zesdex/docs/CODEMAPS/backend.md

<!-- Generated: 2026-07-12 | Files scanned: 124 | Token estimate: ~850 -->

# Backend / Service Layer

## AI Provider

`src/service/provider.rs` (310 lines)
- `LlmClient::new(api_key, model, base_url)` — constructs blocking reqwest client
- `chat_with_tools()` — non-streaming with tool definitions
- `chat_stream()` — SSE streaming, returns `SseParser` yielding `StreamEvent`
- Retry logic: up to 3 attempts on transient errors, exponential backoff

## OAuth

`src/service/oauth/manager.rs` (113 lines) + `loopback.rs` + `pkce.rs`
- PKCE flow: `CodeVerifier` → challenge → browser auth → loopback server → token exchange
- Configurable via `app_config.json` provider definitions (auth URL, token URL, scopes)

## IPC / Daemon

`src/ipc/` (7 files, ~350 lines total)
- Unix domain socket, length-prefixed JSON frames
- Daemon sends `DaemonFrame` (state payload, stream tokens, system notes)
- Clients send `ClientRequest` (key presses, resize, submit, scroll)
- State sync uses full-state push from daemon to client after each action

## Workflow Engine

`src/app/workflow/engine.rs` (648 lines) + `script.rs`
- Inline JS-style DSL executed by a lightweight runtime
- `agent()`, `parallel()`, `pipeline()`, `phase()`, `log()` — spawns sub-agents
- Max concurrency configurable via `workflow_max_concurrency` setting
- Hive-mind orchestrator in `hive_mind.rs`: Core Intelligence compiles a `CognitiveCyclePlan` per task — cycle count and nodes-per-cycle are decided fresh each time based on what the task actually needs

## Sub-Agent System

`src/app/subagent/` (6 files: `spawn.rs`, `engine.rs`, `context.rs`, `event.rs`, `division.rs`, `auto.rs`, ~450 lines total)
- `run_subagent()` — spawns independent agent with its own tool set and context
- Communicates via `mpsc<SubagentEvent>` channel (tool calls, results, completion)
- Uses `LlmClient` (same as main agent) with tool-use API
- Auto-healing: on build/test failure, spawns auto-fix sub-agent
- Node access tiers (`division.rs`'s `tool_scope` module): `read`, `write`, `full` — granted per node by the Core Intelligence based on what its directive needs

## MCP Client

`src/app/mcp/manager.rs` (441+ lines)
- Stdio transport: spawns child process, JSON-RPC via stdin/stdout
- HTTP transport: streaming HTTP with JSON-RPC
- Dynamic tool list refresh and error recovery
- Persistent child handle for stdio (reuses connection across calls)

## Self-Review

`src/app/review/mod.rs` (495 lines)
- Post-tool execution quality check against learned lessons
- Invokes `run_subagent()` with reviewer prompt
- Staleness detection: skips review after N consecutive empty results
- Three review types: code quality, architecture, security

## Background Bash

`src/app/bgbash/` (2 files: `job.rs`, `control.rs`)
- `spawn_bash_job()` — runs `sh -c` in a thread, collects stdout line-by-line
- Channels: output via `mpsc<String>`, PID via `mpsc<u32>`
- Killable via PID (SIGTERM)
- Output buffering capped at 10,000 lines to prevent memory issues

## Gate Guard / Harness

`src/app/harness.rs` (495 lines)
- `Harness::gate_tool_call()` — verdict-based tool gating (allow/block)
- Path traversal, credential read, and destructive command detection
- Pattern detection for stub code, denial language, and assumptions in write/edit content
- Reason validation for mutating tools (minimum 8 characters, rejects generic non-answers)
- Includes 8 unit tests for verdict parsing formats