feat(backends): multi-backend intelligence layer — LLM + Agent backends, persona-loading agents, honest CLI commands
Add IntelligenceBackend abstraction with two categories: - LLMBackend (OllamaLocal, OllamaCloud): CI runs tool loop, provides tools, constructs prompts - AgentBackend (Opencode): agent runs own tool loop, CI serializes request Refactor all 18 agents from hardcoded stubs to persona loaders that delegate to the active backend or fail honestly when no backend is available. Refactor OrchestratorAgent.executeStage() from monolithic switch to agent delegation via STAGE_AGENT_MAP for intelligent stages (research, plan, execute, verify), with mechanical stages (specify, clarify, complete) staying inline. Wire CLI commands with --backend flag and auto-detection (opencode → ollama-local → ollama-cloud). Harden rollback/ship with real git operations. No command returns fake success.
This commit is contained in:
CI
@@ -0,0 +1,317 @@
|
||||
import { execSync } from "node:child_process";
|
||||
import * as fs from "node:fs";
|
||||
import * as path from "node:path";
|
||||
import * as os from "node:os";
|
||||
import {
|
||||
IntelligenceBackend,
|
||||
BackendRequest,
|
||||
BackendResult,
|
||||
BackendType,
|
||||
LLMBackendConfig,
|
||||
TokenUsage,
|
||||
Artifact,
|
||||
emptyTokenUsage,
|
||||
emptyBackendResult,
|
||||
} from "./types.js";
|
||||
import { AgentName, ModelProfile } from "../types/config.js";
|
||||
import { Decision } from "../types/decisions.js";
|
||||
import { Escalation } from "../types/escalation.js";
|
||||
import { ToolRegistry, ToolCall, ToolResult } from "./tool-registry.js";
|
||||
|
||||
const MAX_TOOL_ROUNDS = 50;
|
||||
|
||||
export abstract class OllamaBaseBackend implements IntelligenceBackend {
|
||||
abstract readonly name: string;
|
||||
readonly type: BackendType = "llm";
|
||||
|
||||
protected config: LLMBackendConfig;
|
||||
protected projectPath: string;
|
||||
|
||||
constructor(config: LLMBackendConfig | undefined) {
|
||||
this.config = config || { base_url: "http://localhost:11434", model_profile: "balanced" };
|
||||
this.projectPath = process.cwd();
|
||||
}
|
||||
|
||||
abstract isAvailable(): Promise<boolean>;
|
||||
|
||||
async execute(request: BackendRequest): Promise<BackendResult> {
|
||||
const startTime = Date.now();
|
||||
|
||||
try {
|
||||
const personaContent = this.loadPersona(request.persona);
|
||||
const workflowContent = this.loadWorkflow(request.workflow);
|
||||
const model = this.resolveModel();
|
||||
|
||||
const toolRegistry = new ToolRegistry(request.context.project_path);
|
||||
|
||||
const messages: OllamaMessage[] = [];
|
||||
messages.push({
|
||||
role: "system",
|
||||
content: this.buildSystemPrompt(personaContent, workflowContent, request),
|
||||
});
|
||||
messages.push({
|
||||
role: "user",
|
||||
content: request.task,
|
||||
});
|
||||
|
||||
let totalInputTokens = 0;
|
||||
let totalOutputTokens = 0;
|
||||
let round = 0;
|
||||
const allArtifacts: Artifact[] = [];
|
||||
const allDecisions: Decision[] = [];
|
||||
const allEscalations: Escalation[] = [];
|
||||
|
||||
while (round < MAX_TOOL_ROUNDS) {
|
||||
round++;
|
||||
const response = await this.callModel(messages, model, toolRegistry);
|
||||
|
||||
totalInputTokens += response.usage?.prompt_tokens || 0;
|
||||
totalOutputTokens += response.usage?.completion_tokens || 0;
|
||||
|
||||
const assistantContent = response.choices?.[0]?.message?.content || "";
|
||||
const toolCalls = response.choices?.[0]?.message?.tool_calls;
|
||||
|
||||
messages.push({
|
||||
role: "assistant",
|
||||
content: assistantContent,
|
||||
tool_calls: toolCalls,
|
||||
});
|
||||
|
||||
if (!toolCalls || toolCalls.length === 0) {
|
||||
return this.parseFinalResponse(assistantContent, allArtifacts, allDecisions, allEscalations, {
|
||||
input_tokens: totalInputTokens,
|
||||
output_tokens: totalOutputTokens,
|
||||
total_tokens: totalInputTokens + totalOutputTokens,
|
||||
estimated_cost_usd: 0,
|
||||
});
|
||||
}
|
||||
|
||||
for (const toolCall of toolCalls) {
