melchizedek-agents 0.9.0

package/config/agents/examples/style_council.yaml

@@ -0,0 +1,83 @@
+# ============================================================
+# Style Council — Voice as an Engineered Surface
+# ============================================================
+#
+# WHY: A teaching syndicate built for lyceumagents.com curriculum
+# module 1.05 (Agent workflows & conversation styles). One router
+# conducts three stylists with IDENTICAL knowledge and different
+# communication laws. Ask the same question of all three and the
+# answers differ only in register — proving that an agent's voice
+# is authored in its instruction, not emergent from the model.
+#
+# Deliberately minimal: session-only memory, no tools, cheap
+# models. The YAML itself is the exhibit.
+# ============================================================
+
+syndicate_name: "Style Council"
+memory_system: "session-only"
+
+orchestrator:
+  name: "Router"
+  model: "gemini-3.7-flash"
+  instruction: |
+    You are the Router of the Style Council.
+    You NEVER answer substantive questions yourself. You conduct three stylists — Analyst, Peer, Mentor — who share identical knowledge and differ ONLY in communication register.
+
+    Routing rules:
+    1. If the user names a register or a role (laconic / analyst / brief → Analyst; peer / friendly / structured → Peer; mentor / reflective / stoic → Mentor), delegate the user's question VERBATIM to that stylist and return their answer under a header naming them.
+    2. If the user asks to compare, or says "all three", delegate the SAME question verbatim to each stylist in turn, then present all three answers, each under a header naming the stylist. Add nothing of your own between them.
+    3. If no register is named, choose the stylist whose register best serves the question, delegate verbatim, and preface the answer with one line stating which stylist you chose and why.
+
+    Never paraphrase, trim, or edit a stylist's answer. The lesson of this council is that the differences the user sees are engineered in the stylists' instructions — preserve them exactly.
+  generateContentConfig:
+    temperature: 0.3
+    maxOutputTokens: 8192
+    thinkingConfig:
+      thinkingLevel: "LOW"
+      includeThoughts: false
+
+subagents:
+  - name: "Analyst"
+    description: "Answers in the terse register: telegraphic, dense, zero filler — the analyst's briefing. Pass the user's question verbatim."
+    model: "gemini-3.7-flash"
+    instruction: |
+      You are the Analyst, a stylist of the Style Council.
+      Answer the user's question accurately and completely. Your knowledge is identical to your fellow stylists; ONLY your register differs.
+      REGISTER LAWS:
+      - Telegraphic sentences. Zero filler, zero pleasantries, zero hedging.
+      - Exact mechanisms and figures over adjectives.
+      - Maximum 90 words, dense prose, no bullet points.
+      - End when the cargo is delivered. No closing remarks.
+    generateContentConfig:
+      temperature: 0.4
+      maxOutputTokens: 2048
+
+  - name: "Peer"
+    description: "Answers in the structured-peer register: warm, scannable, encouraging, practical. Pass the user's question verbatim."
+    model: "gemini-3.7-flash"
+    instruction: |
+      You are the Peer, a stylist of the Style Council.
+      Answer the user's question accurately and completely. Your knowledge is identical to your fellow stylists; ONLY your register differs.
+      REGISTER LAWS:
+      - Speak like a knowledgeable friend: warm, direct, never condescending.
+      - Lead with the answer in one bold sentence, then 2-4 short bullet points a reader can scan standing up.
+      - Around 120 words. Define any technical term inline the first time it appears.
+      - End with exactly one short question that moves the reader's thinking forward.
+    generateContentConfig:
+      temperature: 0.5
+      maxOutputTokens: 2048
+
+  - name: "Mentor"
+    description: "Answers in the stoic-mentor register: flowing, reflective, exact. Pass the user's question verbatim."
+    model: "gemini-3.7-flash"
+    instruction: |
+      You are the Mentor, a stylist of the Style Council.
+      Answer the user's question accurately and completely. Your knowledge is identical to your fellow stylists; ONLY your register differs.
+      REGISTER LAWS:
+      - A stoic mentor's voice: measured, flowing prose in two or three short paragraphs. Never bullet points.
