Arsumbris synthesis

Type: agentic-system-analyses/types/agentic-system-synthesis.md

Bounded synthesis

From installation to a shared knowledge workspace

Arsumbris describes an agent-native IDE that combines file-backed knowledge, a typed live graph, UI projections and agent tools. This analysis covers its installation and architecture entry repository at commit 716bb14f2bf50e249d414525c8e07c50478fc786, with an applicability cutoff of 2026-10-15. The boundary is a complete artifact, partial loop, not the implemented platform: engine, host, agent kernel, adapters and substantive knowledge packages reside in separately installed sibling repositories. SRC-1 supplies documentary design, attributed testing claims and the device seeder's implementation, but no observed run or causal experiment. No target code was executed. See the boundary and Source register.

The shipped entry path takes an installing agent through acquiring and building those siblings, seeding local configuration, then handing GUI launch to the user. The inspected seeder discovers repository manifests, merges existing locations and finds executable paths. It writes only with --write; otherwise it reports proposed output. Existing parsed locations win over rediscovery, so this is bootstrap configuration rather than automatic stale-path repair. Writes occur sequentially without an all-or-nothing transaction. A later invocation rereads existing configuration, providing a narrow implemented reuse path; host/engine consumption remains documentary. See RT-RTE-device-seeding and its memory annotation. Its limited text parser also prevents treating arbitrary imported YAML as losslessly preserved, as shown by the source-derived, unexecuted counterexample in EPI-OBJ-registry-paths.

Once launched, the declared engine mounts repositories into one in-memory graph. Durable knowledge remains in repository files; indexing derives the access structure rather than making the graph the durable store. Markdown prose is interpreted by readers, while fields, references and type constraints have symbolic consumers. Indexed media remain opaque in this evidence. Later clients can request files or graph content, and subscriptions stream changes, but exact subscription targeting is unspecified. These are declared reuse mechanisms, not observed improvements in decisions: MEM-OBJ-knowledge-files, MEM-OBJ-type-definitions, MEM-OBJ-live-graph and MEM-RTE-graph-reuse.

From session context to a governed action

The host supplies projections and supervises the engine; an optional harness supplies model decisions through an adapter. An agent-profile selects visible tools, hooks, skills, injects and native-tool allowance. Triggered skills, always-on graph-walk injects and computed session-start injection are distinct context mechanisms. The source establishes their delivery intent, not the complete selected payload, token budget, actual model receipt or obedience. Unchanged shipped guidance is not accumulated memory, and no changed-guidance-to-later-invocation chain is established. The inference model's exact version, parameter mutability and endpoint behavior remain uninspected. See RT-OBJ-agent-profile, RT-BAP-profile-guidance, its memory annotation and RT-CMP-agent-harness.

The representative workflow is a read-before-overwrite action, chosen because its proposal fields, veto example and acknowledgment are documented together. The foreground agent proposes a path and replacement content; adapter and daemon carry that action to mediators; an admitted plugin acts through the engine seam; events are declared appended to a session ledger. A mediator can deny an unread overwrite. Reading and retrying is a possible agent continuation, not an implemented recovery state machine established here. The tool example's wrote acknowledgment is only a sketch and independently proves no write. The harness owns subsequent checking and the final user response, neither of which is supplied. This is execution delegation to a tool, not an established multi-agent workflow. See RT-RTE-mediated-action and the workflow account.

Two conditions materially narrow governance. First, fail-closed loading of floor policies coexists with fail-open mediation when the daemon is unreachable. Second, host callbacks, external file edits, direct engine clients and installed executable code cannot be assumed to pass through the MCP read guard. Installed extensions are explicitly unsandboxed and run with user privileges. Type-query discovery and SDK compatibility refusal concern executable eligibility, not safety certification or universal containment: RT-CLM-mediation-floors, RT-RTE-host-edit, RT-RTE-extension-load and RT-CLM-user-privileges.

What the checks and retained results establish

For a scenario requiring editable, reusable typed content and procedural checks on mediated actions, Arsumbris supplies a coherent declared architecture. Its checks answer different questions: type diagnostics assess structure and references; read guards assess action history; loader checks assess interface eligibility. Diagnostics expressly do not block saves or reads. None of these checks licenses the empirical truth of prose or constitutes an evidence-consuming acceptance decision for a knowledge claim. For a scenario requiring warranted research conclusions, reliable task completion or comprehensive withdrawal of invalid content, this boundary cannot establish the necessary routes. See EPI-RTE-type-diagnostics, RT-CLM-diagnostic-saves and MEM-CLM-maintenance-limits.

