🇪🇺 · 🇫🇷 FR · 🇩🇪 DE · 🇬🇧 ENG · 🇪🇸 ES

🟪 TRINITY Agent

An independent, open-source coding assistant and model API relay for OVHcloud AI Endpoints, DeepSeek and OpenAI Codex, with conversation archives, tool approvals, task resumption, steering and a persistent follow-up queue.

This extension is the client for the self-hosted TRINITY Agent backend. Installing this VSIX alone does not install Python, the agent backend or an OVH account. Use it in the supplied Linux Docker environment, or connect VS Code Remote to that environment. It is not an official OVHcloud or OpenAI extension.

⬛ Preview

Edited demonstration screenshot: conversation titles and the mount path are fictional.

⬛ Install and configure

1. Deploy the companion backend using the repository installation guide and its Docker Compose example. The backend and Python dependencies must be installed on the same host as the VS Code extension.
2. Install this extension in that environment. With VS Code Remote, install it on the remote host.
3. If you use a different installation path, set trinityAgent.backendPath in User settings to the absolute path of scripts/trinity-agent.sh. The launcher must implement the bundled JSON-lines protocol and load the backend dependencies and secrets.
4. Select trinityAgent.provider in User settings (ovhcloud, deepseek or openai) or run TRINITY Agent: Choose the inference provider, then configure that provider's API key through the extension command. The supplied Docker deployment persists it in its configured secret directory; never commit it to Git. A provider change restarts an idle backend and applies to the next message; an active task finishes first.
5. Open a trusted workspace and start a conversation. The backend can read and modify files and run commands with the permissions of its container. Review the access mode and container mounts.

The extension itself is free and licensed under 0BSD. API usage and any hosting are charged by their providers. OpenAI Codex uses the Responses API. Its configured monthly allowance and locally observed token cost are estimates; the OpenAI Platform billing page remains authoritative. No model subscription or API credit is included with this extension.

With DeepSeek, local inference budgets, request-count limits, configured context/output token caps, repeated-tool-failure cutoffs and model request timeouts are disabled. Tool corrections are no longer cut off after two SDK retries. The provider's context/output capacities, account balance and rate limits still apply. File backups, permission checks and verification remain active. Automatic routing starts economically; Flash and Pro are available for explicit selection. Costs use conservative peak/cache-miss estimates and are displayed in euros. Both USD and CNY account balances are fetched from DeepSeek, converted using dated ECB reference rates and summed without counting granted credit twice. Exchange rates are refreshed every six hours; a cached rate up to seven days old can be used during an outage and is marked stale. Missing rates show conversion as unavailable. Native amounts remain unchanged in accounting records. Keys, model caches and usage accounting are separate from OVH.

⬛ Features

• Model selection and context compaction with local Markdown memory.
• Paste screenshots with Ctrl+V / Cmd+V, including clipboard-item images and accompanying text. Up to five images; preparation completes before sending.
• Optional local catalogue evaluations guide automatic routing by task domain, accuracy, measured latency and cost. Results appear in conversation archives and expire after 14 days.
• Perf. Auto / Perf. Max: switch local CPU/RAM allocation from the composer, including during a task. Max replaces fixed cache/reader/search caps with hardware budgets and adapts to container limits and memory pressure. The choice persists; API budgets and verification remain active.
• When a user challenges an answer or reports an error/doubt, prioritize official documentation, then official-repository issues. Detected challenges require a search attempt and consultation/citation of an available source; failed research is explicitly labelled incomplete. Explicit offline requests are respected.
• Executable verification for tracked file edits: syntax checks, saved originals, reference checks before/after changes, and a final-answer gate. Failed unverified edits can be restored without resetting existing user work.
• Check receipts and explicit verification status in the activity log and conversation archives. A passing result covers the configured checks; it does not certify every model claim or arbitrary shell/SSH effects. Without a functional test policy, tracked edits receive backup, syntax and content checks with their limited coverage explicitly reported. An existing config/trinity-agent-checks.json remains authoritative.
• Local syntax highlighting with a dark Monokai theme, including restored conversations; full-width tables with subtle alternating rows.
• Persistent conversation history, deletion and interrupted-task resumption.
• Send a message while a task runs to queue it; edit, remove or promote pending messages to steering. Stop remains the interruption control.
• Drag pending queued/steering tasks up or down, or use Alt + ↑ / ↓ on their grip. Order is saved without restarting a stopped queue. Steering keeps its next-step behavior; queued tasks run afterward in their saved order.
• Steer task applies the instruction before the next model call, after the current call/tool completes. It does not roll back running commands.
• Pending text messages survive restarts and require explicit queue continuation after a stop/restart. Images are supported in ordinary prompts, not in queued messages.
• Access changes apply to the next request; the running task retains its displayed active permissions. Resuming a task uses the currently selected access and model. Older clients that omit access on resume default to restricted access.
• French, English, German and Spanish interface; automatic locale or Docker override.
• Configurable names, colors and custom logos in self-hosted deployments.

⬛ Privacy and storage

Prompts, relevant workspace contents, tool results and images can be sent to the configured AI endpoint. Conversations, task state and the follow-up queue are stored on the backend's persistent volume. Deleting a local conversation does not delete provider-side logs. The extension does not include a separate analytics service.