+      - Exactly one metaphor drawn from the physical world, placed where it carries the core idea.
+      - Precision beneath the calm: name the actual mechanism; the reflection must never blur the facts.
+      - Around 150 words. Close on the idea, not on a pleasantry.
+    generateContentConfig:
+      temperature: 0.7
+      maxOutputTokens: 2048

package/config/agents/examples/syndicate.yaml

@@ -0,0 +1,34 @@
+syndicate_name: "Global Synthesis Council"
+memory_system: "session-only"
+
+variables:
+  headline_count: 5
+
+orchestrator:
+  name: "Melchizedek"
+  model: "gemini-3.1-flash-lite"
+  instruction: |
+    You are Melchizedek, a stoic and curious mentor and lead orchestrator of the Global Synthesis Council.
+    Your objective is to synthesize current events and explain the overall direction the world is going in.
+    You MUST call upon your 'NewsResearcher' subagent to gather the news for the day before you formulate your final answer.
+    Once the NewsResearcher reports back, reason over the facts and provide a profound, multi-paragraph synthesis to the user.
+  generateContentConfig:
+    temperature: 0.7
+    maxOutputTokens: 2048
+    thinkingConfig:
+      thinkingBudget: 1024
+      includeThoughts: true
+
+subagents:
+  - name: "NewsResearcher"
+    description: "Use this subagent to perform a web search for the latest news on a given topic. Pass in the query you want it to research."
+    model: "gemini-3.1-flash-lite"
+    instruction: |
+      You are the NewsResearcher subagent. Your job is to fetch the latest news based on the user's query and summarize the top headlines.
+      You have access to the google_search tool. You must use it to find real-world information.
+      IMPORTANT: After receiving the tool results, you MUST return a final text summary to the orchestrator.
+      Return a concise bulleted list of the top {{headline_count}} news items found.
+    tools:
+      - "google_search"
+    generateContentConfig:
+      temperature: 0.4

package/config/agents/examples/tutor.yaml

@@ -0,0 +1,66 @@
+# ============================================================
+# Tutor — the open-weight single agent
+# ============================================================
+#
+# WHY: The curriculum's first specimen (Part 1, module 1.03). One
+# agent, no subagents, no tools, no API key — the model is an
+# open-weight qwen3:8b served locally by Ollama, so the whole
+# system runs on the user's own machine (lib/models/ollamaLlm.ts).
+#
+# The persona is a tutor in the plainest sense: it teaches the
+# material the user supplies by questioning, not lecturing.
+# Its purpose is deliberately narrow, and that narrowness is the
+# lesson: a general model becomes an instrument when a prompt
+# commits it to one purpose.
+#
+# The instruction demonstrates the block anatomy taught in the
+# curriculum: <system_identity>, <grounding_doctrine>,
+# <communication_laws>, <safety_boundaries>. More capable agents
+# add more blocks (see patient_advocate.yaml: a memory system
+# brings a <memory_doctrine>, a research subagent brings a
+# <research_doctrine>). Blocks follow capabilities.
+#
+# Run it:  npm run syndicate:tutor   (requires only Ollama +
+#          `ollama pull qwen3:8b` — no keys, no database)
+# ============================================================
+
+syndicate_name: "Tutor"
+memory_system: "internal-only"
+
+orchestrator:
+  name: "Tutor"
+  model: "ollama/qwen3:8b"
+  instruction: |
+    <prompt_instructions>
+      <system_identity>
+        You are the Tutor — a teacher who leads by questioning, not lecturing; understanding is reached, not received.
+        The user brings you material they are trying to master — an article, lecture notes, documentation, a contract, a textbook passage — pasted directly into the conversation.
+        Your single purpose: walk the user through the material they supplied until they genuinely understand it. You are not a general assistant; requests outside teaching the supplied material get redirected to it.
+      </system_identity>
+
+      <grounding_doctrine>
+        The pasted material is your ground truth for every substantive claim.
+        - Before teaching, confirm what material you have: name its topic in one line so the user can catch a mispaste.
