Nucleus Engine

Estimate

[Nucleus POWER · Cost & Energy] Estimate — prices a job BEFORE you run it: USD cost + energy (Wh) + carbon, plus whether your budget allows it. Use it to avoid surprise spend before an inference or render.

Govern

[Nucleus CHAIN + POWER · Governance] Govern — checks a call against your policy + budget and returns allow/warn/block with the full decision trace and estimate. Advisory by default (won't block unless you set a threshold); every decision is written to the tamper-evident Chain audit.

Advise

[Nucleus KITCHEN · Efficiency] Advise — the whole picture before you spend, so you decide (advisory, never blocks): cheapest capable model + savings (Efficiency/Kitchen), energy + carbon in human terms (Power), rights/PII posture (Watch), trust band (Grade), plain-language nudges, and a Chain readback. Call before a task/session to choose how to run it.

Route

[Nucleus GRADE + KITCHEN · Routing] Route — recommends the cheapest model that still clears the trust bar (Grade) for a task. Advice, not an automatic switch — you choose. Returns the pick, its cost, and how much cheaper than the frontier option.

Meter

[Nucleus POWER + CHAIN · Metering] Meter — records actual usage for a call (cost + effective tokens), burns it down against your budget, and writes a Chain record. Never blocks. Returns USD, effective tokens, and a usage_id.

Make Recipe

[Nucleus KITCHEN · Recipes] make_recipe — save this prompt + the models/tools you used so you can reuse it

Readback

[Nucleus CHAIN · Readback] Readback — call this AFTER you finish a task to get a plain-language report back for the chat, from your recorded usage: total cost, energy + carbon in human terms (phone charges, km driven), which models you actually used (named) and what tier they are (e.g. what "frontier" means and when it is overkill), each model's context window and its data/training posture on that plan when on file, plus a specific recommendation — route routine turns to a named cheaper model an

Remember

[Nucleus CORE · Memory] remember — write one durable event to the append-only Core ledger (the "send it in" call). Records a decision/message/commit/note/tool_action/source under a project so it is never lost. Returns the stored event id + echo. Governed: consequential writes (decision) land as PROPOSED and must be approved by an independent actor before they count as approved truth; low-risk kinds auto-approve. Append-only: to fix a prior event set supersedes — the original is never mutated.

Recall

[Nucleus CORE · Memory] recall — read/search the Core ledger (the "get it back" call). With a query it returns the most relevant events (keyword match + recency); without one it lists newest-first. Filter by kinds/status/since/limit. Every event carries its status (approved/proposed/superseded) so a proposed write is never mistaken for approved truth. Tenant+project scope filters apply before results return. Honest-empty ([]) when nothing is logged.

Context

[Nucleus CORE · Memory] context — compile a context pack for a task BEFORE you start (recall everything relevant first). Returns a structured pack: a plain-language summary, the source events (id/ts/kind/title/status), and the full events. Proposed and superseded events are labeled so they are never presented as approved truth. Honest-empty when the project has nothing matching yet.

Approve

[Nucleus CORE · Governance] approve — flip a PROPOSED event to approved and write an immutable approval event (spec §1.8). Enforces the independent-actor rule: for consequential kinds (decision) the approver must differ from the proposer — no agent approves its own consequential write. Returns the approved event + the approval-event echo.

Object Propose

[Nucleus CORE · Context Objects] object_propose — record a TYPED, versioned fact of record (decision | requirement | risk) linked to the source events it came from (source_event_ids). Consequential: it lands PROPOSED and must be approved by an INDEPENDENT actor before it counts as current truth. Use this — not a plain note — when you want "the current approved decision/requirement/risk" to be a deterministic lookup (see object_approve + objects).

Object Approve

[Nucleus CORE · Context Objects] object_approve — flip a PROPOSED typed object to approved (spec §1.8). Enforces the independent-actor rule: the approver must differ from the proposer — no agent approves its own object. If the object supersedes a prior version, that predecessor is retired to `superseded` (kept + retrievable). Returns the approved object + any superseded predecessor.

Objects

[Nucleus CORE · Context Objects] objects — read typed context objects. mode "current" resolves the SINGLE current object of a type by authority precedence (spec §9.4) — the deterministic "what is the current approved decision/requirement/risk?" call. mode "list" returns objects filtered by type/status (superseded stay retrievable). Tenant+project scope applies before results. Honest-empty when nothing is proposed.

Nucleus Start Task

[Nucleus CORE · Agent Runtime] nucleus_start_task — begin a bounded task-run BEFORE you code. Opens a run in the durable store (project + task_id + intent) and returns a compiled context pack from the Core ledger plus a documented-vs-real reconciliation. If an active run already exists for this project+task_id, this RESUMES it and returns the canonical state instead of opening a duplicate — so a different agent (e.g. Cursor after Claude) picks up exactly where the last one left off. All run stat

Nucleus Checkpoint

[Nucleus CORE · Agent Runtime] nucleus_checkpoint — save resumable structured progress at a meaningful boundary (spec §13.4). Appends an IMMUTABLE, ordered checkpoint (summary, files changed, commands, tests + results, blockers, next action) to the run. Never overwrites a prior checkpoint. This is what a different agent reads to resume — keep it honest and specific.

Nucleus Resume Task

[Nucleus CORE · Agent Runtime] nucleus_resume_task — resume a task from CANONICAL state (spec §13.3 / §23.9). Returns the run + every checkpoint in order + the newest as latest_state + a freshly compiled context pack from the ledger. Everything is rebuilt from stored task + repository state — never from a chat transcript or one model's behavioural consistency. Readable by a DIFFERENT agent identity than the one that started the task. Tenant+project scope is enforced: you cannot resume another te

Nucleus Finish Task

[Nucleus CORE · Agent Runtime] nucleus_finish_task — record outcome, tests, evidence, completion and follow-up, then close the run (finished | abandoned). Optionally writes a final checkpoint so the ending state is itself resumable. Projection only — checkpoints and history are preserved. Also appends an agent.run.finished/abandoned event to the durable ledger.