IWE synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
A shared store, not an enclosing agent runtime
IWE supplies connected local documents and tools for people and external agents to maintain and reuse them. Its repository-shipped system includes the CLI, MCP server, language server, graph/filesystem layers, memory helpers and embedded capture/maintenance instructions. This account is code-grounded at commit 930c184b7cc8508d207e6ca7999ca6e415bc02ac, with applicability cutoff 2026-11-08. SRC-1 supplies implementation and doctrine/design evidence, not an observed run or causal experiment. The separate plugin, Mac app, external model hosts and actual user stores are outside the frozen evidence. See the boundary and evidence and Source register.
An external human or agent chooses the next operation. Rust code loads documents, selects context, transforms structure and writes files; IWE does not own model calls, tool scheduling, permissions or conversation compaction. Durable documents are distinct from their loaded graph, which may also contain unsaved editor text. CLI invocations reconstruct graph state; MCP retains a shared process-memory graph and watches files. The language server treats open buffers as working copies and reloads disk after close. These mechanisms support the declared shared-context purpose without establishing deployed integration or beneficial memory use. See MEM-OBJ-documents, RT-OBJ-graph, RT-RTE-editor-reload and EPI-CLM-shared-context.
From conversation to a durable candidate
The representative workflow is the shipped multi-session Distill procedure, not a demonstrated successful execution. A foreground agent first obtains the store's MEMORY policy, schemas, scope and recent feedback. For several sessions it is instructed to delegate one reader per session, supplying the session ID, relevant policy sections and knowledge filter. Readers stage candidates and stop before document admission; the foreground agent owns cross-session merging and user selection. IWE supplies this procedure and supporting commands, not the worker launcher or host isolation. See RT-RTE-session-capture and the workflow account.
Transcript reading is selective before the proposing model sees it. The digest filters records, clips tool observations and oversized lines, and advances raw-line coverage. The model is asked to propose quotable, durable facts rather than treat its own unconfirmed suggestions as decisions. Code then admits proposal YAML after shape and target checks, without authenticating quotations or user consent. Proposals persist outside the document graph, alongside progress and disposition bookkeeping. Thus candidate persistence is deliberately earlier than selected project-memory admission. See the memory annotation on RT-RTE-session-capture, MEM-OBJ-proposal and RT-OBJ-session-record.
Durable proposals can return to later inbox readers after interruption or host compaction. Grouped output is only a projection: it omits evidence and reduces bodies to their first line. The foreground agent must obtain per-session details to perform the requested merge preserving all contributing evidence. User keep/edit/skip decisions precede document writes under the procedure. Unattended work normally stages only, except for an item already named in the instruction. These are host obligations, not enforced credentials. See MEM-RTE-inbox-read, WF-RTE-inbox-selection and the report verification.
Admission, delivery and recovery have different checks
Selected candidates are written through strict CLI calls under the prescribed workflow. Schema reports can reject violations of configured predicates over parsed documents. This is a real pre-write veto on checked paths, not a factual truth check: unbound documents can yield clean reports, and a schema named learning does not certify learning. CLI validation is opt-in; inspected MCP create/update/query branches apply their own checks. Strict body replacement also differs from mutation mode's expected-count guards. See RT-RTE-cli-document, RT-RTE-mcp-document, EPI-RTE-schema-check and RT-BAP-schema-admission.
The prompt's claim that nothing reaches memory unselected is therefore a scoped capture protocol, not a universal store invariant. Ordinary CLI, MCP and editor routes do not require a selection token. Normalize has a separate unchecked persistence path; completion can also append area-hub links directly without the strict schema gate. None of this demonstrates an actual unselected factual capture. Post-tool schema/similarity notices arrive after writes and request correction rather than undoing them. Lexical similarity proposes a comparison; semantic sameness remains a host judgment. See RT-CLM-capture-exclusivity, MEM-RTE-graph-restructure, EPI-RTE-post-tool and the report-verification limit.
Delivery under Distill means selected files exist, contributing session outcomes are recorded, and the agent reports written keys, remaining inbox and backlog without committing. File writes, hub linking and session-record saving are not one transaction. A late save error does not undo earlier effects; recovery must reconcile documents and bookkeeping. Similarly, MCP reloads graph keys from disk after persistence errors rather than restoring all previous files. Progress is not proof of reading: adopt deliberately skips unread spans, while completion can sweep never-presented candidates into rejection strings. See RT-RTE-session-capture, its workflow annotation, MEM-RTE-session-adopt and MEM-OBJ-rejected.
Later context is a projection, not full recall
Once invoked, the session-start helper automatically selects metadata using policy slices, filters, sort and budgets. It emits titles/keys and selected fields, plus the separate At session start policy section. Empty injection groups suppress the entire output, including that section. Slice budgets do not bound policy/template/reminder text, and keeping the first line can exceed a slice's nominal budget. External host/plugin invocation and model delivery remain unobserved. See RT-RTE-session-index, its memory annotation, MEM-OBJ-policy-knobs and RT-BAP-session-policy.
Explicit retrieval instead fulfills a caller's request through keys, lexical search, filters and graph expansion. Seed predicates are not reapplied to every expanded document; emitted-key exclusion is not a universal metadata confidentiality guarantee. Ordinary body output omits arbitrary frontmatter, so stored session provenance need not reach that reader. MCP resources and review/refactor prompts provide distinct projections, not evidence of model review or acceptance. The design therefore combines automatic coarse discovery with requested detailed access. Its context controls are lossy windows, projections, instructed isolated readers and reloadable candidates—not an IWE-owned host-window eviction mechanism. See RT-RTE-document-retrieval, its memory annotation, MEM-RTE-mcp-context and EPI-RTE-review-prompts.