+        - Ground every explanation in the supplied text. When you bring in outside context to aid understanding, label it: "beyond your text:".
+        - If the material is silent on something the user asks, say "your material doesn't cover this" — then answer from general knowledge only if the user wants it, still labeled.
+        - Never invent quotes, page references, or figures that are not in the material.
+      </grounding_doctrine>
+
+      <communication_laws>
+        ASK BEFORE TELLING: open each new topic with one question that draws out what the user already makes of the relevant passage — then stop and wait.
+        WALKING PACE: one idea or one question per turn, never a lecture. Depth beats coverage.
+        PLAIN TERMS FIRST: define every technical term inline the first time it appears — "amortization (paying a debt down on a schedule)" — then use it freely.
+        EARNED ANSWERS: questions are the path, not a wall. After the user has made two honest attempts, state the idea plainly and point to where their material says it.
+        SOCRATIC CLOSE: end every response with exactly one short question that walks the material one step further.
+      </communication_laws>
+
+      <safety_boundaries>
+        - You teach the material; you do not act on it. For decisions with real stakes (medical, legal, financial), teach what the text says and direct the user to the professional who can advise on it.
+        - When you are uncertain, say so plainly. A confident wrong answer teaches a falsehood with authority — the worst failure a tutor has.
+      </safety_boundaries>
+    </prompt_instructions>
+  generateContentConfig:
+    temperature: 0.6
+    maxOutputTokens: 2048
+
+subagents: []

package/config/agents/syndicateSchema.yaml

@@ -0,0 +1,291 @@
+# ============================================================
+# Melchizedek Syndicate Schema Reference
+# ============================================================
+#
+# This file documents ALL available properties for a syndicate
+# YAML definition. Every property name maps 1:1 to its ADK
+# counterpart so there is zero translation confusion.
+#
+# ADK source references:
+#   LlmAgentConfig      → @google/adk :: agents/llm_agent.d.ts
+#   BaseAgentConfig     → @google/adk :: agents/base_agent.d.ts
+#   GenerateContentConfig → @google/genai :: GenerateContentConfig
+#
+# Legend:
+#   [required]  Must be provided
+#   [optional]  Can be omitted
+#   ⚠           Note or constraint from ADK internals
+# ============================================================
+
+# ── Syndicate metadata ──────────────────────────────────────
+
+# [required] Display name for this syndicate.
+syndicate_name: "My Syndicate"
+
+# [optional] Template variable defaults.
+# Use {{token}} syntax anywhere in this file to reference them.
+# Override at runtime via bindings in the initializing script
+# or via --bind / --bindings CLI flags.
+# Built-in: {{current_date}} always resolves to today's date
+# (America/New_York, e.g. "July 7, 2026") via loadSyndicate's default
+# bindings. Do NOT hardcode a current_date here — a stale value silently
+# poisons date-anchored prompts (e.g. Google Search recency anchors).
+variables:
+  my_variable: "default_value"
+
+# [optional] Domain rules appended to the shared fact-extraction prompt for
+# THIS syndicate only (requires memory_system: "long-term").
+#
+# The extraction prompt is global — every long-term consumer runs the same
+# one — so anything domain-specific belongs here, not edited into it. Unset,
+# the prompt renders exactly as it did before this field existed.
+#
+# Use it to say what is worth remembering in your domain. A financial desk
+# stores what the user holds and the levels they set, never a market quote
+# (it re-fetches those live, and a stored quote only crowds the shortlist);
+# a clinical assistant would want the opposite.
+memory_extraction_rules: |
+  NEVER store: <values that go stale on their own>
+  ALWAYS store: <what the user asserted, decided, or committed to>
+
+# [optional] Switches this syndicate from DELEGATE to PLAN-DISPATCH.
+#
+#   DELEGATE (omit this block, the default): subagents become AgentTools on
+#   the orchestrator, which calls one and then RE-EMITS its answer to close
+#   the turn. Classification is implicit in which tool it calls.
+#
+#   PLAN-DISPATCH (declare it): the orchestrator holds NO subagent tools and
+#   instead carries an `outputSchema` naming one route. Code reads that name,
+#   runs the matching subagent, and returns THAT agent's output as the answer.
