Tvijo technology map: products, evidence and roadmap
How Workspace, DSHREM, AI OS, knowledge and learning fit together. Verified capabilities and remaining product gates, dated 28 October 2026.
Tvijo connects a workspace, agent execution, spending controls and evidence about the result. The useful unit is a task with a scope, a spending ceiling, an output and acceptance checks. A model name alone does not provide those controls.
Observed 28 October 2026 (UTC). This is an automated editorial report from repository inspection, public HTTP checks and bounded execution receipts. It is not independent certification or a claim that every component is ready for customers.
Open Workspace · Search the knowledge map · Learning exercises · Evaluation method
MCP update — 28 October 2026, 15:47 UTC
MCP 0.2.0 is merged and tagged, but not yet published to npm. It adds tvijo_review: a review of a supplied diff with findings and a spending receipt, without repository access, test execution or editor delegation. npm and the open registry still report 0.1.1; publication run 40713641341 is waiting. GitHub catalogue onboarding was requested, not accepted. PR #2 with optional local snapshots is now merged. The real paid-review receipt, editor integration, safe delegation and settled-payment checks remain separate gates. Full update and sources.
One system, distinct responsibilities
| Component | Responsibility | Evidence and boundary |
|---|---|---|
| Tvijo Workspace | A project entry point, connections, runs and evidence navigation | Public shell reachable. Complete authenticated project coverage has not been audited in this report. |
| Tvijo VS Code extension and TvijoCode | Editor entry points for work on files and commands | Published channels have separate versions; see the download page. An installed editor does not establish successful task completion. |
| DSHREM CLI | Bounded task execution through approved provider routes | Two execution receipts in this audit; spending and acceptance are separate. |
| Tvijo MCP | Makes Tvijo callable by an MCP client | Published 0.1.1 provides consultation. Merged/tagged 0.2.0 adds supplied-diff review and awaits npm publication. Neither is general local-disk execution; see the dated MCP update above. |
| AI OS | Task policies, route admission, metering and evidence services | Existing services reused; no new model router created for this map. |
| Wiki / KB / Learn | Publish approved explanations, sources and exercises | Native Django Wiki with English and Russian articles. Lessons here do not yet provide an LMS progress tracker or certificate. |
| Knowledge retrieval | Retrieve bounded evidence with source and visibility metadata | Existing SQLite FTS5 EvidenceIndex reused for a local scoped index; public search covers approved cards and external URL metadata. This is not a full import of every authenticated Workspace. |
The website has several services: Wiki uses Django, Workspace uses Next.js, and the AI OS transport controls provider calls. FinEdg's portal-go/HTMX surface is another product surface. They should share contracts and receipts, not credentials or unrestricted database access.
A task moves through Workspace/editor/MCP → scoped task → admitted DSHREM route → meter → result → acceptance → approved publication. Private source documents stay behind their own authorization boundary. The public map is not a tenant authentication service.
Our methods, in plain language
| Name | What it does | Status in this map |
|---|---|---|
| ROYAGENTS / AGENTROY | Coordinates scoped agent work, limits and handoffs | Existing execution tools; bounded workers used for this audit. Not an always-running autonomous workforce. |
| AGENTJUDGE | Reviews versioned evidence; missing evidence can block acceptance | Review protocol. Advice from a model is not permission to spend or self-approve a change. |
| ALKNDA | Routes reusable knowledge into articles, learning, KB, news, docs and API surfaces | Publication framework; this map and its original exercises are concrete outputs. |
| ORBITA | Outcomes, Reuse, Bottlenecks, Investment, Traction and Amplification | Prioritization method. A registry entry does not prove customer traction. |
| HITDEP | Chooses a research/test method for uncertainty and checks the result | Documented method; use measurements before scaling a change. |
| LOBE | Reuses an existing component before building another | Existing metering and retrieval reused here. |
| DEMOCRA | Makes useful evidence public after privacy and publication checks | Public methods and aggregate receipts; no secrets, customer records or trading execution details. |
| RULEZ | Combines HITDEP, ALKNDA, ORBITA, LOBE and DEMOCRA | Working rules, not an additional software product. |
| RULEZ2 | A reminder for PROOF, DRIFT, SAFE, COST, OWNER and FUNNEL | References existing rules. It does not introduce a second competing policy set. |
What belongs where
| Destination | Include | Keep out / next gate |
|---|---|---|
| Internal project retrieval | Approved project documents, code evidence, decisions with access scope | Secrets and unrestricted cross-customer search; authorize the tenant on the server before retrieval. |
| Deep research queue | Unresolved questions, hypotheses, conflicting evidence, evaluation designs | Treat findings as proposals until checked; budget each investigation. |
| Public SEO / AEO / GEO | Original explanations, methods, product cases, dated aggregate measurements and source links | Customer details, private paths, invented ratings, unlicensed copied courses. |
| External reading map | Public links and taxonomy with attribution | A sitemap is not a full-text reuse licence. |
The ABAI Wiki source map contains 1,252 sitemap URLs allowed by its robots rules, observed 28 October: 630 docs, 527 learn, 48 journey, 41 CAIO and six other entries. Six navigation pages were sampled. Direct HTTPS worked; no proxy was used. No course body was copied. URL-derived labels in our search are not verified page titles. Inclusion does not imply partnership or endorsement.
