guanghu/docs/adr/0093-shared-cli-agent-runtime-adapters.md
冰朔 8739805f99
Some checks failed
Auto-update PR branches / Update open PR branches (push) Has been cancelled
CI / Frontend Static Quality Checks (push) Has been cancelled
CI / Frontend Tests & Coverage (push) Has been cancelled
CI / Rust Tests & Quality Checks (push) Has been cancelled
CI / Linux build verification (push) Has been cancelled
Release (Alpha) / Compute alpha version (push) Has been cancelled
Release (Alpha) / Build release artifacts (push) Has been cancelled
Release (Alpha) / GitHub Release (alpha) (push) Has been cancelled
Release (Alpha) / Update docs and release pages (push) Has been cancelled
Deploy docs / Build VitePress site (push) Has been cancelled
Deploy docs / Deploy to GitHub Pages (push) Has been cancelled
光湖开源源码快照 · Tolaria AGPL 分叉基线 · 独立更新链
2026-07-05 17:45:16 +08:00

36 lines
2.8 KiB
Markdown

---
type: ADR
id: "0093"
title: "Shared CLI agent runtime adapters"
status: active
date: 2026-04-29
---
## Context
Tolaria supports Claude Code, Codex, OpenCode, and Pi as local CLI agents in the AI panel. Each agent has different command-line arguments, configuration shape, and JSON event schema, but the Rust backend had grown repeated runtime plumbing around those differences: request shapes, prompt wrapping, subprocess launch, stdout JSON reading, stderr capture, exit handling, done events, version probing, and Tolaria MCP server path resolution.
That duplication made small runtime fixes expensive because they had to be repeated across several adapter files. It also kept Codex-specific command and event mapping inside `ai_agents.rs`, making the top-level module both an orchestrator and a bespoke adapter.
## Decision
**Tolaria uses `cli_agent_runtime.rs` as the shared runtime scaffold for app-managed CLI agents, while `ai_agents.rs` only normalizes and dispatches requests to per-agent adapter modules.**
The shared scaffold owns the common agent request shape, system/user prompt wrapping, JSON-line process lifecycle, normalized error/done handling for `AiAgentStreamEvent` adapters, version probing, and Tolaria MCP server path resolution. Per-agent modules keep the provider-specific pieces: binary discovery candidates, command arguments, transient config shape, authentication error wording, and JSON event mapping.
Codex now lives in `codex_cli.rs`, matching the Claude, OpenCode, and Pi adapter boundary. `ai_agents.rs` remains the Tauri-facing orchestrator that chooses an adapter and maps Claude's legacy event enum into the normalized event stream.
## Options Considered
- **Shared runtime scaffold with thin adapters** (chosen): reduces repeated process lifecycle code without hiding provider-specific command/config/event behavior.
- **One trait object per agent**: more uniform on paper, but adds indirection without removing much current complexity.
- **Leave each adapter self-contained**: keeps local readability for a single file, but new process, prompt, and MCP fixes continue to land in parallel.
- **Fully generic event mapping**: over-abstracts the JSON schemas and makes provider-specific edge cases harder to test.
## Consequences
- Runtime lifecycle fixes should usually start in `cli_agent_runtime.rs`.
- New agent adapters should use the shared request/prompt/process helpers and keep only command, config, discovery, and event mapping local.
- `ai_agents.rs` should not grow provider-specific runtime code again; it should normalize the frontend request, dispatch to an adapter, and map any legacy event shape.
- The shared scaffold deliberately does not erase provider differences. Authentication messages, permission semantics, and transient config formats remain adapter-owned and must stay covered by adapter tests.