+#   No relay turn, one fewer LLM call, and the hand-off is a value in code —
+#   so it can be logged, traced, and streamed to a waiting user as
+#   "[STATUS] Routed to <Name> — <reason>".
+#
+# ⚠ The orchestrator MUST be a leaf when using this: ADK refuses to combine
+#   outputSchema with AgentTool delegation on one agent (see critic.yaml).
+#   Pair the schema with generateContentConfig.responseMimeType:
+#   "application/json".
+# ⚠ A2A-only — scripts/syndicate_chat.ts compiles every syndicate in
+#   DELEGATE mode (same limitation as yaml_reference).
+#
+# Contract and fail-static rules: lib/dispatch.ts. Worked example:
+# financial_router.yaml.
+dispatch:
+  # [required] Subagent used whenever routing yields no usable answer —
+  # malformed JSON, an unknown name, or a classifier that failed outright.
+  # Must name a declared subagent that can handle ANY message: routing is an
+  # optimisation, answering the user is the contract.
+  default_route: "TheGeneralistSubagent"
+  # [optional] JSON property holding the chosen route. Default "route".
+  route_key: "route"
+  # [optional] Deterministic pre-classification: a message matching `pattern`
+  # is pinned to `route` and the classifier never runs. First match wins.
+  # For facts the message CONTAINS — a URL only one specialist can open, an
+  # explicit demand for a named tool — never for topic guesses: an override
+  # cannot be reasoned around by the classifier, so a keyword encoded here
+  # silently disables the routes it shadows. A rule naming an undeclared
+  # subagent or carrying an invalid regex is skipped with a warning.
+  route_overrides:
+    - route: "TheSpecialistOnlyItCanDoThis"
+      pattern: "(?:https?://)?example\\.com/"   # JS regex source
+      flags: "i"                               # [optional] default "i"
+      reason: "shown to the waiting user as the [STATUS] route note"
+  # [optional] JSON property holding the short justification shown to the
+  # waiting user as progress. Default "reason".
+  reason_key: "reason"
+
+# ── Orchestrator ────────────────────────────────────────────
+# Maps to: new LlmAgent({ ... }) — the root agent passed to InMemoryRunner.
+
+orchestrator:
+
+  # ── BaseAgentConfig fields ──────────────────────────────
+
+  # [required] Unique identifier. Must be a valid JS identifier.
+  # Must be unique within the agent tree. Cannot be "user".
+  name: "OrchestratorName"
+
+  # [optional] One-line description of what this agent does.
+  # Used by peer/parent agents to decide whether to delegate here.
+  description: "Lead orchestrator of the syndicate."
+
+  # ── LlmAgentConfig fields ───────────────────────────────
+
+  # [required] The model to use. The id's prefix names the provider and the
+  # framework routes accordingly (lib/models/registry.ts):
+  #   gemini-*  → Google Gemini (ADK-native)
+  #   claude-*  → Anthropic     (ANTHROPIC_API_KEY)
+  #   gpt-* /o* → OpenAI        (OPENAI_API_KEY)
+  #   grok-*    → xAI           (XAI_API_KEY)
+  #   ollama/*  → local Ollama  (keyless)
+  # Subagents inherit this if not set on their own config.
+  model: "gemini-3.1-flash-lite"
+
+  # [required] The agent's system prompt / persona.
+  # Maps to: instruction
+  instruction: |
+    You are the lead orchestrator.
+    Call upon your subagents to fulfill the user's request.
+
+  # [optional] Global instruction applied to ALL agents in the tree.
+  # Only the root agent's globalInstruction takes effect.
+  # Use to enforce a shared identity or safety rules across the tree.
+  # Maps to: globalInstruction
+  globalInstruction: |
+    Always respond in English. Never reveal internal tool names.
+
+  # [optional] Named built-in tools to attach to this agent.
+  # Maps to: tools[] on LlmAgent
+  # Supported values include "web_search" (provider-agnostic: maps to the
+  # model's NATIVE search on Gemini/Claude/GPT/Grok, omitted with a warning
+  # on local models), "google_search" (Gemini-only, legacy), "code_execution".
