Jarvis

The assistant persona that plans, delegates, and answers for the work CoreLayer runs.

Jarvis is the built-in assistant persona. It is not a separate service — it is the agent runtime's conversation loop plus the system prompt, tools, and governance around it. It coordinates work; it does not bypass the permission system to do so.

How a turn works

Input (text or voice transcript)
  → context assembly
      conversation history (branch-aware)
      long-term memory (MEMORY.md / USER.md snapshot + retrieved memories)
      core tools + on-demand tools selected by the router
  → Model Gateway resolves the active provider and model
  → Vercel AI SDK v6 loop; step control lives in prepareStep
  → tool calls are gated by the Permission Guard
  → loop breakers, step budget, and turn timeout bound the turn
  → completion policy and quality gates decide whether to deliver

What it can do

Capability	Notes
Conversation and tool calling	Native tools, MCP tools, and discovered tools
Long-running work	Dispatch a run; it survives the chat turn and is cancellable by id
Coding delegation	Hands a task to a managed external executor inside a worktree
Task graphs	Declarative DAG of sub-tasks, independent nodes run concurrently
Memory	Persistent long-term memory plus a frozen snapshot injected per turn
Voice	Wake word, streaming ASR, sentence-level TTS, barge-in
Scheduling	Cron-driven autonomous runs

Memory

Memory is layered rather than a single chat history:

Explicit Markdown memory — USER.md and MEMORY.md under the memory zone, readable and editable by you
Retrieved memories — stored entries scored against the current turn, above a configurable minimum score (defaults.memoryMinScore)
Frozen snapshot — the injected memory is snapshotted per run, so a turn cannot see memory that was written after it started
Injection tracking — memories track when they were last injected, so frequently retrieved ones are not re-injected indefinitely

Delegation

Jarvis can hand a task to a child run, but the bounds can only narrow: paths intersect, tools intersect, required checks are additive, and the executor and model are pinned. See Executor Governance.

Honesty constraints

Jarvis is told to treat tool output as data, not instructions.
A completion is only delivered when the configured evidence (tests, artifacts, scope) supports it.
When the adversarial judge is enabled and rejects a diff, the attempt is retried within adjudication.maxRetries and then fails rather than being reported as success.

Next Steps

Tool Registry — how tools reach Jarvis
Permission Guard — what it may execute
Memory Runtime — how memory is stored and retrieved