Nimbalyst Wiki synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From coding sessions to project memory
Nimbalyst Wiki is a prompt-directed project knowledge system, not an enclosing agent runtime. It asks a coding agent to turn session decisions, reasons and open questions into subject pages, then read relevant pages before later answers or code changes. This account covers the released plugin repository at revision 97c1ba1594d6e3d466a23d33c1bda084026b01b8, with an analysis cutoff of 2026-11-08. The evidence is doc-grounded: shipped instructions and file contracts, executable integration declarations, and attributed product claims. No execution trace or causal experiment is registered. The boundary is complete artifact, partial loop because the CLI, persistence library, transcript reader, browser, agent hosts and hosted team server are outside the frozen source (boundary and Source register).
The runtime map is consequently split. Claude Code and Codex load plugin instructions and declared MCP connections; local and session servers invoke @nimbalyst/cli@0.12.4, while the optional team add-on connects to a hosted endpoint. Claude Code also has SessionStart and Stop hook declarations. These connections are wired as configuration, not verified execution. The foreground host agent interprets the skills, chooses next steps and proposes edits; external tools perform persistence and retrieval. Loading connect or update is a procedural handoff within that agent, not delegation to an independent model worker. The plugin neither specifies a fixed model nor establishes provider training behavior (RT-RTE-host-entry, RT-CMP-host-model).
Choose one destination, then capture meaning
The representative workflow is capture into an existing local wiki. It is an ordered shipped procedure, not an inferred combination of capabilities. The agent first prefers desktop tools if present, otherwise resolves one local or team destination. A migrated local copy is ordinarily an archive, not a second current write target; exceptional local use requires an explicit user choice. The one-destination rule is agent policy, not demonstrated isolation. CLI, ordinary file editors and other clients remain alternate routes (workflow account, MEM-RTE-cli-file).
Before writing, the agent reads Home and the retained writing guide. It supplies changed files to pagesForFiles, reads covering pages and judges whether their meaning changed. It updates affected prose or marks it outdated when it cannot explain the new meaning. Staleness is a maintenance signal, not proof that this session caused the divergence. The agent then selects decisions, alternatives, reasons, questions and useful context, excludes routine progress, searches for existing coverage and extends the owning page rather than duplicating it. “Nothing to record” is a valid endpoint (RT-RTE-session-capture).
People supply the decisions; the model chooses their wording and placement. Ordinary additions apply directly without per-content approval. Approval is separately required for such changes as rewriting another person's text, replacing an existing customized guide, and destructive schema changes. Starting capture is not approval of each synthesized sentence. Available session tools supply citation Markdown to paste unchanged, but missing quotes permit prose attribution rather than rejecting all writes. Codex lacks the documented end-of-session hook reminder and Claude-transcript citation path; explicit capture remains available (RT-RTE-wiki-setup, RT-CLM-session-isolation, EPI-RTE-citation-check, RT-RTE-host-entry).
Retain content without confusing its controls with truth
The local contract stores ordinary files. Natural-language prose carries decisions and reasons; symbolic IDs, fields, marks and coverage globs support location, filtering and typed operations. Stable page IDs are distinct from paths. CSV table storage, editor payloads and sidecars, recovery records and destination metadata have separate consumers. A valid ID or schema does not validate a page's assertions, and a readable display does not establish what an editor payload delivers to a later model (RT-OBJ-wiki-pages, MEM-OBJ-page-metadata, MEM-OBJ-table-types, MEM-OBJ-editor-pages).
Checks serve different purposes. Citation comparison concerns sentence support and attribution, not the truth of an external claim. Setup read-back concerns presence and usability. wikiDoctor diagnoses links, coverage and freshness without changing content; the agent then repairs, merges or marks outdated. Formula examples separately specify conditional arithmetic consequences and assertion predicates, but neither inspected evaluator code nor a downstream acceptance gate is supplied. Retention and operational use therefore do not establish evidence-consuming acceptance of a newly produced ampliative claim. Nor does the lack of a concrete output prove that generated prose is novel conjecture rather than reshaping (EPI-RTE-citation-check, EPI-RTE-doctor-check, EPI-RTE-doctor-repair, EPI-RTE-sheet-derive, EPI-RTE-sheet-check).
Recovery is prescribed rather than demonstrated. A stale-version refusal returns the agent to reread and reapply the change; ambiguous writes require read-back rather than blind repetition. Link checks and doctor checks apply under their stated conditions, not to every prose edit. Local trash and team history are different recovery mechanisms, not a general multi-page rollback. Version protection concerns cooperating writers: CLI checking is optional and ordinary editors remain possible. Delivery is a final list of updated or outdated page links, or the empty-result response; later recall is a separate invocation, not a capture completion condition (workflow account, MEM-RTE-wiki-maintenance, MEM-RTE-cli-file).
Bring selected material into later work
Recall requests marks by keywords, kind or person, searches the area and uses coverage for intended code changes. The agent reads a few full pages—about five at most as advice—because reasons surround the marks. It reports attribution and freshness and separates inference from wiki content. This is foreground pull, including proactive recall before coding; it is not an enforced token budget or host compaction mechanism. Unloaded pages remain reachable through tools. Team and desktop tools do not provide the same local marks/coverage interfaces (RT-RTE-wiki-recall).