⬛ Source, issues and licensing

GitHub · Codeberg · Issues

Original code: 0BSD. Bootstrap Icons: MIT. Highlight.js and its bundled Monokai theme: BSD-3-Clause. The Marketplace icon is the TRINITY logo. The history header uses Font Awesome Free's hexagon-nodes icon under CC BY 4.0. DOMPurify, Markdown-it and other dependencies retain their respective licenses; license files are included in the package.

Display style is configurable through TRINITY_AGENT_EMOJI_MODE:

🚫 0 disables decorative emojis.
✅ 1 allows only ✅ for verified successes, ⭕ for errors and ➡️ for next steps. This is the default.
✨ 2 uses all supported emojis sparingly.
🎉 3 uses emojis expressively.

Code and quotations remain exact. Activity labels do not use emojis; the setting applies to written progress messages and answers. Icons match their actions, the cursor follows the primary color, verification labels and the disclaimer use 50% opacity, and diffs color the full line. These settings and the newer controls also require an up-to-date backend.

⬛ Markdown memory

In the project selector, Edit project memory opens the manually maintained NOTES.md. Aggregate project memories lets you select several projects and opens their AGGREGATE.md, containing notes, recent excerpts and source links. These actions do not call the model.

The MEMORY.md and INDEX.md views are generated; edit NOTES.md or the source notes instead. A compressed local lexical index and RAM cache speed up searches across conversations, while Markdown archives remain on persistent storage. Search stays targeted and can be partial. Notes from another project are not injected automatically, except for user-scoped preferences.

⬛ Conversation organization

History groups only the conversations you choose into collapsible folders. Drag a folder handle or title to reorder it; the marker shows whether it will be inserted before or after the hovered folder. With focus on the handle, Alt + ↑ / ↓ also moves the folder.

Drag a conversation onto a folder, including a collapsed folder, to add or move it. Drop it on Conversations outside projects to remove its folder assignment; this area appears while dragging even when empty. The organization control remains keyboard accessible. Order and assignments are saved automatically, and the target folder opens after the drop. Opening a conversation restores its workspace and context. Adding or reordering a folder does not change conversation context. Opening an archived conversation restores its working directory; a new conversation starts in the initial VS Code workspace and does not inherit the previous directory. Folder colors, titles and paths are applied only after Save. Closing the editor or clicking outside it discards unsaved changes; the folder keeps its saved color.

The project editor includes a red Delete button. Confirmation permanently deletes the history folder and all conversations assigned to it, including their attachments. Project files and manual notes are preserved, as are conversations assigned elsewhere. An active conversation is stopped before deletion; failures are reported without claiming that the entire folder was removed.

The Markdown file list is collapsed by default and expands on demand. Project folders also show root Markdown paths and available memory views. REGRESSIONS.md stores sourced historical check results; rerun the tests to validate the current state. Project files remain their own sources of truth.

⬛ Project memory before inference

Before calling the model, automatic memory retrieval selects sourced user preferences and dated execution receipts. Previous assistant answers, transcripts and hypotheses remain available through explicit memory searches but are not injected automatically. Current project files are labelled as current sources. Instruction files already loaded in full are excluded from automatic recall by their canonical path; they remain indexed and explicitly readable. Changes to known Markdown sources invalidate their cached index immediately. Historical receipts are not proof that today's code still passes its checks.

An exact, standalone informational question matching a Markdown heading in a root .md file or the project's NOTES.md can be answered locally. For example:

## How do I run the project tests?

Run `python3 -m unittest discover` from the project root.

The agent quotes the existing answer with the source path and modification date, records it in conversation history, and reports zero model requests and zero API cost. It does not execute the documented command. Matching only normalizes case and whitespace. Conflicting answers, partial reads, images, task resumption, requests for current verification, or approximate matches use the normal agent path with relevant memory. Source files are reread through metadata-validated caches; generated summaries and old model answers cannot become direct answers. The local lookup is bounded to avoid blocking on slow mounts. No budgets change.

⬛ Contributing to the extension flow

The repository also provides opt-in GEPA prompt optimization and recursive-context (RLM) experiments. They compare candidates against fixed checks and measure OVH calls, usage and latency; they do not activate new chat behavior or apply generated prompts automatically. See the repository research experiments for setup, limits and evaluation commands.

See Architecture and contribution for the message flow, module responsibilities, cache rules and minified frontend generation.

⬛ Provider-based budgets (1.14.3)

Update the companion backend together with the extension. Configure OVH billing access through the balance control before sending paid requests. The budget now uses the OVH Public Cloud project's reported consumption and available credit; local model-token estimates are informational only. This includes other services charged to that same cloud project, not just this agent.

Completed or failed requests release their temporary reservation. Unknown usage is never converted into a maximum-cost expense. The previous usage-eur.json ledger is preserved for audit and is no longer used to authorize requests; the provider observations are stored separately in budgets/usage-provider-eur.json.

The monthly counter comes from OVH. Daily and weekly counters track changes observed since monitoring began, with partial and since metadata: the API's monthly snapshot cannot reconstruct prior daily spending. Top-ups do not count as consumption. Existing daily/weekly/monthly limits and credit expiry remain.

OVH usage is delayed; these checks are not a real-time provider-enforced quota. No requests are authorized from unavailable billing data, a cache older than five minutes, an OVH usage report older than 24 hours, or a report for another month.