soft3/neuron/docs/cell foundations and agent review.md

The 0.1 design supplied a substantial durable execution model, but left the connection to cyb cognition underdefined. It could have produced a reliable generic host while leaving the agent fragmented across context, sessions, tools and learning. The 0.2 draft makes those relationships explicit.

The convergent cell is an addressable, loaded ability with bounded state, authority, lifetime and causal history. Local hosting realizes that same model; protocol profiles supply their stronger guarantees. In a robot, soma supplies the thinking and learning that make the hosted ability an agent.

This review changes specifications. No runtime, feature parity or measured advantage over Hermes is claimed.

flowchart LR
    Input[com / sense / plan] --> S[soma: cognition and tasks]
    Context[now / soul] --> S
    S --> C[cell: bounded execution and recovery]
    C --> W[ward: effect and disclosure authority]
    W --> X[tools / models / devices]
    X -->|observations and outcomes| C
    C -->|transitions and provenance| G[cybergraph + bbg: durable history]
    G -->|context and evaluated experience| S
    G --> V[log / time / brain / memory]

Model decisions and external observations follow the same operation boundary. Vault supplies secret operations; body and sigma supply resource accounting.

source precedence and scope

The user's latest definitions and cyb anatomy govern the 21 organs. Product philosophy, product spec, truth, now, soul, filesystem and scripting provide product constraints. Older robot and soma terminology must converge toward anatomy. The ctx bundle, generated on 2026-05-15, is a derived context snapshot; its packer scores and selects graph pages. It can seed retrieval while current sources govern definitions. Its selection heuristic supplies no guarantee of optimal task context.

Soft3 types, terms, rune execution and the stack ownership map constrain implementation. Cyberia development supplies the development doctrine; quality supplies the review lenses: determinism, bounded locality, arithmetic, cryptography, types, errors, adversarial inputs, composition, naming, size, performance and testability. Existing integration findings retain their pinned source revisions in convergence and history ownership.

The Hermes baseline is source inspection of d15ed4445207dda418b984e8bda0f68f48b8c6f3. It covers the mechanisms below; the complete command/tool/platform inventory remains a prerequisite to claiming full functional parity.

Local source revisions at final verification:

Repository Revision
cyb 63e533157ac3f2656c9669b30e412b43e53c1aa0
soft3 77c780cef2ce07bff29855278a465ff070a91102
rune 67281db727070c1160c12be341c6fd85e39e5380
soma 312177028abeecaf252e0c72d25f8e3fc1dc8985
ctx 688d859ffbd2851f3b6976fae73975966a808176
cyberia bacb7a378b9381b7d3df684d1ceedef73cbe5c79
cyber eecfe635a7309d590d6181b67c5636cd5e2021a5
tru de5e5517ed741d5d633519fa9686c2f242f623c7

Cyb advanced from 7f22732cbb0eeb52550dd18dbaa3621bb73eca7d during this review. The inspected anatomy/product/parts/scripting/context contracts are unchanged; the robot host changes were checked separately. Earlier runtime/storage findings retain the source revisions documented in the convergence review.

findings and specification corrections

Severity describes consequence if implemented as previously written. Every correction below is a draft requirement awaiting executable validation.

Finding Consequence Correction in 0.2
High: application context implicit A restored task could silently use another now, soul or workspace Event/Continuation bind context; explicit steering and source manifests
High: installation dominates the model Every small extension could acquire deployment/proof ceremony Immediate local gate path, automatic descriptors, explicit stronger profiles
High: schema identity hashes a name Different meanings could masquerade behind the same schema particle Content-bound schema manifests and pinned suite; concrete codec gate
High: generic history without semantic projection Durable transcript alone gives weak retrieval and learning Soma publishes task/context/skill provenance through the same graph
High: no complete agency contract Cell conformance could be mistaken for a usable agent Agent composition requirements and complete capability inventory
High: delegation lifetime and budgets underspecified Restart loses child ownership or renews spent budget Durable joins, reservations, cancellation and result adoption
High: checkpoint could be confused with workspace rollback Restored VM state could overwrite later user edits Versioned file observations and authorized conditional compensation
Medium: flat snapshot collections Work may scale with total retained history per transition Persistent stack indexes and bounded graph access
Medium: resource economy absent Many parked cells retain unnecessary VMs or model copies Shared immutable artifacts, parking and measured resource budgets
High: learning evidence contract absent Skill adoption lacks explicit task-level evaluation and conviction rules Extend existing evidence distinctions into evaluated skill promotion
High: superiority lacks falsifiable criteria Architecture language substitutes for user outcomes Paired task, recovery, learning and cost gates

The original authority, crash, finality, content-closure and relocation contracts remain essential. In particular, rune authority propagation, graph atomicity and storage error handling remain implementation blockers documented in integration. Specification repairs do not fix those upstream implementations.

all 21 organs

Every responsibility has one owner. This matrix constrains cell's boundary; it does not decide whether every organ should become a standalone repository.