|
||||
const call: ToolCall = {
|
||||
name: toolCall.function.name,
|
||||
arguments: JSON.parse(toolCall.function.arguments),
|
||||
};
|
||||
const result = toolRegistry.execute(call);
|
||||
messages.push({
|
||||
role: "tool",
|
||||
name: call.name,
|
||||
content: result.content,
|
||||
});
|
||||
|
||||
if (call.name === "writeFile" && !result.isError) {
|
||||
allArtifacts.push({
|
||||
path: String(call.arguments.path),
|
||||
content: String(call.arguments.content),
|
||||
operation: "create",
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
const finalContent = messages
|
||||
.filter((m) => m.role === "assistant" && m.content)
|
||||
.map((m) => m.content)
|
||||
.join("\n");
|
||||
|
||||
return this.parseFinalResponse(
|
||||
`Tool loop reached maximum rounds (${MAX_TOOL_ROUNDS}). Partial progress:\n${finalContent}`,
|
||||
allArtifacts,
|
||||
allDecisions,
|
||||
allEscalations,
|
||||
{ input_tokens: totalInputTokens, output_tokens: totalOutputTokens, total_tokens: totalInputTokens + totalOutputTokens, estimated_cost_usd: 0 }
|
||||
);
|
||||
} catch (err) {
|
||||
return emptyBackendResult(`Backend execution failed: ${err instanceof Error ? err.message : String(err)}`);
|
||||
}
|
||||
}
|
||||
|
||||
protected abstract callModel(
|
||||
messages: OllamaMessage[],
|
||||
model: string,
|
||||
toolRegistry: ToolRegistry
|
||||
): Promise<OllamaChatResponse>;
|
||||
|
||||
protected abstract resolveModel(): string;
|
||||
|
||||
protected buildSystemPrompt(persona: string, workflow: string, request: BackendRequest): string {
|
||||
const parts = [persona];
|
||||
if (workflow) {
|
||||
parts.push("", "## Workflow Instructions", workflow);
|
||||
}
|
||||
parts.push(
|
||||
"",
|
||||
"## Execution Context",
|
||||
`Autonomy level: ${request.autonomy}`,
|
||||
`Project path: ${request.context.project_path}`,
|
||||
`Phase: ${request.context.phase}`,
|
||||
`Stage: ${request.context.stage}`,
|
||||
"",
|
||||
"## Output Format",
|
||||
"When you have completed your task, output a JSON object with this structure:",
|
||||
"```json",
|
||||
'{',
|
||||
' "success": true,',
|
||||
' "output": "Summary of what was accomplished",',
|
||||
' "artifacts": [{"path": "file/path", "content": "...", "operation": "create"}],',
|
||||
' "decisions": [{"id": "D-NNN", "decision": "what", "rationale": "why", "confidence": 0.85, "category": "general", "alternatives_considered": [], "learnship_equivalent": "", "human_override": null, "timestamp": ""}],',
|
||||
' "escalations": []',
|
||||
'}',
|
||||
"```"
|
||||
);
|
||||
return parts.join("\n");
|
||||
}
|
||||
|
||||
protected loadPersona(persona: AgentName): string {
|
||||
const candidates = [
|
||||
path.join(os.homedir(), ".config", "opencode", "agents", `ci-${persona}.md`),
|
||||
path.join(process.cwd(), "opencode", "agents", `ci-${persona}.md`),
|
||||
];
|
||||
for (const candidate of candidates) {
|
||||
if (fs.existsSync(candidate)) {
|
||||
return fs.readFileSync(candidate, "utf-8");
|
||||
}
|
||||
}
|
||||
return `You are the CI ${persona} agent. Execute the requested task thoroughly and autonomously.`;
|
||||
}
|
||||
|
||||
protected loadWorkflow(workflow: string): string {
|
||||
const candidates = [
|
||||
path.join(os.homedir(), ".config", "opencode", "ci", "workflows", `${workflow}.md`),
|
||||
path.join(process.cwd(), "opencode", "workflows", `${workflow}.md`),
|
||||
];
|
||||
for (const candidate of candidates) {
|
||||
if (fs.existsSync(candidate)) {
|
||||
return fs.readFileSync(candidate, "utf-8");
|
||||
}
|
||||
}
|
||||
return "";
|
||||
}
|
||||
|
||||
protected parseFinalResponse(
|
||||
content: string,
|
||||
artifacts: Artifact[],
|
||||
decisions: Decision[],
|
||||
escalations: Escalation[],
|
||||
usage: TokenUsage
|
||||
): BackendResult {
|
||||
const jsonMatch = content.match(/\{[\s\S]*"success"[\s\S]*\}/);
|
||||
if (jsonMatch) {
|
||||
try {
|
||||
const parsed = JSON.parse(jsonMatch[0]);
|
||||
return {
|
||||
success: parsed.success ?? true,
|
||||
output: parsed.output || content,
