Agents and personas

TDE does not talk to models

This is the single most important thing to understand about agents in TDE: there is no in-process model client. TDE does not hold your API keys and call Anthropic or OpenAI itself. Every agent run shells out to a coding CLI you have installed, and that CLI owns the conversation, the tool loop and the billing.

Four CLIs back agent execution:

Two more selectable runtimes reuse the Claude Code binary and only change the model and endpoint environment: TDE Pass and Ollama (Local), the latter pointed at a local daemon on 127.0.0.1:11434.

Connecting a runtime

Control → AI & Automation → LLM lists the four CLIs plus a local runtime card. Its intro is literal about the scope: AI Chat & My Loop run these four official agent CLIs — nothing else.

Each card shows a status pill and one action button:

Auth is whatever the CLI itself uses. Connect runs the CLI's own login command in a terminal window — claude auth login, codex login --device-auth, cursor-agent login, opencode auth login — or lets you paste an API key.

Login state is detected by looking for the CLI's own credential file, without spawning anything:

An injected API key is read from the process environment first, then from the encrypted vault at ~/.tde/secrets.db, and handed to the CLI as ANTHROPIC_API_KEY, OPENAI_API_KEY, CURSOR_API_KEY or OPENCODE_GO_API_KEY. A blank vault value means cleared — TDE will never send an empty key.

Statuses are probed automatically when you open the panel, which also checks each installed binary's version against its registry for updates.

Personas

A persona is a named, reusable agent identity: a base CLI, a model, a reasoning effort, a system prompt, a set of skills and an appearance. Personas live in ~/.tde/personas.db and are edited in Control → AI & Automation → Agents.

Web search is a per-persona policy, not a per-conversation toggle. It applies everywhere that persona runs — pipeline stages, Bot, squads and AI Chat when that persona is selected:

Two personas are bootstrapped and cannot be renamed or deleted:

• Default — created from your existing configuration on first run. Existing configuration always wins; TDE never resets a runtime or model you chose.
• Bot — the reserved name for the agent assigned to scheduled automations. It is an ordinary, fully editable persona; the name is kept stable so persisted bots point at an agent identity rather than at a raw model.

Picking a model

The model list a persona offers depends on its base runtime:

Two rules protect you from silent substitution. A model id you typed with an explicit local/ or cloud/ prefix is preserved verbatim, because falling back to another provider would change who pays for the run. And a raw id from a live catalog is kept while that catalog is empty, so a transient probe failure never rewrites your persona.

At run time the model is resolved in this order: the persona's model, then the pipeline's per-category routing, then the stage default. Near the task budget the model is dropped entirely so the CLI's cheaper default carries the run.

How a persona becomes a run

Each pipeline stage may name a persona. When a task enters that stage, TDE builds a spec from the persona — base CLI, name, system prompt, skills, API key, model, effort, web-search policy — and invokes the CLI:

claude -p "<prompt>" --model claude-sonnet-5 --effort high --no-session-persistence
codex exec --json "<prompt>" -m gpt-5.5 -c model_reasoning_effort="high"
cursor-agent --print --trust --output-format stream-json "<prompt>"
opencode run --format json "<prompt>"

Read-only stages add each CLI's own restriction flags — plan mode and a Read,Glob,Grep,WebFetch,WebSearch tool allowlist for Claude, --sandbox read-only for Codex, --mode ask --sandbox enabled for Cursor, a pure read-only agent profile for OpenCode.

Attachments before the agent starts

When a run is started with image or document attachments — from AI Chat or from a My Loop task — TDE resolves them before the CLI tool loop begins. That matters most on TDE Pass:

• Images (PNG, JPEG, WebP, GIF, up to 20 MiB) are sent as real multimodal content to the Pass gateway. The model id you selected stays unchanged; if that model has no vision, the gateway applies its configured vision fallback. The verified description is then handed to the agent as context, so the agent does not need to Read the image file.
• Documents (PDF, DOCX, XLSX, XLS, ODS) are extracted natively. Small extractions are inlined into the prompt (bounded); large ones are truncated in the prompt and chunked into the task's Recall overlay so later stages can retrieve narrower excerpts.
• Other binaries stay path references for the agent's normal file tools.

Identical tool calls with the same input are circuit-broken after a few repeats, on both AI Chat and My Loop agents, so a stuck loop cannot burn through a Pass budget unnoticed.

Where personas show up

• Pipeline stages — the agent that runs that stage.
• Task detail — the Agent: field, which overrides the stage default for that task.
• Bots — every bot names the persona that carries its action.
• AI Chat — the persona you are talking to. Conversations are independent: one chat can keep streaming while you send in another, Stop cancels only the open conversation, and closing the AI Chat window does not cancel in-flight runs.
• Council — the chairing persona and its counselor bench.

From a shell, tde agent list prints your configured agents with provider and model and marks the default with *. It is a catalog read, unrelated to the desktop Agent API.