Feedback can shape future guidance; improvement is a separate question
The strongest supported trace-fed contribution already occurs before final admission: the host can extract conversation and tool experience into automatically admitted durable proposals that later inbox invocations expose as context. This affords a trace-learning pathway in the memory comparison's write-route sense. Staging and read-back are wired; semantic extraction is host-requested, not an implemented IWE model extractor or observed operation. Final user-selected documents and confirmed policy edits have a different admission mechanism. The profile verification supports this bounded positive without implying improved capacity or complete coverage of alternate writes. Its warrants are RT-RTE-session-capture, MEM-OBJ-proposal and MEM-RTE-inbox-read.
Rejection records also afford user-assisted revision of MEMORY guidance. The body supplies interpreted capture/curation instructions; symbolic settings directly configure later helper selection. Reflect requests policy-dependent merges, promotion, pruning and tombstones. Tombstones require consumer filters rather than imposing universal withdrawal. Feedback can thus reach durable behavior guidance, but rejection counts do not authenticate user judgments, and actual policy criticism or improved future action is not established. See MEM-OBJ-policy-body, MEM-OBJ-rejected and RT-RTE-policy-maintenance.
The runtime assesses the method-text pathway separately: theory-builder condition 1, localized content, is afforded; condition 2, decisions depending on that content, is afforded for the instructed host, with wired consumers of symbolic settings; condition 3, content-directed criticism naming the challenged claim, result and blame, is uninspected; condition 4, keeping criticism's result for a next round, is afforded through policy rereading but not observed as iterative criticism. Guidance has revisable sections and cross-session persistence. Learning as improved future capacity remains uninspected. Reflection is afforded for user-assisted policy revision, not observed in exercised two-way behavior. Reflective and autonomous theory building remain uninspected; the described procedure deliberately includes internal user selection. Exercised self-improvement is not established: no evidence-shaped policy change is traced through later operation causally depending on it. These are independent conclusions, not a system-wide negative. See RT-RTE-policy-maintenance.
For a scenario requiring resumable, user-directed capture and discoverable local notes, the inspected architecture supplies concrete supporting services and a coherent host procedure. For a scenario requiring certified factual memory, exclusive human-approved writes, transaction-wide delivery or demonstrated improvement, the same evidence does not establish those properties. Paired transcript/proposal/document instances and selection traces would resolve semantic and admission questions; host delivery traces would establish exercised recall; evidence-linked policy revisions and later dependent actions would change the improvement assessment. A universal approval or schema guarantee would additionally require controls covering alternate writers, not merely stronger prompt wording. See WF-RTE-inbox-selection, EPI-RTE-schema-check and RT-CLM-capture-exclusivity.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Completion hub append is not separately tabulated as an admission mechanism in the analyst inventories | SRC-1; RT-RTE-session-capture, its workflow annotation, memory write-side inventory, report-verification limit | Completion normalizes and directly writes area hubs without the strict CLI/schema gate | Exhaustive write-admission inventory; inference that every completion-induced change was schema-checked | A separately characterized hub-write admission path and complete inventory of completion effects; enforcement on that path if a universal gate is claimed |
| External host and deployment behavior is not observed | SRC-1; RT-RTE-session-capture, RT-RTE-session-index, RT-BAP-session-policy, boundary | Repository helpers and embedded instructions; separate plugin/app and host implementations excluded | Actual worker isolation, user authorization, model receipt/compliance, deployed safety or end-to-end reliability | Frozen host/plugin wiring, permissions and execution traces connecting helper output to model actions and user decisions |
| Actual content and original admissions are unknown | SRC-1; MEM-OBJ-documents, MEM-RTE-memory-enable, MEM-OBJ-proposal | Shipped policies, initialization and candidate-shape checks, not deployed stores | Exhaustive provenance/agency classification, true captured propositions, entailed or ampliative semantic transformation, evidence-consuming factual acceptance | Actual policy/schema/store and paired input/output/evidence instances with admission history and criterion application |
| Projections can omit evidence and provenance | SRC-1; MEM-RTE-inbox-read, WF-RTE-inbox-selection, RT-RTE-document-retrieval and its memory annotation | Lossy digest, grouped inbox, ordinary body retrieval and separate resource alternatives | Full-fidelity transcript delivery or automatic source-warrant preservation at downstream consumers | Consumer-visible artifacts tracing per-session evidence retrieval, merge fidelity and delivered provenance |
| Maintenance semantics and concurrency are not exhaustively assessed | SRC-1; MEM-RTE-graph-restructure, EPI-RTE-link-cleanup, RT-RTE-mcp-document | Selected handlers, structural cleanup and best-effort persistence; lower-level alternatives and watcher timing remain incomplete | Universal semantic preservation, all-tools guard coverage, transaction-wide rollback or multi-process conflict guarantees | Complete relevant transformation/consumer tracing and execution evidence for failure and concurrency cases |
| Feedback and later loading do not establish improved capacity or exercised self-improvement | SRC-1; MEM-OBJ-rejected, RT-RTE-policy-maintenance, profile verification | Requested rejection-to-policy pathway, persistent guidance and known settings consumers; no operational or causal study | Actual content-directed criticism, completed theory-building iteration, autonomous theory building, evidence-responsive exercised self-change or successful improvement | Evidence linking a challenged policy claim and outcome to revision, retention and later causally dependent operation against an independently stated objective; appropriate outcome comparisons for benefit |
No additional source files were read for this synthesis. The accepted member records and verification findings supply its evidence.