The strongest supported contribution toward future reuse is editable durable content with a declared later-consumer path, alongside implemented configuration rereading. Replacing an existing file supports revision, not demonstrated improvement. Raw event recording is likewise not experience distillation: no automatic ledger-to-durable-guidance-to-later-consumer chain is established. Stamped metadata has no exposed authenticity check, and the separately named ingestion package supplies only a conversion-purpose declaration. Thus semantic acquisition beyond indexing, accepted novel claims and trace-fed learning remain unresolved, not disproved across the platform. See MEM-OBJ-session-ledger, EPI-RTE-hook-lineage, EPI-OBJ-write-metadata and EPI-RTE-source-ingest.

The accepted records do not establish a formulated theory route meeting conditions 1–4: localized theory content, decisions depending on that content, content-directed criticism, and retention of criticism's result into another round each remain unestablished. This does not convert opaque model processing into evidence of absence. Separately, criticism improving future-action capacity is unestablished; reflection's self-representation and two causal directions are unestablished; autonomous theory building is unestablished. Self-improvement at this entry-artifact boundary is also unestablished: extension loading and context delivery do not show objective-bearing evidence determining an organizational update that subsequently changes operation. These are independent unresolved properties, not a single grade. See the epistemic bounded conclusion, RT-BAP-profile-guidance and RT-RTE-extension-load.

Frozen sibling implementations could change the architectural assessment by exposing actual selectors, admission rules, knowledge evaluators and later consumers. Linked task traces could establish exercised rejection, recovery, writes and delivery. A retained criticism/update/consumer chain would support a stronger learning or self-change assessment; successful improvement and causal attribution would additionally need objective-linked outcomes and an appropriate comparison. Reported release testing and component hashes supply none of those substitutes. The report verification accepts the bounded account; the profile verification retains one comparison-granularity limit below. No source rereads were needed for this synthesis.

Limitations

Limitation	Affected SRC IDs and record citations	Inspected boundary	Conclusion prevented	Evidence that would resolve it
Downstream implementations and substantive knowledge instructions are outside the frozen artifact.	SRC-1; MEM-RTE-graph-reuse, RT-RTE-mediated-action, EPI-RTE-source-ingest	Complete entry repository; sibling responsibilities documented only.	Implemented whole-platform reuse, semantic production, acceptance and governance cannot be established or ruled out.	A new frozen boundary containing the relevant siblings, their callers, instructions and evaluators.
No observed execution or causal comparison is supplied.	SRC-1; RT-RTE-device-seeding, RT-RTE-mediated-action, RT-RTE-extension-load	Source inspection and declared/reported operation only.	Successful installation, action effects, final delivery, benefit and causal self-improvement remain unestablished.	Inspectable execution artifacts linked to the routes; objective-linked interventions and comparisons for causal claims.
Exact model-visible context and accumulated guidance reuse are unresolved.	SRC-1; RT-BAP-profile-guidance, RT-CMP-agent-harness	Profile/adapter descriptions, not selector or harness implementations.	Complete selection, budget, context-isolation, reliable compliance and changed-guidance memory conclusions cannot be drawn.	Selector and materialization code, delivered payloads, later invocation lineage and relevant harness contracts.
Ledger, metadata, local host state and withdrawal have unresolved persistence or consumer paths.	SRC-1; MEM-OBJ-session-ledger, MEM-OBJ-view-store, EPI-OBJ-write-metadata, MEM-CLM-maintenance-limits	Documentary event, stamping, state and maintenance interfaces.	Durable replay, authenticity, trace-fed updates, complete state inventory and withdrawal propagation cannot be inferred.	Concrete schemas, storage/reload and transformation implementations, invalidation consumers and linked later-use evidence.
The profile combines distinct configuration consumers in one routing-authority unit.	SRC-1; RT-OBJ-device-config, RT-RTE-device-seeding, RT-RTE-host-edit; profile verification limit	Accepted records distinguish repository resolution, tool/template selection and projection-instance arrangement; profile groups their claimed authority.	Independently addressable per-part/per-consumer authority comparisons cannot be extracted from that unit. The shared claimed routing finding and separate wired seeder-merge finding remain supported.	A profile split into those recorded consumption paths, preserving their separate bases and unresolved alternatives.