|
||||
artifacts: parsed.artifacts?.length ? this.parseArtifacts(parsed.artifacts) : artifacts,
|
||||
decisions: parsed.decisions?.length ? this.parseDecisions(parsed.decisions) : decisions,
|
||||
escalations: parsed.escalations?.length ? this.parseEscalations(parsed.escalations) : escalations,
|
||||
usage,
|
||||
};
|
||||
} catch {}
|
||||
}
|
||||
|
||||
return {
|
||||
success: true,
|
||||
output: content,
|
||||
artifacts,
|
||||
decisions,
|
||||
escalations,
|
||||
usage,
|
||||
};
|
||||
}
|
||||
|
||||
private parseArtifacts(raw: unknown[]): Artifact[] {
|
||||
return raw.filter((a): a is Record<string, unknown> => !!a).map((a) => ({
|
||||
path: String(a.path || ""),
|
||||
content: String(a.content || ""),
|
||||
operation: (a.operation as Artifact["operation"]) || "create",
|
||||
}));
|
||||
}
|
||||
|
||||
private parseDecisions(raw: unknown[]): Decision[] {
|
||||
return raw.filter((d): d is Record<string, unknown> => !!d).map((d) => ({
|
||||
id: String(d.id || "D-000"),
|
||||
decision: String(d.decision || ""),
|
||||
rationale: String(d.rationale || ""),
|
||||
confidence: Number(d.confidence || 0.5),
|
||||
category: (d.category as Decision["category"]) || "general",
|
||||
alternatives_considered: Array.isArray(d.alternatives_considered)
|
||||
? d.alternatives_considered.map((a: unknown) =>
|
||||
typeof a === "string"
|
||||
? { option: a, rejected_reason: "" }
|
||||
: (a as { option: string; rejected_reason: string })
|
||||
)
|
||||
: [],
|
||||
learnship_equivalent: String(d.learnship_equivalent || ""),
|
||||
human_override: d.human_override ? String(d.human_override) : null,
|
||||
timestamp: String(d.timestamp || new Date().toISOString()),
|
||||
}));
|
||||
}
|
||||
|
||||
private parseEscalations(raw: unknown[]): Escalation[] {
|
||||
return raw.filter((e): e is Record<string, unknown> => !!e).map((e) => ({
|
||||
id: String(e.id || "E-000"),
|
||||
timestamp: String(e.timestamp || new Date().toISOString()),
|
||||
type: (e.type as Escalation["type"]) || "specification_ambiguity",
|
||||
phase: String(e.phase || ""),
|
||||
description: String(e.description || ""),
|
||||
context: String(e.context || ""),
|
||||
options: Array.isArray(e.options) ? e.options : [],
|
||||
default_option_id: String(e.default_option_id || ""),
|
||||
resolution: (e.resolution as Escalation["resolution"]) || "pending",
|
||||
audit_file: String(e.audit_file || ""),
|
||||
}));
|
||||
}
|
||||
|
||||
protected modelProfileToModel(profile: ModelProfile, availableModels: string[]): string {
|
||||
if (availableModels.length === 0) return "llama3.1";
|
||||
|
||||
const sorted = [...availableModels].sort((a, b) => a.length - b.length);
|
||||
switch (profile) {
|
||||
case "speed":
|
||||
return sorted[0];
|
||||
case "quality":
|
||||
return sorted[sorted.length - 1];
|
||||
case "balanced":
|
||||
default:
|
||||
return sorted[Math.floor(sorted.length / 2)] || sorted[0];
|
||||
}
|
||||
}
|
||||
|
||||
protected async fetchAvailableModels(): Promise<string[]> {
|
||||
try {
|
||||
const response = await fetch(`${this.config.base_url}/api/tags`);
|
||||
if (!response.ok) return [];
|
||||
const data = await response.json() as { models?: Array<{ name: string }> };
|
||||
return (data.models || []).map((m) => m.name);
|
||||
} catch {
|
||||
return [];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
interface OllamaMessage {
|
||||
role: "system" | "user" | "assistant" | "tool";
|
||||
content: string;
|
||||
name?: string;
|
||||
tool_calls?: Array<{
|
||||
function: { name: string; arguments: string };
|
||||
}>;
|
||||
}
|
||||
|
||||
interface OllamaChatResponse {
|
||||
choices?: Array<{
|
||||
message: {
|
||||
content: string;
|
||||
tool_calls?: Array<{
|
||||
function: { name: string; arguments: string };
|
||||
}>;
|
||||
};
|
||||
}>;
|
||||
usage?: {
|
||||
prompt_tokens: number;
|
||||
completion_tokens: number;
|
||||
total_tokens: number;
|
||||
};
|
||||
}
|
||||
|
||||
export { OllamaMessage, OllamaChatResponse };
|
||||
Reference in New Issue
Block a user