Hephaestus Network

Hephaestus Agent Workforce Network

The active host LLM staffs the task. Agentlas Core federates content menus from registered Local, owner Cloud, and public Hub inventory. No source and no deterministic layer is the decision-maker or a server-side LLM executor.

Source scopes are exact:

• network: Local + Cloud + Hub;
• local: registered Local packages only;
• cloud: the signed-in owner's Cloud packages only;
• hub: public Hub packages only.

Before an unpinned search uses the Local source, Core refreshes every active registered source folder and creates a new immutable release when its safe content snapshot changed. This is discovery freshness; a prepared or goal-bound roster remains pinned to its exact release. network reindex is a rebuildable card-cache operation and is not required to refresh a registered Local release. New folders still require explicit local-register.

Public demos and distribution proof use explicit hub scope. They must not use private Local/Cloud inventory as evidence of public availability.

Resolve the runner and sign in

Network can query owner Cloud and public Hub inventory, so establish the same saved Agentlas session before staffing:

RUNNER=""
for c in \
  "$HOME/.agentlas/runtime/current/bin/hephaestus" \
  ./bin/hephaestus
do [ -x "$c" ] && RUNNER="$c" && break; done
if [ -z "$RUNNER" ]; then
  for cache in \
    "$HOME/.claude/plugins/cache/agentlas-core-engine/hephaestus" \
    "$HOME/.codex/plugins/cache/agentlas-core-engine/hephaestus"; do
    newest="$(ls -d "$cache"/*/bin/hephaestus 2>/dev/null | sort -V | tail -1)"
    [ -n "$newest" ] && [ -x "$newest" ] && RUNNER="$newest" && break
  done
fi
if [ -n "$RUNNER" ] && [ "${HEPHAESTUS_AUTH_AUTOPOPUP:-1}" != "0" ]; then
  "$RUNNER" auth ensure --timeout 180 >/dev/null 2>&1 || true
fi

The browser opens only when no reusable local sign-in exists. In CI or another headless environment, set HEPHAESTUS_AUTH_AUTOPOPUP=0.

Required MCP sequence

First confirm the typed tool menu contains workforce.preflight_work_order and that its _meta.protocolVersion is at least 2026-08-20.1. If the tool is absent, the host is still attached to a preflight-less runtime. Do not fall back to model-authoring the strict wire WorkOrder and do not retry the same invalid call. Return the machine-readable boundary workforce_protocol_upgrade_required; let the runtime's normal verified auto-update finish, then reload the host/MCP session. An explicit hephaestus hep-update remains an operator action, never an implicit skill side effect.

Use the Agentlas Core Workforce contracts in this order:

workforce.preflight_work_order(taskBrief=..., roles=..., edges=...)
workforce.search_candidates(workOrderRef=..., sourceScope="network")
workforce.expand_candidates(selectionSessionId=..., candidates=[{slotId, candidateOrdinal}])
workforce.validate_selection(decision={selectionSessionId, decisionAuthor, assignments})
workforce.prepare_execution(selection=..., federatedSelectionDigest=..., projectDir=..., goalId=activeGoalId?, fullDossier=false)
workforce.validate_execution_receipt(receipt=..., executionPlan=..., toolInventory=...)

Call these exact typed tools directly. Do not enumerate, serialize, print, or search the host's complete ALL_TOOLS registry: the sequence and tool names are already specified here, and dumping unrelated schemas spends context without improving staffing.

The source-internal workforce.fetch_runtime_bundle call is performed by Core from the pinned original source session/digest. The host must not call it directly or replace it with a slug/latest lookup.

The default search response is a projected decision menu. Preserve its source receipts and selectionSessionId, but do not echo that projection as federationResult; Core resolves and revalidates the full pinned result by session. The final receipt call is local, bounded, and read-only. It validates host-produced evidence and never executes workers or creates a receipt.

The current CLI equivalent is workforce search --scope network. A host adapter that does not yet expose typed sourceScope must report that wiring gap; it must not silently call public Hub-only search and label it Network. Do not call the legacy lexical router first. Do not turn install count, ratings, invocation history, source precedence, or a deterministic top score into the staffing decision. If a source is unavailable, preserve its finite failure receipt.

