another (name, email, org role), and how user search is scoped to prevent enumeration. Enforced server-side.
- Portal command bar & pages — the portal's ⌘K command bar (open by URN with lenient input, search, create) and its command pages: check access, merge node, and merge memory, with their inputs, outputs, and permissions.
- PDF service HTTP API — the
/convertendpoints of the internalhadrontool-pdfcapability tool: Markdown↔PDF, auth, request/response shapes, errors, and limits. - GraphQL API — full type, query, and mutation reference, auto-generated from the server schema.
- Breaking API changes — arguments the GraphQL API renames with no alias, one entry per change, with old and new names, who is affected, and whether it is deployed.
- GET /capabilities — the unauthenticated endpoint that says which deployment switches are on (
node-signing,passkey-registration), its caching and rate limit, and how a client uses it before selecting a switch-gated GraphQL field. - Data model — entity reference with ERD diagrams, auto-generated from the Prisma schema.
Memory¶
- Node types — the five
nodeTypevalues (info,abstract,record,system,reference), their search behavior, and how to choose between neighbors. - RAG vector index — per-memory semantic retrieval: what gets embedded, when, how to search it (keyword / vector / hybrid + passages), staleness, and the encrypted-memory disclosure.
- Structured storage and queries — opt-in structured storage: the
Memory.schemashape, theNodeWhereInput(where) grammar,NodePropertySort(sortProperty), theobjectTypefacet, and the surface matrix (what's on GraphQL vs. MCP vs. CLI). - Object store API — the record-oriented projection over structured storage:
createObject/object/updateObject/deleteObject/findObjects(GraphQL + MCP), the flat{ id, type, ...fields }record, reserved keys, andmatch/where/sort.
AI automation¶
- Multi-node run flows — field-level reference for multi-node headless runs: routing-edge selection, the
message.datascope,extractionSpec,tools,output, budgets and failure codes, and the run record shape. - Edge conditions — the condition grammar a headless run evaluates as it walks a flow.
- Asset upload — the bi-directional asset-upload tool: what it does, what it stores, limits, lifecycle, encryption.
Chatbots¶
- Conversation routing — how topics, conversations, stages, edges, and goals fit together.
- Chat API — GraphQL surface for driving a Hadron-managed chat from a third-party integration.
AI coding¶
- hadron CLI — command surface, global flags, stable exit codes, URN rules, write semantics, and the raw-GraphQL escape hatch.
- MCP tools — the
hadron_*tools Hadron exposes through MCP, categorized by purpose. - MCP App selection —
hadron_list_appsandhadron_set_active_app: how multi-App OAuth callers pick an App per MCP session, and the typed-error vocabulary.