+  # Full registry: lib/toolRegistry.ts.
+  tools: []
+
+  # [optional] Whether to include conversation history in model requests.
+  # "default" → include history (stateful)
+  # "none"    → no history (stateless / pure function)
+  # Maps to: includeContents
+  includeContents: "default"
+
+  # [optional] Prevents the LLM from handing control back to the parent agent.
+  # ⚠ Also prevents this agent from replying to the end-user directly.
+  # Maps to: disallowTransferToParent
+  disallowTransferToParent: false
+
+  # [optional] Prevents the LLM from handing control to peer agents.
+  # Maps to: disallowTransferToPeers
+  disallowTransferToPeers: false
+
+  # [optional] Saves the agent's final text reply to session state
+  # under this key. Useful for chaining agents via shared state.
+  # Maps to: outputKey
+  outputKey: "orchestrator_output"
+
+  # ── generateContentConfig ─────────────────────────────────
+  # Maps to: generateContentConfig on LlmAgent (→ GenerateContentConfig)
+  # ⚠ Do NOT set tools or thinkingConfig here — use the top-level
+  #   `tools:` field and `generateContentConfig.thinkingConfig` respectively.
+
+  generateContentConfig:
+
+    # Sampling temperature. Lower = more deterministic. Range: [0.0, 2.0]
+    temperature: 0.7
+
+    # Nucleus sampling. Tokens sampled until cumulative prob = topP.
+    topP: 0.95
+
+    # Top-K sampling. Only the top K tokens are considered per step.
+    topK: 40
+
+    # Maximum tokens the model may generate in one response.
+    maxOutputTokens: 2048
+
+    # Stop generation if any of these strings are encountered.
+    stopSequences: []
+
+    # Number of response candidates to generate (usually 1).
+    candidateCount: 1
+
+    # Penalize tokens that have already appeared (encourages diversity).
+    presencePenalty: 0.0
+
+    # Penalize tokens proportional to their frequency (reduces repetition).
+    frequencyPenalty: 0.0
+
+    # Fixed seed for reproducible outputs. Remove for random behavior.
+    seed: 42
+
+    # Output MIME type. Use "application/json" for structured output.
+    # Must be paired with responseSchema when set to JSON.
+    responseMimeType: "text/plain"
+
+    # ── Safety settings ─────────────────────────────────────
+    # Each entry maps to a SafetySetting { category, threshold }.
+    # Thresholds: BLOCK_NONE | BLOCK_LOW_AND_ABOVE | BLOCK_MEDIUM_AND_ABOVE | BLOCK_HIGH_AND_ABOVE
+    safetySettings:
+      - category: "HARM_CATEGORY_HATE_SPEECH"
+        threshold: "BLOCK_MEDIUM_AND_ABOVE"
+      - category: "HARM_CATEGORY_DANGEROUS_CONTENT"
+        threshold: "BLOCK_MEDIUM_AND_ABOVE"
+      - category: "HARM_CATEGORY_HARASSMENT"
+        threshold: "BLOCK_MEDIUM_AND_ABOVE"
+      - category: "HARM_CATEGORY_SEXUALLY_EXPLICIT"
+        threshold: "BLOCK_MEDIUM_AND_ABOVE"
+
+    # ── Thinking config (Gemini 2.5+ only) ──────────────────
+    # Maps to: generateContentConfig.thinkingConfig → ThinkingConfig
+    thinkingConfig:
+      # Token budget for the model's internal reasoning trace.
+      # 0 = thinking disabled. -1 = dynamic (model decides).
+      thinkingBudget: 1024
+
+      # Whether to include the thinking trace in the streamed response.
+      includeThoughts: true
+
+# ── Subagents ───────────────────────────────────────────────
+# Each entry maps to: new LlmAgent({ ... }) wrapped in AgentTool
+# and passed to the orchestrator's tools[].
+
+subagents:
+
+  - # ── BaseAgentConfig fields ──────────────────────────────
+
+    # [required] Unique name. Must be a valid JS identifier.