1. Perform job analysis

Act as the active top-level orchestrator. Convert the user's task into one compact semantic draft for workforce.preflight_work_order. Core compiles it into the exact redacted agentlas.workforce-work-order.v1, generates finite WorkOrder/slot/artifact identifiers, fills omitted empty arrays, pins the ontology version, validates the privacy boundary, and returns a one-hour workOrderRef. Keep raw local files, secrets, memory, and private prompt details on the host. Create one roles entry per materially distinct responsibility. Each role may identify:

• semantic role/community and required skill or knowledge concepts, written as plain English phrases when no ontology id is obvious — Core normalizes them into schema-valid concept ids and reports every rewrite as normalizedConcepts. Only these semantic communities, roles, skills, and knowledge may narrow menu fit;
• execution requirements such as required MCP/tool capabilities, runtime, language, modality, and required/forbidden authority. Include these only when the requested action genuinely needs host proof. They never filter, rank, or exclude semantic candidates; Core carries them into the ExecutionContext and the host validates them against its actual tool inventory, capability binding plan, permission policy, and invocation receipt after selection;
• collaboration edges by 1-based role ordinal. An edge is a declaration of handoff, never a qualification requirement. Requiring an artifact almost no published agent declares does not narrow a menu, it empties it;
• executable entity-kind constraints;
• cardinality, criticality, and collaboration edges;
• the minimum evidence level: declared, checked, demonstrated, or attested.

Do not create decorative roles. A single specialist is valid for a genuinely single-role task; a composite task should become a real temporary task force. Executable slots allow only agent or team; group is discovery-only until an authoritative group execution contract exists.

Omit unconstrained runtime, authority, language, and modality arrays. Never invent prefixed finite values such as language:ko or custom slot/artifact IDs; the draft schema exposes the finite values and Core owns mechanical IDs. Use the default candidate policy (2 minimum, 30 maximum per role) unless the task has a concrete recall reason to widen it. Protocol receipt verification is already an independent gate, so do not add a decorative verifier role merely to restate it.

The user never has to write the word goal or enable a goal mode. Before a new search, call workforce.goal_context for the current project. If an active binding is the same ongoing work, treat its goalId and roster as incumbent. Only create a new WorkOrder when no active binding covers the work or the incumbent has a real gap.

2. Retrieve the menu, then make the LLM decision

Before calling any remote source, call workforce.preflight_work_order. Its deterministic compiler and WorkOrder boundary cover every schema-declared string and structured identifier, including nested roles, skills, tools, authorities, artifacts, edges, runtime/language, and policy fields—not only prose fields. A path, personal/account identifier, credential URL, private concept identifier, or secret-like value returns only path/class repair evidence with hubCalls=0 and a null rejected-object digest; never trust a draft's prose, echo the rejected value, or compute/display a digest over rejected data. Ask Core for sourceScope="network" with the returned workOrderRef only after that boundary accepts. Do not echo the compiled WorkOrder. Each source returns a bounded shortlist by default; narrow it, then use workforce.expand_candidates only for the candidates worth a full-card comparison. Each source's complete menu remains pinned in Core; Core validates and unions them using canonical identity ordering. It performs no semantic rerank. Read the exact roles, skills, MCP tools, inputs/outputs, authority, eval evidence, communities, release version, package hash, content digest, source receipt, and provenance.

Source precedence is not ranking. It applies only when the same agentDefinitionId has verified identical lineage issuer/digest and the exact same release version, package hash, content digest, and entity kind at multiple sources. Then and only then shadow Local > Cloud > Hub. Similar names/slugs are never deduplicated. Missing lineage or different releases fail closed for that collided identity: Core quarantines the ambiguous definition and preserves unrelated Local/Cloud/Hub candidates with finite aggregate conflict evidence.

You, the active host LLM, choose the ideal roster. Consider complementary coverage and handoffs, not a scalar top-1 score. Return agentlas.workforce-selection.v1 with decisionAuthor.kind = "host_llm", the real host model id, exact slot/release assignments, graph edges, alternatives, and short reason codes. Some nondeterminism in final judgment is intentional; hard constraints are not.

