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.
Tvijo MCP 0.2.0 adds a focused use case: send a diff for a second review before merging it, with an explicit spending limit and a cost receipt. The code is merged and tagged; npm publication is still pending. This is an update about prepared functionality, not a claim that 0.2.0 is already available through the public package channel.
Observed 28 October 2026, 15:47 UTC. English and Russian versions describe the same release boundary. Technology map Β· Agent comparison Β· Knowledge search
What you can use today, and what is prepared
| Channel or capability | Verified state |
|---|---|
| Public npm package | Latest is 0.1.1; the explicit 0.2.0 metadata endpoint returns 404. |
| Open MCP Registry, com.tvijo/dshrem | Active latest version is 0.1.1. |
| MCP 0.2.0 source | PR #3 merged; tag mcp-v0.2.0 identifies the prepared source. |
| Publication workflow | Run 40713641341 is waiting for the protected publication environment. A tag and passing CI are not publication receipts. |
| GitHub MCP catalogue | An onboarding request was submitted; the catalogue lookup still returns no matching server. Submission is not acceptance. |
| Local snapshot tools, PR #2 | Merged on 28 October; absent from the published 0.1.1 package. They are optional, not part of the default seven-tool list in 0.2.0. |
| Editor integration and delegation | Separate future phases with their own acceptance gates. |
Primary checks: npm latest, open registry, merged implementation, publish run, snapshot PR. These links can change after this dated observation.
What tvijo_review reviews
A compatible MCP client supplies a unified diff, its base/head commit identifiers, an optional review focus and a positive max_usd. The prepared tool returns findings with a file, line, severity, issue and supporting evidence. It covers only the supplied diff: it does not fetch the repository, execute tests, edit files or merge a pull request. An editor window is not required for this cloud review path.
The possible outcomes are findings, no_findings_in_reviewed_scope and inconclusive. No findings in that scope is not a safety certificate or permission to merge. Findings outside the supplied diff are filtered from the structured result. Invalid or inconsistent model output becomes inconclusive.
The input diff is limited to 20,000 characters and optional focus to 500. Larger input is rejected before a paid request. The receipt identifies the task and observed spending; unavailable spending stays null, not zero. Submitted diff content is sent to the Tvijo service as model context. Use only material you are authorized to submit.
The practical offer is a separately metered second opinion for an agent's proposed change. Whether it improves acceptance or saves money needs a measured pilot. A real paid example and its redacted receipt are still pending in the tagged README; no synthetic receipt is presented as a customer run.
Where the spending boundary ends
The prepared review uses the existing paid task path: one task creation and one chat request on the normal path, with no automatic retry of a paid operation. Timeout recovery reads status instead of repeating payment-bearing work.
max_usd is mandatory. The configurable local ceiling defaults to$1 per call and has a hard ceiling of $5; account credit and server limits also apply. These limits constrain Tvijo's call, not the calling agent's other model requests or shell actions. Several separately initiated reviews can spend more in total. A process-wide aggregate session budget is not established by the per-call cap.
The new X-Tvijo-Client tag carries bounded channel/host/offer metadata. It is self-reported attribution, not authentication, and does not prove that gateway funnel reporting has already been deployed. Do not infer installs, successful customer onboarding or purchases from the tag alone.
Why editor delegation is a separate phase
MCP exposes Tvijo tools to a client. It does not automatically make Tvijo an executor of every other MCP server. The released consultation tool and the editor's local-agent bridge must remain distinguishable.
The proposed extension integration would register an MCP provider and reuse an appropriate runtime/sign-in path after platform tests. VS Code has an extension API for MCP providers; the existence of that API is not proof that Tvijo's integration has shipped. For Tvijo Code, the plan includes connecting a terminal MCP client such as Claude Code or Codex. A running server needs a client to invoke its tools.
Delegation into an editor window was deferred. Its acceptance must cover duplicate execution, the wrong project/account, recovery races, incompatible old extensions, cancellation and an immutable spending ceiling. Until those checks pass, neither safe editor delegation nor its budget guarantee is advertised as available. Renaming the editor bridge from tvijo_ask to tvijo_delegate is a release requirement in the proposal, not a published rename. MCP tvijo_ask retains its consultation role.
What was actually checked
| Evidence | Scope |
|---|---|
| GitHub CI | Node 22 and 24 jobs completed successfully, including tests, package dry-run and registry-schema validation. |
| Acceptance report | Reports 95/95 tests, publisher CLI validation and seven default tools from a clean package install. These are the accepting developer's checks, not a new production review performed for this article. |
| Earlier implementation report | Recorded 94 tests and a blocked publisher CLI check. Later acceptance reports an added correction and successful validation; the earlier report is retained as history. |
| Client recipes | README documents VS Code, Claude Code, Codex and Cursor. Linux stdio recipe checks do not establish native Windows or complete GUI-client testing. |
| Production receipt and payment | A real review receipt and an end-to-end settled public top-up are still acceptance gates. A created order is not a settled payment. |
The corrected validation preserves useful findings while rejecting the model's own approval stamp. An arithmetic expression inside a finding must not cause an otherwise paid review to disappear. Static tests and CI support that code change; commercial readiness still requires the production gates above. CI run, acceptance PR.
GitHub discovery and next steps
GitHub's discussion#1257 describes initial onboarding as manual curation, followed by version synchronization from the open registry. Tvijo's submitted request describes the currently published 0.1.1. No acceptance date or guaranteed listing is claimed.
The SPEC-237 proposal is still an open documentation PR. The next gates are publication approval and independent package/registry verification; a real redacted review receipt; editor provider integration; safe delegation; gateway attribution; and a verified public payment through credit settlement. These are separate outcomes, not one βdoneβ flag.
Discuss a scoped review pilot. This is an inquiry, not an immediate charge. Machine-readable status snapshot.
π Article connections
Open the full knowledge graph βSame topic
- Tvijo technology map: products, evidence and roadmap
- 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 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.
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.
π¬ Comments