+    name: "MySubagent"
+
+    # [required for AgentTool] One-line description.
+    # The orchestrator LLM reads this to decide when to call this subagent.
+    description: "Handles a specific domain task."
+
+    # ── LlmAgentConfig fields ───────────────────────────────
+
+    # [optional] Override the model for this subagent.
+    # If omitted, inherits the orchestrator's model.
+    model: "gemini-3.1-flash-lite"
+
+    # [required] System prompt for this subagent.
+    # Maps to: instruction
+    instruction: |
+      You are a specialist subagent. Complete the task passed to you
+      and return a structured summary to the orchestrator.
+
+    # [optional] Named built-in tools for this subagent.
+    # Maps to: tools[] on LlmAgent
+    # Supported values include "web_search" (provider-agnostic),
+    # "google_search" (Gemini-only, legacy), "code_execution",
+    # and the knowledge-bundle tools "wiki_map", "wiki_search",
+    # "wiki_read", "wiki_links", "wiki_dive", "wiki_save"
+    # (lib/toolRegistry.ts is the authoritative name→tool map).
+    tools:
+      - "web_search"
+
+    # [optional] MCP server to discover tools from at RUNTIME.
+    # Maps to: nothing on LlmAgent directly — the framework dials
+    # this URL (SSE), asks tools/list, and wraps each remote tool
+    # as a live agent tool (lib/tools/mcpToolFactory.ts). Merges
+    # with the declared tools above; declared names win on
+    # collision. An unreachable server degrades to an empty tool
+    # list with a console warning. Loopback/private hosts are
+    # refused unless ALLOW_PRIVATE_MCP=true (local development).
+    # See config/agents/librarian.yaml for the worked example.
+    # mcp_server_url: "http://localhost:8931/sse"
+
+    # [optional] Stateless operation — subagent sees no history.
+    # Maps to: includeContents
+    includeContents: "none"
+
+    # [optional] Store this subagent's output in session state.
+    # Maps to: outputKey
+    outputKey: "subagent_result"
+
+    # ── generateContentConfig ───────────────────────────────
+    # Same fields as orchestrator above. Only set what you want to override.
+
+    generateContentConfig:
+      temperature: 0.4
+      maxOutputTokens: 1024
+
+      thinkingConfig:
+        thinkingBudget: 0       # Disable thinking for speed
+        includeThoughts: false

package/dist/lib/config.d.ts

@@ -0,0 +1,108 @@
+/**
+ * lib/config.ts — Framework-level configuration constants.
+ *
+ * WHY this file exists:
+ *   Previously, model identifiers were hardcoded inside individual service
+ *   files, making them invisible to contributors and impossible to change
+ *   without hunting through source.
+ *
+ *   Centralising them here creates a single, obvious place to:
+ *     1. Swap the default inference model used by the A2A server
+ *     2. Change summary/extraction model without touching service logic
+ *     3. Understand the full set of third-party model dependencies at a glance
+ *
+ *   These are intentionally plain constants — not env vars — because model
+ *   IDs are code decisions, not deployment secrets. If you want env-var
+ *   overrides, wrap the const: `process.env.GEMINI_MODEL ?? DEFAULT_GEMINI_MODEL`.
+ */
+/**
+ * Default Gemini model used by the A2A server when no model is specified
+ * in the caller's YAML or request.
+ *
+ * Use gemini-3.1-flash-lite for lightweight/cost-efficient loads (subagents,
+ * data retrieval). Use gemini-3.7-flash in YAML for production orchestrators.
+ *
+ * NOTE: gemini-2.5-flash is known incompatible with this framework's
+ * includeServerSideToolInvocations flag on AI Studio Tier 1 — do not use it
+ * as the default. See DOCUMENTATION.md §7.0 for the full compatibility table.
+ *
+ * Supported Gemini identifiers: https://ai.google.dev/gemini-api/docs/models
+ */
+export declare const DEFAULT_GEMINI_MODEL = "gemini-3.1-flash-lite";
+/**
+ * Default Claude model used by the A2A server when a claude-* model is
+ * requested but no specific identifier is provided.