If a required slot has inadequate coverage, use at most two same-host semantic WorkOrder refinements across the whole decision. A provisional Selection may request content expansion through requestExpansionForSlots; the adapter gives the host only aggregate slot/count/gap data, never candidate identities. Never fill a post with a semantically unrelated agent or repeat an exhausted request.

3. Validate and pin exact releases

Send the compact decision — selectionSessionId, decisionAuthor, and one assignments row per post naming the candidate by its menu candidateOrdinal with reason codes. Core resolves the pinned WorkOrder, federated CandidateSet and federation receipt from that session, fills the candidate-set digest and the arrays that are empty in a normal decision, and compiles the exact agentlas.workforce-selection.v1. Re-plan on rejection. The validator may reject constraints, cardinality, cycles, drift, out-of-menu releases, or a source-pin mismatch; it must never pick for you.

An accepted receipt can still carry unmetRequirementCount. That is not a rejection and Core will not choose for you, but it is not noise either: read selectionValidation.unmetRequirements, then either accept the gap on purpose or pick a different candidate. Never report an accepted validation as if nothing were unmet.

Do not send the merged CandidateSet to remote Hub/Cloud validation or preparation: it is not one of their source sessions. Core pins every assignment to the original source's selection session and candidate-set digest, then fetches the exact release/package/content claims from that source. Call Core federated preparation only after acceptance. Preparation must return agentlas.workforce-execution-plan.v5, status prepared, an exact preparationReceiptId, and an executionRoster whose release version, package hash, and content digest match the candidate set. It returns BYOM directiveBundle records. Every row must declare bundleDigestSchema=agentlas.workforce-runtime-bundle-digest.v4; recompute its canonical digest before execution and fail closed on mismatch. Digest values allow only Unicode-scalar strings, booleans, null, arrays, and ASCII-keyed objects; numbers, invalid keys, and __proto__/prototype/constructor fail closed. A row must also carry a nonblank top-level systemPrompt, instructions, or agentMd, a first-class digest-bound permissionPolicy, and an agent-null/team-authoritative executionGraph. Missing permission declarations become an explicit deny-all policy, never inherited host access; incomplete claimed allowlists fail. The plan's digest-bound executionContext must preserve every validated slot demand, WorkOrder/Selection edge and artifact kind, assignment and reason code. Missing or changed releases create unfilled posts; there is no silent substitution.

4. Bind the roster to the durable goal

Exact preparation must include projectDir and the incumbent goalId when one exists. Core rejects a preparation without a project and automatically binds every successful exact plan before it can be executed. On first contact, when the host has no durable Task/conversation id, Core derives a content-free stable id from the already-required WorkOrder id. Therefore “the user did not explicitly say goal” is never a valid reason to omit continuity. Later accepted preparations with the incumbent goalId append only newly recruited exact releases. Never replace or silently remove an incumbent release.

At the start of every later turn, read workforce.goal_context (the installed SessionStart/UserPromptSubmit hooks also project the same bounded context). Pass knownRevisions with the goalId -> rosterRevision pairs already in this conversation: unchanged goals come back as one line instead of a full roster, which is what makes a continuity read cheap enough to actually do every turn. Read pendingExecution in that response — those releases were prepared and never executed, and the session-end checkpoint reports the same fact to the user, so it cannot be quietly carried forward. The active host LLM must choose exactly one turn posture:

• reuse: the bound roster plus local skills can handle this turn;
• local-only: no bound worker invocation is useful for this turn;
• recruit: a real role/tool/modality gap requires another Network cycle, followed by another additive workforce.bind_goal;
• standby: no worker needs to execute, but the roster remains bound;
• blocked: the goal cannot progress safely.

Record that content-free decision with workforce.record_goal_turn. A turn ending, session closing, runtime restarting, context compaction, worker invocation completing, or a 24-hour Hub lease expiring must not release the roster. The Hub/Web account authority alone decides whether the next actual remote preparation is covered by an existing same-account lease or creates a new charge. Never manufacture lease state locally.

standby means a durable roster binding available to later turns. It does not mean a continuously running model, process, socket, or background token burn. Only actual worker invocations accumulate their normal Memory Curator and Experience records.