Organ Relationship to cell and agent
name Resolves named definitions/instances; an NFT is unnecessary for local execution
avatar Renders the robot's appearance; cell carries versioned presentation bindings
soul Owns configuration; each ask pins its version, ward applies live policy
ward Enforces effects, disclosure, grants and delegated limits
soma Owns model use, cognition, task strategy, context selection and learning
brain Renders graph state and relations, including task/skill provenance
memory Presents particles as a filesystem; external file adapters declare consistency
com Admits user asks/steering with now and soul and presents progress
sense Owns conversations and recipient delivery; task execution has its own identity
voice Supplies speech inputs/outputs with media artifacts and declared evidence
vision Supplies camera/screen/world observations through permitted device operations
state Supplies external values with verification tiers preserved in receipts/context
sigma Owns assets/neurons and monetary accounting for authorized actions
vault Holds secrets and signs; checkpoints retain references with rebinding rules
log Presents causal history queried from cybergraph
now Owns the graph context hinge; running tasks retain admitted bindings
plan Owns standing orders, schedule occurrence identities and catch-up policy
time Presents log, now and plan around a temporal cursor
body Owns silicon, devices, process supervision, energy and physical resource accounting
cell Hosts loaded abilities with explicit state, authority, lifetime and recovery
radio Transports declared protocols; authority/admission semantics survive transport changes

Repository, organ and instance counts are independent. A standalone cell is justified by its shared host/model across several consumers. Ward and vault are strong independent boundary candidates because enforcement and secret operations need uniform contracts. The remaining extraction decisions require their own dependency and maintenance audit. This task introduces no 22nd core organ.

what Hermes already supplies

The comparison must respect the actual reference rather than a minimal chat loop.

Inspected mechanism Evidence in Hermes Implication for cyb
Durable sessions Session storage describes persisted messages, compaction archives and delivery/delegation bookkeeping A durable transcript is baseline functionality
Extensible context Context engine plugins can replace the context manager, including DAG approaches Graph-shaped context alone is no exclusive advantage
Learning graph Learning graph implementation relates skills and memory chunks and tracks use Cyb must demonstrate better grounded reuse, not just draw a graph
Delegation Delegation guide documents isolated children; top-level async work is process-local and recommends other mechanisms for durable work Persisted, composable joins are a concrete target to validate
Child permissions Toolset implementation bounds child toolsets by parent availability Ward's advantage must be verified across runtimes and effect scopes
Scheduled attempt history Cron execution ledger persists attempts and recognizes unknown interrupted outcomes Explicit uncertainty is already present in part of the reference
Workspace rollback Checkpoint guide describes shared shadow-git storage and preservation of later user edits Cell execution snapshots alone would be a functional regression
Daily agent surface README describes skills, memory, tools, scheduling and multiple interfaces The Rust reimplementation needs usable end-to-end adapters

These are source findings, not comparative reliability measurements. Missing evidence for a global guarantee does not establish a reference defect. Inventory and runtime tests must resolve exact behavior before a parity claim.

where the proposed design can win

The strongest differentiator is composable continuity: one causal structure connects intent, admitted context, tool attempts, children, artifacts, evaluation and skill adoption. It can travel with the task under explicit authority and availability rules. Every organ can inspect the same structure.

Target Mechanism Evidence needed
More reliable long work Context binding, durable continuations and operation reconciliation Complete more interrupted tasks without duplicate effects or lost constraints
Useful experience transfer Source-linked skills with evaluation, scope and revision history Improve held-out outcomes within the same total budget
Portable agent lifetime Content closure, model-independent task state and fenced relocation Resume supported tasks on another host with correct authority and artifacts
Coherent user experience CLI, cyb, time, memory and sense read the same task state Reconnect, steer and inspect through each surface without divergence
Cheaper extensibility Immediate rune gate path, shared runtime/artifacts, parked instances Better setup time, idle cost and useful work per resource unit

Each target is a hypothesis until measured. A graph-shaped execution record may improve auditability while increasing storage and latency. Rust may reduce host overhead while model cost still dominates. The evaluation includes these costs.

research claims and dependency discipline

Context reasoning and runtime reasoning motivate graph-local selection and compiled-model state reuse. They are research directions. This review does not derive optimal general LLM context or lossless arbitrary-model KV transfer from graph connectivity. CT-0 defines its own compiler and conformance; cell treats its artifacts through declared runtime contracts.

Baseline progress needs local inference and practical external adapters, bounded graph retrieval and honest provenance. Stronger proof/consensus/model mechanisms join through profiles as their prerequisites become implemented and validated.

order of implementation after specification review

Stage Concrete exit condition Owning work
1. Foundations Codec/schema fixtures, graph transaction/content closure and ward regression pass hemera/data, cybergraph, bbg, rune, ward
2. One local ability Source gate starts, performs a permitted effect, survives restart and appears in CLI/cyb cell model/engine/node and rune/prysm adapters
3. One complete agent task Soma ask binds now/soul, uses a model and tools, produces tested artifacts and can be steered/recovered soma, now, com, body, vault, memory
4. Daily autonomy Durable delegation, evaluated skill reuse, schedules and delivery pass end-to-end cases soma, plan, sense, radio, sigma
5. Portability and comparison Backup/relocation checks and frozen matched Hermes suite publish scoped results all participating owners

Capability inventory starts before implementation so missing Hermes functionality enters the plan explicitly. Stronger ledger/knowledge profiles can develop in parallel without being prerequisites for an ordinary local ask.

The revised contracts are foundations, agent composition and evaluation, with updates to model, data, execution, authority, history, evolution, integration, lifecycle, API and conformance. Review these before selecting the first executable slice. The intended result is a complete agent whose claimed advantages are demonstrated by work it actually finishes.

Graph