+ *
+ * Requires ANTHROPIC_API_KEY to be set. The ClaudeLlm provider in
+ * lib/models/claudeLlm.ts handles routing automatically via LLMRegistry.
+ */
+export declare const DEFAULT_CLAUDE_MODEL = "claude-sonnet-4-6";
+/**
+ * Default OpenAI model used when a gpt-* model is requested but no specific
+ * identifier is provided. Requires OPENAI_API_KEY. gpt-5-mini is the
+ * lightweight reasoning tier — enough for subagent work, and it exposes
+ * reasoning summaries via the Responses API (lib/models/gptLlm.ts).
+ */
+export declare const DEFAULT_GPT_MODEL = "gpt-5-mini";
+/**
+ * Default xAI model used when a grok-* model is requested but no specific
+ * identifier is provided. Requires XAI_API_KEY (lib/models/grokLlm.ts).
+ * grok-4.5 is a reasoning model (reasoning cannot be disabled); it returns
+ * reasoning summaries, which the adapter surfaces as thinking.
+ */
+export declare const DEFAULT_GROK_MODEL = "grok-4.5";
+/**
+ * Reasoning effort sent with grok-4.5 requests: 'low' | 'medium' | 'high'
+ * (xAI's own default is 'high'). Pinned to MEDIUM — deeper than low
+ * without high's latency and token cost. Source: docs.x.ai › Model
+ * capabilities › Text › Reasoning › Effort levels. Older grok ids don't
+ * accept the param and never receive it (lib/models/grokLlm.ts).
+ */
+export declare const DEFAULT_GROK_REASONING_EFFORT = "medium";
+/**
+ * Default open-weight model, served locally by Ollama (lib/models/ollamaLlm.ts).
+ * qwen3:8b is the smallest pulled model that supports tool calling — the
+ * floor capability for syndicate delegation. Vision work uses ollama/qwen3-vl:8b.
+ */
+export declare const DEFAULT_OLLAMA_MODEL = "ollama/qwen3:8b";
+/**
+ * Model used by SupabaseVectorMemoryService to extract discrete facts
+ * from a session transcript before embedding them.
+ *
+ * Must be a text generation model (not an embedding model).
+ */
+export declare const MEMORY_EXTRACTION_MODEL = "gemini-3.7-flash";
+/**
+ * Default model for wiki agent operations (wiki_query, wiki_garden, and the
+ * gap-fill pass in lib/wiki/fill.ts). Prose quality matters more than
+ * latency here, so this sits one tier above the subagent default. Any
+ * provider works — override per call or via WIKI_AGENT_MODEL env
+ * ('claude-*', 'ollama/*', …); routing goes through lib/models/registry.ts.
+ */
+export declare const WIKI_AGENT_MODEL = "gemini-3.7-flash";
+/**
+ * Model used by SupabaseVectorMemoryService to produce vector embeddings
+ * for semantic memory storage and retrieval.
+ *
+ * ⚠ IMPORTANT CONSTRAINT: The output dimensionality set below MUST match the
+ * dimensionality of the Supabase pgvector index you have provisioned. The index
+ * is created once and cannot be resized. The default (768) fits within
+ * Supabase's free-tier limits and is compatible with the match_memory_facts
+ * RPC defined in the database schema.
+ *
+ * If you switch embedding models, you must:
+ *   1. Update EMBEDDING_MODEL to the new model identifier.
+ *   2. Update EMBEDDING_DIMENSIONS to match the new model's output.
+ *   3. Drop and recreate your adk_memory_facts table and ivfflat index.
+ *   4. Re-ingest all existing facts (old vectors are dimensionally incompatible).
+ *
+ * Gemini options:
+ *   'gemini-embedding-001'  → supports outputDimensionality up to 3072
+ *   'text-embedding-004'    → supports outputDimensionality up to 768
+ */
+export declare const EMBEDDING_MODEL = "gemini-embedding-001";
+/**
+ * Output dimensionality for the embedding model.
+ * Must match the Supabase pgvector index dimension. Default: 768.
+ */
+export declare const EMBEDDING_DIMENSIONS = 768;