Call workforce.complete_goal only after the user or authoritative host goal state explicitly marks the whole goal completed or cancelled. It requires explicitCompletion=true; lease expiry and successful completion of one model turn are invalid completion signals.

5. Resolve the model for every invocation

Before every bound planner/manager, worker, synthesis, and verifier invocation, advertise the current host's real available sessions and call model.resolve_allocation. The host owns the invocation stage:

• planner, manager-plan, and nested-manager resolve as orchestrator;
• worker, execute, delegate, and task resolve as worker;
• manager-synthesis and synthesis resolve as orchestrator;
• verifier resolves as orchestrator.

Model pins and ceilings come only from the MCP server's operator policy in AGENTLAS_MODEL_ALLOCATION_POLICY_JSON; a task, Hub bundle, or tool argument must not override them. Prefer role-scoped orchestrator and worker policy objects. A missing worker policy inherits the orchestrator for quality; orchestrator never falls through to a cheaper worker policy.

Use the allocation receipt's exact provider, model, session, and effort for the invocation. The pre-invocation receipt has usage: null; record observed usage only in the later invocation/run receipt. A resolved allocation proves policy selection, not that a worker model ran. If the host cannot launch the selected provider/model as a distinct invocation, stop at allocation-only evidence and report that boundary.

6. Execute the real task force

Run the prepared roster through the current host runtime:

1. planner/manager creates structured worker assignments;
2. each selected worker runs in a distinct model invocation with its exact release directive and only the needed local grounding;
3. workers emit explicit handoff artifacts;
4. synthesis runs after dependencies complete;
5. an independent verifier checks the requested result.

When a prepared release is itself a Team, honor its authoritative manager/worker/synthesis graph; do not flatten it into one prompt. Follow the row policy intersected with the host policy for all side effects. Unsupported enforcement blocks execution. zero-tools requires an actually empty tool inventory; a residual primitive isolated by forced read-only/no-filesystem is no-authority-sandbox, not zero-tools.

Snapshot the just-in-time policy-filtered local tool menu as a private agentlas.workforce-tool-inventory.v1 artifact and give only that menu to the executor planner. Never send the snapshot to Hub. The host LLM, not Hub or a lexical rule, records the pair-scoped capabilityBindingPlan; its context, tool-inventory, and planner-bound digests must validate before execution. Every required tool capability maps to an exact snapshot entry and permitted tool. A package policy mentioning a tool is not inventory proof, and a required binding cannot run under no-authority enforcement.

If the host cannot create distinct child invocations, stop at prepared and say so. A route id, bundle id, process exit code, or prose that imitates several roles is not execution proof.

Preparation is not delivery. A turn that pinned a roster and produced no worker output has not answered the user's request, and the session-end checkpoint says so in the user's own view. Either run the roster and record the turn with workforce.record_goal_turn, or tell the user plainly that the team is prepared but not executed and what remains.

7. Truthful receipts

For an executed task force, retain one joined agentlas.workforce-execution-receipt.v2 joined to the exact v5 plan containing:

• selection and preparation receipt ids;
• orchestrator and planner model/invocation ids;
• planner.parseSuccess, planner.fallbackUsed, toolInventoryDigest, and capabilityBindingPlanDigest;
• every roster row's exact release/package/content/bundle/policy/graph digests, capability bindings, and handoff artifact refs;
• either one real direct invocation or a nested receipt proving manager-plan parse/no-fallback, exact declared workers in graph order, and manager-synthesis; never fabricate one aggregate team invocation;
• every actual model/provider/runtime, requested/applied effort with evidence, globally unique invocation id, and permission-enforcement evidence whose exact granted tool ids and inventory digest match the binding plan;
• synthesis and verifier invocation ids and verifier verdict.

Never report success when planner JSON fell back, child receipts are missing, or verification did not pass. In the user-facing summary, name the actual workers and distinguish selected, prepared, and executed.