SessionStart separately claims unsolicited local coverage context, but its selector, delivered parts and budget are external. Stop requests capture rather than itself creating memory. Desktop components afford another distinct route: human-clicked retained prompts can start an agent session, and templates can populate new pages. The transferred context, isolation and return behavior are uninspected (RT-RTE-host-entry, MEM-RTE-component-consumption).
The retained guide has stronger intended authority than ordinary reference prose: writers are told to follow it even where it overrides base writing advice. Ordinary page instructions are simultaneously declared data. This distinction identifies a consequential policy surface; it establishes neither enforced injection resistance nor an observed bypass. Decisions request reasons, but hard-won lessons may omit incident narrative and legacy migration keeps selected history. Universal retention of every originating criticism cannot be inferred (RT-BAP-guide-writing, MEM-OBJ-writing-guide, MEM-RTE-legacy-migration).
Assessment at the supported comparison level
The strongest learning-related contribution is a prescribed trace-fed memory path: session conversation feeds agent-admitted durable page prose intended for later agent context. This supports claimed trace learning in the memory comparison's write-route sense, with session-log origin, not demonstrated improved capacity. The unresolved question is whether faithful capture reaches later work and changes its decisions beneficially. Retained writing policy also has an identified future consumer, but no supplied content-directed criticism and next-round outcome establishes a theory-building improvement loop. The records do not establish exercised reflection, autonomous theory building or self-improvement; these remain independent questions, not consequences of the memory classification (RT-RTE-session-capture, RT-RTE-wiki-recall, RT-BAP-guide-writing, profile verification).
For the selected coding-session scenario, the design connects changed code, attributed decisions, explicit outdated status and later whole-page recall. It does not demonstrate that reasons prevent unwanted reversals, despite that stated purpose. A source-linked session/page pair would change the fidelity assessment; a later invocation tracing selected reasons into an actual decision would change the activation assessment. A controlled comparison would be needed to attribute improved outcomes to those reasons. Inspected tool implementations and route-specific rejection traces would separately change judgments about persistence, isolation and conflict enforcement. Neither release updates nor team migration supplies that improvement evidence (EPI-CLM-reason-reliance, RT-RTE-plugin-update, RT-RTE-team-transition, report verification).
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| External implementations and host behavior | SRC-1; RT-RTE-host-entry, RT-CLM-session-isolation, MEM-RTE-cli-file | Plugin manifests, hooks, skills and file contracts; no CLI, host or service code | No verified end-to-end dispatch, transcript isolation, locking, universal destination exclusivity or access enforcement | Frozen implementations and route-specific execution evidence, including alternate writers and clients |
| No observed capture-to-later-action chain or comparison | SRC-1; RT-RTE-session-capture, RT-RTE-wiki-recall, EPI-CLM-reason-reliance | Prescribed production and recall routes, with no registered run or experiment | No observed fidelity, activation, improved capacity or causal prevention of reversals; claimed trace-fed memory remains supported | Linked session, edits, retrieval and later decision trace; controlled outcome comparison for benefit attribution |
| Policy priority and reason retention are scoped | SRC-1; RT-BAP-guide-writing, MEM-OBJ-writing-guide, MEM-RTE-legacy-migration | Guide override, page-as-data rule and selective migration/history instructions | No enforced guide/data separation or universal preservation of originating evidence and criticism | Consumer enforcement evidence and before/after artifacts tracing retained reasons to later readers |
| Incomplete hook, editor and context-delivery account | SRC-1; RT-RTE-host-entry, MEM-OBJ-editor-pages, MEM-RTE-component-consumption | Integration declarations and selected payload/consumer contracts, not host renderers or compaction | No complete push-selector classification, lossless model-visible payload claim, enforced context budget or separate-session isolation guarantee | Selector and renderer implementations plus exact delivered-context traces; host context-control evidence |
| Checks have bounded licenses; novel-claim acceptance is not established | SRC-1; EPI-RTE-citation-check, EPI-RTE-doctor-check, EPI-RTE-sheet-check | Attribution, structural/freshness diagnostics and formula/assertion doctrine | No universal semantic oracle, system-complete evaluation absence, or established acceptance and subsequent integration of a produced ampliative claim | Concrete candidates, evidence-consuming decisions with criterion/use/scope, and consequential consumers; inspected evaluator alternatives |
| Migration verification is an external report, not independent fidelity evidence | SRC-1; RT-RTE-team-transition, MEM-OBJ-destination-metadata | Dry-run/confirmation, batch-copy and archive-routing instructions | No proof that every local item became team-visible, links survived or migration is resumable; skipped items can remain local | Source/destination comparison including skipped items and interruption/recovery traces |
| Partial curation inventory omits documented forgetting by purge | SRC-1; MEM-OBJ-recovery-records, MEM-RTE-wiki-maintenance; profile verification limit | Explicit irreversible purge of already trashed recovery content in the file contract | Do not treat the profile's operations as exhaustive or infer that decay lacks a documented route; its existing positive findings remain supported | A classification of the recorded purge mechanism as claimed forgetting in a revised comparison inventory |