Tolaria synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From retained knowledge to an agent assignment
Tolaria is a Markdown knowledge-base application that supplies context and file operations to selected external agents or direct model providers. It is not the enclosing runtime of those agents. This account is code-grounded at commit 48f2a70844c1e8005a7aa5f1499c66f7678360de of https://github.com/refactoringhq/tolaria, with an applicability cutoff of 2026-10-14. SRC-1 supplies implementation, design and attributed use claims, but no target execution, observed-run artifacts or causal experiments. The complete repository was frozen; inspection covers a partial loop rather than the whole system. These distinctions bound every operational conclusion below (boundary, Source register).
The strongest supported contribution is an implemented route from editable retained content to later agent work. Vault Markdown persists as files; workspace conversations have a native JSON file, a localStorage cache and frontend state. Later sends reconstruct selected conversation text, while explicit references and MCP tools deliver retained notes. This establishes retention and reuse wiring, not learning benefit or authenticated knowledge provenance (MEM-OBJ-vault-notes, MEM-OBJ-workspace-sessions, MEM-RTE-transcript-replay, MEM-RTE-mcp-read).
A user starts work by choosing a target and submitting a prompt with optional references. Tolaria checks that a vault and ready target exist, excludes concurrent sends, assembles instructions and history, and dispatches the request. On the inspected Codex path it configures the bundled MCP service with active vault roots and launches codex exec --json. The external agent/model owns subsequent tool choices and stopping decisions; Tolaria supplies capabilities, executes its local services and normalizes progress and results. The external CLI owns its authentication and runtime behavior (RT-RTE-workspace-send, RT-CMP-cli-model).
The next invocation does not receive everything the application retained. Code supplies completed user and assistant text, omitting stored reasoning, tool actions and reference bodies. It keeps a newest-fitting suffix under a 100,000-token estimate based on characters divided by four. That estimate excludes the latest prompt, system instructions and references, so it is not a combined request bound. Older turns remain stored but are not automatically reimported by this builder. This is automatic history supply, not generation of a durable summary (RT-RTE-workspace-send, MEM-RTE-transcript-replay).
Explicit references instead request named retained material. Already supplied content bypasses disk hydration. Bodies over 12,000 characters are reduced to an 8,000-character head and 2,000-character tail. These are value copies with omissions, not semantic summaries. MCP offers another requested read path: parsed note bodies, substring-matched snippets, and orientation containing recent-note metadata and guidance. Search defaults to ten results; orientation selects twenty recent notes by modification time. Neither selector establishes semantic relevance or truth. A request to reload omitted material is not proof of reload, particularly on the direct-provider create-only tool branch (MEM-RTE-reference-context, MEM-RTE-mcp-read, EPI-RTE-note-search, RT-RTE-provider-create).
From assignment to a changed note
The representative workflow is revision of an existing note through desktop Codex and MCP. It is connected by shipped configuration, but the agent is not required to read, edit and verify in that order. If it requests get_note, it receives parsed frontmatter, trimmed body and modification time. It can then submit complete replacement text and optionally the read-derived expectedMtime. Tolaria does not restore frontmatter omitted from that replacement (workflow account, RT-RTE-mcp-write).
When supplied, the timestamp can reject a stale replacement and return an instruction to re-read and retry. Omission skips the check. Equality is a sampled stat comparison, not atomic compare-and-swap; another edit can intervene before the write. An admitted update writes a temporary file and renames it over the note. Cleanup handles temporary-file errors, not workflow rollback. The caller owns any retry; the service does not automatically merge or repair. Create and append are separate operations, and append does not inherit this optional conflict check (EPI-RTE-mtime-check, RT-RTE-mcp-write, workflow account).