Product cases and what they prove
- Hotido: a catalog and inquiry surface developed in the shared agent workflow. A dated deployment receipt and anonymous navigation/form checks demonstrate a working surface. A synthetic inquiry is not a sale or proof of customer demand.
- FinEdg: research and paper trading workflows, including portal-go/HTMX presentation. Paper evidence is labelled as such. This case makes no realized-profit or live-order claim.
- Tvijo: editor entry points, metered execution and this native Wiki. Delivery across several surfaces is visible; onboarding, settlement and independent quality evaluation still need dedicated acceptance.
Models: catalogue, route, probe, acceptance
| Model or route | Verified in this audit | Remaining boundary |
|---|---|---|
| ds41f | DeepInfra route; one completed bounded advisory task | Its narrative cost claim was wrong; use the ledger. |
| glm53f | DeepInfra route; one completed bounded advisory task | Its cost-per-accepted formula was wrong; corrected in the evaluation method. Requested glm53d was not an admitted alias and was not silently invented. |
| Gemini 4 Argon | OpenRouter/Google route answered one bounded request; $0.021426 provider-reported cost | Catalogue context is 1,048,576 tokens, not two million. Response reached the output limit; partial advice is not acceptance. |
| Codex Terra / Luna / Sol | Account CLI used for bounded implementation, classification and review | Requested model IDs and CLI receipts are retained. Served identity and per-task subscription cost are not inferred from the request. |
| Sakana Fugu Ultra V2 | Catalogue and local candidate route inspected | Not called: current uncapped-output reservation can exceed the $2 root task ceiling before input. |
| Sakana Fugu Max | Present in the provider catalogue | No admitted local route established by this audit. |
The three paid requests together cost $0.02443503856, with zero pending or held spend at reconciliation. Account CLI subscription allocation is unknown and excluded from that API total. These were different tasks, not a controlled model comparison. Prices quoted per million tokens are not prices per job. Catalogue availability does not mean every model is exposed in the editor.
Tvijo can differentiate through scope enforcement, cost receipts, evidence, private retrieval and human-readable acceptance. Whether that advantage outweighs another product's speed or quality remains an experimental question. Fugu is a model family; Tvijo is a workflow around models. A useful benchmark compares complete task outcomes under equal conditions.
Search visibility: what is actually known
The authorized Search Console query for the 28-day period 30 September–27 October 2026 returned 3 impressions and 0 clicks. URL inspection found the English home page as a duplicate with Google's canonical set to the root. Workspace, Wiki index and the comparison article were reported as unknown to Google. These four selected URLs cannot yield a site-wide indexing percentage.
We have not measured a reproducible Claude or Codex citation rate. A crawler's permission, an indexed page, a search impression and a citation in an AI answer are different observations. No article or robots setting guarantees a ranking, training inclusion or citation. Google's AI features guidance and the crawler descriptions from OpenAI and Anthropic describe those boundaries.
Roadmap with acceptance gates
| Phase | Deliverable | Gate to claim completion |
|---|---|---|
| 1: Evidence foundation | This public map, original learning route, model receipts and bounded local retrieval | EN/RU publication, working navigation, source/visibility tests and reconciled spend. |
| 2: Customer project retrieval and first task | Authenticated Workspace-to-DSHREM retrieval; clear install/login/task/report flow; attributable MCP traffic | Two tenant isolation tests; a new user's first completed task timed; channel events joined to receipts; settled payment tested separately. |
| 3: Controlled quality study | Same task corpus across admitted models and guarded execution presets | Frozen code/environment, repeated runs, independent acceptance, all costs, attempted/effected violations and uncertainty. |
| 4: Learning and discovery | Original tutorials, progress tracking, case studies and public measured results | Lesson completion evidence; rights review; searchable HTML; fresh Search Console sample and repeated AI citation panel. |
Only phase 1 is the current publication. Later phases are proposals with concrete gates, not shipped features. Start with one paid pilot: request scope and acceptance criteria. The form is an inquiry, not an immediate payment or a promise of delivery.
Model metadata sources: OpenRouter catalogue, Argon, Sakana models. See the public receipt summary for scope and limitations.
🔗 Article connections
Open the full knowledge graph ↗Same topic
- Tvijo MCP 0.2.0: diff review, release status and next gates
- Learn Tvijo: six exercises from task to evidence
- How to evaluate agent quality, cost and control
- Tvijo, DSHREM, Codex, Claude Code and Copilot: which fits your work?
- Connections, backlinks and the knowledge graph
- Models and routing: what the composer control changes
Next steps
Related reading
Tvijo MCP 0.2.0: diff review, release status and next gates
Prepared diff review with spending receipts; verified publication and GitHub listing status, plus the boundary between MCP consultation and editor delegation.
Workspace and agentsLearn Tvijo: six exercises from task to evidence
Original exercises for scoped tasks, budgets, source retrieval, review, publication and product measurement.
Workspace and agentsHow to evaluate agent quality, cost and control
A proposed repeatable protocol for comparable tasks, full spending, acceptance and attempted versus effected violations.
Workspace and agentsTvijo, DSHREM, Codex, Claude Code and Copilot: which fits your work?
Compare execution boundaries, workflow ratings and rollout status. Choose services for a company pilot and send an inquiry to the Tvijo team.
💬 Comments