Permission labels do not collapse these effect boundaries. Codex Safe selects a native read-only sandbox, yet the separate Node MCP service still exposes note writes. Prompt prohibitions are interpreted instructions, not OS isolation. The inspected Pi configuration also does not provide an equivalent mode-dependent sandbox; it can inject an unversioned MCP adapter extension without disabling native tools. This supports withholding an application-wide no-shell or read-only guarantee, not claiming that an external tool actually ran (RT-RTE-workspace-send, RT-RTE-mcp-write, RT-RTE-pi-extension, RT-BAP-prompt-policy).
Direct-model requests have their own admission path. With an active vault, OpenAI kind or a catalog-declared tools capability can cause the OpenAI-compatible payload to advertise create_note; the Anthropic payload is different. Returned supported calls execute locally without an inspected per-content confirmation branch. Calls are sequential, earlier writes survive a later failure, and this path does not send tool results into another provider round. Consequently, the documentation's universal chat-only/no-write-tools claim is too broad. This does not establish all-provider write support or direct-model shell access (RT-RTE-provider-create, RT-CLM-direct-chat).
Delivery, later reliance and assessment
File effects, UI actions and conversation completion are separate delivery boundaries. Successful MCP replacement emits a vault-change action; opening the note requires another call. Stream completion can finalize a response containing a failed tool action, and pending-action finalization is not independent outcome verification. Regenerate removes a conversation suffix, not previous file changes. Stop does not undo writes; direct-provider requests receive no abort signal, so suppressing UI callbacks need not stop a potential write (workflow account, RT-RTE-workspace-send, RT-RTE-provider-create).
Later readers can consume retained candidates without an intervening factual acceptance decision in these executors. Path, active-root, exclusive-create and timestamp checks license file operations, not truth. Read parsing and search excerpting can change visible context without adding warrant. The model edge remains semantically indeterminate without candidate instances and premise comparisons: these records establish neither warranted derivation nor accepted ampliative conjecture. Types are declared navigation aids, and the organized flag controls workflow state rather than certifying note content (EPI-RTE-note-read, EPI-RTE-note-search, EPI-RTE-mtime-check, EPI-CLM-types-lenses, EPI-RTE-organized-toggle, epistemic conclusion).
Mutable guidance is a consequential instruction source, not merely ordinary note evidence. Code can refresh recognized managed templates while preserving unrecognized custom content; MCP can return root and legacy guidance directly. This establishes an available policy-change and later-delivery path, not evidence-responsive criticism or compliance. Attach/clone separately expands later MCP root scope without assessing imported claims. Permanent native deletion removes a file at its path without revoking copied transcript or reference content (RT-RTE-guidance-sync, MEM-BAP-retained-context, MEM-RTE-mcp-read, EPI-RTE-vault-register, MEM-RTE-note-withdrawal).
For a scenario requiring editable knowledge to reach an agent and permit scoped note changes, these paths establish implemented support. For a scenario requiring every revision to be factually accepted, every consumer to receive complete evidence, or every client to share a consistent recoverable state, they do not establish the required guarantees. The verified retention inventory is also broader than desktop Markdown and transcripts: collaboration has Yjs binary state plus a Markdown/revision checkpoint and reload consumer, but its complete lifecycle remains unverified (report verification limits).
The unresolved learning question is whether retained changes improve future capacity, not whether files can change. Raw-turn replay and managed template replacement do not establish automatic trace-derived lessons. Theory-builder conditions 1–4 are not separately established by the runtime report; guidance editability must not be substituted for content-dependent consumption, criticism and iteration. Improved capacity from criticism of a consumed theory, reflection, autonomous theory building and exercised self-improvement likewise remain independently unestablished in the assessed paths, not evidenced absent throughout Tolaria. External model state is opaque and model IDs do not pin immutable weights. No comparison supports a benefit or component-effect attribution (RT-RTE-guidance-sync, RT-CMP-cli-model, RT-CMP-api-model, epistemic conclusion, profile verification).
A linked execution trace could establish which read–update–retry and delivery steps actually occurred. Candidate-level evidence and an acceptance criterion could establish a factual license beyond successful I/O. A retained criticism-to-guidance update with a later consumer could establish additional theory or adaptation links; independently specified objectives and subsequent operation causally dependent on the update would be needed for exercised self-improvement. Comparative outcome evidence would then be needed for successful improvement. Full collaborative lifecycle, adapter and editor tracing could change coverage and enforcement assessments without changing the positive desktop findings.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| Retention inventory is partial; collaboration includes additional Yjs state and Markdown/revision checkpoint reload. | SRC-1; MEM-OBJ-vault-notes, MEM-OBJ-workspace-sessions, report verification limits | Desktop note/transcript routes plus verifier inspection of collaboration checkpoint existence, not its complete lifecycle | Complete storage ontology, cross-client retention, checkpoint and recovery comparisons | End-to-end records for collaborative write, reconcile, failure and later-use paths, with recovery evidence |
| Admission and authority remain route-specific, including an additional collaboration admission check. | SRC-1; RT-RTE-workspace-send, RT-RTE-pi-extension, MEM-RTE-note-withdrawal, epistemic coverage, report verification limits | Selected adapters and local services; remaining adapters, editor confirmations, symbolic consumers and watcher/cache/Git paths untraced | Universal enforcement, complete human-control coverage, global revocation or system-wide absence of acceptance | Records tracing each material alternate admission/consumer and copied-content reconciliation path |
| Wiring is not observed operation or benefit. | SRC-1; boundary, RT-RTE-workspace-send, workflow account | Pinned implementation and design, no target execution or causal experiments | Successful workflow delivery, actual model compliance, reliability or causal effectiveness | Linked request/tool/file/UI traces and appropriately controlled outcome comparisons |
| Context budgets and reload are local, not complete-request guarantees. | SRC-1; RT-RTE-workspace-send, MEM-RTE-reference-context | History builder and reference hydration/omission; complete active-panel wiring not retraced | Combined token bound, complete evidence delivery, automatic restoration of omitted context or snapshot freshness | Full panel-to-dispatch and reload records, complete payload accounting and editor/disk provenance |
| Transcript retention and file updates lack established crash/concurrency guarantees. | SRC-1; MEM-RTE-transcript-replay, RT-RTE-mcp-write, EPI-RTE-mtime-check | Best-effort saves, whole-file transcript write, optional stat check and temp-file replacement | Crash durability, completed shutdown writes, cross-window conflict resolution, atomic stale-write exclusion or batch rollback | Persistence/recovery and concurrency records covering these failure windows |
| Source warrant and model content relations remain unresolved. | SRC-1; MEM-OBJ-vault-notes, EPI-RTE-note-read, EPI-RTE-note-search, RT-RTE-provider-create | Readers and write executors, without actual user propositions or provider candidate comparisons | Authenticated factual provenance, general semantic preservation, warranted derivation or accepted novel claims | Candidate/input/evidence instances and evidence-consuming acceptance decisions with criterion, use and scope |
| Guidance refresh and trace retention do not establish learning or self-improvement. | SRC-1; RT-RTE-guidance-sync, MEM-RTE-transcript-replay, epistemic conclusion, profile verification | Managed template replacement, raw replay and generic writes; other maintenance unresolved | Automatic trace-learning classification, theory criticism/iteration, reflection, autonomous theory building, improved future capacity or exercised self-improvement | Separate route evidence for each property; trace-fed durable updates and consumers, criticism and next rounds, self-representation links, objective-directed updates and causally dependent later operation |
| External runtimes and guidance precedence are unresolved. | SRC-1; RT-CMP-cli-model, RT-CMP-api-model, RT-RTE-pi-extension, MEM-RTE-mcp-read | Tolaria invocation/configuration and root/legacy guidance delivery, excluding external implementations | Immutable model identity, internal weight-change conclusions, loaded extension behavior or definitive guidance precedence/compliance | Pinned runtime/model contracts and evidence of actual configuration, guidance selection and consumption |